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) 2013-2019,2020,2021, 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 #include <optional>
  48
  49 #include "gromacs/applied_forces/awh/awh.h"
  50 #include "gromacs/domdec/dlbtiming.h"
  51 #include "gromacs/domdec/domdec.h"
  52 #include "gromacs/domdec/domdec_struct.h"
  53 #include "gromacs/domdec/gpuhaloexchange.h"
  54 #include "gromacs/domdec/partition.h"
  55 #include "gromacs/essentialdynamics/edsam.h"
  56 #include "gromacs/ewald/pme.h"
  57 #include "gromacs/ewald/pme_pp.h"
  58 #include "gromacs/ewald/pme_pp_comm_gpu.h"
  59 #include "gromacs/gmxlib/network.h"
  60 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  61 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  62 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  63 #include "gromacs/gpu_utils/gpu_utils.h"
  64 #include "gromacs/imd/imd.h"
  65 #include "gromacs/listed_forces/disre.h"
  66 #include "gromacs/listed_forces/gpubonded.h"
  67 #include "gromacs/listed_forces/listed_forces.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/dispersioncorrection.h"
  78 #include "gromacs/mdlib/enerdata_utils.h"
  79 #include "gromacs/mdlib/force.h"
  80 #include "gromacs/mdlib/force_flags.h"
  81 #include "gromacs/mdlib/forcerec.h"
  82 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  83 #include "gromacs/mdlib/update.h"
  84 #include "gromacs/mdlib/vsite.h"
  85 #include "gromacs/mdlib/wall.h"
  86 #include "gromacs/mdlib/wholemoleculetransform.h"
  87 #include "gromacs/mdtypes/commrec.h"
  88 #include "gromacs/mdtypes/enerdata.h"
  89 #include "gromacs/mdtypes/forcebuffers.h"
  90 #include "gromacs/mdtypes/forceoutput.h"
  91 #include "gromacs/mdtypes/forcerec.h"
  92 #include "gromacs/mdtypes/iforceprovider.h"
  93 #include "gromacs/mdtypes/inputrec.h"
  94 #include "gromacs/mdtypes/md_enums.h"
  95 #include "gromacs/mdtypes/mdatom.h"
  96 #include "gromacs/mdtypes/multipletimestepping.h"
  97 #include "gromacs/mdtypes/simulation_workload.h"
  98 #include "gromacs/mdtypes/state.h"
  99 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
 100 #include "gromacs/nbnxm/gpu_data_mgmt.h"
 101 #include "gromacs/nbnxm/nbnxm.h"
 102 #include "gromacs/nbnxm/nbnxm_gpu.h"
 103 #include "gromacs/pbcutil/ishift.h"
 104 #include "gromacs/pbcutil/pbc.h"
 105 #include "gromacs/pulling/pull.h"
 106 #include "gromacs/pulling/pull_rotation.h"
 107 #include "gromacs/timing/cyclecounter.h"
 108 #include "gromacs/timing/gpu_timing.h"
 109 #include "gromacs/timing/wallcycle.h"
 110 #include "gromacs/timing/wallcyclereporting.h"
 111 #include "gromacs/timing/walltime_accounting.h"
 112 #include "gromacs/topology/topology.h"
 113 #include "gromacs/utility/arrayref.h"
 114 #include "gromacs/utility/basedefinitions.h"
 115 #include "gromacs/utility/cstringutil.h"
 116 #include "gromacs/utility/exceptions.h"
 117 #include "gromacs/utility/fatalerror.h"
 118 #include "gromacs/utility/fixedcapacityvector.h"
 119 #include "gromacs/utility/gmxassert.h"
 120 #include "gromacs/utility/gmxmpi.h"
 121 #include "gromacs/utility/logger.h"
 122 #include "gromacs/utility/smalloc.h"
 123 #include "gromacs/utility/strconvert.h"
 124 #include "gromacs/utility/sysinfo.h"
 125
 126 #include "gpuforcereduction.h"
 127
 128 using gmx::ArrayRef;
 129 using gmx::AtomLocality;
 130 using gmx::DomainLifetimeWorkload;
 131 using gmx::ForceOutputs;
 132 using gmx::ForceWithShiftForces;
 133 using gmx::InteractionLocality;
 134 using gmx::RVec;
 135 using gmx::SimulationWorkload;
 136 using gmx::StepWorkload;
 137
 138 // TODO: this environment variable allows us to verify before release
 139 // that on less common architectures the total cost of polling is not larger than
 140 // a blocking wait (so polling does not introduce overhead when the static
 141 // PME-first ordering would suffice).
 142 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 143
 144 static void sum_forces(ArrayRef<RVec> f, ArrayRef<const RVec> forceToAdd)
 145 {
 146     GMX_ASSERT(f.size() >= forceToAdd.size(), "Accumulation buffer should be sufficiently large");
 147     const int end = forceToAdd.size();
 148
 149     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 150 #pragma omp parallel for num_threads(nt) schedule(static)
 151     for (int i = 0; i < end; i++)
 152     {
 153         rvec_inc(f[i], forceToAdd[i]);
 154     }
 155 }
 156
 157 static void calc_virial(int                              start,
 158                         int                              homenr,
 159                         const rvec                       x[],
 160                         const gmx::ForceWithShiftForces& forceWithShiftForces,
 161                         tensor                           vir_part,
 162                         const matrix                     box,
 163                         t_nrnb*                          nrnb,
 164                         const t_forcerec*                fr,
 165                         PbcType                          pbcType)
 166 {
 167     /* The short-range virial from surrounding boxes */
 168     const rvec* fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
 169     calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, pbcType == PbcType::Screw, box);
 170     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 171
 172     /* Calculate partial virial, for local atoms only, based on short range.
 173      * Total virial is computed in global_stat, called from do_md
 174      */
 175     const rvec* f = as_rvec_array(forceWithShiftForces.force().data());
 176     f_calc_vir(start, start + homenr, x, f, vir_part, box);
 177     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 178
 179     if (debug)
 180     {
 181         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 182     }
 183 }
 184
 185 static void pull_potential_wrapper(const t_commrec*               cr,
 186                                    const t_inputrec*              ir,
 187                                    const matrix                   box,
 188                                    gmx::ArrayRef<const gmx::RVec> x,
 189                                    gmx::ForceWithVirial*          force,
 190                                    const t_mdatoms*               mdatoms,
 191                                    gmx_enerdata_t*                enerd,
 192                                    pull_t*                        pull_work,
 193                                    const real*                    lambda,
 194                                    double                         t,
 195                                    gmx_wallcycle_t                wcycle)
 196 {
 197     t_pbc pbc;
 198     real  dvdl;
 199
 200     /* Calculate the center of mass forces, this requires communication,
 201      * which is why pull_potential is called close to other communication.
 202      */
 203     wallcycle_start(wcycle, ewcPULLPOT);
 204     set_pbc(&pbc, ir->pbcType, box);
 205     dvdl = 0;
 206     enerd->term[F_COM_PULL] += pull_potential(
 207             pull_work, mdatoms->massT, &pbc, cr, t, lambda[efptRESTRAINT], as_rvec_array(x.data()), force, &dvdl);
 208     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 209     wallcycle_stop(wcycle, ewcPULLPOT);
 210 }
 211
 212 static void pme_receive_force_ener(t_forcerec*           fr,
 213                                    const t_commrec*      cr,
 214                                    gmx::ForceWithVirial* forceWithVirial,
 215                                    gmx_enerdata_t*       enerd,
 216                                    bool                  useGpuPmePpComms,
 217                                    bool                  receivePmeForceToGpu,
 218                                    gmx_wallcycle_t       wcycle)
 219 {
 220     real  e_q, e_lj, dvdl_q, dvdl_lj;
 221     float cycles_ppdpme, cycles_seppme;
 222
 223     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 224     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 225
 226     /* In case of node-splitting, the PP nodes receive the long-range
 227      * forces, virial and energy from the PME nodes here.
 228      */
 229     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 230     dvdl_q  = 0;
 231     dvdl_lj = 0;
 232     gmx_pme_receive_f(fr->pmePpCommGpu.get(),
 233                       cr,
 234                       forceWithVirial,
 235                       &e_q,
 236                       &e_lj,
 237                       &dvdl_q,
 238                       &dvdl_lj,
 239                       useGpuPmePpComms,
 240                       receivePmeForceToGpu,
 241                       &cycles_seppme);
 242     enerd->term[F_COUL_RECIP] += e_q;
 243     enerd->term[F_LJ_RECIP] += e_lj;
 244     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 245     enerd->dvdl_lin[efptVDW] += dvdl_lj;
 246
 247     if (wcycle)
 248     {
 249         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 250     }
 251     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 252 }
 253
 254 static void print_large_forces(FILE*                fp,
 255                                const t_mdatoms*     md,
 256                                const t_commrec*     cr,
 257                                int64_t              step,
 258                                real                 forceTolerance,
 259                                ArrayRef<const RVec> x,
 260                                ArrayRef<const RVec> f)
 261 {
 262     real       force2Tolerance = gmx::square(forceTolerance);
 263     gmx::index numNonFinite    = 0;
 264     for (int i = 0; i < md->homenr; i++)
 265     {
 266         real force2    = norm2(f[i]);
 267         bool nonFinite = !std::isfinite(force2);
 268         if (force2 >= force2Tolerance || nonFinite)
 269         {
 270             fprintf(fp,
 271                     "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 272                     step,
 273                     ddglatnr(cr->dd, i),
 274                     x[i][XX],
 275                     x[i][YY],
 276                     x[i][ZZ],
 277                     std::sqrt(force2));
 278         }
 279         if (nonFinite)
 280         {
 281             numNonFinite++;
 282         }
 283     }
 284     if (numNonFinite > 0)
 285     {
 286         /* Note that with MPI this fatal call on one rank might interrupt
 287          * the printing on other ranks. But we can only avoid that with
 288          * an expensive MPI barrier that we would need at each step.
 289          */
 290         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 291     }
 292 }
 293
 294 //! When necessary, spreads forces on vsites and computes the virial for \p forceOutputs->forceWithShiftForces()
 295 static void postProcessForceWithShiftForces(t_nrnb*                   nrnb,
 296                                             gmx_wallcycle_t           wcycle,
 297                                             const matrix              box,
 298                                             ArrayRef<const RVec>      x,
 299                                             ForceOutputs*             forceOutputs,
 300                                             tensor                    vir_force,
 301                                             const t_mdatoms&          mdatoms,
 302                                             const t_forcerec&         fr,
 303                                             gmx::VirtualSitesHandler* vsite,
 304                                             const StepWorkload&       stepWork)
 305 {
 306     ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
 307
 308     /* If we have NoVirSum forces, but we do not calculate the virial,
 309      * we later sum the forceWithShiftForces buffer together with
 310      * the noVirSum buffer and spread the combined vsite forces at once.
 311      */
 312     if (vsite && (!forceOutputs->haveForceWithVirial() || stepWork.computeVirial))
 313     {
 314         using VirialHandling = gmx::VirtualSitesHandler::VirialHandling;
 315
 316         auto                 f      = forceWithShiftForces.force();
 317         auto                 fshift = forceWithShiftForces.shiftForces();
 318         const VirialHandling virialHandling =
 319                 (stepWork.computeVirial ? VirialHandling::Pbc : VirialHandling::None);
 320         vsite->spreadForces(x, f, virialHandling, fshift, nullptr, nrnb, box, wcycle);
 321         forceWithShiftForces.haveSpreadVsiteForces() = true;
 322     }
 323
 324     if (stepWork.computeVirial)
 325     {
 326         /* Calculation of the virial must be done after vsites! */
 327         calc_virial(
 328                 0, mdatoms.homenr, as_rvec_array(x.data()), forceWithShiftForces, vir_force, box, nrnb, &fr, fr.pbcType);
 329     }
 330 }
 331
 332 //! Spread, compute virial for and sum forces, when necessary
 333 static void postProcessForces(const t_commrec*          cr,
 334                               int64_t                   step,
 335                               t_nrnb*                   nrnb,
 336                               gmx_wallcycle_t           wcycle,
 337                               const matrix              box,
 338                               ArrayRef<const RVec>      x,
 339                               ForceOutputs*             forceOutputs,
 340                               tensor                    vir_force,
 341                               const t_mdatoms*          mdatoms,
 342                               const t_forcerec*         fr,
 343                               gmx::VirtualSitesHandler* vsite,
 344                               const StepWorkload&       stepWork)
 345 {
 346     // Extract the final output force buffer, which is also the buffer for forces with shift forces
 347     ArrayRef<RVec> f = forceOutputs->forceWithShiftForces().force();
 348
 349     if (forceOutputs->haveForceWithVirial())
 350     {
 351         auto& forceWithVirial = forceOutputs->forceWithVirial();
 352
 353         if (vsite)
 354         {
 355             /* Spread the mesh force on virtual sites to the other particles...
 356              * This is parallellized. MPI communication is performed
 357              * if the constructing atoms aren't local.
 358              */
 359             GMX_ASSERT(!stepWork.computeVirial || f.data() != forceWithVirial.force_.data(),
 360                        "We need separate force buffers for shift and virial forces when "
 361                        "computing the virial");
 362             GMX_ASSERT(!stepWork.computeVirial
 363                                || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
 364                        "We should spread the force with shift forces separately when computing "
 365                        "the virial");
 366             const gmx::VirtualSitesHandler::VirialHandling virialHandling =
 367                     (stepWork.computeVirial ? gmx::VirtualSitesHandler::VirialHandling::NonLinear
 368                                             : gmx::VirtualSitesHandler::VirialHandling::None);
 369             matrix virial = { { 0 } };
 370             vsite->spreadForces(x, forceWithVirial.force_, virialHandling, {}, virial, nrnb, box, wcycle);
 371             forceWithVirial.addVirialContribution(virial);
 372         }
 373
 374         if (stepWork.computeVirial)
 375         {
 376             /* Now add the forces, this is local */
 377             sum_forces(f, forceWithVirial.force_);
 378
 379             /* Add the direct virial contributions */
 380             GMX_ASSERT(
 381                     forceWithVirial.computeVirial_,
 382                     "forceWithVirial should request virial computation when we request the virial");
 383             m_add(vir_force, forceWithVirial.getVirial(), vir_force);
 384
 385             if (debug)
 386             {
 387                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 388             }
 389         }
 390     }
 391     else
 392     {
 393         GMX_ASSERT(vsite == nullptr || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
 394                    "We should have spread the vsite forces (earlier)");
 395     }
 396
 397     if (fr->print_force >= 0)
 398     {
 399         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 400     }
 401 }
 402
 403 static void do_nb_verlet(t_forcerec*                fr,
 404                          const interaction_const_t* ic,
 405                          gmx_enerdata_t*            enerd,
 406                          const StepWorkload&        stepWork,
 407                          const InteractionLocality  ilocality,
 408                          const int                  clearF,
 409                          const int64_t              step,
 410                          t_nrnb*                    nrnb,
 411                          gmx_wallcycle_t            wcycle)
 412 {
 413     if (!stepWork.computeNonbondedForces)
 414     {
 415         /* skip non-bonded calculation */
 416         return;
 417     }
 418
 419     nonbonded_verlet_t* nbv = fr->nbv.get();
 420
 421     /* GPU kernel launch overhead is already timed separately */
 422     if (!nbv->useGpu())
 423     {
 424         /* When dynamic pair-list  pruning is requested, we need to prune
 425          * at nstlistPrune steps.
 426          */
 427         if (nbv->isDynamicPruningStepCpu(step))
 428         {
 429             /* Prune the pair-list beyond fr->ic->rlistPrune using
 430              * the current coordinates of the atoms.
 431              */
 432             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 433             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 434             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 435         }
 436     }
 437
 438     nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
 439 }
 440
 441 static inline void clearRVecs(ArrayRef<RVec> v, const bool useOpenmpThreading)
 442 {
 443     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, v.ssize());
 444
 445     /* Note that we would like to avoid this conditional by putting it
 446      * into the omp pragma instead, but then we still take the full
 447      * omp parallel for overhead (at least with gcc5).
 448      */
 449     if (!useOpenmpThreading || nth == 1)
 450     {
 451         for (RVec& elem : v)
 452         {
 453             clear_rvec(elem);
 454         }
 455     }
 456     else
 457     {
 458 #pragma omp parallel for num_threads(nth) schedule(static)
 459         for (gmx::index i = 0; i < v.ssize(); i++)
 460         {
 461             clear_rvec(v[i]);
 462         }
 463     }
 464 }
 465
 466 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 467  *
 468  * \param groupOptions  Group options, containing T-coupling options
 469  */
 470 static real averageKineticEnergyEstimate(const t_grpopts& groupOptions)
 471 {
 472     real nrdfCoupled   = 0;
 473     real nrdfUncoupled = 0;
 474     real kineticEnergy = 0;
 475     for (int g = 0; g < groupOptions.ngtc; g++)
 476     {
 477         if (groupOptions.tau_t[g] >= 0)
 478         {
 479             nrdfCoupled += groupOptions.nrdf[g];
 480             kineticEnergy += groupOptions.nrdf[g] * 0.5 * groupOptions.ref_t[g] * BOLTZ;
 481         }
 482         else
 483         {
 484             nrdfUncoupled += groupOptions.nrdf[g];
 485         }
 486     }
 487
 488     /* This conditional with > also catches nrdf=0 */
 489     if (nrdfCoupled > nrdfUncoupled)
 490     {
 491         return kineticEnergy * (nrdfCoupled + nrdfUncoupled) / nrdfCoupled;
 492     }
 493     else
 494     {
 495         return 0;
 496     }
 497 }
 498
 499 /*! \brief This routine checks that the potential energy is finite.
 500  *
 501  * Always checks that the potential energy is finite. If step equals
 502  * inputrec.init_step also checks that the magnitude of the potential energy
 503  * is reasonable. Terminates with a fatal error when a check fails.
 504  * Note that passing this check does not guarantee finite forces,
 505  * since those use slightly different arithmetics. But in most cases
 506  * there is just a narrow coordinate range where forces are not finite
 507  * and energies are finite.
 508  *
 509  * \param[in] step      The step number, used for checking and printing
 510  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to
 511  * enerd.term[F_EPOT] before calling this routine \param[in] inputrec  The input record
 512  */
 513 static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
 514 {
 515     /* Threshold valid for comparing absolute potential energy against
 516      * the kinetic energy. Normally one should not consider absolute
 517      * potential energy values, but with a factor of one million
 518      * we should never get false positives.
 519      */
 520     constexpr real c_thresholdFactor = 1e6;
 521
 522     bool energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 523     real averageKineticEnergy = 0;
 524     /* We only check for large potential energy at the initial step,
 525      * because that is by far the most likely step for this too occur
 526      * and because computing the average kinetic energy is not free.
 527      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 528      * before they become NaN.
 529      */
 530     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 531     {
 532         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 533     }
 534
 535     if (energyIsNotFinite
 536         || (averageKineticEnergy > 0 && enerd.term[F_EPOT] > c_thresholdFactor * averageKineticEnergy))
 537     {
 538         gmx_fatal(
 539                 FARGS,
 540                 "Step %" PRId64
 541                 ": The total potential energy is %g, which is %s. The LJ and electrostatic "
 542                 "contributions to the energy are %g and %g, respectively. A %s potential energy "
 543                 "can be caused by overlapping interactions in bonded interactions or very large%s "
 544                 "coordinate values. Usually this is caused by a badly- or non-equilibrated initial "
 545                 "configuration, incorrect interactions or parameters in the topology.",
 546                 step,
 547                 enerd.term[F_EPOT],
 548                 energyIsNotFinite ? "not finite" : "extremely high",
 549                 enerd.term[F_LJ],
 550                 enerd.term[F_COUL_SR],
 551                 energyIsNotFinite ? "non-finite" : "very high",
 552                 energyIsNotFinite ? " or Nan" : "");
 553     }
 554 }
 555
 556 /*! \brief Return true if there are special forces computed this step.
 557  *
 558  * The conditionals exactly correspond to those in computeSpecialForces().
 559  */
 560 static bool haveSpecialForces(const t_inputrec&          inputrec,
 561                               const gmx::ForceProviders& forceProviders,
 562                               const pull_t*              pull_work,
 563                               const bool                 computeForces,
 564                               const gmx_edsam*           ed)
 565 {
 566
 567     return ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
 568             (inputrec.bPull && pull_have_potential(*pull_work)) ||  // pull
 569             inputrec.bRot ||                                        // enforced rotation
 570             (ed != nullptr) ||                                      // flooding
 571             (inputrec.bIMD && computeForces));                      // IMD
 572 }
 573
 574 /*! \brief Compute forces and/or energies for special algorithms
 575  *
 576  * The intention is to collect all calls to algorithms that compute
 577  * forces on local atoms only and that do not contribute to the local
 578  * virial sum (but add their virial contribution separately).
 579  * Eventually these should likely all become ForceProviders.
 580  * Within this function the intention is to have algorithms that do
 581  * global communication at the end, so global barriers within the MD loop
 582  * are as close together as possible.
 583  *
 584  * \param[in]     fplog            The log file
 585  * \param[in]     cr               The communication record
 586  * \param[in]     inputrec         The input record
 587  * \param[in]     awh              The Awh module (nullptr if none in use).
 588  * \param[in]     enforcedRotation Enforced rotation module.
 589  * \param[in]     imdSession       The IMD session
 590  * \param[in]     pull_work        The pull work structure.
 591  * \param[in]     step             The current MD step
 592  * \param[in]     t                The current time
 593  * \param[in,out] wcycle           Wallcycle accounting struct
 594  * \param[in,out] forceProviders   Pointer to a list of force providers
 595  * \param[in]     box              The unit cell
 596  * \param[in]     x                The coordinates
 597  * \param[in]     mdatoms          Per atom properties
 598  * \param[in]     lambda           Array of free-energy lambda values
 599  * \param[in]     stepWork         Step schedule flags
 600  * \param[in,out] forceWithVirialMtsLevel0  Force and virial for MTS level0 forces
 601  * \param[in,out] forceWithVirialMtsLevel1  Force and virial for MTS level1 forces, can be nullptr
 602  * \param[in,out] enerd            Energy buffer
 603  * \param[in,out] ed               Essential dynamics pointer
 604  * \param[in]     didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
 605  *
 606  * \todo Remove didNeighborSearch, which is used incorrectly.
 607  * \todo Convert all other algorithms called here to ForceProviders.
 608  */
 609 static void computeSpecialForces(FILE*                          fplog,
 610                                  const t_commrec*               cr,
 611                                  const t_inputrec*              inputrec,
 612                                  gmx::Awh*                      awh,
 613                                  gmx_enfrot*                    enforcedRotation,
 614                                  gmx::ImdSession*               imdSession,
 615                                  pull_t*                        pull_work,
 616                                  int64_t                        step,
 617                                  double                         t,
 618                                  gmx_wallcycle_t                wcycle,
 619                                  gmx::ForceProviders*           forceProviders,
 620                                  const matrix                   box,
 621                                  gmx::ArrayRef<const gmx::RVec> x,
 622                                  const t_mdatoms*               mdatoms,
 623                                  gmx::ArrayRef<const real>      lambda,
 624                                  const StepWorkload&            stepWork,
 625                                  gmx::ForceWithVirial*          forceWithVirialMtsLevel0,
 626                                  gmx::ForceWithVirial*          forceWithVirialMtsLevel1,
 627                                  gmx_enerdata_t*                enerd,
 628                                  gmx_edsam*                     ed,
 629                                  bool                           didNeighborSearch)
 630 {
 631     /* NOTE: Currently all ForceProviders only provide forces.
 632      *       When they also provide energies, remove this conditional.
 633      */
 634     if (stepWork.computeForces)
 635     {
 636         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 637         gmx::ForceProviderOutput forceProviderOutput(forceWithVirialMtsLevel0, enerd);
 638
 639         /* Collect forces from modules */
 640         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 641     }
 642
 643     if (inputrec->bPull && pull_have_potential(*pull_work))
 644     {
 645         const int mtsLevel = forceGroupMtsLevel(inputrec->mtsLevels, gmx::MtsForceGroups::Pull);
 646         if (mtsLevel == 0 || stepWork.computeSlowForces)
 647         {
 648             auto& forceWithVirial = (mtsLevel == 0) ? forceWithVirialMtsLevel0 : forceWithVirialMtsLevel1;
 649             pull_potential_wrapper(
 650                     cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work, lambda.data(), t, wcycle);
 651         }
 652     }
 653     if (awh)
 654     {
 655         const int mtsLevel = forceGroupMtsLevel(inputrec->mtsLevels, gmx::MtsForceGroups::Pull);
 656         if (mtsLevel == 0 || stepWork.computeSlowForces)
 657         {
 658             const bool needForeignEnergyDifferences = awh->needForeignEnergyDifferences(step);
 659             std::vector<double> foreignLambdaDeltaH, foreignLambdaDhDl;
 660             if (needForeignEnergyDifferences)
 661             {
 662                 enerd->foreignLambdaTerms.finalizePotentialContributions(
 663                         enerd->dvdl_lin, lambda, *inputrec->fepvals);
 664                 std::tie(foreignLambdaDeltaH, foreignLambdaDhDl) = enerd->foreignLambdaTerms.getTerms(cr);
 665             }
 666
 667             auto& forceWithVirial = (mtsLevel == 0) ? forceWithVirialMtsLevel0 : forceWithVirialMtsLevel1;
 668             enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(inputrec->pbcType,
 669                                                                          mdatoms->massT,
 670                                                                          foreignLambdaDeltaH,
 671                                                                          foreignLambdaDhDl,
 672                                                                          box,
 673                                                                          forceWithVirial,
 674                                                                          t,
 675                                                                          step,
 676                                                                          wcycle,
 677                                                                          fplog);
 678         }
 679     }
 680
 681     rvec* f = as_rvec_array(forceWithVirialMtsLevel0->force_.data());
 682
 683     /* Add the forces from enforced rotation potentials (if any) */
 684     if (inputrec->bRot)
 685     {
 686         wallcycle_start(wcycle, ewcROTadd);
 687         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 688         wallcycle_stop(wcycle, ewcROTadd);
 689     }
 690
 691     if (ed)
 692     {
 693         /* Note that since init_edsam() is called after the initialization
 694          * of forcerec, edsam doesn't request the noVirSum force buffer.
 695          * Thus if no other algorithm (e.g. PME) requires it, the forces
 696          * here will contribute to the virial.
 697          */
 698         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
 699     }
 700
 701     /* Add forces from interactive molecular dynamics (IMD), if any */
 702     if (inputrec->bIMD && stepWork.computeForces)
 703     {
 704         imdSession->applyForces(f);
 705     }
 706 }
 707
 708 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 709  *
 710  * \param[in]  pmedata              The PME structure
 711  * \param[in]  box                  The box matrix
 712  * \param[in]  stepWork             Step schedule flags
 713  * \param[in]  xReadyOnDevice       Event synchronizer indicating that the coordinates are ready in the device memory.
 714  * \param[in]  lambdaQ              The Coulomb lambda of the current state.
 715  * \param[in]  wcycle               The wallcycle structure
 716  */
 717 static inline void launchPmeGpuSpread(gmx_pme_t*            pmedata,
 718                                       const matrix          box,
 719                                       const StepWorkload&   stepWork,
 720                                       GpuEventSynchronizer* xReadyOnDevice,
 721                                       const real            lambdaQ,
 722                                       gmx_wallcycle_t       wcycle)
 723 {
 724     pme_gpu_prepare_computation(pmedata, box, wcycle, stepWork);
 725     pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle, lambdaQ);
 726 }
 727
 728 /*! \brief Launch the FFT and gather stages of PME GPU
 729  *
 730  * This function only implements setting the output forces (no accumulation).
 731  *
 732  * \param[in]  pmedata        The PME structure
 733  * \param[in]  lambdaQ        The Coulomb lambda of the current system state.
 734  * \param[in]  wcycle         The wallcycle structure
 735  * \param[in]  stepWork       Step schedule flags
 736  */
 737 static void launchPmeGpuFftAndGather(gmx_pme_t*               pmedata,
 738                                      const real               lambdaQ,
 739                                      gmx_wallcycle_t          wcycle,
 740                                      const gmx::StepWorkload& stepWork)
 741 {
 742     pme_gpu_launch_complex_transforms(pmedata, wcycle, stepWork);
 743     pme_gpu_launch_gather(pmedata, wcycle, lambdaQ);
 744 }
 745
 746 /*! \brief
 747  *  Polling wait for either of the PME or nonbonded GPU tasks.
 748  *
 749  * Instead of a static order in waiting for GPU tasks, this function
 750  * polls checking which of the two tasks completes first, and does the
 751  * associated force buffer reduction overlapped with the other task.
 752  * By doing that, unlike static scheduling order, it can always overlap
 753  * one of the reductions, regardless of the GPU task completion order.
 754  *
 755  * \param[in]     nbv              Nonbonded verlet structure
 756  * \param[in,out] pmedata          PME module data
 757  * \param[in,out] forceOutputsNonbonded  Force outputs for the non-bonded forces and shift forces
 758  * \param[in,out] forceOutputsPme  Force outputs for the PME forces and virial
 759  * \param[in,out] enerd            Energy data structure results are reduced into
 760  * \param[in]     lambdaQ          The Coulomb lambda of the current system state.
 761  * \param[in]     stepWork         Step schedule flags
 762  * \param[in]     wcycle           The wallcycle structure
 763  */
 764 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
 765                                         gmx_pme_t*          pmedata,
 766                                         gmx::ForceOutputs*  forceOutputsNonbonded,
 767                                         gmx::ForceOutputs*  forceOutputsPme,
 768                                         gmx_enerdata_t*     enerd,
 769                                         const real          lambdaQ,
 770                                         const StepWorkload& stepWork,
 771                                         gmx_wallcycle_t     wcycle)
 772 {
 773     bool isPmeGpuDone = false;
 774     bool isNbGpuDone  = false;
 775
 776     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 777
 778     while (!isPmeGpuDone || !isNbGpuDone)
 779     {
 780         if (!isPmeGpuDone)
 781         {
 782             GpuTaskCompletion completionType =
 783                     (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 784             isPmeGpuDone = pme_gpu_try_finish_task(
 785                     pmedata, stepWork, wcycle, &forceOutputsPme->forceWithVirial(), enerd, lambdaQ, completionType);
 786         }
 787
 788         if (!isNbGpuDone)
 789         {
 790             auto&             forceBuffersNonbonded = forceOutputsNonbonded->forceWithShiftForces();
 791             GpuTaskCompletion completionType =
 792                     (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 793             isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
 794                                                      stepWork,
 795                                                      AtomLocality::Local,
 796                                                      enerd->grpp.ener[egLJSR].data(),
 797                                                      enerd->grpp.ener[egCOULSR].data(),
 798                                                      forceBuffersNonbonded.shiftForces(),
 799                                                      completionType,
 800                                                      wcycle);
 801
 802             if (isNbGpuDone)
 803             {
 804                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceBuffersNonbonded.force());
 805             }
 806         }
 807     }
 808 }
 809
 810 /*! \brief Set up the different force buffers; also does clearing.
 811  *
 812  * \param[in] forceHelperBuffers  Helper force buffers
 813  * \param[in] force     force array
 814  * \param[in] stepWork  Step schedule flags
 815  * \param[out] wcycle   wallcycle recording structure
 816  *
 817  * \returns             Cleared force output structure
 818  */
 819 static ForceOutputs setupForceOutputs(ForceHelperBuffers*                 forceHelperBuffers,
 820                                       gmx::ArrayRefWithPadding<gmx::RVec> force,
 821                                       const StepWorkload&                 stepWork,
 822                                       gmx_wallcycle_t                     wcycle)
 823 {
 824     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 825
 826     /* NOTE: We assume fr->shiftForces is all zeros here */
 827     gmx::ForceWithShiftForces forceWithShiftForces(
 828             force, stepWork.computeVirial, forceHelperBuffers->shiftForces());
 829
 830     if (stepWork.computeForces)
 831     {
 832         /* Clear the short- and long-range forces */
 833         clearRVecs(forceWithShiftForces.force(), true);
 834
 835         /* Clear the shift forces */
 836         clearRVecs(forceWithShiftForces.shiftForces(), false);
 837     }
 838
 839     /* If we need to compute the virial, we might need a separate
 840      * force buffer for algorithms for which the virial is calculated
 841      * directly, such as PME. Otherwise, forceWithVirial uses the
 842      * the same force (f in legacy calls) buffer as other algorithms.
 843      */
 844     const bool useSeparateForceWithVirialBuffer =
 845             (stepWork.computeForces
 846              && (stepWork.computeVirial && forceHelperBuffers->haveDirectVirialContributions()));
 847     /* forceWithVirial uses the local atom range only */
 848     gmx::ForceWithVirial forceWithVirial(
 849             useSeparateForceWithVirialBuffer ? forceHelperBuffers->forceBufferForDirectVirialContributions()
 850                                              : force.unpaddedArrayRef(),
 851             stepWork.computeVirial);
 852
 853     if (useSeparateForceWithVirialBuffer)
 854     {
 855         /* TODO: update comment
 856          * We only compute forces on local atoms. Note that vsites can
 857          * spread to non-local atoms, but that part of the buffer is
 858          * cleared separately in the vsite spreading code.
 859          */
 860         clearRVecs(forceWithVirial.force_, true);
 861     }
 862
 863     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 864
 865     return ForceOutputs(
 866             forceWithShiftForces, forceHelperBuffers->haveDirectVirialContributions(), forceWithVirial);
 867 }
 868
 869
 870 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
 871  */
 872 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec&         inputrec,
 873                                                           const t_forcerec&         fr,
 874                                                           const pull_t*             pull_work,
 875                                                           const gmx_edsam*          ed,
 876                                                           const t_mdatoms&          mdatoms,
 877                                                           const SimulationWorkload& simulationWork,
 878                                                           const StepWorkload&       stepWork)
 879 {
 880     DomainLifetimeWorkload domainWork;
 881     // Note that haveSpecialForces is constant over the whole run
 882     domainWork.haveSpecialForces =
 883             haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 884     domainWork.haveCpuListedForceWork = false;
 885     domainWork.haveCpuBondedWork      = false;
 886     for (const auto& listedForces : fr.listedForces)
 887     {
 888         if (listedForces.haveCpuListedForces(*fr.fcdata))
 889         {
 890             domainWork.haveCpuListedForceWork = true;
 891         }
 892         if (listedForces.haveCpuBondeds())
 893         {
 894             domainWork.haveCpuBondedWork = true;
 895         }
 896     }
 897     domainWork.haveGpuBondedWork = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
 898     // Note that haveFreeEnergyWork is constant over the whole run
 899     domainWork.haveFreeEnergyWork = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
 900     // We assume we have local force work if there are CPU
 901     // force tasks including PME or nonbondeds.
 902     domainWork.haveCpuLocalForceWork =
 903             domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
 904             || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
 905             || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
 906
 907     return domainWork;
 908 }
 909
 910 /*! \brief Set up force flag stuct from the force bitmask.
 911  *
 912  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 913  * \param[in]      mtsLevels            The multiple time-stepping levels, either empty or 2 levels
 914  * \param[in]      step                 The current MD step
 915  * \param[in]      simulationWork       Simulation workload description.
 916  * \param[in]      rankHasPmeDuty       If this rank computes PME.
 917  *
 918  * \returns New Stepworkload description.
 919  */
 920 static StepWorkload setupStepWorkload(const int                     legacyFlags,
 921                                       ArrayRef<const gmx::MtsLevel> mtsLevels,
 922                                       const int64_t                 step,
 923                                       const SimulationWorkload&     simulationWork,
 924                                       const bool                    rankHasPmeDuty)
 925 {
 926     GMX_ASSERT(mtsLevels.empty() || mtsLevels.size() == 2, "Expect 0 or 2 MTS levels");
 927     const bool computeSlowForces = (mtsLevels.empty() || step % mtsLevels[1].stepFactor == 0);
 928
 929     StepWorkload flags;
 930     flags.stateChanged        = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 931     flags.haveDynamicBox      = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 932     flags.doNeighborSearch    = ((legacyFlags & GMX_FORCE_NS) != 0);
 933     flags.computeSlowForces   = computeSlowForces;
 934     flags.computeVirial       = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 935     flags.computeEnergy       = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 936     flags.computeForces       = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 937     flags.computeListedForces = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 938     flags.computeNonbondedForces =
 939             ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && simulationWork.computeNonbonded
 940             && !(simulationWork.computeNonbondedAtMtsLevel1 && !computeSlowForces);
 941     flags.computeDhdl = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 942
 943     if (simulationWork.useGpuBufferOps)
 944     {
 945         GMX_ASSERT(simulationWork.useGpuNonbonded,
 946                    "Can only offload buffer ops if nonbonded computation is also offloaded");
 947     }
 948     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 949     // on virial steps the CPU reduction path is taken
 950     flags.useGpuFBufferOps = simulationWork.useGpuBufferOps && !flags.computeVirial;
 951     flags.useGpuPmeFReduction = flags.computeSlowForces && flags.useGpuFBufferOps && simulationWork.useGpuPme
 952                                 && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication);
 953     flags.useGpuXHalo = simulationWork.useGpuHaloExchange;
 954     flags.useGpuFHalo = simulationWork.useGpuHaloExchange && flags.useGpuFBufferOps;
 955
 956     return flags;
 957 }
 958
 959
 960 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
 961  *
 962  * TODO: eliminate \p useGpuPmeOnThisRank when this is
 963  * incorporated in DomainLifetimeWorkload.
 964  */
 965 static void launchGpuEndOfStepTasks(nonbonded_verlet_t*               nbv,
 966                                     gmx::GpuBonded*                   gpuBonded,
 967                                     gmx_pme_t*                        pmedata,
 968                                     gmx_enerdata_t*                   enerd,
 969                                     const gmx::MdrunScheduleWorkload& runScheduleWork,
 970                                     bool                              useGpuPmeOnThisRank,
 971                                     int64_t                           step,
 972                                     gmx_wallcycle_t                   wcycle)
 973 {
 974     if (runScheduleWork.simulationWork.useGpuNonbonded && runScheduleWork.stepWork.computeNonbondedForces)
 975     {
 976         /* Launch pruning before buffer clearing because the API overhead of the
 977          * clear kernel launches can leave the GPU idle while it could be running
 978          * the prune kernel.
 979          */
 980         if (nbv->isDynamicPruningStepGpu(step))
 981         {
 982             nbv->dispatchPruneKernelGpu(step);
 983         }
 984
 985         /* now clear the GPU outputs while we finish the step on the CPU */
 986         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 987         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 988         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
 989         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 990         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 991     }
 992
 993     if (useGpuPmeOnThisRank)
 994     {
 995         pme_gpu_reinit_computation(pmedata, wcycle);
 996     }
 997
 998     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
 999     {
1000         // in principle this should be included in the DD balancing region,
1001         // but generally it is infrequent so we'll omit it for the sake of
1002         // simpler code
1003         gpuBonded->waitAccumulateEnergyTerms(enerd);
1004
1005         gpuBonded->clearEnergies();
1006     }
1007 }
1008
1009 //! \brief Data structure to hold dipole-related data and staging arrays
1010 struct DipoleData
1011 {
1012     //! Dipole staging for fast summing over MPI
1013     gmx::DVec muStaging[2] = { { 0.0, 0.0, 0.0 } };
1014     //! Dipole staging for states A and B (index 0 and 1 resp.)
1015     gmx::RVec muStateAB[2] = { { 0.0_real, 0.0_real, 0.0_real } };
1016 };
1017
1018
1019 static void reduceAndUpdateMuTot(DipoleData*                   dipoleData,
1020                                  const t_commrec*              cr,
1021                                  const bool                    haveFreeEnergy,
1022                                  gmx::ArrayRef<const real>     lambda,
1023                                  rvec                          muTotal,
1024                                  const DDBalanceRegionHandler& ddBalanceRegionHandler)
1025 {
1026     if (PAR(cr))
1027     {
1028         gmx_sumd(2 * DIM, dipoleData->muStaging[0], cr);
1029         ddBalanceRegionHandler.reopenRegionCpu();
1030     }
1031     for (int i = 0; i < 2; i++)
1032     {
1033         for (int j = 0; j < DIM; j++)
1034         {
1035             dipoleData->muStateAB[i][j] = dipoleData->muStaging[i][j];
1036         }
1037     }
1038
1039     if (!haveFreeEnergy)
1040     {
1041         copy_rvec(dipoleData->muStateAB[0], muTotal);
1042     }
1043     else
1044     {
1045         for (int j = 0; j < DIM; j++)
1046         {
1047             muTotal[j] = (1.0 - lambda[efptCOUL]) * dipoleData->muStateAB[0][j]
1048                          + lambda[efptCOUL] * dipoleData->muStateAB[1][j];
1049         }
1050     }
1051 }
1052
1053 /*! \brief Combines MTS level0 and level1 force buffes into a full and MTS-combined force buffer.
1054  *
1055  * \param[in]     numAtoms        The number of atoms to combine forces for
1056  * \param[in,out] forceMtsLevel0  Input: F_level0, output: F_level0 + F_level1
1057  * \param[in,out] forceMts        Input: F_level1, output: F_level0 + mtsFactor * F_level1
1058  * \param[in]     mtsFactor       The factor between the level0 and level1 time step
1059  */
1060 static void combineMtsForces(const int      numAtoms,
1061                              ArrayRef<RVec> forceMtsLevel0,
1062                              ArrayRef<RVec> forceMts,
1063                              const real     mtsFactor)
1064 {
1065     const int gmx_unused numThreads = gmx_omp_nthreads_get(emntDefault);
1066 #pragma omp parallel for num_threads(numThreads) schedule(static)
1067     for (int i = 0; i < numAtoms; i++)
1068     {
1069         const RVec forceMtsLevel0Tmp = forceMtsLevel0[i];
1070         forceMtsLevel0[i] += forceMts[i];
1071         forceMts[i] = forceMtsLevel0Tmp + mtsFactor * forceMts[i];
1072     }
1073 }
1074
1075 /*! \brief Setup for the local and non-local GPU force reductions:
1076  * reinitialization plus the registration of forces and dependencies.
1077  *
1078  * \param [in] runScheduleWork               Schedule workload flag structure
1079  * \param [in] cr                            Communication record object
1080  * \param [in] fr                            Force record object
1081  */
1082 static void setupGpuForceReductions(gmx::MdrunScheduleWorkload* runScheduleWork,
1083                                     const t_commrec*            cr,
1084                                     t_forcerec*                 fr)
1085 {
1086
1087     nonbonded_verlet_t*          nbv      = fr->nbv.get();
1088     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1089
1090     // (re-)initialize local GPU force reduction
1091     const bool accumulate =
1092             runScheduleWork->domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr);
1093     const int atomStart = 0;
1094     fr->gpuForceReduction[gmx::AtomLocality::Local]->reinit(stateGpu->getForces(),
1095                                                             nbv->getNumAtoms(AtomLocality::Local),
1096                                                             nbv->getGridIndices(),
1097                                                             atomStart,
1098                                                             accumulate,
1099                                                             stateGpu->fReducedOnDevice());
1100
1101     // register forces and add dependencies
1102     fr->gpuForceReduction[gmx::AtomLocality::Local]->registerNbnxmForce(nbv->getGpuForces());
1103
1104     if (runScheduleWork->simulationWork.useGpuPme
1105         && (thisRankHasDuty(cr, DUTY_PME) || runScheduleWork->simulationWork.useGpuPmePpCommunication))
1106     {
1107         void* forcePtr = thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_device_f(fr->pmedata)
1108                                                        : // PME force buffer on same GPU
1109                                  fr->pmePpCommGpu->getGpuForceStagingPtr(); // buffer received from other GPU
1110         fr->gpuForceReduction[gmx::AtomLocality::Local]->registerRvecForce(forcePtr);
1111
1112         GpuEventSynchronizer* const pmeSynchronizer =
1113                 (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_f_ready_synchronizer(fr->pmedata)
1114                                                : // PME force buffer on same GPU
1115                          fr->pmePpCommGpu->getForcesReadySynchronizer()); // buffer received from other GPU
1116         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(pmeSynchronizer);
1117     }
1118
1119     if ((runScheduleWork->domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr))
1120         && !runScheduleWork->simulationWork.useGpuHaloExchange)
1121     {
1122         auto forcesReadyLocality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
1123         const bool useGpuForceBufferOps = true;
1124         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1125                 stateGpu->getForcesReadyOnDeviceEvent(forcesReadyLocality, useGpuForceBufferOps));
1126     }
1127
1128     if (runScheduleWork->simulationWork.useGpuHaloExchange)
1129     {
1130         fr->gpuForceReduction[gmx::AtomLocality::Local]->addDependency(
1131                 cr->dd->gpuHaloExchange[0][0]->getForcesReadyOnDeviceEvent());
1132     }
1133
1134     if (havePPDomainDecomposition(cr))
1135     {
1136         // (re-)initialize non-local GPU force reduction
1137         const bool accumulate = runScheduleWork->domainWork.haveCpuBondedWork
1138                                 || runScheduleWork->domainWork.haveFreeEnergyWork;
1139         const int atomStart = dd_numHomeAtoms(*cr->dd);
1140         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->reinit(stateGpu->getForces(),
1141                                                                    nbv->getNumAtoms(AtomLocality::NonLocal),
1142                                                                    nbv->getGridIndices(),
1143                                                                    atomStart,
1144                                                                    accumulate);
1145
1146         // register forces and add dependencies
1147         fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->registerNbnxmForce(nbv->getGpuForces());
1148         if (runScheduleWork->domainWork.haveCpuBondedWork || runScheduleWork->domainWork.haveFreeEnergyWork)
1149         {
1150             fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->addDependency(
1151                     stateGpu->getForcesReadyOnDeviceEvent(AtomLocality::NonLocal, true));
1152         }
1153     }
1154 }
1155
1156
1157 void do_force(FILE*                               fplog,
1158               const t_commrec*                    cr,
1159               const gmx_multisim_t*               ms,
1160               const t_inputrec*                   inputrec,
1161               gmx::Awh*                           awh,
1162               gmx_enfrot*                         enforcedRotation,
1163               gmx::ImdSession*                    imdSession,
1164               pull_t*                             pull_work,
1165               int64_t                             step,
1166               t_nrnb*                             nrnb,
1167               gmx_wallcycle_t                     wcycle,
1168               const gmx_localtop_t*               top,
1169               const matrix                        box,
1170               gmx::ArrayRefWithPadding<gmx::RVec> x,
1171               const history_t*                    hist,
1172               gmx::ForceBuffersView*              forceView,
1173               tensor                              vir_force,
1174               const t_mdatoms*                    mdatoms,
1175               gmx_enerdata_t*                     enerd,
1176               gmx::ArrayRef<const real>           lambda,
1177               t_forcerec*                         fr,
1178               gmx::MdrunScheduleWorkload*         runScheduleWork,
1179               gmx::VirtualSitesHandler*           vsite,
1180               rvec                                muTotal,
1181               double                              t,
1182               gmx_edsam*                          ed,
1183               int                                 legacyFlags,
1184               const DDBalanceRegionHandler&       ddBalanceRegionHandler)
1185 {
1186     auto force = forceView->forceWithPadding();
1187     GMX_ASSERT(force.unpaddedArrayRef().ssize() >= fr->natoms_force_constr,
1188                "The size of the force buffer should be at least the number of atoms to compute "
1189                "forces for");
1190
1191     nonbonded_verlet_t*  nbv = fr->nbv.get();
1192     interaction_const_t* ic  = fr->ic;
1193
1194     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1195
1196     const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
1197
1198     runScheduleWork->stepWork = setupStepWorkload(
1199             legacyFlags, inputrec->mtsLevels, step, simulationWork, thisRankHasDuty(cr, DUTY_PME));
1200     const StepWorkload& stepWork = runScheduleWork->stepWork;
1201
1202     const bool useGpuPmeOnThisRank =
1203             simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME) && stepWork.computeSlowForces;
1204
1205     /* At a search step we need to start the first balancing region
1206      * somewhere early inside the step after communication during domain
1207      * decomposition (and not during the previous step as usual).
1208      */
1209     if (stepWork.doNeighborSearch)
1210     {
1211         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
1212     }
1213
1214     clear_mat(vir_force);
1215
1216     if (fr->pbcType != PbcType::No)
1217     {
1218         /* Compute shift vectors every step,
1219          * because of pressure coupling or box deformation!
1220          */
1221         if (stepWork.haveDynamicBox && stepWork.stateChanged)
1222         {
1223             calc_shifts(box, fr->shift_vec);
1224         }
1225
1226         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
1227         const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
1228         if (calcCGCM)
1229         {
1230             put_atoms_in_box_omp(fr->pbcType,
1231                                  box,
1232                                  x.unpaddedArrayRef().subArray(0, mdatoms->homenr),
1233                                  gmx_omp_nthreads_get(emntDefault));
1234             inc_nrnb(nrnb, eNR_SHIFTX, mdatoms->homenr);
1235         }
1236     }
1237
1238     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1239
1240     const bool pmeSendCoordinatesFromGpu =
1241             GMX_MPI && simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1242     const bool reinitGpuPmePpComms =
1243             GMX_MPI && simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1244
1245     const auto localXReadyOnDevice = (useGpuPmeOnThisRank || simulationWork.useGpuBufferOps)
1246                                              ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1247                                                        AtomLocality::Local, simulationWork, stepWork)
1248                                              : nullptr;
1249
1250     // Copy coordinate from the GPU if update is on the GPU and there
1251     // are forces to be computed on the CPU, or for the computation of
1252     // virial, or if host-side data will be transferred from this task
1253     // to a remote task for halo exchange or PME-PP communication. At
1254     // search steps the current coordinates are already on the host,
1255     // hence copy is not needed.
1256     const bool haveHostPmePpComms =
1257             !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication;
1258
1259     GMX_ASSERT(simulationWork.useGpuHaloExchange
1260                        == ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange[0].empty())),
1261                "The GPU halo exchange is active, but it has not been constructed.");
1262     const bool haveHostHaloExchangeComms =
1263             havePPDomainDecomposition(cr) && !simulationWork.useGpuHaloExchange;
1264
1265     bool gmx_used_in_debug haveCopiedXFromGpu = false;
1266     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1267         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1268             || haveHostPmePpComms || haveHostHaloExchangeComms))
1269     {
1270         stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1271         haveCopiedXFromGpu = true;
1272     }
1273
1274     // If coordinates are to be sent to PME task from CPU memory, perform that send here.
1275     // Otherwise the send will occur after H2D coordinate transfer.
1276     if (GMX_MPI && !thisRankHasDuty(cr, DUTY_PME) && !pmeSendCoordinatesFromGpu && stepWork.computeSlowForces)
1277     {
1278         /* Send particle coordinates to the pme nodes */
1279         if (!stepWork.doNeighborSearch && simulationWork.useGpuUpdate)
1280         {
1281             stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1282         }
1283
1284         gmx_pme_send_coordinates(fr,
1285                                  cr,
1286                                  box,
1287                                  as_rvec_array(x.unpaddedArrayRef().data()),
1288                                  lambda[efptCOUL],
1289                                  lambda[efptVDW],
1290                                  (stepWork.computeVirial || stepWork.computeEnergy),
1291                                  step,
1292                                  simulationWork.useGpuPmePpCommunication,
1293                                  reinitGpuPmePpComms,
1294                                  pmeSendCoordinatesFromGpu,
1295                                  localXReadyOnDevice,
1296                                  wcycle);
1297     }
1298
1299     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1300     // The local coordinates can be copied right away.
1301     // NOTE: Consider moving this copy to right after they are updated and constrained,
1302     //       if the later is not offloaded.
1303     if (useGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
1304     {
1305         if (stepWork.doNeighborSearch)
1306         {
1307             // TODO refactor this to do_md, after partitioning.
1308             stateGpu->reinit(mdatoms->homenr,
1309                              cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1310             if (useGpuPmeOnThisRank)
1311             {
1312                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1313                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1314             }
1315         }
1316         // We need to copy coordinates when:
1317         // 1. Update is not offloaded
1318         // 2. The buffers were reinitialized on search step
1319         if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1320         {
1321             GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1322             stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1323         }
1324     }
1325
1326     // If coordinates are to be sent to PME task from GPU memory, perform that send here.
1327     // Otherwise the send will occur before the H2D coordinate transfer.
1328     if (!thisRankHasDuty(cr, DUTY_PME) && pmeSendCoordinatesFromGpu)
1329     {
1330         /* Send particle coordinates to the pme nodes */
1331         gmx_pme_send_coordinates(fr,
1332                                  cr,
1333                                  box,
1334                                  as_rvec_array(x.unpaddedArrayRef().data()),
1335                                  lambda[efptCOUL],
1336                                  lambda[efptVDW],
1337                                  (stepWork.computeVirial || stepWork.computeEnergy),
1338                                  step,
1339                                  simulationWork.useGpuPmePpCommunication,
1340                                  reinitGpuPmePpComms,
1341                                  pmeSendCoordinatesFromGpu,
1342                                  localXReadyOnDevice,
1343                                  wcycle);
1344     }
1345
1346     if (useGpuPmeOnThisRank)
1347     {
1348         launchPmeGpuSpread(fr->pmedata, box, stepWork, localXReadyOnDevice, lambda[efptCOUL], wcycle);
1349     }
1350
1351     const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1352
1353     /* do gridding for pair search */
1354     if (stepWork.doNeighborSearch)
1355     {
1356         if (fr->wholeMoleculeTransform && stepWork.stateChanged)
1357         {
1358             fr->wholeMoleculeTransform->updateForAtomPbcJumps(x.unpaddedArrayRef(), box);
1359         }
1360
1361         // TODO
1362         // - vzero is constant, do we need to pass it?
1363         // - box_diag should be passed directly to nbnxn_put_on_grid
1364         //
1365         rvec vzero;
1366         clear_rvec(vzero);
1367
1368         rvec box_diag;
1369         box_diag[XX] = box[XX][XX];
1370         box_diag[YY] = box[YY][YY];
1371         box_diag[ZZ] = box[ZZ][ZZ];
1372
1373         wallcycle_start(wcycle, ewcNS);
1374         if (!DOMAINDECOMP(cr))
1375         {
1376             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1377             nbnxn_put_on_grid(nbv,
1378                               box,
1379                               0,
1380                               vzero,
1381                               box_diag,
1382                               nullptr,
1383                               { 0, mdatoms->homenr },
1384                               -1,
1385                               fr->cginfo,
1386                               x.unpaddedArrayRef(),
1387                               0,
1388                               nullptr);
1389             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1390         }
1391         else
1392         {
1393             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1394             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->cginfo, x.unpaddedArrayRef());
1395             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1396         }
1397
1398         nbv->setAtomProperties(gmx::constArrayRefFromArray(mdatoms->typeA, mdatoms->nr),
1399                                gmx::constArrayRefFromArray(mdatoms->chargeA, mdatoms->nr),
1400                                fr->cginfo);
1401
1402         wallcycle_stop(wcycle, ewcNS);
1403
1404         /* initialize the GPU nbnxm atom data and bonded data structures */
1405         if (simulationWork.useGpuNonbonded)
1406         {
1407             // Note: cycle counting only nononbondeds, gpuBonded counts internally
1408             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1409             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1410             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1411             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1412             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1413
1414             if (fr->gpuBonded)
1415             {
1416                 /* Now we put all atoms on the grid, we can assign bonded
1417                  * interactions to the GPU, where the grid order is
1418                  * needed. Also the xq, f and fshift device buffers have
1419                  * been reallocated if needed, so the bonded code can
1420                  * learn about them. */
1421                 // TODO the xq, f, and fshift buffers are now shared
1422                 // resources, so they should be maintained by a
1423                 // higher-level object than the nb module.
1424                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbv->getGridIndices(),
1425                                                                       top->idef,
1426                                                                       Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1427                                                                       Nbnxm::gpu_get_f(nbv->gpu_nbv),
1428                                                                       Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1429             }
1430         }
1431
1432         // Need to run after the GPU-offload bonded interaction lists
1433         // are set up to be able to determine whether there is bonded work.
1434         runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1435                 *inputrec, *fr, pull_work, ed, *mdatoms, simulationWork, stepWork);
1436
1437         wallcycle_start_nocount(wcycle, ewcNS);
1438         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1439         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1440         nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1441
1442         nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::Local);
1443
1444         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1445         wallcycle_stop(wcycle, ewcNS);
1446
1447         if (stepWork.useGpuXBufferOps)
1448         {
1449             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1450         }
1451
1452         if (simulationWork.useGpuBufferOps)
1453         {
1454             setupGpuForceReductions(runScheduleWork, cr, fr);
1455         }
1456     }
1457     else if (!EI_TPI(inputrec->eI) && stepWork.computeNonbondedForces)
1458     {
1459         if (stepWork.useGpuXBufferOps)
1460         {
1461             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1462             nbv->convertCoordinatesGpu(
1463                     AtomLocality::Local, false, stateGpu->getCoordinates(), localXReadyOnDevice);
1464         }
1465         else
1466         {
1467             if (simulationWork.useGpuUpdate)
1468             {
1469                 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1470                 GMX_ASSERT(haveCopiedXFromGpu,
1471                            "a wait should only be triggered if copy has been scheduled");
1472                 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1473             }
1474             nbv->convertCoordinates(AtomLocality::Local, false, x.unpaddedArrayRef());
1475         }
1476     }
1477
1478     if (simulationWork.useGpuNonbonded && (stepWork.computeNonbondedForces || domainWork.haveGpuBondedWork))
1479     {
1480         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1481
1482         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1483         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1484         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1485         if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1486         {
1487             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1488         }
1489         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1490         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1491         // with X buffer ops offloaded to the GPU on all but the search steps
1492
1493         // bonded work not split into separate local and non-local, so with DD
1494         // we can only launch the kernel after non-local coordinates have been received.
1495         if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1496         {
1497             fr->gpuBonded->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1498         }
1499
1500         /* launch local nonbonded work on GPU */
1501         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1502         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1503         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1504         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1505         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1506     }
1507
1508     if (useGpuPmeOnThisRank)
1509     {
1510         // In PME GPU and mixed mode we launch FFT / gather after the
1511         // X copy/transform to allow overlap as well as after the GPU NB
1512         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1513         // the nonbonded kernel.
1514         launchPmeGpuFftAndGather(fr->pmedata, lambda[efptCOUL], wcycle, stepWork);
1515     }
1516
1517     /* Communicate coordinates and sum dipole if necessary +
1518        do non-local pair search */
1519     if (havePPDomainDecomposition(cr))
1520     {
1521         if (stepWork.doNeighborSearch)
1522         {
1523             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1524             wallcycle_start_nocount(wcycle, ewcNS);
1525             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1526             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1527             nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1528
1529             nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::NonLocal);
1530             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1531             wallcycle_stop(wcycle, ewcNS);
1532             // TODO refactor this GPU halo exchange re-initialisation
1533             // to location in do_md where GPU halo exchange is
1534             // constructed at partitioning, after above stateGpu
1535             // re-initialization has similarly been refactored
1536             if (simulationWork.useGpuHaloExchange)
1537             {
1538                 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1539             }
1540         }
1541         else
1542         {
1543             if (stepWork.useGpuXHalo)
1544             {
1545                 // The following must be called after local setCoordinates (which records an event
1546                 // when the coordinate data has been copied to the device).
1547                 communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1548
1549                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1550                 {
1551                     // non-local part of coordinate buffer must be copied back to host for CPU work
1552                     stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1553                 }
1554             }
1555             else
1556             {
1557                 if (simulationWork.useGpuUpdate)
1558                 {
1559                     GMX_ASSERT(haveCopiedXFromGpu,
1560                                "a wait should only be triggered if copy has been scheduled");
1561                     stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1562                 }
1563                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1564             }
1565
1566             if (stepWork.useGpuXBufferOps)
1567             {
1568                 if (!useGpuPmeOnThisRank && !stepWork.useGpuXHalo)
1569                 {
1570                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1571                 }
1572                 nbv->convertCoordinatesGpu(AtomLocality::NonLocal,
1573                                            false,
1574                                            stateGpu->getCoordinates(),
1575                                            stateGpu->getCoordinatesReadyOnDeviceEvent(
1576                                                    AtomLocality::NonLocal, simulationWork, stepWork));
1577             }
1578             else
1579             {
1580                 nbv->convertCoordinates(AtomLocality::NonLocal, false, x.unpaddedArrayRef());
1581             }
1582         }
1583
1584         if (simulationWork.useGpuNonbonded)
1585         {
1586
1587             if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1588             {
1589                 wallcycle_start(wcycle, ewcLAUNCH_GPU);
1590                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1591                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1592                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1593                 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1594             }
1595
1596             if (domainWork.haveGpuBondedWork)
1597             {
1598                 fr->gpuBonded->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1599             }
1600
1601             /* launch non-local nonbonded tasks on GPU */
1602             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1603             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1604             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step, nrnb, wcycle);
1605             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1606             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1607         }
1608     }
1609
1610     if (simulationWork.useGpuNonbonded && stepWork.computeNonbondedForces)
1611     {
1612         /* launch D2H copy-back F */
1613         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1614         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1615
1616         if (havePPDomainDecomposition(cr))
1617         {
1618             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1619         }
1620         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1621         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1622
1623         if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1624         {
1625             fr->gpuBonded->launchEnergyTransfer();
1626         }
1627         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1628     }
1629
1630     gmx::ArrayRef<const gmx::RVec> xWholeMolecules;
1631     if (fr->wholeMoleculeTransform)
1632     {
1633         xWholeMolecules = fr->wholeMoleculeTransform->wholeMoleculeCoordinates(x.unpaddedArrayRef(), box);
1634     }
1635
1636     DipoleData dipoleData;
1637
1638     if (simulationWork.computeMuTot)
1639     {
1640         const int start = 0;
1641
1642         /* Calculate total (local) dipole moment in a temporary common array.
1643          * This makes it possible to sum them over nodes faster.
1644          */
1645         gmx::ArrayRef<const gmx::RVec> xRef =
1646                 (xWholeMolecules.empty() ? x.unpaddedArrayRef() : xWholeMolecules);
1647         calc_mu(start,
1648                 mdatoms->homenr,
1649                 xRef,
1650                 mdatoms->chargeA,
1651                 mdatoms->chargeB,
1652                 mdatoms->nChargePerturbed,
1653                 dipoleData.muStaging[0],
1654                 dipoleData.muStaging[1]);
1655
1656         reduceAndUpdateMuTot(&dipoleData, cr, (fr->efep != efepNO), lambda, muTotal, ddBalanceRegionHandler);
1657     }
1658
1659     /* Reset energies */
1660     reset_enerdata(enerd);
1661
1662     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1663     {
1664         wallcycle_start(wcycle, ewcPPDURINGPME);
1665         dd_force_flop_start(cr->dd, nrnb);
1666     }
1667
1668     // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1669     // this wait ensures that the D2H transfer is complete.
1670     if ((simulationWork.useGpuUpdate)
1671         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial))
1672     {
1673         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1674     }
1675
1676     if (inputrec->bRot)
1677     {
1678         wallcycle_start(wcycle, ewcROT);
1679         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, stepWork.doNeighborSearch);
1680         wallcycle_stop(wcycle, ewcROT);
1681     }
1682
1683     /* Start the force cycle counter.
1684      * Note that a different counter is used for dynamic load balancing.
1685      */
1686     wallcycle_start(wcycle, ewcFORCE);
1687
1688     /* Set up and clear force outputs:
1689      * forceOutMtsLevel0:  everything except what is in the other two outputs
1690      * forceOutMtsLevel1:  PME-mesh and listed-forces group 1
1691      * forceOutNonbonded: non-bonded forces
1692      * Without multiple time stepping all point to the same object.
1693      * With multiple time-stepping the use is different for MTS fast (level0 only) and slow steps.
1694      */
1695     ForceOutputs forceOutMtsLevel0 =
1696             setupForceOutputs(&fr->forceHelperBuffers[0], force, stepWork, wcycle);
1697
1698     // Force output for MTS combined forces, only set at level1 MTS steps
1699     std::optional<ForceOutputs> forceOutMts =
1700             (fr->useMts && stepWork.computeSlowForces)
1701                     ? std::optional(setupForceOutputs(&fr->forceHelperBuffers[1],
1702                                                       forceView->forceMtsCombinedWithPadding(),
1703                                                       stepWork,
1704                                                       wcycle))
1705                     : std::nullopt;
1706
1707     ForceOutputs* forceOutMtsLevel1 =
1708             fr->useMts ? (stepWork.computeSlowForces ? &forceOutMts.value() : nullptr) : &forceOutMtsLevel0;
1709
1710     const bool nonbondedAtMtsLevel1 = runScheduleWork->simulationWork.computeNonbondedAtMtsLevel1;
1711
1712     ForceOutputs* forceOutNonbonded = nonbondedAtMtsLevel1 ? forceOutMtsLevel1 : &forceOutMtsLevel0;
1713
1714     if (inputrec->bPull && pull_have_constraint(*pull_work))
1715     {
1716         clear_pull_forces(pull_work);
1717     }
1718
1719     /* We calculate the non-bonded forces, when done on the CPU, here.
1720      * We do this before calling do_force_lowlevel, because in that
1721      * function, the listed forces are calculated before PME, which
1722      * does communication.  With this order, non-bonded and listed
1723      * force calculation imbalance can be balanced out by the domain
1724      * decomposition load balancing.
1725      */
1726
1727     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1728
1729     if (!useOrEmulateGpuNb)
1730     {
1731         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1732     }
1733
1734     if (fr->efep != efepNO && stepWork.computeNonbondedForces)
1735     {
1736         /* Calculate the local and non-local free energy interactions here.
1737          * Happens here on the CPU both with and without GPU.
1738          */
1739         nbv->dispatchFreeEnergyKernel(InteractionLocality::Local,
1740                                       fr,
1741                                       as_rvec_array(x.unpaddedArrayRef().data()),
1742                                       &forceOutNonbonded->forceWithShiftForces(),
1743                                       *mdatoms,
1744                                       inputrec->fepvals,
1745                                       lambda,
1746                                       enerd,
1747                                       stepWork,
1748                                       nrnb);
1749
1750         if (havePPDomainDecomposition(cr))
1751         {
1752             nbv->dispatchFreeEnergyKernel(InteractionLocality::NonLocal,
1753                                           fr,
1754                                           as_rvec_array(x.unpaddedArrayRef().data()),
1755                                           &forceOutNonbonded->forceWithShiftForces(),
1756                                           *mdatoms,
1757                                           inputrec->fepvals,
1758                                           lambda,
1759                                           enerd,
1760                                           stepWork,
1761                                           nrnb);
1762         }
1763     }
1764
1765     if (stepWork.computeNonbondedForces && !useOrEmulateGpuNb)
1766     {
1767         if (havePPDomainDecomposition(cr))
1768         {
1769             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step, nrnb, wcycle);
1770         }
1771
1772         if (stepWork.computeForces)
1773         {
1774             /* Add all the non-bonded force to the normal force array.
1775              * This can be split into a local and a non-local part when overlapping
1776              * communication with calculation with domain decomposition.
1777              */
1778             wallcycle_stop(wcycle, ewcFORCE);
1779             nbv->atomdata_add_nbat_f_to_f(AtomLocality::All,
1780                                           forceOutNonbonded->forceWithShiftForces().force());
1781             wallcycle_start_nocount(wcycle, ewcFORCE);
1782         }
1783
1784         /* If there are multiple fshift output buffers we need to reduce them */
1785         if (stepWork.computeVirial)
1786         {
1787             /* This is not in a subcounter because it takes a
1788                negligible and constant-sized amount of time */
1789             nbnxn_atomdata_add_nbat_fshift_to_fshift(
1790                     *nbv->nbat, forceOutNonbonded->forceWithShiftForces().shiftForces());
1791         }
1792     }
1793
1794     // TODO Force flags should include haveFreeEnergyWork for this domain
1795     if (stepWork.useGpuXHalo && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1796     {
1797         wallcycle_stop(wcycle, ewcFORCE);
1798         /* Wait for non-local coordinate data to be copied from device */
1799         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1800         wallcycle_start_nocount(wcycle, ewcFORCE);
1801     }
1802
1803     // Compute wall interactions, when present.
1804     // Note: should be moved to special forces.
1805     if (inputrec->nwall && stepWork.computeNonbondedForces)
1806     {
1807         /* foreign lambda component for walls */
1808         real dvdl_walls = do_walls(*inputrec,
1809                                    *fr,
1810                                    box,
1811                                    *mdatoms,
1812                                    x.unpaddedConstArrayRef(),
1813                                    &forceOutMtsLevel0.forceWithVirial(),
1814                                    lambda[efptVDW],
1815                                    enerd->grpp.ener[egLJSR].data(),
1816                                    nrnb);
1817         enerd->dvdl_lin[efptVDW] += dvdl_walls;
1818     }
1819
1820     if (stepWork.computeListedForces)
1821     {
1822         /* Check whether we need to take into account PBC in listed interactions */
1823         bool needMolPbc = false;
1824         for (const auto& listedForces : fr->listedForces)
1825         {
1826             if (listedForces.haveCpuListedForces(*fr->fcdata))
1827             {
1828                 needMolPbc = fr->bMolPBC;
1829             }
1830         }
1831
1832         t_pbc pbc;
1833
1834         if (needMolPbc)
1835         {
1836             /* Since all atoms are in the rectangular or triclinic unit-cell,
1837              * only single box vector shifts (2 in x) are required.
1838              */
1839             set_pbc_dd(&pbc, fr->pbcType, DOMAINDECOMP(cr) ? cr->dd->numCells : nullptr, TRUE, box);
1840         }
1841
1842         for (int mtsIndex = 0; mtsIndex < (fr->useMts && stepWork.computeSlowForces ? 2 : 1); mtsIndex++)
1843         {
1844             ListedForces& listedForces = fr->listedForces[mtsIndex];
1845             ForceOutputs& forceOut     = (mtsIndex == 0 ? forceOutMtsLevel0 : *forceOutMtsLevel1);
1846             listedForces.calculate(wcycle,
1847                                    box,
1848                                    inputrec->fepvals,
1849                                    cr,
1850                                    ms,
1851                                    x,
1852                                    xWholeMolecules,
1853                                    fr->fcdata.get(),
1854                                    hist,
1855                                    &forceOut,
1856                                    fr,
1857                                    &pbc,
1858                                    enerd,
1859                                    nrnb,
1860                                    lambda.data(),
1861                                    mdatoms,
1862                                    DOMAINDECOMP(cr) ? cr->dd->globalAtomIndices.data() : nullptr,
1863                                    stepWork);
1864         }
1865     }
1866
1867     if (stepWork.computeSlowForces)
1868     {
1869         calculateLongRangeNonbondeds(fr,
1870                                      inputrec,
1871                                      cr,
1872                                      nrnb,
1873                                      wcycle,
1874                                      mdatoms,
1875                                      x.unpaddedConstArrayRef(),
1876                                      &forceOutMtsLevel1->forceWithVirial(),
1877                                      enerd,
1878                                      box,
1879                                      lambda.data(),
1880                                      as_rvec_array(dipoleData.muStateAB),
1881                                      stepWork,
1882                                      ddBalanceRegionHandler);
1883     }
1884
1885     wallcycle_stop(wcycle, ewcFORCE);
1886
1887     // VdW dispersion correction, only computed on master rank to avoid double counting
1888     if ((stepWork.computeEnergy || stepWork.computeVirial) && fr->dispersionCorrection && MASTER(cr))
1889     {
1890         // Calculate long range corrections to pressure and energy
1891         const DispersionCorrection::Correction correction =
1892                 fr->dispersionCorrection->calculate(box, lambda[efptVDW]);
1893
1894         if (stepWork.computeEnergy)
1895         {
1896             enerd->term[F_DISPCORR] = correction.energy;
1897             enerd->term[F_DVDL_VDW] += correction.dvdl;
1898             enerd->dvdl_lin[efptVDW] += correction.dvdl;
1899         }
1900         if (stepWork.computeVirial)
1901         {
1902             correction.correctVirial(vir_force);
1903             enerd->term[F_PDISPCORR] = correction.pressure;
1904         }
1905     }
1906
1907     computeSpecialForces(fplog,
1908                          cr,
1909                          inputrec,
1910                          awh,
1911                          enforcedRotation,
1912                          imdSession,
1913                          pull_work,
1914                          step,
1915                          t,
1916                          wcycle,
1917                          fr->forceProviders,
1918                          box,
1919                          x.unpaddedArrayRef(),
1920                          mdatoms,
1921                          lambda,
1922                          stepWork,
1923                          &forceOutMtsLevel0.forceWithVirial(),
1924                          forceOutMtsLevel1 ? &forceOutMtsLevel1->forceWithVirial() : nullptr,
1925                          enerd,
1926                          ed,
1927                          stepWork.doNeighborSearch);
1928
1929     if (havePPDomainDecomposition(cr) && stepWork.computeForces && stepWork.useGpuFHalo
1930         && domainWork.haveCpuLocalForceWork)
1931     {
1932         stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(), AtomLocality::Local);
1933     }
1934
1935     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
1936                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
1937     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFHalo),
1938                "The schedule below does not allow for nonbonded MTS with GPU halo exchange");
1939     // Will store the amount of cycles spent waiting for the GPU that
1940     // will be later used in the DLB accounting.
1941     float cycles_wait_gpu = 0;
1942     if (useOrEmulateGpuNb && stepWork.computeNonbondedForces)
1943     {
1944         auto& forceWithShiftForces = forceOutNonbonded->forceWithShiftForces();
1945
1946         /* wait for non-local forces (or calculate in emulation mode) */
1947         if (havePPDomainDecomposition(cr))
1948         {
1949             if (simulationWork.useGpuNonbonded)
1950             {
1951                 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1952                                                                stepWork,
1953                                                                AtomLocality::NonLocal,
1954                                                                enerd->grpp.ener[egLJSR].data(),
1955                                                                enerd->grpp.ener[egCOULSR].data(),
1956                                                                forceWithShiftForces.shiftForces(),
1957                                                                wcycle);
1958             }
1959             else
1960             {
1961                 wallcycle_start_nocount(wcycle, ewcFORCE);
1962                 do_nb_verlet(
1963                         fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes, step, nrnb, wcycle);
1964                 wallcycle_stop(wcycle, ewcFORCE);
1965             }
1966
1967             if (stepWork.useGpuFBufferOps)
1968             {
1969                 // TODO: move this into DomainLifetimeWorkload, including the second part of the
1970                 // condition The bonded and free energy CPU tasks can have non-local force
1971                 // contributions which are a dependency for the GPU force reduction.
1972                 bool haveNonLocalForceContribInCpuBuffer =
1973                         domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1974
1975                 if (haveNonLocalForceContribInCpuBuffer)
1976                 {
1977                     stateGpu->copyForcesToGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
1978                                               AtomLocality::NonLocal);
1979                 }
1980
1981
1982                 fr->gpuForceReduction[gmx::AtomLocality::NonLocal]->execute();
1983
1984                 if (!stepWork.useGpuFHalo)
1985                 {
1986                     // copy from GPU input for dd_move_f()
1987                     stateGpu->copyForcesFromGpu(forceOutMtsLevel0.forceWithShiftForces().force(),
1988                                                 AtomLocality::NonLocal);
1989                 }
1990             }
1991             else
1992             {
1993                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
1994             }
1995
1996             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1997             {
1998                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
1999             }
2000         }
2001     }
2002
2003     /* Combining the forces for multiple time stepping before the halo exchange, when possible,
2004      * avoids an extra halo exchange (when DD is used) and post-processing step.
2005      */
2006     const bool combineMtsForcesBeforeHaloExchange =
2007             (stepWork.computeForces && fr->useMts && stepWork.computeSlowForces
2008              && (legacyFlags & GMX_FORCE_DO_NOT_NEED_NORMAL_FORCE) != 0
2009              && !(stepWork.computeVirial || simulationWork.useGpuNonbonded || useGpuPmeOnThisRank));
2010     if (combineMtsForcesBeforeHaloExchange)
2011     {
2012         const int numAtoms = havePPDomainDecomposition(cr) ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr;
2013         combineMtsForces(numAtoms,
2014                          force.unpaddedArrayRef(),
2015                          forceView->forceMtsCombined(),
2016                          inputrec->mtsLevels[1].stepFactor);
2017     }
2018
2019     if (havePPDomainDecomposition(cr))
2020     {
2021         /* We are done with the CPU compute.
2022          * We will now communicate the non-local forces.
2023          * If we use a GPU this will overlap with GPU work, so in that case
2024          * we do not close the DD force balancing region here.
2025          */
2026         ddBalanceRegionHandler.closeAfterForceComputationCpu();
2027
2028         if (stepWork.computeForces)
2029         {
2030
2031             if (stepWork.useGpuFHalo)
2032             {
2033                 communicateGpuHaloForces(*cr, domainWork.haveCpuLocalForceWork);
2034             }
2035             else
2036             {
2037                 if (stepWork.useGpuFBufferOps)
2038                 {
2039                     stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
2040                 }
2041
2042                 // Without MTS or with MTS at slow steps with uncombined forces we need to
2043                 // communicate the fast forces
2044                 if (!fr->useMts || !combineMtsForcesBeforeHaloExchange)
2045                 {
2046                     dd_move_f(cr->dd, &forceOutMtsLevel0.forceWithShiftForces(), wcycle);
2047                 }
2048                 // With MTS we need to communicate the slow or combined (in forceOutMtsLevel1) forces
2049                 if (fr->useMts && stepWork.computeSlowForces)
2050                 {
2051                     dd_move_f(cr->dd, &forceOutMtsLevel1->forceWithShiftForces(), wcycle);
2052                 }
2053             }
2054         }
2055     }
2056
2057     // With both nonbonded and PME offloaded a GPU on the same rank, we use
2058     // an alternating wait/reduction scheme.
2059     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded
2060                              && !DOMAINDECOMP(cr) && !stepWork.useGpuFBufferOps);
2061     if (alternateGpuWait)
2062     {
2063         alternatePmeNbGpuWaitReduce(
2064                 fr->nbv.get(), fr->pmedata, forceOutNonbonded, forceOutMtsLevel1, enerd, lambda[efptCOUL], stepWork, wcycle);
2065     }
2066
2067     if (!alternateGpuWait && useGpuPmeOnThisRank)
2068     {
2069         pme_gpu_wait_and_reduce(
2070                 fr->pmedata, stepWork, wcycle, &forceOutMtsLevel1->forceWithVirial(), enerd, lambda[efptCOUL]);
2071     }
2072
2073     /* Wait for local GPU NB outputs on the non-alternating wait path */
2074     if (!alternateGpuWait && stepWork.computeNonbondedForces && simulationWork.useGpuNonbonded)
2075     {
2076         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
2077          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
2078          * but even with a step of 0.1 ms the difference is less than 1%
2079          * of the step time.
2080          */
2081         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
2082         const float waitCycles =
2083                 Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
2084                                             stepWork,
2085                                             AtomLocality::Local,
2086                                             enerd->grpp.ener[egLJSR].data(),
2087                                             enerd->grpp.ener[egCOULSR].data(),
2088                                             forceOutNonbonded->forceWithShiftForces().shiftForces(),
2089                                             wcycle);
2090
2091         if (ddBalanceRegionHandler.useBalancingRegion())
2092         {
2093             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
2094             if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
2095             {
2096                 /* We measured few cycles, it could be that the kernel
2097                  * and transfer finished earlier and there was no actual
2098                  * wait time, only API call overhead.
2099                  * Then the actual time could be anywhere between 0 and
2100                  * cycles_wait_est. We will use half of cycles_wait_est.
2101                  */
2102                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
2103             }
2104             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
2105         }
2106     }
2107
2108     if (fr->nbv->emulateGpu())
2109     {
2110         // NOTE: emulation kernel is not included in the balancing region,
2111         // but emulation mode does not target performance anyway
2112         wallcycle_start_nocount(wcycle, ewcFORCE);
2113         do_nb_verlet(fr,
2114                      ic,
2115                      enerd,
2116                      stepWork,
2117                      InteractionLocality::Local,
2118                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
2119                      step,
2120                      nrnb,
2121                      wcycle);
2122         wallcycle_stop(wcycle, ewcFORCE);
2123     }
2124
2125     // If on GPU PME-PP comms path, receive forces from PME before GPU buffer ops
2126     // TODO refactor this and unify with below default-path call to the same function
2127     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && stepWork.computeSlowForces
2128         && simulationWork.useGpuPmePpCommunication)
2129     {
2130         /* In case of node-splitting, the PP nodes receive the long-range
2131          * forces, virial and energy from the PME nodes here.
2132          */
2133         pme_receive_force_ener(fr,
2134                                cr,
2135                                &forceOutMtsLevel1->forceWithVirial(),
2136                                enerd,
2137                                simulationWork.useGpuPmePpCommunication,
2138                                stepWork.useGpuPmeFReduction,
2139                                wcycle);
2140     }
2141
2142
2143     /* Do the nonbonded GPU (or emulation) force buffer reduction
2144      * on the non-alternating path. */
2145     GMX_ASSERT(!(nonbondedAtMtsLevel1 && stepWork.useGpuFBufferOps),
2146                "The schedule below does not allow for nonbonded MTS with GPU buffer ops");
2147     if (useOrEmulateGpuNb && !alternateGpuWait)
2148     {
2149         if (stepWork.useGpuFBufferOps)
2150         {
2151             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
2152
2153             // Flag to specify whether the CPU force buffer has contributions to
2154             // local atoms. This depends on whether there are CPU-based force tasks
2155             // or when DD is active the halo exchange has resulted in contributions
2156             // from the non-local part.
2157             const bool haveLocalForceContribInCpuBuffer =
2158                     (domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr));
2159
2160             // TODO: move these steps as early as possible:
2161             // - CPU f H2D should be as soon as all CPU-side forces are done
2162             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
2163             //   before the next CPU task that consumes the forces: vsite spread or update)
2164             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
2165             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
2166             //   These should be unified.
2167             if (haveLocalForceContribInCpuBuffer && !stepWork.useGpuFHalo)
2168             {
2169                 // Note: AtomLocality::All is used for the non-DD case because, as in this
2170                 // case copyForcesToGpu() uses a separate stream, it allows overlap of
2171                 // CPU force H2D with GPU force tasks on all streams including those in the
2172                 // local stream which would otherwise be implicit dependencies for the
2173                 // transfer and would not overlap.
2174                 auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
2175
2176                 stateGpu->copyForcesToGpu(forceWithShift, locality);
2177             }
2178
2179             if (stepWork.computeNonbondedForces)
2180             {
2181                 fr->gpuForceReduction[gmx::AtomLocality::Local]->execute();
2182             }
2183
2184             // Copy forces to host if they are needed for update or if virtual sites are enabled.
2185             // If there are vsites, we need to copy forces every step to spread vsite forces on host.
2186             // TODO: When the output flags will be included in step workload, this copy can be combined with the
2187             //       copy call done in sim_utils(...) for the output.
2188             // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
2189             //       they should not be copied in do_md(...) for the output.
2190             if (!simulationWork.useGpuUpdate
2191                 || (simulationWork.useGpuUpdate && DOMAINDECOMP(cr) && haveHostPmePpComms) || vsite)
2192             {
2193                 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
2194                 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
2195             }
2196         }
2197         else if (stepWork.computeNonbondedForces)
2198         {
2199             ArrayRef<gmx::RVec> forceWithShift = forceOutNonbonded->forceWithShiftForces().force();
2200             nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
2201         }
2202     }
2203
2204     launchGpuEndOfStepTasks(
2205             nbv, fr->gpuBonded, fr->pmedata, enerd, *runScheduleWork, useGpuPmeOnThisRank, step, wcycle);
2206
2207     if (DOMAINDECOMP(cr))
2208     {
2209         dd_force_flop_stop(cr->dd, nrnb);
2210     }
2211
2212     const bool haveCombinedMtsForces = (stepWork.computeForces && fr->useMts && stepWork.computeSlowForces
2213                                         && combineMtsForcesBeforeHaloExchange);
2214     if (stepWork.computeForces)
2215     {
2216         postProcessForceWithShiftForces(
2217                 nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutMtsLevel0, vir_force, *mdatoms, *fr, vsite, stepWork);
2218
2219         if (fr->useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2220         {
2221             postProcessForceWithShiftForces(
2222                     nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1, vir_force, *mdatoms, *fr, vsite, stepWork);
2223         }
2224     }
2225
2226     // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
2227     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication
2228         && stepWork.computeSlowForces)
2229     {
2230         /* In case of node-splitting, the PP nodes receive the long-range
2231          * forces, virial and energy from the PME nodes here.
2232          */
2233         pme_receive_force_ener(fr,
2234                                cr,
2235                                &forceOutMtsLevel1->forceWithVirial(),
2236                                enerd,
2237                                simulationWork.useGpuPmePpCommunication,
2238                                false,
2239                                wcycle);
2240     }
2241
2242     if (stepWork.computeForces)
2243     {
2244         /* If we don't use MTS or if we already combined the MTS forces before, we only
2245          * need to post-process one ForceOutputs object here, called forceOutCombined,
2246          * otherwise we have to post-process two outputs and then combine them.
2247          */
2248         ForceOutputs& forceOutCombined = (haveCombinedMtsForces ? forceOutMts.value() : forceOutMtsLevel0);
2249         postProcessForces(
2250                 cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOutCombined, vir_force, mdatoms, fr, vsite, stepWork);
2251
2252         if (fr->useMts && stepWork.computeSlowForces && !haveCombinedMtsForces)
2253         {
2254             postProcessForces(
2255                     cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), forceOutMtsLevel1, vir_force, mdatoms, fr, vsite, stepWork);
2256
2257             combineMtsForces(mdatoms->homenr,
2258                              force.unpaddedArrayRef(),
2259                              forceView->forceMtsCombined(),
2260                              inputrec->mtsLevels[1].stepFactor);
2261         }
2262     }
2263
2264     if (stepWork.computeEnergy)
2265     {
2266         /* Compute the final potential energy terms */
2267         accumulatePotentialEnergies(enerd, lambda, inputrec->fepvals);
2268
2269         if (!EI_TPI(inputrec->eI))
2270         {
2271             checkPotentialEnergyValidity(step, *enerd, *inputrec);
2272         }
2273     }
2274
2275     /* In case we don't have constraints and are using GPUs, the next balancing
2276      * region starts here.
2277      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
2278      * virial calculation and COM pulling, is not thus not included in
2279      * the balance timing, which is ok as most tasks do communication.
2280      */
2281     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
2282 }