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