2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2013-2019,2020, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
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.
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.
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.
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.
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.
48 #include "gromacs/awh/awh.h"
49 #include "gromacs/domdec/dlbtiming.h"
50 #include "gromacs/domdec/domdec.h"
51 #include "gromacs/domdec/domdec_struct.h"
52 #include "gromacs/domdec/gpuhaloexchange.h"
53 #include "gromacs/domdec/partition.h"
54 #include "gromacs/essentialdynamics/edsam.h"
55 #include "gromacs/ewald/pme.h"
56 #include "gromacs/ewald/pme_pp.h"
57 #include "gromacs/ewald/pme_pp_comm_gpu.h"
58 #include "gromacs/gmxlib/network.h"
59 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
60 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
61 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
62 #include "gromacs/gpu_utils/gpu_utils.h"
63 #include "gromacs/imd/imd.h"
64 #include "gromacs/listed_forces/disre.h"
65 #include "gromacs/listed_forces/gpubonded.h"
66 #include "gromacs/listed_forces/listed_forces.h"
67 #include "gromacs/listed_forces/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/qmmm.h"
82 #include "gromacs/mdlib/update.h"
83 #include "gromacs/mdtypes/commrec.h"
84 #include "gromacs/mdtypes/enerdata.h"
85 #include "gromacs/mdtypes/forceoutput.h"
86 #include "gromacs/mdtypes/iforceprovider.h"
87 #include "gromacs/mdtypes/inputrec.h"
88 #include "gromacs/mdtypes/md_enums.h"
89 #include "gromacs/mdtypes/simulation_workload.h"
90 #include "gromacs/mdtypes/state.h"
91 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
92 #include "gromacs/nbnxm/gpu_data_mgmt.h"
93 #include "gromacs/nbnxm/nbnxm.h"
94 #include "gromacs/nbnxm/nbnxm_gpu.h"
95 #include "gromacs/pbcutil/ishift.h"
96 #include "gromacs/pbcutil/mshift.h"
97 #include "gromacs/pbcutil/pbc.h"
98 #include "gromacs/pulling/pull.h"
99 #include "gromacs/pulling/pull_rotation.h"
100 #include "gromacs/timing/cyclecounter.h"
101 #include "gromacs/timing/gpu_timing.h"
102 #include "gromacs/timing/wallcycle.h"
103 #include "gromacs/timing/wallcyclereporting.h"
104 #include "gromacs/timing/walltime_accounting.h"
105 #include "gromacs/topology/topology.h"
106 #include "gromacs/utility/arrayref.h"
107 #include "gromacs/utility/basedefinitions.h"
108 #include "gromacs/utility/cstringutil.h"
109 #include "gromacs/utility/exceptions.h"
110 #include "gromacs/utility/fatalerror.h"
111 #include "gromacs/utility/fixedcapacityvector.h"
112 #include "gromacs/utility/gmxassert.h"
113 #include "gromacs/utility/gmxmpi.h"
114 #include "gromacs/utility/logger.h"
115 #include "gromacs/utility/smalloc.h"
116 #include "gromacs/utility/strconvert.h"
117 #include "gromacs/utility/sysinfo.h"
119 using gmx::AtomLocality;
120 using gmx::DomainLifetimeWorkload;
121 using gmx::ForceOutputs;
122 using gmx::InteractionLocality;
123 using gmx::SimulationWorkload;
124 using gmx::StepWorkload;
126 // TODO: this environment variable allows us to verify before release
127 // that on less common architectures the total cost of polling is not larger than
128 // a blocking wait (so polling does not introduce overhead when the static
129 // PME-first ordering would suffice).
130 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
132 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
134 const int end = forceToAdd.size();
136 int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
137 #pragma omp parallel for num_threads(nt) schedule(static)
138 for (int i = 0; i < end; i++)
140 rvec_inc(f[i], forceToAdd[i]);
144 static void calc_virial(int start,
147 const gmx::ForceWithShiftForces& forceWithShiftForces,
149 const t_graph* graph,
152 const t_forcerec* fr,
155 /* The short-range virial from surrounding boxes */
156 const rvec* fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
157 calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, pbcType == PbcType::Screw, box);
158 inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
160 /* Calculate partial virial, for local atoms only, based on short range.
161 * Total virial is computed in global_stat, called from do_md
163 const rvec* f = as_rvec_array(forceWithShiftForces.force().data());
164 f_calc_vir(start, start + homenr, x, f, vir_part, graph, box);
165 inc_nrnb(nrnb, eNR_VIRIAL, homenr);
169 pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
173 static void pull_potential_wrapper(const t_commrec* cr,
174 const t_inputrec* ir,
176 gmx::ArrayRef<const gmx::RVec> x,
177 gmx::ForceWithVirial* force,
178 const t_mdatoms* mdatoms,
179 gmx_enerdata_t* enerd,
183 gmx_wallcycle_t wcycle)
188 /* Calculate the center of mass forces, this requires communication,
189 * which is why pull_potential is called close to other communication.
191 wallcycle_start(wcycle, ewcPULLPOT);
192 set_pbc(&pbc, ir->pbcType, box);
194 enerd->term[F_COM_PULL] += pull_potential(pull_work, mdatoms, &pbc, cr, t, lambda[efptRESTRAINT],
195 as_rvec_array(x.data()), force, &dvdl);
196 enerd->dvdl_lin[efptRESTRAINT] += dvdl;
197 wallcycle_stop(wcycle, ewcPULLPOT);
200 static void pme_receive_force_ener(t_forcerec* fr,
202 gmx::ForceWithVirial* forceWithVirial,
203 gmx_enerdata_t* enerd,
204 bool useGpuPmePpComms,
205 bool receivePmeForceToGpu,
206 gmx_wallcycle_t wcycle)
208 real e_q, e_lj, dvdl_q, dvdl_lj;
209 float cycles_ppdpme, cycles_seppme;
211 cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
212 dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
214 /* In case of node-splitting, the PP nodes receive the long-range
215 * forces, virial and energy from the PME nodes here.
217 wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
220 gmx_pme_receive_f(fr->pmePpCommGpu.get(), cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
221 useGpuPmePpComms, receivePmeForceToGpu, &cycles_seppme);
222 enerd->term[F_COUL_RECIP] += e_q;
223 enerd->term[F_LJ_RECIP] += e_lj;
224 enerd->dvdl_lin[efptCOUL] += dvdl_q;
225 enerd->dvdl_lin[efptVDW] += dvdl_lj;
229 dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
231 wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
234 static void print_large_forces(FILE* fp,
242 real force2Tolerance = gmx::square(forceTolerance);
243 gmx::index numNonFinite = 0;
244 for (int i = 0; i < md->homenr; i++)
246 real force2 = norm2(f[i]);
247 bool nonFinite = !std::isfinite(force2);
248 if (force2 >= force2Tolerance || nonFinite)
250 fprintf(fp, "step %" PRId64 " atom %6d x %8.3f %8.3f %8.3f force %12.5e\n", step,
251 ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
258 if (numNonFinite > 0)
260 /* Note that with MPI this fatal call on one rank might interrupt
261 * the printing on other ranks. But we can only avoid that with
262 * an expensive MPI barrier that we would need at each step.
264 gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
268 static void post_process_forces(const t_commrec* cr,
271 gmx_wallcycle_t wcycle,
272 const gmx_localtop_t* top,
275 ForceOutputs* forceOutputs,
277 const t_mdatoms* mdatoms,
278 const t_graph* graph,
279 const t_forcerec* fr,
280 const gmx_vsite_t* vsite,
281 const StepWorkload& stepWork)
283 rvec* f = as_rvec_array(forceOutputs->forceWithShiftForces().force().data());
285 if (fr->haveDirectVirialContributions)
287 auto& forceWithVirial = forceOutputs->forceWithVirial();
288 rvec* fDirectVir = as_rvec_array(forceWithVirial.force_.data());
292 /* Spread the mesh force on virtual sites to the other particles...
293 * This is parallellized. MPI communication is performed
294 * if the constructing atoms aren't local.
296 matrix virial = { { 0 } };
297 spread_vsite_f(vsite, x, fDirectVir, nullptr, stepWork.computeVirial, virial, nrnb,
298 &top->idef, fr->pbcType, fr->bMolPBC, graph, box, cr, wcycle);
299 forceWithVirial.addVirialContribution(virial);
302 if (stepWork.computeVirial)
304 /* Now add the forces, this is local */
305 sum_forces(f, forceWithVirial.force_);
307 /* Add the direct virial contributions */
309 forceWithVirial.computeVirial_,
310 "forceWithVirial should request virial computation when we request the virial");
311 m_add(vir_force, forceWithVirial.getVirial(), vir_force);
315 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
320 if (fr->print_force >= 0)
322 print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
326 static void do_nb_verlet(t_forcerec* fr,
327 const interaction_const_t* ic,
328 gmx_enerdata_t* enerd,
329 const StepWorkload& stepWork,
330 const InteractionLocality ilocality,
334 gmx_wallcycle_t wcycle)
336 if (!stepWork.computeNonbondedForces)
338 /* skip non-bonded calculation */
342 nonbonded_verlet_t* nbv = fr->nbv.get();
344 /* GPU kernel launch overhead is already timed separately */
345 if (fr->cutoff_scheme != ecutsVERLET)
347 gmx_incons("Invalid cut-off scheme passed!");
352 /* When dynamic pair-list pruning is requested, we need to prune
353 * at nstlistPrune steps.
355 if (nbv->isDynamicPruningStepCpu(step))
357 /* Prune the pair-list beyond fr->ic->rlistPrune using
358 * the current coordinates of the atoms.
360 wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
361 nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
362 wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
366 nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
369 static inline void clear_rvecs_omp(int n, rvec v[])
371 int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
373 /* Note that we would like to avoid this conditional by putting it
374 * into the omp pragma instead, but then we still take the full
375 * omp parallel for overhead (at least with gcc5).
379 for (int i = 0; i < n; i++)
386 #pragma omp parallel for num_threads(nth) schedule(static)
387 for (int i = 0; i < n; i++)
394 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
396 * \param groupOptions Group options, containing T-coupling options
398 static real averageKineticEnergyEstimate(const t_grpopts& groupOptions)
400 real nrdfCoupled = 0;
401 real nrdfUncoupled = 0;
402 real kineticEnergy = 0;
403 for (int g = 0; g < groupOptions.ngtc; g++)
405 if (groupOptions.tau_t[g] >= 0)
407 nrdfCoupled += groupOptions.nrdf[g];
408 kineticEnergy += groupOptions.nrdf[g] * 0.5 * groupOptions.ref_t[g] * BOLTZ;
412 nrdfUncoupled += groupOptions.nrdf[g];
416 /* This conditional with > also catches nrdf=0 */
417 if (nrdfCoupled > nrdfUncoupled)
419 return kineticEnergy * (nrdfCoupled + nrdfUncoupled) / nrdfCoupled;
427 /*! \brief This routine checks that the potential energy is finite.
429 * Always checks that the potential energy is finite. If step equals
430 * inputrec.init_step also checks that the magnitude of the potential energy
431 * is reasonable. Terminates with a fatal error when a check fails.
432 * Note that passing this check does not guarantee finite forces,
433 * since those use slightly different arithmetics. But in most cases
434 * there is just a narrow coordinate range where forces are not finite
435 * and energies are finite.
437 * \param[in] step The step number, used for checking and printing
438 * \param[in] enerd The energy data; the non-bonded group energies need to be added to
439 * enerd.term[F_EPOT] before calling this routine \param[in] inputrec The input record
441 static void checkPotentialEnergyValidity(int64_t step, const gmx_enerdata_t& enerd, const t_inputrec& inputrec)
443 /* Threshold valid for comparing absolute potential energy against
444 * the kinetic energy. Normally one should not consider absolute
445 * potential energy values, but with a factor of one million
446 * we should never get false positives.
448 constexpr real c_thresholdFactor = 1e6;
450 bool energyIsNotFinite = !std::isfinite(enerd.term[F_EPOT]);
451 real averageKineticEnergy = 0;
452 /* We only check for large potential energy at the initial step,
453 * because that is by far the most likely step for this too occur
454 * and because computing the average kinetic energy is not free.
455 * Note: nstcalcenergy >> 1 often does not allow to catch large energies
456 * before they become NaN.
458 if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
460 averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
463 if (energyIsNotFinite
464 || (averageKineticEnergy > 0 && enerd.term[F_EPOT] > c_thresholdFactor * averageKineticEnergy))
469 ": The total potential energy is %g, which is %s. The LJ and electrostatic "
470 "contributions to the energy are %g and %g, respectively. A %s potential energy "
471 "can be caused by overlapping interactions in bonded interactions or very large%s "
472 "coordinate values. Usually this is caused by a badly- or non-equilibrated initial "
473 "configuration, incorrect interactions or parameters in the topology.",
474 step, enerd.term[F_EPOT], energyIsNotFinite ? "not finite" : "extremely high",
475 enerd.term[F_LJ], enerd.term[F_COUL_SR],
476 energyIsNotFinite ? "non-finite" : "very high", energyIsNotFinite ? " or Nan" : "");
480 /*! \brief Return true if there are special forces computed this step.
482 * The conditionals exactly correspond to those in computeSpecialForces().
484 static bool haveSpecialForces(const t_inputrec& inputrec,
485 const gmx::ForceProviders& forceProviders,
486 const pull_t* pull_work,
487 const bool computeForces,
491 return ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
492 (inputrec.bPull && pull_have_potential(pull_work)) || // pull
493 inputrec.bRot || // enforced rotation
494 (ed != nullptr) || // flooding
495 (inputrec.bIMD && computeForces)); // IMD
498 /*! \brief Compute forces and/or energies for special algorithms
500 * The intention is to collect all calls to algorithms that compute
501 * forces on local atoms only and that do not contribute to the local
502 * virial sum (but add their virial contribution separately).
503 * Eventually these should likely all become ForceProviders.
504 * Within this function the intention is to have algorithms that do
505 * global communication at the end, so global barriers within the MD loop
506 * are as close together as possible.
508 * \param[in] fplog The log file
509 * \param[in] cr The communication record
510 * \param[in] inputrec The input record
511 * \param[in] awh The Awh module (nullptr if none in use).
512 * \param[in] enforcedRotation Enforced rotation module.
513 * \param[in] imdSession The IMD session
514 * \param[in] pull_work The pull work structure.
515 * \param[in] step The current MD step
516 * \param[in] t The current time
517 * \param[in,out] wcycle Wallcycle accounting struct
518 * \param[in,out] forceProviders Pointer to a list of force providers
519 * \param[in] box The unit cell
520 * \param[in] x The coordinates
521 * \param[in] mdatoms Per atom properties
522 * \param[in] lambda Array of free-energy lambda values
523 * \param[in] stepWork Step schedule flags
524 * \param[in,out] forceWithVirial Force and virial buffers
525 * \param[in,out] enerd Energy buffer
526 * \param[in,out] ed Essential dynamics pointer
527 * \param[in] didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
529 * \todo Remove didNeighborSearch, which is used incorrectly.
530 * \todo Convert all other algorithms called here to ForceProviders.
532 static void computeSpecialForces(FILE* fplog,
534 const t_inputrec* inputrec,
536 gmx_enfrot* enforcedRotation,
537 gmx::ImdSession* imdSession,
541 gmx_wallcycle_t wcycle,
542 gmx::ForceProviders* forceProviders,
544 gmx::ArrayRef<const gmx::RVec> x,
545 const t_mdatoms* mdatoms,
547 const StepWorkload& stepWork,
548 gmx::ForceWithVirial* forceWithVirial,
549 gmx_enerdata_t* enerd,
551 bool didNeighborSearch)
553 /* NOTE: Currently all ForceProviders only provide forces.
554 * When they also provide energies, remove this conditional.
556 if (stepWork.computeForces)
558 gmx::ForceProviderInput forceProviderInput(x, *mdatoms, t, box, *cr);
559 gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
561 /* Collect forces from modules */
562 forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
565 if (inputrec->bPull && pull_have_potential(pull_work))
567 pull_potential_wrapper(cr, inputrec, box, x, forceWithVirial, mdatoms, enerd, pull_work,
572 enerd->term[F_COM_PULL] += awh->applyBiasForcesAndUpdateBias(
573 inputrec->pbcType, *mdatoms, box, forceWithVirial, t, step, wcycle, fplog);
577 rvec* f = as_rvec_array(forceWithVirial->force_.data());
579 /* Add the forces from enforced rotation potentials (if any) */
582 wallcycle_start(wcycle, ewcROTadd);
583 enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
584 wallcycle_stop(wcycle, ewcROTadd);
589 /* Note that since init_edsam() is called after the initialization
590 * of forcerec, edsam doesn't request the noVirSum force buffer.
591 * Thus if no other algorithm (e.g. PME) requires it, the forces
592 * here will contribute to the virial.
594 do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
597 /* Add forces from interactive molecular dynamics (IMD), if any */
598 if (inputrec->bIMD && stepWork.computeForces)
600 imdSession->applyForces(f);
604 /*! \brief Makes PME flags from StepWorkload data.
606 * \param[in] stepWork Step schedule flags
609 static int makePmeFlags(const StepWorkload& stepWork)
611 return GMX_PME_SPREAD | GMX_PME_SOLVE | (stepWork.computeVirial ? GMX_PME_CALC_ENER_VIR : 0)
612 | (stepWork.computeEnergy ? GMX_PME_CALC_ENER_VIR : 0)
613 | (stepWork.computeForces ? GMX_PME_CALC_F : 0);
616 /*! \brief Launch the prepare_step and spread stages of PME GPU.
618 * \param[in] pmedata The PME structure
619 * \param[in] box The box matrix
620 * \param[in] stepWork Step schedule flags
621 * \param[in] pmeFlags PME flags
622 * \param[in] xReadyOnDevice Event synchronizer indicating that the coordinates are ready in
623 * the device memory. \param[in] wcycle The wallcycle structure
625 static inline void launchPmeGpuSpread(gmx_pme_t* pmedata,
627 const StepWorkload& stepWork,
629 GpuEventSynchronizer* xReadyOnDevice,
630 gmx_wallcycle_t wcycle)
632 pme_gpu_prepare_computation(pmedata, stepWork.haveDynamicBox, box, wcycle, pmeFlags,
633 stepWork.useGpuPmeFReduction);
634 pme_gpu_launch_spread(pmedata, xReadyOnDevice, wcycle);
637 /*! \brief Launch the FFT and gather stages of PME GPU
639 * This function only implements setting the output forces (no accumulation).
641 * \param[in] pmedata The PME structure
642 * \param[in] wcycle The wallcycle structure
644 static void launchPmeGpuFftAndGather(gmx_pme_t* pmedata, gmx_wallcycle_t wcycle)
646 pme_gpu_launch_complex_transforms(pmedata, wcycle);
647 pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
651 * Polling wait for either of the PME or nonbonded GPU tasks.
653 * Instead of a static order in waiting for GPU tasks, this function
654 * polls checking which of the two tasks completes first, and does the
655 * associated force buffer reduction overlapped with the other task.
656 * By doing that, unlike static scheduling order, it can always overlap
657 * one of the reductions, regardless of the GPU task completion order.
659 * \param[in] nbv Nonbonded verlet structure
660 * \param[in,out] pmedata PME module data
661 * \param[in,out] forceOutputs Output buffer for the forces and virial
662 * \param[in,out] enerd Energy data structure results are reduced into
663 * \param[in] stepWork Step schedule flags
664 * \param[in] pmeFlags PME flags
665 * \param[in] wcycle The wallcycle structure
667 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t* nbv,
669 gmx::ForceOutputs* forceOutputs,
670 gmx_enerdata_t* enerd,
671 const StepWorkload& stepWork,
673 gmx_wallcycle_t wcycle)
675 bool isPmeGpuDone = false;
676 bool isNbGpuDone = false;
679 gmx::ForceWithShiftForces& forceWithShiftForces = forceOutputs->forceWithShiftForces();
680 gmx::ForceWithVirial& forceWithVirial = forceOutputs->forceWithVirial();
682 gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
684 while (!isPmeGpuDone || !isNbGpuDone)
688 GpuTaskCompletion completionType =
689 (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
690 isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, &forceWithVirial,
691 enerd, completionType);
696 GpuTaskCompletion completionType =
697 (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
698 isNbGpuDone = Nbnxm::gpu_try_finish_task(
699 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
700 enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(),
701 completionType, wcycle);
705 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShiftForces.force());
711 /*! \brief Set up the different force buffers; also does clearing.
713 * \param[in] fr force record pointer
714 * \param[in] pull_work The pull work object.
715 * \param[in] inputrec input record
716 * \param[in] force force array
717 * \param[in] stepWork Step schedule flags
718 * \param[out] wcycle wallcycle recording structure
720 * \returns Cleared force output structure
722 static ForceOutputs setupForceOutputs(t_forcerec* fr,
724 const t_inputrec& inputrec,
725 gmx::ArrayRefWithPadding<gmx::RVec> force,
726 const StepWorkload& stepWork,
727 gmx_wallcycle_t wcycle)
729 wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
731 /* NOTE: We assume fr->shiftForces is all zeros here */
732 gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial, fr->shiftForces);
734 if (stepWork.computeForces)
736 /* Clear the short- and long-range forces */
737 clear_rvecs_omp(fr->natoms_force_constr, as_rvec_array(forceWithShiftForces.force().data()));
740 /* If we need to compute the virial, we might need a separate
741 * force buffer for algorithms for which the virial is calculated
742 * directly, such as PME. Otherwise, forceWithVirial uses the
743 * the same force (f in legacy calls) buffer as other algorithms.
745 const bool useSeparateForceWithVirialBuffer =
746 (stepWork.computeForces && (stepWork.computeVirial && fr->haveDirectVirialContributions));
747 /* forceWithVirial uses the local atom range only */
748 gmx::ForceWithVirial forceWithVirial(useSeparateForceWithVirialBuffer ? fr->forceBufferForDirectVirialContributions
749 : force.unpaddedArrayRef(),
750 stepWork.computeVirial);
752 if (useSeparateForceWithVirialBuffer)
754 /* TODO: update comment
755 * We only compute forces on local atoms. Note that vsites can
756 * spread to non-local atoms, but that part of the buffer is
757 * cleared separately in the vsite spreading code.
759 clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
762 if (inputrec.bPull && pull_have_constraint(pull_work))
764 clear_pull_forces(pull_work);
767 wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
769 return ForceOutputs(forceWithShiftForces, forceWithVirial);
773 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
775 static DomainLifetimeWorkload setupDomainLifetimeWorkload(const t_inputrec& inputrec,
776 const t_forcerec& fr,
777 const pull_t* pull_work,
781 const t_mdatoms& mdatoms,
782 const SimulationWorkload& simulationWork,
783 const StepWorkload& stepWork)
785 DomainLifetimeWorkload domainWork;
786 // Note that haveSpecialForces is constant over the whole run
787 domainWork.haveSpecialForces =
788 haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
789 domainWork.haveCpuBondedWork = haveCpuBondeds(fr);
790 domainWork.haveGpuBondedWork = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
791 domainWork.haveRestraintsWork = haveRestraints(idef, fcd);
792 domainWork.haveCpuListedForceWork = haveCpuListedForces(fr, idef, fcd);
793 // Note that haveFreeEnergyWork is constant over the whole run
794 domainWork.haveFreeEnergyWork = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
795 // We assume we have local force work if there are CPU
796 // force tasks including PME or nonbondeds.
797 domainWork.haveCpuLocalForceWork =
798 domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork
799 || domainWork.haveFreeEnergyWork || simulationWork.useCpuNonbonded || simulationWork.useCpuPme
800 || simulationWork.haveEwaldSurfaceContribution || inputrec.nwall > 0;
805 /*! \brief Set up force flag stuct from the force bitmask.
807 * \param[in] legacyFlags Force bitmask flags used to construct the new flags
808 * \param[in] isNonbondedOn Global override, if false forces to turn off all nonbonded calculation.
809 * \param[in] simulationWork Simulation workload description.
810 * \param[in] rankHasPmeDuty If this rank computes PME.
812 * \returns New Stepworkload description.
814 static StepWorkload setupStepWorkload(const int legacyFlags,
815 const bool isNonbondedOn,
816 const SimulationWorkload& simulationWork,
817 const bool rankHasPmeDuty)
820 flags.stateChanged = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
821 flags.haveDynamicBox = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
822 flags.doNeighborSearch = ((legacyFlags & GMX_FORCE_NS) != 0);
823 flags.computeVirial = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
824 flags.computeEnergy = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
825 flags.computeForces = ((legacyFlags & GMX_FORCE_FORCES) != 0);
826 flags.computeListedForces = ((legacyFlags & GMX_FORCE_LISTED) != 0);
827 flags.computeNonbondedForces = ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && isNonbondedOn;
828 flags.computeDhdl = ((legacyFlags & GMX_FORCE_DHDL) != 0);
830 if (simulationWork.useGpuBufferOps)
832 GMX_ASSERT(simulationWork.useGpuNonbonded,
833 "Can only offload buffer ops if nonbonded computation is also offloaded");
835 flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
836 // on virial steps the CPU reduction path is taken
837 flags.useGpuFBufferOps = simulationWork.useGpuBufferOps && !flags.computeVirial;
838 flags.useGpuPmeFReduction = flags.useGpuFBufferOps
839 && (simulationWork.useGpuPme
840 && (rankHasPmeDuty || simulationWork.useGpuPmePpCommunication));
846 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
848 * TODO: eliminate \p useGpuPmeOnThisRank when this is
849 * incorporated in DomainLifetimeWorkload.
851 static void launchGpuEndOfStepTasks(nonbonded_verlet_t* nbv,
852 gmx::GpuBonded* gpuBonded,
854 gmx_enerdata_t* enerd,
855 const gmx::MdrunScheduleWorkload& runScheduleWork,
856 bool useGpuPmeOnThisRank,
858 gmx_wallcycle_t wcycle)
860 if (runScheduleWork.simulationWork.useGpuNonbonded)
862 /* Launch pruning before buffer clearing because the API overhead of the
863 * clear kernel launches can leave the GPU idle while it could be running
866 if (nbv->isDynamicPruningStepGpu(step))
868 nbv->dispatchPruneKernelGpu(step);
871 /* now clear the GPU outputs while we finish the step on the CPU */
872 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
873 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
874 Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
875 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
876 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
879 if (useGpuPmeOnThisRank)
881 pme_gpu_reinit_computation(pmedata, wcycle);
884 if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
886 // in principle this should be included in the DD balancing region,
887 // but generally it is infrequent so we'll omit it for the sake of
889 gpuBonded->waitAccumulateEnergyTerms(enerd);
891 gpuBonded->clearEnergies();
896 void do_force(FILE* fplog,
898 const gmx_multisim_t* ms,
899 const t_inputrec* inputrec,
901 gmx_enfrot* enforcedRotation,
902 gmx::ImdSession* imdSession,
906 gmx_wallcycle_t wcycle,
907 const gmx_localtop_t* top,
909 gmx::ArrayRefWithPadding<gmx::RVec> x,
911 gmx::ArrayRefWithPadding<gmx::RVec> force,
913 const t_mdatoms* mdatoms,
914 gmx_enerdata_t* enerd,
916 gmx::ArrayRef<real> lambda,
919 gmx::MdrunScheduleWorkload* runScheduleWork,
920 const gmx_vsite_t* vsite,
925 const DDBalanceRegionHandler& ddBalanceRegionHandler)
929 nonbonded_verlet_t* nbv = fr->nbv.get();
930 interaction_const_t* ic = fr->ic;
931 gmx::StatePropagatorDataGpu* stateGpu = fr->stateGpu;
933 const SimulationWorkload& simulationWork = runScheduleWork->simulationWork;
936 runScheduleWork->stepWork = setupStepWorkload(legacyFlags, fr->bNonbonded, simulationWork,
937 thisRankHasDuty(cr, DUTY_PME));
938 const StepWorkload& stepWork = runScheduleWork->stepWork;
941 const bool useGpuPmeOnThisRank = simulationWork.useGpuPme && thisRankHasDuty(cr, DUTY_PME);
942 const int pmeFlags = makePmeFlags(stepWork);
944 /* At a search step we need to start the first balancing region
945 * somewhere early inside the step after communication during domain
946 * decomposition (and not during the previous step as usual).
948 if (stepWork.doNeighborSearch)
950 ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
954 const int homenr = mdatoms->homenr;
956 clear_mat(vir_force);
958 if (stepWork.stateChanged && simulationWork.computeMuTot)
960 /* Calculate total (local) dipole moment in a temporary common array.
961 * This makes it possible to sum them over nodes faster.
963 calc_mu(start, homenr, x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB,
964 mdatoms->nChargePerturbed, mu, mu + DIM);
967 if (fr->pbcType != PbcType::No)
969 /* Compute shift vectors every step,
970 * because of pressure coupling or box deformation!
972 if (stepWork.haveDynamicBox && stepWork.stateChanged)
974 calc_shifts(box, fr->shift_vec);
977 const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
978 const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
981 put_atoms_in_box_omp(fr->pbcType, box, x.unpaddedArrayRef().subArray(0, homenr),
982 gmx_omp_nthreads_get(emntDefault));
983 inc_nrnb(nrnb, eNR_SHIFTX, homenr);
985 else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
987 unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
991 nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox, fr->shift_vec, nbv->nbat.get());
993 // Coordinates on the device are needed if PME or BufferOps are offloaded.
994 // The local coordinates can be copied right away.
995 // NOTE: Consider moving this copy to right after they are updated and constrained,
996 // if the later is not offloaded.
997 if (useGpuPmeOnThisRank || stepWork.useGpuXBufferOps)
999 if (stepWork.doNeighborSearch)
1001 // TODO refactor this to do_md, after partitioning.
1002 stateGpu->reinit(mdatoms->homenr,
1003 cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1004 if (useGpuPmeOnThisRank)
1006 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1007 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1010 // We need to copy coordinates when:
1011 // 1. Update is not offloaded
1012 // 2. The buffers were reinitialized on search step
1013 if (!simulationWork.useGpuUpdate || stepWork.doNeighborSearch)
1015 stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1019 // TODO Update this comment when introducing SimulationWorkload
1021 // The conditions for gpuHaloExchange e.g. using GPU buffer
1022 // operations were checked before construction, so here we can
1023 // just use it and assert upon any conditions.
1024 const bool ddUsesGpuDirectCommunication =
1025 ((cr->dd != nullptr) && (!cr->dd->gpuHaloExchange.empty()));
1026 GMX_ASSERT(!ddUsesGpuDirectCommunication || stepWork.useGpuXBufferOps,
1027 "Must use coordinate buffer ops with GPU halo exchange");
1028 const bool useGpuForcesHaloExchange = ddUsesGpuDirectCommunication && stepWork.useGpuFBufferOps;
1030 // Copy coordinate from the GPU if update is on the GPU and there
1031 // are forces to be computed on the CPU, or for the computation of
1032 // virial, or if host-side data will be transferred from this task
1033 // to a remote task for halo exchange or PME-PP communication. At
1034 // search steps the current coordinates are already on the host,
1035 // hence copy is not needed.
1036 const bool haveHostPmePpComms =
1037 !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication;
1038 const bool haveHostHaloExchangeComms = havePPDomainDecomposition(cr) && !ddUsesGpuDirectCommunication;
1040 bool gmx_used_in_debug haveCopiedXFromGpu = false;
1041 if (simulationWork.useGpuUpdate && !stepWork.doNeighborSearch
1042 && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial
1043 || haveHostPmePpComms || haveHostHaloExchangeComms))
1045 stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::Local);
1046 haveCopiedXFromGpu = true;
1049 const auto localXReadyOnDevice = (stateGpu != nullptr)
1050 ? stateGpu->getCoordinatesReadyOnDeviceEvent(
1051 AtomLocality::Local, simulationWork, stepWork)
1055 if (!thisRankHasDuty(cr, DUTY_PME))
1057 /* Send particle coordinates to the pme nodes.
1058 * Since this is only implemented for domain decomposition
1059 * and domain decomposition does not use the graph,
1060 * we do not need to worry about shifting.
1062 bool reinitGpuPmePpComms = simulationWork.useGpuPmePpCommunication && (stepWork.doNeighborSearch);
1063 bool sendCoordinatesFromGpu =
1064 simulationWork.useGpuPmePpCommunication && !(stepWork.doNeighborSearch);
1066 if (!stepWork.doNeighborSearch && simulationWork.useGpuUpdate && !sendCoordinatesFromGpu)
1068 GMX_RELEASE_ASSERT(false,
1069 "GPU update and separate PME ranks are only supported with GPU "
1070 "direct communication!");
1071 // TODO: when this code-path becomes supported add:
1072 // stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1075 gmx_pme_send_coordinates(fr, cr, box, as_rvec_array(x.unpaddedArrayRef().data()), lambda[efptCOUL],
1076 lambda[efptVDW], (stepWork.computeVirial || stepWork.computeEnergy),
1077 step, simulationWork.useGpuPmePpCommunication, reinitGpuPmePpComms,
1078 sendCoordinatesFromGpu, localXReadyOnDevice, wcycle);
1080 #endif /* GMX_MPI */
1082 if (useGpuPmeOnThisRank)
1084 launchPmeGpuSpread(fr->pmedata, box, stepWork, pmeFlags, localXReadyOnDevice, wcycle);
1087 /* do gridding for pair search */
1088 if (stepWork.doNeighborSearch)
1090 if (graph && stepWork.stateChanged)
1092 /* Calculate intramolecular shift vectors to make molecules whole */
1093 mk_mshift(fplog, graph, fr->pbcType, box, as_rvec_array(x.unpaddedArrayRef().data()));
1097 // - vzero is constant, do we need to pass it?
1098 // - box_diag should be passed directly to nbnxn_put_on_grid
1104 box_diag[XX] = box[XX][XX];
1105 box_diag[YY] = box[YY][YY];
1106 box_diag[ZZ] = box[ZZ][ZZ];
1108 wallcycle_start(wcycle, ewcNS);
1109 if (!DOMAINDECOMP(cr))
1111 wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1112 nbnxn_put_on_grid(nbv, box, 0, vzero, box_diag, nullptr, { 0, mdatoms->homenr }, -1,
1113 fr->cginfo, x.unpaddedArrayRef(), 0, nullptr);
1114 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1118 wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1119 nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd), fr->cginfo, x.unpaddedArrayRef());
1120 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1123 nbv->setAtomProperties(*mdatoms, fr->cginfo);
1125 wallcycle_stop(wcycle, ewcNS);
1127 /* initialize the GPU nbnxm atom data and bonded data structures */
1128 if (simulationWork.useGpuNonbonded)
1130 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1132 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1133 Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1134 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1138 /* Now we put all atoms on the grid, we can assign bonded
1139 * interactions to the GPU, where the grid order is
1140 * needed. Also the xq, f and fshift device buffers have
1141 * been reallocated if needed, so the bonded code can
1142 * learn about them. */
1143 // TODO the xq, f, and fshift buffers are now shared
1144 // resources, so they should be maintained by a
1145 // higher-level object than the nb module.
1146 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(
1147 nbv->getGridIndices(), top->idef, Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1148 Nbnxm::gpu_get_f(nbv->gpu_nbv), Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1150 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1153 // Need to run after the GPU-offload bonded interaction lists
1154 // are set up to be able to determine whether there is bonded work.
1155 runScheduleWork->domainWork = setupDomainLifetimeWorkload(
1156 *inputrec, *fr, pull_work, ed, top->idef, *fcd, *mdatoms, simulationWork, stepWork);
1158 wallcycle_start_nocount(wcycle, ewcNS);
1159 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1160 /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1161 nbv->constructPairlist(InteractionLocality::Local, top->excls, step, nrnb);
1163 nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::Local);
1165 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1166 wallcycle_stop(wcycle, ewcNS);
1168 if (stepWork.useGpuXBufferOps)
1170 nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1172 // For force buffer ops, we use the below conditon rather than
1173 // useGpuFBufferOps to ensure that init is performed even if this
1174 // NS step is also a virial step (on which f buf ops are deactivated).
1175 if (simulationWork.useGpuBufferOps && simulationWork.useGpuNonbonded && (GMX_GPU == GMX_GPU_CUDA))
1177 GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1178 nbv->atomdata_init_add_nbat_f_to_f_gpu(stateGpu->fReducedOnDevice());
1181 else if (!EI_TPI(inputrec->eI))
1183 if (stepWork.useGpuXBufferOps)
1185 GMX_ASSERT(stateGpu, "stateGpu should be valid when buffer ops are offloaded");
1186 nbv->convertCoordinatesGpu(AtomLocality::Local, false, stateGpu->getCoordinates(),
1187 localXReadyOnDevice);
1191 if (simulationWork.useGpuUpdate)
1193 GMX_ASSERT(stateGpu, "need a valid stateGpu object");
1194 GMX_ASSERT(haveCopiedXFromGpu,
1195 "a wait should only be triggered if copy has been scheduled");
1196 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1198 nbv->convertCoordinates(AtomLocality::Local, false, x.unpaddedArrayRef());
1202 const gmx::DomainLifetimeWorkload& domainWork = runScheduleWork->domainWork;
1204 if (simulationWork.useGpuNonbonded)
1206 ddBalanceRegionHandler.openBeforeForceComputationGpu();
1208 wallcycle_start(wcycle, ewcLAUNCH_GPU);
1210 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1211 Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1212 if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1214 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::Local);
1216 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1217 // with X buffer ops offloaded to the GPU on all but the search steps
1219 // bonded work not split into separate local and non-local, so with DD
1220 // we can only launch the kernel after non-local coordinates have been received.
1221 if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1223 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1224 fr->gpuBonded->launchKernel(fr, stepWork, box);
1225 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1228 /* launch local nonbonded work on GPU */
1229 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1230 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFNo, step, nrnb, wcycle);
1231 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1232 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1235 if (useGpuPmeOnThisRank)
1237 // In PME GPU and mixed mode we launch FFT / gather after the
1238 // X copy/transform to allow overlap as well as after the GPU NB
1239 // launch to avoid FFT launch overhead hijacking the CPU and delaying
1240 // the nonbonded kernel.
1241 launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1244 /* Communicate coordinates and sum dipole if necessary +
1245 do non-local pair search */
1246 if (havePPDomainDecomposition(cr))
1248 if (stepWork.doNeighborSearch)
1250 // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1251 wallcycle_start_nocount(wcycle, ewcNS);
1252 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1253 /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1254 nbv->constructPairlist(InteractionLocality::NonLocal, top->excls, step, nrnb);
1256 nbv->setupGpuShortRangeWork(fr->gpuBonded, InteractionLocality::NonLocal);
1257 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1258 wallcycle_stop(wcycle, ewcNS);
1259 // TODO refactor this GPU halo exchange re-initialisation
1260 // to location in do_md where GPU halo exchange is
1261 // constructed at partitioning, after above stateGpu
1262 // re-initialization has similarly been refactored
1263 if (ddUsesGpuDirectCommunication)
1265 reinitGpuHaloExchange(*cr, stateGpu->getCoordinates(), stateGpu->getForces());
1270 if (ddUsesGpuDirectCommunication)
1272 // The following must be called after local setCoordinates (which records an event
1273 // when the coordinate data has been copied to the device).
1274 communicateGpuHaloCoordinates(*cr, box, localXReadyOnDevice);
1276 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1278 // non-local part of coordinate buffer must be copied back to host for CPU work
1279 stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1284 // Note: GPU update + DD without direct communication is not supported,
1285 // a waitCoordinatesReadyOnHost() should be issued if it will be.
1286 GMX_ASSERT(!simulationWork.useGpuUpdate,
1287 "GPU update is not supported with CPU halo exchange");
1288 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1291 if (stepWork.useGpuXBufferOps)
1293 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1295 stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), AtomLocality::NonLocal);
1297 nbv->convertCoordinatesGpu(AtomLocality::NonLocal, false, stateGpu->getCoordinates(),
1298 stateGpu->getCoordinatesReadyOnDeviceEvent(
1299 AtomLocality::NonLocal, simulationWork, stepWork));
1303 nbv->convertCoordinates(AtomLocality::NonLocal, false, x.unpaddedArrayRef());
1307 if (simulationWork.useGpuNonbonded)
1309 wallcycle_start(wcycle, ewcLAUNCH_GPU);
1311 if (stepWork.doNeighborSearch || !stepWork.useGpuXBufferOps)
1313 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1314 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(), AtomLocality::NonLocal);
1315 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1318 if (domainWork.haveGpuBondedWork)
1320 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1321 fr->gpuBonded->launchKernel(fr, stepWork, box);
1322 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1325 /* launch non-local nonbonded tasks on GPU */
1326 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1327 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1329 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1331 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1335 if (simulationWork.useGpuNonbonded)
1337 /* launch D2H copy-back F */
1338 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1339 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1341 if (havePPDomainDecomposition(cr))
1343 Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::NonLocal);
1345 Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(), stepWork, AtomLocality::Local);
1346 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1348 if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1350 fr->gpuBonded->launchEnergyTransfer();
1352 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1355 if (stepWork.stateChanged && simulationWork.computeMuTot)
1359 gmx_sumd(2 * DIM, mu, cr);
1361 ddBalanceRegionHandler.reopenRegionCpu();
1364 for (i = 0; i < 2; i++)
1366 for (j = 0; j < DIM; j++)
1368 fr->mu_tot[i][j] = mu[i * DIM + j];
1372 if (mdatoms->nChargePerturbed == 0)
1374 copy_rvec(fr->mu_tot[0], mu_tot);
1378 for (j = 0; j < DIM; j++)
1380 mu_tot[j] = (1.0 - lambda[efptCOUL]) * fr->mu_tot[0][j] + lambda[efptCOUL] * fr->mu_tot[1][j];
1384 /* Reset energies */
1385 reset_enerdata(enerd);
1386 /* Clear the shift forces */
1387 // TODO: This should be linked to the shift force buffer in use, or cleared before use instead
1388 for (gmx::RVec& elem : fr->shiftForces)
1390 elem = { 0.0_real, 0.0_real, 0.0_real };
1393 if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1395 wallcycle_start(wcycle, ewcPPDURINGPME);
1396 dd_force_flop_start(cr->dd, nrnb);
1399 // For the rest of the CPU tasks that depend on GPU-update produced coordinates,
1400 // this wait ensures that the D2H transfer is complete.
1401 if ((simulationWork.useGpuUpdate)
1402 && (runScheduleWork->domainWork.haveCpuLocalForceWork || stepWork.computeVirial))
1404 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::Local);
1409 wallcycle_start(wcycle, ewcROT);
1410 do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step,
1411 stepWork.doNeighborSearch);
1412 wallcycle_stop(wcycle, ewcROT);
1415 /* Start the force cycle counter.
1416 * Note that a different counter is used for dynamic load balancing.
1418 wallcycle_start(wcycle, ewcFORCE);
1420 // Set up and clear force outputs.
1421 // We use std::move to keep the compiler happy, it has no effect.
1422 ForceOutputs forceOut =
1423 setupForceOutputs(fr, pull_work, *inputrec, std::move(force), stepWork, wcycle);
1425 /* We calculate the non-bonded forces, when done on the CPU, here.
1426 * We do this before calling do_force_lowlevel, because in that
1427 * function, the listed forces are calculated before PME, which
1428 * does communication. With this order, non-bonded and listed
1429 * force calculation imbalance can be balanced out by the domain
1430 * decomposition load balancing.
1433 const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1435 if (!useOrEmulateGpuNb)
1437 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local, enbvClearFYes, step, nrnb, wcycle);
1440 if (fr->efep != efepNO)
1442 /* Calculate the local and non-local free energy interactions here.
1443 * Happens here on the CPU both with and without GPU.
1445 nbv->dispatchFreeEnergyKernel(InteractionLocality::Local, fr,
1446 as_rvec_array(x.unpaddedArrayRef().data()),
1447 &forceOut.forceWithShiftForces(), *mdatoms, inputrec->fepvals,
1448 lambda.data(), enerd, stepWork, nrnb);
1450 if (havePPDomainDecomposition(cr))
1452 nbv->dispatchFreeEnergyKernel(InteractionLocality::NonLocal, fr,
1453 as_rvec_array(x.unpaddedArrayRef().data()),
1454 &forceOut.forceWithShiftForces(), *mdatoms,
1455 inputrec->fepvals, lambda.data(), enerd, stepWork, nrnb);
1459 if (!useOrEmulateGpuNb)
1461 if (havePPDomainDecomposition(cr))
1463 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFNo, step,
1467 if (stepWork.computeForces)
1469 /* Add all the non-bonded force to the normal force array.
1470 * This can be split into a local and a non-local part when overlapping
1471 * communication with calculation with domain decomposition.
1473 wallcycle_stop(wcycle, ewcFORCE);
1474 nbv->atomdata_add_nbat_f_to_f(AtomLocality::All, forceOut.forceWithShiftForces().force());
1475 wallcycle_start_nocount(wcycle, ewcFORCE);
1478 /* If there are multiple fshift output buffers we need to reduce them */
1479 if (stepWork.computeVirial)
1481 /* This is not in a subcounter because it takes a
1482 negligible and constant-sized amount of time */
1483 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1484 forceOut.forceWithShiftForces().shiftForces());
1488 /* update QMMMrec, if necessary */
1491 update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1494 // TODO Force flags should include haveFreeEnergyWork for this domain
1495 if (ddUsesGpuDirectCommunication && (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1497 /* Wait for non-local coordinate data to be copied from device */
1498 stateGpu->waitCoordinatesReadyOnHost(AtomLocality::NonLocal);
1500 /* Compute the bonded and non-bonded energies and optionally forces */
1501 do_force_lowlevel(fr, inputrec, &(top->idef), cr, ms, nrnb, wcycle, mdatoms, x, hist, &forceOut, enerd,
1502 fcd, box, lambda.data(), graph, fr->mu_tot, stepWork, ddBalanceRegionHandler);
1504 wallcycle_stop(wcycle, ewcFORCE);
1506 computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation, imdSession, pull_work, step, t,
1507 wcycle, fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1508 stepWork, &forceOut.forceWithVirial(), enerd, ed, stepWork.doNeighborSearch);
1511 // Will store the amount of cycles spent waiting for the GPU that
1512 // will be later used in the DLB accounting.
1513 float cycles_wait_gpu = 0;
1514 if (useOrEmulateGpuNb)
1516 auto& forceWithShiftForces = forceOut.forceWithShiftForces();
1518 /* wait for non-local forces (or calculate in emulation mode) */
1519 if (havePPDomainDecomposition(cr))
1521 if (simulationWork.useGpuNonbonded)
1523 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(
1524 nbv->gpu_nbv, stepWork, AtomLocality::NonLocal, enerd->grpp.ener[egLJSR].data(),
1525 enerd->grpp.ener[egCOULSR].data(), forceWithShiftForces.shiftForces(), wcycle);
1529 wallcycle_start_nocount(wcycle, ewcFORCE);
1530 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::NonLocal, enbvClearFYes,
1531 step, nrnb, wcycle);
1532 wallcycle_stop(wcycle, ewcFORCE);
1535 if (stepWork.useGpuFBufferOps)
1537 gmx::FixedCapacityVector<GpuEventSynchronizer*, 1> dependencyList;
1539 // TODO: move this into DomainLifetimeWorkload, including the second part of the
1540 // condition The bonded and free energy CPU tasks can have non-local force
1541 // contributions which are a dependency for the GPU force reduction.
1542 bool haveNonLocalForceContribInCpuBuffer =
1543 domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1545 if (haveNonLocalForceContribInCpuBuffer)
1547 stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(),
1548 AtomLocality::NonLocal);
1549 dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(
1550 AtomLocality::NonLocal, stepWork.useGpuFBufferOps));
1553 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::NonLocal, stateGpu->getForces(),
1554 pme_gpu_get_device_f(fr->pmedata), dependencyList,
1555 false, haveNonLocalForceContribInCpuBuffer);
1556 if (!useGpuForcesHaloExchange)
1558 // copy from GPU input for dd_move_f()
1559 stateGpu->copyForcesFromGpu(forceOut.forceWithShiftForces().force(),
1560 AtomLocality::NonLocal);
1565 nbv->atomdata_add_nbat_f_to_f(AtomLocality::NonLocal, forceWithShiftForces.force());
1569 if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1571 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat, forceWithShiftForces.shiftForces());
1576 if (havePPDomainDecomposition(cr))
1578 /* We are done with the CPU compute.
1579 * We will now communicate the non-local forces.
1580 * If we use a GPU this will overlap with GPU work, so in that case
1581 * we do not close the DD force balancing region here.
1583 ddBalanceRegionHandler.closeAfterForceComputationCpu();
1585 if (stepWork.computeForces)
1588 if (useGpuForcesHaloExchange)
1590 if (domainWork.haveCpuLocalForceWork)
1592 stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), AtomLocality::Local);
1594 communicateGpuHaloForces(*cr, domainWork.haveCpuLocalForceWork);
1598 if (stepWork.useGpuFBufferOps)
1600 stateGpu->waitForcesReadyOnHost(AtomLocality::NonLocal);
1602 dd_move_f(cr->dd, &forceOut.forceWithShiftForces(), wcycle);
1607 // With both nonbonded and PME offloaded a GPU on the same rank, we use
1608 // an alternating wait/reduction scheme.
1609 bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded
1610 && !DOMAINDECOMP(cr) && !stepWork.useGpuFBufferOps);
1611 if (alternateGpuWait)
1613 alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &forceOut, enerd, stepWork, pmeFlags, wcycle);
1616 if (!alternateGpuWait && useGpuPmeOnThisRank)
1618 pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial(), enerd);
1621 /* Wait for local GPU NB outputs on the non-alternating wait path */
1622 if (!alternateGpuWait && simulationWork.useGpuNonbonded)
1624 /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1625 * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1626 * but even with a step of 0.1 ms the difference is less than 1%
1629 const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1630 const float waitCycles = Nbnxm::gpu_wait_finish_task(
1631 nbv->gpu_nbv, stepWork, AtomLocality::Local, enerd->grpp.ener[egLJSR].data(),
1632 enerd->grpp.ener[egCOULSR].data(), forceOut.forceWithShiftForces().shiftForces(), wcycle);
1634 if (ddBalanceRegionHandler.useBalancingRegion())
1636 DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1637 if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
1639 /* We measured few cycles, it could be that the kernel
1640 * and transfer finished earlier and there was no actual
1641 * wait time, only API call overhead.
1642 * Then the actual time could be anywhere between 0 and
1643 * cycles_wait_est. We will use half of cycles_wait_est.
1645 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1647 ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1651 if (fr->nbv->emulateGpu())
1653 // NOTE: emulation kernel is not included in the balancing region,
1654 // but emulation mode does not target performance anyway
1655 wallcycle_start_nocount(wcycle, ewcFORCE);
1656 do_nb_verlet(fr, ic, enerd, stepWork, InteractionLocality::Local,
1657 DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes, step, nrnb, wcycle);
1658 wallcycle_stop(wcycle, ewcFORCE);
1661 // If on GPU PME-PP comms or GPU update path, receive forces from PME before GPU buffer ops
1662 // TODO refactor this and unify with below default-path call to the same function
1663 if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME)
1664 && (simulationWork.useGpuPmePpCommunication || simulationWork.useGpuUpdate))
1666 /* In case of node-splitting, the PP nodes receive the long-range
1667 * forces, virial and energy from the PME nodes here.
1669 pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1670 simulationWork.useGpuPmePpCommunication,
1671 stepWork.useGpuPmeFReduction, wcycle);
1675 /* Do the nonbonded GPU (or emulation) force buffer reduction
1676 * on the non-alternating path. */
1677 if (useOrEmulateGpuNb && !alternateGpuWait)
1679 // TODO simplify the below conditionals. Pass buffer and sync pointers at init stage rather than here. Unify getter fns for sameGPU/otherGPU cases.
1681 stepWork.useGpuPmeFReduction
1682 ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_device_f(fr->pmedata)
1683 : // PME force buffer on same GPU
1684 fr->pmePpCommGpu->getGpuForceStagingPtr()) // buffer received from other GPU
1685 : nullptr; // PME reduction not active on GPU
1687 GpuEventSynchronizer* const pmeSynchronizer =
1688 stepWork.useGpuPmeFReduction
1689 ? (thisRankHasDuty(cr, DUTY_PME) ? pme_gpu_get_f_ready_synchronizer(fr->pmedata)
1690 : // PME force buffer on same GPU
1691 static_cast<GpuEventSynchronizer*>(
1692 fr->pmePpCommGpu->getForcesReadySynchronizer())) // buffer received from other GPU
1693 : nullptr; // PME reduction not active on GPU
1695 gmx::FixedCapacityVector<GpuEventSynchronizer*, 3> dependencyList;
1697 if (stepWork.useGpuPmeFReduction)
1699 dependencyList.push_back(pmeSynchronizer);
1702 gmx::ArrayRef<gmx::RVec> forceWithShift = forceOut.forceWithShiftForces().force();
1704 if (stepWork.useGpuFBufferOps)
1706 // Flag to specify whether the CPU force buffer has contributions to
1707 // local atoms. This depends on whether there are CPU-based force tasks
1708 // or when DD is active the halo exchange has resulted in contributions
1709 // from the non-local part.
1710 const bool haveLocalForceContribInCpuBuffer =
1711 (domainWork.haveCpuLocalForceWork || havePPDomainDecomposition(cr));
1713 // TODO: move these steps as early as possible:
1714 // - CPU f H2D should be as soon as all CPU-side forces are done
1715 // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1716 // before the next CPU task that consumes the forces: vsite spread or update)
1717 // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
1718 // of the halo exchange. In that case the copy is instead performed above, before the exchange.
1719 // These should be unified.
1720 if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
1722 // Note: AtomLocality::All is used for the non-DD case because, as in this
1723 // case copyForcesToGpu() uses a separate stream, it allows overlap of
1724 // CPU force H2D with GPU force tasks on all streams including those in the
1725 // local stream which would otherwise be implicit dependencies for the
1726 // transfer and would not overlap.
1727 auto locality = havePPDomainDecomposition(cr) ? AtomLocality::Local : AtomLocality::All;
1729 stateGpu->copyForcesToGpu(forceWithShift, locality);
1730 dependencyList.push_back(
1731 stateGpu->getForcesReadyOnDeviceEvent(locality, stepWork.useGpuFBufferOps));
1733 if (useGpuForcesHaloExchange)
1735 dependencyList.push_back(cr->dd->gpuHaloExchange[0]->getForcesReadyOnDeviceEvent());
1737 nbv->atomdata_add_nbat_f_to_f_gpu(AtomLocality::Local, stateGpu->getForces(), pmeForcePtr,
1738 dependencyList, stepWork.useGpuPmeFReduction,
1739 haveLocalForceContribInCpuBuffer);
1740 // Copy forces to host if they are needed for update or if virtual sites are enabled.
1741 // If there are vsites, we need to copy forces every step to spread vsite forces on host.
1742 // TODO: When the output flags will be included in step workload, this copy can be combined with the
1743 // copy call done in sim_utils(...) for the output.
1744 // NOTE: If there are virtual sites, the forces are modified on host after this D2H copy. Hence,
1745 // they should not be copied in do_md(...) for the output.
1746 if (!simulationWork.useGpuUpdate || vsite)
1748 stateGpu->copyForcesFromGpu(forceWithShift, AtomLocality::Local);
1749 stateGpu->waitForcesReadyOnHost(AtomLocality::Local);
1754 nbv->atomdata_add_nbat_f_to_f(AtomLocality::Local, forceWithShift);
1758 launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd, *runScheduleWork,
1759 useGpuPmeOnThisRank, step, wcycle);
1761 if (DOMAINDECOMP(cr))
1763 dd_force_flop_stop(cr->dd, nrnb);
1766 if (stepWork.computeForces)
1768 rvec* f = as_rvec_array(forceOut.forceWithShiftForces().force().data());
1770 /* If we have NoVirSum forces, but we do not calculate the virial,
1771 * we sum fr->f_novirsum=forceOut.f later.
1773 if (vsite && !(fr->haveDirectVirialContributions && !stepWork.computeVirial))
1775 rvec* fshift = as_rvec_array(forceOut.forceWithShiftForces().shiftForces().data());
1776 spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), f, fshift, FALSE, nullptr,
1777 nrnb, &top->idef, fr->pbcType, fr->bMolPBC, graph, box, cr, wcycle);
1780 if (stepWork.computeVirial)
1782 /* Calculation of the virial must be done after vsites! */
1783 calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()),
1784 forceOut.forceWithShiftForces(), vir_force, graph, box, nrnb, fr,
1789 // TODO refactor this and unify with above GPU PME-PP / GPU update path call to the same function
1790 if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME) && !simulationWork.useGpuPmePpCommunication
1791 && !simulationWork.useGpuUpdate)
1793 /* In case of node-splitting, the PP nodes receive the long-range
1794 * forces, virial and energy from the PME nodes here.
1796 pme_receive_force_ener(fr, cr, &forceOut.forceWithVirial(), enerd,
1797 simulationWork.useGpuPmePpCommunication, false, wcycle);
1800 if (stepWork.computeForces)
1802 post_process_forces(cr, step, nrnb, wcycle, top, box, as_rvec_array(x.unpaddedArrayRef().data()),
1803 &forceOut, vir_force, mdatoms, graph, fr, vsite, stepWork);
1806 if (stepWork.computeEnergy)
1808 /* Sum the potential energy terms from group contributions */
1809 sum_epot(&(enerd->grpp), enerd->term);
1811 if (!EI_TPI(inputrec->eI))
1813 checkPotentialEnergyValidity(step, *enerd, *inputrec);
1817 /* In case we don't have constraints and are using GPUs, the next balancing
1818 * region starts here.
1819 * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1820 * virial calculation and COM pulling, is not thus not included in
1821 * the balance timing, which is ok as most tasks do communication.
1823 ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);