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