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/ewald/pme_pp_comm_gpu.h"
  57 #include "gromacs/gmxlib/network.h"
  58 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  59 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  60 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  61 #include "gromacs/gpu_utils/gpu_utils.h"
  62 #include "gromacs/imd/imd.h"
  63 #include "gromacs/listed_forces/disre.h"
  64 #include "gromacs/listed_forces/gpubonded.h"
  65 #include "gromacs/listed_forces/listed_forces.h"
  66 #include "gromacs/listed_forces/manage_threading.h"
  67 #include "gromacs/listed_forces/orires.h"
  68 #include "gromacs/math/arrayrefwithpadding.h"
  69 #include "gromacs/math/functions.h"
  70 #include "gromacs/math/units.h"
  71 #include "gromacs/math/vec.h"
  72 #include "gromacs/math/vecdump.h"
  73 #include "gromacs/mdlib/calcmu.h"
  74 #include "gromacs/mdlib/calcvir.h"
  75 #include "gromacs/mdlib/constr.h"
  76 #include "gromacs/mdlib/enerdata_utils.h"
  77 #include "gromacs/mdlib/force.h"
  78 #include "gromacs/mdlib/forcerec.h"
  79 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  80 #include "gromacs/mdlib/qmmm.h"
  81 #include "gromacs/mdlib/update.h"
  82 #include "gromacs/mdtypes/commrec.h"
  83 #include "gromacs/mdtypes/enerdata.h"
  84 #include "gromacs/mdtypes/forceoutput.h"
  85 #include "gromacs/mdtypes/iforceprovider.h"
  86 #include "gromacs/mdtypes/inputrec.h"
  87 #include "gromacs/mdtypes/md_enums.h"
  88 #include "gromacs/mdtypes/simulation_workload.h"
  89 #include "gromacs/mdtypes/state.h"
  90 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  91 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  92 #include "gromacs/nbnxm/nbnxm.h"
  93 #include "gromacs/pbcutil/ishift.h"
  94 #include "gromacs/pbcutil/mshift.h"
  95 #include "gromacs/pbcutil/pbc.h"
  96 #include "gromacs/pulling/pull.h"
  97 #include "gromacs/pulling/pull_rotation.h"
  98 #include "gromacs/timing/cyclecounter.h"
  99 #include "gromacs/timing/gpu_timing.h"
 100 #include "gromacs/timing/wallcycle.h"
 101 #include "gromacs/timing/wallcyclereporting.h"
 102 #include "gromacs/timing/walltime_accounting.h"
 103 #include "gromacs/topology/topology.h"
 104 #include "gromacs/utility/arrayref.h"
 105 #include "gromacs/utility/basedefinitions.h"
 106 #include "gromacs/utility/cstringutil.h"
 107 #include "gromacs/utility/exceptions.h"
 108 #include "gromacs/utility/fatalerror.h"
 109 #include "gromacs/utility/fixedcapacityvector.h"
 110 #include "gromacs/utility/gmxassert.h"
 111 #include "gromacs/utility/gmxmpi.h"
 112 #include "gromacs/utility/logger.h"
 113 #include "gromacs/utility/smalloc.h"
 114 #include "gromacs/utility/strconvert.h"
 115 #include "gromacs/utility/sysinfo.h"
 116
 117 using gmx::ForceOutputs;
 118 using gmx::StepWorkload;
 119 using gmx::DomainLifetimeWorkload;
 120 using gmx::SimulationWorkload;
 121 using gmx::AtomLocality;
 122 using gmx::InteractionLocality;
 123
 124 // TODO: this environment variable allows us to verify before release
 125 // that on less common architectures the total cost of polling is not larger than
 126 // a blocking wait (so polling does not introduce overhead when the static
 127 // PME-first ordering would suffice).
 128 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 129
 130 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 131 {
 132     const int      end = forceToAdd.size();
 133
 134     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 135 #pragma omp parallel for num_threads(nt) schedule(static)
 136     for (int i = 0; i < end; i++)
 137     {
 138         rvec_inc(f[i], forceToAdd[i]);
 139     }
 140 }
 141
 142 static void calc_virial(int start, int homenr, const rvec x[],
 143                         const gmx::ForceWithShiftForces &forceWithShiftForces,
 144                         tensor vir_part, const t_graph *graph, const matrix box,
 145                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 146 {
 147     /* The short-range virial from surrounding boxes */
 148     const rvec *fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
 149     calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, ePBC == epbcSCREW, box);
 150     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 151
 152     /* Calculate partial virial, for local atoms only, based on short range.
 153      * Total virial is computed in global_stat, called from do_md
 154      */
 155     const rvec *f = as_rvec_array(forceWithShiftForces.force().data());
 156     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 157     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 158
 159     if (debug)
 160     {
 161         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 162     }
 163 }
 164
 165 static void pull_potential_wrapper(const t_commrec *cr,
 166                                    const t_inputrec *ir,
 167                                    const matrix box, gmx::ArrayRef<const gmx::RVec> x,
 168                                    gmx::ForceWithVirial *force,
 169                                    const t_mdatoms *mdatoms,
 170                                    gmx_enerdata_t *enerd,
 171                                    pull_t *pull_work,
 172                                    const real *lambda,
 173                                    double t,
 174                                    gmx_wallcycle_t wcycle)
 175 {
 176     t_pbc  pbc;
 177     real   dvdl;
 178
 179     /* Calculate the center of mass forces, this requires communication,
 180      * which is why pull_potential is called close to other communication.
 181      */
 182     wallcycle_start(wcycle, ewcPULLPOT);
 183     set_pbc(&pbc, ir->ePBC, box);
 184     dvdl                     = 0;
 185     enerd->term[F_COM_PULL] +=
 186         pull_potential(pull_work, mdatoms, &pbc,
 187                        cr, t, lambda[efptRESTRAINT], as_rvec_array(x.data()), force, &dvdl);
 188     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 189     wallcycle_stop(wcycle, ewcPULLPOT);
 190 }
 191
 192 static void pme_receive_force_ener(t_forcerec           *fr,
 193                                    const t_commrec      *cr,
 194                                    gmx::ForceWithVirial *forceWithVirial,
 195                                    gmx_enerdata_t       *enerd,
 196                                    bool                  useGpuPmePpComms,
 197                                    bool                  receivePmeForceToGpu,
 198                                    gmx_wallcycle_t       wcycle)
 199 {
 200     real   e_q, e_lj, dvdl_q, dvdl_lj;
 201     float  cycles_ppdpme, cycles_seppme;
 202
 203     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 204     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 205
 206     /* In case of node-splitting, the PP nodes receive the long-range
 207      * forces, virial and energy from the PME nodes here.
 208      */
 209     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 210     dvdl_q  = 0;
 211     dvdl_lj = 0;
 212     gmx_pme_receive_f(fr->pmePpCommGpu.get(),
 213                       cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 214                       useGpuPmePpComms, receivePmeForceToGpu, &cycles_seppme);
 215     enerd->term[F_COUL_RECIP] += e_q;
 216     enerd->term[F_LJ_RECIP]   += e_lj;
 217     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 218     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 219
 220     if (wcycle)
 221     {
 222         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 223     }
 224     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 225 }
 226
 227 static void print_large_forces(FILE            *fp,
 228                                const t_mdatoms *md,
 229                                const t_commrec *cr,
 230                                int64_t          step,
 231                                real             forceTolerance,
 232                                const rvec      *x,
 233                                const rvec      *f)
 234 {
 235     real           force2Tolerance = gmx::square(forceTolerance);
 236     gmx::index     numNonFinite    = 0;
 237     for (int i = 0; i < md->homenr; i++)
 238     {
 239         real force2    = norm2(f[i]);
 240         bool nonFinite = !std::isfinite(force2);
 241         if (force2 >= force2Tolerance || nonFinite)
 242         {
 243             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 244                     step,
 245                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 246         }
 247         if (nonFinite)
 248         {
 249             numNonFinite++;
 250         }
 251     }
 252     if (numNonFinite > 0)
 253     {
 254         /* Note that with MPI this fatal call on one rank might interrupt
 255          * the printing on other ranks. But we can only avoid that with
 256          * an expensive MPI barrier that we would need at each step.
 257          */
 258         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 259     }
 260 }
 261
 262 static void post_process_forces(const t_commrec       *cr,
 263                                 int64_t                step,
 264                                 t_nrnb                *nrnb,
 265                                 gmx_wallcycle_t        wcycle,
 266                                 const gmx_localtop_t  *top,
 267                                 const matrix           box,
 268                                 const rvec             x[],
 269                                 ForceOutputs          *forceOutputs,
 270                                 tensor                 vir_force,
 271                                 const t_mdatoms       *mdatoms,
 272                                 const t_graph         *graph,
 273                                 const t_forcerec      *fr,
 274                                 const gmx_vsite_t     *vsite,
 275                                 const StepWorkload    &stepWork)
 276 {
 277     rvec *f = as_rvec_array(forceOutputs->forceWithShiftForces().force().data());
 278
 279     if (fr->haveDirectVirialContributions)
 280     {
 281         auto &forceWithVirial = forceOutputs->forceWithVirial();
 282         rvec *fDirectVir      = as_rvec_array(forceWithVirial.force_.data());
 283
 284         if (vsite)
 285         {
 286             /* Spread the mesh force on virtual sites to the other particles...
 287              * This is parallellized. MPI communication is performed
 288              * if the constructing atoms aren't local.
 289              */
 290             matrix virial = { { 0 } };
 291             spread_vsite_f(vsite, x, fDirectVir, nullptr,
 292                            stepWork.computeVirial, virial,
 293                            nrnb,
 294                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
 295             forceWithVirial.addVirialContribution(virial);
 296         }
 297
 298         if (stepWork.computeVirial)
 299         {
 300             /* Now add the forces, this is local */
 301             sum_forces(f, forceWithVirial.force_);
 302
 303             /* Add the direct virial contributions */
 304             GMX_ASSERT(forceWithVirial.computeVirial_, "forceWithVirial should request virial computation when we request the virial");
 305             m_add(vir_force, forceWithVirial.getVirial(), vir_force);
 306
 307             if (debug)
 308             {
 309                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 310             }
 311         }
 312     }
 313
 314     if (fr->print_force >= 0)
 315     {
 316         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 317     }
 318 }
 319
 320 static void do_nb_verlet(t_forcerec                       *fr,
 321                          const interaction_const_t        *ic,
 322                          gmx_enerdata_t                   *enerd,
 323                          const StepWorkload               &stepWork,
 324                          const InteractionLocality         ilocality,
 325                          const int                         clearF,
 326                          const int64_t                     step,
 327                          t_nrnb                           *nrnb,
 328                          gmx_wallcycle_t                   wcycle)
 329 {
 330     if (!stepWork.computeNonbondedForces)
 331     {
 332         /* skip non-bonded calculation */
 333         return;
 334     }
 335
 336     nonbonded_verlet_t *nbv  = fr->nbv.get();
 337
 338     /* GPU kernel launch overhead is already timed separately */
 339     if (fr->cutoff_scheme != ecutsVERLET)
 340     {
 341         gmx_incons("Invalid cut-off scheme passed!");
 342     }
 343
 344     if (!nbv->useGpu())
 345     {
 346         /* When dynamic pair-list  pruning is requested, we need to prune
 347          * at nstlistPrune steps.
 348          */
 349         if (nbv->isDynamicPruningStepCpu(step))
 350         {
 351             /* Prune the pair-list beyond fr->ic->rlistPrune using
 352              * the current coordinates of the atoms.
 353              */
 354             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 355             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 356             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 357         }
 358     }
 359
 360     nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
 361 }
 362
 363 static inline void clear_rvecs_omp(int n, rvec v[])
 364 {
 365     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 366
 367     /* Note that we would like to avoid this conditional by putting it
 368      * into the omp pragma instead, but then we still take the full
 369      * omp parallel for overhead (at least with gcc5).
 370      */
 371     if (nth == 1)
 372     {
 373         for (int i = 0; i < n; i++)
 374         {
 375             clear_rvec(v[i]);
 376         }
 377     }
 378     else
 379     {
 380 #pragma omp parallel for num_threads(nth) schedule(static)
 381         for (int i = 0; i < n; i++)
 382         {
 383             clear_rvec(v[i]);
 384         }
 385     }
 386 }
 387
 388 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 389  *
 390  * \param groupOptions  Group options, containing T-coupling options
 391  */
 392 static real averageKineticEnergyEstimate(const t_grpopts &groupOptions)
 393 {
 394     real nrdfCoupled   = 0;
 395     real nrdfUncoupled = 0;
 396     real kineticEnergy = 0;
 397     for (int g = 0; g < groupOptions.ngtc; g++)
 398     {
 399         if (groupOptions.tau_t[g] >= 0)
 400         {
 401             nrdfCoupled   += groupOptions.nrdf[g];
 402             kineticEnergy += groupOptions.nrdf[g]*0.5*groupOptions.ref_t[g]*BOLTZ;
 403         }
 404         else
 405         {
 406             nrdfUncoupled += groupOptions.nrdf[g];
 407         }
 408     }
 409
 410     /* This conditional with > also catches nrdf=0 */
 411     if (nrdfCoupled > nrdfUncoupled)
 412     {
 413         return kineticEnergy*(nrdfCoupled + nrdfUncoupled)/nrdfCoupled;
 414     }
 415     else
 416     {
 417         return 0;
 418     }
 419 }
 420
 421 /*! \brief This routine checks that the potential energy is finite.
 422  *
 423  * Always checks that the potential energy is finite. If step equals
 424  * inputrec.init_step also checks that the magnitude of the potential energy
 425  * is reasonable. Terminates with a fatal error when a check fails.
 426  * Note that passing this check does not guarantee finite forces,
 427  * since those use slightly different arithmetics. But in most cases
 428  * there is just a narrow coordinate range where forces are not finite
 429  * and energies are finite.
 430  *
 431  * \param[in] step      The step number, used for checking and printing
 432  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to enerd.term[F_EPOT] before calling this routine
 433  * \param[in] inputrec  The input record
 434  */
 435 static void checkPotentialEnergyValidity(int64_t               step,
 436                                          const gmx_enerdata_t &enerd,
 437                                          const t_inputrec     &inputrec)
 438 {
 439     /* Threshold valid for comparing absolute potential energy against
 440      * the kinetic energy. Normally one should not consider absolute
 441      * potential energy values, but with a factor of one million
 442      * we should never get false positives.
 443      */
 444     constexpr real c_thresholdFactor = 1e6;
 445
 446     bool           energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 447     real           averageKineticEnergy = 0;
 448     /* We only check for large potential energy at the initial step,
 449      * because that is by far the most likely step for this too occur
 450      * and because computing the average kinetic energy is not free.
 451      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 452      * before they become NaN.
 453      */
 454     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 455     {
 456         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 457     }
 458
 459     if (energyIsNotFinite || (averageKineticEnergy > 0 &&
 460                               enerd.term[F_EPOT] > c_thresholdFactor*averageKineticEnergy))
 461     {
 462         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.",
 463                   step,
 464                   enerd.term[F_EPOT],
 465                   energyIsNotFinite ? "not finite" : "extremely high",
 466                   enerd.term[F_LJ],
 467                   enerd.term[F_COUL_SR],
 468                   energyIsNotFinite ? "non-finite" : "very high",
 469                   energyIsNotFinite ? " or Nan" : "");
 470     }
 471 }
 472
 473 /*! \brief Return true if there are special forces computed this step.
 474  *
 475  * The conditionals exactly correspond to those in computeSpecialForces().
 476  */
 477 static bool
 478 haveSpecialForces(const t_inputrec              &inputrec,
 479                   const gmx::ForceProviders     &forceProviders,
 480                   const pull_t                  *pull_work,
 481                   const bool                     computeForces,
 482                   const gmx_edsam               *ed)
 483 {
 484
 485     return
 486         ((computeForces && forceProviders.hasForceProvider()) ||         // forceProviders
 487          (inputrec.bPull && pull_have_potential(pull_work)) ||           // pull
 488          inputrec.bRot ||                                                // enforced rotation
 489          (ed != nullptr) ||                                              // flooding
 490          (inputrec.bIMD && computeForces));                              // IMD
 491 }
 492
 493 /*! \brief Compute forces and/or energies for special algorithms
 494  *
 495  * The intention is to collect all calls to algorithms that compute
 496  * forces on local atoms only and that do not contribute to the local
 497  * virial sum (but add their virial contribution separately).
 498  * Eventually these should likely all become ForceProviders.
 499  * Within this function the intention is to have algorithms that do
 500  * global communication at the end, so global barriers within the MD loop
 501  * are as close together as possible.
 502  *
 503  * \param[in]     fplog            The log file
 504  * \param[in]     cr               The communication record
 505  * \param[in]     inputrec         The input record
 506  * \param[in]     awh              The Awh module (nullptr if none in use).
 507  * \param[in]     enforcedRotation Enforced rotation module.
 508  * \param[in]     imdSession       The IMD session
 509  * \param[in]     pull_work        The pull work structure.
 510  * \param[in]     step             The current MD step
 511  * \param[in]     t                The current time
 512  * \param[in,out] wcycle           Wallcycle accounting struct
 513  * \param[in,out] forceProviders   Pointer to a list of force providers
 514  * \param[in]     box              The unit cell
 515  * \param[in]     x                The coordinates
 516  * \param[in]     mdatoms          Per atom properties
 517  * \param[in]     lambda           Array of free-energy lambda values
 518  * \param[in]     stepWork         Step schedule flags
 519  * \param[in,out] forceWithVirial  Force and virial buffers
 520  * \param[in,out] enerd            Energy buffer
 521  * \param[in,out] ed               Essential dynamics pointer
 522  * \param[in]     didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
 523  *
 524  * \todo Remove didNeighborSearch, which is used incorrectly.
 525  * \todo Convert all other algorithms called here to ForceProviders.
 526  */
 527 static void
 528 computeSpecialForces(FILE                          *fplog,
 529                      const t_commrec               *cr,
 530                      const t_inputrec              *inputrec,
 531                      gmx::Awh                      *awh,
 532                      gmx_enfrot                    *enforcedRotation,
 533                      gmx::ImdSession               *imdSession,
 534                      pull_t                        *pull_work,
 535                      int64_t                        step,
 536                      double                         t,
 537                      gmx_wallcycle_t                wcycle,
 538                      gmx::ForceProviders           *forceProviders,
 539                      const matrix                   box,
 540                      gmx::ArrayRef<const gmx::RVec> x,
 541                      const t_mdatoms               *mdatoms,
 542                      real                          *lambda,
 543                      const StepWorkload            &stepWork,
 544                      gmx::ForceWithVirial          *forceWithVirial,
 545                      gmx_enerdata_t                *enerd,
 546                      gmx_edsam                     *ed,
 547                      bool                           didNeighborSearch)
 548 {
 549     /* NOTE: Currently all ForceProviders only provide forces.
 550      *       When they also provide energies, remove this conditional.
 551      */
 552     if (stepWork.computeForces)
 553     {
 554         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 555         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 556
 557         /* Collect forces from modules */
 558         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 559     }
 560
 561     if (inputrec->bPull && pull_have_potential(pull_work))
 562     {
 563         pull_potential_wrapper(cr, inputrec, box, x,
 564                                forceWithVirial,
 565                                mdatoms, enerd, pull_work, lambda, t,
 566                                wcycle);
 567
 568         if (awh)
 569         {
 570             enerd->term[F_COM_PULL] +=
 571                 awh->applyBiasForcesAndUpdateBias(inputrec->ePBC, *mdatoms, box,
 572                                                   forceWithVirial,
 573                                                   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 |
 612            (stepWork.computeVirial ? GMX_PME_CALC_ENER_VIR : 0) |
 613            (stepWork.computeEnergy ? GMX_PME_CALC_ENER_VIR : 0) |
 614            (stepWork.computeForces ? GMX_PME_CALC_F        : 0);
 615 }
 616
 617 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 618  *
 619  * \param[in]  pmedata              The PME structure
 620  * \param[in]  box                  The box matrix
 621  * \param[in]  stepWork             Step schedule flags
 622  * \param[in]  pmeFlags             PME flags
 623  * \param[in]  xReadyOnDevice       Event synchronizer indicating that the coordinates are ready in the device memory.
 624  * \param[in]  wcycle               The wallcycle structure
 625  */
 626 static inline void launchPmeGpuSpread(gmx_pme_t            *pmedata,
 627                                       const matrix          box,
 628                                       const StepWorkload   &stepWork,
 629                                       int                   pmeFlags,
 630                                       GpuEventSynchronizer *xReadyOnDevice,
 631                                       gmx_wallcycle_t       wcycle)
 632 {
 633     pme_gpu_prepare_computation(pmedata, stepWork.haveDynamicBox, box, wcycle, pmeFlags, 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,
 645                                      gmx_wallcycle_t   wcycle)
 646 {
 647     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 648     pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
 649 }
 650
 651 /*! \brief
 652  *  Polling wait for either of the PME or nonbonded GPU tasks.
 653  *
 654  * Instead of a static order in waiting for GPU tasks, this function
 655  * polls checking which of the two tasks completes first, and does the
 656  * associated force buffer reduction overlapped with the other task.
 657  * By doing that, unlike static scheduling order, it can always overlap
 658  * one of the reductions, regardless of the GPU task completion order.
 659  *
 660  * \param[in]     nbv              Nonbonded verlet structure
 661  * \param[in,out] pmedata          PME module data
 662  * \param[in,out] forceOutputs     Output buffer for the forces and virial
 663  * \param[in,out] enerd            Energy data structure results are reduced into
 664  * \param[in]     stepWork         Step schedule flags
 665  * \param[in]     pmeFlags         PME flags
 666  * \param[in]     wcycle           The wallcycle structure
 667  */
 668 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t *nbv,
 669                                         gmx_pme_t          *pmedata,
 670                                         gmx::ForceOutputs  *forceOutputs,
 671                                         gmx_enerdata_t     *enerd,
 672                                         const StepWorkload &stepWork,
 673                                         int                 pmeFlags,
 674                                         gmx_wallcycle_t     wcycle)
 675 {
 676     bool isPmeGpuDone = false;
 677     bool isNbGpuDone  = false;
 678
 679
 680
 681     gmx::ForceWithShiftForces      &forceWithShiftForces = forceOutputs->forceWithShiftForces();
 682     gmx::ForceWithVirial           &forceWithVirial      = forceOutputs->forceWithVirial();
 683
 684     gmx::ArrayRef<const gmx::RVec>  pmeGpuForces;
 685
 686     while (!isPmeGpuDone || !isNbGpuDone)
 687     {
 688         if (!isPmeGpuDone)
 689         {
 690             GpuTaskCompletion completionType = (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 691             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, &forceWithVirial, enerd, completionType);
 692         }
 693
 694         if (!isNbGpuDone)
 695         {
 696             GpuTaskCompletion completionType = (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 697             isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
 698                                                      stepWork,
 699                                                      AtomLocality::Local,
 700                                                      enerd->grpp.ener[egLJSR].data(),
 701                                                      enerd->grpp.ener[egCOULSR].data(),
 702                                                      forceWithShiftForces.shiftForces(), completionType, wcycle);
 703
 704             if (isNbGpuDone)
 705             {
 706                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local,
 707                                               forceWithShiftForces.force());
 708             }
 709         }
 710     }
 711 }
 712
 713 /*! \brief Set up the different force buffers; also does clearing.
 714  *
 715  * \param[in] fr        force record pointer
 716  * \param[in] pull_work The pull work object.
 717  * \param[in] inputrec  input record
 718  * \param[in] force     force array
 719  * \param[in] stepWork  Step schedule flags
 720  * \param[out] wcycle   wallcycle recording structure
 721  *
 722  * \returns             Cleared force output structure
 723  */
 724 static ForceOutputs
 725 setupForceOutputs(t_forcerec                          *fr,
 726                   pull_t                              *pull_work,
 727                   const t_inputrec                    &inputrec,
 728                   gmx::ArrayRefWithPadding<gmx::RVec>  force,
 729                   const StepWorkload                  &stepWork,
 730                   gmx_wallcycle_t                      wcycle)
 731 {
 732     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 733
 734     /* NOTE: We assume fr->shiftForces is all zeros here */
 735     gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial, fr->shiftForces);
 736
 737     if (stepWork.computeForces)
 738     {
 739         /* Clear the short- and long-range forces */
 740         clear_rvecs_omp(fr->natoms_force_constr,
 741                         as_rvec_array(forceWithShiftForces.force().data()));
 742     }
 743
 744     /* If we need to compute the virial, we might need a separate
 745      * force buffer for algorithms for which the virial is calculated
 746      * directly, such as PME. Otherwise, forceWithVirial uses the
 747      * the same force (f in legacy calls) buffer as other algorithms.
 748      */
 749     const bool useSeparateForceWithVirialBuffer = (stepWork.computeForces &&
 750                                                    (stepWork.computeVirial && fr->haveDirectVirialContributions));
 751     /* forceWithVirial uses the local atom range only */
 752     gmx::ForceWithVirial forceWithVirial(useSeparateForceWithVirialBuffer ?
 753                                          fr->forceBufferForDirectVirialContributions : force.unpaddedArrayRef(),
 754                                          stepWork.computeVirial);
 755
 756     if (useSeparateForceWithVirialBuffer)
 757     {
 758         /* TODO: update comment
 759          * We only compute forces on local atoms. Note that vsites can
 760          * spread to non-local atoms, but that part of the buffer is
 761          * cleared separately in the vsite spreading code.
 762          */
 763         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 764     }
 765
 766     if (inputrec.bPull && pull_have_constraint(pull_work))
 767     {
 768         clear_pull_forces(pull_work);
 769     }
 770
 771     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 772
 773     return ForceOutputs(forceWithShiftForces, forceWithVirial);
 774 }
 775
 776
 777 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
 778  */
 779 static DomainLifetimeWorkload
 780 setupDomainLifetimeWorkload(const t_inputrec         &inputrec,
 781                             const t_forcerec         &fr,
 782                             const pull_t             *pull_work,
 783                             const gmx_edsam          *ed,
 784                             const t_idef             &idef,
 785                             const t_fcdata           &fcd,
 786                             const t_mdatoms          &mdatoms,
 787                             const SimulationWorkload &simulationWork,
 788                             const StepWorkload       &stepWork)
 789 {
 790     DomainLifetimeWorkload domainWork;
 791     // Note that haveSpecialForces is constant over the whole run
 792     domainWork.haveSpecialForces      = haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 793     domainWork.haveCpuBondedWork      = haveCpuBondeds(fr);
 794     domainWork.haveGpuBondedWork      = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
 795     domainWork.haveRestraintsWork     = haveRestraints(idef, fcd);
 796     domainWork.haveCpuListedForceWork = haveCpuListedForces(fr, idef, fcd);
 797     // Note that haveFreeEnergyWork is constant over the whole run
 798     domainWork.haveFreeEnergyWork     = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
 799     // We assume we have local force work if there are CPU
 800     // force tasks including PME or nonbondeds.
 801     domainWork.haveCpuLocalForceWork  = domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork || domainWork.haveFreeEnergyWork ||
 802         simulationWork.useCpuNonbonded || simulationWork.useCpuPme ||
 803         (EEL_PME_EWALD(inputrec.coulombtype) && (inputrec.ewald_geometry == eewg3DC ||
 804                                                  inputrec.epsilon_surface != 0)) ||
 805         inputrec.nwall > 0;
 806     return domainWork;
 807 }
 808
 809 /*! \brief Set up force flag stuct from the force bitmask.
 810  *
 811  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 812  * \param[in]      isNonbondedOn        Global override, if false forces to turn off all nonbonded calculation.
 813  * \param[in]      simulationWork       Simulation workload description.
 814  * \param[in]      rankHasPmeDuty       If this rank computes PME.
 815  *
 816  * \returns New Stepworkload description.
 817  */
 818 static StepWorkload
 819 setupStepWorkload(const int                 legacyFlags,
 820                   const bool                isNonbondedOn,
 821                   const SimulationWorkload &simulationWork,
 822                   const bool                rankHasPmeDuty)
 823 {
 824     StepWorkload flags;
 825     flags.stateChanged           = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 826     flags.haveDynamicBox         = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 827     flags.doNeighborSearch       = ((legacyFlags & GMX_FORCE_NS) != 0);
 828     flags.computeVirial          = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 829     flags.computeEnergy          = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 830     flags.computeForces          = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 831     flags.computeListedForces    = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 832     flags.computeNonbondedForces = ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && isNonbondedOn;
 833     flags.computeDhdl            = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 834
 835     if (simulationWork.useGpuBufferOps)
 836     {
 837         GMX_ASSERT(simulationWork.useGpuNonbonded, "Can only offload buffer ops if nonbonded computation is also offloaded");
 838     }
 839     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 840     // on virial steps the CPU reduction path is taken
 841     // TODO: remove flags.computeEnergy, ref #3128
 842     flags.useGpuFBufferOps    = simulationWork.useGpuBufferOps && !(flags.computeVirial || flags.computeEnergy);
 843     flags.useGpuPmeFReduction = flags.useGpuFBufferOps && (simulationWork.useGpuPme &&
 844                                                            (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication));
 845
 846     return flags;
 847 }
 848
 849
 850 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
 851  *
 852  * TODO: eliminate the \p useGpuNonbonded and \p useGpuNonbonded when these are
 853  * incorporated in DomainLifetimeWorkload.
 854  */
 855 static void
 856 launchGpuEndOfStepTasks(nonbonded_verlet_t               *nbv,
 857                         gmx::GpuBonded                   *gpuBonded,
 858                         gmx_pme_t                        *pmedata,
 859                         gmx_enerdata_t                   *enerd,
 860                         const gmx::MdrunScheduleWorkload &runScheduleWork,
 861                         bool                              useGpuNonbonded,
 862                         bool                              useGpuPme,
 863                         int64_t                           step,
 864                         gmx_wallcycle_t                   wcycle)
 865 {
 866     if (useGpuNonbonded)
 867     {
 868         /* Launch pruning before buffer clearing because the API overhead of the
 869          * clear kernel launches can leave the GPU idle while it could be running
 870          * the prune kernel.
 871          */
 872         if (nbv->isDynamicPruningStepGpu(step))
 873         {
 874             nbv->dispatchPruneKernelGpu(step);
 875         }
 876
 877         /* now clear the GPU outputs while we finish the step on the CPU */
 878         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 879         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 880         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
 881         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 882         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 883     }
 884
 885     if (useGpuPme)
 886     {
 887         pme_gpu_reinit_computation(pmedata, wcycle);
 888     }
 889
 890     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
 891     {
 892         // in principle this should be included in the DD balancing region,
 893         // but generally it is infrequent so we'll omit it for the sake of
 894         // simpler code
 895         gpuBonded->waitAccumulateEnergyTerms(enerd);
 896
 897         gpuBonded->clearEnergies();
 898     }
 899 }
 900
 901
 902 void do_force(FILE                                     *fplog,
 903               const t_commrec                          *cr,
 904               const gmx_multisim_t                     *ms,
 905               const t_inputrec                         *inputrec,
 906               gmx::Awh                                 *awh,
 907               gmx_enfrot                               *enforcedRotation,
 908               gmx::ImdSession                          *imdSession,
 909               pull_t                                   *pull_work,
 910               int64_t                                   step,
 911               t_nrnb                                   *nrnb,
 912               gmx_wallcycle_t                           wcycle,
 913               const gmx_localtop_t                     *top,
 914               const matrix                              box,
 915               gmx::ArrayRefWithPadding<gmx::RVec>       x,
 916               history_t                                *hist,
 917               gmx::ArrayRefWithPadding<gmx::RVec>       force,
 918               tensor                                    vir_force,
 919               const t_mdatoms                          *mdatoms,
 920               gmx_enerdata_t                           *enerd,
 921               t_fcdata                                 *fcd,
 922               gmx::ArrayRef<real>                       lambda,
 923               t_graph                                  *graph,
 924               t_forcerec                               *fr,
 925               gmx::MdrunScheduleWorkload               *runScheduleWork,
 926               const gmx_vsite_t                        *vsite,
 927               rvec                                      mu_tot,
 928               double                                    t,
 929               gmx_edsam                                *ed,
 930               int                                       legacyFlags,
 931               const DDBalanceRegionHandler             &ddBalanceRegionHandler)
 932 {
 933     int                          i, j;
 934     double                       mu[2*DIM];
 935     nonbonded_verlet_t          *nbv      = fr->nbv.get();
 936     interaction_const_t         *ic       = fr->ic;
 937     gmx::StatePropagatorDataGpu *stateGpu = fr->stateGpu;
 938
 939     // TODO remove the code below when the legacy flags are not in use anymore
 940     /* modify force flag if not doing nonbonded */
 941     if (!fr->bNonbonded)
 942     {
 943         legacyFlags &= ~GMX_FORCE_NONBONDED;
 944     }
 945
 946     const SimulationWorkload &simulationWork = runScheduleWork->simulationWork;
 947
 948
 949     runScheduleWork->stepWork = setupStepWorkload(legacyFlags, fr->bNonbonded,
 950                                                   simulationWork, thisRankHasDuty(cr, DUTY_PME));
 951     const StepWorkload &stepWork = runScheduleWork->stepWork;
 952
 953
 954     const bool useGpuPmeOnThisRank = simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME);
 955     const int  pmeFlags            = makePmeFlags(stepWork);
 956
 957     // Switches on whether to use GPU for position and force buffer operations
 958     // TODO consider all possible combinations of triggers, and how to combine optimally in each case.
 959     const BufferOpsUseGpu useGpuXBufOps = stepWork.useGpuXBufferOps ? BufferOpsUseGpu::True : BufferOpsUseGpu::False;
 960     // GPU Force buffer ops are disabled on virial steps, because the virial calc is not yet ported to GPU
 961     const BufferOpsUseGpu useGpuFBufOps = stepWork.useGpuFBufferOps ? BufferOpsUseGpu::True : BufferOpsUseGpu::False;
 962
 963     /* At a search step we need to start the first balancing region
 964      * somewhere early inside the step after communication during domain
 965      * decomposition (and not during the previous step as usual).
 966      */
 967     if (stepWork.doNeighborSearch)
 968     {
 969         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 970     }
 971
 972     const int start  = 0;
 973     const int homenr = mdatoms->homenr;
 974
 975     clear_mat(vir_force);
 976
 977     if (stepWork.stateChanged)
 978     {
 979         if (inputrecNeedMutot(inputrec))
 980         {
 981             /* Calculate total (local) dipole moment in a temporary common array.
 982              * This makes it possible to sum them over nodes faster.
 983              */
 984             calc_mu(start, homenr,
 985                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 986                     mu, mu+DIM);
 987         }
 988     }
 989
 990     if (fr->ePBC != epbcNONE)
 991     {
 992         /* Compute shift vectors every step,
 993          * because of pressure coupling or box deformation!
 994          */
 995         if (stepWork.haveDynamicBox && stepWork.stateChanged)
 996         {
 997             calc_shifts(box, fr->shift_vec);
 998         }
 999
1000         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
1001         const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
1002         if (calcCGCM)
1003         {
1004             put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr), gmx_omp_nthreads_get(emntDefault));
1005             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
1006         }
1007         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1008         {
1009             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
1010         }
1011     }
1012
1013     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox,
1014                                  fr->shift_vec, nbv->nbat.get());
1015
1016     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1017     // The local coordinates can be copied right away.
1018     // NOTE: Consider moving this copy to right after they are updated and constrained,
1019     //       if the later is not offloaded.
1020     if (useGpuPmeOnThisRank || useGpuXBufOps == BufferOpsUseGpu::True)
1021     {
1022         if (stepWork.doNeighborSearch)
1023         {
1024             stateGpu->reinit(mdatoms->homenr, cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1025             if (useGpuPmeOnThisRank)
1026             {
1027                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1028                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1029             }
1030         }
1031         // We need to copy coordinates when:
1032         // 1. Update is not offloaded
1033         // 2. The buffers were reinitialized on search step
1034         if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1035         {
1036             stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1037         }
1038     }
1039
1040     // Copy coordinate from the GPU if update is on the GPU and there are forces to be computed on the CPU. At search steps the
1041     // current coordinates are already on the host, hence copy is not needed.
1042     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch &&
1043         runScheduleWork->domainWork.haveCpuLocalForceWork)
1044     {
1045         stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1046         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1047     }
1048
1049     const auto localXReadyOnDevice = (stateGpu != nullptr) ? stateGpu->getCoordinatesReadyOnDeviceEvent(AtomLocality::Local,
1050                                                                                                         simulationWork, stepWork) : nullptr;
1051
1052 #if GMX_MPI
1053     if (!thisRankHasDuty(cr, DUTY_PME))
1054     {
1055         /* Send particle coordinates to the pme nodes.
1056          * Since this is only implemented for domain decomposition
1057          * and domain decomposition does not use the graph,
1058          * we do not need to worry about shifting.
1059          */
1060         bool reinitGpuPmePpComms    = simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1061         bool sendCoordinatesFromGpu = simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1062         gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
1063                                  lambda[efptCOUL], lambda[efptVDW],
1064                                  (stepWork.computeVirial || stepWork.computeEnergy),
1065                                  step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1066                                  sendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1067     }
1068 #endif /* GMX_MPI */
1069
1070     if (useGpuPmeOnThisRank)
1071     {
1072         launchPmeGpuSpread(fr->pmedata, box, stepWork, pmeFlags,
1073                            localXReadyOnDevice, wcycle);
1074     }
1075
1076     /* do gridding for pair search */
1077     if (stepWork.doNeighborSearch)
1078     {
1079         if (graph && stepWork.stateChanged)
1080         {
1081             /* Calculate intramolecular shift vectors to make molecules whole */
1082             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
1083         }
1084
1085         // TODO
1086         // - vzero is constant, do we need to pass it?
1087         // - box_diag should be passed directly to nbnxn_put_on_grid
1088         //
1089         rvec vzero;
1090         clear_rvec(vzero);
1091
1092         rvec box_diag;
1093         box_diag[XX] = box[XX][XX];
1094         box_diag[YY] = box[YY][YY];
1095         box_diag[ZZ] = box[ZZ][ZZ];
1096
1097         wallcycle_start(wcycle, ewcNS);
1098         if (!DOMAINDECOMP(cr))
1099         {
1100             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1101             nbnxn_put_on_grid(nbv, box,
1102                               0, vzero, box_diag,
1103                               nullptr, { 0, mdatoms->homenr }, -1,
1104                               fr->cginfo, x.unpaddedArrayRef(),
1105                               0, nullptr);
1106             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1107         }
1108         else
1109         {
1110             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1111             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
1112                                        fr->cginfo, x.unpaddedArrayRef());
1113             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1114         }
1115
1116         nbv->setAtomProperties(*mdatoms, fr->cginfo);
1117
1118         wallcycle_stop(wcycle, ewcNS);
1119
1120         /* initialize the GPU nbnxm atom data and bonded data structures */
1121         if (simulationWork.useGpuNonbonded)
1122         {
1123             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1124
1125             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1126             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1127             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1128
1129             if (fr->gpuBonded)
1130             {
1131                 /* Now we put all atoms on the grid, we can assign bonded
1132                  * interactions to the GPU, where the grid order is
1133                  * needed. Also the xq, f and fshift device buffers have
1134                  * been reallocated if needed, so the bonded code can
1135                  * learn about them. */
1136                 // TODO the xq, f, and fshift buffers are now shared
1137                 // resources, so they should be maintained by a
1138                 // higher-level object than the nb module.
1139                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbv->getGridIndices(),
1140                                                                       top->idef,
1141                                                                       Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1142                                                                       Nbnxm::gpu_get_f(nbv->gpu_nbv),
1143                                                                       Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1144             }
1145             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1146         }
1147     }
1148
1149     if (stepWork.doNeighborSearch)
1150     {
1151         // Need to run after the GPU-offload bonded interaction lists
1152         // are set up to be able to determine whether there is bonded work.
1153         runScheduleWork->domainWork =
1154             setupDomainLifetimeWorkload(*inputrec,
1155                                         *fr,
1156                                         pull_work,
1157                                         ed,
1158                                         top->idef,
1159                                         *fcd,
1160                                         *mdatoms,
1161                                         simulationWork,
1162                                         stepWork);
1163
1164         wallcycle_start_nocount(wcycle, ewcNS);
1165         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1166         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1167         nbv->constructPairlist(InteractionLocality::Local,
1168                                &top->excls, step, nrnb);
1169
1170         nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::Local);
1171
1172         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1173         wallcycle_stop(wcycle, ewcNS);
1174
1175         if (useGpuXBufOps == BufferOpsUseGpu::True)
1176         {
1177             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1178         }
1179         // For force buffer ops, we use the below conditon rather than
1180         // useGpuFBufOps to ensure that init is performed even if this
1181         // NS step is also a virial step (on which f buf ops are deactivated).
1182         if (simulationWork.useGpuBufferOps && simulationWork.useGpuNonbonded && (GMX_GPU == GMX_GPU_CUDA))
1183         {
1184             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1185             nbv->atomdata_init_add_nbat_f_to_f_gpu(stateGpu->fReducedOnDevice());
1186         }
1187     }
1188     else if (!EI_TPI(inputrec->eI))
1189     {
1190         if (useGpuXBufOps == BufferOpsUseGpu::True)
1191         {
1192             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1193             nbv->convertCoordinatesGpu(AtomLocality::Local, false,
1194                                        stateGpu->getCoordinates(),
1195                                        localXReadyOnDevice);
1196         }
1197         else
1198         {
1199             nbv->convertCoordinates(AtomLocality::Local, false,
1200                                     x.unpaddedArrayRef());
1201         }
1202     }
1203
1204     const gmx::DomainLifetimeWorkload &domainWork = runScheduleWork->domainWork;
1205
1206     if (simulationWork.useGpuNonbonded)
1207     {
1208         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1209
1210         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1211
1212         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1213         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1214         if (stepWork.doNeighborSearch || (useGpuXBufOps == BufferOpsUseGpu::False))
1215         {
1216             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1217                                       AtomLocality::Local);
1218         }
1219         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1220         // with X buffer ops offloaded to the GPU on all but the search steps
1221
1222         // bonded work not split into separate local and non-local, so with DD
1223         // we can only launch the kernel after non-local coordinates have been received.
1224         if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1225         {
1226             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1227             fr->gpuBonded->launchKernel(fr, stepWork, box);
1228             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1229         }
1230
1231         /* launch local nonbonded work on GPU */
1232         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1233         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo,
1234                      step, nrnb, wcycle);
1235         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1236         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1237     }
1238
1239     if (useGpuPmeOnThisRank)
1240     {
1241         // In PME GPU and mixed mode we launch FFT / gather after the
1242         // X copy/transform to allow overlap as well as after the GPU NB
1243         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1244         // the nonbonded kernel.
1245         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1246     }
1247
1248     // TODO Update this comment when introducing SimulationWorkload
1249     //
1250     // The conditions for gpuHaloExchange e.g. using GPU buffer
1251     // operations were checked before construction, so here we can
1252     // just use it and assert upon any conditions.
1253     gmx::GpuHaloExchange *gpuHaloExchange              = (havePPDomainDecomposition(cr) ? cr->dd->gpuHaloExchange.get() : nullptr);
1254     const bool            ddUsesGpuDirectCommunication = (gpuHaloExchange != nullptr);
1255     GMX_ASSERT(!ddUsesGpuDirectCommunication || (useGpuXBufOps == BufferOpsUseGpu::True),
1256                "Must use coordinate buffer ops with GPU halo exchange");
1257     const bool useGpuForcesHaloExchange = ddUsesGpuDirectCommunication && (useGpuFBufOps == BufferOpsUseGpu::True);
1258
1259     /* Communicate coordinates and sum dipole if necessary +
1260        do non-local pair search */
1261     if (havePPDomainDecomposition(cr))
1262     {
1263         if (stepWork.doNeighborSearch)
1264         {
1265             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1266             wallcycle_start_nocount(wcycle, ewcNS);
1267             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1268             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1269             nbv->constructPairlist(InteractionLocality::NonLocal,
1270                                    &top->excls, step, nrnb);
1271
1272             nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::NonLocal);
1273             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1274             wallcycle_stop(wcycle, ewcNS);
1275             if (ddUsesGpuDirectCommunication)
1276             {
1277                 gpuHaloExchange->reinitHalo(stateGpu->getCoordinates(), stateGpu->getForces());
1278             }
1279         }
1280         else
1281         {
1282             if (ddUsesGpuDirectCommunication)
1283             {
1284                 // The following must be called after local setCoordinates (which records an event
1285                 // when the coordinate data has been copied to the device).
1286                 gpuHaloExchange->communicateHaloCoordinates(box, localXReadyOnDevice);
1287
1288                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1289                 {
1290                     //non-local part of coordinate buffer must be copied back to host for CPU work
1291                     stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1292                 }
1293             }
1294             else
1295             {
1296                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1297             }
1298
1299             if (useGpuXBufOps == BufferOpsUseGpu::True)
1300             {
1301                 // The condition here was (pme != nullptr && pme_gpu_get_device_x(fr->pmedata) != nullptr)
1302                 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1303                 {
1304                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1305                 }
1306                 nbv->convertCoordinatesGpu(AtomLocality::NonLocal, false,
1307                                            stateGpu->getCoordinates(),
1308                                            stateGpu->getCoordinatesReadyOnDeviceEvent(AtomLocality::NonLocal,
1309                                                                                       simulationWork, stepWork));
1310             }
1311             else
1312             {
1313                 nbv->convertCoordinates(AtomLocality::NonLocal, false,
1314                                         x.unpaddedArrayRef());
1315             }
1316
1317         }
1318
1319         if (simulationWork.useGpuNonbonded)
1320         {
1321             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1322
1323             if (stepWork.doNeighborSearch || (useGpuXBufOps == BufferOpsUseGpu::False))
1324             {
1325                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1326                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1327                                           AtomLocality::NonLocal);
1328                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1329             }
1330
1331             if (domainWork.haveGpuBondedWork)
1332             {
1333                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1334                 fr->gpuBonded->launchKernel(fr, stepWork, box);
1335                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1336             }
1337
1338             /* launch non-local nonbonded tasks on GPU */
1339             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1340             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo,
1341                          step, nrnb, wcycle);
1342             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1343
1344             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1345         }
1346     }
1347
1348     if (simulationWork.useGpuNonbonded)
1349     {
1350         /* launch D2H copy-back F */
1351         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1352         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1353
1354         if (havePPDomainDecomposition(cr))
1355         {
1356             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1357                                       stepWork, AtomLocality::NonLocal);
1358         }
1359         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1360                                   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] =
1396                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1397                 lambda[efptCOUL]*fr->mu_tot[1][j];
1398         }
1399     }
1400
1401     /* Reset energies */
1402     reset_enerdata(enerd);
1403     /* Clear the shift forces */
1404     // TODO: This should be linked to the shift force buffer in use, or cleared before use instead
1405     for (gmx::RVec &elem : fr->shiftForces)
1406     {
1407         elem = { 0.0_real, 0.0_real, 0.0_real };
1408     }
1409
1410     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1411     {
1412         wallcycle_start(wcycle, ewcPPDURINGPME);
1413         dd_force_flop_start(cr->dd, nrnb);
1414     }
1415
1416     if (inputrec->bRot)
1417     {
1418         wallcycle_start(wcycle, ewcROT);
1419         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, stepWork.doNeighborSearch);
1420         wallcycle_stop(wcycle, ewcROT);
1421     }
1422
1423     /* Start the force cycle counter.
1424      * Note that a different counter is used for dynamic load balancing.
1425      */
1426     wallcycle_start(wcycle, ewcFORCE);
1427
1428     // Set up and clear force outputs.
1429     // We use std::move to keep the compiler happy, it has no effect.
1430     ForceOutputs forceOut = setupForceOutputs(fr, pull_work, *inputrec, std::move(force), stepWork, wcycle);
1431
1432     /* We calculate the non-bonded forces, when done on the CPU, here.
1433      * We do this before calling do_force_lowlevel, because in that
1434      * function, the listed forces are calculated before PME, which
1435      * does communication.  With this order, non-bonded and listed
1436      * force calculation imbalance can be balanced out by the domain
1437      * decomposition load balancing.
1438      */
1439
1440     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1441
1442     if (!useOrEmulateGpuNb)
1443     {
1444         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes,
1445                      step, nrnb, wcycle);
1446     }
1447
1448     if (fr->efep != efepNO)
1449     {
1450         /* Calculate the local and non-local free energy interactions here.
1451          * Happens here on the CPU both with and without GPU.
1452          */
1453         nbv->dispatchFreeEnergyKernel(InteractionLocality::Local,
1454                                       fr, as_rvec_array(x.unpaddedArrayRef().data()), &forceOut.forceWithShiftForces(), *mdatoms,
1455                                       inputrec->fepvals, lambda.data(),
1456                                       enerd, stepWork, nrnb);
1457
1458         if (havePPDomainDecomposition(cr))
1459         {
1460             nbv->dispatchFreeEnergyKernel(InteractionLocality::NonLocal,
1461                                           fr, as_rvec_array(x.unpaddedArrayRef().data()), &forceOut.forceWithShiftForces(), *mdatoms,
1462                                           inputrec->fepvals, lambda.data(),
1463                                           enerd, stepWork, nrnb);
1464         }
1465     }
1466
1467     if (!useOrEmulateGpuNb)
1468     {
1469         if (havePPDomainDecomposition(cr))
1470         {
1471             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo,
1472                          step, nrnb, wcycle);
1473         }
1474
1475         if (stepWork.computeForces)
1476         {
1477             /* Add all the non-bonded force to the normal force array.
1478              * This can be split into a local and a non-local part when overlapping
1479              * communication with calculation with domain decomposition.
1480              */
1481             wallcycle_stop(wcycle, ewcFORCE);
1482             nbv->atomdata_add_nbat_f_to_f(AtomLocality::All, forceOut.forceWithShiftForces().force());
1483             wallcycle_start_nocount(wcycle, ewcFORCE);
1484         }
1485
1486         /* If there are multiple fshift output buffers we need to reduce them */
1487         if (stepWork.computeVirial)
1488         {
1489             /* This is not in a subcounter because it takes a
1490                negligible and constant-sized amount of time */
1491             nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1492                                                      forceOut.forceWithShiftForces().shiftForces());
1493         }
1494     }
1495
1496     /* update QMMMrec, if necessary */
1497     if (fr->bQMMM)
1498     {
1499         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1500     }
1501
1502     // TODO Force flags should include haveFreeEnergyWork for this domain
1503     if (ddUsesGpuDirectCommunication &&
1504         (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1505     {
1506         /* Wait for non-local coordinate data to be copied from device */
1507         nbv->wait_nonlocal_x_copy_D2H_done();
1508     }
1509     /* Compute the bonded and non-bonded energies and optionally forces */
1510     do_force_lowlevel(fr, inputrec, &(top->idef),
1511                       cr, ms, nrnb, wcycle, mdatoms,
1512                       x, hist, &forceOut, enerd, fcd,
1513                       box, lambda.data(), graph, fr->mu_tot,
1514                       stepWork,
1515                       ddBalanceRegionHandler);
1516
1517     wallcycle_stop(wcycle, ewcFORCE);
1518
1519     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1520                          imdSession, pull_work, step, t, wcycle,
1521                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1522                          stepWork, &forceOut.forceWithVirial(), enerd,
1523                          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(nbv->gpu_nbv,
1539                                                                stepWork, AtomLocality::NonLocal,
1540                                                                enerd->grpp.ener[egLJSR].data(),
1541                                                                enerd->grpp.ener[egCOULSR].data(),
1542                                                                forceWithShiftForces.shiftForces(),
1543                                                                wcycle);
1544             }
1545             else
1546             {
1547                 wallcycle_start_nocount(wcycle, ewcFORCE);
1548                 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes,
1549                              step, nrnb, wcycle);
1550                 wallcycle_stop(wcycle, ewcFORCE);
1551             }
1552
1553             if (useGpuFBufOps == BufferOpsUseGpu::True)
1554             {
1555                 gmx::FixedCapacityVector<GpuEventSynchronizer*, 1> dependencyList;
1556
1557                 // TODO: move this into DomainLifetimeWorkload, including the second part of the condition
1558                 // The bonded and free energy CPU tasks can have non-local force contributions
1559                 // which are a dependency for the GPU force reduction.
1560                 bool  haveNonLocalForceContribInCpuBuffer = domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1561
1562                 if (haveNonLocalForceContribInCpuBuffer)
1563                 {
1564                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), AtomLocality::NonLocal);
1565                     dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(AtomLocality::NonLocal,
1566                                                                                    useGpuFBufOps == BufferOpsUseGpu::True));
1567                 }
1568
1569                 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::NonLocal,
1570                                                   stateGpu->getForces(),
1571                                                   pme_gpu_get_device_f(fr->pmedata),
1572                                                   dependencyList,
1573                                                   false, haveNonLocalForceContribInCpuBuffer);
1574                 if (!useGpuForcesHaloExchange)
1575                 {
1576                     // copy from GPU input for dd_move_f()
1577                     stateGpu->copyForcesFromGpu(forceOut.forceWithShiftForces().force(), AtomLocality::NonLocal);
1578                 }
1579             }
1580             else
1581             {
1582                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal,
1583                                               forceWithShiftForces.force());
1584             }
1585
1586
1587             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1588             {
1589                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1590                                                          forceWithShiftForces.shiftForces());
1591             }
1592         }
1593     }
1594
1595     if (havePPDomainDecomposition(cr))
1596     {
1597         /* We are done with the CPU compute.
1598          * We will now communicate the non-local forces.
1599          * If we use a GPU this will overlap with GPU work, so in that case
1600          * we do not close the DD force balancing region here.
1601          */
1602         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1603
1604         if (stepWork.computeForces)
1605         {
1606
1607             if (useGpuForcesHaloExchange)
1608             {
1609                 if (domainWork.haveCpuLocalForceWork)
1610                 {
1611                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), AtomLocality::Local);
1612                 }
1613                 gpuHaloExchange->communicateHaloForces(domainWork.haveCpuLocalForceWork);
1614             }
1615             else
1616             {
1617                 if (useGpuFBufOps == BufferOpsUseGpu::True)
1618                 {
1619                     stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
1620                 }
1621                 dd_move_f(cr->dd, &forceOut.forceWithShiftForces(), wcycle);
1622             }
1623
1624         }
1625     }
1626
1627     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1628     // an alternating wait/reduction scheme.
1629     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded && !DOMAINDECOMP(cr) &&
1630                              (useGpuFBufOps == BufferOpsUseGpu::False));
1631     if (alternateGpuWait)
1632     {
1633         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &forceOut, enerd,
1634                                     stepWork, pmeFlags, wcycle);
1635     }
1636
1637     if (!alternateGpuWait && useGpuPmeOnThisRank)
1638     {
1639         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial(), enerd);
1640     }
1641
1642     /* Wait for local GPU NB outputs on the non-alternating wait path */
1643     if (!alternateGpuWait && simulationWork.useGpuNonbonded)
1644     {
1645         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1646          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1647          * but even with a step of 0.1 ms the difference is less than 1%
1648          * of the step time.
1649          */
1650         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1651         const float waitCycles               =
1652             Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1653                                         stepWork, AtomLocality::Local,
1654                                         enerd->grpp.ener[egLJSR].data(),
1655                                         enerd->grpp.ener[egCOULSR].data(),
1656                                         forceOut.forceWithShiftForces().shiftForces(),
1657                                         wcycle);
1658
1659         if (ddBalanceRegionHandler.useBalancingRegion())
1660         {
1661             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1662             if (stepWork.computeForces &&  waitCycles <= gpuWaitApiOverheadMargin)
1663             {
1664                 /* We measured few cycles, it could be that the kernel
1665                  * and transfer finished earlier and there was no actual
1666                  * wait time, only API call overhead.
1667                  * Then the actual time could be anywhere between 0 and
1668                  * cycles_wait_est. We will use half of cycles_wait_est.
1669                  */
1670                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1671             }
1672             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1673         }
1674     }
1675
1676     if (fr->nbv->emulateGpu())
1677     {
1678         // NOTE: emulation kernel is not included in the balancing region,
1679         // but emulation mode does not target performance anyway
1680         wallcycle_start_nocount(wcycle, ewcFORCE);
1681         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local,
1682                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1683                      step, nrnb, wcycle);
1684         wallcycle_stop(wcycle, ewcFORCE);
1685     }
1686
1687     // If on GPU PME-PP comms path, receive forces from PME before GPU buffer ops
1688     // TODO refoactor this and unify with below default-path call to the same function
1689     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && simulationWork.useGpuPmePpCommunication)
1690     {
1691         /* In case of node-splitting, the PP nodes receive the long-range
1692          * forces, virial and energy from the PME nodes here.
1693          */
1694         pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd, simulationWork.useGpuPmePpCommunication, stepWork.useGpuPmeFReduction, wcycle);
1695     }
1696
1697
1698     /* Do the nonbonded GPU (or emulation) force buffer reduction
1699      * on the non-alternating path. */
1700     if (useOrEmulateGpuNb && !alternateGpuWait)
1701     {
1702         //TODO simplify the below conditionals. Pass buffer and sync pointers at init stage rather than here. Unify getter fns for sameGPU/otherGPU cases.
1703         void* pmeForcePtr = stepWork.useGpuPmeFReduction ?
1704             (thisRankHasDuty(cr, DUTY_PME) ?
1705              pme_gpu_get_device_f(fr->pmedata) :        // PME force buffer on same GPU
1706              fr->pmePpCommGpu->getGpuForceStagingPtr()) // buffer received from other GPU
1707             : nullptr;                                  // PME reduction not active on GPU
1708
1709         GpuEventSynchronizer* const pmeSynchronizer = stepWork.useGpuPmeFReduction ?
1710             (thisRankHasDuty(cr, DUTY_PME) ?
1711              pme_gpu_get_f_ready_synchronizer(fr->pmedata) :   // PME force buffer on same GPU
1712              static_cast<GpuEventSynchronizer*>
1713              (fr->pmePpCommGpu->getForcesReadySynchronizer())) // buffer received from other GPU
1714             : nullptr;                                         // PME reduction not active on GPU
1715
1716         gmx::FixedCapacityVector<GpuEventSynchronizer*, 2> dependencyList;
1717
1718         if (stepWork.useGpuPmeFReduction)
1719         {
1720             dependencyList.push_back(pmeSynchronizer);
1721         }
1722
1723         gmx::ArrayRef<gmx::RVec>  forceWithShift = forceOut.forceWithShiftForces().force();
1724
1725         if (useGpuFBufOps == BufferOpsUseGpu::True)
1726         {
1727             // Flag to specify whether the CPU force buffer has contributions to
1728             // local atoms. This depends on whether there are CPU-based force tasks
1729             // or when DD is active the halo exchange has resulted in contributions
1730             // from the non-local part.
1731             const bool haveLocalForceContribInCpuBuffer = (domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr));
1732
1733             // TODO: move these steps as early as possible:
1734             // - CPU f H2D should be as soon as all CPU-side forces are done
1735             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1736             //   before the next CPU task that consumes the forces: vsite spread or update)
1737             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
1738             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
1739             //   These should be unified.
1740             if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
1741             {
1742                 stateGpu->copyForcesToGpu(forceWithShift, AtomLocality::Local);
1743                 dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(AtomLocality::Local,
1744                                                                                useGpuFBufOps == BufferOpsUseGpu::True));
1745             }
1746             if (useGpuForcesHaloExchange)
1747             {
1748                 // Add a stream synchronization to satisfy a dependency
1749                 // for the local buffer ops on the result of GPU halo
1750                 // exchange, which operates in the non-local stream and
1751                 // writes to to local parf og the force buffer.
1752                 //
1753                 // TODO improve this through use of an event - see Redmine #3093
1754                 //      push the event into the dependencyList
1755                 nbv->stream_local_wait_for_nonlocal();
1756             }
1757             nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::Local,
1758                                               stateGpu->getForces(),
1759                                               pmeForcePtr,
1760                                               dependencyList,
1761                                               stepWork.useGpuPmeFReduction, haveLocalForceContribInCpuBuffer);
1762             // Copy forces to host if they are needed for update or if virtual sites are enabled.
1763             // If there are vsites, we need to copy forces every step to spread vsite forces on host.
1764             // TODO: When the output flags will be included in step workload, this copy can be combined with the
1765             //       copy call done in sim_utils(...) for the output.
1766             // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
1767             //       they should not be copied in do_md(...) for the output.
1768             if (!simulationWork.useGpuUpdate || vsite)
1769             {
1770                 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
1771                 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
1772             }
1773         }
1774         else
1775         {
1776             nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
1777         }
1778
1779     }
1780
1781     launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd,
1782                             *runScheduleWork,
1783                             simulationWork.useGpuNonbonded, useGpuPmeOnThisRank,
1784                             step,
1785                             wcycle);
1786
1787     if (DOMAINDECOMP(cr))
1788     {
1789         dd_force_flop_stop(cr->dd, nrnb);
1790     }
1791
1792     if (stepWork.computeForces)
1793     {
1794         rvec *f = as_rvec_array(forceOut.forceWithShiftForces().force().data());
1795
1796         /* If we have NoVirSum forces, but we do not calculate the virial,
1797          * we sum fr->f_novirsum=forceOut.f later.
1798          */
1799         if (vsite && !(fr->haveDirectVirialContributions && !stepWork.computeVirial))
1800         {
1801             rvec *fshift = as_rvec_array(forceOut.forceWithShiftForces().shiftForces().data());
1802             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), f, fshift, FALSE, nullptr, nrnb,
1803                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1804         }
1805
1806         if (stepWork.computeVirial)
1807         {
1808             /* Calculation of the virial must be done after vsites! */
1809             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()),
1810                         forceOut.forceWithShiftForces(),
1811                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1812         }
1813     }
1814
1815     // TODO refoactor this and unify with above PME-PP GPU communication path call to the same function
1816     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication)
1817     {
1818         /* In case of node-splitting, the PP nodes receive the long-range
1819          * forces, virial and energy from the PME nodes here.
1820          */
1821         pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1822                                simulationWork.useGpuPmePpCommunication, false, wcycle);
1823     }
1824
1825     if (stepWork.computeForces)
1826     {
1827         post_process_forces(cr, step, nrnb, wcycle,
1828                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), &forceOut,
1829                             vir_force, mdatoms, graph, fr, vsite,
1830                             stepWork);
1831     }
1832
1833     if (stepWork.computeEnergy)
1834     {
1835         /* Sum the potential energy terms from group contributions */
1836         sum_epot(&(enerd->grpp), enerd->term);
1837
1838         if (!EI_TPI(inputrec->eI))
1839         {
1840             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1841         }
1842     }
1843
1844     /* In case we don't have constraints and are using GPUs, the next balancing
1845      * region starts here.
1846      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1847      * virial calculation and COM pulling, is not thus not included in
1848      * the balance timing, which is ok as most tasks do communication.
1849      */
1850     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1851 }