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