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/enerdata_utils.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 // environment variable to enable GPU buffer ops, to alow incremental and optional
 124 // introduction of this functionality.
 125 // TODO eventially tie this in with other existing GPU flags.
 126 static const bool c_enableGpuBufOps = (getenv("GMX_USE_GPU_BUFFER_OPS") != nullptr);
 127
 128
 129
 130 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
 131 {
 132     const int      end = forceToAdd.size();
 133
 134     int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
 135 #pragma omp parallel for num_threads(nt) schedule(static)
 136     for (int i = 0; i < end; i++)
 137     {
 138         rvec_inc(f[i], forceToAdd[i]);
 139     }
 140 }
 141
 142 static void calc_virial(int start, int homenr, const rvec x[], const rvec f[],
 143                         tensor vir_part, const t_graph *graph, const matrix box,
 144                         t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
 145 {
 146     /* The short-range virial from surrounding boxes */
 147     calc_vir(SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
 148     inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
 149
 150     /* Calculate partial virial, for local atoms only, based on short range.
 151      * Total virial is computed in global_stat, called from do_md
 152      */
 153     f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
 154     inc_nrnb(nrnb, eNR_VIRIAL, homenr);
 155
 156     if (debug)
 157     {
 158         pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
 159     }
 160 }
 161
 162 static void pull_potential_wrapper(const t_commrec *cr,
 163                                    const t_inputrec *ir,
 164                                    const matrix box, gmx::ArrayRef<const gmx::RVec> x,
 165                                    gmx::ForceWithVirial *force,
 166                                    const t_mdatoms *mdatoms,
 167                                    gmx_enerdata_t *enerd,
 168                                    pull_t *pull_work,
 169                                    const real *lambda,
 170                                    double t,
 171                                    gmx_wallcycle_t wcycle)
 172 {
 173     t_pbc  pbc;
 174     real   dvdl;
 175
 176     /* Calculate the center of mass forces, this requires communication,
 177      * which is why pull_potential is called close to other communication.
 178      */
 179     wallcycle_start(wcycle, ewcPULLPOT);
 180     set_pbc(&pbc, ir->ePBC, box);
 181     dvdl                     = 0;
 182     enerd->term[F_COM_PULL] +=
 183         pull_potential(pull_work, mdatoms, &pbc,
 184                        cr, t, lambda[efptRESTRAINT], as_rvec_array(x.data()), force, &dvdl);
 185     enerd->dvdl_lin[efptRESTRAINT] += dvdl;
 186     wallcycle_stop(wcycle, ewcPULLPOT);
 187 }
 188
 189 static void pme_receive_force_ener(const t_commrec      *cr,
 190                                    gmx::ForceWithVirial *forceWithVirial,
 191                                    gmx_enerdata_t       *enerd,
 192                                    gmx_wallcycle_t       wcycle)
 193 {
 194     real   e_q, e_lj, dvdl_q, dvdl_lj;
 195     float  cycles_ppdpme, cycles_seppme;
 196
 197     cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
 198     dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
 199
 200     /* In case of node-splitting, the PP nodes receive the long-range
 201      * forces, virial and energy from the PME nodes here.
 202      */
 203     wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
 204     dvdl_q  = 0;
 205     dvdl_lj = 0;
 206     gmx_pme_receive_f(cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
 207                       &cycles_seppme);
 208     enerd->term[F_COUL_RECIP] += e_q;
 209     enerd->term[F_LJ_RECIP]   += e_lj;
 210     enerd->dvdl_lin[efptCOUL] += dvdl_q;
 211     enerd->dvdl_lin[efptVDW]  += dvdl_lj;
 212
 213     if (wcycle)
 214     {
 215         dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
 216     }
 217     wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
 218 }
 219
 220 static void print_large_forces(FILE            *fp,
 221                                const t_mdatoms *md,
 222                                const t_commrec *cr,
 223                                int64_t          step,
 224                                real             forceTolerance,
 225                                const rvec      *x,
 226                                const rvec      *f)
 227 {
 228     real           force2Tolerance = gmx::square(forceTolerance);
 229     gmx::index     numNonFinite    = 0;
 230     for (int i = 0; i < md->homenr; i++)
 231     {
 232         real force2    = norm2(f[i]);
 233         bool nonFinite = !std::isfinite(force2);
 234         if (force2 >= force2Tolerance || nonFinite)
 235         {
 236             fprintf(fp, "step %" PRId64 " atom %6d  x %8.3f %8.3f %8.3f  force %12.5e\n",
 237                     step,
 238                     ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
 239         }
 240         if (nonFinite)
 241         {
 242             numNonFinite++;
 243         }
 244     }
 245     if (numNonFinite > 0)
 246     {
 247         /* Note that with MPI this fatal call on one rank might interrupt
 248          * the printing on other ranks. But we can only avoid that with
 249          * an expensive MPI barrier that we would need at each step.
 250          */
 251         gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
 252     }
 253 }
 254
 255 static void post_process_forces(const t_commrec           *cr,
 256                                 int64_t                    step,
 257                                 t_nrnb                    *nrnb,
 258                                 gmx_wallcycle_t            wcycle,
 259                                 const gmx_localtop_t      *top,
 260                                 const matrix               box,
 261                                 const rvec                 x[],
 262                                 rvec                       f[],
 263                                 gmx::ForceWithVirial      *forceWithVirial,
 264                                 tensor                     vir_force,
 265                                 const t_mdatoms           *mdatoms,
 266                                 const t_graph             *graph,
 267                                 const t_forcerec          *fr,
 268                                 const gmx_vsite_t         *vsite,
 269                                 int                        flags)
 270 {
 271     if (fr->haveDirectVirialContributions)
 272     {
 273         rvec *fDirectVir = as_rvec_array(forceWithVirial->force_.data());
 274
 275         if (vsite)
 276         {
 277             /* Spread the mesh force on virtual sites to the other particles...
 278              * This is parallellized. MPI communication is performed
 279              * if the constructing atoms aren't local.
 280              */
 281             matrix virial = { { 0 } };
 282             spread_vsite_f(vsite, x, fDirectVir, nullptr,
 283                            (flags & GMX_FORCE_VIRIAL) != 0, virial,
 284                            nrnb,
 285                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
 286             forceWithVirial->addVirialContribution(virial);
 287         }
 288
 289         if (flags & GMX_FORCE_VIRIAL)
 290         {
 291             /* Now add the forces, this is local */
 292             sum_forces(f, forceWithVirial->force_);
 293
 294             /* Add the direct virial contributions */
 295             GMX_ASSERT(forceWithVirial->computeVirial_, "forceWithVirial should request virial computation when we request the virial");
 296             m_add(vir_force, forceWithVirial->getVirial(), vir_force);
 297
 298             if (debug)
 299             {
 300                 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
 301             }
 302         }
 303     }
 304
 305     if (fr->print_force >= 0)
 306     {
 307         print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
 308     }
 309 }
 310
 311 static void do_nb_verlet(t_forcerec                       *fr,
 312                          const interaction_const_t        *ic,
 313                          gmx_enerdata_t                   *enerd,
 314                          const int                         flags,
 315                          const Nbnxm::InteractionLocality  ilocality,
 316                          const int                         clearF,
 317                          const int64_t                     step,
 318                          t_nrnb                           *nrnb,
 319                          gmx_wallcycle_t                   wcycle)
 320 {
 321     if (!(flags & GMX_FORCE_NONBONDED))
 322     {
 323         /* skip non-bonded calculation */
 324         return;
 325     }
 326
 327     nonbonded_verlet_t *nbv  = fr->nbv.get();
 328
 329     /* GPU kernel launch overhead is already timed separately */
 330     if (fr->cutoff_scheme != ecutsVERLET)
 331     {
 332         gmx_incons("Invalid cut-off scheme passed!");
 333     }
 334
 335     if (!nbv->useGpu())
 336     {
 337         /* When dynamic pair-list  pruning is requested, we need to prune
 338          * at nstlistPrune steps.
 339          */
 340         if (nbv->isDynamicPruningStepCpu(step))
 341         {
 342             /* Prune the pair-list beyond fr->ic->rlistPrune using
 343              * the current coordinates of the atoms.
 344              */
 345             wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
 346             nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
 347             wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
 348         }
 349     }
 350
 351     nbv->dispatchNonbondedKernel(ilocality, *ic, flags, clearF, *fr, enerd, nrnb, wcycle);
 352 }
 353
 354 static inline void clear_rvecs_omp(int n, rvec v[])
 355 {
 356     int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
 357
 358     /* Note that we would like to avoid this conditional by putting it
 359      * into the omp pragma instead, but then we still take the full
 360      * omp parallel for overhead (at least with gcc5).
 361      */
 362     if (nth == 1)
 363     {
 364         for (int i = 0; i < n; i++)
 365         {
 366             clear_rvec(v[i]);
 367         }
 368     }
 369     else
 370     {
 371 #pragma omp parallel for num_threads(nth) schedule(static)
 372         for (int i = 0; i < n; i++)
 373         {
 374             clear_rvec(v[i]);
 375         }
 376     }
 377 }
 378
 379 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
 380  *
 381  * \param groupOptions  Group options, containing T-coupling options
 382  */
 383 static real averageKineticEnergyEstimate(const t_grpopts &groupOptions)
 384 {
 385     real nrdfCoupled   = 0;
 386     real nrdfUncoupled = 0;
 387     real kineticEnergy = 0;
 388     for (int g = 0; g < groupOptions.ngtc; g++)
 389     {
 390         if (groupOptions.tau_t[g] >= 0)
 391         {
 392             nrdfCoupled   += groupOptions.nrdf[g];
 393             kineticEnergy += groupOptions.nrdf[g]*0.5*groupOptions.ref_t[g]*BOLTZ;
 394         }
 395         else
 396         {
 397             nrdfUncoupled += groupOptions.nrdf[g];
 398         }
 399     }
 400
 401     /* This conditional with > also catches nrdf=0 */
 402     if (nrdfCoupled > nrdfUncoupled)
 403     {
 404         return kineticEnergy*(nrdfCoupled + nrdfUncoupled)/nrdfCoupled;
 405     }
 406     else
 407     {
 408         return 0;
 409     }
 410 }
 411
 412 /*! \brief This routine checks that the potential energy is finite.
 413  *
 414  * Always checks that the potential energy is finite. If step equals
 415  * inputrec.init_step also checks that the magnitude of the potential energy
 416  * is reasonable. Terminates with a fatal error when a check fails.
 417  * Note that passing this check does not guarantee finite forces,
 418  * since those use slightly different arithmetics. But in most cases
 419  * there is just a narrow coordinate range where forces are not finite
 420  * and energies are finite.
 421  *
 422  * \param[in] step      The step number, used for checking and printing
 423  * \param[in] enerd     The energy data; the non-bonded group energies need to be added to enerd.term[F_EPOT] before calling this routine
 424  * \param[in] inputrec  The input record
 425  */
 426 static void checkPotentialEnergyValidity(int64_t               step,
 427                                          const gmx_enerdata_t &enerd,
 428                                          const t_inputrec     &inputrec)
 429 {
 430     /* Threshold valid for comparing absolute potential energy against
 431      * the kinetic energy. Normally one should not consider absolute
 432      * potential energy values, but with a factor of one million
 433      * we should never get false positives.
 434      */
 435     constexpr real c_thresholdFactor = 1e6;
 436
 437     bool           energyIsNotFinite    = !std::isfinite(enerd.term[F_EPOT]);
 438     real           averageKineticEnergy = 0;
 439     /* We only check for large potential energy at the initial step,
 440      * because that is by far the most likely step for this too occur
 441      * and because computing the average kinetic energy is not free.
 442      * Note: nstcalcenergy >> 1 often does not allow to catch large energies
 443      * before they become NaN.
 444      */
 445     if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
 446     {
 447         averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
 448     }
 449
 450     if (energyIsNotFinite || (averageKineticEnergy > 0 &&
 451                               enerd.term[F_EPOT] > c_thresholdFactor*averageKineticEnergy))
 452     {
 453         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.",
 454                   step,
 455                   enerd.term[F_EPOT],
 456                   energyIsNotFinite ? "not finite" : "extremely high",
 457                   enerd.term[F_LJ],
 458                   enerd.term[F_COUL_SR],
 459                   energyIsNotFinite ? "non-finite" : "very high",
 460                   energyIsNotFinite ? " or Nan" : "");
 461     }
 462 }
 463
 464 /*! \brief Return true if there are special forces computed this step.
 465  *
 466  * The conditionals exactly correspond to those in computeSpecialForces().
 467  */
 468 static bool
 469 haveSpecialForces(const t_inputrec              *inputrec,
 470                   ForceProviders                *forceProviders,
 471                   const pull_t                  *pull_work,
 472                   int                            forceFlags,
 473                   const gmx_edsam               *ed)
 474 {
 475     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 476
 477     return
 478         ((computeForces && forceProviders->hasForceProvider()) ||         // forceProviders
 479          (inputrec->bPull && pull_have_potential(pull_work)) ||           // pull
 480          inputrec->bRot ||                                                // enforced rotation
 481          (ed != nullptr) ||                                               // flooding
 482          (inputrec->bIMD && computeForces));                              // IMD
 483 }
 484
 485 /*! \brief Compute forces and/or energies for special algorithms
 486  *
 487  * The intention is to collect all calls to algorithms that compute
 488  * forces on local atoms only and that do not contribute to the local
 489  * virial sum (but add their virial contribution separately).
 490  * Eventually these should likely all become ForceProviders.
 491  * Within this function the intention is to have algorithms that do
 492  * global communication at the end, so global barriers within the MD loop
 493  * are as close together as possible.
 494  *
 495  * \param[in]     fplog            The log file
 496  * \param[in]     cr               The communication record
 497  * \param[in]     inputrec         The input record
 498  * \param[in]     awh              The Awh module (nullptr if none in use).
 499  * \param[in]     enforcedRotation Enforced rotation module.
 500  * \param[in]     imdSession       The IMD session
 501  * \param[in]     pull_work        The pull work structure.
 502  * \param[in]     step             The current MD step
 503  * \param[in]     t                The current time
 504  * \param[in,out] wcycle           Wallcycle accounting struct
 505  * \param[in,out] forceProviders   Pointer to a list of force providers
 506  * \param[in]     box              The unit cell
 507  * \param[in]     x                The coordinates
 508  * \param[in]     mdatoms          Per atom properties
 509  * \param[in]     lambda           Array of free-energy lambda values
 510  * \param[in]     forceFlags       Flags that tell whether we should compute forces/energies/virial
 511  * \param[in,out] forceWithVirial  Force and virial buffers
 512  * \param[in,out] enerd            Energy buffer
 513  * \param[in,out] ed               Essential dynamics pointer
 514  * \param[in]     bNS              Tells if we did neighbor searching this step, used for ED sampling
 515  *
 516  * \todo Remove bNS, which is used incorrectly.
 517  * \todo Convert all other algorithms called here to ForceProviders.
 518  */
 519 static void
 520 computeSpecialForces(FILE                          *fplog,
 521                      const t_commrec               *cr,
 522                      const t_inputrec              *inputrec,
 523                      gmx::Awh                      *awh,
 524                      gmx_enfrot                    *enforcedRotation,
 525                      gmx::ImdSession               *imdSession,
 526                      pull_t                        *pull_work,
 527                      int64_t                        step,
 528                      double                         t,
 529                      gmx_wallcycle_t                wcycle,
 530                      ForceProviders                *forceProviders,
 531                      matrix                         box,
 532                      gmx::ArrayRef<const gmx::RVec> x,
 533                      const t_mdatoms               *mdatoms,
 534                      real                          *lambda,
 535                      int                            forceFlags,
 536                      gmx::ForceWithVirial          *forceWithVirial,
 537                      gmx_enerdata_t                *enerd,
 538                      gmx_edsam                     *ed,
 539                      gmx_bool                       bNS)
 540 {
 541     const bool computeForces = (forceFlags & GMX_FORCE_FORCES) != 0;
 542
 543     /* NOTE: Currently all ForceProviders only provide forces.
 544      *       When they also provide energies, remove this conditional.
 545      */
 546     if (computeForces)
 547     {
 548         gmx::ForceProviderInput  forceProviderInput(x, *mdatoms, t, box, *cr);
 549         gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
 550
 551         /* Collect forces from modules */
 552         forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
 553     }
 554
 555     if (inputrec->bPull && pull_have_potential(pull_work))
 556     {
 557         pull_potential_wrapper(cr, inputrec, box, x,
 558                                forceWithVirial,
 559                                mdatoms, enerd, pull_work, lambda, t,
 560                                wcycle);
 561
 562         if (awh)
 563         {
 564             enerd->term[F_COM_PULL] +=
 565                 awh->applyBiasForcesAndUpdateBias(inputrec->ePBC, *mdatoms, box,
 566                                                   forceWithVirial,
 567                                                   t, step, wcycle, fplog);
 568         }
 569     }
 570
 571     rvec *f = as_rvec_array(forceWithVirial->force_.data());
 572
 573     /* Add the forces from enforced rotation potentials (if any) */
 574     if (inputrec->bRot)
 575     {
 576         wallcycle_start(wcycle, ewcROTadd);
 577         enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
 578         wallcycle_stop(wcycle, ewcROTadd);
 579     }
 580
 581     if (ed)
 582     {
 583         /* Note that since init_edsam() is called after the initialization
 584          * of forcerec, edsam doesn't request the noVirSum force buffer.
 585          * Thus if no other algorithm (e.g. PME) requires it, the forces
 586          * here will contribute to the virial.
 587          */
 588         do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, bNS);
 589     }
 590
 591     /* Add forces from interactive molecular dynamics (IMD), if any */
 592     if (inputrec->bIMD && computeForces)
 593     {
 594         imdSession->applyForces(f);
 595     }
 596 }
 597
 598 /*! \brief Launch the prepare_step and spread stages of PME GPU.
 599  *
 600  * \param[in]  pmedata       The PME structure
 601  * \param[in]  box           The box matrix
 602  * \param[in]  x             Coordinate array
 603  * \param[in]  flags         Force flags
 604  * \param[in]  pmeFlags      PME flags
 605  * \param[in]  wcycle        The wallcycle structure
 606  */
 607 static inline void launchPmeGpuSpread(gmx_pme_t      *pmedata,
 608                                       matrix          box,
 609                                       rvec            x[],
 610                                       int             flags,
 611                                       int             pmeFlags,
 612                                       gmx_wallcycle_t wcycle)
 613 {
 614     pme_gpu_prepare_computation(pmedata, (flags & GMX_FORCE_DYNAMICBOX) != 0, box, wcycle, pmeFlags);
 615     pme_gpu_launch_spread(pmedata, x, wcycle);
 616 }
 617
 618 /*! \brief Launch the FFT and gather stages of PME GPU
 619  *
 620  * This function only implements setting the output forces (no accumulation).
 621  *
 622  * \param[in]  pmedata        The PME structure
 623  * \param[in]  wcycle         The wallcycle structure
 624  */
 625 static void launchPmeGpuFftAndGather(gmx_pme_t        *pmedata,
 626                                      gmx_wallcycle_t   wcycle)
 627 {
 628     pme_gpu_launch_complex_transforms(pmedata, wcycle);
 629     pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
 630 }
 631
 632 /*! \brief
 633  *  Polling wait for either of the PME or nonbonded GPU tasks.
 634  *
 635  * Instead of a static order in waiting for GPU tasks, this function
 636  * polls checking which of the two tasks completes first, and does the
 637  * associated force buffer reduction overlapped with the other task.
 638  * By doing that, unlike static scheduling order, it can always overlap
 639  * one of the reductions, regardless of the GPU task completion order.
 640  *
 641  * \param[in]     nbv              Nonbonded verlet structure
 642  * \param[in,out] pmedata          PME module data
 643  * \param[in,out] force            Force array to reduce task outputs into.
 644  * \param[in,out] forceWithVirial  Force and virial buffers
 645  * \param[in,out] fshift           Shift force output vector results are reduced into
 646  * \param[in,out] enerd            Energy data structure results are reduced into
 647  * \param[in]     flags            Force flags
 648  * \param[in]     pmeFlags         PME flags
 649  * \param[in]     haveOtherWork    Tells whether there is other work than non-bonded in the stream(s)
 650  * \param[in]     wcycle           The wallcycle structure
 651  */
 652 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t                  *nbv,
 653                                         gmx_pme_t                           *pmedata,
 654                                         gmx::ArrayRefWithPadding<gmx::RVec> *force,
 655                                         gmx::ForceWithVirial                *forceWithVirial,
 656                                         rvec                                 fshift[],
 657                                         gmx_enerdata_t                      *enerd,
 658                                         int                                  flags,
 659                                         int                                  pmeFlags,
 660                                         bool                                 haveOtherWork,
 661                                         gmx_wallcycle_t                      wcycle)
 662 {
 663     bool isPmeGpuDone = false;
 664     bool isNbGpuDone  = false;
 665
 666
 667     gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
 668
 669     while (!isPmeGpuDone || !isNbGpuDone)
 670     {
 671         if (!isPmeGpuDone)
 672         {
 673             GpuTaskCompletion completionType = (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 674             isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, forceWithVirial, enerd, completionType);
 675         }
 676
 677         if (!isNbGpuDone)
 678         {
 679             GpuTaskCompletion completionType = (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
 680             wallcycle_start_nocount(wcycle, ewcWAIT_GPU_NB_L);
 681             isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
 682                                                      flags,
 683                                                      Nbnxm::AtomLocality::Local,
 684                                                      haveOtherWork,
 685                                                      enerd->grpp.ener[egLJSR].data(),
 686                                                      enerd->grpp.ener[egCOULSR].data(),
 687                                                      fshift, completionType);
 688             wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 689             // To get the call count right, when the task finished we
 690             // issue a start/stop.
 691             // TODO: move the ewcWAIT_GPU_NB_L cycle counting into nbnxn_gpu_try_finish_task()
 692             // and ewcNB_XF_BUF_OPS counting into nbnxn_atomdata_add_nbat_f_to_f().
 693             if (isNbGpuDone)
 694             {
 695                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
 696                 wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
 697
 698                 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
 699                                               as_rvec_array(force->unpaddedArrayRef().data()), wcycle);
 700             }
 701         }
 702     }
 703 }
 704
 705 /*! \brief Hack structure with force ouput buffers for do_force for the home atoms for this domain */
 706 struct ForceOutputs
 707 {
 708     //! Constructor
 709     ForceOutputs(rvec *f, gmx::ForceWithVirial const forceWithVirial) :
 710         f(f),
 711         forceWithVirial(forceWithVirial) {}
 712
 713     //! Force output buffer used by legacy modules (without SIMD padding)
 714     rvec                 *const f;
 715     //! Force with direct virial contribution (if there are any; without SIMD padding)
 716     gmx::ForceWithVirial        forceWithVirial;
 717 };
 718
 719 /*! \brief Set up the different force buffers; also does clearing.
 720  *
 721  * \param[in] fr        force record pointer
 722  * \param[in] pull_work The pull work object.
 723  * \param[in] inputrec  input record
 724  * \param[in] force     force array
 725  * \param[in] bDoForces True if force are computed this step
 726  * \param[in] doVirial  True if virial is computed this step
 727  * \param[out] wcycle   wallcycle recording structure
 728  *
 729  * \returns             Cleared force output structure
 730  */
 731 static ForceOutputs
 732 setupForceOutputs(const t_forcerec                    *fr,
 733                   pull_t                              *pull_work,
 734                   const t_inputrec                    &inputrec,
 735                   gmx::ArrayRefWithPadding<gmx::RVec>  force,
 736                   const bool                           bDoForces,
 737                   const bool                           doVirial,
 738                   gmx_wallcycle_t                      wcycle)
 739 {
 740     wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
 741
 742     /* Temporary solution until all routines take PaddedRVecVector */
 743     rvec *const f = as_rvec_array(force.unpaddedArrayRef().data());
 744     if (bDoForces)
 745     {
 746         /* Clear the short- and long-range forces */
 747         clear_rvecs_omp(fr->natoms_force_constr, f);
 748     }
 749
 750     /* If we need to compute the virial, we might need a separate
 751      * force buffer for algorithms for which the virial is calculated
 752      * directly, such as PME. Otherwise, forceWithVirial uses the
 753      * the same force (f in legacy calls) buffer as other algorithms.
 754      */
 755     const bool useSeparateForceWithVirialBuffer = (bDoForces && (doVirial && fr->haveDirectVirialContributions));
 756
 757
 758     /* forceWithVirial uses the local atom range only */
 759     gmx::ForceWithVirial forceWithVirial (useSeparateForceWithVirialBuffer ?
 760                                           *fr->forceBufferForDirectVirialContributions : force.unpaddedArrayRef(),
 761                                           doVirial);
 762
 763     if (useSeparateForceWithVirialBuffer)
 764     {
 765         /* TODO: update comment
 766          * We only compute forces on local atoms. Note that vsites can
 767          * spread to non-local atoms, but that part of the buffer is
 768          * cleared separately in the vsite spreading code.
 769          */
 770         clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
 771     }
 772
 773     if (inputrec.bPull && pull_have_constraint(pull_work))
 774     {
 775         clear_pull_forces(pull_work);
 776     }
 777
 778     wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
 779
 780     return ForceOutputs(f, forceWithVirial);
 781 }
 782
 783
 784 /*! \brief Set up flags that indicate what type of work is there to compute.
 785  *
 786  * Currently we only update it at search steps,
 787  * but some properties may change more frequently (e.g. virial/non-virial step),
 788  * so when including those either the frequency of update (per-step) or the scope
 789  * of a flag will change (i.e. a set of flags for nstlist steps).
 790  *
 791  */
 792 static void
 793 setupForceWorkload(gmx::PpForceWorkload *forceWork,
 794                    const t_inputrec     *inputrec,
 795                    const t_forcerec     *fr,
 796                    const pull_t         *pull_work,
 797                    const gmx_edsam      *ed,
 798                    const t_idef         &idef,
 799                    const t_fcdata       *fcd,
 800                    const int             forceFlags
 801                    )
 802 {
 803     forceWork->haveSpecialForces      = haveSpecialForces(inputrec, fr->forceProviders, pull_work, forceFlags, ed);
 804     forceWork->haveCpuBondedWork      = haveCpuBondeds(*fr);
 805     forceWork->haveGpuBondedWork      = ((fr->gpuBonded != nullptr) && fr->gpuBonded->haveInteractions());
 806     forceWork->haveRestraintsWork     = havePositionRestraints(idef, *fcd);
 807     forceWork->haveCpuListedForceWork = haveCpuListedForces(*fr, idef, *fcd);
 808 }
 809
 810 void do_force(FILE                                     *fplog,
 811               const t_commrec                          *cr,
 812               const gmx_multisim_t                     *ms,
 813               const t_inputrec                         *inputrec,
 814               gmx::Awh                                 *awh,
 815               gmx_enfrot                               *enforcedRotation,
 816               gmx::ImdSession                          *imdSession,
 817               pull_t                                   *pull_work,
 818               int64_t                                   step,
 819               t_nrnb                                   *nrnb,
 820               gmx_wallcycle_t                           wcycle,
 821               const gmx_localtop_t                     *top,
 822               matrix                                    box,
 823               gmx::ArrayRefWithPadding<gmx::RVec>       x,     //NOLINT(performance-unnecessary-value-param)
 824               history_t                                *hist,
 825               gmx::ArrayRefWithPadding<gmx::RVec>       force, //NOLINT(performance-unnecessary-value-param)
 826               tensor                                    vir_force,
 827               const t_mdatoms                          *mdatoms,
 828               gmx_enerdata_t                           *enerd,
 829               t_fcdata                                 *fcd,
 830               gmx::ArrayRef<real>                       lambda,
 831               t_graph                                  *graph,
 832               t_forcerec                               *fr,
 833               gmx::PpForceWorkload                     *ppForceWorkload,
 834               const gmx_vsite_t                        *vsite,
 835               rvec                                      mu_tot,
 836               double                                    t,
 837               gmx_edsam                                *ed,
 838               int                                       flags,
 839               const DDBalanceRegionHandler             &ddBalanceRegionHandler)
 840 {
 841     int                  i, j;
 842     double               mu[2*DIM];
 843     gmx_bool             bStateChanged, bNS, bFillGrid, bCalcCGCM;
 844     gmx_bool             bDoForces, bUseGPU, bUseOrEmulGPU;
 845     nonbonded_verlet_t  *nbv = fr->nbv.get();
 846     interaction_const_t *ic  = fr->ic;
 847
 848     /* modify force flag if not doing nonbonded */
 849     if (!fr->bNonbonded)
 850     {
 851         flags &= ~GMX_FORCE_NONBONDED;
 852     }
 853     bStateChanged = ((flags & GMX_FORCE_STATECHANGED) != 0);
 854     bNS           = ((flags & GMX_FORCE_NS) != 0);
 855     bFillGrid     = (bNS && bStateChanged);
 856     bCalcCGCM     = (bFillGrid && !DOMAINDECOMP(cr));
 857     bDoForces     = ((flags & GMX_FORCE_FORCES) != 0);
 858     bUseGPU       = fr->nbv->useGpu();
 859     bUseOrEmulGPU = bUseGPU || fr->nbv->emulateGpu();
 860
 861     const auto pmeRunMode = fr->pmedata ? pme_run_mode(fr->pmedata) : PmeRunMode::CPU;
 862     // TODO slim this conditional down - inputrec and duty checks should mean the same in proper code!
 863     const bool useGpuPme  = EEL_PME(fr->ic->eeltype) && thisRankHasDuty(cr, DUTY_PME) &&
 864         ((pmeRunMode == PmeRunMode::GPU) || (pmeRunMode == PmeRunMode::Mixed));
 865     const int  pmeFlags = GMX_PME_SPREAD | GMX_PME_SOLVE |
 866         ((flags & GMX_FORCE_VIRIAL) ? GMX_PME_CALC_ENER_VIR : 0) |
 867         ((flags & GMX_FORCE_ENERGY) ? GMX_PME_CALC_ENER_VIR : 0) |
 868         ((flags & GMX_FORCE_FORCES) ? GMX_PME_CALC_F : 0);
 869
 870     const bool useGpuXBufOps = (c_enableGpuBufOps && bUseGPU && (GMX_GPU == GMX_GPU_CUDA));
 871
 872     /* At a search step we need to start the first balancing region
 873      * somewhere early inside the step after communication during domain
 874      * decomposition (and not during the previous step as usual).
 875      */
 876     if (bNS)
 877     {
 878         ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
 879     }
 880
 881     const int start  = 0;
 882     const int homenr = mdatoms->homenr;
 883
 884     clear_mat(vir_force);
 885
 886     if (bStateChanged)
 887     {
 888         update_forcerec(fr, box);
 889
 890         if (inputrecNeedMutot(inputrec))
 891         {
 892             /* Calculate total (local) dipole moment in a temporary common array.
 893              * This makes it possible to sum them over nodes faster.
 894              */
 895             calc_mu(start, homenr,
 896                     x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
 897                     mu, mu+DIM);
 898         }
 899     }
 900
 901     if (fr->ePBC != epbcNONE)
 902     {
 903         /* Compute shift vectors every step,
 904          * because of pressure coupling or box deformation!
 905          */
 906         if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
 907         {
 908             calc_shifts(box, fr->shift_vec);
 909         }
 910
 911         if (bCalcCGCM)
 912         {
 913             put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr), gmx_omp_nthreads_get(emntDefault));
 914             inc_nrnb(nrnb, eNR_SHIFTX, homenr);
 915         }
 916         else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
 917         {
 918             unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
 919         }
 920     }
 921
 922     nbnxn_atomdata_copy_shiftvec((flags & GMX_FORCE_DYNAMICBOX) != 0,
 923                                  fr->shift_vec, nbv->nbat.get());
 924
 925 #if GMX_MPI
 926     if (!thisRankHasDuty(cr, DUTY_PME))
 927     {
 928         /* Send particle coordinates to the pme nodes.
 929          * Since this is only implemented for domain decomposition
 930          * and domain decomposition does not use the graph,
 931          * we do not need to worry about shifting.
 932          */
 933         gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
 934                                  lambda[efptCOUL], lambda[efptVDW],
 935                                  (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)) != 0,
 936                                  step, wcycle);
 937     }
 938 #endif /* GMX_MPI */
 939
 940     if (useGpuPme)
 941     {
 942         launchPmeGpuSpread(fr->pmedata, box, as_rvec_array(x.unpaddedArrayRef().data()), flags, pmeFlags, wcycle);
 943     }
 944
 945     /* do gridding for pair search */
 946     if (bNS)
 947     {
 948         if (graph && bStateChanged)
 949         {
 950             /* Calculate intramolecular shift vectors to make molecules whole */
 951             mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
 952         }
 953
 954         // TODO
 955         // - vzero is constant, do we need to pass it?
 956         // - box_diag should be passed directly to nbnxn_put_on_grid
 957         //
 958         rvec vzero;
 959         clear_rvec(vzero);
 960
 961         rvec box_diag;
 962         box_diag[XX] = box[XX][XX];
 963         box_diag[YY] = box[YY][YY];
 964         box_diag[ZZ] = box[ZZ][ZZ];
 965
 966         wallcycle_start(wcycle, ewcNS);
 967         if (!DOMAINDECOMP(cr))
 968         {
 969             wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
 970             nbnxn_put_on_grid(nbv, box,
 971                               0, vzero, box_diag,
 972                               nullptr, 0, mdatoms->homenr, -1,
 973                               fr->cginfo, x.unpaddedArrayRef(),
 974                               0, nullptr);
 975             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
 976         }
 977         else
 978         {
 979             wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
 980             nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
 981                                        fr->cginfo, x.unpaddedArrayRef());
 982             wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
 983         }
 984
 985         nbv->setAtomProperties(*mdatoms, fr->cginfo);
 986
 987         wallcycle_stop(wcycle, ewcNS);
 988
 989         /* initialize the GPU nbnxm atom data and bonded data structures */
 990         if (bUseGPU)
 991         {
 992             wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
 993
 994             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 995             Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
 996             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
 997
 998             if (fr->gpuBonded)
 999             {
1000                 /* Now we put all atoms on the grid, we can assign bonded
1001                  * interactions to the GPU, where the grid order is
1002                  * needed. Also the xq, f and fshift device buffers have
1003                  * been reallocated if needed, so the bonded code can
1004                  * learn about them. */
1005                 // TODO the xq, f, and fshift buffers are now shared
1006                 // resources, so they should be maintained by a
1007                 // higher-level object than the nb module.
1008                 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbv->getGridIndices(),
1009                                                                       top->idef,
1010                                                                       Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1011                                                                       Nbnxm::gpu_get_f(nbv->gpu_nbv),
1012                                                                       Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1013             }
1014             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1015         }
1016
1017         // Need to run after the GPU-offload bonded interaction lists
1018         // are set up to be able to determine whether there is bonded work.
1019         setupForceWorkload(ppForceWorkload,
1020                            inputrec,
1021                            fr,
1022                            pull_work,
1023                            ed,
1024                            top->idef,
1025                            fcd,
1026                            flags);
1027     }
1028
1029     /* do local pair search */
1030     if (bNS)
1031     {
1032         // TODO: fuse this branch with the above bNS block
1033         wallcycle_start_nocount(wcycle, ewcNS);
1034         wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1035         /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1036         nbv->constructPairlist(Nbnxm::InteractionLocality::Local,
1037                                &top->excls, step, nrnb);
1038         wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1039         wallcycle_stop(wcycle, ewcNS);
1040
1041         if (useGpuXBufOps)
1042         {
1043             nbv->atomdata_init_copy_x_to_nbat_x_gpu( Nbnxm::AtomLocality::Local);
1044         }
1045
1046     }
1047     else
1048     {
1049         nbv->setCoordinates(Nbnxm::AtomLocality::Local, false,
1050                             x.unpaddedArrayRef(), useGpuXBufOps, pme_gpu_get_device_x(fr->pmedata), wcycle);
1051     }
1052
1053     if (bUseGPU)
1054     {
1055         ddBalanceRegionHandler.openBeforeForceComputationGpu();
1056
1057         wallcycle_start(wcycle, ewcLAUNCH_GPU);
1058
1059         wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1060         Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1061         if (bNS || !useGpuXBufOps)
1062         {
1063             Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1064                                       Nbnxm::AtomLocality::Local,
1065                                       ppForceWorkload->haveGpuBondedWork);
1066         }
1067         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1068         // with X buffer ops offloaded to the GPU on all but the search steps
1069
1070         // bonded work not split into separate local and non-local, so with DD
1071         // we can only launch the kernel after non-local coordinates have been received.
1072         if (ppForceWorkload->haveGpuBondedWork && !havePPDomainDecomposition(cr))
1073         {
1074             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1075             fr->gpuBonded->launchKernel(fr, flags, box);
1076             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1077         }
1078
1079         /* launch local nonbonded work on GPU */
1080         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1081         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFNo,
1082                      step, nrnb, wcycle);
1083         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1084         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1085     }
1086
1087     if (useGpuPme)
1088     {
1089         // In PME GPU and mixed mode we launch FFT / gather after the
1090         // X copy/transform to allow overlap as well as after the GPU NB
1091         // launch to avoid FFT launch overhead hijacking the CPU and delaying
1092         // the nonbonded kernel.
1093         launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1094     }
1095
1096     /* Communicate coordinates and sum dipole if necessary +
1097        do non-local pair search */
1098     if (havePPDomainDecomposition(cr))
1099     {
1100         if (bNS)
1101         {
1102             // TODO: fuse this branch with the above large bNS block
1103             wallcycle_start_nocount(wcycle, ewcNS);
1104             wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1105             /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1106             nbv->constructPairlist(Nbnxm::InteractionLocality::NonLocal,
1107                                    &top->excls, step, nrnb);
1108             wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1109             wallcycle_stop(wcycle, ewcNS);
1110
1111             if (useGpuXBufOps)
1112             {
1113
1114                 nbv->atomdata_init_copy_x_to_nbat_x_gpu( Nbnxm::AtomLocality::NonLocal);
1115             }
1116         }
1117         else
1118         {
1119             dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1120
1121             nbv->setCoordinates(Nbnxm::AtomLocality::NonLocal, false,
1122                                 x.unpaddedArrayRef(), useGpuXBufOps, pme_gpu_get_device_x(fr->pmedata), wcycle);
1123
1124         }
1125
1126         if (bUseGPU)
1127         {
1128             wallcycle_start(wcycle, ewcLAUNCH_GPU);
1129
1130             if (bNS || !useGpuXBufOps)
1131             {
1132                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1133                 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1134                                           Nbnxm::AtomLocality::NonLocal,
1135                                           ppForceWorkload->haveGpuBondedWork);
1136                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1137             }
1138
1139             if (ppForceWorkload->haveGpuBondedWork)
1140             {
1141                 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1142                 fr->gpuBonded->launchKernel(fr, flags, box);
1143                 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1144             }
1145
1146             /* launch non-local nonbonded tasks on GPU */
1147             wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1148             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1149                          step, nrnb, wcycle);
1150             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1151
1152             wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1153         }
1154     }
1155
1156     if (bUseGPU)
1157     {
1158         /* launch D2H copy-back F */
1159         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1160         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1161         if (havePPDomainDecomposition(cr))
1162         {
1163             Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1164                                       flags, Nbnxm::AtomLocality::NonLocal, ppForceWorkload->haveGpuBondedWork);
1165         }
1166         Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1167                                   flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork);
1168         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1169
1170         if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1171         {
1172             wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1173             fr->gpuBonded->launchEnergyTransfer();
1174             wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1175         }
1176         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1177     }
1178
1179     if (bStateChanged && inputrecNeedMutot(inputrec))
1180     {
1181         if (PAR(cr))
1182         {
1183             gmx_sumd(2*DIM, mu, cr);
1184
1185             ddBalanceRegionHandler.reopenRegionCpu();
1186         }
1187
1188         for (i = 0; i < 2; i++)
1189         {
1190             for (j = 0; j < DIM; j++)
1191             {
1192                 fr->mu_tot[i][j] = mu[i*DIM + j];
1193             }
1194         }
1195     }
1196     if (fr->efep == efepNO)
1197     {
1198         copy_rvec(fr->mu_tot[0], mu_tot);
1199     }
1200     else
1201     {
1202         for (j = 0; j < DIM; j++)
1203         {
1204             mu_tot[j] =
1205                 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1206                 lambda[efptCOUL]*fr->mu_tot[1][j];
1207         }
1208     }
1209
1210     /* Reset energies */
1211     reset_enerdata(enerd);
1212     clear_rvecs(SHIFTS, fr->fshift);
1213
1214     if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1215     {
1216         wallcycle_start(wcycle, ewcPPDURINGPME);
1217         dd_force_flop_start(cr->dd, nrnb);
1218     }
1219
1220     if (inputrec->bRot)
1221     {
1222         wallcycle_start(wcycle, ewcROT);
1223         do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, bNS);
1224         wallcycle_stop(wcycle, ewcROT);
1225     }
1226
1227     /* Start the force cycle counter.
1228      * Note that a different counter is used for dynamic load balancing.
1229      */
1230     wallcycle_start(wcycle, ewcFORCE);
1231
1232     // set up and clear force outputs
1233     struct ForceOutputs forceOut = setupForceOutputs(fr, pull_work, *inputrec, force, bDoForces,
1234                                                      ((flags & GMX_FORCE_VIRIAL) != 0), wcycle);
1235
1236     /* We calculate the non-bonded forces, when done on the CPU, here.
1237      * We do this before calling do_force_lowlevel, because in that
1238      * function, the listed forces are calculated before PME, which
1239      * does communication.  With this order, non-bonded and listed
1240      * force calculation imbalance can be balanced out by the domain
1241      * decomposition load balancing.
1242      */
1243
1244     if (!bUseOrEmulGPU)
1245     {
1246         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local, enbvClearFYes,
1247                      step, nrnb, wcycle);
1248     }
1249
1250     if (fr->efep != efepNO)
1251     {
1252         /* Calculate the local and non-local free energy interactions here.
1253          * Happens here on the CPU both with and without GPU.
1254          */
1255         nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::Local,
1256                                       fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1257                                       inputrec->fepvals, lambda.data(),
1258                                       enerd, flags, nrnb, wcycle);
1259
1260         if (havePPDomainDecomposition(cr))
1261         {
1262             nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::NonLocal,
1263                                           fr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, *mdatoms,
1264                                           inputrec->fepvals, lambda.data(),
1265                                           enerd, flags, nrnb, wcycle);
1266         }
1267     }
1268
1269     if (!bUseOrEmulGPU)
1270     {
1271         if (havePPDomainDecomposition(cr))
1272         {
1273             do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1274                          step, nrnb, wcycle);
1275         }
1276
1277         /* Add all the non-bonded force to the normal force array.
1278          * This can be split into a local and a non-local part when overlapping
1279          * communication with calculation with domain decomposition.
1280          */
1281         wallcycle_stop(wcycle, ewcFORCE);
1282
1283         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::All, forceOut.f, wcycle);
1284
1285         wallcycle_start_nocount(wcycle, ewcFORCE);
1286
1287         /* If there are multiple fshift output buffers we need to reduce them */
1288         if (flags & GMX_FORCE_VIRIAL)
1289         {
1290             /* This is not in a subcounter because it takes a
1291                negligible and constant-sized amount of time */
1292             nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat.get(),
1293                                                      fr->fshift);
1294         }
1295     }
1296
1297     /* update QMMMrec, if necessary */
1298     if (fr->bQMMM)
1299     {
1300         update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1301     }
1302
1303     /* Compute the bonded and non-bonded energies and optionally forces */
1304     do_force_lowlevel(fr, inputrec, &(top->idef),
1305                       cr, ms, nrnb, wcycle, mdatoms,
1306                       x, hist, forceOut.f, &forceOut.forceWithVirial, enerd, fcd,
1307                       box, lambda.data(), graph, fr->mu_tot,
1308                       flags,
1309                       ddBalanceRegionHandler);
1310
1311     wallcycle_stop(wcycle, ewcFORCE);
1312
1313     computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1314                          imdSession, pull_work, step, t, wcycle,
1315                          fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1316                          flags, &forceOut.forceWithVirial, enerd,
1317                          ed, bNS);
1318
1319     // Will store the amount of cycles spent waiting for the GPU that
1320     // will be later used in the DLB accounting.
1321     float cycles_wait_gpu = 0;
1322     if (bUseOrEmulGPU)
1323     {
1324         /* wait for non-local forces (or calculate in emulation mode) */
1325         if (havePPDomainDecomposition(cr))
1326         {
1327             if (bUseGPU)
1328             {
1329                 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1330                 Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1331                                             flags, Nbnxm::AtomLocality::NonLocal,
1332                                             ppForceWorkload->haveGpuBondedWork,
1333                                             enerd->grpp.ener[egLJSR].data(),
1334                                             enerd->grpp.ener[egCOULSR].data(),
1335                                             fr->fshift);
1336                 cycles_wait_gpu += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1337             }
1338             else
1339             {
1340                 wallcycle_start_nocount(wcycle, ewcFORCE);
1341                 do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::NonLocal, enbvClearFYes,
1342                              step, nrnb, wcycle);
1343                 wallcycle_stop(wcycle, ewcFORCE);
1344             }
1345
1346             nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::NonLocal,
1347                                           forceOut.f, wcycle);
1348
1349             if (fr->nbv->emulateGpu() && (flags & GMX_FORCE_VIRIAL))
1350             {
1351                 nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->nbat.get(),
1352                                                          fr->fshift);
1353             }
1354         }
1355     }
1356
1357     if (havePPDomainDecomposition(cr))
1358     {
1359         /* We are done with the CPU compute.
1360          * We will now communicate the non-local forces.
1361          * If we use a GPU this will overlap with GPU work, so in that case
1362          * we do not close the DD force balancing region here.
1363          */
1364         ddBalanceRegionHandler.closeAfterForceComputationCpu();
1365
1366         if (bDoForces)
1367         {
1368             dd_move_f(cr->dd, force.unpaddedArrayRef(), fr->fshift, wcycle);
1369         }
1370     }
1371
1372     // With both nonbonded and PME offloaded a GPU on the same rank, we use
1373     // an alternating wait/reduction scheme.
1374     bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPme && bUseGPU && !DOMAINDECOMP(cr));
1375     if (alternateGpuWait)
1376     {
1377         alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &force, &forceOut.forceWithVirial, fr->fshift, enerd,
1378                                     flags, pmeFlags, ppForceWorkload->haveGpuBondedWork, wcycle);
1379     }
1380
1381     if (!alternateGpuWait && useGpuPme)
1382     {
1383         pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial, enerd);
1384     }
1385
1386     /* Wait for local GPU NB outputs on the non-alternating wait path */
1387     if (!alternateGpuWait && bUseGPU)
1388     {
1389         /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1390          * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1391          * but even with a step of 0.1 ms the difference is less than 1%
1392          * of the step time.
1393          */
1394         const float gpuWaitApiOverheadMargin = 2e6f; /* cycles */
1395
1396         wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1397         Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1398                                     flags, Nbnxm::AtomLocality::Local, ppForceWorkload->haveGpuBondedWork,
1399                                     enerd->grpp.ener[egLJSR].data(),
1400                                     enerd->grpp.ener[egCOULSR].data(),
1401                                     fr->fshift);
1402         float cycles_tmp = wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1403
1404         if (ddBalanceRegionHandler.useBalancingRegion())
1405         {
1406             DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1407             if (bDoForces && cycles_tmp <= gpuWaitApiOverheadMargin)
1408             {
1409                 /* We measured few cycles, it could be that the kernel
1410                  * and transfer finished earlier and there was no actual
1411                  * wait time, only API call overhead.
1412                  * Then the actual time could be anywhere between 0 and
1413                  * cycles_wait_est. We will use half of cycles_wait_est.
1414                  */
1415                 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1416             }
1417             ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1418         }
1419     }
1420
1421     if (fr->nbv->emulateGpu())
1422     {
1423         // NOTE: emulation kernel is not included in the balancing region,
1424         // but emulation mode does not target performance anyway
1425         wallcycle_start_nocount(wcycle, ewcFORCE);
1426         do_nb_verlet(fr, ic, enerd, flags, Nbnxm::InteractionLocality::Local,
1427                      DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1428                      step, nrnb, wcycle);
1429         wallcycle_stop(wcycle, ewcFORCE);
1430     }
1431
1432     if (useGpuPme)
1433     {
1434         pme_gpu_reinit_computation(fr->pmedata, wcycle);
1435     }
1436
1437     if (bUseGPU)
1438     {
1439         /* now clear the GPU outputs while we finish the step on the CPU */
1440         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1441         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1442         Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, flags);
1443
1444         if (nbv->isDynamicPruningStepGpu(step))
1445         {
1446             nbv->dispatchPruneKernelGpu(step);
1447         }
1448         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1449         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1450     }
1451
1452     if (ppForceWorkload->haveGpuBondedWork && (flags & GMX_FORCE_ENERGY))
1453     {
1454         wallcycle_start(wcycle, ewcWAIT_GPU_BONDED);
1455         // in principle this should be included in the DD balancing region,
1456         // but generally it is infrequent so we'll omit it for the sake of
1457         // simpler code
1458         fr->gpuBonded->accumulateEnergyTerms(enerd);
1459         wallcycle_stop(wcycle, ewcWAIT_GPU_BONDED);
1460
1461         wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1462         wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_BONDED);
1463         fr->gpuBonded->clearEnergies();
1464         wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1465         wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1466     }
1467
1468     /* Do the nonbonded GPU (or emulation) force buffer reduction
1469      * on the non-alternating path. */
1470     if (bUseOrEmulGPU && !alternateGpuWait)
1471     {
1472         nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
1473                                       forceOut.f, wcycle);
1474     }
1475     if (DOMAINDECOMP(cr))
1476     {
1477         dd_force_flop_stop(cr->dd, nrnb);
1478     }
1479
1480     if (bDoForces)
1481     {
1482         /* If we have NoVirSum forces, but we do not calculate the virial,
1483          * we sum fr->f_novirsum=forceOut.f later.
1484          */
1485         if (vsite && !(fr->haveDirectVirialContributions && !(flags & GMX_FORCE_VIRIAL)))
1486         {
1487             spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, fr->fshift, FALSE, nullptr, nrnb,
1488                            &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1489         }
1490
1491         if (flags & GMX_FORCE_VIRIAL)
1492         {
1493             /* Calculation of the virial must be done after vsites! */
1494             calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f,
1495                         vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1496         }
1497     }
1498
1499     if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1500     {
1501         /* In case of node-splitting, the PP nodes receive the long-range
1502          * forces, virial and energy from the PME nodes here.
1503          */
1504         pme_receive_force_ener(cr, &forceOut.forceWithVirial, enerd, wcycle);
1505     }
1506
1507     if (bDoForces)
1508     {
1509         post_process_forces(cr, step, nrnb, wcycle,
1510                             top, box, as_rvec_array(x.unpaddedArrayRef().data()), forceOut.f, &forceOut.forceWithVirial,
1511                             vir_force, mdatoms, graph, fr, vsite,
1512                             flags);
1513     }
1514
1515     if (flags & GMX_FORCE_ENERGY)
1516     {
1517         /* Sum the potential energy terms from group contributions */
1518         sum_epot(&(enerd->grpp), enerd->term);
1519
1520         if (!EI_TPI(inputrec->eI))
1521         {
1522             checkPotentialEnergyValidity(step, *enerd, *inputrec);
1523         }
1524     }
1525
1526     /* In case we don't have constraints and are using GPUs, the next balancing
1527      * region starts here.
1528      * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1529      * virial calculation and COM pulling, is not thus not included in
1530      * the balance timing, which is ok as most tasks do communication.
1531      */
1532     ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);
1533 }