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
  48 #include "gromacs/awh/awh.h"
  49 #include "gromacs/domdec/dlbtiming.h"
  50 #include "gromacs/domdec/domdec.h"
  51 #include "gromacs/domdec/domdec_struct.h"
  52 #include "gromacs/domdec/gpuhaloexchange.h"
  53 #include "gromacs/domdec/partition.h"
  54 #include "gromacs/essentialdynamics/edsam.h"
  55 #include "gromacs/ewald/pme.h"
  56 #include "gromacs/ewald/pme_pp.h"
  57 #include "gromacs/ewald/pme_pp_comm_gpu.h"
  58 #include "gromacs/gmxlib/network.h"
  59 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  60 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  61 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  62 #include "gromacs/gpu_utils/gpu_utils.h"
  63 #include "gromacs/imd/imd.h"
  64 #include "gromacs/listed_forces/disre.h"
  65 #include "gromacs/listed_forces/gpubonded.h"
  66 #include "gromacs/listed_forces/listed_forces.h"
  67 #include "gromacs/listed_forces/manage_threading.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/enerdata_utils.h"
  78 #include "gromacs/mdlib/force.h"
  79 #include "gromacs/mdlib/forcerec.h"
  80 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  81 #include "gromacs/mdlib/qmmm.h"
  82 #include "gromacs/mdlib/update.h"
  83 #include "gromacs/mdtypes/commrec.h"
  84 #include "gromacs/mdtypes/enerdata.h"
  85 #include "gromacs/mdtypes/forceoutput.h"
  86 #include "gromacs/mdtypes/iforceprovider.h"
  87 #include "gromacs/mdtypes/inputrec.h"
  88 #include "gromacs/mdtypes/md_enums.h"
  89 #include "gromacs/mdtypes/simulation_workload.h"
  90 #include "gromacs/mdtypes/state.h"
  91 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  92 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  93 #include "gromacs/nbnxm/nbnxm.h"
  94 #include "gromacs/nbnxm/nbnxm_gpu.h"
  95 #include "gromacs/pbcutil/ishift.h"
  96 #include "gromacs/pbcutil/mshift.h"
  97 #include "gromacs/pbcutil/pbc.h"
  98 #include "gromacs/pulling/pull.h"
  99 #include "gromacs/pulling/pull_rotation.h"
 100 #include "gromacs/timing/cyclecounter.h"
 101 #include "gromacs/timing/gpu_timing.h"
 102 #include "gromacs/timing/wallcycle.h"
 103 #include "gromacs/timing/wallcyclereporting.h"
 104 #include "gromacs/timing/walltime_accounting.h"
 105 #include "gromacs/topology/topology.h"
 106 #include "gromacs/utility/arrayref.h"
 107 #include "gromacs/utility/basedefinitions.h"
 108 #include "gromacs/utility/cstringutil.h"
 109 #include "gromacs/utility/exceptions.h"
 110 #include "gromacs/utility/fatalerror.h"
 111 #include "gromacs/utility/fixedcapacityvector.h"
 112 #include "gromacs/utility/gmxassert.h"
 113 #include "gromacs/utility/gmxmpi.h"
 114 #include "gromacs/utility/logger.h"
 115 #include "gromacs/utility/smalloc.h"
 116 #include "gromacs/utility/strconvert.h"
 117 #include "gromacs/utility/sysinfo.h"
 118
 119 using gmx::AtomLocality;
 120 using gmx::DomainLifetimeWorkload;
 121 using gmx::ForceOutputs;
 122 using gmx::InteractionLocality;
 123 using gmx::SimulationWorkload;
 124 using gmx::StepWorkload;
 125
 126 // TODO: this environment variable allows us to verify before release
 127 // that on less common architectures the total cost of polling is not larger than
 128 // a blocking wait (so polling does not introduce overhead when the static
 129 // PME-first ordering would suffice).
 130 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 131
 132 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 133 {
 134     const int end = forceToAdd.size();
 135
 136     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 137 #pragma omp parallel for num_threads(nt) schedule(static)
 138     for (int i = 0; i < end; i++)
 139     {
 140         rvec_inc(f[i], forceToAdd[i]);
 141     }
 142 }
 143
 144 static void calc_virial(int                              start,
 145                         int                              homenr,
 146                         const rvec                       x[],
 147                         const gmx::ForceWithShiftForces& forceWithShiftForces,
 148                         tensor                           vir_part,
 149                         const t_graph*                   graph,
 150                         const matrix                     box,
 151                         t_nrnb*                          nrnb,
 152                         const t_forcerec*                fr,
 153                         PbcType                          pbcType)
 154 {
 155     /* The short-range virial from surrounding boxes */
 156     const rvec* fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
 157     calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, pbcType == PbcType::Screw, box);
 158     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 159
 160     /* Calculate partial virial, for local atoms only, based on short range.
 161      * Total virial is computed in global_stat, called from do_md
 162      */
 163     const rvec* f = as_rvec_array(forceWithShiftForces.force().data());
 164     f_calc_vir(start, start + homenr, x, f, vir_part, graph, box);
 165     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 166
 167     if (debug)
 168     {
 169         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 170     }
 171 }
 172
 173 static void pull_potential_wrapper(const t_commrec*               cr,
 174                                    const t_inputrec*              ir,
 175                                    const matrix                   box,
 176                                    gmx::ArrayRef<const gmx::RVec> x,
 177                                    gmx::ForceWithVirial*          force,
 178                                    const t_mdatoms*               mdatoms,
 179                                    gmx_enerdata_t*                enerd,
 180                                    pull_t*                        pull_work,
 181                                    const real*                    lambda,
 182                                    double                         t,
 183                                    gmx_wallcycle_t                wcycle)
 184 {
 185     t_pbc pbc;
 186     real  dvdl;
 187
 188     /* Calculate the center of mass forces, this requires communication,
 189      * which is why pull_potential is called close to other communication.
 190      */
 191     wallcycle_start(wcycle, ewcPULLPOT);
 192     set_pbc(&pbc, ir->pbcType, box);
 193     dvdl = 0;
 194     enerd->term[F_COM_PULL] += pull_potential(pull_work, mdatoms, &pbc, cr, t, lambda[efptRESTRAINT],
 195                                               as_rvec_array(x.data()), force, &dvdl);
 196     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 197     wallcycle_stop(wcycle, ewcPULLPOT);
 198 }
 199
 200 static void pme_receive_force_ener(t_forcerec*           fr,
 201                                    const t_commrec*      cr,
 202                                    gmx::ForceWithVirial* forceWithVirial,
 203                                    gmx_enerdata_t*       enerd,
 204                                    bool                  useGpuPmePpComms,
 205                                    bool                  receivePmeForceToGpu,
 206                                    gmx_wallcycle_t       wcycle)
 207 {
 208     real  e_q, e_lj, dvdl_q, dvdl_lj;
 209     float cycles_ppdpme, cycles_seppme;
 210
 211     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 212     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 213
 214     /* In case of node-splitting, the PP nodes receive the long-range
 215      * forces, virial and energy from the PME nodes here.
 216      */
 217     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 218     dvdl_q  = 0;
 219     dvdl_lj = 0;
 220     gmx_pme_receive_f(fr->pmePpCommGpu.get(), cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 221                       useGpuPmePpComms, receivePmeForceToGpu, &cycles_seppme);
 222     enerd->term[F_COUL_RECIP] += e_q;
 223     enerd->term[F_LJ_RECIP] += e_lj;
 224     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 225     enerd->dvdl_lin[efptVDW] += dvdl_lj;
 226
 227     if (wcycle)
 228     {
 229         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 230     }
 231     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 232 }
 233
 234 static void print_large_forces(FILE*            fp,
 235                                const t_mdatoms* md,
 236                                const t_commrec* cr,
 237                                int64_t          step,
 238                                real             forceTolerance,
 239                                const rvec*      x,
 240                                const rvec*      f)
 241 {
 242     real       force2Tolerance = gmx::square(forceTolerance);
 243     gmx::index numNonFinite    = 0;
 244     for (int i = 0; i < md->homenr; i++)
 245     {
 246         real force2    = norm2(f[i]);
 247         bool nonFinite = !std::isfinite(force2);
 248         if (force2 >= force2Tolerance || nonFinite)
 249         {
 250             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n", step,
 251                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 252         }
 253         if (nonFinite)
 254         {
 255             numNonFinite++;
 256         }
 257     }
 258     if (numNonFinite > 0)
 259     {
 260         /* Note that with MPI this fatal call on one rank might interrupt
 261          * the printing on other ranks. But we can only avoid that with
 262          * an expensive MPI barrier that we would need at each step.
 263          */
 264         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 265     }
 266 }
 267
 268 static void post_process_forces(const t_commrec*      cr,
 269                                 int64_t               step,
 270                                 t_nrnb*               nrnb,
 271                                 gmx_wallcycle_t       wcycle,
 272                                 const gmx_localtop_t* top,
 273                                 const matrix          box,
 274                                 const rvec            x[],
 275                                 ForceOutputs*         forceOutputs,
 276                                 tensor                vir_force,
 277                                 const t_mdatoms*      mdatoms,
 278                                 const t_graph*        graph,
 279                                 const t_forcerec*     fr,
 280                                 const gmx_vsite_t*    vsite,
 281                                 const StepWorkload&   stepWork)
 282 {
 283     rvec* f = as_rvec_array(forceOutputs->forceWithShiftForces().force().data());
 284
 285     if (fr->haveDirectVirialContributions)
 286     {
 287         auto& forceWithVirial = forceOutputs->forceWithVirial();
 288         rvec* fDirectVir      = as_rvec_array(forceWithVirial.force_.data());
 289
 290         if (vsite)
 291         {
 292             /* Spread the mesh force on virtual sites to the other particles...
 293              * This is parallellized. MPI communication is performed
 294              * if the constructing atoms aren't local.
 295              */
 296             matrix virial = { { 0 } };
 297             spread_vsite_f(vsite, x, fDirectVir, nullptr, stepWork.computeVirial, virial, nrnb,
 298                            &top->idef, fr->pbcType, fr->bMolPBC, graph, box, cr, wcycle);
 299             forceWithVirial.addVirialContribution(virial);
 300         }
 301
 302         if (stepWork.computeVirial)
 303         {
 304             /* Now add the forces, this is local */
 305             sum_forces(f, forceWithVirial.force_);
 306
 307             /* Add the direct virial contributions */
 308             GMX_ASSERT(
 309                     forceWithVirial.computeVirial_,
 310                     "forceWithVirial should request virial computation when we request the virial");
 311             m_add(vir_force, forceWithVirial.getVirial(), vir_force);
 312
 313             if (debug)
 314             {
 315                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 316             }
 317         }
 318     }
 319
 320     if (fr->print_force >= 0)
 321     {
 322         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 323     }
 324 }
 325
 326 static void do_nb_verlet(t_forcerec*                fr,
 327                          const interaction_const_t* ic,
 328                          gmx_enerdata_t*            enerd,
 329                          const StepWorkload&        stepWork,
 330                          const InteractionLocality  ilocality,
 331                          const int                  clearF,
 332                          const int64_t              step,
 333                          t_nrnb*                    nrnb,
 334                          gmx_wallcycle_t            wcycle)
 335 {
 336     if (!stepWork.computeNonbondedForces)
 337     {
 338         /* skip non-bonded calculation */
 339         return;
 340     }
 341
 342     nonbonded_verlet_t* nbv = fr->nbv.get();
 343
 344     /* GPU kernel launch overhead is already timed separately */
 345     if (fr->cutoff_scheme != ecutsVERLET)
 346     {
 347         gmx_incons("Invalid cut-off scheme passed!");
 348     }
 349
 350     if (!nbv->useGpu())
 351     {
 352         /* When dynamic pair-list  pruning is requested, we need to prune
 353          * at nstlistPrune steps.
 354          */
 355         if (nbv->isDynamicPruningStepCpu(step))
 356         {
 357             /* Prune the pair-list beyond fr->ic->rlistPrune using
 358              * the current coordinates of the atoms.
 359              */
 360             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 361             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 362             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 363         }
 364     }
 365
 366     nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
 367 }
 368
 369 static inline void clear_rvecs_omp(int n, rvec v[])
 370 {
 371     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 372
 373     /* Note that we would like to avoid this conditional by putting it
 374      * into the omp pragma instead, but then we still take the full
 375      * omp parallel for overhead (at least with gcc5).
 376      */
 377     if (nth == 1)
 378     {
 379         for (int i = 0; i < n; i++)
 380         {
 381             clear_rvec(v[i]);
 382         }
 383     }
 384     else
 385     {
 386 #pragma omp parallel for num_threads(nth) schedule(static)
 387         for (int i = 0; i < n; i++)
 388         {
 389             clear_rvec(v[i]);
 390         }
 391     }
 392 }
 393
 394 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 395  *
 396  * \param groupOptions  Group options, containing T-coupling options
 397  */
 398 static real averageKineticEnergyEstimate(const t_grpopts& groupOptions)
 399 {
 400     real nrdfCoupled   = 0;
 401     real nrdfUncoupled = 0;
 402     real kineticEnergy = 0;
 403     for (int g = 0; g < groupOptions.ngtc; g++)
 404     {
 405         if (groupOptions.tau_t[g] >= 0)
 406         {
 407             nrdfCoupled += groupOptions.nrdf[g];
 408             kineticEnergy += groupOptions.nrdf[g] * 0.5 * groupOptions.ref_t[g] * BOLTZ;
 409         }
 410         else
 411         {
 412             nrdfUncoupled += groupOptions.nrdf[g];
 413         }
 414     }
 415
 416     /* This conditional with > also catches nrdf=0 */
 417     if (nrdfCoupled > nrdfUncoupled)
 418     {
 419         return kineticEnergy * (nrdfCoupled + nrdfUncoupled) / nrdfCoupled;
 420     }
 421     else
 422     {
 423         return 0;
 424     }
 425 }
 426
 427 /*! \brief This routine checks that the potential energy is finite.
 428  *
 429  * Always checks that the potential energy is finite. If step equals
 430  * inputrec.init_step also checks that the magnitude of the potential energy
 431  * is reasonable. Terminates with a fatal error when a check fails.
 432  * Note that passing this check does not guarantee finite forces,
 433  * since those use slightly different arithmetics. But in most cases
 434  * there is just a narrow coordinate range where forces are not finite
 435  * and energies are finite.
 436  *
 437  * \param[in] step      The step number, used for checking and printing
 438  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to
 439  * enerd.term[F_EPOT] before calling this routine \param[in] inputrec  The input record
 440  */
 441 static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
 442 {
 443     /* Threshold valid for comparing absolute potential energy against
 444      * the kinetic energy. Normally one should not consider absolute
 445      * potential energy values, but with a factor of one million
 446      * we should never get false positives.
 447      */
 448     constexpr real c_thresholdFactor = 1e6;
 449
 450     bool energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 451     real averageKineticEnergy = 0;
 452     /* We only check for large potential energy at the initial step,
 453      * because that is by far the most likely step for this too occur
 454      * and because computing the average kinetic energy is not free.
 455      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 456      * before they become NaN.
 457      */
 458     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 459     {
 460         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 461     }
 462
 463     if (energyIsNotFinite
 464         || (averageKineticEnergy > 0 && enerd.term[F_EPOT] > c_thresholdFactor * averageKineticEnergy))
 465     {
 466         gmx_fatal(
 467                 FARGS,
 468                 "Step %" PRId64
 469                 ": The total potential energy is %g, which is %s. The LJ and electrostatic "
 470                 "contributions to the energy are %g and %g, respectively. A %s potential energy "
 471                 "can be caused by overlapping interactions in bonded interactions or very large%s "
 472                 "coordinate values. Usually this is caused by a badly- or non-equilibrated initial "
 473                 "configuration, incorrect interactions or parameters in the topology.",
 474                 step, enerd.term[F_EPOT], energyIsNotFinite ? "not finite" : "extremely high",
 475                 enerd.term[F_LJ], enerd.term[F_COUL_SR],
 476                 energyIsNotFinite ? "non-finite" : "very high", energyIsNotFinite ? " or Nan" : "");
 477     }
 478 }
 479
 480 /*! \brief Return true if there are special forces computed this step.
 481  *
 482  * The conditionals exactly correspond to those in computeSpecialForces().
 483  */
 484 static bool haveSpecialForces(const t_inputrec&          inputrec,
 485                               const gmx::ForceProviders& forceProviders,
 486                               const pull_t*              pull_work,
 487                               const bool                 computeForces,
 488                               const gmx_edsam*           ed)
 489 {
 490
 491     return ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
 492             (inputrec.bPull && pull_have_potential(pull_work)) ||   // pull
 493             inputrec.bRot ||                                        // enforced rotation
 494             (ed != nullptr) ||                                      // flooding
 495             (inputrec.bIMD && computeForces));                      // IMD
 496 }
 497
 498 /*! \brief Compute forces and/or energies for special algorithms
 499  *
 500  * The intention is to collect all calls to algorithms that compute
 501  * forces on local atoms only and that do not contribute to the local
 502  * virial sum (but add their virial contribution separately).
 503  * Eventually these should likely all become ForceProviders.
 504  * Within this function the intention is to have algorithms that do
 505  * global communication at the end, so global barriers within the MD loop
 506  * are as close together as possible.
 507  *
 508  * \param[in]     fplog            The log file
 509  * \param[in]     cr               The communication record
 510  * \param[in]     inputrec         The input record
 511  * \param[in]     awh              The Awh module (nullptr if none in use).
 512  * \param[in]     enforcedRotation Enforced rotation module.
 513  * \param[in]     imdSession       The IMD session
 514  * \param[in]     pull_work        The pull work structure.
 515  * \param[in]     step             The current MD step
 516  * \param[in]     t                The current time
 517  * \param[in,out] wcycle           Wallcycle accounting struct
 518  * \param[in,out] forceProviders   Pointer to a list of force providers
 519  * \param[in]     box              The unit cell
 520  * \param[in]     x                The coordinates
 521  * \param[in]     mdatoms          Per atom properties
 522  * \param[in]     lambda           Array of free-energy lambda values
 523  * \param[in]     stepWork         Step schedule flags
 524  * \param[in,out] forceWithVirial  Force and virial buffers
 525  * \param[in,out] enerd            Energy buffer
 526  * \param[in,out] ed               Essential dynamics pointer
 527  * \param[in]     didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
 528  *
 529  * \todo Remove didNeighborSearch, which is used incorrectly.
 530  * \todo Convert all other algorithms called here to ForceProviders.
 531  */
 532 static void computeSpecialForces(FILE*                          fplog,
 533                                  const t_commrec*               cr,
 534                                  const t_inputrec*              inputrec,
 535                                  gmx::Awh*                      awh,
 536                                  gmx_enfrot*                    enforcedRotation,
 537                                  gmx::ImdSession*               imdSession,
 538                                  pull_t*                        pull_work,
 539                                  int64_t                        step,
 540                                  double                         t,
 541                                  gmx_wallcycle_t                wcycle,
 542                                  gmx::ForceProviders*           forceProviders,
 543                                  const matrix                   box,
 544                                  gmx::ArrayRef<const gmx::RVec> x,
 545                                  const t_mdatoms*               mdatoms,
 546                                  real*                          lambda,
 547                                  const StepWorkload&            stepWork,
 548                                  gmx::ForceWithVirial*          forceWithVirial,
 549                                  gmx_enerdata_t*                enerd,
 550                                  gmx_edsam*                     ed,
 551                                  bool                           didNeighborSearch)
 552 {
 553     /* NOTE: Currently all ForceProviders only provide forces.
 554      *       When they also provide energies, remove this conditional.
 555      */
 556     if (stepWork.computeForces)
 557     {
 558         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 559         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 560
 561         /* Collect forces from modules */
 562         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 563     }
 564
 565     if (inputrec->bPull && pull_have_potential(pull_work))
 566     {
 567         pull_potential_wrapper(cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work,
 568                                lambda, t, wcycle);
 569
 570         if (awh)
 571         {
 572             enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(
 573                     inputrec->pbcType, *mdatoms, box, forceWithVirial, t, step, wcycle, fplog);
 574         }
 575     }
 576
 577     rvec* f = as_rvec_array(forceWithVirial->force_.data());
 578
 579     /* Add the forces from enforced rotation potentials (if any) */
 580     if (inputrec->bRot)
 581     {
 582         wallcycle_start(wcycle, ewcROTadd);
 583         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 584         wallcycle_stop(wcycle, ewcROTadd);
 585     }
 586
 587     if (ed)
 588     {
 589         /* Note that since init_edsam() is called after the initialization
 590          * of forcerec, edsam doesn't request the noVirSum force buffer.
 591          * Thus if no other algorithm (e.g. PME) requires it, the forces
 592          * here will contribute to the virial.
 593          */
 594         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
 595     }
 596
 597     /* Add forces from interactive molecular dynamics (IMD), if any */
 598     if (inputrec->bIMD && stepWork.computeForces)
 599     {
 600         imdSession->applyForces(f);
 601     }
 602 }
 603
 604 /*! \brief Makes PME flags from StepWorkload data.
 605  *
 606  * \param[in]  stepWork     Step schedule flags
 607  * \returns                 PME flags
 608  */
 609 static int makePmeFlags(const StepWorkload& stepWork)
 610 {
 611     return GMX_PME_SPREAD | GMX_PME_SOLVE | (stepWork.computeVirial ? GMX_PME_CALC_ENER_VIR : 0)
 612            | (stepWork.computeEnergy ? GMX_PME_CALC_ENER_VIR : 0)
 613            | (stepWork.computeForces ? GMX_PME_CALC_F : 0);
 614 }
 615
 616 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 617  *
 618  * \param[in]  pmedata              The PME structure
 619  * \param[in]  box                  The box matrix
 620  * \param[in]  stepWork             Step schedule flags
 621  * \param[in]  pmeFlags             PME flags
 622  * \param[in]  xReadyOnDevice       Event synchronizer indicating that the coordinates are ready in
 623  * the device memory. \param[in]  wcycle               The wallcycle structure
 624  */
 625 static inline void launchPmeGpuSpread(gmx_pme_t*            pmedata,
 626                                       const matrix          box,
 627                                       const StepWorkload&   stepWork,
 628                                       int                   pmeFlags,
 629                                       GpuEventSynchronizer* xReadyOnDevice,
 630                                       gmx_wallcycle_t       wcycle)
 631 {
 632     pme_gpu_prepare_computation(pmedata, stepWork.haveDynamicBox, box, wcycle, pmeFlags,
 633                                 stepWork.useGpuPmeFReduction);
 634     pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle);
 635 }
 636
 637 /*! \brief Launch the FFT and gather stages of PME GPU
 638  *
 639  * This function only implements setting the output forces (no accumulation).
 640  *
 641  * \param[in]  pmedata        The PME structure
 642  * \param[in]  wcycle         The wallcycle structure
 643  */
 644 static void launchPmeGpuFftAndGather(gmx_pme_t* pmedata, gmx_wallcycle_t wcycle)
 645 {
 646     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 647     pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
 648 }
 649
 650 /*! \brief
 651  *  Polling wait for either of the PME or nonbonded GPU tasks.
 652  *
 653  * Instead of a static order in waiting for GPU tasks, this function
 654  * polls checking which of the two tasks completes first, and does the
 655  * associated force buffer reduction overlapped with the other task.
 656  * By doing that, unlike static scheduling order, it can always overlap
 657  * one of the reductions, regardless of the GPU task completion order.
 658  *
 659  * \param[in]     nbv              Nonbonded verlet structure
 660  * \param[in,out] pmedata          PME module data
 661  * \param[in,out] forceOutputs     Output buffer for the forces and virial
 662  * \param[in,out] enerd            Energy data structure results are reduced into
 663  * \param[in]     stepWork         Step schedule flags
 664  * \param[in]     pmeFlags         PME flags
 665  * \param[in]     wcycle           The wallcycle structure
 666  */
 667 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
 668                                         gmx_pme_t*          pmedata,
 669                                         gmx::ForceOutputs*  forceOutputs,
 670                                         gmx_enerdata_t*     enerd,
 671                                         const StepWorkload& stepWork,
 672                                         int                 pmeFlags,
 673                                         gmx_wallcycle_t     wcycle)
 674 {
 675     bool isPmeGpuDone = false;
 676     bool isNbGpuDone  = false;
 677
 678
 679     gmx::ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
 680     gmx::ForceWithVirial&      forceWithVirial      = forceOutputs->forceWithVirial();
 681
 682     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 683
 684     while (!isPmeGpuDone || !isNbGpuDone)
 685     {
 686         if (!isPmeGpuDone)
 687         {
 688             GpuTaskCompletion completionType =
 689                     (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 690             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, &forceWithVirial,
 691                                                    enerd, completionType);
 692         }
 693
 694         if (!isNbGpuDone)
 695         {
 696             GpuTaskCompletion completionType =
 697                     (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 698             isNbGpuDone = Nbnxm::gpu_try_finish_task(
 699                     nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
 700                     enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(),
 701                     completionType, wcycle);
 702
 703             if (isNbGpuDone)
 704             {
 705                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShiftForces.force());
 706             }
 707         }
 708     }
 709 }
 710
 711 /*! \brief Set up the different force buffers; also does clearing.
 712  *
 713  * \param[in] fr        force record pointer
 714  * \param[in] pull_work The pull work object.
 715  * \param[in] inputrec  input record
 716  * \param[in] force     force array
 717  * \param[in] stepWork  Step schedule flags
 718  * \param[out] wcycle   wallcycle recording structure
 719  *
 720  * \returns             Cleared force output structure
 721  */
 722 static ForceOutputs setupForceOutputs(t_forcerec*                         fr,
 723                                       pull_t*                             pull_work,
 724                                       const t_inputrec&                   inputrec,
 725                                       gmx::ArrayRefWithPadding<gmx::RVec> force,
 726                                       const StepWorkload&                 stepWork,
 727                                       gmx_wallcycle_t                     wcycle)
 728 {
 729     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 730
 731     /* NOTE: We assume fr->shiftForces is all zeros here */
 732     gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial, fr->shiftForces);
 733
 734     if (stepWork.computeForces)
 735     {
 736         /* Clear the short- and long-range forces */
 737         clear_rvecs_omp(fr->natoms_force_constr, as_rvec_array(forceWithShiftForces.force().data()));
 738     }
 739
 740     /* If we need to compute the virial, we might need a separate
 741      * force buffer for algorithms for which the virial is calculated
 742      * directly, such as PME. Otherwise, forceWithVirial uses the
 743      * the same force (f in legacy calls) buffer as other algorithms.
 744      */
 745     const bool useSeparateForceWithVirialBuffer =
 746             (stepWork.computeForces && (stepWork.computeVirial && fr->haveDirectVirialContributions));
 747     /* forceWithVirial uses the local atom range only */
 748     gmx::ForceWithVirial forceWithVirial(useSeparateForceWithVirialBuffer ? fr->forceBufferForDirectVirialContributions
 749                                                                           : force.unpaddedArrayRef(),
 750                                          stepWork.computeVirial);
 751
 752     if (useSeparateForceWithVirialBuffer)
 753     {
 754         /* TODO: update comment
 755          * We only compute forces on local atoms. Note that vsites can
 756          * spread to non-local atoms, but that part of the buffer is
 757          * cleared separately in the vsite spreading code.
 758          */
 759         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 760     }
 761
 762     if (inputrec.bPull && pull_have_constraint(pull_work))
 763     {
 764         clear_pull_forces(pull_work);
 765     }
 766
 767     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 768
 769     return ForceOutputs(forceWithShiftForces, forceWithVirial);
 770 }
 771
 772
 773 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
 774  */
 775 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec&         inputrec,
 776                                                           const t_forcerec&         fr,
 777                                                           const pull_t*             pull_work,
 778                                                           const gmx_edsam*          ed,
 779                                                           const t_idef&             idef,
 780                                                           const t_fcdata&           fcd,
 781                                                           const t_mdatoms&          mdatoms,
 782                                                           const SimulationWorkload& simulationWork,
 783                                                           const StepWorkload&       stepWork)
 784 {
 785     DomainLifetimeWorkload domainWork;
 786     // Note that haveSpecialForces is constant over the whole run
 787     domainWork.haveSpecialForces =
 788             haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 789     domainWork.haveCpuBondedWork  = haveCpuBondeds(fr);
 790     domainWork.haveGpuBondedWork  = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
 791     domainWork.haveRestraintsWork = haveRestraints(idef, fcd);
 792     domainWork.haveCpuListedForceWork = haveCpuListedForces(fr, idef, fcd);
 793     // Note that haveFreeEnergyWork is constant over the whole run
 794     domainWork.haveFreeEnergyWork = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
 795     // We assume we have local force work if there are CPU
 796     // force tasks including PME or nonbondeds.
 797     domainWork.haveCpuLocalForceWork =
 798             domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
 799             || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
 800             || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
 801
 802     return domainWork;
 803 }
 804
 805 /*! \brief Set up force flag stuct from the force bitmask.
 806  *
 807  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 808  * \param[in]      isNonbondedOn        Global override, if false forces to turn off all nonbonded calculation.
 809  * \param[in]      simulationWork       Simulation workload description.
 810  * \param[in]      rankHasPmeDuty       If this rank computes PME.
 811  *
 812  * \returns New Stepworkload description.
 813  */
 814 static StepWorkload setupStepWorkload(const int                 legacyFlags,
 815                                       const bool                isNonbondedOn,
 816                                       const SimulationWorkload& simulationWork,
 817                                       const bool                rankHasPmeDuty)
 818 {
 819     StepWorkload flags;
 820     flags.stateChanged           = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 821     flags.haveDynamicBox         = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 822     flags.doNeighborSearch       = ((legacyFlags & GMX_FORCE_NS) != 0);
 823     flags.computeVirial          = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 824     flags.computeEnergy          = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 825     flags.computeForces          = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 826     flags.computeListedForces    = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 827     flags.computeNonbondedForces = ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && isNonbondedOn;
 828     flags.computeDhdl            = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 829
 830     if (simulationWork.useGpuBufferOps)
 831     {
 832         GMX_ASSERT(simulationWork.useGpuNonbonded,
 833                    "Can only offload buffer ops if nonbonded computation is also offloaded");
 834     }
 835     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 836     // on virial steps the CPU reduction path is taken
 837     flags.useGpuFBufferOps    = simulationWork.useGpuBufferOps && !flags.computeVirial;
 838     flags.useGpuPmeFReduction = flags.useGpuFBufferOps
 839                                 && (simulationWork.useGpuPme
 840                                     && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication));
 841
 842     return flags;
 843 }
 844
 845
 846 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
 847  *
 848  * TODO: eliminate \p useGpuPmeOnThisRank when this is
 849  * incorporated in DomainLifetimeWorkload.
 850  */
 851 static void launchGpuEndOfStepTasks(nonbonded_verlet_t*               nbv,
 852                                     gmx::GpuBonded*                   gpuBonded,
 853                                     gmx_pme_t*                        pmedata,
 854                                     gmx_enerdata_t*                   enerd,
 855                                     const gmx::MdrunScheduleWorkload& runScheduleWork,
 856                                     bool                              useGpuPmeOnThisRank,
 857                                     int64_t                           step,
 858                                     gmx_wallcycle_t                   wcycle)
 859 {
 860     if (runScheduleWork.simulationWork.useGpuNonbonded)
 861     {
 862         /* Launch pruning before buffer clearing because the API overhead of the
 863          * clear kernel launches can leave the GPU idle while it could be running
 864          * the prune kernel.
 865          */
 866         if (nbv->isDynamicPruningStepGpu(step))
 867         {
 868             nbv->dispatchPruneKernelGpu(step);
 869         }
 870
 871         /* now clear the GPU outputs while we finish the step on the CPU */
 872         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 873         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 874         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
 875         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 876         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 877     }
 878
 879     if (useGpuPmeOnThisRank)
 880     {
 881         pme_gpu_reinit_computation(pmedata, wcycle);
 882     }
 883
 884     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
 885     {
 886         // in principle this should be included in the DD balancing region,
 887         // but generally it is infrequent so we'll omit it for the sake of
 888         // simpler code
 889         gpuBonded->waitAccumulateEnergyTerms(enerd);
 890
 891         gpuBonded->clearEnergies();
 892     }
 893 }
 894
 895
 896 void do_force(FILE*                               fplog,
 897               const t_commrec*                    cr,
 898               const gmx_multisim_t*               ms,
 899               const t_inputrec*                   inputrec,
 900               gmx::Awh*                           awh,
 901               gmx_enfrot*                         enforcedRotation,
 902               gmx::ImdSession*                    imdSession,
 903               pull_t*                             pull_work,
 904               int64_t                             step,
 905               t_nrnb*                             nrnb,
 906               gmx_wallcycle_t                     wcycle,
 907               const gmx_localtop_t*               top,
 908               const matrix                        box,
 909               gmx::ArrayRefWithPadding<gmx::RVec> x,
 910               history_t*                          hist,
 911               gmx::ArrayRefWithPadding<gmx::RVec> force,
 912               tensor                              vir_force,
 913               const t_mdatoms*                    mdatoms,
 914               gmx_enerdata_t*                     enerd,
 915               t_fcdata*                           fcd,
 916               gmx::ArrayRef<real>                 lambda,
 917               t_graph*                            graph,
 918               t_forcerec*                         fr,
 919               gmx::MdrunScheduleWorkload*         runScheduleWork,
 920               const gmx_vsite_t*                  vsite,
 921               rvec                                mu_tot,
 922               double                              t,
 923               gmx_edsam*                          ed,
 924               int                                 legacyFlags,
 925               const DDBalanceRegionHandler&       ddBalanceRegionHandler)
 926 {
 927     int                          i, j;
 928     double                       mu[2 * DIM];
 929     nonbonded_verlet_t*          nbv      = fr->nbv.get();
 930     interaction_const_t*         ic       = fr->ic;
 931     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
 932
 933     // TODO remove the code below when the legacy flags are not in use anymore
 934     /* modify force flag if not doing nonbonded */
 935     if (!fr->bNonbonded)
 936     {
 937         legacyFlags &= ~GMX_FORCE_NONBONDED;
 938     }
 939
 940     const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
 941
 942
 943     runScheduleWork->stepWork    = setupStepWorkload(legacyFlags, fr->bNonbonded, simulationWork,
 944                                                   thisRankHasDuty(cr, DUTY_PME));
 945     const StepWorkload& stepWork = runScheduleWork->stepWork;
 946
 947
 948     const bool useGpuPmeOnThisRank = simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME);
 949     const int  pmeFlags            = makePmeFlags(stepWork);
 950
 951     // Switches on whether to use GPU for position and force buffer operations
 952     // TODO consider all possible combinations of triggers, and how to combine optimally in each case.
 953     const BufferOpsUseGpu useGpuXBufOps =
 954             stepWork.useGpuXBufferOps ? BufferOpsUseGpu::True : BufferOpsUseGpu::False;
 955     // GPU Force buffer ops are disabled on virial steps, because the virial calc is not yet ported to GPU
 956     const BufferOpsUseGpu useGpuFBufOps =
 957             stepWork.useGpuFBufferOps ? BufferOpsUseGpu::True : BufferOpsUseGpu::False;
 958
 959     /* At a search step we need to start the first balancing region
 960      * somewhere early inside the step after communication during domain
 961      * decomposition (and not during the previous step as usual).
 962      */
 963     if (stepWork.doNeighborSearch)
 964     {
 965         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 966     }
 967
 968     const int start  = 0;
 969     const int homenr = mdatoms->homenr;
 970
 971     clear_mat(vir_force);
 972
 973     if (stepWork.stateChanged)
 974     {
 975         if (inputrecNeedMutot(inputrec))
 976         {
 977             /* Calculate total (local) dipole moment in a temporary common array.
 978              * This makes it possible to sum them over nodes faster.
 979              */
 980             calc_mu(start, homenr, x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB,
 981                     mdatoms->nChargePerturbed, mu, mu + DIM);
 982         }
 983     }
 984
 985     if (fr->pbcType != PbcType::No)
 986     {
 987         /* Compute shift vectors every step,
 988          * because of pressure coupling or box deformation!
 989          */
 990         if (stepWork.haveDynamicBox && stepWork.stateChanged)
 991         {
 992             calc_shifts(box, fr->shift_vec);
 993         }
 994
 995         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
 996         const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
 997         if (calcCGCM)
 998         {
 999             put_atoms_in_box_omp(fr->pbcType, box, x.unpaddedArrayRef().subArray(0, homenr),
1000                                  gmx_omp_nthreads_get(emntDefault));
1001             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
1002         }
1003         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1004         {
1005             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
1006         }
1007     }
1008
1009     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1010
1011     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1012     // The local coordinates can be copied right away.
1013     // NOTE: Consider moving this copy to right after they are updated and constrained,
1014     //       if the later is not offloaded.
1015     if (useGpuPmeOnThisRank || useGpuXBufOps == BufferOpsUseGpu::True)
1016     {
1017         if (stepWork.doNeighborSearch)
1018         {
1019             stateGpu->reinit(mdatoms->homenr,
1020                              cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1021             if (useGpuPmeOnThisRank)
1022             {
1023                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1024                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1025             }
1026         }
1027         // We need to copy coordinates when:
1028         // 1. Update is not offloaded
1029         // 2. The buffers were reinitialized on search step
1030         if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1031         {
1032             stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1033         }
1034     }
1035
1036     // TODO Update this comment when introducing SimulationWorkload
1037     //
1038     // The conditions for gpuHaloExchange e.g. using GPU buffer
1039     // operations were checked before construction, so here we can
1040     // just use it and assert upon any conditions.
1041     gmx::GpuHaloExchange* gpuHaloExchange =
1042             (havePPDomainDecomposition(cr) ? cr->dd->gpuHaloExchange.get() : nullptr);
1043     const bool ddUsesGpuDirectCommunication = (gpuHaloExchange != nullptr);
1044     GMX_ASSERT(!ddUsesGpuDirectCommunication || (useGpuXBufOps == BufferOpsUseGpu::True),
1045                "Must use coordinate buffer ops with GPU halo exchange");
1046     const bool useGpuForcesHaloExchange =
1047             ddUsesGpuDirectCommunication && (useGpuFBufOps == BufferOpsUseGpu::True);
1048
1049     // Copy coordinate from the GPU if update is on the GPU and there
1050     // are forces to be computed on the CPU, or for the computation of
1051     // virial, or if host-side data will be transferred from this task
1052     // to a remote task for halo exchange or PME-PP communication. At
1053     // search steps the current coordinates are already on the host,
1054     // hence copy is not needed.
1055     const bool haveHostPmePpComms =
1056             !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication;
1057     const bool haveHostHaloExchangeComms = havePPDomainDecomposition(cr) && !ddUsesGpuDirectCommunication;
1058
1059     bool gmx_used_in_debug haveCopiedXFromGpu = false;
1060     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1061         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1062             || haveHostPmePpComms || haveHostHaloExchangeComms))
1063     {
1064         stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1065         haveCopiedXFromGpu = true;
1066     }
1067
1068     const auto localXReadyOnDevice = (stateGpu != nullptr)
1069                                              ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1070                                                        AtomLocality::Local, simulationWork, stepWork)
1071                                              : nullptr;
1072
1073 #if GMX_MPI
1074     if (!thisRankHasDuty(cr, DUTY_PME))
1075     {
1076         /* Send particle coordinates to the pme nodes.
1077          * Since this is only implemented for domain decomposition
1078          * and domain decomposition does not use the graph,
1079          * we do not need to worry about shifting.
1080          */
1081         bool reinitGpuPmePpComms = simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1082         bool sendCoordinatesFromGpu =
1083                 simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1084
1085         if (!stepWork.doNeighborSearch && simulationWork.useGpuUpdate && !sendCoordinatesFromGpu)
1086         {
1087             GMX_RELEASE_ASSERT(false,
1088                                "GPU update and separate PME ranks are only supported with GPU "
1089                                "direct communication!");
1090             // TODO: when this code-path becomes supported add:
1091             // stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1092         }
1093
1094         gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1095                                  lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1096                                  step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1097                                  sendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1098     }
1099 #endif /* GMX_MPI */
1100
1101     if (useGpuPmeOnThisRank)
1102     {
1103         launchPmeGpuSpread(fr->pmedata, box, stepWork, pmeFlags, localXReadyOnDevice, wcycle);
1104     }
1105
1106     /* do gridding for pair search */
1107     if (stepWork.doNeighborSearch)
1108     {
1109         if (graph && stepWork.stateChanged)
1110         {
1111             /* Calculate intramolecular shift vectors to make molecules whole */
1112             mk_mshift(fplog, graph, fr->pbcType, box, as_rvec_array(x.unpaddedArrayRef().data()));
1113         }
1114
1115         // TODO
1116         // - vzero is constant, do we need to pass it?
1117         // - box_diag should be passed directly to nbnxn_put_on_grid
1118         //
1119         rvec vzero;
1120         clear_rvec(vzero);
1121
1122         rvec box_diag;
1123         box_diag[XX] = box[XX][XX];
1124         box_diag[YY] = box[YY][YY];
1125         box_diag[ZZ] = box[ZZ][ZZ];
1126
1127         wallcycle_start(wcycle, ewcNS);
1128         if (!DOMAINDECOMP(cr))
1129         {
1130             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1131             nbnxn_put_on_grid(nbv, box, 0, vzero, box_diag, nullptr, { 0, mdatoms->homenr }, -1,
1132                               fr->cginfo, x.unpaddedArrayRef(), 0, nullptr);
1133             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1134         }
1135         else
1136         {
1137             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1138             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->cginfo, x.unpaddedArrayRef());
1139             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1140         }
1141
1142         nbv->setAtomProperties(*mdatoms, fr->cginfo);
1143
1144         wallcycle_stop(wcycle, ewcNS);
1145
1146         /* initialize the GPU nbnxm atom data and bonded data structures */
1147         if (simulationWork.useGpuNonbonded)
1148         {
1149             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1150
1151             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1152             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1153             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1154
1155             if (fr->gpuBonded)
1156             {
1157                 /* Now we put all atoms on the grid, we can assign bonded
1158                  * interactions to the GPU, where the grid order is
1159                  * needed. Also the xq, f and fshift device buffers have
1160                  * been reallocated if needed, so the bonded code can
1161                  * learn about them. */
1162                 // TODO the xq, f, and fshift buffers are now shared
1163                 // resources, so they should be maintained by a
1164                 // higher-level object than the nb module.
1165                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(
1166                         nbv->getGridIndices(), top->idef, Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1167                         Nbnxm::gpu_get_f(nbv->gpu_nbv), Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1168             }
1169             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1170         }
1171
1172         // Need to run after the GPU-offload bonded interaction lists
1173         // are set up to be able to determine whether there is bonded work.
1174         runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1175                 *inputrec, *fr, pull_work, ed, top->idef, *fcd, *mdatoms, simulationWork, stepWork);
1176
1177         wallcycle_start_nocount(wcycle, ewcNS);
1178         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1179         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1180         nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1181
1182         nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::Local);
1183
1184         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1185         wallcycle_stop(wcycle, ewcNS);
1186
1187         if (useGpuXBufOps == BufferOpsUseGpu::True)
1188         {
1189             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1190         }
1191         // For force buffer ops, we use the below conditon rather than
1192         // useGpuFBufOps to ensure that init is performed even if this
1193         // NS step is also a virial step (on which f buf ops are deactivated).
1194         if (simulationWork.useGpuBufferOps && simulationWork.useGpuNonbonded && (GMX_GPU == GMX_GPU_CUDA))
1195         {
1196             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1197             nbv->atomdata_init_add_nbat_f_to_f_gpu(stateGpu->fReducedOnDevice());
1198         }
1199     }
1200     else if (!EI_TPI(inputrec->eI))
1201     {
1202         if (useGpuXBufOps == BufferOpsUseGpu::True)
1203         {
1204             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1205             nbv->convertCoordinatesGpu(AtomLocality::Local, false, stateGpu->getCoordinates(),
1206                                        localXReadyOnDevice);
1207         }
1208         else
1209         {
1210             if (simulationWork.useGpuUpdate)
1211             {
1212                 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1213                 GMX_ASSERT(haveCopiedXFromGpu,
1214                            "a wait should only be triggered if copy has been scheduled");
1215                 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1216             }
1217             nbv->convertCoordinates(AtomLocality::Local, false, x.unpaddedArrayRef());
1218         }
1219     }
1220
1221     const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1222
1223     if (simulationWork.useGpuNonbonded)
1224     {
1225         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1226
1227         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1228
1229         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1230         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1231         if (stepWork.doNeighborSearch || (useGpuXBufOps == BufferOpsUseGpu::False))
1232         {
1233             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1234         }
1235         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1236         // with X buffer ops offloaded to the GPU on all but the search steps
1237
1238         // bonded work not split into separate local and non-local, so with DD
1239         // we can only launch the kernel after non-local coordinates have been received.
1240         if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1241         {
1242             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1243             fr->gpuBonded->launchKernel(fr, stepWork, box);
1244             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1245         }
1246
1247         /* launch local nonbonded work on GPU */
1248         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1249         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1250         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1251         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1252     }
1253
1254     if (useGpuPmeOnThisRank)
1255     {
1256         // In PME GPU and mixed mode we launch FFT / gather after the
1257         // X copy/transform to allow overlap as well as after the GPU NB
1258         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1259         // the nonbonded kernel.
1260         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1261     }
1262
1263     /* Communicate coordinates and sum dipole if necessary +
1264        do non-local pair search */
1265     if (havePPDomainDecomposition(cr))
1266     {
1267         if (stepWork.doNeighborSearch)
1268         {
1269             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1270             wallcycle_start_nocount(wcycle, ewcNS);
1271             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1272             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1273             nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1274
1275             nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::NonLocal);
1276             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1277             wallcycle_stop(wcycle, ewcNS);
1278             if (ddUsesGpuDirectCommunication)
1279             {
1280                 gpuHaloExchange->reinitHalo(stateGpu->getCoordinates(), stateGpu->getForces());
1281             }
1282         }
1283         else
1284         {
1285             if (ddUsesGpuDirectCommunication)
1286             {
1287                 // The following must be called after local setCoordinates (which records an event
1288                 // when the coordinate data has been copied to the device).
1289                 gpuHaloExchange->communicateHaloCoordinates(box, localXReadyOnDevice);
1290
1291                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1292                 {
1293                     // non-local part of coordinate buffer must be copied back to host for CPU work
1294                     stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1295                 }
1296             }
1297             else
1298             {
1299                 // Note: GPU update + DD without direct communication is not supported,
1300                 // a waitCoordinatesReadyOnHost() should be issued if it will be.
1301                 GMX_ASSERT(!simulationWork.useGpuUpdate,
1302                            "GPU update is not supported with CPU halo exchange");
1303                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1304             }
1305
1306             if (useGpuXBufOps == BufferOpsUseGpu::True)
1307             {
1308                 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1309                 {
1310                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1311                 }
1312                 nbv->convertCoordinatesGpu(AtomLocality::NonLocal, false, stateGpu->getCoordinates(),
1313                                            stateGpu->getCoordinatesReadyOnDeviceEvent(
1314                                                    AtomLocality::NonLocal, simulationWork, stepWork));
1315             }
1316             else
1317             {
1318                 nbv->convertCoordinates(AtomLocality::NonLocal, false, x.unpaddedArrayRef());
1319             }
1320         }
1321
1322         if (simulationWork.useGpuNonbonded)
1323         {
1324             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1325
1326             if (stepWork.doNeighborSearch || (useGpuXBufOps == BufferOpsUseGpu::False))
1327             {
1328                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1329                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1330                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1331             }
1332
1333             if (domainWork.haveGpuBondedWork)
1334             {
1335                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1336                 fr->gpuBonded->launchKernel(fr, stepWork, box);
1337                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1338             }
1339
1340             /* launch non-local nonbonded tasks on GPU */
1341             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1342             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1343                          nrnb, wcycle);
1344             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1345
1346             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1347         }
1348     }
1349
1350     if (simulationWork.useGpuNonbonded)
1351     {
1352         /* launch D2H copy-back F */
1353         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1354         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1355
1356         if (havePPDomainDecomposition(cr))
1357         {
1358             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1359         }
1360         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1361         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1362
1363         if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1364         {
1365             fr->gpuBonded->launchEnergyTransfer();
1366         }
1367         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1368     }
1369
1370     if (stepWork.stateChanged && inputrecNeedMutot(inputrec))
1371     {
1372         if (PAR(cr))
1373         {
1374             gmx_sumd(2 * DIM, mu, cr);
1375
1376             ddBalanceRegionHandler.reopenRegionCpu();
1377         }
1378
1379         for (i = 0; i < 2; i++)
1380         {
1381             for (j = 0; j < DIM; j++)
1382             {
1383                 fr->mu_tot[i][j] = mu[i * DIM + j];
1384             }
1385         }
1386     }
1387     if (mdatoms->nChargePerturbed == 0)
1388     {
1389         copy_rvec(fr->mu_tot[0], mu_tot);
1390     }
1391     else
1392     {
1393         for (j = 0; j < DIM; j++)
1394         {
1395             mu_tot[j] = (1.0 - lambda[efptCOUL]) * fr->mu_tot[0][j] + lambda[efptCOUL] * fr->mu_tot[1][j];
1396         }
1397     }
1398
1399     /* Reset energies */
1400     reset_enerdata(enerd);
1401     /* Clear the shift forces */
1402     // TODO: This should be linked to the shift force buffer in use, or cleared before use instead
1403     for (gmx::RVec& elem : fr->shiftForces)
1404     {
1405         elem = { 0.0_real, 0.0_real, 0.0_real };
1406     }
1407
1408     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1409     {
1410         wallcycle_start(wcycle, ewcPPDURINGPME);
1411         dd_force_flop_start(cr->dd, nrnb);
1412     }
1413
1414     // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1415     // this wait ensures that the D2H transfer is complete.
1416     if ((simulationWork.useGpuUpdate)
1417         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial))
1418     {
1419         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1420     }
1421
1422     if (inputrec->bRot)
1423     {
1424         wallcycle_start(wcycle, ewcROT);
1425         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step,
1426                     stepWork.doNeighborSearch);
1427         wallcycle_stop(wcycle, ewcROT);
1428     }
1429
1430     /* Start the force cycle counter.
1431      * Note that a different counter is used for dynamic load balancing.
1432      */
1433     wallcycle_start(wcycle, ewcFORCE);
1434
1435     // Set up and clear force outputs.
1436     // We use std::move to keep the compiler happy, it has no effect.
1437     ForceOutputs forceOut =
1438             setupForceOutputs(fr, pull_work, *inputrec, std::move(force), stepWork, wcycle);
1439
1440     /* We calculate the non-bonded forces, when done on the CPU, here.
1441      * We do this before calling do_force_lowlevel, because in that
1442      * function, the listed forces are calculated before PME, which
1443      * does communication.  With this order, non-bonded and listed
1444      * force calculation imbalance can be balanced out by the domain
1445      * decomposition load balancing.
1446      */
1447
1448     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1449
1450     if (!useOrEmulateGpuNb)
1451     {
1452         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1453     }
1454
1455     if (fr->efep != efepNO)
1456     {
1457         /* Calculate the local and non-local free energy interactions here.
1458          * Happens here on the CPU both with and without GPU.
1459          */
1460         nbv->dispatchFreeEnergyKernel(InteractionLocality::Local, fr,
1461                                       as_rvec_array(x.unpaddedArrayRef().data()),
1462                                       &forceOut.forceWithShiftForces(), *mdatoms, inputrec->fepvals,
1463                                       lambda.data(), enerd, stepWork, nrnb);
1464
1465         if (havePPDomainDecomposition(cr))
1466         {
1467             nbv->dispatchFreeEnergyKernel(InteractionLocality::NonLocal, fr,
1468                                           as_rvec_array(x.unpaddedArrayRef().data()),
1469                                           &forceOut.forceWithShiftForces(), *mdatoms,
1470                                           inputrec->fepvals, lambda.data(), enerd, stepWork, nrnb);
1471         }
1472     }
1473
1474     if (!useOrEmulateGpuNb)
1475     {
1476         if (havePPDomainDecomposition(cr))
1477         {
1478             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1479                          nrnb, wcycle);
1480         }
1481
1482         if (stepWork.computeForces)
1483         {
1484             /* Add all the non-bonded force to the normal force array.
1485              * This can be split into a local and a non-local part when overlapping
1486              * communication with calculation with domain decomposition.
1487              */
1488             wallcycle_stop(wcycle, ewcFORCE);
1489             nbv->atomdata_add_nbat_f_to_f(AtomLocality::All, forceOut.forceWithShiftForces().force());
1490             wallcycle_start_nocount(wcycle, ewcFORCE);
1491         }
1492
1493         /* If there are multiple fshift output buffers we need to reduce them */
1494         if (stepWork.computeVirial)
1495         {
1496             /* This is not in a subcounter because it takes a
1497                negligible and constant-sized amount of time */
1498             nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1499                                                      forceOut.forceWithShiftForces().shiftForces());
1500         }
1501     }
1502
1503     /* update QMMMrec, if necessary */
1504     if (fr->bQMMM)
1505     {
1506         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1507     }
1508
1509     // TODO Force flags should include haveFreeEnergyWork for this domain
1510     if (ddUsesGpuDirectCommunication && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1511     {
1512         /* Wait for non-local coordinate data to be copied from device */
1513         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1514     }
1515     /* Compute the bonded and non-bonded energies and optionally forces */
1516     do_force_lowlevel(fr, inputrec, &(top->idef), cr, ms, nrnb, wcycle, mdatoms, x, hist, &forceOut, enerd,
1517                       fcd, box, lambda.data(), graph, fr->mu_tot, stepWork, ddBalanceRegionHandler);
1518
1519     wallcycle_stop(wcycle, ewcFORCE);
1520
1521     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation, imdSession, pull_work, step, t,
1522                          wcycle, fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1523                          stepWork, &forceOut.forceWithVirial(), enerd, ed, stepWork.doNeighborSearch);
1524
1525
1526     // Will store the amount of cycles spent waiting for the GPU that
1527     // will be later used in the DLB accounting.
1528     float cycles_wait_gpu = 0;
1529     if (useOrEmulateGpuNb)
1530     {
1531         auto& forceWithShiftForces = forceOut.forceWithShiftForces();
1532
1533         /* wait for non-local forces (or calculate in emulation mode) */
1534         if (havePPDomainDecomposition(cr))
1535         {
1536             if (simulationWork.useGpuNonbonded)
1537             {
1538                 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
1539                         nbv->gpu_nbv, stepWork, AtomLocality::NonLocal, enerd->grpp.ener[egLJSR].data(),
1540                         enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(), wcycle);
1541             }
1542             else
1543             {
1544                 wallcycle_start_nocount(wcycle, ewcFORCE);
1545                 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes,
1546                              step, nrnb, wcycle);
1547                 wallcycle_stop(wcycle, ewcFORCE);
1548             }
1549
1550             if (useGpuFBufOps == BufferOpsUseGpu::True)
1551             {
1552                 gmx::FixedCapacityVector<GpuEventSynchronizer*, 1> dependencyList;
1553
1554                 // TODO: move this into DomainLifetimeWorkload, including the second part of the
1555                 // condition The bonded and free energy CPU tasks can have non-local force
1556                 // contributions which are a dependency for the GPU force reduction.
1557                 bool haveNonLocalForceContribInCpuBuffer =
1558                         domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1559
1560                 if (haveNonLocalForceContribInCpuBuffer)
1561                 {
1562                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(),
1563                                               AtomLocality::NonLocal);
1564                     dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(
1565                             AtomLocality::NonLocal, useGpuFBufOps == BufferOpsUseGpu::True));
1566                 }
1567
1568                 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::NonLocal, stateGpu->getForces(),
1569                                                   pme_gpu_get_device_f(fr->pmedata), dependencyList,
1570                                                   false, haveNonLocalForceContribInCpuBuffer);
1571                 if (!useGpuForcesHaloExchange)
1572                 {
1573                     // copy from GPU input for dd_move_f()
1574                     stateGpu->copyForcesFromGpu(forceOut.forceWithShiftForces().force(),
1575                                                 AtomLocality::NonLocal);
1576                 }
1577             }
1578             else
1579             {
1580                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
1581             }
1582
1583
1584             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1585             {
1586                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
1587             }
1588         }
1589     }
1590
1591     if (havePPDomainDecomposition(cr))
1592     {
1593         /* We are done with the CPU compute.
1594          * We will now communicate the non-local forces.
1595          * If we use a GPU this will overlap with GPU work, so in that case
1596          * we do not close the DD force balancing region here.
1597          */
1598         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1599
1600         if (stepWork.computeForces)
1601         {
1602
1603             if (useGpuForcesHaloExchange)
1604             {
1605                 if (domainWork.haveCpuLocalForceWork)
1606                 {
1607                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), AtomLocality::Local);
1608                 }
1609                 gpuHaloExchange->communicateHaloForces(domainWork.haveCpuLocalForceWork);
1610             }
1611             else
1612             {
1613                 if (useGpuFBufOps == BufferOpsUseGpu::True)
1614                 {
1615                     stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
1616                 }
1617                 dd_move_f(cr->dd, &forceOut.forceWithShiftForces(), wcycle);
1618             }
1619         }
1620     }
1621
1622     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1623     // an alternating wait/reduction scheme.
1624     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded
1625                              && !DOMAINDECOMP(cr) && (useGpuFBufOps == BufferOpsUseGpu::False));
1626     if (alternateGpuWait)
1627     {
1628         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &forceOut, enerd, stepWork, pmeFlags, wcycle);
1629     }
1630
1631     if (!alternateGpuWait && useGpuPmeOnThisRank)
1632     {
1633         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial(), enerd);
1634     }
1635
1636     /* Wait for local GPU NB outputs on the non-alternating wait path */
1637     if (!alternateGpuWait && simulationWork.useGpuNonbonded)
1638     {
1639         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1640          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1641          * but even with a step of 0.1 ms the difference is less than 1%
1642          * of the step time.
1643          */
1644         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1645         const float waitCycles               = Nbnxm::gpu_wait_finish_task(
1646                 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
1647                 enerd->grpp.ener[egCOULSR].data(), forceOut.forceWithShiftForces().shiftForces(), wcycle);
1648
1649         if (ddBalanceRegionHandler.useBalancingRegion())
1650         {
1651             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1652             if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
1653             {
1654                 /* We measured few cycles, it could be that the kernel
1655                  * and transfer finished earlier and there was no actual
1656                  * wait time, only API call overhead.
1657                  * Then the actual time could be anywhere between 0 and
1658                  * cycles_wait_est. We will use half of cycles_wait_est.
1659                  */
1660                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1661             }
1662             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1663         }
1664     }
1665
1666     if (fr->nbv->emulateGpu())
1667     {
1668         // NOTE: emulation kernel is not included in the balancing region,
1669         // but emulation mode does not target performance anyway
1670         wallcycle_start_nocount(wcycle, ewcFORCE);
1671         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local,
1672                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes, step, nrnb, wcycle);
1673         wallcycle_stop(wcycle, ewcFORCE);
1674     }
1675
1676     // If on GPU PME-PP comms or GPU update path, receive forces from PME before GPU buffer ops
1677     // TODO refactor this and unify with below default-path call to the same function
1678     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME)
1679         && (simulationWork.useGpuPmePpCommunication || simulationWork.useGpuUpdate))
1680     {
1681         /* In case of node-splitting, the PP nodes receive the long-range
1682          * forces, virial and energy from the PME nodes here.
1683          */
1684         pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1685                                simulationWork.useGpuPmePpCommunication,
1686                                stepWork.useGpuPmeFReduction, wcycle);
1687     }
1688
1689
1690     /* Do the nonbonded GPU (or emulation) force buffer reduction
1691      * on the non-alternating path. */
1692     if (useOrEmulateGpuNb && !alternateGpuWait)
1693     {
1694         // TODO simplify the below conditionals. Pass buffer and sync pointers at init stage rather than here. Unify getter fns for sameGPU/otherGPU cases.
1695         void* pmeForcePtr =
1696                 stepWork.useGpuPmeFReduction
1697                         ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_device_f(fr->pmedata)
1698                                                          : // PME force buffer on same GPU
1699                                    fr->pmePpCommGpu->getGpuForceStagingPtr()) // buffer received from other GPU
1700                         : nullptr; // PME reduction not active on GPU
1701
1702         GpuEventSynchronizer* const pmeSynchronizer =
1703                 stepWork.useGpuPmeFReduction
1704                         ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_f_ready_synchronizer(fr->pmedata)
1705                                                          : // PME force buffer on same GPU
1706                                    static_cast<GpuEventSynchronizer*>(
1707                                            fr->pmePpCommGpu->getForcesReadySynchronizer())) // buffer received from other GPU
1708                         : nullptr; // PME reduction not active on GPU
1709
1710         gmx::FixedCapacityVector<GpuEventSynchronizer*, 3> dependencyList;
1711
1712         if (stepWork.useGpuPmeFReduction)
1713         {
1714             dependencyList.push_back(pmeSynchronizer);
1715         }
1716
1717         gmx::ArrayRef<gmx::RVec> forceWithShift = forceOut.forceWithShiftForces().force();
1718
1719         if (useGpuFBufOps == BufferOpsUseGpu::True)
1720         {
1721             // Flag to specify whether the CPU force buffer has contributions to
1722             // local atoms. This depends on whether there are CPU-based force tasks
1723             // or when DD is active the halo exchange has resulted in contributions
1724             // from the non-local part.
1725             const bool haveLocalForceContribInCpuBuffer =
1726                     (domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr));
1727
1728             // TODO: move these steps as early as possible:
1729             // - CPU f H2D should be as soon as all CPU-side forces are done
1730             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1731             //   before the next CPU task that consumes the forces: vsite spread or update)
1732             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
1733             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
1734             //   These should be unified.
1735             if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
1736             {
1737                 // Note: AtomLocality::All is used for the non-DD case because, as in this
1738                 // case copyForcesToGpu() uses a separate stream, it allows overlap of
1739                 // CPU force H2D with GPU force tasks on all streams including those in the
1740                 // local stream which would otherwise be implicit dependencies for the
1741                 // transfer and would not overlap.
1742                 auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
1743
1744                 stateGpu->copyForcesToGpu(forceWithShift, locality);
1745                 dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(
1746                         locality, useGpuFBufOps == BufferOpsUseGpu::True));
1747             }
1748             if (useGpuForcesHaloExchange)
1749             {
1750                 dependencyList.push_back(gpuHaloExchange->getForcesReadyOnDeviceEvent());
1751             }
1752             nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::Local, stateGpu->getForces(), pmeForcePtr,
1753                                               dependencyList, stepWork.useGpuPmeFReduction,
1754                                               haveLocalForceContribInCpuBuffer);
1755             // Copy forces to host if they are needed for update or if virtual sites are enabled.
1756             // If there are vsites, we need to copy forces every step to spread vsite forces on host.
1757             // TODO: When the output flags will be included in step workload, this copy can be combined with the
1758             //       copy call done in sim_utils(...) for the output.
1759             // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
1760             //       they should not be copied in do_md(...) for the output.
1761             if (!simulationWork.useGpuUpdate || vsite)
1762             {
1763                 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
1764                 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
1765             }
1766         }
1767         else
1768         {
1769             nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
1770         }
1771     }
1772
1773     launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd, *runScheduleWork,
1774                             useGpuPmeOnThisRank, step, wcycle);
1775
1776     if (DOMAINDECOMP(cr))
1777     {
1778         dd_force_flop_stop(cr->dd, nrnb);
1779     }
1780
1781     if (stepWork.computeForces)
1782     {
1783         rvec* f = as_rvec_array(forceOut.forceWithShiftForces().force().data());
1784
1785         /* If we have NoVirSum forces, but we do not calculate the virial,
1786          * we sum fr->f_novirsum=forceOut.f later.
1787          */
1788         if (vsite && !(fr->haveDirectVirialContributions && !stepWork.computeVirial))
1789         {
1790             rvec* fshift = as_rvec_array(forceOut.forceWithShiftForces().shiftForces().data());
1791             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), f, fshift, FALSE, nullptr,
1792                            nrnb, &top->idef, fr->pbcType, fr->bMolPBC, graph, box, cr, wcycle);
1793         }
1794
1795         if (stepWork.computeVirial)
1796         {
1797             /* Calculation of the virial must be done after vsites! */
1798             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()),
1799                         forceOut.forceWithShiftForces(), vir_force, graph, box, nrnb, fr,
1800                         inputrec->pbcType);
1801         }
1802     }
1803
1804     // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
1805     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication
1806         && !simulationWork.useGpuUpdate)
1807     {
1808         /* In case of node-splitting, the PP nodes receive the long-range
1809          * forces, virial and energy from the PME nodes here.
1810          */
1811         pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1812                                simulationWork.useGpuPmePpCommunication, false, wcycle);
1813     }
1814
1815     if (stepWork.computeForces)
1816     {
1817         post_process_forces(cr, step, nrnb, wcycle, top, box, as_rvec_array(x.unpaddedArrayRef().data()),
1818                             &forceOut, vir_force, mdatoms, graph, fr, vsite, stepWork);
1819     }
1820
1821     if (stepWork.computeEnergy)
1822     {
1823         /* Sum the potential energy terms from group contributions */
1824         sum_epot(&(enerd->grpp), enerd->term);
1825
1826         if (!EI_TPI(inputrec->eI))
1827         {
1828             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1829         }
1830     }
1831
1832     /* In case we don't have constraints and are using GPUs, the next balancing
1833      * region starts here.
1834      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1835      * virial calculation and COM pulling, is not thus not included in
1836      * the balance timing, which is ok as most tasks do communication.
1837      */
1838     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1839 }