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 "sim_util.h"
  40
  41 #include "config.h"
  42
  43 #include <cmath>
  44 #include <cstdint>
  45 #include <cstdio>
  46 #include <cstring>
  47
  48 #include <array>
  49
  50 #include "gromacs/awh/awh.h"
  51 #include "gromacs/domdec/dlbtiming.h"
  52 #include "gromacs/domdec/domdec.h"
  53 #include "gromacs/domdec/domdec_struct.h"
  54 #include "gromacs/domdec/partition.h"
  55 #include "gromacs/essentialdynamics/edsam.h"
  56 #include "gromacs/ewald/pme.h"
  57 #include "gromacs/gmxlib/chargegroup.h"
  58 #include "gromacs/gmxlib/network.h"
  59 #include "gromacs/gmxlib/nrnb.h"
  60 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
  61 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
  62 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
  63 #include "gromacs/gpu_utils/gpu_utils.h"
  64 #include "gromacs/imd/imd.h"
  65 #include "gromacs/listed_forces/bonded.h"
  66 #include "gromacs/listed_forces/disre.h"
  67 #include "gromacs/listed_forces/gpubonded.h"
  68 #include "gromacs/listed_forces/manage_threading.h"
  69 #include "gromacs/listed_forces/orires.h"
  70 #include "gromacs/math/arrayrefwithpadding.h"
  71 #include "gromacs/math/functions.h"
  72 #include "gromacs/math/units.h"
  73 #include "gromacs/math/vec.h"
  74 #include "gromacs/math/vecdump.h"
  75 #include "gromacs/mdlib/calcmu.h"
  76 #include "gromacs/mdlib/calcvir.h"
  77 #include "gromacs/mdlib/constr.h"
  78 #include "gromacs/mdlib/force.h"
  79 #include "gromacs/mdlib/forcerec.h"
  80 #include "gromacs/mdlib/gmx_omp_nthreads.h"
  81 #include "gromacs/mdlib/ppforceworkload.h"
  82 #include "gromacs/mdlib/qmmm.h"
  83 #include "gromacs/mdlib/update.h"
  84 #include "gromacs/mdtypes/commrec.h"
  85 #include "gromacs/mdtypes/enerdata.h"
  86 #include "gromacs/mdtypes/forceoutput.h"
  87 #include "gromacs/mdtypes/iforceprovider.h"
  88 #include "gromacs/mdtypes/inputrec.h"
  89 #include "gromacs/mdtypes/md_enums.h"
  90 #include "gromacs/mdtypes/state.h"
  91 #include "gromacs/nbnxm/atomdata.h"
  92 #include "gromacs/nbnxm/gpu_data_mgmt.h"
  93 #include "gromacs/nbnxm/nbnxm.h"
  94 #include "gromacs/pbcutil/ishift.h"
  95 #include "gromacs/pbcutil/mshift.h"
  96 #include "gromacs/pbcutil/pbc.h"
  97 #include "gromacs/pulling/pull.h"
  98 #include "gromacs/pulling/pull_rotation.h"
  99 #include "gromacs/timing/cyclecounter.h"
 100 #include "gromacs/timing/gpu_timing.h"
 101 #include "gromacs/timing/wallcycle.h"
 102 #include "gromacs/timing/wallcyclereporting.h"
 103 #include "gromacs/timing/walltime_accounting.h"
 104 #include "gromacs/topology/topology.h"
 105 #include "gromacs/utility/arrayref.h"
 106 #include "gromacs/utility/basedefinitions.h"
 107 #include "gromacs/utility/cstringutil.h"
 108 #include "gromacs/utility/exceptions.h"
 109 #include "gromacs/utility/fatalerror.h"
 110 #include "gromacs/utility/gmxassert.h"
 111 #include "gromacs/utility/gmxmpi.h"
 112 #include "gromacs/utility/logger.h"
 113 #include "gromacs/utility/smalloc.h"
 114 #include "gromacs/utility/strconvert.h"
 115 #include "gromacs/utility/sysinfo.h"
 116
 117 // TODO: this environment variable allows us to verify before release
 118 // that on less common architectures the total cost of polling is not larger than
 119 // a blocking wait (so polling does not introduce overhead when the static
 120 // PME-first ordering would suffice).
 121 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
 122
 123
 124 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 125 {
 126     const int      end = forceToAdd.size();
 127
 128     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 129 #pragma omp parallel for num_threads(nt) schedule(static)
 130     for (int i = 0; i < end; i++)
 131     {
 132         rvec_inc(f[i], forceToAdd[i]);
 133     }
 134 }
 135
 136 static void calc_virial(int start, int homenr, const rvec x[], const rvec f[],
 137                         tensor vir_part, const t_graph *graph, const matrix box,
 138                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 139 {
 140     /* The short-range virial from surrounding boxes */
 141     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 142     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 143
 144     /* Calculate partial virial, for local atoms only, based on short range.
 145      * Total virial is computed in global_stat, called from do_md
 146      */
 147     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 148     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 149
 150     if (debug)
 151     {
 152         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 153     }
 154 }
 155
 156 static void pull_potential_wrapper(const t_commrec *cr,
 157                                    const t_inputrec *ir,
 158                                    const matrix box, gmx::ArrayRef<const gmx::RVec> x,
 159                                    gmx::ForceWithVirial *force,
 160                                    const t_mdatoms *mdatoms,
 161                                    gmx_enerdata_t *enerd,
 162                                    const real *lambda,
 163                                    double t,
 164                                    gmx_wallcycle_t wcycle)
 165 {
 166     t_pbc  pbc;
 167     real   dvdl;
 168
 169     /* Calculate the center of mass forces, this requires communication,
 170      * which is why pull_potential is called close to other communication.
 171      */
 172     wallcycle_start(wcycle, ewcPULLPOT);
 173     set_pbc(&pbc, ir->ePBC, box);
 174     dvdl                     = 0;
 175     enerd->term[F_COM_PULL] +=
 176         pull_potential(ir->pull_work, mdatoms, &pbc,
 177                        cr, t, lambda[efptRESTRAINT], as_rvec_array(x.data()), force, &dvdl);
 178     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 179     wallcycle_stop(wcycle, ewcPULLPOT);
 180 }
 181
 182 static void pme_receive_force_ener(const t_commrec      *cr,
 183                                    gmx::ForceWithVirial *forceWithVirial,
 184                                    gmx_enerdata_t       *enerd,
 185                                    gmx_wallcycle_t       wcycle)
 186 {
 187     real   e_q, e_lj, dvdl_q, dvdl_lj;
 188     float  cycles_ppdpme, cycles_seppme;
 189
 190     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 191     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 192
 193     /* In case of node-splitting, the PP nodes receive the long-range
 194      * forces, virial and energy from the PME nodes here.
 195      */
 196     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 197     dvdl_q  = 0;
 198     dvdl_lj = 0;
 199     gmx_pme_receive_f(cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 200                       &cycles_seppme);
 201     enerd->term[F_COUL_RECIP] += e_q;
 202     enerd->term[F_LJ_RECIP]   += e_lj;
 203     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 204     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 205
 206     if (wcycle)
 207     {
 208         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 209     }
 210     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 211 }
 212
 213 static void print_large_forces(FILE            *fp,
 214                                const t_mdatoms *md,
 215                                const t_commrec *cr,
 216                                int64_t          step,
 217                                real             forceTolerance,
 218                                const rvec      *x,
 219                                const rvec      *f)
 220 {
 221     real           force2Tolerance = gmx::square(forceTolerance);
 222     gmx::index     numNonFinite    = 0;
 223     for (int i = 0; i < md->homenr; i++)
 224     {
 225         real force2    = norm2(f[i]);
 226         bool nonFinite = !std::isfinite(force2);
 227         if (force2 >= force2Tolerance || nonFinite)
 228         {
 229             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 230                     step,
 231                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 232         }
 233         if (nonFinite)
 234         {
 235             numNonFinite++;
 236         }
 237     }
 238     if (numNonFinite > 0)
 239     {
 240         /* Note that with MPI this fatal call on one rank might interrupt
 241          * the printing on other ranks. But we can only avoid that with
 242          * an expensive MPI barrier that we would need at each step.
 243          */
 244         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 245     }
 246 }
 247
 248 static void post_process_forces(const t_commrec           *cr,
 249                                 int64_t                    step,
 250                                 t_nrnb                    *nrnb,
 251                                 gmx_wallcycle_t            wcycle,
 252                                 const gmx_localtop_t      *top,
 253                                 const matrix               box,
 254                                 const rvec                 x[],
 255                                 rvec                       f[],
 256                                 gmx::ForceWithVirial      *forceWithVirial,
 257                                 tensor                     vir_force,
 258                                 const t_mdatoms           *mdatoms,
 259                                 const t_graph             *graph,
 260                                 const t_forcerec          *fr,
 261                                 const gmx_vsite_t         *vsite,
 262                                 int                        flags)
 263 {
 264     if (fr->haveDirectVirialContributions)
 265     {
 266         rvec *fDirectVir = as_rvec_array(forceWithVirial->force_.data());
 267
 268         if (vsite)
 269         {
 270             /* Spread the mesh force on virtual sites to the other particles...
 271              * This is parallellized. MPI communication is performed
 272              * if the constructing atoms aren't local.
 273              */
 274             matrix virial = { { 0 } };
 275             spread_vsite_f(vsite, x, fDirectVir, nullptr,
 276                            (flags & GMX_FORCE_VIRIAL) != 0, virial,
 277                            nrnb,
 278                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
 279             forceWithVirial->addVirialContribution(virial);
 280         }
 281
 282         if (flags & GMX_FORCE_VIRIAL)
 283         {
 284             /* Now add the forces, this is local */
 285             sum_forces(f, forceWithVirial->force_);
 286
 287             /* Add the direct virial contributions */
 288             GMX_ASSERT(forceWithVirial->computeVirial_, "forceWithVirial should request virial computation when we request the virial");
 289             m_add(vir_force, forceWithVirial->getVirial(), vir_force);
 290
 291             if (debug)
 292             {
 293                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 294             }
 295         }
 296     }
 297
 298     if (fr->print_force >= 0)
 299     {
 300         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 301     }
 302 }
 303
 304 static void do_nb_verlet(t_forcerec                       *fr,
 305                          const interaction_const_t        *ic,
 306                          gmx_enerdata_t                   *enerd,
 307                          const int                         flags,
 308                          const Nbnxm::InteractionLocality  ilocality,
 309                          const int                         clearF,
 310                          const int64_t                     step,
 311                          t_nrnb                           *nrnb,
 312                          gmx_wallcycle_t                   wcycle)
 313 {
 314     if (!(flags & GMX_FORCE_NONBONDED))
 315     {
 316         /* skip non-bonded calculation */
 317         return;
 318     }
 319
 320     nonbonded_verlet_t *nbv  = fr->nbv.get();
 321
 322     /* GPU kernel launch overhead is already timed separately */
 323     if (fr->cutoff_scheme != ecutsVERLET)
 324     {
 325         gmx_incons("Invalid cut-off scheme passed!");
 326     }
 327
 328     if (!nbv->useGpu())
 329     {
 330         /* When dynamic pair-list  pruning is requested, we need to prune
 331          * at nstlistPrune steps.
 332          */
 333         if (nbv->pairlistSets().isDynamicPruningStepCpu(step))
 334         {
 335             /* Prune the pair-list beyond fr->ic->rlistPrune using
 336              * the current coordinates of the atoms.
 337              */
 338             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 339             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 340             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 341         }
 342
 343         wallcycle_sub_start(wcycle, ewcsNONBONDED);
 344     }
 345
 346     nbv->dispatchNonbondedKernel(ilocality, *ic, flags, clearF, fr, enerd, nrnb);
 347
 348     if (!nbv->useGpu())
 349     {
 350         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
 351     }
 352 }
 353
 354 gmx_bool use_GPU(const nonbonded_verlet_t *nbv)
 355 {
 356     return nbv != nullptr && nbv->useGpu();
 357 }
 358
 359 static inline void clear_rvecs_omp(int n, rvec v[])
 360 {
 361     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 362
 363     /* Note that we would like to avoid this conditional by putting it
 364      * into the omp pragma instead, but then we still take the full
 365      * omp parallel for overhead (at least with gcc5).
 366      */
 367     if (nth == 1)
 368     {
 369         for (int i = 0; i < n; i++)
 370         {
 371             clear_rvec(v[i]);
 372         }
 373     }
 374     else
 375     {
 376 #pragma omp parallel for num_threads(nth) schedule(static)
 377         for (int i = 0; i < n; i++)
 378         {
 379             clear_rvec(v[i]);
 380         }
 381     }
 382 }
 383
 384 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 385  *
 386  * \param groupOptions  Group options, containing T-coupling options
 387  */
 388 static real averageKineticEnergyEstimate(const t_grpopts &groupOptions)
 389 {
 390     real nrdfCoupled   = 0;
 391     real nrdfUncoupled = 0;
 392     real kineticEnergy = 0;
 393     for (int g = 0; g < groupOptions.ngtc; g++)
 394     {
 395         if (groupOptions.tau_t[g] >= 0)
 396         {
 397             nrdfCoupled   += groupOptions.nrdf[g];
 398             kineticEnergy += groupOptions.nrdf[g]*0.5*groupOptions.ref_t[g]*BOLTZ;
 399         }
 400         else
 401         {
 402             nrdfUncoupled += groupOptions.nrdf[g];
 403         }
 404     }
 405
 406     /* This conditional with > also catches nrdf=0 */
 407     if (nrdfCoupled > nrdfUncoupled)
 408     {
 409         return kineticEnergy*(nrdfCoupled + nrdfUncoupled)/nrdfCoupled;
 410     }
 411     else
 412     {
 413         return 0;
 414     }
 415 }
 416
 417 /*! \brief This routine checks that the potential energy is finite.
 418  *
 419  * Always checks that the potential energy is finite. If step equals
 420  * inputrec.init_step also checks that the magnitude of the potential energy
 421  * is reasonable. Terminates with a fatal error when a check fails.
 422  * Note that passing this check does not guarantee finite forces,
 423  * since those use slightly different arithmetics. But in most cases
 424  * there is just a narrow coordinate range where forces are not finite
 425  * and energies are finite.
 426  *
 427  * \param[in] step      The step number, used for checking and printing
 428  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to enerd.term[F_EPOT] before calling this routine
 429  * \param[in] inputrec  The input record
 430  */
 431 static void checkPotentialEnergyValidity(int64_t               step,
 432                                          const gmx_enerdata_t &enerd,
 433                                          const t_inputrec     &inputrec)
 434 {
 435     /* Threshold valid for comparing absolute potential energy against
 436      * the kinetic energy. Normally one should not consider absolute
 437      * potential energy values, but with a factor of one million
 438      * we should never get false positives.
 439      */
 440     constexpr real c_thresholdFactor = 1e6;
 441
 442     bool           energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 443     real           averageKineticEnergy = 0;
 444     /* We only check for large potential energy at the initial step,
 445      * because that is by far the most likely step for this too occur
 446      * and because computing the average kinetic energy is not free.
 447      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 448      * before they become NaN.
 449      */
 450     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 451     {
 452         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 453     }
 454
 455     if (energyIsNotFinite || (averageKineticEnergy > 0 &&
 456                               enerd.term[F_EPOT] > c_thresholdFactor*averageKineticEnergy))
 457     {
 458         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.",
 459                   step,
 460                   enerd.term[F_EPOT],
 461                   energyIsNotFinite ? "not finite" : "extremely high",
 462                   enerd.term[F_LJ],
 463                   enerd.term[F_COUL_SR],
 464                   energyIsNotFinite ? "non-finite" : "very high",
 465                   energyIsNotFinite ? " or Nan" : "");
 466     }
 467 }
 468
 469 /*! \brief Compute forces and/or energies for special algorithms
 470  *
 471  * The intention is to collect all calls to algorithms that compute
 472  * forces on local atoms only and that do not contribute to the local
 473  * virial sum (but add their virial contribution separately).
 474  * Eventually these should likely all become ForceProviders.
 475  * Within this function the intention is to have algorithms that do
 476  * global communication at the end, so global barriers within the MD loop
 477  * are as close together as possible.
 478  *
 479  * \param[in]     fplog            The log file
 480  * \param[in]     cr               The communication record
 481  * \param[in]     inputrec         The input record
 482  * \param[in]     awh              The Awh module (nullptr if none in use).
 483  * \param[in]     enforcedRotation Enforced rotation module.
 484  * \param[in]     step             The current MD step
 485  * \param[in]     t                The current time
 486  * \param[in,out] wcycle           Wallcycle accounting struct
 487  * \param[in,out] forceProviders   Pointer to a list of force providers
 488  * \param[in]     box              The unit cell
 489  * \param[in]     x                The coordinates
 490  * \param[in]     mdatoms          Per atom properties
 491  * \param[in]     lambda           Array of free-energy lambda values
 492  * \param[in]     forceFlags       Flags that tell whether we should compute forces/energies/virial
 493  * \param[in,out] forceWithVirial  Force and virial buffers
 494  * \param[in,out] enerd            Energy buffer
 495  * \param[in,out] ed               Essential dynamics pointer
 496  * \param[in]     bNS              Tells if we did neighbor searching this step, used for ED sampling
 497  *
 498  * \todo Remove bNS, which is used incorrectly.
 499  * \todo Convert all other algorithms called here to ForceProviders.
 500  */
 501 static void
 502 computeSpecialForces(FILE                          *fplog,
 503                      const t_commrec               *cr,
 504                      const t_inputrec              *inputrec,
 505                      gmx::Awh                      *awh,
 506                      gmx_enfrot                    *enforcedRotation,
 507                      int64_t                        step,
 508                      double                         t,
 509                      gmx_wallcycle_t                wcycle,
 510                      ForceProviders                *forceProviders,
 511                      matrix                         box,
 512                      gmx::ArrayRef<const gmx::RVec> x,
 513                      const t_mdatoms               *mdatoms,
 514                      real                          *lambda,
 515                      int                            forceFlags,
 516                      gmx::ForceWithVirial          *forceWithVirial,
 517                      gmx_enerdata_t                *enerd,
 518                      gmx_edsam                     *ed,
 519                      gmx_bool                       bNS)
 520 {
 521     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 522
 523     /* NOTE: Currently all ForceProviders only provide forces.
 524      *       When they also provide energies, remove this conditional.
 525      */
 526     if (computeForces)
 527     {
 528         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 529         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 530
 531         /* Collect forces from modules */
 532         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 533     }
 534
 535     if (inputrec->bPull && pull_have_potential(inputrec->pull_work))
 536     {
 537         pull_potential_wrapper(cr, inputrec, box, x,
 538                                forceWithVirial,
 539                                mdatoms, enerd, lambda, t,
 540                                wcycle);
 541
 542         if (awh)
 543         {
 544             enerd->term[F_COM_PULL] +=
 545                 awh->applyBiasForcesAndUpdateBias(inputrec->ePBC, *mdatoms, box,
 546                                                   forceWithVirial,
 547                                                   t, step, wcycle, fplog);
 548         }
 549     }
 550
 551     rvec *f = as_rvec_array(forceWithVirial->force_.data());
 552
 553     /* Add the forces from enforced rotation potentials (if any) */
 554     if (inputrec->bRot)
 555     {
 556         wallcycle_start(wcycle, ewcROTadd);
 557         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 558         wallcycle_stop(wcycle, ewcROTadd);
 559     }
 560
 561     if (ed)
 562     {
 563         /* Note that since init_edsam() is called after the initialization
 564          * of forcerec, edsam doesn't request the noVirSum force buffer.
 565          * Thus if no other algorithm (e.g. PME) requires it, the forces
 566          * here will contribute to the virial.
 567          */
 568         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, bNS);
 569     }
 570
 571     /* Add forces from interactive molecular dynamics (IMD), if bIMD == TRUE. */
 572     if (inputrec->bIMD && computeForces)
 573     {
 574         IMD_apply_forces(inputrec->bIMD, inputrec->imd, cr, f, wcycle);
 575     }
 576 }
 577
 578 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 579  *
 580  * \param[in]  pmedata       The PME structure
 581  * \param[in]  box           The box matrix
 582  * \param[in]  x             Coordinate array
 583  * \param[in]  flags         Force flags
 584  * \param[in]  pmeFlags      PME flags
 585  * \param[in]  wcycle        The wallcycle structure
 586  */
 587 static inline void launchPmeGpuSpread(gmx_pme_t      *pmedata,
 588                                       matrix          box,
 589                                       rvec            x[],
 590                                       int             flags,
 591                                       int             pmeFlags,
 592                                       gmx_wallcycle_t wcycle)
 593 {
 594     pme_gpu_prepare_computation(pmedata, (flags & GMX_FORCE_DYNAMICBOX) != 0, box, wcycle, pmeFlags);
 595     pme_gpu_launch_spread(pmedata, x, wcycle);
 596 }
 597
 598 /*! \brief Launch the FFT and gather stages of PME GPU
 599  *
 600  * This function only implements setting the output forces (no accumulation).
 601  *
 602  * \param[in]  pmedata        The PME structure
 603  * \param[in]  wcycle         The wallcycle structure
 604  */
 605 static void launchPmeGpuFftAndGather(gmx_pme_t        *pmedata,
 606                                      gmx_wallcycle_t   wcycle)
 607 {
 608     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 609     pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
 610 }
 611
 612 /*! \brief
 613  *  Polling wait for either of the PME or nonbonded GPU tasks.
 614  *
 615  * Instead of a static order in waiting for GPU tasks, this function
 616  * polls checking which of the two tasks completes first, and does the
 617  * associated force buffer reduction overlapped with the other task.
 618  * By doing that, unlike static scheduling order, it can always overlap
 619  * one of the reductions, regardless of the GPU task completion order.
 620  *
 621  * \param[in]     nbv              Nonbonded verlet structure
 622  * \param[in,out] pmedata          PME module data
 623  * \param[in,out] force            Force array to reduce task outputs into.
 624  * \param[in,out] forceWithVirial  Force and virial buffers
 625  * \param[in,out] fshift           Shift force output vector results are reduced into
 626  * \param[in,out] enerd            Energy data structure results are reduced into
 627  * \param[in]     flags            Force flags
 628  * \param[in]     pmeFlags         PME flags
 629  * \param[in]     haveOtherWork    Tells whether there is other work than non-bonded in the stream(s)
 630  * \param[in]     wcycle           The wallcycle structure
 631  */
 632 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t                  *nbv,
 633                                         gmx_pme_t                           *pmedata,
 634                                         gmx::ArrayRefWithPadding<gmx::RVec> *force,
 635                                         gmx::ForceWithVirial                *forceWithVirial,
 636                                         rvec                                 fshift[],
 637                                         gmx_enerdata_t                      *enerd,
 638                                         int                                  flags,
 639                                         int                                  pmeFlags,
 640                                         bool                                 haveOtherWork,
 641                                         gmx_wallcycle_t                      wcycle)
 642 {
 643     bool isPmeGpuDone = false;
 644     bool isNbGpuDone  = false;
 645
 646
 647     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 648
 649     while (!isPmeGpuDone || !isNbGpuDone)
 650     {
 651         if (!isPmeGpuDone)
 652         {
 653             GpuTaskCompletion completionType = (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 654             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, forceWithVirial, enerd, completionType);
 655         }
 656
 657         if (!isNbGpuDone)
 658         {
 659             GpuTaskCompletion completionType = (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 660             wallcycle_start_nocount(wcycle, ewcWAIT_GPU_NB_L);
 661             isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
 662                                                      flags,
 663                                                      Nbnxm::AtomLocality::Local,
 664                                                      haveOtherWork,
 665                                                      enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
 666                                                      fshift, completionType);
 667             wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 668             // To get the call count right, when the task finished we
 669             // issue a start/stop.
 670             // TODO: move the ewcWAIT_GPU_NB_L cycle counting into nbnxn_gpu_try_finish_task()
 671             // and ewcNB_XF_BUF_OPS counting into nbnxn_atomdata_add_nbat_f_to_f().
 672             if (isNbGpuDone)
 673             {
 674                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
 675                 wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 676
 677                 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
 678                                               as_rvec_array(force->unpaddedArrayRef().data()), wcycle);
 679             }
 680         }
 681     }
 682 }
 683
 684 /*! \brief Hack structure with force ouput buffers for do_force */
 685 struct ForceOutputs
 686 {
 687     //! Constructor
 688     ForceOutputs(rvec *f, gmx::ForceWithVirial const forceWithVirial) :
 689         f(f),
 690         forceWithVirial(forceWithVirial) {}
 691
 692     //! Force output buffer used by legacy modules
 693     rvec                 *const f;
 694     //! Force with direct virial contribution (if there are any)
 695     gmx::ForceWithVirial        forceWithVirial;
 696 };
 697
 698 /*! \brief Set up the different force buffers; also does clearing.
 699  *
 700  * \param[in] fr        force record pointer
 701  * \param[in] inputrec  input record
 702  * \param[in] force     force array
 703  * \param[in] bDoForces True if force are computed this step
 704  * \param[in] doVirial  True if virial is computed this step
 705  * \param[out] wcycle   wallcycle recording structure
 706  *
 707  * \returns             Cleared force output structure
 708  */
 709 static ForceOutputs
 710 setupForceOutputs(const t_forcerec                    *fr,
 711                   const t_inputrec                    &inputrec,
 712                   gmx::ArrayRefWithPadding<gmx::RVec>  force,
 713                   const bool                           bDoForces,
 714                   const bool                           doVirial,
 715                   gmx_wallcycle_t                      wcycle)
 716 {
 717     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 718
 719     /* Temporary solution until all routines take PaddedRVecVector */
 720     rvec *const f = as_rvec_array(force.unpaddedArrayRef().data());
 721     if (bDoForces)
 722     {
 723         /* Clear the short- and long-range forces */
 724         clear_rvecs_omp(fr->natoms_force_constr, f);
 725     }
 726
 727     /* If we need to compute the virial, we might need a separate
 728      * force buffer for algorithms for which the virial is calculated
 729      * directly, such as PME. Otherwise, forceWithVirial uses the
 730      * the same force (f in legacy calls) buffer as other algorithms.
 731      */
 732     const bool useSeparateForceWithVirialBuffer = (bDoForces && (doVirial && fr->haveDirectVirialContributions));
 733
 734
 735     /* forceWithVirial uses the local atom range only */
 736     gmx::ForceWithVirial forceWithVirial (useSeparateForceWithVirialBuffer ?
 737                                           *fr->forceBufferForDirectVirialContributions : force.unpaddedArrayRef(),
 738                                           doVirial);
 739
 740     if (useSeparateForceWithVirialBuffer)
 741     {
 742         /* TODO: update comment
 743          * We only compute forces on local atoms. Note that vsites can
 744          * spread to non-local atoms, but that part of the buffer is
 745          * cleared separately in the vsite spreading code.
 746          */
 747         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 748     }
 749
 750     if (inputrec.bPull && pull_have_constraint(inputrec.pull_work))
 751     {
 752         clear_pull_forces(inputrec.pull_work);
 753     }
 754
 755     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 756
 757     return ForceOutputs(f, forceWithVirial);
 758 }
 759
 760
 761 static void do_force_cutsVERLET(FILE *fplog,
 762                                 const t_commrec *cr,
 763                                 const gmx_multisim_t *ms,
 764                                 const t_inputrec *inputrec,
 765                                 gmx::Awh *awh,
 766                                 gmx_enfrot *enforcedRotation,
 767                                 int64_t step,
 768                                 t_nrnb *nrnb,
 769                                 gmx_wallcycle_t wcycle,
 770                                 const gmx_localtop_t *top,
 771                                 const gmx_groups_t * /* groups */,
 772                                 matrix box, gmx::ArrayRefWithPadding<gmx::RVec> x,
 773                                 history_t *hist,
 774                                 gmx::ArrayRefWithPadding<gmx::RVec> force,
 775                                 tensor vir_force,
 776                                 const t_mdatoms *mdatoms,
 777                                 gmx_enerdata_t *enerd, t_fcdata *fcd,
 778                                 real *lambda,
 779                                 t_graph *graph,
 780                                 t_forcerec *fr,
 781                                 gmx::PpForceWorkload *ppForceWorkload,
 782                                 interaction_const_t *ic,
 783                                 const gmx_vsite_t *vsite,
 784                                 rvec mu_tot,
 785                                 double t,
 786                                 gmx_edsam *ed,
 787                                 const int flags,
 788                                 const DDBalanceRegionHandler &ddBalanceRegionHandler)
 789 {
 790     int                 cg1, i, j;
 791     double              mu[2*DIM];
 792     gmx_bool            bStateChanged, bNS, bFillGrid, bCalcCGCM;
 793     gmx_bool            bDoForces, bUseGPU, bUseOrEmulGPU;
 794     rvec                vzero, box_diag;
 795     float               cycles_pme, cycles_wait_gpu;
 796     nonbonded_verlet_t *nbv = fr->nbv.get();
 797
 798     bStateChanged = ((flags & GMX_FORCE_STATECHANGED) != 0);
 799     bNS           = ((flags & GMX_FORCE_NS) != 0);
 800     bFillGrid     = (bNS && bStateChanged);
 801     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 802     bDoForces     = ((flags & GMX_FORCE_FORCES) != 0);
 803     bUseGPU       = fr->nbv->useGpu();
 804     bUseOrEmulGPU = bUseGPU || fr->nbv->emulateGpu();
 805
 806     const auto pmeRunMode = fr->pmedata ? pme_run_mode(fr->pmedata) : PmeRunMode::CPU;
 807     // TODO slim this conditional down - inputrec and duty checks should mean the same in proper code!
 808     const bool useGpuPme  = EEL_PME(fr->ic->eeltype) && thisRankHasDuty(cr, DUTY_PME) &&
 809         ((pmeRunMode == PmeRunMode::GPU) || (pmeRunMode == PmeRunMode::Mixed));
 810     const int  pmeFlags = GMX_PME_SPREAD | GMX_PME_SOLVE |
 811         ((flags & GMX_FORCE_VIRIAL) ? GMX_PME_CALC_ENER_VIR : 0) |
 812         ((flags & GMX_FORCE_ENERGY) ? GMX_PME_CALC_ENER_VIR : 0) |
 813         ((flags & GMX_FORCE_FORCES) ? GMX_PME_CALC_F : 0);
 814
 815     /* At a search step we need to start the first balancing region
 816      * somewhere early inside the step after communication during domain
 817      * decomposition (and not during the previous step as usual).
 818      */
 819     if (bNS)
 820     {
 821         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 822     }
 823
 824     cycles_wait_gpu = 0;
 825
 826     const int start  = 0;
 827     const int homenr = mdatoms->homenr;
 828
 829     clear_mat(vir_force);
 830
 831     if (DOMAINDECOMP(cr))
 832     {
 833         cg1 = cr->dd->globalAtomGroupIndices.size();
 834     }
 835     else
 836     {
 837         cg1 = top->cgs.nr;
 838     }
 839     if (fr->n_tpi > 0)
 840     {
 841         cg1--;
 842     }
 843
 844     if (bStateChanged)
 845     {
 846         update_forcerec(fr, box);
 847
 848         if (inputrecNeedMutot(inputrec))
 849         {
 850             /* Calculate total (local) dipole moment in a temporary common array.
 851              * This makes it possible to sum them over nodes faster.
 852              */
 853             calc_mu(start, homenr,
 854                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 855                     mu, mu+DIM);
 856         }
 857     }
 858
 859     if (fr->ePBC != epbcNONE)
 860     {
 861         /* Compute shift vectors every step,
 862          * because of pressure coupling or box deformation!
 863          */
 864         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 865         {
 866             calc_shifts(box, fr->shift_vec);
 867         }
 868
 869         if (bCalcCGCM)
 870         {
 871             put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr));
 872             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 873         }
 874         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 875         {
 876             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
 877         }
 878     }
 879
 880     nbnxn_atomdata_copy_shiftvec((flags & GMX_FORCE_DYNAMICBOX) != 0,
 881                                  fr->shift_vec, nbv->nbat.get());
 882
 883 #if GMX_MPI
 884     if (!thisRankHasDuty(cr, DUTY_PME))
 885     {
 886         /* Send particle coordinates to the pme nodes.
 887          * Since this is only implemented for domain decomposition
 888          * and domain decomposition does not use the graph,
 889          * we do not need to worry about shifting.
 890          */
 891         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
 892                                  lambda[efptCOUL], lambda[efptVDW],
 893                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
 894                                  step, wcycle);
 895     }
 896 #endif /* GMX_MPI */
 897
 898     if (useGpuPme)
 899     {
 900         launchPmeGpuSpread(fr->pmedata, box, as_rvec_array(x.unpaddedArrayRef().data()), flags, pmeFlags, wcycle);
 901     }
 902
 903     /* do gridding for pair search */
 904     if (bNS)
 905     {
 906         if (graph && bStateChanged)
 907         {
 908             /* Calculate intramolecular shift vectors to make molecules whole */
 909             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
 910         }
 911
 912         clear_rvec(vzero);
 913         box_diag[XX] = box[XX][XX];
 914         box_diag[YY] = box[YY][YY];
 915         box_diag[ZZ] = box[ZZ][ZZ];
 916
 917         wallcycle_start(wcycle, ewcNS);
 918         if (!DOMAINDECOMP(cr))
 919         {
 920             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 921             nbnxn_put_on_grid(nbv, box,
 922                               0, vzero, box_diag,
 923                               nullptr, 0, mdatoms->homenr, -1,
 924                               fr->cginfo, x.unpaddedArrayRef(),
 925                               0, nullptr);
 926             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 927         }
 928         else
 929         {
 930             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 931             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
 932                                        fr->cginfo, x.unpaddedArrayRef());
 933             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 934         }
 935
 936         nbnxn_atomdata_set(nbv->nbat.get(), nbv->nbs.get(), mdatoms, fr->cginfo);
 937
 938         wallcycle_stop(wcycle, ewcNS);
 939     }
 940
 941     /* initialize the GPU atom data and copy shift vector */
 942     if (bUseGPU)
 943     {
 944         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 945         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 946
 947         if (bNS)
 948         {
 949             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
 950         }
 951
 952         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
 953
 954         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 955
 956         if (bNS && fr->gpuBonded)
 957         {
 958             /* Now we put all atoms on the grid, we can assign bonded
 959              * interactions to the GPU, where the grid order is
 960              * needed. Also the xq, f and fshift device buffers have
 961              * been reallocated if needed, so the bonded code can
 962              * learn about them. */
 963             // TODO the xq, f, and fshift buffers are now shared
 964             // resources, so they should be maintained by a
 965             // higher-level object than the nb module.
 966             fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbnxn_get_gridindices(fr->nbv->nbs.get()),
 967                                                                   top->idef,
 968                                                                   Nbnxm::gpu_get_xq(nbv->gpu_nbv),
 969                                                                   Nbnxm::gpu_get_f(nbv->gpu_nbv),
 970                                                                   Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
 971             ppForceWorkload->haveGpuBondedWork = fr->gpuBonded->haveInteractions();
 972         }
 973
 974         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
 975     }
 976
 977     /* do local pair search */
 978     if (bNS)
 979     {
 980         wallcycle_start_nocount(wcycle, ewcNS);
 981         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
 982         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
 983         nbv->constructPairlist(Nbnxm::InteractionLocality::Local,
 984                                &top->excls, step, nrnb);
 985         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
 986         wallcycle_stop(wcycle, ewcNS);
 987     }
 988     else
 989     {
 990         nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs.get(), Nbnxm::AtomLocality::Local,
 991                                         FALSE, as_rvec_array(x.unpaddedArrayRef().data()),
 992                                         nbv->nbat.get(), wcycle);
 993     }
 994
 995     if (bUseGPU)
 996     {
 997         ddBalanceRegionHandler.openBeforeForceComputationGpu();
 998
 999         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1000
1001         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1002         Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1003                                   Nbnxm::AtomLocality::Local,
1004                                   ppForceWorkload->haveGpuBondedWork);
1005         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1006
1007         // bonded work not split into separate local and non-local, so with DD
1008         // we can only launch the kernel after non-local coordinates have been received.
1009         if (ppForceWorkload->haveGpuBondedWork && !havePPDomainDecomposition(cr))
1010         {
1011             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1012             fr->gpuBonded->launchKernels(fr, flags, box);
1013             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1014         }
1015
1016         /* launch local nonbonded work on GPU */
1017         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1018         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFNo,
1019                      step, nrnb, wcycle);
1020         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1021         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1022     }
1023
1024     if (useGpuPme)
1025     {
1026         // In PME GPU and mixed mode we launch FFT / gather after the
1027         // X copy/transform to allow overlap as well as after the GPU NB
1028         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1029         // the nonbonded kernel.
1030         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1031     }
1032
1033     /* Communicate coordinates and sum dipole if necessary +
1034        do non-local pair search */
1035     if (havePPDomainDecomposition(cr))
1036     {
1037         if (bNS)
1038         {
1039             wallcycle_start_nocount(wcycle, ewcNS);
1040             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1041             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1042             nbv->constructPairlist(Nbnxm::InteractionLocality::NonLocal,
1043                                    &top->excls, step, nrnb);
1044             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1045             wallcycle_stop(wcycle, ewcNS);
1046         }
1047         else
1048         {
1049             dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1050
1051             nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs.get(), Nbnxm::AtomLocality::NonLocal,
1052                                             FALSE, as_rvec_array(x.unpaddedArrayRef().data()),
1053                                             nbv->nbat.get(), wcycle);
1054         }
1055
1056         if (bUseGPU)
1057         {
1058             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1059
1060             /* launch non-local nonbonded tasks on GPU */
1061             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1062             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1063                                       Nbnxm::AtomLocality::NonLocal,
1064                                       ppForceWorkload->haveGpuBondedWork);
1065             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1066
1067             if (ppForceWorkload->haveGpuBondedWork)
1068             {
1069                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1070                 fr->gpuBonded->launchKernels(fr, flags, box);
1071                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1072             }
1073
1074             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1075             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1076                          step, nrnb, wcycle);
1077             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1078
1079             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1080         }
1081     }
1082
1083     if (bUseGPU)
1084     {
1085         /* launch D2H copy-back F */
1086         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1087         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1088         if (havePPDomainDecomposition(cr))
1089         {
1090             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1091                                       flags, Nbnxm::AtomLocality::NonLocal, ppForceWorkload->haveGpuBondedWork);
1092         }
1093         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1094                                   flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork);
1095         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1096
1097         if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1098         {
1099             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1100             fr->gpuBonded->launchEnergyTransfer();
1101             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1102         }
1103         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1104     }
1105
1106     if (bStateChanged && inputrecNeedMutot(inputrec))
1107     {
1108         if (PAR(cr))
1109         {
1110             gmx_sumd(2*DIM, mu, cr);
1111
1112             ddBalanceRegionHandler.reopenRegionCpu();
1113         }
1114
1115         for (i = 0; i < 2; i++)
1116         {
1117             for (j = 0; j < DIM; j++)
1118             {
1119                 fr->mu_tot[i][j] = mu[i*DIM + j];
1120             }
1121         }
1122     }
1123     if (fr->efep == efepNO)
1124     {
1125         copy_rvec(fr->mu_tot[0], mu_tot);
1126     }
1127     else
1128     {
1129         for (j = 0; j < DIM; j++)
1130         {
1131             mu_tot[j] =
1132                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1133                 lambda[efptCOUL]*fr->mu_tot[1][j];
1134         }
1135     }
1136
1137     /* Reset energies */
1138     reset_enerdata(enerd);
1139     clear_rvecs(SHIFTS, fr->fshift);
1140
1141     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1142     {
1143         wallcycle_start(wcycle, ewcPPDURINGPME);
1144         dd_force_flop_start(cr->dd, nrnb);
1145     }
1146
1147     if (inputrec->bRot)
1148     {
1149         wallcycle_start(wcycle, ewcROT);
1150         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1151         wallcycle_stop(wcycle, ewcROT);
1152     }
1153
1154     /* Start the force cycle counter.
1155      * Note that a different counter is used for dynamic load balancing.
1156      */
1157     wallcycle_start(wcycle, ewcFORCE);
1158
1159     // set up and clear force outputs
1160     struct ForceOutputs forceOut = setupForceOutputs(fr, *inputrec, force, bDoForces, ((flags & GMX_FORCE_VIRIAL) != 0), wcycle);
1161
1162     /* We calculate the non-bonded forces, when done on the CPU, here.
1163      * We do this before calling do_force_lowlevel, because in that
1164      * function, the listed forces are calculated before PME, which
1165      * does communication.  With this order, non-bonded and listed
1166      * force calculation imbalance can be balanced out by the domain
1167      * decomposition load balancing.
1168      */
1169
1170     if (!bUseOrEmulGPU)
1171     {
1172         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFYes,
1173                      step, nrnb, wcycle);
1174     }
1175
1176     if (fr->efep != efepNO)
1177     {
1178         /* Calculate the local and non-local free energy interactions here.
1179          * Happens here on the CPU both with and without GPU.
1180          */
1181         wallcycle_sub_start(wcycle, ewcsNONBONDED);
1182         nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::Local,
1183                                       fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1184                                       inputrec->fepvals, lambda,
1185                                       enerd, flags, nrnb);
1186
1187         if (havePPDomainDecomposition(cr))
1188         {
1189             nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::NonLocal,
1190                                           fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1191                                           inputrec->fepvals, lambda,
1192                                           enerd, flags, nrnb);
1193         }
1194         wallcycle_sub_stop(wcycle, ewcsNONBONDED);
1195     }
1196
1197     if (!bUseOrEmulGPU)
1198     {
1199         if (havePPDomainDecomposition(cr))
1200         {
1201             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1202                          step, nrnb, wcycle);
1203         }
1204
1205         /* Add all the non-bonded force to the normal force array.
1206          * This can be split into a local and a non-local part when overlapping
1207          * communication with calculation with domain decomposition.
1208          */
1209         wallcycle_stop(wcycle, ewcFORCE);
1210
1211         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::All, forceOut.f, wcycle);
1212
1213         wallcycle_start_nocount(wcycle, ewcFORCE);
1214
1215         /* If there are multiple fshift output buffers we need to reduce them */
1216         if (flags & GMX_FORCE_VIRIAL)
1217         {
1218             /* This is not in a subcounter because it takes a
1219                negligible and constant-sized amount of time */
1220             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat.get(),
1221                                                      fr->fshift);
1222         }
1223     }
1224
1225     /* update QMMMrec, if necessary */
1226     if (fr->bQMMM)
1227     {
1228         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1229     }
1230
1231     /* Compute the bonded and non-bonded energies and optionally forces */
1232     do_force_lowlevel(fr, inputrec, &(top->idef),
1233                       cr, ms, nrnb, wcycle, mdatoms,
1234                       as_rvec_array(x.unpaddedArrayRef().data()), hist, forceOut.f, &forceOut.forceWithVirial, enerd, fcd,
1235                       box, inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1236                       flags,
1237                       &cycles_pme, ddBalanceRegionHandler);
1238
1239     wallcycle_stop(wcycle, ewcFORCE);
1240
1241     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1242                          step, t, wcycle,
1243                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda,
1244                          flags, &forceOut.forceWithVirial, enerd,
1245                          ed, bNS);
1246
1247     if (bUseOrEmulGPU)
1248     {
1249         /* wait for non-local forces (or calculate in emulation mode) */
1250         if (havePPDomainDecomposition(cr))
1251         {
1252             if (bUseGPU)
1253             {
1254                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1255                 Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1256                                             flags, Nbnxm::AtomLocality::NonLocal,
1257                                             ppForceWorkload->haveGpuBondedWork,
1258                                             enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1259                                             fr->fshift);
1260                 cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1261             }
1262             else
1263             {
1264                 wallcycle_start_nocount(wcycle, ewcFORCE);
1265                 do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFYes,
1266                              step, nrnb, wcycle);
1267                 wallcycle_stop(wcycle, ewcFORCE);
1268             }
1269
1270             nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::NonLocal,
1271                                           forceOut.f, wcycle);
1272         }
1273     }
1274
1275     if (havePPDomainDecomposition(cr))
1276     {
1277         /* We are done with the CPU compute.
1278          * We will now communicate the non-local forces.
1279          * If we use a GPU this will overlap with GPU work, so in that case
1280          * we do not close the DD force balancing region here.
1281          */
1282         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1283
1284         if (bDoForces)
1285         {
1286             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1287         }
1288     }
1289
1290     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1291     // an alternating wait/reduction scheme.
1292     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPme && bUseGPU && !DOMAINDECOMP(cr));
1293     if (alternateGpuWait)
1294     {
1295         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &force, &forceOut.forceWithVirial, fr->fshift, enerd,
1296                                     flags, pmeFlags, ppForceWorkload->haveGpuBondedWork, wcycle);
1297     }
1298
1299     if (!alternateGpuWait && useGpuPme)
1300     {
1301         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial, enerd);
1302     }
1303
1304     /* Wait for local GPU NB outputs on the non-alternating wait path */
1305     if (!alternateGpuWait && bUseGPU)
1306     {
1307         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1308          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1309          * but even with a step of 0.1 ms the difference is less than 1%
1310          * of the step time.
1311          */
1312         const float gpuWaitApiOverheadMargin = 2e6f; /* cycles */
1313
1314         wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1315         Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1316                                     flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork,
1317                                     enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1318                                     fr->fshift);
1319         float cycles_tmp = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1320
1321         if (ddBalanceRegionHandler.useBalancingRegion())
1322         {
1323             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1324             if (bDoForces && cycles_tmp <= gpuWaitApiOverheadMargin)
1325             {
1326                 /* We measured few cycles, it could be that the kernel
1327                  * and transfer finished earlier and there was no actual
1328                  * wait time, only API call overhead.
1329                  * Then the actual time could be anywhere between 0 and
1330                  * cycles_wait_est. We will use half of cycles_wait_est.
1331                  */
1332                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1333             }
1334             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1335         }
1336     }
1337
1338     if (fr->nbv->emulateGpu())
1339     {
1340         // NOTE: emulation kernel is not included in the balancing region,
1341         // but emulation mode does not target performance anyway
1342         wallcycle_start_nocount(wcycle, ewcFORCE);
1343         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local,
1344                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1345                      step, nrnb, wcycle);
1346         wallcycle_stop(wcycle, ewcFORCE);
1347     }
1348
1349     if (useGpuPme)
1350     {
1351         pme_gpu_reinit_computation(fr->pmedata, wcycle);
1352     }
1353
1354     if (bUseGPU)
1355     {
1356         /* now clear the GPU outputs while we finish the step on the CPU */
1357         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1358         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1359         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, flags);
1360
1361         if (nbv->pairlistSets().isDynamicPruningStepGpu(step))
1362         {
1363             nbv->dispatchPruneKernelGpu(step);
1364         }
1365         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1366         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1367     }
1368
1369     if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1370     {
1371         wallcycle_start(wcycle, ewcWAIT_GPU_BONDED);
1372         // in principle this should be included in the DD balancing region,
1373         // but generally it is infrequent so we'll omit it for the sake of
1374         // simpler code
1375         fr->gpuBonded->accumulateEnergyTerms(enerd);
1376         wallcycle_stop(wcycle, ewcWAIT_GPU_BONDED);
1377
1378         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1379         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1380         fr->gpuBonded->clearEnergies();
1381         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1382         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1383     }
1384
1385     /* Do the nonbonded GPU (or emulation) force buffer reduction
1386      * on the non-alternating path. */
1387     if (bUseOrEmulGPU && !alternateGpuWait)
1388     {
1389         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
1390                                       forceOut.f, wcycle);
1391     }
1392     if (DOMAINDECOMP(cr))
1393     {
1394         dd_force_flop_stop(cr->dd, nrnb);
1395     }
1396
1397     if (bDoForces)
1398     {
1399         /* If we have NoVirSum forces, but we do not calculate the virial,
1400          * we sum fr->f_novirsum=forceOut.f later.
1401          */
1402         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1403         {
1404             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, fr->fshift, FALSE, nullptr, nrnb,
1405                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1406         }
1407
1408         if (flags & GMX_FORCE_VIRIAL)
1409         {
1410             /* Calculation of the virial must be done after vsites! */
1411             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f,
1412                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1413         }
1414     }
1415
1416     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1417     {
1418         /* In case of node-splitting, the PP nodes receive the long-range
1419          * forces, virial and energy from the PME nodes here.
1420          */
1421         pme_receive_force_ener(cr, &forceOut.forceWithVirial, enerd, wcycle);
1422     }
1423
1424     if (bDoForces)
1425     {
1426         post_process_forces(cr, step, nrnb, wcycle,
1427                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, &forceOut.forceWithVirial,
1428                             vir_force, mdatoms, graph, fr, vsite,
1429                             flags);
1430     }
1431
1432     if (flags & GMX_FORCE_ENERGY)
1433     {
1434         /* Sum the potential energy terms from group contributions */
1435         sum_epot(&(enerd->grpp), enerd->term);
1436
1437         if (!EI_TPI(inputrec->eI))
1438         {
1439             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1440         }
1441     }
1442 }
1443
1444 static void do_force_cutsGROUP(FILE *fplog,
1445                                const t_commrec *cr,
1446                                const gmx_multisim_t *ms,
1447                                const t_inputrec *inputrec,
1448                                gmx::Awh *awh,
1449                                gmx_enfrot *enforcedRotation,
1450                                int64_t step,
1451                                t_nrnb *nrnb,
1452                                gmx_wallcycle_t wcycle,
1453                                gmx_localtop_t *top,
1454                                const gmx_groups_t *groups,
1455                                matrix box, gmx::ArrayRefWithPadding<gmx::RVec> x,
1456                                history_t *hist,
1457                                gmx::ArrayRefWithPadding<gmx::RVec> force,
1458                                tensor vir_force,
1459                                const t_mdatoms *mdatoms,
1460                                gmx_enerdata_t *enerd,
1461                                t_fcdata *fcd,
1462                                real *lambda,
1463                                t_graph *graph,
1464                                t_forcerec *fr,
1465                                const gmx_vsite_t *vsite,
1466                                rvec mu_tot,
1467                                double t,
1468                                gmx_edsam *ed,
1469                                int flags,
1470                                const DDBalanceRegionHandler &ddBalanceRegionHandler)
1471 {
1472     int        cg0, cg1, i, j;
1473     double     mu[2*DIM];
1474     gmx_bool   bStateChanged, bNS, bFillGrid, bCalcCGCM;
1475     gmx_bool   bDoForces;
1476     float      cycles_pme;
1477
1478     const int  start  = 0;
1479     const int  homenr = mdatoms->homenr;
1480
1481     clear_mat(vir_force);
1482
1483     cg0 = 0;
1484     if (DOMAINDECOMP(cr))
1485     {
1486         cg1 = cr->dd->globalAtomGroupIndices.size();
1487     }
1488     else
1489     {
1490         cg1 = top->cgs.nr;
1491     }
1492     if (fr->n_tpi > 0)
1493     {
1494         cg1--;
1495     }
1496
1497     bStateChanged  = ((flags & GMX_FORCE_STATECHANGED) != 0);
1498     bNS            = ((flags & GMX_FORCE_NS) != 0);
1499     /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1500     bFillGrid      = (bNS && bStateChanged);
1501     bCalcCGCM      = (bFillGrid && !DOMAINDECOMP(cr));
1502     bDoForces      = ((flags & GMX_FORCE_FORCES) != 0);
1503
1504     if (bStateChanged)
1505     {
1506         update_forcerec(fr, box);
1507
1508         if (inputrecNeedMutot(inputrec))
1509         {
1510             /* Calculate total (local) dipole moment in a temporary common array.
1511              * This makes it possible to sum them over nodes faster.
1512              */
1513             calc_mu(start, homenr,
1514                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1515                     mu, mu+DIM);
1516         }
1517     }
1518
1519     if (fr->ePBC != epbcNONE)
1520     {
1521         /* Compute shift vectors every step,
1522          * because of pressure coupling or box deformation!
1523          */
1524         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1525         {
1526             calc_shifts(box, fr->shift_vec);
1527         }
1528
1529         if (bCalcCGCM)
1530         {
1531             put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1532                                      &(top->cgs), as_rvec_array(x.unpaddedArrayRef().data()), fr->cg_cm);
1533             inc_nrnb(nrnb, eNR_CGCM, homenr);
1534             inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1535         }
1536         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1537         {
1538             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
1539         }
1540     }
1541     else if (bCalcCGCM)
1542     {
1543         calc_cgcm(fplog, cg0, cg1, &(top->cgs), as_rvec_array(x.unpaddedArrayRef().data()), fr->cg_cm);
1544         inc_nrnb(nrnb, eNR_CGCM, homenr);
1545     }
1546
1547     if (bCalcCGCM && gmx_debug_at)
1548     {
1549         pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1550     }
1551
1552 #if GMX_MPI
1553     if (!thisRankHasDuty(cr, DUTY_PME))
1554     {
1555         /* Send particle coordinates to the pme nodes.
1556          * Since this is only implemented for domain decomposition
1557          * and domain decomposition does not use the graph,
1558          * we do not need to worry about shifting.
1559          */
1560         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
1561                                  lambda[efptCOUL], lambda[efptVDW],
1562                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
1563                                  step, wcycle);
1564     }
1565 #endif /* GMX_MPI */
1566
1567     /* Communicate coordinates and sum dipole if necessary */
1568     if (DOMAINDECOMP(cr))
1569     {
1570         dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1571
1572         /* No GPU support, no move_x overlap, so reopen the balance region here */
1573         ddBalanceRegionHandler.reopenRegionCpu();
1574     }
1575
1576     if (inputrecNeedMutot(inputrec))
1577     {
1578         if (bStateChanged)
1579         {
1580             if (PAR(cr))
1581             {
1582                 gmx_sumd(2*DIM, mu, cr);
1583
1584                 ddBalanceRegionHandler.reopenRegionCpu();
1585             }
1586             for (i = 0; i < 2; i++)
1587             {
1588                 for (j = 0; j < DIM; j++)
1589                 {
1590                     fr->mu_tot[i][j] = mu[i*DIM + j];
1591                 }
1592             }
1593         }
1594         if (fr->efep == efepNO)
1595         {
1596             copy_rvec(fr->mu_tot[0], mu_tot);
1597         }
1598         else
1599         {
1600             for (j = 0; j < DIM; j++)
1601             {
1602                 mu_tot[j] =
1603                     (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1604             }
1605         }
1606     }
1607
1608     /* Reset energies */
1609     reset_enerdata(enerd);
1610     clear_rvecs(SHIFTS, fr->fshift);
1611
1612     if (bNS)
1613     {
1614         wallcycle_start(wcycle, ewcNS);
1615
1616         if (graph && bStateChanged)
1617         {
1618             /* Calculate intramolecular shift vectors to make molecules whole */
1619             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
1620         }
1621
1622         /* Do the actual neighbour searching */
1623         ns(fplog, fr, box,
1624            groups, top, mdatoms,
1625            cr, nrnb, bFillGrid);
1626
1627         wallcycle_stop(wcycle, ewcNS);
1628     }
1629
1630     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1631     {
1632         wallcycle_start(wcycle, ewcPPDURINGPME);
1633         dd_force_flop_start(cr->dd, nrnb);
1634     }
1635
1636     if (inputrec->bRot)
1637     {
1638         wallcycle_start(wcycle, ewcROT);
1639         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1640         wallcycle_stop(wcycle, ewcROT);
1641     }
1642
1643
1644     /* Start the force cycle counter.
1645      * Note that a different counter is used for dynamic load balancing.
1646      */
1647     wallcycle_start(wcycle, ewcFORCE);
1648
1649     // set up and clear force outputs
1650     struct ForceOutputs forceOut = setupForceOutputs(fr, *inputrec, force, bDoForces, ((flags & GMX_FORCE_VIRIAL) != 0), wcycle);
1651
1652     if (inputrec->bPull && pull_have_constraint(inputrec->pull_work))
1653     {
1654         clear_pull_forces(inputrec->pull_work);
1655     }
1656
1657     /* update QMMMrec, if necessary */
1658     if (fr->bQMMM)
1659     {
1660         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1661     }
1662
1663     /* Compute the bonded and non-bonded energies and optionally forces */
1664     do_force_lowlevel(fr, inputrec, &(top->idef),
1665                       cr, ms, nrnb, wcycle, mdatoms,
1666                       as_rvec_array(x.unpaddedArrayRef().data()), hist, forceOut.f, &forceOut.forceWithVirial, enerd, fcd,
1667                       box, inputrec->fepvals, lambda,
1668                       graph, &(top->excls), fr->mu_tot,
1669                       flags,
1670                       &cycles_pme, ddBalanceRegionHandler);
1671
1672     wallcycle_stop(wcycle, ewcFORCE);
1673
1674     if (DOMAINDECOMP(cr))
1675     {
1676         dd_force_flop_stop(cr->dd, nrnb);
1677
1678         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1679     }
1680
1681     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1682                          step, t, wcycle,
1683                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda,
1684                          flags, &forceOut.forceWithVirial, enerd,
1685                          ed, bNS);
1686
1687     if (bDoForces)
1688     {
1689         /* Communicate the forces */
1690         if (DOMAINDECOMP(cr))
1691         {
1692             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1693             /* Do we need to communicate the separate force array
1694              * for terms that do not contribute to the single sum virial?
1695              * Position restraints and electric fields do not introduce
1696              * inter-cg forces, only full electrostatics methods do.
1697              * When we do not calculate the virial, fr->f_novirsum = forceOut.f,
1698              * so we have already communicated these forces.
1699              */
1700             if (EEL_FULL(fr->ic->eeltype) && cr->dd->n_intercg_excl &&
1701                 (flags & GMX_FORCE_VIRIAL))
1702             {
1703                 dd_move_f(cr->dd, forceOut.forceWithVirial.force_, nullptr, wcycle);
1704             }
1705         }
1706
1707         /* If we have NoVirSum forces, but we do not calculate the virial,
1708          * we sum fr->f_novirsum=forceOut.f later.
1709          */
1710         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1711         {
1712             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, fr->fshift, FALSE, nullptr, nrnb,
1713                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1714         }
1715
1716         if (flags & GMX_FORCE_VIRIAL)
1717         {
1718             /* Calculation of the virial must be done after vsites! */
1719             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f,
1720                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1721         }
1722     }
1723
1724     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1725     {
1726         /* In case of node-splitting, the PP nodes receive the long-range
1727          * forces, virial and energy from the PME nodes here.
1728          */
1729         pme_receive_force_ener(cr, &forceOut.forceWithVirial, enerd, wcycle);
1730     }
1731
1732     if (bDoForces)
1733     {
1734         post_process_forces(cr, step, nrnb, wcycle,
1735                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, &forceOut.forceWithVirial,
1736                             vir_force, mdatoms, graph, fr, vsite,
1737                             flags);
1738     }
1739
1740     if (flags & GMX_FORCE_ENERGY)
1741     {
1742         /* Sum the potential energy terms from group contributions */
1743         sum_epot(&(enerd->grpp), enerd->term);
1744
1745         if (!EI_TPI(inputrec->eI))
1746         {
1747             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1748         }
1749     }
1750
1751 }
1752
1753 void do_force(FILE                                     *fplog,
1754               const t_commrec                          *cr,
1755               const gmx_multisim_t                     *ms,
1756               const t_inputrec                         *inputrec,
1757               gmx::Awh                                 *awh,
1758               gmx_enfrot                               *enforcedRotation,
1759               int64_t                                   step,
1760               t_nrnb                                   *nrnb,
1761               gmx_wallcycle_t                           wcycle,
1762               gmx_localtop_t                           *top,
1763               const gmx_groups_t                       *groups,
1764               matrix                                    box,
1765               gmx::ArrayRefWithPadding<gmx::RVec>       x,     //NOLINT(performance-unnecessary-value-param)
1766               history_t                                *hist,
1767               gmx::ArrayRefWithPadding<gmx::RVec>       force, //NOLINT(performance-unnecessary-value-param)
1768               tensor                                    vir_force,
1769               const t_mdatoms                          *mdatoms,
1770               gmx_enerdata_t                           *enerd,
1771               t_fcdata                                 *fcd,
1772               gmx::ArrayRef<real>                       lambda,
1773               t_graph                                  *graph,
1774               t_forcerec                               *fr,
1775               gmx::PpForceWorkload                     *ppForceWorkload,
1776               const gmx_vsite_t                        *vsite,
1777               rvec                                      mu_tot,
1778               double                                    t,
1779               gmx_edsam                                *ed,
1780               int                                       flags,
1781               const DDBalanceRegionHandler             &ddBalanceRegionHandler)
1782 {
1783     /* modify force flag if not doing nonbonded */
1784     if (!fr->bNonbonded)
1785     {
1786         flags &= ~GMX_FORCE_NONBONDED;
1787     }
1788
1789     switch (inputrec->cutoff_scheme)
1790     {
1791         case ecutsVERLET:
1792             do_force_cutsVERLET(fplog, cr, ms, inputrec,
1793                                 awh, enforcedRotation, step, nrnb, wcycle,
1794                                 top,
1795                                 groups,
1796                                 box, x, hist,
1797                                 force, vir_force,
1798                                 mdatoms,
1799                                 enerd, fcd,
1800                                 lambda.data(), graph,
1801                                 fr,
1802                                 ppForceWorkload,
1803                                 fr->ic,
1804                                 vsite, mu_tot,
1805                                 t, ed,
1806                                 flags,
1807                                 ddBalanceRegionHandler);
1808             break;
1809         case ecutsGROUP:
1810             do_force_cutsGROUP(fplog, cr, ms, inputrec,
1811                                awh, enforcedRotation, step, nrnb, wcycle,
1812                                top,
1813                                groups,
1814                                box, x, hist,
1815                                force, vir_force,
1816                                mdatoms,
1817                                enerd, fcd,
1818                                lambda.data(), graph,
1819                                fr, vsite, mu_tot,
1820                                t, ed,
1821                                flags,
1822                                ddBalanceRegionHandler);
1823             break;
1824         default:
1825             gmx_incons("Invalid cut-off scheme passed!");
1826     }
1827
1828     /* In case we don't have constraints and are using GPUs, the next balancing
1829      * region starts here.
1830      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1831      * virial calculation and COM pulling, is not thus not included in
1832      * the balance timing, which is ok as most tasks do communication.
1833      */
1834     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1835 }
1836
1837
1838 void do_constrain_first(FILE *fplog, gmx::Constraints *constr,
1839                         const t_inputrec *ir, const t_mdatoms *md,
1840                         t_state *state)
1841 {
1842     int             i, m, start, end;
1843     int64_t         step;
1844     real            dt = ir->delta_t;
1845     real            dvdl_dum;
1846     rvec           *savex;
1847
1848     /* We need to allocate one element extra, since we might use
1849      * (unaligned) 4-wide SIMD loads to access rvec entries.
1850      */
1851     snew(savex, state->natoms + 1);
1852
1853     start = 0;
1854     end   = md->homenr;
1855
1856     if (debug)
1857     {
1858         fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
1859                 start, md->homenr, end);
1860     }
1861     /* Do a first constrain to reset particles... */
1862     step = ir->init_step;
1863     if (fplog)
1864     {
1865         char buf[STEPSTRSIZE];
1866         fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
1867                 gmx_step_str(step, buf));
1868     }
1869     dvdl_dum = 0;
1870
1871     /* constrain the current position */
1872     constr->apply(TRUE, FALSE,
1873                   step, 0, 1.0,
1874                   state->x.rvec_array(), state->x.rvec_array(), nullptr,
1875                   state->box,
1876                   state->lambda[efptBONDED], &dvdl_dum,
1877                   nullptr, nullptr, gmx::ConstraintVariable::Positions);
1878     if (EI_VV(ir->eI))
1879     {
1880         /* constrain the inital velocity, and save it */
1881         /* also may be useful if we need the ekin from the halfstep for velocity verlet */
1882         constr->apply(TRUE, FALSE,
1883                       step, 0, 1.0,
1884                       state->x.rvec_array(), state->v.rvec_array(), state->v.rvec_array(),
1885                       state->box,
1886                       state->lambda[efptBONDED], &dvdl_dum,
1887                       nullptr, nullptr, gmx::ConstraintVariable::Velocities);
1888     }
1889     /* constrain the inital velocities at t-dt/2 */
1890     if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
1891     {
1892         auto x = makeArrayRef(state->x).subArray(start, end);
1893         auto v = makeArrayRef(state->v).subArray(start, end);
1894         for (i = start; (i < end); i++)
1895         {
1896             for (m = 0; (m < DIM); m++)
1897             {
1898                 /* Reverse the velocity */
1899                 v[i][m] = -v[i][m];
1900                 /* Store the position at t-dt in buf */
1901                 savex[i][m] = x[i][m] + dt*v[i][m];
1902             }
1903         }
1904         /* Shake the positions at t=-dt with the positions at t=0
1905          * as reference coordinates.
1906          */
1907         if (fplog)
1908         {
1909             char buf[STEPSTRSIZE];
1910             fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
1911                     gmx_step_str(step, buf));
1912         }
1913         dvdl_dum = 0;
1914         constr->apply(TRUE, FALSE,
1915                       step, -1, 1.0,
1916                       state->x.rvec_array(), savex, nullptr,
1917                       state->box,
1918                       state->lambda[efptBONDED], &dvdl_dum,
1919                       state->v.rvec_array(), nullptr, gmx::ConstraintVariable::Positions);
1920
1921         for (i = start; i < end; i++)
1922         {
1923             for (m = 0; m < DIM; m++)
1924             {
1925                 /* Re-reverse the velocities */
1926                 v[i][m] = -v[i][m];
1927             }
1928         }
1929     }
1930     sfree(savex);
1931 }
1932
1933 void put_atoms_in_box_omp(int ePBC, const matrix box, gmx::ArrayRef<gmx::RVec> x)
1934 {
1935     int t, nth;
1936     nth = gmx_omp_nthreads_get(emntDefault);
1937
1938 #pragma omp parallel for num_threads(nth) schedule(static)
1939     for (t = 0; t < nth; t++)
1940     {
1941         try
1942         {
1943             size_t natoms = x.size();
1944             size_t offset = (natoms*t    )/nth;
1945             size_t len    = (natoms*(t + 1))/nth - offset;
1946             put_atoms_in_box(ePBC, box, x.subArray(offset, len));
1947         }
1948         GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
1949     }
1950 }
1951
1952 void initialize_lambdas(FILE               *fplog,
1953                         const t_inputrec   &ir,
1954                         bool                isMaster,
1955                         int                *fep_state,
1956                         gmx::ArrayRef<real> lambda,
1957                         double             *lam0)
1958 {
1959     /* TODO: Clean up initialization of fep_state and lambda in
1960        t_state.  This function works, but could probably use a logic
1961        rewrite to keep all the different types of efep straight. */
1962
1963     if ((ir.efep == efepNO) && (!ir.bSimTemp))
1964     {
1965         return;
1966     }
1967
1968     const t_lambda *fep = ir.fepvals;
1969     if (isMaster)
1970     {
1971         *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
1972                                              if checkpoint is set -- a kludge is in for now
1973                                              to prevent this.*/
1974     }
1975
1976     for (int i = 0; i < efptNR; i++)
1977     {
1978         double thisLambda;
1979         /* overwrite lambda state with init_lambda for now for backwards compatibility */
1980         if (fep->init_lambda >= 0) /* if it's -1, it was never initialized */
1981         {
1982             thisLambda = fep->init_lambda;
1983         }
1984         else
1985         {
1986             thisLambda = fep->all_lambda[i][fep->init_fep_state];
1987         }
1988         if (isMaster)
1989         {
1990             lambda[i] = thisLambda;
1991         }
1992         if (lam0 != nullptr)
1993         {
1994             lam0[i] = thisLambda;
1995         }
1996     }
1997     if (ir.bSimTemp)
1998     {
1999         /* need to rescale control temperatures to match current state */
2000         for (int i = 0; i < ir.opts.ngtc; i++)
2001         {
2002             if (ir.opts.ref_t[i] > 0)
2003             {
2004                 ir.opts.ref_t[i] = ir.simtempvals->temperatures[fep->init_fep_state];
2005             }
2006         }
2007     }
2008
2009     /* Send to the log the information on the current lambdas */
2010     if (fplog != nullptr)
2011     {
2012         fprintf(fplog, "Initial vector of lambda components:[ ");
2013         for (const auto &l : lambda)
2014         {
2015             fprintf(fplog, "%10.4f ", l);
2016         }
2017         fprintf(fplog, "]\n");
2018     }
2019 }