src/gromacs/mdlib/sim_util.cpp

   1 /*
   2  * This file is part of the GROMACS molecular simulation package.
   3  *
   4  * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
   5  * Copyright (c) 2001-2004, The GROMACS development team.
   6  * Copyright (c) 2013-2019,2020, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "config.h"
  40
  41 #include <cmath>
  42 #include <cstdint>
  43 #include <cstdio>
  44 #include <cstring>
  45
  46 #include <array>
  47
  48 #include "gromacs/awh/awh.h"
  49 #include "gromacs/domdec/dlbtiming.h"
  50 #include "gromacs/domdec/domdec.h"
  51 #include "gromacs/domdec/domdec_struct.h"
  52 #include "gromacs/domdec/gpuhaloexchange.h"
  53 #include "gromacs/domdec/partition.h"
  54 #include "gromacs/essentialdynamics/edsam.h"
  55 #include "gromacs/ewald/pme.h"
  56 #include "gromacs/ewald/pme_pp.h"
  57 #include "gromacs/ewald/pme_pp_comm_gpu.h"
  58 #include "gromacs/gmxlib/network.h"
  59 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  60 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  61 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  62 #include "gromacs/gpu_utils/gpu_utils.h"
  63 #include "gromacs/imd/imd.h"
  64 #include "gromacs/listed_forces/disre.h"
  65 #include "gromacs/listed_forces/gpubonded.h"
  66 #include "gromacs/listed_forces/listed_forces.h"
  67 #include "gromacs/listed_forces/orires.h"
  68 #include "gromacs/math/arrayrefwithpadding.h"
  69 #include "gromacs/math/functions.h"
  70 #include "gromacs/math/units.h"
  71 #include "gromacs/math/vec.h"
  72 #include "gromacs/math/vecdump.h"
  73 #include "gromacs/mdlib/calcmu.h"
  74 #include "gromacs/mdlib/calcvir.h"
  75 #include "gromacs/mdlib/constr.h"
  76 #include "gromacs/mdlib/dispersioncorrection.h"
  77 #include "gromacs/mdlib/enerdata_utils.h"
  78 #include "gromacs/mdlib/force.h"
  79 #include "gromacs/mdlib/force_flags.h"
  80 #include "gromacs/mdlib/forcerec.h"
  81 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  82 #include "gromacs/mdlib/update.h"
  83 #include "gromacs/mdlib/vsite.h"
  84 #include "gromacs/mdlib/wholemoleculetransform.h"
  85 #include "gromacs/mdtypes/commrec.h"
  86 #include "gromacs/mdtypes/enerdata.h"
  87 #include "gromacs/mdtypes/forceoutput.h"
  88 #include "gromacs/mdtypes/forcerec.h"
  89 #include "gromacs/mdtypes/iforceprovider.h"
  90 #include "gromacs/mdtypes/inputrec.h"
  91 #include "gromacs/mdtypes/md_enums.h"
  92 #include "gromacs/mdtypes/mdatom.h"
  93 #include "gromacs/mdtypes/simulation_workload.h"
  94 #include "gromacs/mdtypes/state.h"
  95 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
  96 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  97 #include "gromacs/nbnxm/nbnxm.h"
  98 #include "gromacs/nbnxm/nbnxm_gpu.h"
  99 #include "gromacs/pbcutil/ishift.h"
 100 #include "gromacs/pbcutil/pbc.h"
 101 #include "gromacs/pulling/pull.h"
 102 #include "gromacs/pulling/pull_rotation.h"
 103 #include "gromacs/timing/cyclecounter.h"
 104 #include "gromacs/timing/gpu_timing.h"
 105 #include "gromacs/timing/wallcycle.h"
 106 #include "gromacs/timing/wallcyclereporting.h"
 107 #include "gromacs/timing/walltime_accounting.h"
 108 #include "gromacs/topology/topology.h"
 109 #include "gromacs/utility/arrayref.h"
 110 #include "gromacs/utility/basedefinitions.h"
 111 #include "gromacs/utility/cstringutil.h"
 112 #include "gromacs/utility/exceptions.h"
 113 #include "gromacs/utility/fatalerror.h"
 114 #include "gromacs/utility/fixedcapacityvector.h"
 115 #include "gromacs/utility/gmxassert.h"
 116 #include "gromacs/utility/gmxmpi.h"
 117 #include "gromacs/utility/logger.h"
 118 #include "gromacs/utility/smalloc.h"
 119 #include "gromacs/utility/strconvert.h"
 120 #include "gromacs/utility/sysinfo.h"
 121
 122 using gmx::ArrayRef;
 123 using gmx::AtomLocality;
 124 using gmx::DomainLifetimeWorkload;
 125 using gmx::ForceOutputs;
 126 using gmx::ForceWithShiftForces;
 127 using gmx::InteractionLocality;
 128 using gmx::RVec;
 129 using gmx::SimulationWorkload;
 130 using gmx::StepWorkload;
 131
 132 // TODO: this environment variable allows us to verify before release
 133 // that on less common architectures the total cost of polling is not larger than
 134 // a blocking wait (so polling does not introduce overhead when the static
 135 // PME-first ordering would suffice).
 136 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 137
 138 static void sum_forces(ArrayRef<RVec> f, ArrayRef<const RVec> forceToAdd)
 139 {
 140     GMX_ASSERT(f.size() >= forceToAdd.size(), "Accumulation buffer should be sufficiently large");
 141     const int end = forceToAdd.size();
 142
 143     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 144 #pragma omp parallel for num_threads(nt) schedule(static)
 145     for (int i = 0; i < end; i++)
 146     {
 147         rvec_inc(f[i], forceToAdd[i]);
 148     }
 149 }
 150
 151 static void calc_virial(int                              start,
 152                         int                              homenr,
 153                         const rvec                       x[],
 154                         const gmx::ForceWithShiftForces& forceWithShiftForces,
 155                         tensor                           vir_part,
 156                         const matrix                     box,
 157                         t_nrnb*                          nrnb,
 158                         const t_forcerec*                fr,
 159                         PbcType                          pbcType)
 160 {
 161     /* The short-range virial from surrounding boxes */
 162     const rvec* fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
 163     calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, pbcType == PbcType::Screw, box);
 164     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 165
 166     /* Calculate partial virial, for local atoms only, based on short range.
 167      * Total virial is computed in global_stat, called from do_md
 168      */
 169     const rvec* f = as_rvec_array(forceWithShiftForces.force().data());
 170     f_calc_vir(start, start + homenr, x, f, vir_part, box);
 171     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 172
 173     if (debug)
 174     {
 175         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 176     }
 177 }
 178
 179 static void pull_potential_wrapper(const t_commrec*               cr,
 180                                    const t_inputrec*              ir,
 181                                    const matrix                   box,
 182                                    gmx::ArrayRef<const gmx::RVec> x,
 183                                    gmx::ForceWithVirial*          force,
 184                                    const t_mdatoms*               mdatoms,
 185                                    gmx_enerdata_t*                enerd,
 186                                    pull_t*                        pull_work,
 187                                    const real*                    lambda,
 188                                    double                         t,
 189                                    gmx_wallcycle_t                wcycle)
 190 {
 191     t_pbc pbc;
 192     real  dvdl;
 193
 194     /* Calculate the center of mass forces, this requires communication,
 195      * which is why pull_potential is called close to other communication.
 196      */
 197     wallcycle_start(wcycle, ewcPULLPOT);
 198     set_pbc(&pbc, ir->pbcType, box);
 199     dvdl = 0;
 200     enerd->term[F_COM_PULL] +=
 201             pull_potential(pull_work, mdatoms->massT, &pbc, cr, t, lambda[efptRESTRAINT],
 202                            as_rvec_array(x.data()), force, &dvdl);
 203     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 204     wallcycle_stop(wcycle, ewcPULLPOT);
 205 }
 206
 207 static void pme_receive_force_ener(t_forcerec*           fr,
 208                                    const t_commrec*      cr,
 209                                    gmx::ForceWithVirial* forceWithVirial,
 210                                    gmx_enerdata_t*       enerd,
 211                                    bool                  useGpuPmePpComms,
 212                                    bool                  receivePmeForceToGpu,
 213                                    gmx_wallcycle_t       wcycle)
 214 {
 215     real  e_q, e_lj, dvdl_q, dvdl_lj;
 216     float cycles_ppdpme, cycles_seppme;
 217
 218     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 219     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 220
 221     /* In case of node-splitting, the PP nodes receive the long-range
 222      * forces, virial and energy from the PME nodes here.
 223      */
 224     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 225     dvdl_q  = 0;
 226     dvdl_lj = 0;
 227     gmx_pme_receive_f(fr->pmePpCommGpu.get(), cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 228                       useGpuPmePpComms, receivePmeForceToGpu, &cycles_seppme);
 229     enerd->term[F_COUL_RECIP] += e_q;
 230     enerd->term[F_LJ_RECIP] += e_lj;
 231     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 232     enerd->dvdl_lin[efptVDW] += dvdl_lj;
 233
 234     if (wcycle)
 235     {
 236         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 237     }
 238     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 239 }
 240
 241 static void print_large_forces(FILE*                fp,
 242                                const t_mdatoms*     md,
 243                                const t_commrec*     cr,
 244                                int64_t              step,
 245                                real                 forceTolerance,
 246                                ArrayRef<const RVec> x,
 247                                ArrayRef<const RVec> f)
 248 {
 249     real       force2Tolerance = gmx::square(forceTolerance);
 250     gmx::index numNonFinite    = 0;
 251     for (int i = 0; i < md->homenr; i++)
 252     {
 253         real force2    = norm2(f[i]);
 254         bool nonFinite = !std::isfinite(force2);
 255         if (force2 >= force2Tolerance || nonFinite)
 256         {
 257             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n", step,
 258                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 259         }
 260         if (nonFinite)
 261         {
 262             numNonFinite++;
 263         }
 264     }
 265     if (numNonFinite > 0)
 266     {
 267         /* Note that with MPI this fatal call on one rank might interrupt
 268          * the printing on other ranks. But we can only avoid that with
 269          * an expensive MPI barrier that we would need at each step.
 270          */
 271         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 272     }
 273 }
 274
 275 //! When necessary, spreads forces on vsites and computes the virial for \p forceOutputs->forceWithShiftForces()
 276 static void postProcessForceWithShiftForces(t_nrnb*                   nrnb,
 277                                             gmx_wallcycle_t           wcycle,
 278                                             const matrix              box,
 279                                             ArrayRef<const RVec>      x,
 280                                             ForceOutputs*             forceOutputs,
 281                                             tensor                    vir_force,
 282                                             const t_mdatoms&          mdatoms,
 283                                             const t_forcerec&         fr,
 284                                             gmx::VirtualSitesHandler* vsite,
 285                                             const StepWorkload&       stepWork)
 286 {
 287     ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
 288
 289     /* If we have NoVirSum forces, but we do not calculate the virial,
 290      * we later sum the forceWithShiftForces buffer together with
 291      * the noVirSum buffer and spread the combined vsite forces at once.
 292      */
 293     if (vsite && (!forceOutputs->haveForceWithVirial() || stepWork.computeVirial))
 294     {
 295         using VirialHandling = gmx::VirtualSitesHandler::VirialHandling;
 296
 297         auto                 f      = forceWithShiftForces.force();
 298         auto                 fshift = forceWithShiftForces.shiftForces();
 299         const VirialHandling virialHandling =
 300                 (stepWork.computeVirial ? VirialHandling::Pbc : VirialHandling::None);
 301         vsite->spreadForces(x, f, virialHandling, fshift, nullptr, nrnb, box, wcycle);
 302         forceWithShiftForces.haveSpreadVsiteForces() = true;
 303     }
 304
 305     if (stepWork.computeVirial)
 306     {
 307         /* Calculation of the virial must be done after vsites! */
 308         calc_virial(0, mdatoms.homenr, as_rvec_array(x.data()), forceWithShiftForces, vir_force,
 309                     box, nrnb, &fr, fr.pbcType);
 310     }
 311 }
 312
 313 //! Spread, compute virial for and sum forces, when necessary
 314 static void postProcessForces(const t_commrec*          cr,
 315                               int64_t                   step,
 316                               t_nrnb*                   nrnb,
 317                               gmx_wallcycle_t           wcycle,
 318                               const matrix              box,
 319                               ArrayRef<const RVec>      x,
 320                               ForceOutputs*             forceOutputs,
 321                               tensor                    vir_force,
 322                               const t_mdatoms*          mdatoms,
 323                               const t_forcerec*         fr,
 324                               gmx::VirtualSitesHandler* vsite,
 325                               const StepWorkload&       stepWork)
 326 {
 327     // Extract the final output force buffer, which is also the buffer for forces with shift forces
 328     ArrayRef<RVec> f = forceOutputs->forceWithShiftForces().force();
 329
 330     if (forceOutputs->haveForceWithVirial())
 331     {
 332         auto& forceWithVirial = forceOutputs->forceWithVirial();
 333
 334         if (vsite)
 335         {
 336             /* Spread the mesh force on virtual sites to the other particles...
 337              * This is parallellized. MPI communication is performed
 338              * if the constructing atoms aren't local.
 339              */
 340             GMX_ASSERT(!stepWork.computeVirial || f.data() != forceWithVirial.force_.data(),
 341                        "We need separate force buffers for shift and virial forces when "
 342                        "computing the virial");
 343             GMX_ASSERT(!stepWork.computeVirial
 344                                || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
 345                        "We should spread the force with shift forces separately when computing "
 346                        "the virial");
 347             const gmx::VirtualSitesHandler::VirialHandling virialHandling =
 348                     (stepWork.computeVirial ? gmx::VirtualSitesHandler::VirialHandling::NonLinear
 349                                             : gmx::VirtualSitesHandler::VirialHandling::None);
 350             matrix virial = { { 0 } };
 351             vsite->spreadForces(x, forceWithVirial.force_, virialHandling, {}, virial, nrnb, box, wcycle);
 352             forceWithVirial.addVirialContribution(virial);
 353         }
 354
 355         if (stepWork.computeVirial)
 356         {
 357             /* Now add the forces, this is local */
 358             sum_forces(f, forceWithVirial.force_);
 359
 360             /* Add the direct virial contributions */
 361             GMX_ASSERT(
 362                     forceWithVirial.computeVirial_,
 363                     "forceWithVirial should request virial computation when we request the virial");
 364             m_add(vir_force, forceWithVirial.getVirial(), vir_force);
 365
 366             if (debug)
 367             {
 368                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 369             }
 370         }
 371     }
 372     else
 373     {
 374         GMX_ASSERT(vsite == nullptr || forceOutputs->forceWithShiftForces().haveSpreadVsiteForces(),
 375                    "We should have spread the vsite forces (earlier)");
 376     }
 377
 378     if (fr->print_force >= 0)
 379     {
 380         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 381     }
 382 }
 383
 384 static void do_nb_verlet(t_forcerec*                fr,
 385                          const interaction_const_t* ic,
 386                          gmx_enerdata_t*            enerd,
 387                          const StepWorkload&        stepWork,
 388                          const InteractionLocality  ilocality,
 389                          const int                  clearF,
 390                          const int64_t              step,
 391                          t_nrnb*                    nrnb,
 392                          gmx_wallcycle_t            wcycle)
 393 {
 394     if (!stepWork.computeNonbondedForces)
 395     {
 396         /* skip non-bonded calculation */
 397         return;
 398     }
 399
 400     nonbonded_verlet_t* nbv = fr->nbv.get();
 401
 402     /* GPU kernel launch overhead is already timed separately */
 403     if (fr->cutoff_scheme != ecutsVERLET)
 404     {
 405         gmx_incons("Invalid cut-off scheme passed!");
 406     }
 407
 408     if (!nbv->useGpu())
 409     {
 410         /* When dynamic pair-list  pruning is requested, we need to prune
 411          * at nstlistPrune steps.
 412          */
 413         if (nbv->isDynamicPruningStepCpu(step))
 414         {
 415             /* Prune the pair-list beyond fr->ic->rlistPrune using
 416              * the current coordinates of the atoms.
 417              */
 418             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 419             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 420             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 421         }
 422     }
 423
 424     nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
 425 }
 426
 427 static inline void clearRVecs(ArrayRef<RVec> v, const bool useOpenmpThreading)
 428 {
 429     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, v.ssize());
 430
 431     /* Note that we would like to avoid this conditional by putting it
 432      * into the omp pragma instead, but then we still take the full
 433      * omp parallel for overhead (at least with gcc5).
 434      */
 435     if (!useOpenmpThreading || nth == 1)
 436     {
 437         for (RVec& elem : v)
 438         {
 439             clear_rvec(elem);
 440         }
 441     }
 442     else
 443     {
 444 #pragma omp parallel for num_threads(nth) schedule(static)
 445         for (gmx::index i = 0; i < v.ssize(); i++)
 446         {
 447             clear_rvec(v[i]);
 448         }
 449     }
 450 }
 451
 452 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 453  *
 454  * \param groupOptions  Group options, containing T-coupling options
 455  */
 456 static real averageKineticEnergyEstimate(const t_grpopts& groupOptions)
 457 {
 458     real nrdfCoupled   = 0;
 459     real nrdfUncoupled = 0;
 460     real kineticEnergy = 0;
 461     for (int g = 0; g < groupOptions.ngtc; g++)
 462     {
 463         if (groupOptions.tau_t[g] >= 0)
 464         {
 465             nrdfCoupled += groupOptions.nrdf[g];
 466             kineticEnergy += groupOptions.nrdf[g] * 0.5 * groupOptions.ref_t[g] * BOLTZ;
 467         }
 468         else
 469         {
 470             nrdfUncoupled += groupOptions.nrdf[g];
 471         }
 472     }
 473
 474     /* This conditional with > also catches nrdf=0 */
 475     if (nrdfCoupled > nrdfUncoupled)
 476     {
 477         return kineticEnergy * (nrdfCoupled + nrdfUncoupled) / nrdfCoupled;
 478     }
 479     else
 480     {
 481         return 0;
 482     }
 483 }
 484
 485 /*! \brief This routine checks that the potential energy is finite.
 486  *
 487  * Always checks that the potential energy is finite. If step equals
 488  * inputrec.init_step also checks that the magnitude of the potential energy
 489  * is reasonable. Terminates with a fatal error when a check fails.
 490  * Note that passing this check does not guarantee finite forces,
 491  * since those use slightly different arithmetics. But in most cases
 492  * there is just a narrow coordinate range where forces are not finite
 493  * and energies are finite.
 494  *
 495  * \param[in] step      The step number, used for checking and printing
 496  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to
 497  * enerd.term[F_EPOT] before calling this routine \param[in] inputrec  The input record
 498  */
 499 static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
 500 {
 501     /* Threshold valid for comparing absolute potential energy against
 502      * the kinetic energy. Normally one should not consider absolute
 503      * potential energy values, but with a factor of one million
 504      * we should never get false positives.
 505      */
 506     constexpr real c_thresholdFactor = 1e6;
 507
 508     bool energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 509     real averageKineticEnergy = 0;
 510     /* We only check for large potential energy at the initial step,
 511      * because that is by far the most likely step for this too occur
 512      * and because computing the average kinetic energy is not free.
 513      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 514      * before they become NaN.
 515      */
 516     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 517     {
 518         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 519     }
 520
 521     if (energyIsNotFinite
 522         || (averageKineticEnergy > 0 && enerd.term[F_EPOT] > c_thresholdFactor * averageKineticEnergy))
 523     {
 524         gmx_fatal(
 525                 FARGS,
 526                 "Step %" PRId64
 527                 ": The total potential energy is %g, which is %s. The LJ and electrostatic "
 528                 "contributions to the energy are %g and %g, respectively. A %s potential energy "
 529                 "can be caused by overlapping interactions in bonded interactions or very large%s "
 530                 "coordinate values. Usually this is caused by a badly- or non-equilibrated initial "
 531                 "configuration, incorrect interactions or parameters in the topology.",
 532                 step, enerd.term[F_EPOT], energyIsNotFinite ? "not finite" : "extremely high",
 533                 enerd.term[F_LJ], enerd.term[F_COUL_SR],
 534                 energyIsNotFinite ? "non-finite" : "very high", energyIsNotFinite ? " or Nan" : "");
 535     }
 536 }
 537
 538 /*! \brief Return true if there are special forces computed this step.
 539  *
 540  * The conditionals exactly correspond to those in computeSpecialForces().
 541  */
 542 static bool haveSpecialForces(const t_inputrec&          inputrec,
 543                               const gmx::ForceProviders& forceProviders,
 544                               const pull_t*              pull_work,
 545                               const bool                 computeForces,
 546                               const gmx_edsam*           ed)
 547 {
 548
 549     return ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
 550             (inputrec.bPull && pull_have_potential(pull_work)) ||   // pull
 551             inputrec.bRot ||                                        // enforced rotation
 552             (ed != nullptr) ||                                      // flooding
 553             (inputrec.bIMD && computeForces));                      // IMD
 554 }
 555
 556 /*! \brief Compute forces and/or energies for special algorithms
 557  *
 558  * The intention is to collect all calls to algorithms that compute
 559  * forces on local atoms only and that do not contribute to the local
 560  * virial sum (but add their virial contribution separately).
 561  * Eventually these should likely all become ForceProviders.
 562  * Within this function the intention is to have algorithms that do
 563  * global communication at the end, so global barriers within the MD loop
 564  * are as close together as possible.
 565  *
 566  * \param[in]     fplog            The log file
 567  * \param[in]     cr               The communication record
 568  * \param[in]     inputrec         The input record
 569  * \param[in]     awh              The Awh module (nullptr if none in use).
 570  * \param[in]     enforcedRotation Enforced rotation module.
 571  * \param[in]     imdSession       The IMD session
 572  * \param[in]     pull_work        The pull work structure.
 573  * \param[in]     step             The current MD step
 574  * \param[in]     t                The current time
 575  * \param[in,out] wcycle           Wallcycle accounting struct
 576  * \param[in,out] forceProviders   Pointer to a list of force providers
 577  * \param[in]     box              The unit cell
 578  * \param[in]     x                The coordinates
 579  * \param[in]     mdatoms          Per atom properties
 580  * \param[in]     lambda           Array of free-energy lambda values
 581  * \param[in]     stepWork         Step schedule flags
 582  * \param[in,out] forceWithVirial  Force and virial buffers
 583  * \param[in,out] enerd            Energy buffer
 584  * \param[in,out] ed               Essential dynamics pointer
 585  * \param[in]     didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
 586  *
 587  * \todo Remove didNeighborSearch, which is used incorrectly.
 588  * \todo Convert all other algorithms called here to ForceProviders.
 589  */
 590 static void computeSpecialForces(FILE*                          fplog,
 591                                  const t_commrec*               cr,
 592                                  const t_inputrec*              inputrec,
 593                                  gmx::Awh*                      awh,
 594                                  gmx_enfrot*                    enforcedRotation,
 595                                  gmx::ImdSession*               imdSession,
 596                                  pull_t*                        pull_work,
 597                                  int64_t                        step,
 598                                  double                         t,
 599                                  gmx_wallcycle_t                wcycle,
 600                                  gmx::ForceProviders*           forceProviders,
 601                                  const matrix                   box,
 602                                  gmx::ArrayRef<const gmx::RVec> x,
 603                                  const t_mdatoms*               mdatoms,
 604                                  gmx::ArrayRef<const real>      lambda,
 605                                  const StepWorkload&            stepWork,
 606                                  gmx::ForceWithVirial*          forceWithVirial,
 607                                  gmx_enerdata_t*                enerd,
 608                                  gmx_edsam*                     ed,
 609                                  bool                           didNeighborSearch)
 610 {
 611     /* NOTE: Currently all ForceProviders only provide forces.
 612      *       When they also provide energies, remove this conditional.
 613      */
 614     if (stepWork.computeForces)
 615     {
 616         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 617         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 618
 619         /* Collect forces from modules */
 620         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 621     }
 622
 623     if (inputrec->bPull && pull_have_potential(pull_work))
 624     {
 625         pull_potential_wrapper(cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work,
 626                                lambda.data(), t, wcycle);
 627
 628         if (awh)
 629         {
 630             enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(
 631                     inputrec->pbcType, mdatoms->massT, box, forceWithVirial, t, step, wcycle, fplog);
 632         }
 633     }
 634
 635     rvec* f = as_rvec_array(forceWithVirial->force_.data());
 636
 637     /* Add the forces from enforced rotation potentials (if any) */
 638     if (inputrec->bRot)
 639     {
 640         wallcycle_start(wcycle, ewcROTadd);
 641         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 642         wallcycle_stop(wcycle, ewcROTadd);
 643     }
 644
 645     if (ed)
 646     {
 647         /* Note that since init_edsam() is called after the initialization
 648          * of forcerec, edsam doesn't request the noVirSum force buffer.
 649          * Thus if no other algorithm (e.g. PME) requires it, the forces
 650          * here will contribute to the virial.
 651          */
 652         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
 653     }
 654
 655     /* Add forces from interactive molecular dynamics (IMD), if any */
 656     if (inputrec->bIMD && stepWork.computeForces)
 657     {
 658         imdSession->applyForces(f);
 659     }
 660 }
 661
 662 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 663  *
 664  * \param[in]  pmedata              The PME structure
 665  * \param[in]  box                  The box matrix
 666  * \param[in]  stepWork             Step schedule flags
 667  * \param[in]  xReadyOnDevice       Event synchronizer indicating that the coordinates are ready in the device memory.
 668  * \param[in]  lambdaQ              The Coulomb lambda of the current state.
 669  * \param[in]  wcycle               The wallcycle structure
 670  */
 671 static inline void launchPmeGpuSpread(gmx_pme_t*            pmedata,
 672                                       const matrix          box,
 673                                       const StepWorkload&   stepWork,
 674                                       GpuEventSynchronizer* xReadyOnDevice,
 675                                       const real            lambdaQ,
 676                                       gmx_wallcycle_t       wcycle)
 677 {
 678     pme_gpu_prepare_computation(pmedata, box, wcycle, stepWork);
 679     pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle, lambdaQ);
 680 }
 681
 682 /*! \brief Launch the FFT and gather stages of PME GPU
 683  *
 684  * This function only implements setting the output forces (no accumulation).
 685  *
 686  * \param[in]  pmedata        The PME structure
 687  * \param[in]  lambdaQ        The Coulomb lambda of the current system state.
 688  * \param[in]  wcycle         The wallcycle structure
 689  * \param[in]  stepWork       Step schedule flags
 690  */
 691 static void launchPmeGpuFftAndGather(gmx_pme_t*               pmedata,
 692                                      const real               lambdaQ,
 693                                      gmx_wallcycle_t          wcycle,
 694                                      const gmx::StepWorkload& stepWork)
 695 {
 696     pme_gpu_launch_complex_transforms(pmedata, wcycle, stepWork);
 697     pme_gpu_launch_gather(pmedata, wcycle, lambdaQ);
 698 }
 699
 700 /*! \brief
 701  *  Polling wait for either of the PME or nonbonded GPU tasks.
 702  *
 703  * Instead of a static order in waiting for GPU tasks, this function
 704  * polls checking which of the two tasks completes first, and does the
 705  * associated force buffer reduction overlapped with the other task.
 706  * By doing that, unlike static scheduling order, it can always overlap
 707  * one of the reductions, regardless of the GPU task completion order.
 708  *
 709  * \param[in]     nbv              Nonbonded verlet structure
 710  * \param[in,out] pmedata          PME module data
 711  * \param[in,out] forceOutputs     Output buffer for the forces and virial
 712  * \param[in,out] enerd            Energy data structure results are reduced into
 713  * \param[in]     lambdaQ          The Coulomb lambda of the current system state.
 714  * \param[in]     stepWork         Step schedule flags
 715  * \param[in]     wcycle           The wallcycle structure
 716  */
 717 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
 718                                         gmx_pme_t*          pmedata,
 719                                         gmx::ForceOutputs*  forceOutputs,
 720                                         gmx_enerdata_t*     enerd,
 721                                         const real          lambdaQ,
 722                                         const StepWorkload& stepWork,
 723                                         gmx_wallcycle_t     wcycle)
 724 {
 725     bool isPmeGpuDone = false;
 726     bool isNbGpuDone  = false;
 727
 728
 729     gmx::ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
 730     gmx::ForceWithVirial&      forceWithVirial      = forceOutputs->forceWithVirial();
 731
 732     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 733
 734     while (!isPmeGpuDone || !isNbGpuDone)
 735     {
 736         if (!isPmeGpuDone)
 737         {
 738             GpuTaskCompletion completionType =
 739                     (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 740             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, stepWork, wcycle, &forceWithVirial,
 741                                                    enerd, lambdaQ, completionType);
 742         }
 743
 744         if (!isNbGpuDone)
 745         {
 746             GpuTaskCompletion completionType =
 747                     (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 748             isNbGpuDone = Nbnxm::gpu_try_finish_task(
 749                     nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
 750                     enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(),
 751                     completionType, wcycle);
 752
 753             if (isNbGpuDone)
 754             {
 755                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShiftForces.force());
 756             }
 757         }
 758     }
 759 }
 760
 761 /*! \brief Set up the different force buffers; also does clearing.
 762  *
 763  * \param[in] forceHelperBuffers  Helper force buffers
 764  * \param[in] pull_work The pull work object.
 765  * \param[in] inputrec  input record
 766  * \param[in] force     force array
 767  * \param[in] stepWork  Step schedule flags
 768  * \param[out] wcycle   wallcycle recording structure
 769  *
 770  * \returns             Cleared force output structure
 771  */
 772 static ForceOutputs setupForceOutputs(ForceHelperBuffers*                 forceHelperBuffers,
 773                                       pull_t*                             pull_work,
 774                                       const t_inputrec&                   inputrec,
 775                                       gmx::ArrayRefWithPadding<gmx::RVec> force,
 776                                       const StepWorkload&                 stepWork,
 777                                       gmx_wallcycle_t                     wcycle)
 778 {
 779     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 780
 781     /* NOTE: We assume fr->shiftForces is all zeros here */
 782     gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial,
 783                                                    forceHelperBuffers->shiftForces());
 784
 785     if (stepWork.computeForces)
 786     {
 787         /* Clear the short- and long-range forces */
 788         clearRVecs(forceWithShiftForces.force(), true);
 789
 790         /* Clear the shift forces */
 791         clearRVecs(forceWithShiftForces.shiftForces(), false);
 792     }
 793
 794     /* If we need to compute the virial, we might need a separate
 795      * force buffer for algorithms for which the virial is calculated
 796      * directly, such as PME. Otherwise, forceWithVirial uses the
 797      * the same force (f in legacy calls) buffer as other algorithms.
 798      */
 799     const bool useSeparateForceWithVirialBuffer =
 800             (stepWork.computeForces
 801              && (stepWork.computeVirial && forceHelperBuffers->haveDirectVirialContributions()));
 802     /* forceWithVirial uses the local atom range only */
 803     gmx::ForceWithVirial forceWithVirial(
 804             useSeparateForceWithVirialBuffer ? forceHelperBuffers->forceBufferForDirectVirialContributions()
 805                                              : force.unpaddedArrayRef(),
 806             stepWork.computeVirial);
 807
 808     if (useSeparateForceWithVirialBuffer)
 809     {
 810         /* TODO: update comment
 811          * We only compute forces on local atoms. Note that vsites can
 812          * spread to non-local atoms, but that part of the buffer is
 813          * cleared separately in the vsite spreading code.
 814          */
 815         clearRVecs(forceWithVirial.force_, true);
 816     }
 817
 818     if (inputrec.bPull && pull_have_constraint(pull_work))
 819     {
 820         clear_pull_forces(pull_work);
 821     }
 822
 823     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 824
 825     return ForceOutputs(forceWithShiftForces, forceHelperBuffers->haveDirectVirialContributions(),
 826                         forceWithVirial);
 827 }
 828
 829
 830 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
 831  */
 832 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec&         inputrec,
 833                                                           const t_forcerec&         fr,
 834                                                           const pull_t*             pull_work,
 835                                                           const gmx_edsam*          ed,
 836                                                           const t_mdatoms&          mdatoms,
 837                                                           const SimulationWorkload& simulationWork,
 838                                                           const StepWorkload&       stepWork)
 839 {
 840     DomainLifetimeWorkload domainWork;
 841     // Note that haveSpecialForces is constant over the whole run
 842     domainWork.haveSpecialForces =
 843             haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
 844     domainWork.haveCpuBondedWork  = fr.listedForces->haveCpuBondeds();
 845     domainWork.haveGpuBondedWork  = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
 846     domainWork.haveRestraintsWork = fr.listedForces->haveRestraints();
 847     domainWork.haveCpuListedForceWork = fr.listedForces->haveCpuListedForces();
 848     // Note that haveFreeEnergyWork is constant over the whole run
 849     domainWork.haveFreeEnergyWork = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
 850     // We assume we have local force work if there are CPU
 851     // force tasks including PME or nonbondeds.
 852     domainWork.haveCpuLocalForceWork =
 853             domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
 854             || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
 855             || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
 856
 857     return domainWork;
 858 }
 859
 860 /*! \brief Set up force flag stuct from the force bitmask.
 861  *
 862  * \param[in]      legacyFlags          Force bitmask flags used to construct the new flags
 863  * \param[in]      isNonbondedOn        Global override, if false forces to turn off all nonbonded calculation.
 864  * \param[in]      simulationWork       Simulation workload description.
 865  * \param[in]      rankHasPmeDuty       If this rank computes PME.
 866  *
 867  * \returns New Stepworkload description.
 868  */
 869 static StepWorkload setupStepWorkload(const int                 legacyFlags,
 870                                       const bool                isNonbondedOn,
 871                                       const SimulationWorkload& simulationWork,
 872                                       const bool                rankHasPmeDuty)
 873 {
 874     StepWorkload flags;
 875     flags.stateChanged           = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
 876     flags.haveDynamicBox         = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
 877     flags.doNeighborSearch       = ((legacyFlags & GMX_FORCE_NS) != 0);
 878     flags.computeVirial          = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
 879     flags.computeEnergy          = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
 880     flags.computeForces          = ((legacyFlags & GMX_FORCE_FORCES) != 0);
 881     flags.computeListedForces    = ((legacyFlags & GMX_FORCE_LISTED) != 0);
 882     flags.computeNonbondedForces = ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && isNonbondedOn;
 883     flags.computeDhdl            = ((legacyFlags & GMX_FORCE_DHDL) != 0);
 884
 885     if (simulationWork.useGpuBufferOps)
 886     {
 887         GMX_ASSERT(simulationWork.useGpuNonbonded,
 888                    "Can only offload buffer ops if nonbonded computation is also offloaded");
 889     }
 890     flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
 891     // on virial steps the CPU reduction path is taken
 892     flags.useGpuFBufferOps    = simulationWork.useGpuBufferOps && !flags.computeVirial;
 893     flags.useGpuPmeFReduction = flags.useGpuFBufferOps
 894                                 && (simulationWork.useGpuPme
 895                                     && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication));
 896
 897     return flags;
 898 }
 899
 900
 901 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
 902  *
 903  * TODO: eliminate \p useGpuPmeOnThisRank when this is
 904  * incorporated in DomainLifetimeWorkload.
 905  */
 906 static void launchGpuEndOfStepTasks(nonbonded_verlet_t*               nbv,
 907                                     gmx::GpuBonded*                   gpuBonded,
 908                                     gmx_pme_t*                        pmedata,
 909                                     gmx_enerdata_t*                   enerd,
 910                                     const gmx::MdrunScheduleWorkload& runScheduleWork,
 911                                     bool                              useGpuPmeOnThisRank,
 912                                     int64_t                           step,
 913                                     gmx_wallcycle_t                   wcycle)
 914 {
 915     if (runScheduleWork.simulationWork.useGpuNonbonded)
 916     {
 917         /* Launch pruning before buffer clearing because the API overhead of the
 918          * clear kernel launches can leave the GPU idle while it could be running
 919          * the prune kernel.
 920          */
 921         if (nbv->isDynamicPruningStepGpu(step))
 922         {
 923             nbv->dispatchPruneKernelGpu(step);
 924         }
 925
 926         /* now clear the GPU outputs while we finish the step on the CPU */
 927         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 928         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 929         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
 930         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 931         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 932     }
 933
 934     if (useGpuPmeOnThisRank)
 935     {
 936         pme_gpu_reinit_computation(pmedata, wcycle);
 937     }
 938
 939     if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
 940     {
 941         // in principle this should be included in the DD balancing region,
 942         // but generally it is infrequent so we'll omit it for the sake of
 943         // simpler code
 944         gpuBonded->waitAccumulateEnergyTerms(enerd);
 945
 946         gpuBonded->clearEnergies();
 947     }
 948 }
 949
 950 //! \brief Data structure to hold dipole-related data and staging arrays
 951 struct DipoleData
 952 {
 953     //! Dipole staging for fast summing over MPI
 954     gmx::DVec muStaging[2] = { { 0.0, 0.0, 0.0 } };
 955     //! Dipole staging for states A and B (index 0 and 1 resp.)
 956     gmx::RVec muStateAB[2] = { { 0.0_real, 0.0_real, 0.0_real } };
 957 };
 958
 959
 960 static void reduceAndUpdateMuTot(DipoleData*                   dipoleData,
 961                                  const t_commrec*              cr,
 962                                  const bool                    haveFreeEnergy,
 963                                  gmx::ArrayRef<const real>     lambda,
 964                                  rvec                          muTotal,
 965                                  const DDBalanceRegionHandler& ddBalanceRegionHandler)
 966 {
 967     if (PAR(cr))
 968     {
 969         gmx_sumd(2 * DIM, dipoleData->muStaging[0], cr);
 970         ddBalanceRegionHandler.reopenRegionCpu();
 971     }
 972     for (int i = 0; i < 2; i++)
 973     {
 974         for (int j = 0; j < DIM; j++)
 975         {
 976             dipoleData->muStateAB[i][j] = dipoleData->muStaging[i][j];
 977         }
 978     }
 979
 980     if (!haveFreeEnergy)
 981     {
 982         copy_rvec(dipoleData->muStateAB[0], muTotal);
 983     }
 984     else
 985     {
 986         for (int j = 0; j < DIM; j++)
 987         {
 988             muTotal[j] = (1.0 - lambda[efptCOUL]) * dipoleData->muStateAB[0][j]
 989                          + lambda[efptCOUL] * dipoleData->muStateAB[1][j];
 990         }
 991     }
 992 }
 993
 994 void do_force(FILE*                               fplog,
 995               const t_commrec*                    cr,
 996               const gmx_multisim_t*               ms,
 997               const t_inputrec*                   inputrec,
 998               gmx::Awh*                           awh,
 999               gmx_enfrot*                         enforcedRotation,
1000               gmx::ImdSession*                    imdSession,
1001               pull_t*                             pull_work,
1002               int64_t                             step,
1003               t_nrnb*                             nrnb,
1004               gmx_wallcycle_t                     wcycle,
1005               const gmx_localtop_t*               top,
1006               const matrix                        box,
1007               gmx::ArrayRefWithPadding<gmx::RVec> x,
1008               history_t*                          hist,
1009               gmx::ArrayRefWithPadding<gmx::RVec> force,
1010               tensor                              vir_force,
1011               const t_mdatoms*                    mdatoms,
1012               gmx_enerdata_t*                     enerd,
1013               gmx::ArrayRef<const real>           lambda,
1014               t_forcerec*                         fr,
1015               gmx::MdrunScheduleWorkload*         runScheduleWork,
1016               gmx::VirtualSitesHandler*           vsite,
1017               rvec                                muTotal,
1018               double                              t,
1019               gmx_edsam*                          ed,
1020               int                                 legacyFlags,
1021               const DDBalanceRegionHandler&       ddBalanceRegionHandler)
1022 {
1023     GMX_ASSERT(force.unpaddedArrayRef().ssize() >= fr->natoms_force_constr,
1024                "The size of the force buffer should be at least the number of atoms to compute "
1025                "forces for");
1026
1027     nonbonded_verlet_t*          nbv      = fr->nbv.get();
1028     interaction_const_t*         ic       = fr->ic;
1029     gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
1030
1031     const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
1032
1033
1034     runScheduleWork->stepWork    = setupStepWorkload(legacyFlags, fr->bNonbonded, simulationWork,
1035                                                   thisRankHasDuty(cr, DUTY_PME));
1036     const StepWorkload& stepWork = runScheduleWork->stepWork;
1037
1038
1039     const bool useGpuPmeOnThisRank = simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME);
1040
1041     /* At a search step we need to start the first balancing region
1042      * somewhere early inside the step after communication during domain
1043      * decomposition (and not during the previous step as usual).
1044      */
1045     if (stepWork.doNeighborSearch)
1046     {
1047         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
1048     }
1049
1050     clear_mat(vir_force);
1051
1052     if (fr->pbcType != PbcType::No)
1053     {
1054         /* Compute shift vectors every step,
1055          * because of pressure coupling or box deformation!
1056          */
1057         if (stepWork.haveDynamicBox && stepWork.stateChanged)
1058         {
1059             calc_shifts(box, fr->shift_vec);
1060         }
1061
1062         const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
1063         const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
1064         if (calcCGCM)
1065         {
1066             put_atoms_in_box_omp(fr->pbcType, box, x.unpaddedArrayRef().subArray(0, mdatoms->homenr),
1067                                  gmx_omp_nthreads_get(emntDefault));
1068             inc_nrnb(nrnb, eNR_SHIFTX, mdatoms->homenr);
1069         }
1070     }
1071
1072     nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
1073
1074     const bool pmeSendCoordinatesFromGpu =
1075             GMX_MPI && simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1076     const bool reinitGpuPmePpComms =
1077             GMX_MPI && simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1078
1079     const auto localXReadyOnDevice = (useGpuPmeOnThisRank || simulationWork.useGpuBufferOps)
1080                                              ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1081                                                        AtomLocality::Local, simulationWork, stepWork)
1082                                              : nullptr;
1083
1084     // If coordinates are to be sent to PME task from CPU memory, perform that send here.
1085     // Otherwise the send will occur after H2D coordinate transfer.
1086     if (GMX_MPI && !thisRankHasDuty(cr, DUTY_PME) && !pmeSendCoordinatesFromGpu)
1087     {
1088         /* Send particle coordinates to the pme nodes */
1089         if (!stepWork.doNeighborSearch && simulationWork.useGpuUpdate)
1090         {
1091             GMX_RELEASE_ASSERT(false,
1092                                "GPU update and separate PME ranks are only supported with GPU "
1093                                "direct communication!");
1094             // TODO: when this code-path becomes supported add:
1095             // stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1096         }
1097
1098         gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1099                                  lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1100                                  step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1101                                  pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1102     }
1103
1104     // Coordinates on the device are needed if PME or BufferOps are offloaded.
1105     // The local coordinates can be copied right away.
1106     // NOTE: Consider moving this copy to right after they are updated and constrained,
1107     //       if the later is not offloaded.
1108     if (useGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
1109     {
1110         if (stepWork.doNeighborSearch)
1111         {
1112             // TODO refactor this to do_md, after partitioning.
1113             stateGpu->reinit(mdatoms->homenr,
1114                              cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1115             if (useGpuPmeOnThisRank)
1116             {
1117                 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1118                 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1119             }
1120         }
1121         // We need to copy coordinates when:
1122         // 1. Update is not offloaded
1123         // 2. The buffers were reinitialized on search step
1124         if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1125         {
1126             GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1127             stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1128         }
1129     }
1130
1131     // TODO Update this comment when introducing SimulationWorkload
1132     //
1133     // The conditions for gpuHaloExchange e.g. using GPU buffer
1134     // operations were checked before construction, so here we can
1135     // just use it and assert upon any conditions.
1136     const bool ddUsesGpuDirectCommunication =
1137             ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange.empty()));
1138     GMX_ASSERT(!ddUsesGpuDirectCommunication || stepWork.useGpuXBufferOps,
1139                "Must use coordinate buffer ops with GPU halo exchange");
1140     const bool useGpuForcesHaloExchange = ddUsesGpuDirectCommunication && stepWork.useGpuFBufferOps;
1141
1142     // Copy coordinate from the GPU if update is on the GPU and there
1143     // are forces to be computed on the CPU, or for the computation of
1144     // virial, or if host-side data will be transferred from this task
1145     // to a remote task for halo exchange or PME-PP communication. At
1146     // search steps the current coordinates are already on the host,
1147     // hence copy is not needed.
1148     const bool haveHostPmePpComms =
1149             !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication;
1150     const bool haveHostHaloExchangeComms = havePPDomainDecomposition(cr) && !ddUsesGpuDirectCommunication;
1151
1152     bool gmx_used_in_debug haveCopiedXFromGpu = false;
1153     if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1154         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1155             || haveHostPmePpComms || haveHostHaloExchangeComms))
1156     {
1157         GMX_ASSERT(stateGpu != nullptr, "stateGpu should not be null");
1158         stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1159         haveCopiedXFromGpu = true;
1160     }
1161
1162     // If coordinates are to be sent to PME task from GPU memory, perform that send here.
1163     // Otherwise the send will occur before the H2D coordinate transfer.
1164     if (!thisRankHasDuty(cr, DUTY_PME) && pmeSendCoordinatesFromGpu)
1165     {
1166         /* Send particle coordinates to the pme nodes */
1167         gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1168                                  lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1169                                  step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1170                                  pmeSendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1171     }
1172
1173     if (useGpuPmeOnThisRank)
1174     {
1175         launchPmeGpuSpread(fr->pmedata, box, stepWork, localXReadyOnDevice, lambda[efptCOUL], wcycle);
1176     }
1177
1178     /* do gridding for pair search */
1179     if (stepWork.doNeighborSearch)
1180     {
1181         if (fr->wholeMoleculeTransform && stepWork.stateChanged)
1182         {
1183             fr->wholeMoleculeTransform->updateForAtomPbcJumps(x.unpaddedArrayRef(), box);
1184         }
1185
1186         // TODO
1187         // - vzero is constant, do we need to pass it?
1188         // - box_diag should be passed directly to nbnxn_put_on_grid
1189         //
1190         rvec vzero;
1191         clear_rvec(vzero);
1192
1193         rvec box_diag;
1194         box_diag[XX] = box[XX][XX];
1195         box_diag[YY] = box[YY][YY];
1196         box_diag[ZZ] = box[ZZ][ZZ];
1197
1198         wallcycle_start(wcycle, ewcNS);
1199         if (!DOMAINDECOMP(cr))
1200         {
1201             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1202             nbnxn_put_on_grid(nbv, box, 0, vzero, box_diag, nullptr, { 0, mdatoms->homenr }, -1,
1203                               fr->cginfo, x.unpaddedArrayRef(), 0, nullptr);
1204             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1205         }
1206         else
1207         {
1208             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1209             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->cginfo, x.unpaddedArrayRef());
1210             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1211         }
1212
1213         nbv->setAtomProperties(gmx::constArrayRefFromArray(mdatoms->typeA, mdatoms->nr),
1214                                gmx::constArrayRefFromArray(mdatoms->chargeA, mdatoms->nr), fr->cginfo);
1215
1216         wallcycle_stop(wcycle, ewcNS);
1217
1218         /* initialize the GPU nbnxm atom data and bonded data structures */
1219         if (simulationWork.useGpuNonbonded)
1220         {
1221             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1222
1223             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1224             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1225             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1226
1227             if (fr->gpuBonded)
1228             {
1229                 /* Now we put all atoms on the grid, we can assign bonded
1230                  * interactions to the GPU, where the grid order is
1231                  * needed. Also the xq, f and fshift device buffers have
1232                  * been reallocated if needed, so the bonded code can
1233                  * learn about them. */
1234                 // TODO the xq, f, and fshift buffers are now shared
1235                 // resources, so they should be maintained by a
1236                 // higher-level object than the nb module.
1237                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(
1238                         nbv->getGridIndices(), top->idef, Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1239                         Nbnxm::gpu_get_f(nbv->gpu_nbv), Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1240             }
1241             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1242         }
1243
1244         // Need to run after the GPU-offload bonded interaction lists
1245         // are set up to be able to determine whether there is bonded work.
1246         runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1247                 *inputrec, *fr, pull_work, ed, *mdatoms, simulationWork, stepWork);
1248
1249         wallcycle_start_nocount(wcycle, ewcNS);
1250         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1251         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1252         nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1253
1254         nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::Local);
1255
1256         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1257         wallcycle_stop(wcycle, ewcNS);
1258
1259         if (stepWork.useGpuXBufferOps)
1260         {
1261             nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1262         }
1263         // For force buffer ops, we use the below conditon rather than
1264         // useGpuFBufferOps to ensure that init is performed even if this
1265         // NS step is also a virial step (on which f buf ops are deactivated).
1266         if (simulationWork.useGpuBufferOps && simulationWork.useGpuNonbonded && (GMX_GPU == GMX_GPU_CUDA))
1267         {
1268             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1269             nbv->atomdata_init_add_nbat_f_to_f_gpu(stateGpu->fReducedOnDevice());
1270         }
1271     }
1272     else if (!EI_TPI(inputrec->eI))
1273     {
1274         if (stepWork.useGpuXBufferOps)
1275         {
1276             GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1277             nbv->convertCoordinatesGpu(AtomLocality::Local, false, stateGpu->getCoordinates(),
1278                                        localXReadyOnDevice);
1279         }
1280         else
1281         {
1282             if (simulationWork.useGpuUpdate)
1283             {
1284                 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1285                 GMX_ASSERT(haveCopiedXFromGpu,
1286                            "a wait should only be triggered if copy has been scheduled");
1287                 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1288             }
1289             nbv->convertCoordinates(AtomLocality::Local, false, x.unpaddedArrayRef());
1290         }
1291     }
1292
1293     const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1294
1295     if (simulationWork.useGpuNonbonded)
1296     {
1297         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1298
1299         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1300
1301         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1302         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1303         if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1304         {
1305             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1306         }
1307         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1308         // with X buffer ops offloaded to the GPU on all but the search steps
1309
1310         // bonded work not split into separate local and non-local, so with DD
1311         // we can only launch the kernel after non-local coordinates have been received.
1312         if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1313         {
1314             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1315             fr->gpuBonded->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1316             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1317         }
1318
1319         /* launch local nonbonded work on GPU */
1320         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1321         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1322         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1323         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1324     }
1325
1326     if (useGpuPmeOnThisRank)
1327     {
1328         // In PME GPU and mixed mode we launch FFT / gather after the
1329         // X copy/transform to allow overlap as well as after the GPU NB
1330         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1331         // the nonbonded kernel.
1332         launchPmeGpuFftAndGather(fr->pmedata, lambda[efptCOUL], wcycle, stepWork);
1333     }
1334
1335     /* Communicate coordinates and sum dipole if necessary +
1336        do non-local pair search */
1337     if (havePPDomainDecomposition(cr))
1338     {
1339         if (stepWork.doNeighborSearch)
1340         {
1341             // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1342             wallcycle_start_nocount(wcycle, ewcNS);
1343             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1344             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1345             nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1346
1347             nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::NonLocal);
1348             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1349             wallcycle_stop(wcycle, ewcNS);
1350             // TODO refactor this GPU halo exchange re-initialisation
1351             // to location in do_md where GPU halo exchange is
1352             // constructed at partitioning, after above stateGpu
1353             // re-initialization has similarly been refactored
1354             if (ddUsesGpuDirectCommunication)
1355             {
1356                 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1357             }
1358         }
1359         else
1360         {
1361             if (ddUsesGpuDirectCommunication)
1362             {
1363                 // The following must be called after local setCoordinates (which records an event
1364                 // when the coordinate data has been copied to the device).
1365                 communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1366
1367                 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1368                 {
1369                     // non-local part of coordinate buffer must be copied back to host for CPU work
1370                     stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1371                 }
1372             }
1373             else
1374             {
1375                 // Note: GPU update + DD without direct communication is not supported,
1376                 // a waitCoordinatesReadyOnHost() should be issued if it will be.
1377                 GMX_ASSERT(!simulationWork.useGpuUpdate,
1378                            "GPU update is not supported with CPU halo exchange");
1379                 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1380             }
1381
1382             if (stepWork.useGpuXBufferOps)
1383             {
1384                 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1385                 {
1386                     stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1387                 }
1388                 nbv->convertCoordinatesGpu(AtomLocality::NonLocal, false, stateGpu->getCoordinates(),
1389                                            stateGpu->getCoordinatesReadyOnDeviceEvent(
1390                                                    AtomLocality::NonLocal, simulationWork, stepWork));
1391             }
1392             else
1393             {
1394                 nbv->convertCoordinates(AtomLocality::NonLocal, false, x.unpaddedArrayRef());
1395             }
1396         }
1397
1398         if (simulationWork.useGpuNonbonded)
1399         {
1400             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1401
1402             if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1403             {
1404                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1405                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1406                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1407             }
1408
1409             if (domainWork.haveGpuBondedWork)
1410             {
1411                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1412                 fr->gpuBonded->setPbcAndlaunchKernel(fr->pbcType, box, fr->bMolPBC, stepWork);
1413                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1414             }
1415
1416             /* launch non-local nonbonded tasks on GPU */
1417             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1418             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1419                          nrnb, wcycle);
1420             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1421
1422             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1423         }
1424     }
1425
1426     if (simulationWork.useGpuNonbonded)
1427     {
1428         /* launch D2H copy-back F */
1429         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1430         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1431
1432         if (havePPDomainDecomposition(cr))
1433         {
1434             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1435         }
1436         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1437         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1438
1439         if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1440         {
1441             fr->gpuBonded->launchEnergyTransfer();
1442         }
1443         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1444     }
1445
1446     gmx::ArrayRef<const gmx::RVec> xWholeMolecules;
1447     if (fr->wholeMoleculeTransform)
1448     {
1449         xWholeMolecules = fr->wholeMoleculeTransform->wholeMoleculeCoordinates(x.unpaddedArrayRef(), box);
1450     }
1451
1452     DipoleData dipoleData;
1453
1454     if (simulationWork.computeMuTot)
1455     {
1456         const int start = 0;
1457
1458         /* Calculate total (local) dipole moment in a temporary common array.
1459          * This makes it possible to sum them over nodes faster.
1460          */
1461         gmx::ArrayRef<const gmx::RVec> xRef =
1462                 (xWholeMolecules.empty() ? x.unpaddedArrayRef() : xWholeMolecules);
1463         calc_mu(start, mdatoms->homenr, xRef, mdatoms->chargeA, mdatoms->chargeB,
1464                 mdatoms->nChargePerturbed, dipoleData.muStaging[0], dipoleData.muStaging[1]);
1465
1466         reduceAndUpdateMuTot(&dipoleData, cr, (fr->efep != efepNO), lambda, muTotal, ddBalanceRegionHandler);
1467     }
1468
1469     /* Reset energies */
1470     reset_enerdata(enerd);
1471
1472     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1473     {
1474         wallcycle_start(wcycle, ewcPPDURINGPME);
1475         dd_force_flop_start(cr->dd, nrnb);
1476     }
1477
1478     // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1479     // this wait ensures that the D2H transfer is complete.
1480     if ((simulationWork.useGpuUpdate)
1481         && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial))
1482     {
1483         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1484     }
1485
1486     if (inputrec->bRot)
1487     {
1488         wallcycle_start(wcycle, ewcROT);
1489         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step,
1490                     stepWork.doNeighborSearch);
1491         wallcycle_stop(wcycle, ewcROT);
1492     }
1493
1494     /* Start the force cycle counter.
1495      * Note that a different counter is used for dynamic load balancing.
1496      */
1497     wallcycle_start(wcycle, ewcFORCE);
1498
1499     // Set up and clear force outputs.
1500     // We use std::move to keep the compiler happy, it has no effect.
1501     ForceOutputs forceOut = setupForceOutputs(fr->forceHelperBuffers.get(), pull_work, *inputrec,
1502                                               std::move(force), stepWork, wcycle);
1503
1504     /* We calculate the non-bonded forces, when done on the CPU, here.
1505      * We do this before calling do_force_lowlevel, because in that
1506      * function, the listed forces are calculated before PME, which
1507      * does communication.  With this order, non-bonded and listed
1508      * force calculation imbalance can be balanced out by the domain
1509      * decomposition load balancing.
1510      */
1511
1512     const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1513
1514     if (!useOrEmulateGpuNb)
1515     {
1516         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1517     }
1518
1519     if (fr->efep != efepNO)
1520     {
1521         /* Calculate the local and non-local free energy interactions here.
1522          * Happens here on the CPU both with and without GPU.
1523          */
1524         nbv->dispatchFreeEnergyKernel(InteractionLocality::Local, fr,
1525                                       as_rvec_array(x.unpaddedArrayRef().data()),
1526                                       &forceOut.forceWithShiftForces(), *mdatoms, inputrec->fepvals,
1527                                       lambda, enerd, stepWork, nrnb);
1528
1529         if (havePPDomainDecomposition(cr))
1530         {
1531             nbv->dispatchFreeEnergyKernel(InteractionLocality::NonLocal, fr,
1532                                           as_rvec_array(x.unpaddedArrayRef().data()),
1533                                           &forceOut.forceWithShiftForces(), *mdatoms,
1534                                           inputrec->fepvals, lambda, enerd, stepWork, nrnb);
1535         }
1536     }
1537
1538     if (!useOrEmulateGpuNb)
1539     {
1540         if (havePPDomainDecomposition(cr))
1541         {
1542             do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1543                          nrnb, wcycle);
1544         }
1545
1546         if (stepWork.computeForces)
1547         {
1548             /* Add all the non-bonded force to the normal force array.
1549              * This can be split into a local and a non-local part when overlapping
1550              * communication with calculation with domain decomposition.
1551              */
1552             wallcycle_stop(wcycle, ewcFORCE);
1553             nbv->atomdata_add_nbat_f_to_f(AtomLocality::All, forceOut.forceWithShiftForces().force());
1554             wallcycle_start_nocount(wcycle, ewcFORCE);
1555         }
1556
1557         /* If there are multiple fshift output buffers we need to reduce them */
1558         if (stepWork.computeVirial)
1559         {
1560             /* This is not in a subcounter because it takes a
1561                negligible and constant-sized amount of time */
1562             nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1563                                                      forceOut.forceWithShiftForces().shiftForces());
1564         }
1565     }
1566
1567     // TODO Force flags should include haveFreeEnergyWork for this domain
1568     if (ddUsesGpuDirectCommunication && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1569     {
1570         /* Wait for non-local coordinate data to be copied from device */
1571         stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1572     }
1573     /* Compute the bonded and non-bonded energies and optionally forces */
1574     do_force_lowlevel(fr, inputrec, cr, ms, nrnb, wcycle, mdatoms, x, xWholeMolecules, hist,
1575                       &forceOut, enerd, box, lambda.data(), as_rvec_array(dipoleData.muStateAB),
1576                       stepWork, ddBalanceRegionHandler);
1577
1578     wallcycle_stop(wcycle, ewcFORCE);
1579
1580     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation, imdSession, pull_work, step, t,
1581                          wcycle, fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda,
1582                          stepWork, &forceOut.forceWithVirial(), enerd, ed, stepWork.doNeighborSearch);
1583
1584
1585     // Will store the amount of cycles spent waiting for the GPU that
1586     // will be later used in the DLB accounting.
1587     float cycles_wait_gpu = 0;
1588     if (useOrEmulateGpuNb)
1589     {
1590         auto& forceWithShiftForces = forceOut.forceWithShiftForces();
1591
1592         /* wait for non-local forces (or calculate in emulation mode) */
1593         if (havePPDomainDecomposition(cr))
1594         {
1595             if (simulationWork.useGpuNonbonded)
1596             {
1597                 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
1598                         nbv->gpu_nbv, stepWork, AtomLocality::NonLocal, enerd->grpp.ener[egLJSR].data(),
1599                         enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(), wcycle);
1600             }
1601             else
1602             {
1603                 wallcycle_start_nocount(wcycle, ewcFORCE);
1604                 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes,
1605                              step, nrnb, wcycle);
1606                 wallcycle_stop(wcycle, ewcFORCE);
1607             }
1608
1609             if (stepWork.useGpuFBufferOps)
1610             {
1611                 gmx::FixedCapacityVector<GpuEventSynchronizer*, 1> dependencyList;
1612
1613                 // TODO: move this into DomainLifetimeWorkload, including the second part of the
1614                 // condition The bonded and free energy CPU tasks can have non-local force
1615                 // contributions which are a dependency for the GPU force reduction.
1616                 bool haveNonLocalForceContribInCpuBuffer =
1617                         domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1618
1619                 if (haveNonLocalForceContribInCpuBuffer)
1620                 {
1621                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(),
1622                                               AtomLocality::NonLocal);
1623                     dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(
1624                             AtomLocality::NonLocal, stepWork.useGpuFBufferOps));
1625                 }
1626
1627                 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::NonLocal, stateGpu->getForces(),
1628                                                   pme_gpu_get_device_f(fr->pmedata), dependencyList,
1629                                                   false, haveNonLocalForceContribInCpuBuffer);
1630                 if (!useGpuForcesHaloExchange)
1631                 {
1632                     // copy from GPU input for dd_move_f()
1633                     stateGpu->copyForcesFromGpu(forceOut.forceWithShiftForces().force(),
1634                                                 AtomLocality::NonLocal);
1635                 }
1636             }
1637             else
1638             {
1639                 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
1640             }
1641
1642
1643             if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1644             {
1645                 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
1646             }
1647         }
1648     }
1649
1650     if (havePPDomainDecomposition(cr))
1651     {
1652         /* We are done with the CPU compute.
1653          * We will now communicate the non-local forces.
1654          * If we use a GPU this will overlap with GPU work, so in that case
1655          * we do not close the DD force balancing region here.
1656          */
1657         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1658
1659         if (stepWork.computeForces)
1660         {
1661
1662             if (useGpuForcesHaloExchange)
1663             {
1664                 if (domainWork.haveCpuLocalForceWork)
1665                 {
1666                     stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), AtomLocality::Local);
1667                 }
1668                 communicateGpuHaloForces(*cr, domainWork.haveCpuLocalForceWork);
1669             }
1670             else
1671             {
1672                 if (stepWork.useGpuFBufferOps)
1673                 {
1674                     stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
1675                 }
1676                 dd_move_f(cr->dd, &forceOut.forceWithShiftForces(), wcycle);
1677             }
1678         }
1679     }
1680
1681     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1682     // an alternating wait/reduction scheme.
1683     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded
1684                              && !DOMAINDECOMP(cr) && !stepWork.useGpuFBufferOps);
1685     if (alternateGpuWait)
1686     {
1687         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &forceOut, enerd, lambda[efptCOUL],
1688                                     stepWork, wcycle);
1689     }
1690
1691     if (!alternateGpuWait && useGpuPmeOnThisRank)
1692     {
1693         pme_gpu_wait_and_reduce(fr->pmedata, stepWork, wcycle, &forceOut.forceWithVirial(), enerd,
1694                                 lambda[efptCOUL]);
1695     }
1696
1697     /* Wait for local GPU NB outputs on the non-alternating wait path */
1698     if (!alternateGpuWait && simulationWork.useGpuNonbonded)
1699     {
1700         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1701          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1702          * but even with a step of 0.1 ms the difference is less than 1%
1703          * of the step time.
1704          */
1705         const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1706         const float waitCycles               = Nbnxm::gpu_wait_finish_task(
1707                 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
1708                 enerd->grpp.ener[egCOULSR].data(), forceOut.forceWithShiftForces().shiftForces(), wcycle);
1709
1710         if (ddBalanceRegionHandler.useBalancingRegion())
1711         {
1712             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1713             if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
1714             {
1715                 /* We measured few cycles, it could be that the kernel
1716                  * and transfer finished earlier and there was no actual
1717                  * wait time, only API call overhead.
1718                  * Then the actual time could be anywhere between 0 and
1719                  * cycles_wait_est. We will use half of cycles_wait_est.
1720                  */
1721                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1722             }
1723             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1724         }
1725     }
1726
1727     if (fr->nbv->emulateGpu())
1728     {
1729         // NOTE: emulation kernel is not included in the balancing region,
1730         // but emulation mode does not target performance anyway
1731         wallcycle_start_nocount(wcycle, ewcFORCE);
1732         do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local,
1733                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes, step, nrnb, wcycle);
1734         wallcycle_stop(wcycle, ewcFORCE);
1735     }
1736
1737     // If on GPU PME-PP comms or GPU update path, receive forces from PME before GPU buffer ops
1738     // TODO refactor this and unify with below default-path call to the same function
1739     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME)
1740         && (simulationWork.useGpuPmePpCommunication || simulationWork.useGpuUpdate))
1741     {
1742         /* In case of node-splitting, the PP nodes receive the long-range
1743          * forces, virial and energy from the PME nodes here.
1744          */
1745         pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1746                                simulationWork.useGpuPmePpCommunication,
1747                                stepWork.useGpuPmeFReduction, wcycle);
1748     }
1749
1750
1751     /* Do the nonbonded GPU (or emulation) force buffer reduction
1752      * on the non-alternating path. */
1753     if (useOrEmulateGpuNb && !alternateGpuWait)
1754     {
1755         // TODO simplify the below conditionals. Pass buffer and sync pointers at init stage rather than here. Unify getter fns for sameGPU/otherGPU cases.
1756         void* pmeForcePtr =
1757                 stepWork.useGpuPmeFReduction
1758                         ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_device_f(fr->pmedata)
1759                                                          : // PME force buffer on same GPU
1760                                    fr->pmePpCommGpu->getGpuForceStagingPtr()) // buffer received from other GPU
1761                         : nullptr; // PME reduction not active on GPU
1762
1763         GpuEventSynchronizer* const pmeSynchronizer =
1764                 stepWork.useGpuPmeFReduction
1765                         ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_f_ready_synchronizer(fr->pmedata)
1766                                                          : // PME force buffer on same GPU
1767                                    static_cast<GpuEventSynchronizer*>(
1768                                            fr->pmePpCommGpu->getForcesReadySynchronizer())) // buffer received from other GPU
1769                         : nullptr; // PME reduction not active on GPU
1770
1771         gmx::FixedCapacityVector<GpuEventSynchronizer*, 3> dependencyList;
1772
1773         if (stepWork.useGpuPmeFReduction)
1774         {
1775             dependencyList.push_back(pmeSynchronizer);
1776         }
1777
1778         gmx::ArrayRef<gmx::RVec> forceWithShift = forceOut.forceWithShiftForces().force();
1779
1780         if (stepWork.useGpuFBufferOps)
1781         {
1782             // Flag to specify whether the CPU force buffer has contributions to
1783             // local atoms. This depends on whether there are CPU-based force tasks
1784             // or when DD is active the halo exchange has resulted in contributions
1785             // from the non-local part.
1786             const bool haveLocalForceContribInCpuBuffer =
1787                     (domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr));
1788
1789             // TODO: move these steps as early as possible:
1790             // - CPU f H2D should be as soon as all CPU-side forces are done
1791             // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1792             //   before the next CPU task that consumes the forces: vsite spread or update)
1793             // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
1794             //   of the halo exchange. In that case the copy is instead performed above, before the exchange.
1795             //   These should be unified.
1796             if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
1797             {
1798                 // Note: AtomLocality::All is used for the non-DD case because, as in this
1799                 // case copyForcesToGpu() uses a separate stream, it allows overlap of
1800                 // CPU force H2D with GPU force tasks on all streams including those in the
1801                 // local stream which would otherwise be implicit dependencies for the
1802                 // transfer and would not overlap.
1803                 auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
1804
1805                 stateGpu->copyForcesToGpu(forceWithShift, locality);
1806                 dependencyList.push_back(
1807                         stateGpu->getForcesReadyOnDeviceEvent(locality, stepWork.useGpuFBufferOps));
1808             }
1809             if (useGpuForcesHaloExchange)
1810             {
1811                 dependencyList.push_back(cr->dd->gpuHaloExchange[0]->getForcesReadyOnDeviceEvent());
1812             }
1813             nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::Local, stateGpu->getForces(), pmeForcePtr,
1814                                               dependencyList, stepWork.useGpuPmeFReduction,
1815                                               haveLocalForceContribInCpuBuffer);
1816             // Copy forces to host if they are needed for update or if virtual sites are enabled.
1817             // If there are vsites, we need to copy forces every step to spread vsite forces on host.
1818             // TODO: When the output flags will be included in step workload, this copy can be combined with the
1819             //       copy call done in sim_utils(...) for the output.
1820             // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
1821             //       they should not be copied in do_md(...) for the output.
1822             if (!simulationWork.useGpuUpdate || vsite)
1823             {
1824                 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
1825                 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
1826             }
1827         }
1828         else
1829         {
1830             nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
1831         }
1832     }
1833
1834     launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd, *runScheduleWork,
1835                             useGpuPmeOnThisRank, step, wcycle);
1836
1837     if (DOMAINDECOMP(cr))
1838     {
1839         dd_force_flop_stop(cr->dd, nrnb);
1840     }
1841
1842     if (stepWork.computeForces)
1843     {
1844         postProcessForceWithShiftForces(nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOut,
1845                                         vir_force, *mdatoms, *fr, vsite, stepWork);
1846     }
1847
1848     // VdW dispersion correction, only computed on master rank to avoid double counting
1849     if ((stepWork.computeEnergy || stepWork.computeVirial) && fr->dispersionCorrection && MASTER(cr))
1850     {
1851         // Calculate long range corrections to pressure and energy
1852         const DispersionCorrection::Correction correction =
1853                 fr->dispersionCorrection->calculate(box, lambda[efptVDW]);
1854
1855         if (stepWork.computeEnergy)
1856         {
1857             enerd->term[F_DISPCORR] = correction.energy;
1858             enerd->term[F_DVDL_VDW] += correction.dvdl;
1859             enerd->dvdl_lin[efptVDW] += correction.dvdl;
1860         }
1861         if (stepWork.computeVirial)
1862         {
1863             correction.correctVirial(vir_force);
1864             enerd->term[F_PDISPCORR] = correction.pressure;
1865         }
1866     }
1867
1868     // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
1869     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication
1870         && !simulationWork.useGpuUpdate)
1871     {
1872         /* In case of node-splitting, the PP nodes receive the long-range
1873          * forces, virial and energy from the PME nodes here.
1874          */
1875         pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1876                                simulationWork.useGpuPmePpCommunication, false, wcycle);
1877     }
1878
1879     if (stepWork.computeForces)
1880     {
1881         postProcessForces(cr, step, nrnb, wcycle, box, x.unpaddedArrayRef(), &forceOut, vir_force,
1882                           mdatoms, fr, vsite, stepWork);
1883     }
1884
1885     if (stepWork.computeEnergy)
1886     {
1887         /* Compute the final potential energy terms */
1888         accumulatePotentialEnergies(enerd, lambda, inputrec->fepvals);
1889
1890         if (!EI_TPI(inputrec->eI))
1891         {
1892             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1893         }
1894     }
1895
1896     /* In case we don't have constraints and are using GPUs, the next balancing
1897      * region starts here.
1898      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1899      * virial calculation and COM pulling, is not thus not included in
1900      * the balance timing, which is ok as most tasks do communication.
1901      */
1902     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1903 }