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) 2013,2014,2015,2016,2017,2018,2019, by the GROMACS development team, led by
   7  * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
   8  * and including many others, as listed in the AUTHORS file in the
   9  * top-level source directory and at http://www.gromacs.org.
  10  *
  11  * GROMACS is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public License
  13  * as published by the Free Software Foundation; either version 2.1
  14  * of the License, or (at your option) any later version.
  15  *
  16  * GROMACS is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with GROMACS; if not, see
  23  * http://www.gnu.org/licenses, or write to the Free Software Foundation,
  24  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
  25  *
  26  * If you want to redistribute modifications to GROMACS, please
  27  * consider that scientific software is very special. Version
  28  * control is crucial - bugs must be traceable. We will be happy to
  29  * consider code for inclusion in the official distribution, but
  30  * derived work must not be called official GROMACS. Details are found
  31  * in the README & COPYING files - if they are missing, get the
  32  * official version at http://www.gromacs.org.
  33  *
  34  * To help us fund GROMACS development, we humbly ask that you cite
  35  * the research papers on the package. Check out http://www.gromacs.org.
  36  */
  37 #include "gmxpre.h"
  38
  39 #include "config.h"
  40
  41 #include <cmath>
  42 #include <cstdint>
  43 #include <cstdio>
  44 #include <cstring>
  45
  46 #include <array>
  47
  48 #include "gromacs/awh/awh.h"
  49 #include "gromacs/domdec/dlbtiming.h"
  50 #include "gromacs/domdec/domdec.h"
  51 #include "gromacs/domdec/domdec_struct.h"
  52 #include "gromacs/domdec/partition.h"
  53 #include "gromacs/essentialdynamics/edsam.h"
  54 #include "gromacs/ewald/pme.h"
  55 #include "gromacs/gmxlib/chargegroup.h"
  56 #include "gromacs/gmxlib/network.h"
  57 #include "gromacs/gmxlib/nrnb.h"
  58 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  59 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  60 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  61 #include "gromacs/gpu_utils/gpu_utils.h"
  62 #include "gromacs/imd/imd.h"
  63 #include "gromacs/listed_forces/bonded.h"
  64 #include "gromacs/listed_forces/disre.h"
  65 #include "gromacs/listed_forces/gpubonded.h"
  66 #include "gromacs/listed_forces/listed_forces.h"
  67 #include "gromacs/listed_forces/manage_threading.h"
  68 #include "gromacs/listed_forces/orires.h"
  69 #include "gromacs/math/arrayrefwithpadding.h"
  70 #include "gromacs/math/functions.h"
  71 #include "gromacs/math/units.h"
  72 #include "gromacs/math/vec.h"
  73 #include "gromacs/math/vecdump.h"
  74 #include "gromacs/mdlib/calcmu.h"
  75 #include "gromacs/mdlib/calcvir.h"
  76 #include "gromacs/mdlib/constr.h"
  77 #include "gromacs/mdlib/force.h"
  78 #include "gromacs/mdlib/forcerec.h"
  79 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  80 #include "gromacs/mdlib/ppforceworkload.h"
  81 #include "gromacs/mdlib/qmmm.h"
  82 #include "gromacs/mdlib/update.h"
  83 #include "gromacs/mdtypes/commrec.h"
  84 #include "gromacs/mdtypes/enerdata.h"
  85 #include "gromacs/mdtypes/forceoutput.h"
  86 #include "gromacs/mdtypes/iforceprovider.h"
  87 #include "gromacs/mdtypes/inputrec.h"
  88 #include "gromacs/mdtypes/md_enums.h"
  89 #include "gromacs/mdtypes/state.h"
  90 #include "gromacs/nbnxm/atomdata.h"
  91 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  92 #include "gromacs/nbnxm/nbnxm.h"
  93 #include "gromacs/pbcutil/ishift.h"
  94 #include "gromacs/pbcutil/mshift.h"
  95 #include "gromacs/pbcutil/pbc.h"
  96 #include "gromacs/pulling/pull.h"
  97 #include "gromacs/pulling/pull_rotation.h"
  98 #include "gromacs/timing/cyclecounter.h"
  99 #include "gromacs/timing/gpu_timing.h"
 100 #include "gromacs/timing/wallcycle.h"
 101 #include "gromacs/timing/wallcyclereporting.h"
 102 #include "gromacs/timing/walltime_accounting.h"
 103 #include "gromacs/topology/topology.h"
 104 #include "gromacs/utility/arrayref.h"
 105 #include "gromacs/utility/basedefinitions.h"
 106 #include "gromacs/utility/cstringutil.h"
 107 #include "gromacs/utility/exceptions.h"
 108 #include "gromacs/utility/fatalerror.h"
 109 #include "gromacs/utility/gmxassert.h"
 110 #include "gromacs/utility/gmxmpi.h"
 111 #include "gromacs/utility/logger.h"
 112 #include "gromacs/utility/smalloc.h"
 113 #include "gromacs/utility/strconvert.h"
 114 #include "gromacs/utility/sysinfo.h"
 115
 116 // TODO: this environment variable allows us to verify before release
 117 // that on less common architectures the total cost of polling is not larger than
 118 // a blocking wait (so polling does not introduce overhead when the static
 119 // PME-first ordering would suffice).
 120 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 121
 122
 123 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 124 {
 125     const int      end = forceToAdd.size();
 126
 127     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 128 #pragma omp parallel for num_threads(nt) schedule(static)
 129     for (int i = 0; i < end; i++)
 130     {
 131         rvec_inc(f[i], forceToAdd[i]);
 132     }
 133 }
 134
 135 static void calc_virial(int start, int homenr, const rvec x[], const rvec f[],
 136                         tensor vir_part, const t_graph *graph, const matrix box,
 137                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 138 {
 139     /* The short-range virial from surrounding boxes */
 140     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 141     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 142
 143     /* Calculate partial virial, for local atoms only, based on short range.
 144      * Total virial is computed in global_stat, called from do_md
 145      */
 146     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 147     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 148
 149     if (debug)
 150     {
 151         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 152     }
 153 }
 154
 155 static void pull_potential_wrapper(const t_commrec *cr,
 156                                    const t_inputrec *ir,
 157                                    const matrix box, gmx::ArrayRef<const gmx::RVec> x,
 158                                    gmx::ForceWithVirial *force,
 159                                    const t_mdatoms *mdatoms,
 160                                    gmx_enerdata_t *enerd,
 161                                    const real *lambda,
 162                                    double t,
 163                                    gmx_wallcycle_t wcycle)
 164 {
 165     t_pbc  pbc;
 166     real   dvdl;
 167
 168     /* Calculate the center of mass forces, this requires communication,
 169      * which is why pull_potential is called close to other communication.
 170      */
 171     wallcycle_start(wcycle, ewcPULLPOT);
 172     set_pbc(&pbc, ir->ePBC, box);
 173     dvdl                     = 0;
 174     enerd->term[F_COM_PULL] +=
 175         pull_potential(ir->pull_work, mdatoms, &pbc,
 176                        cr, t, lambda[efptRESTRAINT], as_rvec_array(x.data()), force, &dvdl);
 177     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 178     wallcycle_stop(wcycle, ewcPULLPOT);
 179 }
 180
 181 static void pme_receive_force_ener(const t_commrec      *cr,
 182                                    gmx::ForceWithVirial *forceWithVirial,
 183                                    gmx_enerdata_t       *enerd,
 184                                    gmx_wallcycle_t       wcycle)
 185 {
 186     real   e_q, e_lj, dvdl_q, dvdl_lj;
 187     float  cycles_ppdpme, cycles_seppme;
 188
 189     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 190     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 191
 192     /* In case of node-splitting, the PP nodes receive the long-range
 193      * forces, virial and energy from the PME nodes here.
 194      */
 195     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 196     dvdl_q  = 0;
 197     dvdl_lj = 0;
 198     gmx_pme_receive_f(cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 199                       &cycles_seppme);
 200     enerd->term[F_COUL_RECIP] += e_q;
 201     enerd->term[F_LJ_RECIP]   += e_lj;
 202     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 203     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 204
 205     if (wcycle)
 206     {
 207         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 208     }
 209     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 210 }
 211
 212 static void print_large_forces(FILE            *fp,
 213                                const t_mdatoms *md,
 214                                const t_commrec *cr,
 215                                int64_t          step,
 216                                real             forceTolerance,
 217                                const rvec      *x,
 218                                const rvec      *f)
 219 {
 220     real           force2Tolerance = gmx::square(forceTolerance);
 221     gmx::index     numNonFinite    = 0;
 222     for (int i = 0; i < md->homenr; i++)
 223     {
 224         real force2    = norm2(f[i]);
 225         bool nonFinite = !std::isfinite(force2);
 226         if (force2 >= force2Tolerance || nonFinite)
 227         {
 228             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 229                     step,
 230                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 231         }
 232         if (nonFinite)
 233         {
 234             numNonFinite++;
 235         }
 236     }
 237     if (numNonFinite > 0)
 238     {
 239         /* Note that with MPI this fatal call on one rank might interrupt
 240          * the printing on other ranks. But we can only avoid that with
 241          * an expensive MPI barrier that we would need at each step.
 242          */
 243         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 244     }
 245 }
 246
 247 static void post_process_forces(const t_commrec           *cr,
 248                                 int64_t                    step,
 249                                 t_nrnb                    *nrnb,
 250                                 gmx_wallcycle_t            wcycle,
 251                                 const gmx_localtop_t      *top,
 252                                 const matrix               box,
 253                                 const rvec                 x[],
 254                                 rvec                       f[],
 255                                 gmx::ForceWithVirial      *forceWithVirial,
 256                                 tensor                     vir_force,
 257                                 const t_mdatoms           *mdatoms,
 258                                 const t_graph             *graph,
 259                                 const t_forcerec          *fr,
 260                                 const gmx_vsite_t         *vsite,
 261                                 int                        flags)
 262 {
 263     if (fr->haveDirectVirialContributions)
 264     {
 265         rvec *fDirectVir = as_rvec_array(forceWithVirial->force_.data());
 266
 267         if (vsite)
 268         {
 269             /* Spread the mesh force on virtual sites to the other particles...
 270              * This is parallellized. MPI communication is performed
 271              * if the constructing atoms aren't local.
 272              */
 273             matrix virial = { { 0 } };
 274             spread_vsite_f(vsite, x, fDirectVir, nullptr,
 275                            (flags & GMX_FORCE_VIRIAL) != 0, virial,
 276                            nrnb,
 277                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
 278             forceWithVirial->addVirialContribution(virial);
 279         }
 280
 281         if (flags & GMX_FORCE_VIRIAL)
 282         {
 283             /* Now add the forces, this is local */
 284             sum_forces(f, forceWithVirial->force_);
 285
 286             /* Add the direct virial contributions */
 287             GMX_ASSERT(forceWithVirial->computeVirial_, "forceWithVirial should request virial computation when we request the virial");
 288             m_add(vir_force, forceWithVirial->getVirial(), vir_force);
 289
 290             if (debug)
 291             {
 292                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 293             }
 294         }
 295     }
 296
 297     if (fr->print_force >= 0)
 298     {
 299         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 300     }
 301 }
 302
 303 static void do_nb_verlet(t_forcerec                       *fr,
 304                          const interaction_const_t        *ic,
 305                          gmx_enerdata_t                   *enerd,
 306                          const int                         flags,
 307                          const Nbnxm::InteractionLocality  ilocality,
 308                          const int                         clearF,
 309                          const int64_t                     step,
 310                          t_nrnb                           *nrnb,
 311                          gmx_wallcycle_t                   wcycle)
 312 {
 313     if (!(flags & GMX_FORCE_NONBONDED))
 314     {
 315         /* skip non-bonded calculation */
 316         return;
 317     }
 318
 319     nonbonded_verlet_t *nbv  = fr->nbv.get();
 320
 321     /* GPU kernel launch overhead is already timed separately */
 322     if (fr->cutoff_scheme != ecutsVERLET)
 323     {
 324         gmx_incons("Invalid cut-off scheme passed!");
 325     }
 326
 327     if (!nbv->useGpu())
 328     {
 329         /* When dynamic pair-list  pruning is requested, we need to prune
 330          * at nstlistPrune steps.
 331          */
 332         if (nbv->isDynamicPruningStepCpu(step))
 333         {
 334             /* Prune the pair-list beyond fr->ic->rlistPrune using
 335              * the current coordinates of the atoms.
 336              */
 337             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 338             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 339             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 340         }
 341
 342         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 343     }
 344
 345     nbv->dispatchNonbondedKernel(ilocality, *ic, flags, clearF, fr, enerd, nrnb);
 346
 347     if (!nbv->useGpu())
 348     {
 349         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 350     }
 351 }
 352
 353 static inline void clear_rvecs_omp(int n, rvec v[])
 354 {
 355     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 356
 357     /* Note that we would like to avoid this conditional by putting it
 358      * into the omp pragma instead, but then we still take the full
 359      * omp parallel for overhead (at least with gcc5).
 360      */
 361     if (nth == 1)
 362     {
 363         for (int i = 0; i < n; i++)
 364         {
 365             clear_rvec(v[i]);
 366         }
 367     }
 368     else
 369     {
 370 #pragma omp parallel for num_threads(nth) schedule(static)
 371         for (int i = 0; i < n; i++)
 372         {
 373             clear_rvec(v[i]);
 374         }
 375     }
 376 }
 377
 378 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 379  *
 380  * \param groupOptions  Group options, containing T-coupling options
 381  */
 382 static real averageKineticEnergyEstimate(const t_grpopts &groupOptions)
 383 {
 384     real nrdfCoupled   = 0;
 385     real nrdfUncoupled = 0;
 386     real kineticEnergy = 0;
 387     for (int g = 0; g < groupOptions.ngtc; g++)
 388     {
 389         if (groupOptions.tau_t[g] >= 0)
 390         {
 391             nrdfCoupled   += groupOptions.nrdf[g];
 392             kineticEnergy += groupOptions.nrdf[g]*0.5*groupOptions.ref_t[g]*BOLTZ;
 393         }
 394         else
 395         {
 396             nrdfUncoupled += groupOptions.nrdf[g];
 397         }
 398     }
 399
 400     /* This conditional with > also catches nrdf=0 */
 401     if (nrdfCoupled > nrdfUncoupled)
 402     {
 403         return kineticEnergy*(nrdfCoupled + nrdfUncoupled)/nrdfCoupled;
 404     }
 405     else
 406     {
 407         return 0;
 408     }
 409 }
 410
 411 /*! \brief This routine checks that the potential energy is finite.
 412  *
 413  * Always checks that the potential energy is finite. If step equals
 414  * inputrec.init_step also checks that the magnitude of the potential energy
 415  * is reasonable. Terminates with a fatal error when a check fails.
 416  * Note that passing this check does not guarantee finite forces,
 417  * since those use slightly different arithmetics. But in most cases
 418  * there is just a narrow coordinate range where forces are not finite
 419  * and energies are finite.
 420  *
 421  * \param[in] step      The step number, used for checking and printing
 422  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to enerd.term[F_EPOT] before calling this routine
 423  * \param[in] inputrec  The input record
 424  */
 425 static void checkPotentialEnergyValidity(int64_t               step,
 426                                          const gmx_enerdata_t &enerd,
 427                                          const t_inputrec     &inputrec)
 428 {
 429     /* Threshold valid for comparing absolute potential energy against
 430      * the kinetic energy. Normally one should not consider absolute
 431      * potential energy values, but with a factor of one million
 432      * we should never get false positives.
 433      */
 434     constexpr real c_thresholdFactor = 1e6;
 435
 436     bool           energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 437     real           averageKineticEnergy = 0;
 438     /* We only check for large potential energy at the initial step,
 439      * because that is by far the most likely step for this too occur
 440      * and because computing the average kinetic energy is not free.
 441      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 442      * before they become NaN.
 443      */
 444     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 445     {
 446         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 447     }
 448
 449     if (energyIsNotFinite || (averageKineticEnergy > 0 &&
 450                               enerd.term[F_EPOT] > c_thresholdFactor*averageKineticEnergy))
 451     {
 452         gmx_fatal(FARGS, "Step %" PRId64 ": The total potential energy is %g, which is %s. The LJ and electrostatic contributions to the energy are %g and %g, respectively. A %s potential energy can be caused by overlapping interactions in bonded interactions or very large%s coordinate values. Usually this is caused by a badly- or non-equilibrated initial configuration, incorrect interactions or parameters in the topology.",
 453                   step,
 454                   enerd.term[F_EPOT],
 455                   energyIsNotFinite ? "not finite" : "extremely high",
 456                   enerd.term[F_LJ],
 457                   enerd.term[F_COUL_SR],
 458                   energyIsNotFinite ? "non-finite" : "very high",
 459                   energyIsNotFinite ? " or Nan" : "");
 460     }
 461 }
 462
 463 /*! \brief Return true if there are special forces computed this step.
 464  *
 465  * The conditionals exactly correspond to those in computeSpecialForces().
 466  */
 467 static bool
 468 haveSpecialForces(const t_inputrec              *inputrec,
 469                   ForceProviders                *forceProviders,
 470                   int                            forceFlags,
 471                   const gmx_edsam               *ed)
 472 {
 473     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 474
 475     return
 476         ((computeForces && forceProviders->hasForceProvider()) ||         // forceProviders
 477          (inputrec->bPull && pull_have_potential(inputrec->pull_work)) || // pull
 478          inputrec->bRot ||                                                // enforced rotation
 479          (ed != nullptr) ||                                               // flooding
 480          (inputrec->bIMD && computeForces));                              // IMD
 481 }
 482
 483 /*! \brief Compute forces and/or energies for special algorithms
 484  *
 485  * The intention is to collect all calls to algorithms that compute
 486  * forces on local atoms only and that do not contribute to the local
 487  * virial sum (but add their virial contribution separately).
 488  * Eventually these should likely all become ForceProviders.
 489  * Within this function the intention is to have algorithms that do
 490  * global communication at the end, so global barriers within the MD loop
 491  * are as close together as possible.
 492  *
 493  * \param[in]     fplog            The log file
 494  * \param[in]     cr               The communication record
 495  * \param[in]     inputrec         The input record
 496  * \param[in]     awh              The Awh module (nullptr if none in use).
 497  * \param[in]     enforcedRotation Enforced rotation module.
 498  * \param[in]     imdSession       The IMD session
 499  * \param[in]     step             The current MD step
 500  * \param[in]     t                The current time
 501  * \param[in,out] wcycle           Wallcycle accounting struct
 502  * \param[in,out] forceProviders   Pointer to a list of force providers
 503  * \param[in]     box              The unit cell
 504  * \param[in]     x                The coordinates
 505  * \param[in]     mdatoms          Per atom properties
 506  * \param[in]     lambda           Array of free-energy lambda values
 507  * \param[in]     forceFlags       Flags that tell whether we should compute forces/energies/virial
 508  * \param[in,out] forceWithVirial  Force and virial buffers
 509  * \param[in,out] enerd            Energy buffer
 510  * \param[in,out] ed               Essential dynamics pointer
 511  * \param[in]     bNS              Tells if we did neighbor searching this step, used for ED sampling
 512  *
 513  * \todo Remove bNS, which is used incorrectly.
 514  * \todo Convert all other algorithms called here to ForceProviders.
 515  */
 516 static void
 517 computeSpecialForces(FILE                          *fplog,
 518                      const t_commrec               *cr,
 519                      const t_inputrec              *inputrec,
 520                      gmx::Awh                      *awh,
 521                      gmx_enfrot                    *enforcedRotation,
 522                      gmx::ImdSession               *imdSession,
 523                      int64_t                        step,
 524                      double                         t,
 525                      gmx_wallcycle_t                wcycle,
 526                      ForceProviders                *forceProviders,
 527                      matrix                         box,
 528                      gmx::ArrayRef<const gmx::RVec> x,
 529                      const t_mdatoms               *mdatoms,
 530                      real                          *lambda,
 531                      int                            forceFlags,
 532                      gmx::ForceWithVirial          *forceWithVirial,
 533                      gmx_enerdata_t                *enerd,
 534                      gmx_edsam                     *ed,
 535                      gmx_bool                       bNS)
 536 {
 537     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 538
 539     /* NOTE: Currently all ForceProviders only provide forces.
 540      *       When they also provide energies, remove this conditional.
 541      */
 542     if (computeForces)
 543     {
 544         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 545         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 546
 547         /* Collect forces from modules */
 548         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 549     }
 550
 551     if (inputrec->bPull && pull_have_potential(inputrec->pull_work))
 552     {
 553         pull_potential_wrapper(cr, inputrec, box, x,
 554                                forceWithVirial,
 555                                mdatoms, enerd, lambda, t,
 556                                wcycle);
 557
 558         if (awh)
 559         {
 560             enerd->term[F_COM_PULL] +=
 561                 awh->applyBiasForcesAndUpdateBias(inputrec->ePBC, *mdatoms, box,
 562                                                   forceWithVirial,
 563                                                   t, step, wcycle, fplog);
 564         }
 565     }
 566
 567     rvec *f = as_rvec_array(forceWithVirial->force_.data());
 568
 569     /* Add the forces from enforced rotation potentials (if any) */
 570     if (inputrec->bRot)
 571     {
 572         wallcycle_start(wcycle, ewcROTadd);
 573         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 574         wallcycle_stop(wcycle, ewcROTadd);
 575     }
 576
 577     if (ed)
 578     {
 579         /* Note that since init_edsam() is called after the initialization
 580          * of forcerec, edsam doesn't request the noVirSum force buffer.
 581          * Thus if no other algorithm (e.g. PME) requires it, the forces
 582          * here will contribute to the virial.
 583          */
 584         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, bNS);
 585     }
 586
 587     /* Add forces from interactive molecular dynamics (IMD), if any */
 588     if (inputrec->bIMD && computeForces)
 589     {
 590         imdSession->applyForces(f);
 591     }
 592 }
 593
 594 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 595  *
 596  * \param[in]  pmedata       The PME structure
 597  * \param[in]  box           The box matrix
 598  * \param[in]  x             Coordinate array
 599  * \param[in]  flags         Force flags
 600  * \param[in]  pmeFlags      PME flags
 601  * \param[in]  wcycle        The wallcycle structure
 602  */
 603 static inline void launchPmeGpuSpread(gmx_pme_t      *pmedata,
 604                                       matrix          box,
 605                                       rvec            x[],
 606                                       int             flags,
 607                                       int             pmeFlags,
 608                                       gmx_wallcycle_t wcycle)
 609 {
 610     pme_gpu_prepare_computation(pmedata, (flags & GMX_FORCE_DYNAMICBOX) != 0, box, wcycle, pmeFlags);
 611     pme_gpu_launch_spread(pmedata, x, wcycle);
 612 }
 613
 614 /*! \brief Launch the FFT and gather stages of PME GPU
 615  *
 616  * This function only implements setting the output forces (no accumulation).
 617  *
 618  * \param[in]  pmedata        The PME structure
 619  * \param[in]  wcycle         The wallcycle structure
 620  */
 621 static void launchPmeGpuFftAndGather(gmx_pme_t        *pmedata,
 622                                      gmx_wallcycle_t   wcycle)
 623 {
 624     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 625     pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
 626 }
 627
 628 /*! \brief
 629  *  Polling wait for either of the PME or nonbonded GPU tasks.
 630  *
 631  * Instead of a static order in waiting for GPU tasks, this function
 632  * polls checking which of the two tasks completes first, and does the
 633  * associated force buffer reduction overlapped with the other task.
 634  * By doing that, unlike static scheduling order, it can always overlap
 635  * one of the reductions, regardless of the GPU task completion order.
 636  *
 637  * \param[in]     nbv              Nonbonded verlet structure
 638  * \param[in,out] pmedata          PME module data
 639  * \param[in,out] force            Force array to reduce task outputs into.
 640  * \param[in,out] forceWithVirial  Force and virial buffers
 641  * \param[in,out] fshift           Shift force output vector results are reduced into
 642  * \param[in,out] enerd            Energy data structure results are reduced into
 643  * \param[in]     flags            Force flags
 644  * \param[in]     pmeFlags         PME flags
 645  * \param[in]     haveOtherWork    Tells whether there is other work than non-bonded in the stream(s)
 646  * \param[in]     wcycle           The wallcycle structure
 647  */
 648 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t                  *nbv,
 649                                         gmx_pme_t                           *pmedata,
 650                                         gmx::ArrayRefWithPadding<gmx::RVec> *force,
 651                                         gmx::ForceWithVirial                *forceWithVirial,
 652                                         rvec                                 fshift[],
 653                                         gmx_enerdata_t                      *enerd,
 654                                         int                                  flags,
 655                                         int                                  pmeFlags,
 656                                         bool                                 haveOtherWork,
 657                                         gmx_wallcycle_t                      wcycle)
 658 {
 659     bool isPmeGpuDone = false;
 660     bool isNbGpuDone  = false;
 661
 662
 663     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 664
 665     while (!isPmeGpuDone || !isNbGpuDone)
 666     {
 667         if (!isPmeGpuDone)
 668         {
 669             GpuTaskCompletion completionType = (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 670             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, forceWithVirial, enerd, completionType);
 671         }
 672
 673         if (!isNbGpuDone)
 674         {
 675             GpuTaskCompletion completionType = (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 676             wallcycle_start_nocount(wcycle, ewcWAIT_GPU_NB_L);
 677             isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
 678                                                      flags,
 679                                                      Nbnxm::AtomLocality::Local,
 680                                                      haveOtherWork,
 681                                                      enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
 682                                                      fshift, completionType);
 683             wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 684             // To get the call count right, when the task finished we
 685             // issue a start/stop.
 686             // TODO: move the ewcWAIT_GPU_NB_L cycle counting into nbnxn_gpu_try_finish_task()
 687             // and ewcNB_XF_BUF_OPS counting into nbnxn_atomdata_add_nbat_f_to_f().
 688             if (isNbGpuDone)
 689             {
 690                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
 691                 wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 692
 693                 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
 694                                               as_rvec_array(force->unpaddedArrayRef().data()), wcycle);
 695             }
 696         }
 697     }
 698 }
 699
 700 /*! \brief Hack structure with force ouput buffers for do_force for the home atoms for this domain */
 701 struct ForceOutputs
 702 {
 703     //! Constructor
 704     ForceOutputs(rvec *f, gmx::ForceWithVirial const forceWithVirial) :
 705         f(f),
 706         forceWithVirial(forceWithVirial) {}
 707
 708     //! Force output buffer used by legacy modules (without SIMD padding)
 709     rvec                 *const f;
 710     //! Force with direct virial contribution (if there are any; without SIMD padding)
 711     gmx::ForceWithVirial        forceWithVirial;
 712 };
 713
 714 /*! \brief Set up the different force buffers; also does clearing.
 715  *
 716  * \param[in] fr        force record pointer
 717  * \param[in] inputrec  input record
 718  * \param[in] force     force array
 719  * \param[in] bDoForces True if force are computed this step
 720  * \param[in] doVirial  True if virial is computed this step
 721  * \param[out] wcycle   wallcycle recording structure
 722  *
 723  * \returns             Cleared force output structure
 724  */
 725 static ForceOutputs
 726 setupForceOutputs(const t_forcerec                    *fr,
 727                   const t_inputrec                    &inputrec,
 728                   gmx::ArrayRefWithPadding<gmx::RVec>  force,
 729                   const bool                           bDoForces,
 730                   const bool                           doVirial,
 731                   gmx_wallcycle_t                      wcycle)
 732 {
 733     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 734
 735     /* Temporary solution until all routines take PaddedRVecVector */
 736     rvec *const f = as_rvec_array(force.unpaddedArrayRef().data());
 737     if (bDoForces)
 738     {
 739         /* Clear the short- and long-range forces */
 740         clear_rvecs_omp(fr->natoms_force_constr, f);
 741     }
 742
 743     /* If we need to compute the virial, we might need a separate
 744      * force buffer for algorithms for which the virial is calculated
 745      * directly, such as PME. Otherwise, forceWithVirial uses the
 746      * the same force (f in legacy calls) buffer as other algorithms.
 747      */
 748     const bool useSeparateForceWithVirialBuffer = (bDoForces && (doVirial && fr->haveDirectVirialContributions));
 749
 750
 751     /* forceWithVirial uses the local atom range only */
 752     gmx::ForceWithVirial forceWithVirial (useSeparateForceWithVirialBuffer ?
 753                                           *fr->forceBufferForDirectVirialContributions : force.unpaddedArrayRef(),
 754                                           doVirial);
 755
 756     if (useSeparateForceWithVirialBuffer)
 757     {
 758         /* TODO: update comment
 759          * We only compute forces on local atoms. Note that vsites can
 760          * spread to non-local atoms, but that part of the buffer is
 761          * cleared separately in the vsite spreading code.
 762          */
 763         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 764     }
 765
 766     if (inputrec.bPull && pull_have_constraint(inputrec.pull_work))
 767     {
 768         clear_pull_forces(inputrec.pull_work);
 769     }
 770
 771     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 772
 773     return ForceOutputs(f, forceWithVirial);
 774 }
 775
 776
 777 /*! \brief Set up flags that indicate what type of work is there to compute.
 778  *
 779  * Currently we only update it at search steps,
 780  * but some properties may change more frequently (e.g. virial/non-virial step),
 781  * so when including those either the frequency of update (per-step) or the scope
 782  * of a flag will change (i.e. a set of flags for nstlist steps).
 783  *
 784  */
 785 static void
 786 setupForceWorkload(gmx::PpForceWorkload *forceWork,
 787                    const t_inputrec     *inputrec,
 788                    const t_forcerec     *fr,
 789                    const gmx_edsam      *ed,
 790                    const t_idef         &idef,
 791                    const t_fcdata       *fcd,
 792                    const int             forceFlags
 793                    )
 794 {
 795     forceWork->haveSpecialForces      = haveSpecialForces(inputrec, fr->forceProviders, forceFlags, ed);
 796     forceWork->haveCpuBondedWork      = haveCpuBondeds(*fr);
 797     forceWork->haveGpuBondedWork      = ((fr->gpuBonded != nullptr) && fr->gpuBonded->haveInteractions());
 798     forceWork->haveRestraintsWork     = havePositionRestraints(idef, *fcd);
 799     forceWork->haveCpuListedForceWork = haveCpuListedForces(*fr, idef, *fcd);
 800 }
 801
 802 void do_force(FILE                                     *fplog,
 803               const t_commrec                          *cr,
 804               const gmx_multisim_t                     *ms,
 805               const t_inputrec                         *inputrec,
 806               gmx::Awh                                 *awh,
 807               gmx_enfrot                               *enforcedRotation,
 808               gmx::ImdSession                          *imdSession,
 809               int64_t                                   step,
 810               t_nrnb                                   *nrnb,
 811               gmx_wallcycle_t                           wcycle,
 812               const gmx_localtop_t                     *top,
 813               matrix                                    box,
 814               gmx::ArrayRefWithPadding<gmx::RVec>       x,     //NOLINT(performance-unnecessary-value-param)
 815               history_t                                *hist,
 816               gmx::ArrayRefWithPadding<gmx::RVec>       force, //NOLINT(performance-unnecessary-value-param)
 817               tensor                                    vir_force,
 818               const t_mdatoms                          *mdatoms,
 819               gmx_enerdata_t                           *enerd,
 820               t_fcdata                                 *fcd,
 821               gmx::ArrayRef<real>                       lambda,
 822               t_graph                                  *graph,
 823               t_forcerec                               *fr,
 824               gmx::PpForceWorkload                     *ppForceWorkload,
 825               const gmx_vsite_t                        *vsite,
 826               rvec                                      mu_tot,
 827               double                                    t,
 828               gmx_edsam                                *ed,
 829               int                                       flags,
 830               const DDBalanceRegionHandler             &ddBalanceRegionHandler)
 831 {
 832     int                  i, j;
 833     double               mu[2*DIM];
 834     gmx_bool             bStateChanged, bNS, bFillGrid, bCalcCGCM;
 835     gmx_bool             bDoForces, bUseGPU, bUseOrEmulGPU;
 836     nonbonded_verlet_t  *nbv = fr->nbv.get();
 837     interaction_const_t *ic  = fr->ic;
 838
 839     /* modify force flag if not doing nonbonded */
 840     if (!fr->bNonbonded)
 841     {
 842         flags &= ~GMX_FORCE_NONBONDED;
 843     }
 844     bStateChanged = ((flags & GMX_FORCE_STATECHANGED) != 0);
 845     bNS           = ((flags & GMX_FORCE_NS) != 0);
 846     bFillGrid     = (bNS && bStateChanged);
 847     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 848     bDoForces     = ((flags & GMX_FORCE_FORCES) != 0);
 849     bUseGPU       = fr->nbv->useGpu();
 850     bUseOrEmulGPU = bUseGPU || fr->nbv->emulateGpu();
 851
 852     const auto pmeRunMode = fr->pmedata ? pme_run_mode(fr->pmedata) : PmeRunMode::CPU;
 853     // TODO slim this conditional down - inputrec and duty checks should mean the same in proper code!
 854     const bool useGpuPme  = EEL_PME(fr->ic->eeltype) && thisRankHasDuty(cr, DUTY_PME) &&
 855         ((pmeRunMode == PmeRunMode::GPU) || (pmeRunMode == PmeRunMode::Mixed));
 856     const int  pmeFlags = GMX_PME_SPREAD | GMX_PME_SOLVE |
 857         ((flags & GMX_FORCE_VIRIAL) ? GMX_PME_CALC_ENER_VIR : 0) |
 858         ((flags & GMX_FORCE_ENERGY) ? GMX_PME_CALC_ENER_VIR : 0) |
 859         ((flags & GMX_FORCE_FORCES) ? GMX_PME_CALC_F : 0);
 860
 861     /* At a search step we need to start the first balancing region
 862      * somewhere early inside the step after communication during domain
 863      * decomposition (and not during the previous step as usual).
 864      */
 865     if (bNS)
 866     {
 867         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 868     }
 869
 870     const int start  = 0;
 871     const int homenr = mdatoms->homenr;
 872
 873     clear_mat(vir_force);
 874
 875     if (bStateChanged)
 876     {
 877         update_forcerec(fr, box);
 878
 879         if (inputrecNeedMutot(inputrec))
 880         {
 881             /* Calculate total (local) dipole moment in a temporary common array.
 882              * This makes it possible to sum them over nodes faster.
 883              */
 884             calc_mu(start, homenr,
 885                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 886                     mu, mu+DIM);
 887         }
 888     }
 889
 890     if (fr->ePBC != epbcNONE)
 891     {
 892         /* Compute shift vectors every step,
 893          * because of pressure coupling or box deformation!
 894          */
 895         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 896         {
 897             calc_shifts(box, fr->shift_vec);
 898         }
 899
 900         if (bCalcCGCM)
 901         {
 902             put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr), gmx_omp_nthreads_get(emntDefault));
 903             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 904         }
 905         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 906         {
 907             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
 908         }
 909     }
 910
 911     nbnxn_atomdata_copy_shiftvec((flags & GMX_FORCE_DYNAMICBOX) != 0,
 912                                  fr->shift_vec, nbv->nbat.get());
 913
 914 #if GMX_MPI
 915     if (!thisRankHasDuty(cr, DUTY_PME))
 916     {
 917         /* Send particle coordinates to the pme nodes.
 918          * Since this is only implemented for domain decomposition
 919          * and domain decomposition does not use the graph,
 920          * we do not need to worry about shifting.
 921          */
 922         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
 923                                  lambda[efptCOUL], lambda[efptVDW],
 924                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
 925                                  step, wcycle);
 926     }
 927 #endif /* GMX_MPI */
 928
 929     if (useGpuPme)
 930     {
 931         launchPmeGpuSpread(fr->pmedata, box, as_rvec_array(x.unpaddedArrayRef().data()), flags, pmeFlags, wcycle);
 932     }
 933
 934     /* do gridding for pair search */
 935     if (bNS)
 936     {
 937         if (graph && bStateChanged)
 938         {
 939             /* Calculate intramolecular shift vectors to make molecules whole */
 940             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
 941         }
 942
 943         // TODO
 944         // - vzero is constant, do we need to pass it?
 945         // - box_diag should be passed directly to nbnxn_put_on_grid
 946         //
 947         rvec vzero;
 948         clear_rvec(vzero);
 949
 950         rvec box_diag;
 951         box_diag[XX] = box[XX][XX];
 952         box_diag[YY] = box[YY][YY];
 953         box_diag[ZZ] = box[ZZ][ZZ];
 954
 955         wallcycle_start(wcycle, ewcNS);
 956         if (!DOMAINDECOMP(cr))
 957         {
 958             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 959             nbnxn_put_on_grid(nbv, box,
 960                               0, vzero, box_diag,
 961                               nullptr, 0, mdatoms->homenr, -1,
 962                               fr->cginfo, x.unpaddedArrayRef(),
 963                               0, nullptr);
 964             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 965         }
 966         else
 967         {
 968             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 969             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
 970                                        fr->cginfo, x.unpaddedArrayRef());
 971             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 972         }
 973
 974         nbv->setAtomProperties(*mdatoms, *fr->cginfo);
 975
 976         wallcycle_stop(wcycle, ewcNS);
 977
 978         /* initialize the GPU nbnxm atom data and bonded data structures */
 979         if (bUseGPU)
 980         {
 981             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 982
 983             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 984             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
 985             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 986
 987             if (fr->gpuBonded)
 988             {
 989                 /* Now we put all atoms on the grid, we can assign bonded
 990                  * interactions to the GPU, where the grid order is
 991                  * needed. Also the xq, f and fshift device buffers have
 992                  * been reallocated if needed, so the bonded code can
 993                  * learn about them. */
 994                 // TODO the xq, f, and fshift buffers are now shared
 995                 // resources, so they should be maintained by a
 996                 // higher-level object than the nb module.
 997                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbv->getGridIndices(),
 998                                                                       top->idef,
 999                                                                       Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1000                                                                       Nbnxm::gpu_get_f(nbv->gpu_nbv),
1001                                                                       Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1002             }
1003             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1004         }
1005
1006         // Need to run after the GPU-offload bonded interaction lists
1007         // are set up to be able to determine whether there is bonded work.
1008         setupForceWorkload(ppForceWorkload,
1009                            inputrec,
1010                            fr,
1011                            ed,
1012                            top->idef,
1013                            fcd,
1014                            flags);
1015     }
1016
1017     /* do local pair search */
1018     if (bNS)
1019     {
1020         // TODO: fuse this branch with the above bNS block
1021         wallcycle_start_nocount(wcycle, ewcNS);
1022         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1023         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1024         nbv->constructPairlist(Nbnxm::InteractionLocality::Local,
1025                                &top->excls, step, nrnb);
1026         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1027         wallcycle_stop(wcycle, ewcNS);
1028     }
1029     else
1030     {
1031         nbv->setCoordinates(Nbnxm::AtomLocality::Local, false,
1032                             x.unpaddedArrayRef(), wcycle);
1033     }
1034
1035     if (bUseGPU)
1036     {
1037         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1038
1039         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1040
1041         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1042         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1043         Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1044                                   Nbnxm::AtomLocality::Local,
1045                                   ppForceWorkload->haveGpuBondedWork);
1046         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1047
1048         // bonded work not split into separate local and non-local, so with DD
1049         // we can only launch the kernel after non-local coordinates have been received.
1050         if (ppForceWorkload->haveGpuBondedWork && !havePPDomainDecomposition(cr))
1051         {
1052             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1053             fr->gpuBonded->launchKernels(fr, flags, box);
1054             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1055         }
1056
1057         /* launch local nonbonded work on GPU */
1058         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1059         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFNo,
1060                      step, nrnb, wcycle);
1061         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1062         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1063     }
1064
1065     if (useGpuPme)
1066     {
1067         // In PME GPU and mixed mode we launch FFT / gather after the
1068         // X copy/transform to allow overlap as well as after the GPU NB
1069         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1070         // the nonbonded kernel.
1071         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1072     }
1073
1074     /* Communicate coordinates and sum dipole if necessary +
1075        do non-local pair search */
1076     if (havePPDomainDecomposition(cr))
1077     {
1078         if (bNS)
1079         {
1080             // TODO: fuse this branch with the above large bNS block
1081             wallcycle_start_nocount(wcycle, ewcNS);
1082             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1083             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1084             nbv->constructPairlist(Nbnxm::InteractionLocality::NonLocal,
1085                                    &top->excls, step, nrnb);
1086             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1087             wallcycle_stop(wcycle, ewcNS);
1088         }
1089         else
1090         {
1091             dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1092
1093             nbv->setCoordinates(Nbnxm::AtomLocality::NonLocal, false,
1094                                 x.unpaddedArrayRef(), wcycle);
1095         }
1096
1097         if (bUseGPU)
1098         {
1099             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1100
1101             /* launch non-local nonbonded tasks on GPU */
1102             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1103             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1104                                       Nbnxm::AtomLocality::NonLocal,
1105                                       ppForceWorkload->haveGpuBondedWork);
1106             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1107
1108             if (ppForceWorkload->haveGpuBondedWork)
1109             {
1110                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1111                 fr->gpuBonded->launchKernels(fr, flags, box);
1112                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1113             }
1114
1115             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1116             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1117                          step, nrnb, wcycle);
1118             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1119
1120             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1121         }
1122     }
1123
1124     if (bUseGPU)
1125     {
1126         /* launch D2H copy-back F */
1127         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1128         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1129         if (havePPDomainDecomposition(cr))
1130         {
1131             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1132                                       flags, Nbnxm::AtomLocality::NonLocal, ppForceWorkload->haveGpuBondedWork);
1133         }
1134         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1135                                   flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork);
1136         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1137
1138         if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1139         {
1140             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1141             fr->gpuBonded->launchEnergyTransfer();
1142             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1143         }
1144         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1145     }
1146
1147     if (bStateChanged && inputrecNeedMutot(inputrec))
1148     {
1149         if (PAR(cr))
1150         {
1151             gmx_sumd(2*DIM, mu, cr);
1152
1153             ddBalanceRegionHandler.reopenRegionCpu();
1154         }
1155
1156         for (i = 0; i < 2; i++)
1157         {
1158             for (j = 0; j < DIM; j++)
1159             {
1160                 fr->mu_tot[i][j] = mu[i*DIM + j];
1161             }
1162         }
1163     }
1164     if (fr->efep == efepNO)
1165     {
1166         copy_rvec(fr->mu_tot[0], mu_tot);
1167     }
1168     else
1169     {
1170         for (j = 0; j < DIM; j++)
1171         {
1172             mu_tot[j] =
1173                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1174                 lambda[efptCOUL]*fr->mu_tot[1][j];
1175         }
1176     }
1177
1178     /* Reset energies */
1179     reset_enerdata(enerd);
1180     clear_rvecs(SHIFTS, fr->fshift);
1181
1182     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1183     {
1184         wallcycle_start(wcycle, ewcPPDURINGPME);
1185         dd_force_flop_start(cr->dd, nrnb);
1186     }
1187
1188     if (inputrec->bRot)
1189     {
1190         wallcycle_start(wcycle, ewcROT);
1191         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1192         wallcycle_stop(wcycle, ewcROT);
1193     }
1194
1195     /* Start the force cycle counter.
1196      * Note that a different counter is used for dynamic load balancing.
1197      */
1198     wallcycle_start(wcycle, ewcFORCE);
1199
1200     // set up and clear force outputs
1201     struct ForceOutputs forceOut = setupForceOutputs(fr, *inputrec, force, bDoForces, ((flags & GMX_FORCE_VIRIAL) != 0), wcycle);
1202
1203     /* We calculate the non-bonded forces, when done on the CPU, here.
1204      * We do this before calling do_force_lowlevel, because in that
1205      * function, the listed forces are calculated before PME, which
1206      * does communication.  With this order, non-bonded and listed
1207      * force calculation imbalance can be balanced out by the domain
1208      * decomposition load balancing.
1209      */
1210
1211     if (!bUseOrEmulGPU)
1212     {
1213         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFYes,
1214                      step, nrnb, wcycle);
1215     }
1216
1217     if (fr->efep != efepNO)
1218     {
1219         /* Calculate the local and non-local free energy interactions here.
1220          * Happens here on the CPU both with and without GPU.
1221          */
1222         wallcycle_sub_start(wcycle, ewcsNONBONDED);
1223         nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::Local,
1224                                       fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1225                                       inputrec->fepvals, lambda.data(),
1226                                       enerd, flags, nrnb);
1227
1228         if (havePPDomainDecomposition(cr))
1229         {
1230             nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::NonLocal,
1231                                           fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1232                                           inputrec->fepvals, lambda.data(),
1233                                           enerd, flags, nrnb);
1234         }
1235         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
1236     }
1237
1238     if (!bUseOrEmulGPU)
1239     {
1240         if (havePPDomainDecomposition(cr))
1241         {
1242             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1243                          step, nrnb, wcycle);
1244         }
1245
1246         /* Add all the non-bonded force to the normal force array.
1247          * This can be split into a local and a non-local part when overlapping
1248          * communication with calculation with domain decomposition.
1249          */
1250         wallcycle_stop(wcycle, ewcFORCE);
1251
1252         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::All, forceOut.f, wcycle);
1253
1254         wallcycle_start_nocount(wcycle, ewcFORCE);
1255
1256         /* If there are multiple fshift output buffers we need to reduce them */
1257         if (flags & GMX_FORCE_VIRIAL)
1258         {
1259             /* This is not in a subcounter because it takes a
1260                negligible and constant-sized amount of time */
1261             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat.get(),
1262                                                      fr->fshift);
1263         }
1264     }
1265
1266     /* update QMMMrec, if necessary */
1267     if (fr->bQMMM)
1268     {
1269         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1270     }
1271
1272     /* Compute the bonded and non-bonded energies and optionally forces */
1273     do_force_lowlevel(fr, inputrec, &(top->idef),
1274                       cr, ms, nrnb, wcycle, mdatoms,
1275                       as_rvec_array(x.unpaddedArrayRef().data()), hist, forceOut.f, &forceOut.forceWithVirial, enerd, fcd,
1276                       box, inputrec->fepvals, lambda.data(), graph, &(top->excls), fr->mu_tot,
1277                       flags,
1278                       ddBalanceRegionHandler);
1279
1280     wallcycle_stop(wcycle, ewcFORCE);
1281
1282     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1283                          imdSession, step, t, wcycle,
1284                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1285                          flags, &forceOut.forceWithVirial, enerd,
1286                          ed, bNS);
1287
1288     // Will store the amount of cycles spent waiting for the GPU that
1289     // will be later used in the DLB accounting.
1290     float cycles_wait_gpu = 0;
1291     if (bUseOrEmulGPU)
1292     {
1293         /* wait for non-local forces (or calculate in emulation mode) */
1294         if (havePPDomainDecomposition(cr))
1295         {
1296             if (bUseGPU)
1297             {
1298                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1299                 Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1300                                             flags, Nbnxm::AtomLocality::NonLocal,
1301                                             ppForceWorkload->haveGpuBondedWork,
1302                                             enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1303                                             fr->fshift);
1304                 cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1305             }
1306             else
1307             {
1308                 wallcycle_start_nocount(wcycle, ewcFORCE);
1309                 do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFYes,
1310                              step, nrnb, wcycle);
1311                 wallcycle_stop(wcycle, ewcFORCE);
1312             }
1313
1314             nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::NonLocal,
1315                                           forceOut.f, wcycle);
1316         }
1317     }
1318
1319     if (havePPDomainDecomposition(cr))
1320     {
1321         /* We are done with the CPU compute.
1322          * We will now communicate the non-local forces.
1323          * If we use a GPU this will overlap with GPU work, so in that case
1324          * we do not close the DD force balancing region here.
1325          */
1326         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1327
1328         if (bDoForces)
1329         {
1330             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1331         }
1332     }
1333
1334     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1335     // an alternating wait/reduction scheme.
1336     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPme && bUseGPU && !DOMAINDECOMP(cr));
1337     if (alternateGpuWait)
1338     {
1339         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &force, &forceOut.forceWithVirial, fr->fshift, enerd,
1340                                     flags, pmeFlags, ppForceWorkload->haveGpuBondedWork, wcycle);
1341     }
1342
1343     if (!alternateGpuWait && useGpuPme)
1344     {
1345         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial, enerd);
1346     }
1347
1348     /* Wait for local GPU NB outputs on the non-alternating wait path */
1349     if (!alternateGpuWait && bUseGPU)
1350     {
1351         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1352          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1353          * but even with a step of 0.1 ms the difference is less than 1%
1354          * of the step time.
1355          */
1356         const float gpuWaitApiOverheadMargin = 2e6f; /* cycles */
1357
1358         wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1359         Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1360                                     flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork,
1361                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1362                                     fr->fshift);
1363         float cycles_tmp = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1364
1365         if (ddBalanceRegionHandler.useBalancingRegion())
1366         {
1367             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1368             if (bDoForces && cycles_tmp <= gpuWaitApiOverheadMargin)
1369             {
1370                 /* We measured few cycles, it could be that the kernel
1371                  * and transfer finished earlier and there was no actual
1372                  * wait time, only API call overhead.
1373                  * Then the actual time could be anywhere between 0 and
1374                  * cycles_wait_est. We will use half of cycles_wait_est.
1375                  */
1376                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1377             }
1378             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1379         }
1380     }
1381
1382     if (fr->nbv->emulateGpu())
1383     {
1384         // NOTE: emulation kernel is not included in the balancing region,
1385         // but emulation mode does not target performance anyway
1386         wallcycle_start_nocount(wcycle, ewcFORCE);
1387         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local,
1388                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1389                      step, nrnb, wcycle);
1390         wallcycle_stop(wcycle, ewcFORCE);
1391     }
1392
1393     if (useGpuPme)
1394     {
1395         pme_gpu_reinit_computation(fr->pmedata, wcycle);
1396     }
1397
1398     if (bUseGPU)
1399     {
1400         /* now clear the GPU outputs while we finish the step on the CPU */
1401         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1402         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1403         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, flags);
1404
1405         if (nbv->isDynamicPruningStepGpu(step))
1406         {
1407             nbv->dispatchPruneKernelGpu(step);
1408         }
1409         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1410         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1411     }
1412
1413     if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1414     {
1415         wallcycle_start(wcycle, ewcWAIT_GPU_BONDED);
1416         // in principle this should be included in the DD balancing region,
1417         // but generally it is infrequent so we'll omit it for the sake of
1418         // simpler code
1419         fr->gpuBonded->accumulateEnergyTerms(enerd);
1420         wallcycle_stop(wcycle, ewcWAIT_GPU_BONDED);
1421
1422         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1423         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1424         fr->gpuBonded->clearEnergies();
1425         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1426         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1427     }
1428
1429     /* Do the nonbonded GPU (or emulation) force buffer reduction
1430      * on the non-alternating path. */
1431     if (bUseOrEmulGPU && !alternateGpuWait)
1432     {
1433         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
1434                                       forceOut.f, wcycle);
1435     }
1436     if (DOMAINDECOMP(cr))
1437     {
1438         dd_force_flop_stop(cr->dd, nrnb);
1439     }
1440
1441     if (bDoForces)
1442     {
1443         /* If we have NoVirSum forces, but we do not calculate the virial,
1444          * we sum fr->f_novirsum=forceOut.f later.
1445          */
1446         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1447         {
1448             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, fr->fshift, FALSE, nullptr, nrnb,
1449                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1450         }
1451
1452         if (flags & GMX_FORCE_VIRIAL)
1453         {
1454             /* Calculation of the virial must be done after vsites! */
1455             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f,
1456                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1457         }
1458     }
1459
1460     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1461     {
1462         /* In case of node-splitting, the PP nodes receive the long-range
1463          * forces, virial and energy from the PME nodes here.
1464          */
1465         pme_receive_force_ener(cr, &forceOut.forceWithVirial, enerd, wcycle);
1466     }
1467
1468     if (bDoForces)
1469     {
1470         post_process_forces(cr, step, nrnb, wcycle,
1471                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, &forceOut.forceWithVirial,
1472                             vir_force, mdatoms, graph, fr, vsite,
1473                             flags);
1474     }
1475
1476     if (flags & GMX_FORCE_ENERGY)
1477     {
1478         /* Sum the potential energy terms from group contributions */
1479         sum_epot(&(enerd->grpp), enerd->term);
1480
1481         if (!EI_TPI(inputrec->eI))
1482         {
1483             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1484         }
1485     }
1486
1487     /* In case we don't have constraints and are using GPUs, the next balancing
1488      * region starts here.
1489      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1490      * virial calculation and COM pulling, is not thus not included in
1491      * the balance timing, which is ok as most tasks do communication.
1492      */
1493     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1494 }