src/gromacs/mdlib/sim_util.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013-2019,2020, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "config.h"
  40
  41 #include <cmath>
  42 #include <cstdint>
  43 #include <cstdio>
  44 #include <cstring>
  45
  46 #include <array>
  47
  48 #include "gromacs/awh/awh.h"
  49 #include "gromacs/domdec/dlbtiming.h"
  50 #include "gromacs/domdec/domdec.h"
  51 #include "gromacs/domdec/domdec_struct.h"
  52 #include "gromacs/domdec/gpuhaloexchange.h"
  53 #include "gromacs/domdec/partition.h"
  54 #include "gromacs/essentialdynamics/edsam.h"
  55 #include "gromacs/ewald/pme.h"
  56 #include "gromacs/ewald/pme_pp.h"
  57 #include "gromacs/ewald/pme_pp_comm_gpu.h"
  58 #include "gromacs/gmxlib/network.h"
  59 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  60 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  61 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  62 #include "gromacs/gpu_utils/gpu_utils.h"
  63 #include "gromacs/imd/imd.h"
  64 #include "gromacs/listed_forces/disre.h"
  65 #include "gromacs/listed_forces/gpubonded.h"
  66 #include "gromacs/listed_forces/listed_forces.h"
  67 #include "gromacs/listed_forces/manage_threading.h"
  68 #include "gromacs/listed_forces/orires.h"
  69 #include "gromacs/math/arrayrefwithpadding.h"
  70 #include "gromacs/math/functions.h"
  71 #include "gromacs/math/units.h"
  72 #include "gromacs/math/vec.h"
  73 #include "gromacs/math/vecdump.h"
  74 #include "gromacs/mdlib/calcmu.h"
  75 #include "gromacs/mdlib/calcvir.h"
  76 #include "gromacs/mdlib/constr.h"
  77 #include "gromacs/mdlib/enerdata_utils.h"
  78 #include "gromacs/mdlib/force.h"
  79 #include "gromacs/mdlib/force_flags.h"
  80 #include "gromacs/mdlib/forcerec.h"
  81 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  82 #include "gromacs/mdlib/qmmm.h"
  83 #include "gromacs/mdlib/update.h"
  84 #include "gromacs/mdlib/vsite.h"
  85 #include "gromacs/mdtypes/commrec.h"
  86 #include "gromacs/mdtypes/enerdata.h"
  87 #include "gromacs/mdtypes/forceoutput.h"
  88 #include "gromacs/mdtypes/forcerec.h"
  89 #include "gromacs/mdtypes/iforceprovider.h"
  90 #include "gromacs/mdtypes/inputrec.h"
  91 #include "gromacs/mdtypes/md_enums.h"
  92 #include "gromacs/mdtypes/mdatom.h"
  93 #include "gromacs/mdtypes/simulation_workload.h"
  94 #include "gromacs/mdtypes/state.h"
  95 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  96 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  97 #include "gromacs/nbnxm/nbnxm.h"
  98 #include "gromacs/nbnxm/nbnxm_gpu.h"
  99 #include "gromacs/pbcutil/ishift.h"
 100 #include "gromacs/pbcutil/mshift.h"
 101 #include "gromacs/pbcutil/pbc.h"
 102 #include "gromacs/pulling/pull.h"
 103 #include "gromacs/pulling/pull_rotation.h"
 104 #include "gromacs/timing/cyclecounter.h"
 105 #include "gromacs/timing/gpu_timing.h"
 106 #include "gromacs/timing/wallcycle.h"
 107 #include "gromacs/timing/wallcyclereporting.h"
 108 #include "gromacs/timing/walltime_accounting.h"
 109 #include "gromacs/topology/topology.h"
 110 #include "gromacs/utility/arrayref.h"
 111 #include "gromacs/utility/basedefinitions.h"
 112 #include "gromacs/utility/cstringutil.h"
 113 #include "gromacs/utility/exceptions.h"
 114 #include "gromacs/utility/fatalerror.h"
 115 #include "gromacs/utility/fixedcapacityvector.h"
 116 #include "gromacs/utility/gmxassert.h"
 117 #include "gromacs/utility/gmxmpi.h"
 118 #include "gromacs/utility/logger.h"
 119 #include "gromacs/utility/smalloc.h"
 120 #include "gromacs/utility/strconvert.h"
 121 #include "gromacs/utility/sysinfo.h"
 122
 123 using gmx::AtomLocality;
 124 using gmx::DomainLifetimeWorkload;
 125 using gmx::ForceOutputs;
 126 using gmx::InteractionLocality;
 127 using gmx::SimulationWorkload;
 128 using gmx::StepWorkload;
 129
 130 // TODO: this environment variable allows us to verify before release
 131 // that on less common architectures the total cost of polling is not larger than
 132 // a blocking wait (so polling does not introduce overhead when the static
 133 // PME-first ordering would suffice).
 134 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 135
 136 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 137 {
 138     const int end = forceToAdd.size();
 139
 140     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 141 #pragma omp parallel for num_threads(nt) schedule(static)
 142     for (int i = 0; i < end; i++)
 143     {
 144         rvec_inc(f[i], forceToAdd[i]);
 145     }
 146 }
 147
 148 static void calc_virial(int                              start,
 149                         int                              homenr,
 150                         const rvec                       x[],
 151                         const gmx::ForceWithShiftForces& forceWithShiftForces,
 152                         tensor                           vir_part,
 153                         const t_graph*                   graph,
 154                         const matrix                     box,
 155                         t_nrnb*                          nrnb,
 156                         const t_forcerec*                fr,
 157                         PbcType                          pbcType)
 158 {
 159     /* The short-range virial from surrounding boxes */
 160     const rvec* fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
 161     calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, pbcType == PbcType::Screw, box);
 162     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 163
 164     /* Calculate partial virial, for local atoms only, based on short range.
 165      * Total virial is computed in global_stat, called from do_md
 166      */
 167     const rvec* f = as_rvec_array(forceWithShiftForces.force().data());
 168     f_calc_vir(start, start + homenr, x, f, vir_part, graph, box);
 169     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 170
 171     if (debug)
 172     {
 173         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 174     }
 175 }
 176
 177 static void pull_potential_wrapper(const t_commrec*               cr,
 178                                    const t_inputrec*              ir,
 179                                    const matrix                   box,
 180                                    gmx::ArrayRef<const gmx::RVec> x,
 181                                    gmx::ForceWithVirial*          force,
 182                                    const t_mdatoms*               mdatoms,
 183                                    gmx_enerdata_t*                enerd,
 184                                    pull_t*                        pull_work,
 185                                    const real*                    lambda,
 186                                    double                         t,
 187                                    gmx_wallcycle_t                wcycle)
 188 {
 189     t_pbc pbc;
 190     real  dvdl;
 191
 192     /* Calculate the center of mass forces, this requires communication,
 193      * which is why pull_potential is called close to other communication.
 194      */
 195     wallcycle_start(wcycle, ewcPULLPOT);
 196     set_pbc(&pbc, ir->pbcType, box);
 197     dvdl = 0;
 198     enerd->term[F_COM_PULL] += pull_potential(pull_work, mdatoms, &pbc, cr, t, lambda[efptRESTRAINT],
 199                                               as_rvec_array(x.data()), force, &dvdl);
 200     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 201     for (auto& dhdl : enerd->dhdlLambda)
 202     {
 203         dhdl += dvdl;
 204     }
 205     wallcycle_stop(wcycle, ewcPULLPOT);
 206 }
 207
 208 static void pme_receive_force_ener(t_forcerec*           fr,
 209                                    const t_commrec*      cr,
 210                                    gmx::ForceWithVirial* forceWithVirial,
 211                                    gmx_enerdata_t*       enerd,
 212                                    bool                  useGpuPmePpComms,
 213                                    bool                  receivePmeForceToGpu,
 214                                    gmx_wallcycle_t       wcycle)
 215 {
 216     real  e_q, e_lj, dvdl_q, dvdl_lj;
 217     float cycles_ppdpme, cycles_seppme;
 218
 219     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 220     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 221
 222     /* In case of node-splitting, the PP nodes receive the long-range
 223      * forces, virial and energy from the PME nodes here.
 224      */
 225     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 226     dvdl_q  = 0;
 227     dvdl_lj = 0;
 228     gmx_pme_receive_f(fr->pmePpCommGpu.get(), cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 229                       useGpuPmePpComms, receivePmeForceToGpu, &cycles_seppme);
 230     enerd->term[F_COUL_RECIP] += e_q;
 231     enerd->term[F_LJ_RECIP] += e_lj;
 232     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 233     enerd->dvdl_lin[efptVDW] += dvdl_lj;
 234
 235     for (auto& dhdl : enerd->dhdlLambda)
 236     {
 237         dhdl += dvdl_q + dvdl_lj;
 238     }
 239
 240     if (wcycle)
 241     {
 242         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 243     }
 244     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 245 }
 246
 247 static void print_large_forces(FILE*            fp,
 248                                const t_mdatoms* md,
 249                                const t_commrec* cr,
 250                                int64_t          step,
 251                                real             forceTolerance,
 252                                const rvec*      x,
 253                                const rvec*      f)
 254 {
 255     real       force2Tolerance = gmx::square(forceTolerance);
 256     gmx::index numNonFinite    = 0;
 257     for (int i = 0; i < md->homenr; i++)
 258     {
 259         real force2    = norm2(f[i]);
 260         bool nonFinite = !std::isfinite(force2);
 261         if (force2 >= force2Tolerance || nonFinite)
 262         {
 263             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n", step,
 264                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 265         }
 266         if (nonFinite)
 267         {
 268             numNonFinite++;
 269         }
 270     }
 271     if (numNonFinite > 0)
 272     {
 273         /* Note that with MPI this fatal call on one rank might interrupt
 274          * the printing on other ranks. But we can only avoid that with
 275          * an expensive MPI barrier that we would need at each step.
 276          */
 277         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 278     }
 279 }
 280
 281 static void post_process_forces(const t_commrec*      cr,
 282                                 int64_t               step,
 283                                 t_nrnb*               nrnb,
 284                                 gmx_wallcycle_t       wcycle,
 285                                 const gmx_localtop_t* top,
 286                                 const matrix          box,
 287                                 const rvec            x[],
 288                                 ForceOutputs*         forceOutputs,
 289                                 tensor                vir_force,
 290                                 const t_mdatoms*      mdatoms,
 291                                 const t_graph*        graph,
 292                                 const t_forcerec*     fr,
 293                                 const gmx_vsite_t*    vsite,
 294                                 const StepWorkload&   stepWork)
 295 {
 296     rvec* f = as_rvec_array(forceOutputs->forceWithShiftForces().force().data());
 297
 298     if (fr->haveDirectVirialContributions)
 299     {
 300         auto& forceWithVirial = forceOutputs->forceWithVirial();
 301         rvec* fDirectVir      = as_rvec_array(forceWithVirial.force_.data());
 302
 303         if (vsite)
 304         {
 305             /* Spread the mesh force on virtual sites to the other particles...
 306              * This is parallellized. MPI communication is performed
 307              * if the constructing atoms aren't local.
 308              */
 309             matrix virial = { { 0 } };
 310             spread_vsite_f(vsite, x, fDirectVir, nullptr, stepWork.computeVirial, virial, nrnb,
 311                            top->idef, fr->pbcType, fr->bMolPBC, graph, box, cr, wcycle);
 312             forceWithVirial.addVirialContribution(virial);
 313         }
 314
 315         if (stepWork.computeVirial)
 316         {
 317             /* Now add the forces, this is local */
 318             sum_forces(f, forceWithVirial.force_);
 319
 320             /* Add the direct virial contributions */
 321             GMX_ASSERT(
 322                     forceWithVirial.computeVirial_,
 323                     "forceWithVirial should request virial computation when we request the virial");
 324             m_add(vir_force, forceWithVirial.getVirial(), vir_force);
 325
 326             if (debug)
 327             {
 328                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 329             }
 330         }
 331     }
 332
 333     if (fr->print_force >= 0)
 334     {
 335         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 336     }
 337 }
 338
 339 static void do_nb_verlet(t_forcerec*                fr,
 340                          const interaction_const_t* ic,
 341                          gmx_enerdata_t*            enerd,
 342                          const StepWorkload&        stepWork,
 343                          const InteractionLocality  ilocality,
 344                          const int                  clearF,
 345                          const int64_t              step,
 346                          t_nrnb*                    nrnb,
 347                          gmx_wallcycle_t            wcycle)
 348 {
 349     if (!stepWork.computeNonbondedForces)
 350     {
 351         /* skip non-bonded calculation */
 352         return;
 353     }
 354
 355     nonbonded_verlet_t* nbv = fr->nbv.get();
 356
 357     /* GPU kernel launch overhead is already timed separately */
 358     if (fr->cutoff_scheme != ecutsVERLET)
 359     {
 360         gmx_incons("Invalid cut-off scheme passed!");
 361     }
 362
 363     if (!nbv->useGpu())
 364     {
 365         /* When dynamic pair-list  pruning is requested, we need to prune
 366          * at nstlistPrune steps.
 367          */
 368         if (nbv->isDynamicPruningStepCpu(step))
 369         {
 370             /* Prune the pair-list beyond fr->ic->rlistPrune using
 371              * the current coordinates of the atoms.
 372              */
 373             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 374             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 375             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 376         }
 377     }
 378
 379     nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
 380 }
 381
 382 static inline void clear_rvecs_omp(int n, rvec v[])
 383 {
 384     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 385
 386     /* Note that we would like to avoid this conditional by putting it
 387      * into the omp pragma instead, but then we still take the full
 388      * omp parallel for overhead (at least with gcc5).
 389      */
 390     if (nth == 1)
 391     {
 392         for (int i = 0; i < n; i++)
 393         {
 394             clear_rvec(v[i]);
 395         }
 396     }
 397     else
 398     {
 399 #pragma omp parallel for num_threads(nth) schedule(static)
 400         for (int i = 0; i < n; i++)
 401         {
 402             clear_rvec(v[i]);
 403         }
 404     }
 405 }
 406
 407 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 408  *
 409  * \param groupOptions  Group options, containing T-coupling options
 410  */
 411 static real averageKineticEnergyEstimate(const t_grpopts& groupOptions)
 412 {
 413     real nrdfCoupled   = 0;
 414     real nrdfUncoupled = 0;
 415     real kineticEnergy = 0;
 416     for (int g = 0; g < groupOptions.ngtc; g++)
 417     {
 418         if (groupOptions.tau_t[g] >= 0)
 419         {
 420             nrdfCoupled += groupOptions.nrdf[g];
 421             kineticEnergy += groupOptions.nrdf[g] * 0.5 * groupOptions.ref_t[g] * BOLTZ;
 422         }
 423         else
 424         {
 425             nrdfUncoupled += groupOptions.nrdf[g];
 426         }
 427     }
 428
 429     /* This conditional with > also catches nrdf=0 */
 430     if (nrdfCoupled > nrdfUncoupled)
 431     {
 432         return kineticEnergy * (nrdfCoupled + nrdfUncoupled) / nrdfCoupled;
 433     }
 434     else
 435     {
 436         return 0;
 437     }
 438 }
 439
 440 /*! \brief This routine checks that the potential energy is finite.
 441  *
 442  * Always checks that the potential energy is finite. If step equals
 443  * inputrec.init_step also checks that the magnitude of the potential energy
 444  * is reasonable. Terminates with a fatal error when a check fails.
 445  * Note that passing this check does not guarantee finite forces,
 446  * since those use slightly different arithmetics. But in most cases
 447  * there is just a narrow coordinate range where forces are not finite
 448  * and energies are finite.
 449  *
 450  * \param[in] step      The step number, used for checking and printing
 451  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to
 452  * enerd.term[F_EPOT] before calling this routine \param[in] inputrec  The input record
 453  */
 454 static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
 455 {
 456     /* Threshold valid for comparing absolute potential energy against
 457      * the kinetic energy. Normally one should not consider absolute
 458      * potential energy values, but with a factor of one million
 459      * we should never get false positives.
 460      */
 461     constexpr real c_thresholdFactor = 1e6;
 462
 463     bool energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 464     real averageKineticEnergy = 0;
 465     /* We only check for large potential energy at the initial step,
 466      * because that is by far the most likely step for this too occur
 467      * and because computing the average kinetic energy is not free.
 468      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 469      * before they become NaN.
 470      */
 471     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 472     {
 473         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 474     }
 475
 476     if (energyIsNotFinite
 477         || (averageKineticEnergy > 0 && enerd.term[F_EPOT] > c_thresholdFactor * averageKineticEnergy))
 478     {
 479         gmx_fatal(
 480                 FARGS,
 481                 "Step %" PRId64
 482                 ": The total potential energy is %g, which is %s. The LJ and electrostatic "
 483                 "contributions to the energy are %g and %g, respectively. A %s potential energy "
 484                 "can be caused by overlapping interactions in bonded interactions or very large%s "
 485                 "coordinate values. Usually this is caused by a badly- or non-equilibrated initial "
 486                 "configuration, incorrect interactions or parameters in the topology.",
 487                 step, enerd.term[F_EPOT], energyIsNotFinite ? "not finite" : "extremely high",
 488                 enerd.term[F_LJ], enerd.term[F_COUL_SR],
 489                 energyIsNotFinite ? "non-finite" : "very high", energyIsNotFinite ? " or Nan" : "");
 490     }
 491 }
 492
 493 /*! \brief Return true if there are special forces computed this step.
 494  *
 495  * The conditionals exactly correspond to those in computeSpecialForces().
 496  */
 497 static bool haveSpecialForces(const t_inputrec&          inputrec,
 498                               const gmx::ForceProviders& forceProviders,
 499                               const pull_t*              pull_work,
 500                               const bool                 computeForces,
 501                               const gmx_edsam*           ed)
 502 {
 503
 504     return ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
 505             (inputrec.bPull && pull_have_potential(pull_work)) ||   // pull
 506             inputrec.bRot ||                                        // enforced rotation
 507             (ed != nullptr) ||                                      // flooding
 508             (inputrec.bIMD && computeForces));                      // IMD
 509 }
 510
 511 /*! \brief Compute forces and/or energies for special algorithms
 512  *
 513  * The intention is to collect all calls to algorithms that compute
 514  * forces on local atoms only and that do not contribute to the local
 515  * virial sum (but add their virial contribution separately).
 516  * Eventually these should likely all become ForceProviders.
 517  * Within this function the intention is to have algorithms that do
 518  * global communication at the end, so global barriers within the MD loop
 519  * are as close together as possible.
 520  *
 521  * \param[in]     fplog            The log file
 522  * \param[in]     cr               The communication record
 523  * \param[in]     inputrec         The input record
 524  * \param[in]     awh              The Awh module (nullptr if none in use).
 525  * \param[in]     enforcedRotation Enforced rotation module.
 526  * \param[in]     imdSession       The IMD session
 527  * \param[in]     pull_work        The pull work structure.
 528  * \param[in]     step             The current MD step
 529  * \param[in]     t                The current time
 530  * \param[in,out] wcycle           Wallcycle accounting struct
 531  * \param[in,out] forceProviders   Pointer to a list of force providers
 532  * \param[in]     box              The unit cell
 533  * \param[in]     x                The coordinates
 534  * \param[in]     mdatoms          Per atom properties
 535  * \param[in]     lambda           Array of free-energy lambda values
 536  * \param[in]     stepWork         Step schedule flags
 537  * \param[in,out] forceWithVirial  Force and virial buffers
 538  * \param[in,out] enerd            Energy buffer
 539  * \param[in,out] ed               Essential dynamics pointer
 540  * \param[in]     didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
 541  *
 542  * \todo Remove didNeighborSearch, which is used incorrectly.
 543  * \todo Convert all other algorithms called here to ForceProviders.
 544  */
 545 static void computeSpecialForces(FILE*                          fplog,
 546                                  const t_commrec*               cr,
 547                                  const t_inputrec*              inputrec,
 548                                  gmx::Awh*                      awh,
 549                                  gmx_enfrot*                    enforcedRotation,
 550                                  gmx::ImdSession*               imdSession,
 551                                  pull_t*                        pull_work,
 552                                  int64_t                        step,
 553                                  double                         t,
 554                                  gmx_wallcycle_t                wcycle,
 555                                  gmx::ForceProviders*           forceProviders,
 556                                  const matrix                   box,
 557                                  gmx::ArrayRef<const gmx::RVec> x,
 558                                  const t_mdatoms*               mdatoms,
 559                                  real*                          lambda,
 560                                  const StepWorkload&            stepWork,
 561                                  gmx::ForceWithVirial*          forceWithVirial,
 562                                  gmx_enerdata_t*                enerd,
 563                                  gmx_edsam*                     ed,
 564                                  bool                           didNeighborSearch)
 565 {
 566     /* NOTE: Currently all ForceProviders only provide forces.
 567      *       When they also provide energies, remove this conditional.
 568      */
 569     if (stepWork.computeForces)
 570     {
 571         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 572         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 573
 574         /* Collect forces from modules */
 575         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 576     }
 577
 578     if (inputrec->bPull && pull_have_potential(pull_work))
 579     {
 580         pull_potential_wrapper(cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work,
 581                                lambda, t, wcycle);
 582
 583         if (awh)
 584         {
 585             enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(
 586                     inputrec->pbcType, *mdatoms, box, forceWithVirial, t, step, wcycle, fplog);
 587         }
 588     }
 589
 590     rvec* f = as_rvec_array(forceWithVirial->force_.data());
 591
 592     /* Add the forces from enforced rotation potentials (if any) */
 593     if (inputrec->bRot)
 594     {
 595         wallcycle_start(wcycle, ewcROTadd);
 596         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 597         wallcycle_stop(wcycle, ewcROTadd);
 598     }
 599
 600     if (ed)
 601     {
 602         /* Note that since init_edsam() is called after the initialization
 603          * of forcerec, edsam doesn't request the noVirSum force buffer.
 604          * Thus if no other algorithm (e.g. PME) requires it, the forces
 605          * here will contribute to the virial.
 606          */
 607         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
 608     }
 609
 610     /* Add forces from interactive molecular dynamics (IMD), if any */
 611     if (inputrec->bIMD && stepWork.computeForces)
 612     {
 613         imdSession->applyForces(f);
 614     }
 615 }
 616
 617 /*! \brief Makes PME flags from StepWorkload data.
 618  *
 619  * \param[in]  stepWork     Step schedule flags
 620  * \returns                 PME flags
 621  */
 622 static int makePmeFlags(const StepWorkload& stepWork)
 623 {
 624     return GMX_PME_SPREAD | GMX_PME_SOLVE | (stepWork.computeVirial ? GMX_PME_CALC_ENER_VIR : 0)
 625            | (stepWork.computeEnergy ? GMX_PME_CALC_ENER_VIR : 0)
 626            | (stepWork.computeForces ? GMX_PME_CALC_F : 0);
 627 }
 628
 629 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 630  *
 631  * \param[in]  pmedata              The PME structure
 632  * \param[in]  box                  The box matrix
 633  * \param[in]  stepWork             Step schedule flags
 634  * \param[in]  pmeFlags             PME flags
 635  * \param[in]  xReadyOnDevice       Event synchronizer indicating that the coordinates are ready in
 636  * the device memory. \param[in]  wcycle               The wallcycle structure
 637  */
 638 static inline void launchPmeGpuSpread(gmx_pme_t*            pmedata,
 639                                       const matrix          box,
 640                                       const StepWorkload&   stepWork,
 641                                       int                   pmeFlags,
 642                                       GpuEventSynchronizer* xReadyOnDevice,
 643                                       gmx_wallcycle_t       wcycle)
 644 {
 645     pme_gpu_prepare_computation(pmedata, stepWork.haveDynamicBox, box, wcycle, pmeFlags,
 646                                 stepWork.useGpuPmeFReduction);
 647     pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle);
 648 }
 649
 650 /*! \brief Launch the FFT and gather stages of PME GPU
 651  *
 652  * This function only implements setting the output forces (no accumulation).
 653  *
 654  * \param[in]  pmedata        The PME structure
 655  * \param[in]  wcycle         The wallcycle structure
 656  */
 657 static void launchPmeGpuFftAndGather(gmx_pme_t* pmedata, gmx_wallcycle_t wcycle)
 658 {
 659     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 660     pme_gpu_launch_gather(pmedata, wcycle);
 661 }
 662
 663 /*! \brief
 664  *  Polling wait for either of the PME or nonbonded GPU tasks.
 665  *
 666  * Instead of a static order in waiting for GPU tasks, this function
 667  * polls checking which of the two tasks completes first, and does the
 668  * associated force buffer reduction overlapped with the other task.
 669  * By doing that, unlike static scheduling order, it can always overlap
 670  * one of the reductions, regardless of the GPU task completion order.
 671  *
 672  * \param[in]     nbv              Nonbonded verlet structure
 673  * \param[in,out] pmedata          PME module data
 674  * \param[in,out] forceOutputs     Output buffer for the forces and virial
 675  * \param[in,out] enerd            Energy data structure results are reduced into
 676  * \param[in]     stepWork         Step schedule flags
 677  * \param[in]     pmeFlags         PME flags
 678  * \param[in]     wcycle           The wallcycle structure
 679  */
 680 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
 681                                         gmx_pme_t*          pmedata,
 682                                         gmx::ForceOutputs*  forceOutputs,
 683                                         gmx_enerdata_t*     enerd,
 684                                         const StepWorkload& stepWork,
 685                                         int                 pmeFlags,
 686                                         gmx_wallcycle_t     wcycle)
 687 {
 688     bool isPmeGpuDone = false;
 689     bool isNbGpuDone  = false;
 690
 691
 692     gmx::ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
 693     gmx::ForceWithVirial&      forceWithVirial      = forceOutputs->forceWithVirial();
 694
 695     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 696
 697     while (!isPmeGpuDone || !isNbGpuDone)
 698     {
 699         if (!isPmeGpuDone)
 700         {
 701             GpuTaskCompletion completionType =
 702                     (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 703             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, &forceWithVirial,
 704                                                    enerd, completionType);
 705         }
 706
 707         if (!isNbGpuDone)
 708         {
 709             GpuTaskCompletion completionType =
 710                     (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 711             isNbGpuDone = Nbnxm::gpu_try_finish_task(
 712                     nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
 713                     enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(),
 714                     completionType, wcycle);
 715
 716             if (isNbGpuDone)
 717             {
 718                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShiftForces.force());
 719             }
 720         }
 721     }
 722 }
 723
 724 /*! \brief Set up the different force buffers; also does clearing.
 725  *
 726  * \param[in] fr        force record pointer
 727  * \param[in] pull_work The pull work object.
 728  * \param[in] inputrec  input record
 729  * \param[in] force     force array
 730  * \param[in] stepWork  Step schedule flags
 731  * \param[out] wcycle   wallcycle recording structure
 732  *
 733  * \returns             Cleared force output structure
 734  */
 735 static ForceOutputs setupForceOutputs(t_forcerec*                         fr,
 736                                       pull_t*                             pull_work,
 737                                       const t_inputrec&                   inputrec,
 738                                       gmx::ArrayRefWithPadding<gmx::RVec> force,
 739                                       const StepWorkload&                 stepWork,
 740                                       gmx_wallcycle_t                     wcycle)
 741 {
 742     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 743
 744     /* NOTE: We assume fr->shiftForces is all zeros here */
 745     gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial, fr->shiftForces);
 746
 747     if (stepWork.computeForces)
 748     {
 749         /* Clear the short- and long-range forces */
 750         clear_rvecs_omp(fr->natoms_force_constr, as_rvec_array(forceWithShiftForces.force().data()));
 751     }
 752
 753     /* If we need to compute the virial, we might need a separate
 754      * force buffer for algorithms for which the virial is calculated
 755      * directly, such as PME. Otherwise, forceWithVirial uses the
 756      * the same force (f in legacy calls) buffer as other algorithms.
 757      */
 758     const bool useSeparateForceWithVirialBuffer =
 759             (stepWork.computeForces && (stepWork.computeVirial && fr->haveDirectVirialContributions));
 760     /* forceWithVirial uses the local atom range only */
 761     gmx::ForceWithVirial forceWithVirial(useSeparateForceWithVirialBuffer ? fr->forceBufferForDirectVirialContributions
 762                                                                           : force.unpaddedArrayRef(),
 763                                          stepWork.computeVirial);
 764
 765     if (useSeparateForceWithVirialBuffer)
 766     {
 767         /* TODO: update comment
 768          * We only compute forces on local atoms. Note that vsites can
 769          * spread to non-local atoms, but that part of the buffer is
 770          * cleared separately in the vsite spreading code.
 771          */
 772         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 773     }
 774
 775     if (inputrec.bPull && pull_have_constraint(pull_work))
 776     {
 777         clear_pull_forces(pull_work);
 778     }
 779
 780     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 781
 782     return ForceOutputs(forceWithShiftForces, forceWithVirial);
 783 }
 784
 785
 786 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
 787  */
 788 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec&             inputrec,
 789                                                           const t_forcerec&             fr,
 790                                                           const pull_t*                 pull_work,
 791                                                           const gmx_edsam*              ed,
 792                                                           const InteractionDefinitions& idef,
 793                                                           const t_fcdata&               fcd,
 794                                                           const t_mdatoms&              mdatoms,
 795                                                           const SimulationWorkload& simulationWork,
 796                                                           const StepWorkload&       stepWork)
 797 {
 798     DomainLifetimeWorkload domainWork;
 799     // Note that haveSpecialForces is constant over the whole run
 800     domainWork.haveSpecialForces =
 801             haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 802     domainWork.haveCpuBondedWork  = haveCpuBondeds(fr);
 803     domainWork.haveGpuBondedWork  = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
 804     domainWork.haveRestraintsWork = haveRestraints(idef, fcd);
 805     domainWork.haveCpuListedForceWork = haveCpuListedForces(fr, idef, fcd);
 806     // Note that haveFreeEnergyWork is constant over the whole run
 807     domainWork.haveFreeEnergyWork = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
 808     // We assume we have local force work if there are CPU
 809     // force tasks including PME or nonbondeds.
 810     domainWork.haveCpuLocalForceWork =
 811             domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
 812             || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
 813             || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
 814
 815     return domainWork;
 816 }
 817
 818 /*! \brief Set up force flag stuct from the force bitmask.
 819  *
 820  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 821  * \param[in]      isNonbondedOn        Global override, if false forces to turn off all nonbonded calculation.
 822  * \param[in]      simulationWork       Simulation workload description.
 823  * \param[in]      rankHasPmeDuty       If this rank computes PME.
 824  *
 825  * \returns New Stepworkload description.
 826  */
 827 static StepWorkload setupStepWorkload(const int                 legacyFlags,
 828                                       const bool                isNonbondedOn,
 829                                       const SimulationWorkload& simulationWork,
 830                                       const bool                rankHasPmeDuty)
 831 {
 832     StepWorkload flags;
 833     flags.stateChanged           = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 834     flags.haveDynamicBox         = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 835     flags.doNeighborSearch       = ((legacyFlags & GMX_FORCE_NS) != 0);
 836     flags.computeVirial          = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 837     flags.computeEnergy          = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 838     flags.computeForces          = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 839     flags.computeListedForces    = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 840     flags.computeNonbondedForces = ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && isNonbondedOn;
 841     flags.computeDhdl            = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 842
 843     if (simulationWork.useGpuBufferOps)
 844     {
 845         GMX_ASSERT(simulationWork.useGpuNonbonded,
 846                    "Can only offload buffer ops if nonbonded computation is also offloaded");
 847     }
 848     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 849     // on virial steps the CPU reduction path is taken
 850     flags.useGpuFBufferOps    = simulationWork.useGpuBufferOps && !flags.computeVirial;
 851     flags.useGpuPmeFReduction = flags.useGpuFBufferOps
 852                                 && (simulationWork.useGpuPme
 853                                     && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication));
 854
 855     return flags;
 856 }
 857
 858
 859 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
 860  *
 861  * TODO: eliminate \p useGpuPmeOnThisRank when this is
 862  * incorporated in DomainLifetimeWorkload.
 863  */
 864 static void launchGpuEndOfStepTasks(nonbonded_verlet_t*               nbv,
 865                                     gmx::GpuBonded*                   gpuBonded,
 866                                     gmx_pme_t*                        pmedata,
 867                                     gmx_enerdata_t*                   enerd,
 868                                     const gmx::MdrunScheduleWorkload& runScheduleWork,
 869                                     bool                              useGpuPmeOnThisRank,
 870                                     int64_t                           step,
 871                                     gmx_wallcycle_t                   wcycle)
 872 {
 873     if (runScheduleWork.simulationWork.useGpuNonbonded)
 874     {
 875         /* Launch pruning before buffer clearing because the API overhead of the
 876          * clear kernel launches can leave the GPU idle while it could be running
 877          * the prune kernel.
 878          */
 879         if (nbv->isDynamicPruningStepGpu(step))
 880         {
 881             nbv->dispatchPruneKernelGpu(step);
 882         }
 883
 884         /* now clear the GPU outputs while we finish the step on the CPU */
 885         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 886         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 887         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
 888         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 889         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 890     }
 891
 892     if (useGpuPmeOnThisRank)
 893     {
 894         pme_gpu_reinit_computation(pmedata, wcycle);
 895     }
 896
 897     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
 898     {
 899         // in principle this should be included in the DD balancing region,
 900         // but generally it is infrequent so we'll omit it for the sake of
 901         // simpler code
 902         gpuBonded->waitAccumulateEnergyTerms(enerd);
 903
 904         gpuBonded->clearEnergies();
 905     }
 906 }
 907
 908
 909 void do_force(FILE*                               fplog,
 910               const t_commrec*                    cr,
 911               const gmx_multisim_t*               ms,
 912               const t_inputrec*                   inputrec,
 913               gmx::Awh*                           awh,
 914               gmx_enfrot*                         enforcedRotation,
 915               gmx::ImdSession*                    imdSession,
 916               pull_t*                             pull_work,
 917               int64_t                             step,
 918               t_nrnb*                             nrnb,
 919               gmx_wallcycle_t                     wcycle,
 920               const gmx_localtop_t*               top,
 921               const matrix                        box,
 922               gmx::ArrayRefWithPadding<gmx::RVec> x,
 923               history_t*                          hist,
 924               gmx::ArrayRefWithPadding<gmx::RVec> force,
 925               tensor                              vir_force,
 926               const t_mdatoms*                    mdatoms,
 927               gmx_enerdata_t*                     enerd,
 928               t_fcdata*                           fcd,
 929               gmx::ArrayRef<real>                 lambda,
 930               t_graph*                            graph,
 931               t_forcerec*                         fr,
 932               gmx::MdrunScheduleWorkload*         runScheduleWork,
 933               const gmx_vsite_t*                  vsite,
 934               rvec                                mu_tot,
 935               double                              t,
 936               gmx_edsam*                          ed,
 937               int                                 legacyFlags,
 938               const DDBalanceRegionHandler&       ddBalanceRegionHandler)
 939 {
 940     int                          i, j;
 941     double                       mu[2 * DIM];
 942     nonbonded_verlet_t*          nbv      = fr->nbv.get();
 943     interaction_const_t*         ic       = fr->ic;
 944     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
 945
 946     const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
 947
 948
 949     runScheduleWork->stepWork    = setupStepWorkload(legacyFlags, fr->bNonbonded, simulationWork,
 950                                                   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     /* At a search step we need to start the first balancing region
 958      * somewhere early inside the step after communication during domain
 959      * decomposition (and not during the previous step as usual).
 960      */
 961     if (stepWork.doNeighborSearch)
 962     {
 963         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 964     }
 965
 966     const int start  = 0;
 967     const int homenr = mdatoms->homenr;
 968
 969     clear_mat(vir_force);
 970
 971     if (stepWork.stateChanged && simulationWork.computeMuTot)
 972     {
 973         /* Calculate total (local) dipole moment in a temporary common array.
 974          * This makes it possible to sum them over nodes faster.
 975          */
 976         calc_mu(start, homenr, x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB,
 977                 mdatoms->nChargePerturbed, mu, mu + DIM);
 978     }
 979
 980     if (fr->pbcType != PbcType::No)
 981     {
 982         /* Compute shift vectors every step,
 983          * because of pressure coupling or box deformation!
 984          */
 985         if (stepWork.haveDynamicBox && stepWork.stateChanged)
 986         {
 987             calc_shifts(box, fr->shift_vec);
 988         }
 989
 990         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
 991         const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
 992         if (calcCGCM)
 993         {
 994             put_atoms_in_box_omp(fr->pbcType, box, x.unpaddedArrayRef().subArray(0, homenr),
 995                                  gmx_omp_nthreads_get(emntDefault));
 996             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 997         }
 998         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 999         {
1000             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
1001         }
1002     }
1003
1004     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1005
1006 #if GMX_MPI
1007     const bool pmeSendCoordinatesFromGpu =
1008             simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1009     const bool reinitGpuPmePpComms =
1010             simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1011 #endif
1012
1013     const auto localXReadyOnDevice = (stateGpu != nullptr)
1014                                              ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1015                                                        AtomLocality::Local, simulationWork, stepWork)
1016                                              : nullptr;
1017
1018 #if GMX_MPI
1019     // If coordinates are to be sent to PME task from CPU memory, perform that send here.
1020     // Otherwise the send will occur after H2D coordinate transfer.
1021     if (!thisRankHasDuty(cr, DUTY_PME) && !pmeSendCoordinatesFromGpu)
1022     {
1023         /* Send particle coordinates to the pme nodes.
1024          * Since this is only implemented for domain decomposition
1025          * and domain decomposition does not use the graph,
1026          * we do not need to worry about shifting.
1027          */
1028         if (!stepWork.doNeighborSearch && simulationWork.useGpuUpdate)
1029         {
1030             GMX_RELEASE_ASSERT(false,
1031                                "GPU update and separate PME ranks are only supported with GPU "
1032                                "direct communication!");
1033             // TODO: when this code-path becomes supported add:
1034             // stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1035         }
1036
1037         gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1038                                  lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1039                                  step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1040                                  pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1041     }
1042 #endif /* GMX_MPI */
1043
1044     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1045     // The local coordinates can be copied right away.
1046     // NOTE: Consider moving this copy to right after they are updated and constrained,
1047     //       if the later is not offloaded.
1048     if (useGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
1049     {
1050         if (stepWork.doNeighborSearch)
1051         {
1052             // TODO refactor this to do_md, after partitioning.
1053             stateGpu->reinit(mdatoms->homenr,
1054                              cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1055             if (useGpuPmeOnThisRank)
1056             {
1057                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1058                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1059             }
1060         }
1061         // We need to copy coordinates when:
1062         // 1. Update is not offloaded
1063         // 2. The buffers were reinitialized on search step
1064         if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1065         {
1066             GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1067             stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1068         }
1069     }
1070
1071     // TODO Update this comment when introducing SimulationWorkload
1072     //
1073     // The conditions for gpuHaloExchange e.g. using GPU buffer
1074     // operations were checked before construction, so here we can
1075     // just use it and assert upon any conditions.
1076     const bool ddUsesGpuDirectCommunication =
1077             ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange.empty()));
1078     GMX_ASSERT(!ddUsesGpuDirectCommunication || stepWork.useGpuXBufferOps,
1079                "Must use coordinate buffer ops with GPU halo exchange");
1080     const bool useGpuForcesHaloExchange = ddUsesGpuDirectCommunication && stepWork.useGpuFBufferOps;
1081
1082     // Copy coordinate from the GPU if update is on the GPU and there
1083     // are forces to be computed on the CPU, or for the computation of
1084     // virial, or if host-side data will be transferred from this task
1085     // to a remote task for halo exchange or PME-PP communication. At
1086     // search steps the current coordinates are already on the host,
1087     // hence copy is not needed.
1088     const bool haveHostPmePpComms =
1089             !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication;
1090     const bool haveHostHaloExchangeComms = havePPDomainDecomposition(cr) && !ddUsesGpuDirectCommunication;
1091
1092     bool gmx_used_in_debug haveCopiedXFromGpu = false;
1093     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1094         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1095             || haveHostPmePpComms || haveHostHaloExchangeComms))
1096     {
1097         GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1098         stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1099         haveCopiedXFromGpu = true;
1100     }
1101
1102 #if GMX_MPI
1103     // If coordinates are to be sent to PME task from GPU memory, perform that send here.
1104     // Otherwise the send will occur before the H2D coordinate transfer.
1105     if (pmeSendCoordinatesFromGpu)
1106     {
1107         /* Send particle coordinates to the pme nodes.
1108          * Since this is only implemented for domain decomposition
1109          * and domain decomposition does not use the graph,
1110          * we do not need to worry about shifting.
1111          */
1112         gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1113                                  lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1114                                  step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1115                                  pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1116     }
1117 #endif /* GMX_MPI */
1118
1119     if (useGpuPmeOnThisRank)
1120     {
1121         launchPmeGpuSpread(fr->pmedata, box, stepWork, pmeFlags, localXReadyOnDevice, wcycle);
1122     }
1123
1124     /* do gridding for pair search */
1125     if (stepWork.doNeighborSearch)
1126     {
1127         if (graph && stepWork.stateChanged)
1128         {
1129             /* Calculate intramolecular shift vectors to make molecules whole */
1130             mk_mshift(fplog, graph, fr->pbcType, box, as_rvec_array(x.unpaddedArrayRef().data()));
1131         }
1132
1133         // TODO
1134         // - vzero is constant, do we need to pass it?
1135         // - box_diag should be passed directly to nbnxn_put_on_grid
1136         //
1137         rvec vzero;
1138         clear_rvec(vzero);
1139
1140         rvec box_diag;
1141         box_diag[XX] = box[XX][XX];
1142         box_diag[YY] = box[YY][YY];
1143         box_diag[ZZ] = box[ZZ][ZZ];
1144
1145         wallcycle_start(wcycle, ewcNS);
1146         if (!DOMAINDECOMP(cr))
1147         {
1148             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1149             nbnxn_put_on_grid(nbv, box, 0, vzero, box_diag, nullptr, { 0, mdatoms->homenr }, -1,
1150                               fr->cginfo, x.unpaddedArrayRef(), 0, nullptr);
1151             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1152         }
1153         else
1154         {
1155             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1156             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->cginfo, x.unpaddedArrayRef());
1157             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1158         }
1159
1160         nbv->setAtomProperties(*mdatoms, fr->cginfo);
1161
1162         wallcycle_stop(wcycle, ewcNS);
1163
1164         /* initialize the GPU nbnxm atom data and bonded data structures */
1165         if (simulationWork.useGpuNonbonded)
1166         {
1167             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1168
1169             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1170             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1171             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1172
1173             if (fr->gpuBonded)
1174             {
1175                 /* Now we put all atoms on the grid, we can assign bonded
1176                  * interactions to the GPU, where the grid order is
1177                  * needed. Also the xq, f and fshift device buffers have
1178                  * been reallocated if needed, so the bonded code can
1179                  * learn about them. */
1180                 // TODO the xq, f, and fshift buffers are now shared
1181                 // resources, so they should be maintained by a
1182                 // higher-level object than the nb module.
1183                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(
1184                         nbv->getGridIndices(), top->idef, Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1185                         Nbnxm::gpu_get_f(nbv->gpu_nbv), Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1186             }
1187             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1188         }
1189
1190         // Need to run after the GPU-offload bonded interaction lists
1191         // are set up to be able to determine whether there is bonded work.
1192         runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1193                 *inputrec, *fr, pull_work, ed, top->idef, *fcd, *mdatoms, simulationWork, stepWork);
1194
1195         wallcycle_start_nocount(wcycle, ewcNS);
1196         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1197         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1198         nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1199
1200         nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::Local);
1201
1202         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1203         wallcycle_stop(wcycle, ewcNS);
1204
1205         if (stepWork.useGpuXBufferOps)
1206         {
1207             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1208         }
1209         // For force buffer ops, we use the below conditon rather than
1210         // useGpuFBufferOps to ensure that init is performed even if this
1211         // NS step is also a virial step (on which f buf ops are deactivated).
1212         if (simulationWork.useGpuBufferOps && simulationWork.useGpuNonbonded && (GMX_GPU == GMX_GPU_CUDA))
1213         {
1214             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1215             nbv->atomdata_init_add_nbat_f_to_f_gpu(stateGpu->fReducedOnDevice());
1216         }
1217     }
1218     else if (!EI_TPI(inputrec->eI))
1219     {
1220         if (stepWork.useGpuXBufferOps)
1221         {
1222             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1223             nbv->convertCoordinatesGpu(AtomLocality::Local, false, stateGpu->getCoordinates(),
1224                                        localXReadyOnDevice);
1225         }
1226         else
1227         {
1228             if (simulationWork.useGpuUpdate)
1229             {
1230                 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1231                 GMX_ASSERT(haveCopiedXFromGpu,
1232                            "a wait should only be triggered if copy has been scheduled");
1233                 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1234             }
1235             nbv->convertCoordinates(AtomLocality::Local, false, x.unpaddedArrayRef());
1236         }
1237     }
1238
1239     const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1240
1241     if (simulationWork.useGpuNonbonded)
1242     {
1243         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1244
1245         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1246
1247         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1248         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1249         if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1250         {
1251             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1252         }
1253         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1254         // with X buffer ops offloaded to the GPU on all but the search steps
1255
1256         // bonded work not split into separate local and non-local, so with DD
1257         // we can only launch the kernel after non-local coordinates have been received.
1258         if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1259         {
1260             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1261             fr->gpuBonded->launchKernel(fr, stepWork, box);
1262             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1263         }
1264
1265         /* launch local nonbonded work on GPU */
1266         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1267         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1268         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1269         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1270     }
1271
1272     if (useGpuPmeOnThisRank)
1273     {
1274         // In PME GPU and mixed mode we launch FFT / gather after the
1275         // X copy/transform to allow overlap as well as after the GPU NB
1276         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1277         // the nonbonded kernel.
1278         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1279     }
1280
1281     /* Communicate coordinates and sum dipole if necessary +
1282        do non-local pair search */
1283     if (havePPDomainDecomposition(cr))
1284     {
1285         if (stepWork.doNeighborSearch)
1286         {
1287             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1288             wallcycle_start_nocount(wcycle, ewcNS);
1289             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1290             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1291             nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1292
1293             nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::NonLocal);
1294             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1295             wallcycle_stop(wcycle, ewcNS);
1296             // TODO refactor this GPU halo exchange re-initialisation
1297             // to location in do_md where GPU halo exchange is
1298             // constructed at partitioning, after above stateGpu
1299             // re-initialization has similarly been refactored
1300             if (ddUsesGpuDirectCommunication)
1301             {
1302                 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1303             }
1304         }
1305         else
1306         {
1307             if (ddUsesGpuDirectCommunication)
1308             {
1309                 // The following must be called after local setCoordinates (which records an event
1310                 // when the coordinate data has been copied to the device).
1311                 communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1312
1313                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1314                 {
1315                     // non-local part of coordinate buffer must be copied back to host for CPU work
1316                     stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1317                 }
1318             }
1319             else
1320             {
1321                 // Note: GPU update + DD without direct communication is not supported,
1322                 // a waitCoordinatesReadyOnHost() should be issued if it will be.
1323                 GMX_ASSERT(!simulationWork.useGpuUpdate,
1324                            "GPU update is not supported with CPU halo exchange");
1325                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1326             }
1327
1328             if (stepWork.useGpuXBufferOps)
1329             {
1330                 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1331                 {
1332                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1333                 }
1334                 nbv->convertCoordinatesGpu(AtomLocality::NonLocal, false, stateGpu->getCoordinates(),
1335                                            stateGpu->getCoordinatesReadyOnDeviceEvent(
1336                                                    AtomLocality::NonLocal, simulationWork, stepWork));
1337             }
1338             else
1339             {
1340                 nbv->convertCoordinates(AtomLocality::NonLocal, false, x.unpaddedArrayRef());
1341             }
1342         }
1343
1344         if (simulationWork.useGpuNonbonded)
1345         {
1346             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1347
1348             if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1349             {
1350                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1351                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1352                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1353             }
1354
1355             if (domainWork.haveGpuBondedWork)
1356             {
1357                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1358                 fr->gpuBonded->launchKernel(fr, stepWork, box);
1359                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1360             }
1361
1362             /* launch non-local nonbonded tasks on GPU */
1363             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1364             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1365                          nrnb, wcycle);
1366             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1367
1368             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1369         }
1370     }
1371
1372     if (simulationWork.useGpuNonbonded)
1373     {
1374         /* launch D2H copy-back F */
1375         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1376         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1377
1378         if (havePPDomainDecomposition(cr))
1379         {
1380             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1381         }
1382         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1383         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1384
1385         if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1386         {
1387             fr->gpuBonded->launchEnergyTransfer();
1388         }
1389         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1390     }
1391
1392     if (stepWork.stateChanged && simulationWork.computeMuTot)
1393     {
1394         if (PAR(cr))
1395         {
1396             gmx_sumd(2 * DIM, mu, cr);
1397
1398             ddBalanceRegionHandler.reopenRegionCpu();
1399         }
1400
1401         for (i = 0; i < 2; i++)
1402         {
1403             for (j = 0; j < DIM; j++)
1404             {
1405                 fr->mu_tot[i][j] = mu[i * DIM + j];
1406             }
1407         }
1408     }
1409     if (mdatoms->nChargePerturbed == 0)
1410     {
1411         copy_rvec(fr->mu_tot[0], mu_tot);
1412     }
1413     else
1414     {
1415         for (j = 0; j < DIM; j++)
1416         {
1417             mu_tot[j] = (1.0 - lambda[efptCOUL]) * fr->mu_tot[0][j] + lambda[efptCOUL] * fr->mu_tot[1][j];
1418         }
1419     }
1420
1421     /* Reset energies */
1422     reset_enerdata(enerd);
1423     /* Clear the shift forces */
1424     // TODO: This should be linked to the shift force buffer in use, or cleared before use instead
1425     for (gmx::RVec& elem : fr->shiftForces)
1426     {
1427         elem = { 0.0_real, 0.0_real, 0.0_real };
1428     }
1429
1430     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1431     {
1432         wallcycle_start(wcycle, ewcPPDURINGPME);
1433         dd_force_flop_start(cr->dd, nrnb);
1434     }
1435
1436     // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1437     // this wait ensures that the D2H transfer is complete.
1438     if ((simulationWork.useGpuUpdate)
1439         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial))
1440     {
1441         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1442     }
1443
1444     if (inputrec->bRot)
1445     {
1446         wallcycle_start(wcycle, ewcROT);
1447         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step,
1448                     stepWork.doNeighborSearch);
1449         wallcycle_stop(wcycle, ewcROT);
1450     }
1451
1452     /* Start the force cycle counter.
1453      * Note that a different counter is used for dynamic load balancing.
1454      */
1455     wallcycle_start(wcycle, ewcFORCE);
1456
1457     // Set up and clear force outputs.
1458     // We use std::move to keep the compiler happy, it has no effect.
1459     ForceOutputs forceOut =
1460             setupForceOutputs(fr, pull_work, *inputrec, std::move(force), stepWork, wcycle);
1461
1462     /* We calculate the non-bonded forces, when done on the CPU, here.
1463      * We do this before calling do_force_lowlevel, because in that
1464      * function, the listed forces are calculated before PME, which
1465      * does communication.  With this order, non-bonded and listed
1466      * force calculation imbalance can be balanced out by the domain
1467      * decomposition load balancing.
1468      */
1469
1470     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1471
1472     if (!useOrEmulateGpuNb)
1473     {
1474         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1475     }
1476
1477     if (fr->efep != efepNO)
1478     {
1479         /* Calculate the local and non-local free energy interactions here.
1480          * Happens here on the CPU both with and without GPU.
1481          */
1482         nbv->dispatchFreeEnergyKernel(InteractionLocality::Local, fr,
1483                                       as_rvec_array(x.unpaddedArrayRef().data()),
1484                                       &forceOut.forceWithShiftForces(), *mdatoms, inputrec->fepvals,
1485                                       lambda.data(), enerd, stepWork, nrnb);
1486
1487         if (havePPDomainDecomposition(cr))
1488         {
1489             nbv->dispatchFreeEnergyKernel(InteractionLocality::NonLocal, fr,
1490                                           as_rvec_array(x.unpaddedArrayRef().data()),
1491                                           &forceOut.forceWithShiftForces(), *mdatoms,
1492                                           inputrec->fepvals, lambda.data(), enerd, stepWork, nrnb);
1493         }
1494     }
1495
1496     if (!useOrEmulateGpuNb)
1497     {
1498         if (havePPDomainDecomposition(cr))
1499         {
1500             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1501                          nrnb, wcycle);
1502         }
1503
1504         if (stepWork.computeForces)
1505         {
1506             /* Add all the non-bonded force to the normal force array.
1507              * This can be split into a local and a non-local part when overlapping
1508              * communication with calculation with domain decomposition.
1509              */
1510             wallcycle_stop(wcycle, ewcFORCE);
1511             nbv->atomdata_add_nbat_f_to_f(AtomLocality::All, forceOut.forceWithShiftForces().force());
1512             wallcycle_start_nocount(wcycle, ewcFORCE);
1513         }
1514
1515         /* If there are multiple fshift output buffers we need to reduce them */
1516         if (stepWork.computeVirial)
1517         {
1518             /* This is not in a subcounter because it takes a
1519                negligible and constant-sized amount of time */
1520             nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1521                                                      forceOut.forceWithShiftForces().shiftForces());
1522         }
1523     }
1524
1525     /* update QMMMrec, if necessary */
1526     if (fr->bQMMM)
1527     {
1528         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1529     }
1530
1531     // TODO Force flags should include haveFreeEnergyWork for this domain
1532     if (ddUsesGpuDirectCommunication && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1533     {
1534         /* Wait for non-local coordinate data to be copied from device */
1535         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1536     }
1537     /* Compute the bonded and non-bonded energies and optionally forces */
1538     do_force_lowlevel(fr, inputrec, top->idef, cr, ms, nrnb, wcycle, mdatoms, x, hist, &forceOut, enerd,
1539                       fcd, box, lambda.data(), graph, fr->mu_tot, stepWork, ddBalanceRegionHandler);
1540
1541     wallcycle_stop(wcycle, ewcFORCE);
1542
1543     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation, imdSession, pull_work, step, t,
1544                          wcycle, fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1545                          stepWork, &forceOut.forceWithVirial(), enerd, ed, stepWork.doNeighborSearch);
1546
1547
1548     // Will store the amount of cycles spent waiting for the GPU that
1549     // will be later used in the DLB accounting.
1550     float cycles_wait_gpu = 0;
1551     if (useOrEmulateGpuNb)
1552     {
1553         auto& forceWithShiftForces = forceOut.forceWithShiftForces();
1554
1555         /* wait for non-local forces (or calculate in emulation mode) */
1556         if (havePPDomainDecomposition(cr))
1557         {
1558             if (simulationWork.useGpuNonbonded)
1559             {
1560                 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
1561                         nbv->gpu_nbv, stepWork, AtomLocality::NonLocal, enerd->grpp.ener[egLJSR].data(),
1562                         enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(), wcycle);
1563             }
1564             else
1565             {
1566                 wallcycle_start_nocount(wcycle, ewcFORCE);
1567                 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes,
1568                              step, nrnb, wcycle);
1569                 wallcycle_stop(wcycle, ewcFORCE);
1570             }
1571
1572             if (stepWork.useGpuFBufferOps)
1573             {
1574                 gmx::FixedCapacityVector<GpuEventSynchronizer*, 1> dependencyList;
1575
1576                 // TODO: move this into DomainLifetimeWorkload, including the second part of the
1577                 // condition The bonded and free energy CPU tasks can have non-local force
1578                 // contributions which are a dependency for the GPU force reduction.
1579                 bool haveNonLocalForceContribInCpuBuffer =
1580                         domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1581
1582                 if (haveNonLocalForceContribInCpuBuffer)
1583                 {
1584                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(),
1585                                               AtomLocality::NonLocal);
1586                     dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(
1587                             AtomLocality::NonLocal, stepWork.useGpuFBufferOps));
1588                 }
1589
1590                 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::NonLocal, stateGpu->getForces(),
1591                                                   pme_gpu_get_device_f(fr->pmedata), dependencyList,
1592                                                   false, haveNonLocalForceContribInCpuBuffer);
1593                 if (!useGpuForcesHaloExchange)
1594                 {
1595                     // copy from GPU input for dd_move_f()
1596                     stateGpu->copyForcesFromGpu(forceOut.forceWithShiftForces().force(),
1597                                                 AtomLocality::NonLocal);
1598                 }
1599             }
1600             else
1601             {
1602                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
1603             }
1604
1605
1606             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1607             {
1608                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
1609             }
1610         }
1611     }
1612
1613     if (havePPDomainDecomposition(cr))
1614     {
1615         /* We are done with the CPU compute.
1616          * We will now communicate the non-local forces.
1617          * If we use a GPU this will overlap with GPU work, so in that case
1618          * we do not close the DD force balancing region here.
1619          */
1620         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1621
1622         if (stepWork.computeForces)
1623         {
1624
1625             if (useGpuForcesHaloExchange)
1626             {
1627                 if (domainWork.haveCpuLocalForceWork)
1628                 {
1629                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), AtomLocality::Local);
1630                 }
1631                 communicateGpuHaloForces(*cr, domainWork.haveCpuLocalForceWork);
1632             }
1633             else
1634             {
1635                 if (stepWork.useGpuFBufferOps)
1636                 {
1637                     stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
1638                 }
1639                 dd_move_f(cr->dd, &forceOut.forceWithShiftForces(), wcycle);
1640             }
1641         }
1642     }
1643
1644     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1645     // an alternating wait/reduction scheme.
1646     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded
1647                              && !DOMAINDECOMP(cr) && !stepWork.useGpuFBufferOps);
1648     if (alternateGpuWait)
1649     {
1650         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &forceOut, enerd, stepWork, pmeFlags, wcycle);
1651     }
1652
1653     if (!alternateGpuWait && useGpuPmeOnThisRank)
1654     {
1655         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial(), enerd);
1656     }
1657
1658     /* Wait for local GPU NB outputs on the non-alternating wait path */
1659     if (!alternateGpuWait && simulationWork.useGpuNonbonded)
1660     {
1661         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1662          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1663          * but even with a step of 0.1 ms the difference is less than 1%
1664          * of the step time.
1665          */
1666         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1667         const float waitCycles               = Nbnxm::gpu_wait_finish_task(
1668                 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
1669                 enerd->grpp.ener[egCOULSR].data(), forceOut.forceWithShiftForces().shiftForces(), wcycle);
1670
1671         if (ddBalanceRegionHandler.useBalancingRegion())
1672         {
1673             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1674             if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
1675             {
1676                 /* We measured few cycles, it could be that the kernel
1677                  * and transfer finished earlier and there was no actual
1678                  * wait time, only API call overhead.
1679                  * Then the actual time could be anywhere between 0 and
1680                  * cycles_wait_est. We will use half of cycles_wait_est.
1681                  */
1682                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1683             }
1684             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1685         }
1686     }
1687
1688     if (fr->nbv->emulateGpu())
1689     {
1690         // NOTE: emulation kernel is not included in the balancing region,
1691         // but emulation mode does not target performance anyway
1692         wallcycle_start_nocount(wcycle, ewcFORCE);
1693         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local,
1694                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes, step, nrnb, wcycle);
1695         wallcycle_stop(wcycle, ewcFORCE);
1696     }
1697
1698     // If on GPU PME-PP comms or GPU update path, receive forces from PME before GPU buffer ops
1699     // TODO refactor this and unify with below default-path call to the same function
1700     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME)
1701         && (simulationWork.useGpuPmePpCommunication || simulationWork.useGpuUpdate))
1702     {
1703         /* In case of node-splitting, the PP nodes receive the long-range
1704          * forces, virial and energy from the PME nodes here.
1705          */
1706         pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1707                                simulationWork.useGpuPmePpCommunication,
1708                                stepWork.useGpuPmeFReduction, wcycle);
1709     }
1710
1711
1712     /* Do the nonbonded GPU (or emulation) force buffer reduction
1713      * on the non-alternating path. */
1714     if (useOrEmulateGpuNb && !alternateGpuWait)
1715     {
1716         // TODO simplify the below conditionals. Pass buffer and sync pointers at init stage rather than here. Unify getter fns for sameGPU/otherGPU cases.
1717         void* pmeForcePtr =
1718                 stepWork.useGpuPmeFReduction
1719                         ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_device_f(fr->pmedata)
1720                                                          : // PME force buffer on same GPU
1721                                    fr->pmePpCommGpu->getGpuForceStagingPtr()) // buffer received from other GPU
1722                         : nullptr; // PME reduction not active on GPU
1723
1724         GpuEventSynchronizer* const pmeSynchronizer =
1725                 stepWork.useGpuPmeFReduction
1726                         ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_f_ready_synchronizer(fr->pmedata)
1727                                                          : // PME force buffer on same GPU
1728                                    static_cast<GpuEventSynchronizer*>(
1729                                            fr->pmePpCommGpu->getForcesReadySynchronizer())) // buffer received from other GPU
1730                         : nullptr; // PME reduction not active on GPU
1731
1732         gmx::FixedCapacityVector<GpuEventSynchronizer*, 3> dependencyList;
1733
1734         if (stepWork.useGpuPmeFReduction)
1735         {
1736             dependencyList.push_back(pmeSynchronizer);
1737         }
1738
1739         gmx::ArrayRef<gmx::RVec> forceWithShift = forceOut.forceWithShiftForces().force();
1740
1741         if (stepWork.useGpuFBufferOps)
1742         {
1743             // Flag to specify whether the CPU force buffer has contributions to
1744             // local atoms. This depends on whether there are CPU-based force tasks
1745             // or when DD is active the halo exchange has resulted in contributions
1746             // from the non-local part.
1747             const bool haveLocalForceContribInCpuBuffer =
1748                     (domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr));
1749
1750             // TODO: move these steps as early as possible:
1751             // - CPU f H2D should be as soon as all CPU-side forces are done
1752             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1753             //   before the next CPU task that consumes the forces: vsite spread or update)
1754             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
1755             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
1756             //   These should be unified.
1757             if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
1758             {
1759                 // Note: AtomLocality::All is used for the non-DD case because, as in this
1760                 // case copyForcesToGpu() uses a separate stream, it allows overlap of
1761                 // CPU force H2D with GPU force tasks on all streams including those in the
1762                 // local stream which would otherwise be implicit dependencies for the
1763                 // transfer and would not overlap.
1764                 auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
1765
1766                 stateGpu->copyForcesToGpu(forceWithShift, locality);
1767                 dependencyList.push_back(
1768                         stateGpu->getForcesReadyOnDeviceEvent(locality, stepWork.useGpuFBufferOps));
1769             }
1770             if (useGpuForcesHaloExchange)
1771             {
1772                 dependencyList.push_back(cr->dd->gpuHaloExchange[0]->getForcesReadyOnDeviceEvent());
1773             }
1774             nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::Local, stateGpu->getForces(), pmeForcePtr,
1775                                               dependencyList, stepWork.useGpuPmeFReduction,
1776                                               haveLocalForceContribInCpuBuffer);
1777             // Copy forces to host if they are needed for update or if virtual sites are enabled.
1778             // If there are vsites, we need to copy forces every step to spread vsite forces on host.
1779             // TODO: When the output flags will be included in step workload, this copy can be combined with the
1780             //       copy call done in sim_utils(...) for the output.
1781             // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
1782             //       they should not be copied in do_md(...) for the output.
1783             if (!simulationWork.useGpuUpdate || vsite)
1784             {
1785                 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
1786                 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
1787             }
1788         }
1789         else
1790         {
1791             nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
1792         }
1793     }
1794
1795     launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd, *runScheduleWork,
1796                             useGpuPmeOnThisRank, step, wcycle);
1797
1798     if (DOMAINDECOMP(cr))
1799     {
1800         dd_force_flop_stop(cr->dd, nrnb);
1801     }
1802
1803     if (stepWork.computeForces)
1804     {
1805         rvec* f = as_rvec_array(forceOut.forceWithShiftForces().force().data());
1806
1807         /* If we have NoVirSum forces, but we do not calculate the virial,
1808          * we sum fr->f_novirsum=forceOut.f later.
1809          */
1810         if (vsite && !(fr->haveDirectVirialContributions && !stepWork.computeVirial))
1811         {
1812             rvec* fshift = as_rvec_array(forceOut.forceWithShiftForces().shiftForces().data());
1813             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), f, fshift, FALSE,
1814                            nullptr, nrnb, top->idef, fr->pbcType, fr->bMolPBC, graph, box, cr, wcycle);
1815         }
1816
1817         if (stepWork.computeVirial)
1818         {
1819             /* Calculation of the virial must be done after vsites! */
1820             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()),
1821                         forceOut.forceWithShiftForces(), vir_force, graph, box, nrnb, fr,
1822                         inputrec->pbcType);
1823         }
1824     }
1825
1826     // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
1827     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication
1828         && !simulationWork.useGpuUpdate)
1829     {
1830         /* In case of node-splitting, the PP nodes receive the long-range
1831          * forces, virial and energy from the PME nodes here.
1832          */
1833         pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1834                                simulationWork.useGpuPmePpCommunication, false, wcycle);
1835     }
1836
1837     if (stepWork.computeForces)
1838     {
1839         post_process_forces(cr, step, nrnb, wcycle, top, box, as_rvec_array(x.unpaddedArrayRef().data()),
1840                             &forceOut, vir_force, mdatoms, graph, fr, vsite, stepWork);
1841     }
1842
1843     if (stepWork.computeEnergy)
1844     {
1845         /* Sum the potential energy terms from group contributions */
1846         sum_epot(&(enerd->grpp), enerd->term);
1847
1848         if (!EI_TPI(inputrec->eI))
1849         {
1850             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1851         }
1852     }
1853
1854     /* In case we don't have constraints and are using GPUs, the next balancing
1855      * region starts here.
1856      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1857      * virial calculation and COM pulling, is not thus not included in
1858      * the balance timing, which is ok as most tasks do communication.
1859      */
1860     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1861 }