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) 2013,2014,2015,2016,2017,2018,2019, 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/gmxlib/network.h"
  57 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  58 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  59 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  60 #include "gromacs/gpu_utils/gpu_utils.h"
  61 #include "gromacs/imd/imd.h"
  62 #include "gromacs/listed_forces/disre.h"
  63 #include "gromacs/listed_forces/gpubonded.h"
  64 #include "gromacs/listed_forces/listed_forces.h"
  65 #include "gromacs/listed_forces/manage_threading.h"
  66 #include "gromacs/listed_forces/orires.h"
  67 #include "gromacs/math/arrayrefwithpadding.h"
  68 #include "gromacs/math/functions.h"
  69 #include "gromacs/math/units.h"
  70 #include "gromacs/math/vec.h"
  71 #include "gromacs/math/vecdump.h"
  72 #include "gromacs/mdlib/calcmu.h"
  73 #include "gromacs/mdlib/calcvir.h"
  74 #include "gromacs/mdlib/constr.h"
  75 #include "gromacs/mdlib/enerdata_utils.h"
  76 #include "gromacs/mdlib/force.h"
  77 #include "gromacs/mdlib/forcerec.h"
  78 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  79 #include "gromacs/mdlib/qmmm.h"
  80 #include "gromacs/mdlib/update.h"
  81 #include "gromacs/mdtypes/commrec.h"
  82 #include "gromacs/mdtypes/enerdata.h"
  83 #include "gromacs/mdtypes/forceoutput.h"
  84 #include "gromacs/mdtypes/iforceprovider.h"
  85 #include "gromacs/mdtypes/inputrec.h"
  86 #include "gromacs/mdtypes/md_enums.h"
  87 #include "gromacs/mdtypes/simulation_workload.h"
  88 #include "gromacs/mdtypes/state.h"
  89 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  90 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  91 #include "gromacs/nbnxm/nbnxm.h"
  92 #include "gromacs/pbcutil/ishift.h"
  93 #include "gromacs/pbcutil/mshift.h"
  94 #include "gromacs/pbcutil/pbc.h"
  95 #include "gromacs/pulling/pull.h"
  96 #include "gromacs/pulling/pull_rotation.h"
  97 #include "gromacs/timing/cyclecounter.h"
  98 #include "gromacs/timing/gpu_timing.h"
  99 #include "gromacs/timing/wallcycle.h"
 100 #include "gromacs/timing/wallcyclereporting.h"
 101 #include "gromacs/timing/walltime_accounting.h"
 102 #include "gromacs/topology/topology.h"
 103 #include "gromacs/utility/arrayref.h"
 104 #include "gromacs/utility/basedefinitions.h"
 105 #include "gromacs/utility/cstringutil.h"
 106 #include "gromacs/utility/exceptions.h"
 107 #include "gromacs/utility/fatalerror.h"
 108 #include "gromacs/utility/fixedcapacityvector.h"
 109 #include "gromacs/utility/gmxassert.h"
 110 #include "gromacs/utility/gmxmpi.h"
 111 #include "gromacs/utility/logger.h"
 112 #include "gromacs/utility/smalloc.h"
 113 #include "gromacs/utility/strconvert.h"
 114 #include "gromacs/utility/sysinfo.h"
 115
 116 using gmx::ForceOutputs;
 117 using gmx::StepWorkload;
 118 using gmx::DomainLifetimeWorkload;
 119 using gmx::SimulationWorkload;
 120
 121 // TODO: this environment variable allows us to verify before release
 122 // that on less common architectures the total cost of polling is not larger than
 123 // a blocking wait (so polling does not introduce overhead when the static
 124 // PME-first ordering would suffice).
 125 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 126
 127 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 128 {
 129     const int      end = forceToAdd.size();
 130
 131     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 132 #pragma omp parallel for num_threads(nt) schedule(static)
 133     for (int i = 0; i < end; i++)
 134     {
 135         rvec_inc(f[i], forceToAdd[i]);
 136     }
 137 }
 138
 139 static void calc_virial(int start, int homenr, const rvec x[],
 140                         const gmx::ForceWithShiftForces &forceWithShiftForces,
 141                         tensor vir_part, const t_graph *graph, const matrix box,
 142                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 143 {
 144     /* The short-range virial from surrounding boxes */
 145     const rvec *fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
 146     calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, ePBC == epbcSCREW, box);
 147     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 148
 149     /* Calculate partial virial, for local atoms only, based on short range.
 150      * Total virial is computed in global_stat, called from do_md
 151      */
 152     const rvec *f = as_rvec_array(forceWithShiftForces.force().data());
 153     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 154     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 155
 156     if (debug)
 157     {
 158         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 159     }
 160 }
 161
 162 static void pull_potential_wrapper(const t_commrec *cr,
 163                                    const t_inputrec *ir,
 164                                    const matrix box, gmx::ArrayRef<const gmx::RVec> x,
 165                                    gmx::ForceWithVirial *force,
 166                                    const t_mdatoms *mdatoms,
 167                                    gmx_enerdata_t *enerd,
 168                                    pull_t *pull_work,
 169                                    const real *lambda,
 170                                    double t,
 171                                    gmx_wallcycle_t wcycle)
 172 {
 173     t_pbc  pbc;
 174     real   dvdl;
 175
 176     /* Calculate the center of mass forces, this requires communication,
 177      * which is why pull_potential is called close to other communication.
 178      */
 179     wallcycle_start(wcycle, ewcPULLPOT);
 180     set_pbc(&pbc, ir->ePBC, box);
 181     dvdl                     = 0;
 182     enerd->term[F_COM_PULL] +=
 183         pull_potential(pull_work, mdatoms, &pbc,
 184                        cr, t, lambda[efptRESTRAINT], as_rvec_array(x.data()), force, &dvdl);
 185     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 186     wallcycle_stop(wcycle, ewcPULLPOT);
 187 }
 188
 189 static void pme_receive_force_ener(const t_commrec      *cr,
 190                                    gmx::ForceWithVirial *forceWithVirial,
 191                                    gmx_enerdata_t       *enerd,
 192                                    gmx_wallcycle_t       wcycle)
 193 {
 194     real   e_q, e_lj, dvdl_q, dvdl_lj;
 195     float  cycles_ppdpme, cycles_seppme;
 196
 197     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 198     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 199
 200     /* In case of node-splitting, the PP nodes receive the long-range
 201      * forces, virial and energy from the PME nodes here.
 202      */
 203     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 204     dvdl_q  = 0;
 205     dvdl_lj = 0;
 206     gmx_pme_receive_f(cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 207                       &cycles_seppme);
 208     enerd->term[F_COUL_RECIP] += e_q;
 209     enerd->term[F_LJ_RECIP]   += e_lj;
 210     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 211     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 212
 213     if (wcycle)
 214     {
 215         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 216     }
 217     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 218 }
 219
 220 static void print_large_forces(FILE            *fp,
 221                                const t_mdatoms *md,
 222                                const t_commrec *cr,
 223                                int64_t          step,
 224                                real             forceTolerance,
 225                                const rvec      *x,
 226                                const rvec      *f)
 227 {
 228     real           force2Tolerance = gmx::square(forceTolerance);
 229     gmx::index     numNonFinite    = 0;
 230     for (int i = 0; i < md->homenr; i++)
 231     {
 232         real force2    = norm2(f[i]);
 233         bool nonFinite = !std::isfinite(force2);
 234         if (force2 >= force2Tolerance || nonFinite)
 235         {
 236             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 237                     step,
 238                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 239         }
 240         if (nonFinite)
 241         {
 242             numNonFinite++;
 243         }
 244     }
 245     if (numNonFinite > 0)
 246     {
 247         /* Note that with MPI this fatal call on one rank might interrupt
 248          * the printing on other ranks. But we can only avoid that with
 249          * an expensive MPI barrier that we would need at each step.
 250          */
 251         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 252     }
 253 }
 254
 255 static void post_process_forces(const t_commrec       *cr,
 256                                 int64_t                step,
 257                                 t_nrnb                *nrnb,
 258                                 gmx_wallcycle_t        wcycle,
 259                                 const gmx_localtop_t  *top,
 260                                 const matrix           box,
 261                                 const rvec             x[],
 262                                 ForceOutputs          *forceOutputs,
 263                                 tensor                 vir_force,
 264                                 const t_mdatoms       *mdatoms,
 265                                 const t_graph         *graph,
 266                                 const t_forcerec      *fr,
 267                                 const gmx_vsite_t     *vsite,
 268                                 const StepWorkload    &stepWork)
 269 {
 270     rvec *f = as_rvec_array(forceOutputs->forceWithShiftForces().force().data());
 271
 272     if (fr->haveDirectVirialContributions)
 273     {
 274         auto &forceWithVirial = forceOutputs->forceWithVirial();
 275         rvec *fDirectVir      = as_rvec_array(forceWithVirial.force_.data());
 276
 277         if (vsite)
 278         {
 279             /* Spread the mesh force on virtual sites to the other particles...
 280              * This is parallellized. MPI communication is performed
 281              * if the constructing atoms aren't local.
 282              */
 283             matrix virial = { { 0 } };
 284             spread_vsite_f(vsite, x, fDirectVir, nullptr,
 285                            stepWork.computeVirial, virial,
 286                            nrnb,
 287                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
 288             forceWithVirial.addVirialContribution(virial);
 289         }
 290
 291         if (stepWork.computeVirial)
 292         {
 293             /* Now add the forces, this is local */
 294             sum_forces(f, forceWithVirial.force_);
 295
 296             /* Add the direct virial contributions */
 297             GMX_ASSERT(forceWithVirial.computeVirial_, "forceWithVirial should request virial computation when we request the virial");
 298             m_add(vir_force, forceWithVirial.getVirial(), vir_force);
 299
 300             if (debug)
 301             {
 302                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 303             }
 304         }
 305     }
 306
 307     if (fr->print_force >= 0)
 308     {
 309         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 310     }
 311 }
 312
 313 static void do_nb_verlet(t_forcerec                       *fr,
 314                          const interaction_const_t        *ic,
 315                          gmx_enerdata_t                   *enerd,
 316                          const StepWorkload               &stepWork,
 317                          const Nbnxm::InteractionLocality  ilocality,
 318                          const int                         clearF,
 319                          const int64_t                     step,
 320                          t_nrnb                           *nrnb,
 321                          gmx_wallcycle_t                   wcycle)
 322 {
 323     if (!stepWork.computeNonbondedForces)
 324     {
 325         /* skip non-bonded calculation */
 326         return;
 327     }
 328
 329     nonbonded_verlet_t *nbv  = fr->nbv.get();
 330
 331     /* GPU kernel launch overhead is already timed separately */
 332     if (fr->cutoff_scheme != ecutsVERLET)
 333     {
 334         gmx_incons("Invalid cut-off scheme passed!");
 335     }
 336
 337     if (!nbv->useGpu())
 338     {
 339         /* When dynamic pair-list  pruning is requested, we need to prune
 340          * at nstlistPrune steps.
 341          */
 342         if (nbv->isDynamicPruningStepCpu(step))
 343         {
 344             /* Prune the pair-list beyond fr->ic->rlistPrune using
 345              * the current coordinates of the atoms.
 346              */
 347             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 348             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 349             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 350         }
 351     }
 352
 353     nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
 354 }
 355
 356 static inline void clear_rvecs_omp(int n, rvec v[])
 357 {
 358     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 359
 360     /* Note that we would like to avoid this conditional by putting it
 361      * into the omp pragma instead, but then we still take the full
 362      * omp parallel for overhead (at least with gcc5).
 363      */
 364     if (nth == 1)
 365     {
 366         for (int i = 0; i < n; i++)
 367         {
 368             clear_rvec(v[i]);
 369         }
 370     }
 371     else
 372     {
 373 #pragma omp parallel for num_threads(nth) schedule(static)
 374         for (int i = 0; i < n; i++)
 375         {
 376             clear_rvec(v[i]);
 377         }
 378     }
 379 }
 380
 381 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 382  *
 383  * \param groupOptions  Group options, containing T-coupling options
 384  */
 385 static real averageKineticEnergyEstimate(const t_grpopts &groupOptions)
 386 {
 387     real nrdfCoupled   = 0;
 388     real nrdfUncoupled = 0;
 389     real kineticEnergy = 0;
 390     for (int g = 0; g < groupOptions.ngtc; g++)
 391     {
 392         if (groupOptions.tau_t[g] >= 0)
 393         {
 394             nrdfCoupled   += groupOptions.nrdf[g];
 395             kineticEnergy += groupOptions.nrdf[g]*0.5*groupOptions.ref_t[g]*BOLTZ;
 396         }
 397         else
 398         {
 399             nrdfUncoupled += groupOptions.nrdf[g];
 400         }
 401     }
 402
 403     /* This conditional with > also catches nrdf=0 */
 404     if (nrdfCoupled > nrdfUncoupled)
 405     {
 406         return kineticEnergy*(nrdfCoupled + nrdfUncoupled)/nrdfCoupled;
 407     }
 408     else
 409     {
 410         return 0;
 411     }
 412 }
 413
 414 /*! \brief This routine checks that the potential energy is finite.
 415  *
 416  * Always checks that the potential energy is finite. If step equals
 417  * inputrec.init_step also checks that the magnitude of the potential energy
 418  * is reasonable. Terminates with a fatal error when a check fails.
 419  * Note that passing this check does not guarantee finite forces,
 420  * since those use slightly different arithmetics. But in most cases
 421  * there is just a narrow coordinate range where forces are not finite
 422  * and energies are finite.
 423  *
 424  * \param[in] step      The step number, used for checking and printing
 425  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to enerd.term[F_EPOT] before calling this routine
 426  * \param[in] inputrec  The input record
 427  */
 428 static void checkPotentialEnergyValidity(int64_t               step,
 429                                          const gmx_enerdata_t &enerd,
 430                                          const t_inputrec     &inputrec)
 431 {
 432     /* Threshold valid for comparing absolute potential energy against
 433      * the kinetic energy. Normally one should not consider absolute
 434      * potential energy values, but with a factor of one million
 435      * we should never get false positives.
 436      */
 437     constexpr real c_thresholdFactor = 1e6;
 438
 439     bool           energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 440     real           averageKineticEnergy = 0;
 441     /* We only check for large potential energy at the initial step,
 442      * because that is by far the most likely step for this too occur
 443      * and because computing the average kinetic energy is not free.
 444      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 445      * before they become NaN.
 446      */
 447     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 448     {
 449         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 450     }
 451
 452     if (energyIsNotFinite || (averageKineticEnergy > 0 &&
 453                               enerd.term[F_EPOT] > c_thresholdFactor*averageKineticEnergy))
 454     {
 455         gmx_fatal(FARGS, "Step %" PRId64 ": The total potential energy is %g, which is %s. The LJ and electrostatic contributions to the energy are %g and %g, respectively. A %s potential energy can be caused by overlapping interactions in bonded interactions or very large%s coordinate values. Usually this is caused by a badly- or non-equilibrated initial configuration, incorrect interactions or parameters in the topology.",
 456                   step,
 457                   enerd.term[F_EPOT],
 458                   energyIsNotFinite ? "not finite" : "extremely high",
 459                   enerd.term[F_LJ],
 460                   enerd.term[F_COUL_SR],
 461                   energyIsNotFinite ? "non-finite" : "very high",
 462                   energyIsNotFinite ? " or Nan" : "");
 463     }
 464 }
 465
 466 /*! \brief Return true if there are special forces computed this step.
 467  *
 468  * The conditionals exactly correspond to those in computeSpecialForces().
 469  */
 470 static bool
 471 haveSpecialForces(const t_inputrec              &inputrec,
 472                   const gmx::ForceProviders     &forceProviders,
 473                   const pull_t                  *pull_work,
 474                   const bool                     computeForces,
 475                   const gmx_edsam               *ed)
 476 {
 477
 478     return
 479         ((computeForces && forceProviders.hasForceProvider()) ||         // forceProviders
 480          (inputrec.bPull && pull_have_potential(pull_work)) ||           // pull
 481          inputrec.bRot ||                                                // enforced rotation
 482          (ed != nullptr) ||                                              // flooding
 483          (inputrec.bIMD && computeForces));                              // IMD
 484 }
 485
 486 /*! \brief Compute forces and/or energies for special algorithms
 487  *
 488  * The intention is to collect all calls to algorithms that compute
 489  * forces on local atoms only and that do not contribute to the local
 490  * virial sum (but add their virial contribution separately).
 491  * Eventually these should likely all become ForceProviders.
 492  * Within this function the intention is to have algorithms that do
 493  * global communication at the end, so global barriers within the MD loop
 494  * are as close together as possible.
 495  *
 496  * \param[in]     fplog            The log file
 497  * \param[in]     cr               The communication record
 498  * \param[in]     inputrec         The input record
 499  * \param[in]     awh              The Awh module (nullptr if none in use).
 500  * \param[in]     enforcedRotation Enforced rotation module.
 501  * \param[in]     imdSession       The IMD session
 502  * \param[in]     pull_work        The pull work structure.
 503  * \param[in]     step             The current MD step
 504  * \param[in]     t                The current time
 505  * \param[in,out] wcycle           Wallcycle accounting struct
 506  * \param[in,out] forceProviders   Pointer to a list of force providers
 507  * \param[in]     box              The unit cell
 508  * \param[in]     x                The coordinates
 509  * \param[in]     mdatoms          Per atom properties
 510  * \param[in]     lambda           Array of free-energy lambda values
 511  * \param[in]     stepWork         Step schedule flags
 512  * \param[in,out] forceWithVirial  Force and virial buffers
 513  * \param[in,out] enerd            Energy buffer
 514  * \param[in,out] ed               Essential dynamics pointer
 515  * \param[in]     didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
 516  *
 517  * \todo Remove didNeighborSearch, which is used incorrectly.
 518  * \todo Convert all other algorithms called here to ForceProviders.
 519  */
 520 static void
 521 computeSpecialForces(FILE                          *fplog,
 522                      const t_commrec               *cr,
 523                      const t_inputrec              *inputrec,
 524                      gmx::Awh                      *awh,
 525                      gmx_enfrot                    *enforcedRotation,
 526                      gmx::ImdSession               *imdSession,
 527                      pull_t                        *pull_work,
 528                      int64_t                        step,
 529                      double                         t,
 530                      gmx_wallcycle_t                wcycle,
 531                      gmx::ForceProviders           *forceProviders,
 532                      const matrix                   box,
 533                      gmx::ArrayRef<const gmx::RVec> x,
 534                      const t_mdatoms               *mdatoms,
 535                      real                          *lambda,
 536                      const StepWorkload            &stepWork,
 537                      gmx::ForceWithVirial          *forceWithVirial,
 538                      gmx_enerdata_t                *enerd,
 539                      gmx_edsam                     *ed,
 540                      bool                           didNeighborSearch)
 541 {
 542     /* NOTE: Currently all ForceProviders only provide forces.
 543      *       When they also provide energies, remove this conditional.
 544      */
 545     if (stepWork.computeForces)
 546     {
 547         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 548         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 549
 550         /* Collect forces from modules */
 551         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 552     }
 553
 554     if (inputrec->bPull && pull_have_potential(pull_work))
 555     {
 556         pull_potential_wrapper(cr, inputrec, box, x,
 557                                forceWithVirial,
 558                                mdatoms, enerd, pull_work, lambda, t,
 559                                wcycle);
 560
 561         if (awh)
 562         {
 563             enerd->term[F_COM_PULL] +=
 564                 awh->applyBiasForcesAndUpdateBias(inputrec->ePBC, *mdatoms, box,
 565                                                   forceWithVirial,
 566                                                   t, step, wcycle, fplog);
 567         }
 568     }
 569
 570     rvec *f = as_rvec_array(forceWithVirial->force_.data());
 571
 572     /* Add the forces from enforced rotation potentials (if any) */
 573     if (inputrec->bRot)
 574     {
 575         wallcycle_start(wcycle, ewcROTadd);
 576         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 577         wallcycle_stop(wcycle, ewcROTadd);
 578     }
 579
 580     if (ed)
 581     {
 582         /* Note that since init_edsam() is called after the initialization
 583          * of forcerec, edsam doesn't request the noVirSum force buffer.
 584          * Thus if no other algorithm (e.g. PME) requires it, the forces
 585          * here will contribute to the virial.
 586          */
 587         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
 588     }
 589
 590     /* Add forces from interactive molecular dynamics (IMD), if any */
 591     if (inputrec->bIMD && stepWork.computeForces)
 592     {
 593         imdSession->applyForces(f);
 594     }
 595 }
 596
 597 /*! \brief Makes PME flags from StepWorkload data.
 598  *
 599  * \param[in]  stepWork     Step schedule flags
 600  * \returns                 PME flags
 601  */
 602 static int makePmeFlags(const StepWorkload &stepWork)
 603 {
 604     return GMX_PME_SPREAD | GMX_PME_SOLVE |
 605            (stepWork.computeVirial ? GMX_PME_CALC_ENER_VIR : 0) |
 606            (stepWork.computeEnergy ? GMX_PME_CALC_ENER_VIR : 0) |
 607            (stepWork.computeForces ? GMX_PME_CALC_F        : 0);
 608 }
 609
 610 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 611  *
 612  * \param[in]  pmedata              The PME structure
 613  * \param[in]  box                  The box matrix
 614  * \param[in]  stepWork             Step schedule flags
 615  * \param[in]  pmeFlags             PME flags
 616  * \param[in]  xReadyOnDevice       Event synchronizer indicating that the coordinates are ready in the device memory.
 617  * \param[in]  wcycle               The wallcycle structure
 618  */
 619 static inline void launchPmeGpuSpread(gmx_pme_t            *pmedata,
 620                                       const matrix          box,
 621                                       const StepWorkload   &stepWork,
 622                                       int                   pmeFlags,
 623                                       GpuEventSynchronizer *xReadyOnDevice,
 624                                       gmx_wallcycle_t       wcycle)
 625 {
 626     pme_gpu_prepare_computation(pmedata, stepWork.haveDynamicBox, box, wcycle, pmeFlags, stepWork.useGpuPmeFReduction);
 627     pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle);
 628 }
 629
 630 /*! \brief Launch the FFT and gather stages of PME GPU
 631  *
 632  * This function only implements setting the output forces (no accumulation).
 633  *
 634  * \param[in]  pmedata        The PME structure
 635  * \param[in]  wcycle         The wallcycle structure
 636  */
 637 static void launchPmeGpuFftAndGather(gmx_pme_t        *pmedata,
 638                                      gmx_wallcycle_t   wcycle)
 639 {
 640     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 641     pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
 642 }
 643
 644 /*! \brief
 645  *  Polling wait for either of the PME or nonbonded GPU tasks.
 646  *
 647  * Instead of a static order in waiting for GPU tasks, this function
 648  * polls checking which of the two tasks completes first, and does the
 649  * associated force buffer reduction overlapped with the other task.
 650  * By doing that, unlike static scheduling order, it can always overlap
 651  * one of the reductions, regardless of the GPU task completion order.
 652  *
 653  * \param[in]     nbv              Nonbonded verlet structure
 654  * \param[in,out] pmedata          PME module data
 655  * \param[in,out] forceOutputs     Output buffer for the forces and virial
 656  * \param[in,out] enerd            Energy data structure results are reduced into
 657  * \param[in]     stepWork         Step schedule flags
 658  * \param[in]     pmeFlags         PME flags
 659  * \param[in]     wcycle           The wallcycle structure
 660  */
 661 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t *nbv,
 662                                         gmx_pme_t          *pmedata,
 663                                         gmx::ForceOutputs  *forceOutputs,
 664                                         gmx_enerdata_t     *enerd,
 665                                         const StepWorkload &stepWork,
 666                                         int                 pmeFlags,
 667                                         gmx_wallcycle_t     wcycle)
 668 {
 669     bool isPmeGpuDone = false;
 670     bool isNbGpuDone  = false;
 671
 672
 673
 674     gmx::ForceWithShiftForces      &forceWithShiftForces = forceOutputs->forceWithShiftForces();
 675     gmx::ForceWithVirial           &forceWithVirial      = forceOutputs->forceWithVirial();
 676
 677     gmx::ArrayRef<const gmx::RVec>  pmeGpuForces;
 678
 679     while (!isPmeGpuDone || !isNbGpuDone)
 680     {
 681         if (!isPmeGpuDone)
 682         {
 683             GpuTaskCompletion completionType = (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 684             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, &forceWithVirial, enerd, completionType);
 685         }
 686
 687         if (!isNbGpuDone)
 688         {
 689             GpuTaskCompletion completionType = (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 690             isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
 691                                                      stepWork,
 692                                                      Nbnxm::AtomLocality::Local,
 693                                                      enerd->grpp.ener[egLJSR].data(),
 694                                                      enerd->grpp.ener[egCOULSR].data(),
 695                                                      forceWithShiftForces.shiftForces(), completionType, wcycle);
 696
 697             if (isNbGpuDone)
 698             {
 699                 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
 700                                               forceWithShiftForces.force());
 701             }
 702         }
 703     }
 704 }
 705
 706 /*! \brief Set up the different force buffers; also does clearing.
 707  *
 708  * \param[in] fr        force record pointer
 709  * \param[in] pull_work The pull work object.
 710  * \param[in] inputrec  input record
 711  * \param[in] force     force array
 712  * \param[in] stepWork  Step schedule flags
 713  * \param[out] wcycle   wallcycle recording structure
 714  *
 715  * \returns             Cleared force output structure
 716  */
 717 static ForceOutputs
 718 setupForceOutputs(t_forcerec                          *fr,
 719                   pull_t                              *pull_work,
 720                   const t_inputrec                    &inputrec,
 721                   gmx::ArrayRefWithPadding<gmx::RVec>  force,
 722                   const StepWorkload                  &stepWork,
 723                   gmx_wallcycle_t                      wcycle)
 724 {
 725     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 726
 727     /* NOTE: We assume fr->shiftForces is all zeros here */
 728     gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial, fr->shiftForces);
 729
 730     if (stepWork.computeForces)
 731     {
 732         /* Clear the short- and long-range forces */
 733         clear_rvecs_omp(fr->natoms_force_constr,
 734                         as_rvec_array(forceWithShiftForces.force().data()));
 735     }
 736
 737     /* If we need to compute the virial, we might need a separate
 738      * force buffer for algorithms for which the virial is calculated
 739      * directly, such as PME. Otherwise, forceWithVirial uses the
 740      * the same force (f in legacy calls) buffer as other algorithms.
 741      */
 742     const bool useSeparateForceWithVirialBuffer = (stepWork.computeForces &&
 743                                                    (stepWork.computeVirial && fr->haveDirectVirialContributions));
 744     /* forceWithVirial uses the local atom range only */
 745     gmx::ForceWithVirial forceWithVirial(useSeparateForceWithVirialBuffer ?
 746                                          fr->forceBufferForDirectVirialContributions : force.unpaddedArrayRef(),
 747                                          stepWork.computeVirial);
 748
 749     if (useSeparateForceWithVirialBuffer)
 750     {
 751         /* TODO: update comment
 752          * We only compute forces on local atoms. Note that vsites can
 753          * spread to non-local atoms, but that part of the buffer is
 754          * cleared separately in the vsite spreading code.
 755          */
 756         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 757     }
 758
 759     if (inputrec.bPull && pull_have_constraint(pull_work))
 760     {
 761         clear_pull_forces(pull_work);
 762     }
 763
 764     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 765
 766     return ForceOutputs(forceWithShiftForces, forceWithVirial);
 767 }
 768
 769
 770 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
 771  */
 772 static DomainLifetimeWorkload
 773 setupDomainLifetimeWorkload(const t_inputrec       &inputrec,
 774                             const t_forcerec       &fr,
 775                             const pull_t           *pull_work,
 776                             const gmx_edsam        *ed,
 777                             const t_idef           &idef,
 778                             const t_fcdata         &fcd,
 779                             const t_mdatoms        &mdatoms,
 780                             const StepWorkload     &stepWork)
 781 {
 782     DomainLifetimeWorkload domainWork;
 783     // Note that haveSpecialForces is constant over the whole run
 784     domainWork.haveSpecialForces      = haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 785     domainWork.haveCpuBondedWork      = haveCpuBondeds(fr);
 786     domainWork.haveGpuBondedWork      = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
 787     domainWork.haveRestraintsWork     = havePositionRestraints(idef, fcd);
 788     domainWork.haveCpuListedForceWork = haveCpuListedForces(fr, idef, fcd);
 789     // Note that haveFreeEnergyWork is constant over the whole run
 790     domainWork.haveFreeEnergyWork     = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
 791     return domainWork;
 792 }
 793
 794 /*! \brief Set up force flag stuct from the force bitmask.
 795  *
 796  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 797  * \param[in]      isNonbondedOn        Global override, if false forces to turn off all nonbonded calculation.
 798  * \param[in]      simulationWork       Simulation workload description.
 799  * \param[in]      rankHasPmeDuty       If this rank computes PME.
 800  *
 801  * \returns New Stepworkload description.
 802  */
 803 static StepWorkload
 804 setupStepWorkload(const int                 legacyFlags,
 805                   const bool                isNonbondedOn,
 806                   const SimulationWorkload &simulationWork,
 807                   const bool                rankHasPmeDuty)
 808 {
 809     StepWorkload flags;
 810     flags.stateChanged           = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 811     flags.haveDynamicBox         = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 812     flags.doNeighborSearch       = ((legacyFlags & GMX_FORCE_NS) != 0);
 813     flags.computeVirial          = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 814     flags.computeEnergy          = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 815     flags.computeForces          = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 816     flags.computeListedForces    = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 817     flags.computeNonbondedForces = ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && isNonbondedOn;
 818     flags.computeDhdl            = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 819
 820     if (simulationWork.useGpuBufferOps)
 821     {
 822         GMX_ASSERT(simulationWork.useGpuNonbonded, "Can only offload buffer ops if nonbonded computation is also offloaded");
 823     }
 824     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 825     // on virial steps the CPU reduction path is taken
 826     // TODO: remove flags.computeEnergy, ref #3128
 827     flags.useGpuFBufferOps    = simulationWork.useGpuBufferOps && !(flags.computeVirial || flags.computeEnergy);
 828     flags.useGpuPmeFReduction = flags.useGpuFBufferOps && (simulationWork.usePmeGpu && rankHasPmeDuty);
 829
 830     return flags;
 831 }
 832
 833
 834 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
 835  *
 836  * TODO: eliminate the \p useGpuNonbonded and \p useGpuNonbonded when these are
 837  * incorporated in DomainLifetimeWorkload.
 838  */
 839 static void
 840 launchGpuEndOfStepTasks(nonbonded_verlet_t               *nbv,
 841                         gmx::GpuBonded                   *gpuBonded,
 842                         gmx_pme_t                        *pmedata,
 843                         gmx_enerdata_t                   *enerd,
 844                         const gmx::MdrunScheduleWorkload &runScheduleWork,
 845                         bool                              useGpuNonbonded,
 846                         bool                              useGpuPme,
 847                         int64_t                           step,
 848                         gmx_wallcycle_t                   wcycle)
 849 {
 850     if (useGpuNonbonded)
 851     {
 852         /* Launch pruning before buffer clearing because the API overhead of the
 853          * clear kernel launches can leave the GPU idle while it could be running
 854          * the prune kernel.
 855          */
 856         if (nbv->isDynamicPruningStepGpu(step))
 857         {
 858             nbv->dispatchPruneKernelGpu(step);
 859         }
 860
 861         /* now clear the GPU outputs while we finish the step on the CPU */
 862         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 863         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 864         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
 865         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 866         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 867     }
 868
 869     if (useGpuPme)
 870     {
 871         pme_gpu_reinit_computation(pmedata, wcycle);
 872     }
 873
 874     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
 875     {
 876         // in principle this should be included in the DD balancing region,
 877         // but generally it is infrequent so we'll omit it for the sake of
 878         // simpler code
 879         gpuBonded->waitAccumulateEnergyTerms(enerd);
 880
 881         gpuBonded->clearEnergies();
 882     }
 883 }
 884
 885
 886 void do_force(FILE                                     *fplog,
 887               const t_commrec                          *cr,
 888               const gmx_multisim_t                     *ms,
 889               const t_inputrec                         *inputrec,
 890               gmx::Awh                                 *awh,
 891               gmx_enfrot                               *enforcedRotation,
 892               gmx::ImdSession                          *imdSession,
 893               pull_t                                   *pull_work,
 894               int64_t                                   step,
 895               t_nrnb                                   *nrnb,
 896               gmx_wallcycle_t                           wcycle,
 897               const gmx_localtop_t                     *top,
 898               const matrix                              box,
 899               gmx::ArrayRefWithPadding<gmx::RVec>       x,
 900               history_t                                *hist,
 901               gmx::ArrayRefWithPadding<gmx::RVec>       force,
 902               tensor                                    vir_force,
 903               const t_mdatoms                          *mdatoms,
 904               gmx_enerdata_t                           *enerd,
 905               t_fcdata                                 *fcd,
 906               gmx::ArrayRef<real>                       lambda,
 907               t_graph                                  *graph,
 908               t_forcerec                               *fr,
 909               gmx::MdrunScheduleWorkload               *runScheduleWork,
 910               const gmx_vsite_t                        *vsite,
 911               rvec                                      mu_tot,
 912               double                                    t,
 913               gmx_edsam                                *ed,
 914               int                                       legacyFlags,
 915               const DDBalanceRegionHandler             &ddBalanceRegionHandler)
 916 {
 917     int                          i, j;
 918     double                       mu[2*DIM];
 919     nonbonded_verlet_t          *nbv      = fr->nbv.get();
 920     interaction_const_t         *ic       = fr->ic;
 921     gmx::StatePropagatorDataGpu *stateGpu = fr->stateGpu;
 922
 923     // TODO remove the code below when the legacy flags are not in use anymore
 924     /* modify force flag if not doing nonbonded */
 925     if (!fr->bNonbonded)
 926     {
 927         legacyFlags &= ~GMX_FORCE_NONBONDED;
 928     }
 929
 930     const SimulationWorkload &simulationWork = runScheduleWork->simulationWork;
 931
 932
 933     runScheduleWork->stepWork = setupStepWorkload(legacyFlags, fr->bNonbonded,
 934                                                   simulationWork, thisRankHasDuty(cr, DUTY_PME));
 935     const StepWorkload &stepWork = runScheduleWork->stepWork;
 936
 937
 938     const bool useGpuPmeOnThisRank = simulationWork.usePmeGpu && thisRankHasDuty(cr, DUTY_PME);
 939     const int  pmeFlags            = makePmeFlags(stepWork);
 940
 941     // Switches on whether to use GPU for position and force buffer operations
 942     // TODO consider all possible combinations of triggers, and how to combine optimally in each case.
 943     const BufferOpsUseGpu useGpuXBufOps = stepWork.useGpuXBufferOps ? BufferOpsUseGpu::True : BufferOpsUseGpu::False;
 944     // GPU Force buffer ops are disabled on virial steps, because the virial calc is not yet ported to GPU
 945     const BufferOpsUseGpu useGpuFBufOps = stepWork.useGpuFBufferOps ? BufferOpsUseGpu::True : BufferOpsUseGpu::False;
 946
 947     /* At a search step we need to start the first balancing region
 948      * somewhere early inside the step after communication during domain
 949      * decomposition (and not during the previous step as usual).
 950      */
 951     if (stepWork.doNeighborSearch)
 952     {
 953         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 954     }
 955
 956     const int start  = 0;
 957     const int homenr = mdatoms->homenr;
 958
 959     clear_mat(vir_force);
 960
 961     if (stepWork.stateChanged)
 962     {
 963         if (inputrecNeedMutot(inputrec))
 964         {
 965             /* Calculate total (local) dipole moment in a temporary common array.
 966              * This makes it possible to sum them over nodes faster.
 967              */
 968             calc_mu(start, homenr,
 969                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 970                     mu, mu+DIM);
 971         }
 972     }
 973
 974     if (fr->ePBC != epbcNONE)
 975     {
 976         /* Compute shift vectors every step,
 977          * because of pressure coupling or box deformation!
 978          */
 979         if (stepWork.haveDynamicBox && stepWork.stateChanged)
 980         {
 981             calc_shifts(box, fr->shift_vec);
 982         }
 983
 984         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
 985         const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
 986         if (calcCGCM)
 987         {
 988             put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr), gmx_omp_nthreads_get(emntDefault));
 989             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 990         }
 991         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 992         {
 993             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
 994         }
 995     }
 996
 997     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox,
 998                                  fr->shift_vec, nbv->nbat.get());
 999
1000 #if GMX_MPI
1001     if (!thisRankHasDuty(cr, DUTY_PME))
1002     {
1003         /* Send particle coordinates to the pme nodes.
1004          * Since this is only implemented for domain decomposition
1005          * and domain decomposition does not use the graph,
1006          * we do not need to worry about shifting.
1007          */
1008         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
1009                                  lambda[efptCOUL], lambda[efptVDW],
1010                                  (stepWork.computeVirial || stepWork.computeEnergy),
1011                                  step, wcycle);
1012     }
1013 #endif /* GMX_MPI */
1014
1015     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1016     // The local coordinates can be copied right away.
1017     // NOTE: Consider moving this copy to right after they are updated and constrained,
1018     //       if the later is not offloaded.
1019     if (useGpuPmeOnThisRank || useGpuXBufOps == BufferOpsUseGpu::True)
1020     {
1021         if (stepWork.doNeighborSearch)
1022         {
1023             stateGpu->reinit(mdatoms->homenr, cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1024             if (useGpuPmeOnThisRank)
1025             {
1026                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1027                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1028             }
1029         }
1030         stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), gmx::StatePropagatorDataGpu::AtomLocality::Local);
1031     }
1032
1033     const auto localXReadyOnDevice = (stateGpu != nullptr) ? stateGpu->getCoordinatesReadyOnDeviceEvent(gmx::StatePropagatorDataGpu::AtomLocality::Local,
1034                                                                                                         simulationWork, stepWork) : nullptr;
1035     if (useGpuPmeOnThisRank)
1036     {
1037         launchPmeGpuSpread(fr->pmedata, box, stepWork, pmeFlags,
1038                            localXReadyOnDevice, wcycle);
1039     }
1040
1041     /* do gridding for pair search */
1042     if (stepWork.doNeighborSearch)
1043     {
1044         if (graph && stepWork.stateChanged)
1045         {
1046             /* Calculate intramolecular shift vectors to make molecules whole */
1047             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
1048         }
1049
1050         // TODO
1051         // - vzero is constant, do we need to pass it?
1052         // - box_diag should be passed directly to nbnxn_put_on_grid
1053         //
1054         rvec vzero;
1055         clear_rvec(vzero);
1056
1057         rvec box_diag;
1058         box_diag[XX] = box[XX][XX];
1059         box_diag[YY] = box[YY][YY];
1060         box_diag[ZZ] = box[ZZ][ZZ];
1061
1062         wallcycle_start(wcycle, ewcNS);
1063         if (!DOMAINDECOMP(cr))
1064         {
1065             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1066             nbnxn_put_on_grid(nbv, box,
1067                               0, vzero, box_diag,
1068                               nullptr, 0, mdatoms->homenr, -1,
1069                               fr->cginfo, x.unpaddedArrayRef(),
1070                               0, nullptr);
1071             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1072         }
1073         else
1074         {
1075             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1076             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
1077                                        fr->cginfo, x.unpaddedArrayRef());
1078             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1079         }
1080
1081         nbv->setAtomProperties(*mdatoms, fr->cginfo);
1082
1083         wallcycle_stop(wcycle, ewcNS);
1084
1085         /* initialize the GPU nbnxm atom data and bonded data structures */
1086         if (simulationWork.useGpuNonbonded)
1087         {
1088             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1089
1090             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1091             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1092             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1093
1094             if (fr->gpuBonded)
1095             {
1096                 /* Now we put all atoms on the grid, we can assign bonded
1097                  * interactions to the GPU, where the grid order is
1098                  * needed. Also the xq, f and fshift device buffers have
1099                  * been reallocated if needed, so the bonded code can
1100                  * learn about them. */
1101                 // TODO the xq, f, and fshift buffers are now shared
1102                 // resources, so they should be maintained by a
1103                 // higher-level object than the nb module.
1104                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbv->getGridIndices(),
1105                                                                       top->idef,
1106                                                                       Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1107                                                                       Nbnxm::gpu_get_f(nbv->gpu_nbv),
1108                                                                       Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1109             }
1110             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1111         }
1112     }
1113
1114     if (stepWork.doNeighborSearch)
1115     {
1116         // Need to run after the GPU-offload bonded interaction lists
1117         // are set up to be able to determine whether there is bonded work.
1118         runScheduleWork->domainWork =
1119             setupDomainLifetimeWorkload(*inputrec,
1120                                         *fr,
1121                                         pull_work,
1122                                         ed,
1123                                         top->idef,
1124                                         *fcd,
1125                                         *mdatoms,
1126                                         stepWork);
1127
1128         wallcycle_start_nocount(wcycle, ewcNS);
1129         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1130         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1131         nbv->constructPairlist(Nbnxm::InteractionLocality::Local,
1132                                &top->excls, step, nrnb);
1133
1134         nbv->setupGpuShortRangeWork(fr->gpuBonded, Nbnxm::InteractionLocality::Local);
1135
1136         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1137         wallcycle_stop(wcycle, ewcNS);
1138
1139         if (useGpuXBufOps == BufferOpsUseGpu::True)
1140         {
1141             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1142         }
1143         // For force buffer ops, we use the below conditon rather than
1144         // useGpuFBufOps to ensure that init is performed even if this
1145         // NS step is also a virial step (on which f buf ops are deactivated).
1146         if (simulationWork.useGpuBufferOps && simulationWork.useGpuNonbonded && (GMX_GPU == GMX_GPU_CUDA))
1147         {
1148             nbv->atomdata_init_add_nbat_f_to_f_gpu();
1149         }
1150     }
1151     else if (!EI_TPI(inputrec->eI))
1152     {
1153         if (useGpuXBufOps == BufferOpsUseGpu::True)
1154         {
1155             nbv->convertCoordinatesGpu(Nbnxm::AtomLocality::Local, false,
1156                                        stateGpu->getCoordinates(),
1157                                        localXReadyOnDevice);
1158         }
1159         else
1160         {
1161             nbv->convertCoordinates(Nbnxm::AtomLocality::Local, false,
1162                                     x.unpaddedArrayRef());
1163         }
1164     }
1165
1166     const gmx::DomainLifetimeWorkload &domainWork = runScheduleWork->domainWork;
1167
1168     if (simulationWork.useGpuNonbonded)
1169     {
1170         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1171
1172         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1173
1174         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1175         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1176         if (stepWork.doNeighborSearch || (useGpuXBufOps == BufferOpsUseGpu::False))
1177         {
1178             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1179                                       Nbnxm::AtomLocality::Local);
1180         }
1181         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1182         // with X buffer ops offloaded to the GPU on all but the search steps
1183
1184         // bonded work not split into separate local and non-local, so with DD
1185         // we can only launch the kernel after non-local coordinates have been received.
1186         if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1187         {
1188             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1189             fr->gpuBonded->launchKernel(fr, stepWork, box);
1190             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1191         }
1192
1193         /* launch local nonbonded work on GPU */
1194         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1195         do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::Local, enbvClearFNo,
1196                      step, nrnb, wcycle);
1197         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1198         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1199     }
1200
1201     if (useGpuPmeOnThisRank)
1202     {
1203         // In PME GPU and mixed mode we launch FFT / gather after the
1204         // X copy/transform to allow overlap as well as after the GPU NB
1205         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1206         // the nonbonded kernel.
1207         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1208     }
1209
1210     // TODO Update this comment when introducing SimulationWorkload
1211     //
1212     // The conditions for gpuHaloExchange e.g. using GPU buffer
1213     // operations were checked before construction, so here we can
1214     // just use it and assert upon any conditions.
1215     gmx::GpuHaloExchange *gpuHaloExchange              = (havePPDomainDecomposition(cr) ? cr->dd->gpuHaloExchange.get() : nullptr);
1216     const bool            ddUsesGpuDirectCommunication = (gpuHaloExchange != nullptr);
1217     GMX_ASSERT(!ddUsesGpuDirectCommunication || (useGpuXBufOps == BufferOpsUseGpu::True),
1218                "Must use coordinate buffer ops with GPU halo exchange");
1219
1220     /* Communicate coordinates and sum dipole if necessary +
1221        do non-local pair search */
1222     if (havePPDomainDecomposition(cr))
1223     {
1224         if (stepWork.doNeighborSearch)
1225         {
1226             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1227             wallcycle_start_nocount(wcycle, ewcNS);
1228             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1229             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1230             nbv->constructPairlist(Nbnxm::InteractionLocality::NonLocal,
1231                                    &top->excls, step, nrnb);
1232
1233             nbv->setupGpuShortRangeWork(fr->gpuBonded, Nbnxm::InteractionLocality::NonLocal);
1234             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1235             wallcycle_stop(wcycle, ewcNS);
1236             if (ddUsesGpuDirectCommunication)
1237             {
1238                 gpuHaloExchange->reinitHalo(stateGpu->getCoordinates(), stateGpu->getForces());
1239             }
1240         }
1241         else
1242         {
1243             if (ddUsesGpuDirectCommunication)
1244             {
1245                 // The following must be called after local setCoordinates (which records an event
1246                 // when the coordinate data has been copied to the device).
1247                 gpuHaloExchange->communicateHaloCoordinates(box);
1248
1249                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1250                 {
1251                     //non-local part of coordinate buffer must be copied back to host for CPU work
1252                     stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1253                 }
1254             }
1255             else
1256             {
1257                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1258             }
1259
1260             if (useGpuXBufOps == BufferOpsUseGpu::True)
1261             {
1262                 // The condition here was (pme != nullptr && pme_gpu_get_device_x(fr->pmedata) != nullptr)
1263                 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1264                 {
1265                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1266                 }
1267                 nbv->convertCoordinatesGpu(Nbnxm::AtomLocality::NonLocal, false,
1268                                            stateGpu->getCoordinates(),
1269                                            stateGpu->getCoordinatesReadyOnDeviceEvent(gmx::StatePropagatorDataGpu::AtomLocality::NonLocal,
1270                                                                                       simulationWork, stepWork));
1271             }
1272             else
1273             {
1274                 nbv->convertCoordinates(Nbnxm::AtomLocality::NonLocal, false,
1275                                         x.unpaddedArrayRef());
1276             }
1277
1278         }
1279
1280         if (simulationWork.useGpuNonbonded)
1281         {
1282             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1283
1284             if (stepWork.doNeighborSearch || (useGpuXBufOps == BufferOpsUseGpu::False))
1285             {
1286                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1287                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1288                                           Nbnxm::AtomLocality::NonLocal);
1289                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1290             }
1291
1292             if (domainWork.haveGpuBondedWork)
1293             {
1294                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1295                 fr->gpuBonded->launchKernel(fr, stepWork, box);
1296                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1297             }
1298
1299             /* launch non-local nonbonded tasks on GPU */
1300             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1301             do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1302                          step, nrnb, wcycle);
1303             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1304
1305             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1306         }
1307     }
1308
1309     if (simulationWork.useGpuNonbonded)
1310     {
1311         /* launch D2H copy-back F */
1312         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1313         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1314
1315         bool copyBackNbForce  = (useGpuFBufOps == BufferOpsUseGpu::False);
1316
1317         if (havePPDomainDecomposition(cr))
1318         {
1319             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1320                                       stepWork, Nbnxm::AtomLocality::NonLocal, copyBackNbForce);
1321         }
1322         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1323                                   stepWork, Nbnxm::AtomLocality::Local, copyBackNbForce);
1324         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1325
1326         if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1327         {
1328             fr->gpuBonded->launchEnergyTransfer();
1329         }
1330         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1331     }
1332
1333     if (stepWork.stateChanged && inputrecNeedMutot(inputrec))
1334     {
1335         if (PAR(cr))
1336         {
1337             gmx_sumd(2*DIM, mu, cr);
1338
1339             ddBalanceRegionHandler.reopenRegionCpu();
1340         }
1341
1342         for (i = 0; i < 2; i++)
1343         {
1344             for (j = 0; j < DIM; j++)
1345             {
1346                 fr->mu_tot[i][j] = mu[i*DIM + j];
1347             }
1348         }
1349     }
1350     if (mdatoms->nChargePerturbed == 0)
1351     {
1352         copy_rvec(fr->mu_tot[0], mu_tot);
1353     }
1354     else
1355     {
1356         for (j = 0; j < DIM; j++)
1357         {
1358             mu_tot[j] =
1359                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1360                 lambda[efptCOUL]*fr->mu_tot[1][j];
1361         }
1362     }
1363
1364     /* Reset energies */
1365     reset_enerdata(enerd);
1366     /* Clear the shift forces */
1367     // TODO: This should be linked to the shift force buffer in use, or cleared before use instead
1368     for (gmx::RVec &elem : fr->shiftForces)
1369     {
1370         elem = { 0.0_real, 0.0_real, 0.0_real };
1371     }
1372
1373     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1374     {
1375         wallcycle_start(wcycle, ewcPPDURINGPME);
1376         dd_force_flop_start(cr->dd, nrnb);
1377     }
1378
1379     if (inputrec->bRot)
1380     {
1381         wallcycle_start(wcycle, ewcROT);
1382         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, stepWork.doNeighborSearch);
1383         wallcycle_stop(wcycle, ewcROT);
1384     }
1385
1386     /* Start the force cycle counter.
1387      * Note that a different counter is used for dynamic load balancing.
1388      */
1389     wallcycle_start(wcycle, ewcFORCE);
1390
1391     // Set up and clear force outputs.
1392     // We use std::move to keep the compiler happy, it has no effect.
1393     ForceOutputs forceOut = setupForceOutputs(fr, pull_work, *inputrec, std::move(force), stepWork, wcycle);
1394
1395     /* We calculate the non-bonded forces, when done on the CPU, here.
1396      * We do this before calling do_force_lowlevel, because in that
1397      * function, the listed forces are calculated before PME, which
1398      * does communication.  With this order, non-bonded and listed
1399      * force calculation imbalance can be balanced out by the domain
1400      * decomposition load balancing.
1401      */
1402
1403     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1404
1405     if (!useOrEmulateGpuNb)
1406     {
1407         do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::Local, enbvClearFYes,
1408                      step, nrnb, wcycle);
1409     }
1410
1411     if (fr->efep != efepNO)
1412     {
1413         /* Calculate the local and non-local free energy interactions here.
1414          * Happens here on the CPU both with and without GPU.
1415          */
1416         nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::Local,
1417                                       fr, as_rvec_array(x.unpaddedArrayRef().data()), &forceOut.forceWithShiftForces(), *mdatoms,
1418                                       inputrec->fepvals, lambda.data(),
1419                                       enerd, stepWork, nrnb);
1420
1421         if (havePPDomainDecomposition(cr))
1422         {
1423             nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::NonLocal,
1424                                           fr, as_rvec_array(x.unpaddedArrayRef().data()), &forceOut.forceWithShiftForces(), *mdatoms,
1425                                           inputrec->fepvals, lambda.data(),
1426                                           enerd, stepWork, nrnb);
1427         }
1428     }
1429
1430     if (!useOrEmulateGpuNb)
1431     {
1432         if (havePPDomainDecomposition(cr))
1433         {
1434             do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1435                          step, nrnb, wcycle);
1436         }
1437
1438         if (stepWork.computeForces)
1439         {
1440             /* Add all the non-bonded force to the normal force array.
1441              * This can be split into a local and a non-local part when overlapping
1442              * communication with calculation with domain decomposition.
1443              */
1444             wallcycle_stop(wcycle, ewcFORCE);
1445             nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::All, forceOut.forceWithShiftForces().force());
1446             wallcycle_start_nocount(wcycle, ewcFORCE);
1447         }
1448
1449         /* If there are multiple fshift output buffers we need to reduce them */
1450         if (stepWork.computeVirial)
1451         {
1452             /* This is not in a subcounter because it takes a
1453                negligible and constant-sized amount of time */
1454             nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1455                                                      forceOut.forceWithShiftForces().shiftForces());
1456         }
1457     }
1458
1459     /* update QMMMrec, if necessary */
1460     if (fr->bQMMM)
1461     {
1462         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1463     }
1464
1465     // TODO Force flags should include haveFreeEnergyWork for this domain
1466     if (ddUsesGpuDirectCommunication &&
1467         (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1468     {
1469         /* Wait for non-local coordinate data to be copied from device */
1470         nbv->wait_nonlocal_x_copy_D2H_done();
1471     }
1472     /* Compute the bonded and non-bonded energies and optionally forces */
1473     do_force_lowlevel(fr, inputrec, &(top->idef),
1474                       cr, ms, nrnb, wcycle, mdatoms,
1475                       x, hist, &forceOut, enerd, fcd,
1476                       box, lambda.data(), graph, fr->mu_tot,
1477                       stepWork,
1478                       ddBalanceRegionHandler);
1479
1480     wallcycle_stop(wcycle, ewcFORCE);
1481
1482     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1483                          imdSession, pull_work, step, t, wcycle,
1484                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1485                          stepWork, &forceOut.forceWithVirial(), enerd,
1486                          ed, stepWork.doNeighborSearch);
1487
1488
1489     // Will store the amount of cycles spent waiting for the GPU that
1490     // will be later used in the DLB accounting.
1491     float cycles_wait_gpu = 0;
1492     if (useOrEmulateGpuNb)
1493     {
1494         auto &forceWithShiftForces = forceOut.forceWithShiftForces();
1495
1496         /* wait for non-local forces (or calculate in emulation mode) */
1497         if (havePPDomainDecomposition(cr))
1498         {
1499             if (simulationWork.useGpuNonbonded)
1500             {
1501                 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1502                                                                stepWork, Nbnxm::AtomLocality::NonLocal,
1503                                                                enerd->grpp.ener[egLJSR].data(),
1504                                                                enerd->grpp.ener[egCOULSR].data(),
1505                                                                forceWithShiftForces.shiftForces(),
1506                                                                wcycle);
1507             }
1508             else
1509             {
1510                 wallcycle_start_nocount(wcycle, ewcFORCE);
1511                 do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::NonLocal, enbvClearFYes,
1512                              step, nrnb, wcycle);
1513                 wallcycle_stop(wcycle, ewcFORCE);
1514             }
1515
1516             if (useGpuFBufOps == BufferOpsUseGpu::True)
1517             {
1518                 gmx::FixedCapacityVector<GpuEventSynchronizer*, 1> dependencyList;
1519
1520                 // TODO: move this into DomainLifetimeWorkload, including the second part of the condition
1521                 // The bonded and free energy CPU tasks can have non-local force contributions
1522                 // which are a dependency for the GPU force reduction.
1523                 bool  haveNonLocalForceContribInCpuBuffer = domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1524
1525                 if (haveNonLocalForceContribInCpuBuffer)
1526                 {
1527                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1528                     dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(gmx::StatePropagatorDataGpu::AtomLocality::NonLocal,
1529                                                                                    useGpuFBufOps == BufferOpsUseGpu::True));
1530                 }
1531
1532                 nbv->atomdata_add_nbat_f_to_f_gpu(Nbnxm::AtomLocality::NonLocal,
1533                                                   stateGpu->getForces(),
1534                                                   pme_gpu_get_device_f(fr->pmedata),
1535                                                   dependencyList,
1536                                                   false, haveNonLocalForceContribInCpuBuffer);
1537                 stateGpu->copyForcesFromGpu(forceOut.forceWithShiftForces().force(), gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1538             }
1539             else
1540             {
1541                 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::NonLocal,
1542                                               forceWithShiftForces.force());
1543             }
1544
1545
1546             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1547             {
1548                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1549                                                          forceWithShiftForces.shiftForces());
1550             }
1551         }
1552     }
1553
1554     const bool useGpuForcesHaloExchange = ddUsesGpuDirectCommunication && (useGpuFBufOps == BufferOpsUseGpu::True);
1555     const bool useCpuPmeFReduction      = thisRankHasDuty(cr, DUTY_PME) && !stepWork.useGpuPmeFReduction;
1556     // TODO: move this into DomainLifetimeWorkload, including the second part of the condition
1557     const bool haveCpuLocalForces     = (domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork || useCpuPmeFReduction ||
1558                                          (fr->efep != efepNO));
1559
1560     if (havePPDomainDecomposition(cr))
1561     {
1562         /* We are done with the CPU compute.
1563          * We will now communicate the non-local forces.
1564          * If we use a GPU this will overlap with GPU work, so in that case
1565          * we do not close the DD force balancing region here.
1566          */
1567         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1568
1569         if (stepWork.computeForces)
1570         {
1571
1572             if (useGpuForcesHaloExchange)
1573             {
1574                 if (haveCpuLocalForces)
1575                 {
1576                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), gmx::StatePropagatorDataGpu::AtomLocality::Local);
1577                 }
1578                 gpuHaloExchange->communicateHaloForces(haveCpuLocalForces);
1579             }
1580             else
1581             {
1582                 if (useGpuFBufOps == BufferOpsUseGpu::True)
1583                 {
1584                     stateGpu->waitForcesReadyOnHost(gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1585                 }
1586                 dd_move_f(cr->dd, &forceOut.forceWithShiftForces(), wcycle);
1587             }
1588
1589         }
1590     }
1591
1592     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1593     // an alternating wait/reduction scheme.
1594     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded && !DOMAINDECOMP(cr) &&
1595                              (useGpuFBufOps == BufferOpsUseGpu::False));
1596     if (alternateGpuWait)
1597     {
1598         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &forceOut, enerd,
1599                                     stepWork, pmeFlags, wcycle);
1600     }
1601
1602     if (!alternateGpuWait && useGpuPmeOnThisRank)
1603     {
1604         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial(), enerd);
1605     }
1606
1607     /* Wait for local GPU NB outputs on the non-alternating wait path */
1608     if (!alternateGpuWait && simulationWork.useGpuNonbonded)
1609     {
1610         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1611          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1612          * but even with a step of 0.1 ms the difference is less than 1%
1613          * of the step time.
1614          */
1615         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1616         const float waitCycles               =
1617             Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1618                                         stepWork, Nbnxm::AtomLocality::Local,
1619                                         enerd->grpp.ener[egLJSR].data(),
1620                                         enerd->grpp.ener[egCOULSR].data(),
1621                                         forceOut.forceWithShiftForces().shiftForces(),
1622                                         wcycle);
1623
1624         if (ddBalanceRegionHandler.useBalancingRegion())
1625         {
1626             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1627             if (stepWork.computeForces &&  waitCycles <= gpuWaitApiOverheadMargin)
1628             {
1629                 /* We measured few cycles, it could be that the kernel
1630                  * and transfer finished earlier and there was no actual
1631                  * wait time, only API call overhead.
1632                  * Then the actual time could be anywhere between 0 and
1633                  * cycles_wait_est. We will use half of cycles_wait_est.
1634                  */
1635                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1636             }
1637             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1638         }
1639     }
1640
1641     if (fr->nbv->emulateGpu())
1642     {
1643         // NOTE: emulation kernel is not included in the balancing region,
1644         // but emulation mode does not target performance anyway
1645         wallcycle_start_nocount(wcycle, ewcFORCE);
1646         do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::Local,
1647                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1648                      step, nrnb, wcycle);
1649         wallcycle_stop(wcycle, ewcFORCE);
1650     }
1651
1652     /* Do the nonbonded GPU (or emulation) force buffer reduction
1653      * on the non-alternating path. */
1654     if (useOrEmulateGpuNb && !alternateGpuWait)
1655     {
1656         gmx::FixedCapacityVector<GpuEventSynchronizer*, 2> dependencyList;
1657
1658         if (stepWork.useGpuPmeFReduction)
1659         {
1660             dependencyList.push_back(pme_gpu_get_f_ready_synchronizer(fr->pmedata));
1661         }
1662
1663         gmx::ArrayRef<gmx::RVec>  forceWithShift = forceOut.forceWithShiftForces().force();
1664
1665         if (useGpuFBufOps == BufferOpsUseGpu::True)
1666         {
1667             // Flag to specify whether the CPU force buffer has contributions to
1668             // local atoms. This depends on whether there are CPU-based force tasks
1669             // or when DD is active the halo exchange has resulted in contributions
1670             // from the non-local part.
1671             const bool haveLocalForceContribInCpuBuffer = (haveCpuLocalForces || havePPDomainDecomposition(cr));
1672
1673             // TODO: move these steps as early as possible:
1674             // - CPU f H2D should be as soon as all CPU-side forces are done
1675             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1676             //   before the next CPU task that consumes the forces: vsite spread or update)
1677             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
1678             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
1679             //   These should be unified.
1680             if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
1681             {
1682                 stateGpu->copyForcesToGpu(forceWithShift, gmx::StatePropagatorDataGpu::AtomLocality::Local);
1683                 dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(gmx::StatePropagatorDataGpu::AtomLocality::Local,
1684                                                                                useGpuFBufOps == BufferOpsUseGpu::True));
1685             }
1686             if (useGpuForcesHaloExchange)
1687             {
1688                 // Add a stream synchronization to satisfy a dependency
1689                 // for the local buffer ops on the result of GPU halo
1690                 // exchange, which operates in the non-local stream and
1691                 // writes to to local parf og the force buffer.
1692                 //
1693                 // TODO improve this through use of an event - see Redmine #3093
1694                 //      push the event into the dependencyList
1695                 nbv->stream_local_wait_for_nonlocal();
1696             }
1697             nbv->atomdata_add_nbat_f_to_f_gpu(Nbnxm::AtomLocality::Local,
1698                                               stateGpu->getForces(),
1699                                               pme_gpu_get_device_f(fr->pmedata),
1700                                               dependencyList,
1701                                               stepWork.useGpuPmeFReduction, haveLocalForceContribInCpuBuffer);
1702             stateGpu->copyForcesFromGpu(forceWithShift, gmx::StatePropagatorDataGpu::AtomLocality::Local);
1703             stateGpu->waitForcesReadyOnHost(gmx::StatePropagatorDataGpu::AtomLocality::Local);
1704         }
1705         else
1706         {
1707             nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local, forceWithShift);
1708         }
1709
1710     }
1711
1712     launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd,
1713                             *runScheduleWork,
1714                             simulationWork.useGpuNonbonded, useGpuPmeOnThisRank,
1715                             step,
1716                             wcycle);
1717
1718     if (DOMAINDECOMP(cr))
1719     {
1720         dd_force_flop_stop(cr->dd, nrnb);
1721     }
1722
1723     if (stepWork.computeForces)
1724     {
1725         rvec *f = as_rvec_array(forceOut.forceWithShiftForces().force().data());
1726
1727         /* If we have NoVirSum forces, but we do not calculate the virial,
1728          * we sum fr->f_novirsum=forceOut.f later.
1729          */
1730         if (vsite && !(fr->haveDirectVirialContributions && !stepWork.computeVirial))
1731         {
1732             rvec *fshift = as_rvec_array(forceOut.forceWithShiftForces().shiftForces().data());
1733             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), f, fshift, FALSE, nullptr, nrnb,
1734                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1735         }
1736
1737         if (stepWork.computeVirial)
1738         {
1739             /* Calculation of the virial must be done after vsites! */
1740             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()),
1741                         forceOut.forceWithShiftForces(),
1742                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1743         }
1744     }
1745
1746     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1747     {
1748         /* In case of node-splitting, the PP nodes receive the long-range
1749          * forces, virial and energy from the PME nodes here.
1750          */
1751         pme_receive_force_ener(cr, &forceOut.forceWithVirial(), enerd, wcycle);
1752     }
1753
1754     if (stepWork.computeForces)
1755     {
1756         post_process_forces(cr, step, nrnb, wcycle,
1757                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), &forceOut,
1758                             vir_force, mdatoms, graph, fr, vsite,
1759                             stepWork);
1760     }
1761
1762     if (stepWork.computeEnergy)
1763     {
1764         /* Sum the potential energy terms from group contributions */
1765         sum_epot(&(enerd->grpp), enerd->term);
1766
1767         if (!EI_TPI(inputrec->eI))
1768         {
1769             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1770         }
1771     }
1772
1773     /* In case we don't have constraints and are using GPUs, the next balancing
1774      * region starts here.
1775      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1776      * virial calculation and COM pulling, is not thus not included in
1777      * the balance timing, which is ok as most tasks do communication.
1778      */
1779     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1780 }