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) 2013,2014,2015,2016,2017,2018,2019, 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.
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/gmxlib/network.h"
57 #include "gromacs/gmxlib/nonbonded/nb_free_energy.h"
58 #include "gromacs/gmxlib/nonbonded/nb_kernel.h"
59 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
60 #include "gromacs/gpu_utils/gpu_utils.h"
61 #include "gromacs/imd/imd.h"
62 #include "gromacs/listed_forces/disre.h"
63 #include "gromacs/listed_forces/gpubonded.h"
64 #include "gromacs/listed_forces/listed_forces.h"
65 #include "gromacs/listed_forces/manage_threading.h"
66 #include "gromacs/listed_forces/orires.h"
67 #include "gromacs/math/arrayrefwithpadding.h"
68 #include "gromacs/math/functions.h"
69 #include "gromacs/math/units.h"
70 #include "gromacs/math/vec.h"
71 #include "gromacs/math/vecdump.h"
72 #include "gromacs/mdlib/calcmu.h"
73 #include "gromacs/mdlib/calcvir.h"
74 #include "gromacs/mdlib/constr.h"
75 #include "gromacs/mdlib/enerdata_utils.h"
76 #include "gromacs/mdlib/force.h"
77 #include "gromacs/mdlib/forcerec.h"
78 #include "gromacs/mdlib/gmx_omp_nthreads.h"
79 #include "gromacs/mdlib/qmmm.h"
80 #include "gromacs/mdlib/update.h"
81 #include "gromacs/mdtypes/commrec.h"
82 #include "gromacs/mdtypes/enerdata.h"
83 #include "gromacs/mdtypes/forceoutput.h"
84 #include "gromacs/mdtypes/iforceprovider.h"
85 #include "gromacs/mdtypes/inputrec.h"
86 #include "gromacs/mdtypes/md_enums.h"
87 #include "gromacs/mdtypes/simulation_workload.h"
88 #include "gromacs/mdtypes/state.h"
89 #include "gromacs/mdtypes/state_propagator_data_gpu.h"
90 #include "gromacs/nbnxm/gpu_data_mgmt.h"
91 #include "gromacs/nbnxm/nbnxm.h"
92 #include "gromacs/pbcutil/ishift.h"
93 #include "gromacs/pbcutil/mshift.h"
94 #include "gromacs/pbcutil/pbc.h"
95 #include "gromacs/pulling/pull.h"
96 #include "gromacs/pulling/pull_rotation.h"
97 #include "gromacs/timing/cyclecounter.h"
98 #include "gromacs/timing/gpu_timing.h"
99 #include "gromacs/timing/wallcycle.h"
100 #include "gromacs/timing/wallcyclereporting.h"
101 #include "gromacs/timing/walltime_accounting.h"
102 #include "gromacs/topology/topology.h"
103 #include "gromacs/utility/arrayref.h"
104 #include "gromacs/utility/basedefinitions.h"
105 #include "gromacs/utility/cstringutil.h"
106 #include "gromacs/utility/exceptions.h"
107 #include "gromacs/utility/fatalerror.h"
108 #include "gromacs/utility/fixedcapacityvector.h"
109 #include "gromacs/utility/gmxassert.h"
110 #include "gromacs/utility/gmxmpi.h"
111 #include "gromacs/utility/logger.h"
112 #include "gromacs/utility/smalloc.h"
113 #include "gromacs/utility/strconvert.h"
114 #include "gromacs/utility/sysinfo.h"
116 using gmx::ForceOutputs;
117 using gmx::StepWorkload;
118 using gmx::DomainLifetimeWorkload;
119 using gmx::SimulationWorkload;
121 // TODO: this environment variable allows us to verify before release
122 // that on less common architectures the total cost of polling is not larger than
123 // a blocking wait (so polling does not introduce overhead when the static
124 // PME-first ordering would suffice).
125 static const bool c_disableAlternatingWait = (getenv("GMX_DISABLE_ALTERNATING_GPU_WAIT") != nullptr);
127 static void sum_forces(rvec f[], gmx::ArrayRef<const gmx::RVec> forceToAdd)
129 const int end = forceToAdd.size();
131 int gmx_unused nt = gmx_omp_nthreads_get(emntDefault);
132 #pragma omp parallel for num_threads(nt) schedule(static)
133 for (int i = 0; i < end; i++)
135 rvec_inc(f[i], forceToAdd[i]);
139 static void calc_virial(int start, int homenr, const rvec x[],
140 const gmx::ForceWithShiftForces &forceWithShiftForces,
141 tensor vir_part, const t_graph *graph, const matrix box,
142 t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
144 /* The short-range virial from surrounding boxes */
145 const rvec *fshift = as_rvec_array(forceWithShiftForces.shiftForces().data());
146 calc_vir(SHIFTS, fr->shift_vec, fshift, vir_part, ePBC == epbcSCREW, box);
147 inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
149 /* Calculate partial virial, for local atoms only, based on short range.
150 * Total virial is computed in global_stat, called from do_md
152 const rvec *f = as_rvec_array(forceWithShiftForces.force().data());
153 f_calc_vir(start, start+homenr, x, f, vir_part, graph, box);
154 inc_nrnb(nrnb, eNR_VIRIAL, homenr);
158 pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
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,
171 gmx_wallcycle_t wcycle)
176 /* Calculate the center of mass forces, this requires communication,
177 * which is why pull_potential is called close to other communication.
179 wallcycle_start(wcycle, ewcPULLPOT);
180 set_pbc(&pbc, ir->ePBC, box);
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);
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)
194 real e_q, e_lj, dvdl_q, dvdl_lj;
195 float cycles_ppdpme, cycles_seppme;
197 cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
198 dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
200 /* In case of node-splitting, the PP nodes receive the long-range
201 * forces, virial and energy from the PME nodes here.
203 wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
206 gmx_pme_receive_f(cr, forceWithVirial, &e_q, &e_lj, &dvdl_q, &dvdl_lj,
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;
215 dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
217 wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
220 static void print_large_forces(FILE *fp,
228 real force2Tolerance = gmx::square(forceTolerance);
229 gmx::index numNonFinite = 0;
230 for (int i = 0; i < md->homenr; i++)
232 real force2 = norm2(f[i]);
233 bool nonFinite = !std::isfinite(force2);
234 if (force2 >= force2Tolerance || nonFinite)
236 fprintf(fp, "step %" PRId64 " atom %6d x %8.3f %8.3f %8.3f force %12.5e\n",
238 ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], std::sqrt(force2));
245 if (numNonFinite > 0)
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.
251 gmx_fatal(FARGS, "At step %" PRId64 " detected non-finite forces on %td atoms", step, numNonFinite);
255 static void post_process_forces(const t_commrec *cr,
258 gmx_wallcycle_t wcycle,
259 const gmx_localtop_t *top,
262 ForceOutputs *forceOutputs,
264 const t_mdatoms *mdatoms,
265 const t_graph *graph,
266 const t_forcerec *fr,
267 const gmx_vsite_t *vsite,
268 const StepWorkload &stepWork)
270 rvec *f = as_rvec_array(forceOutputs->forceWithShiftForces().force().data());
272 if (fr->haveDirectVirialContributions)
274 auto &forceWithVirial = forceOutputs->forceWithVirial();
275 rvec *fDirectVir = as_rvec_array(forceWithVirial.force_.data());
279 /* Spread the mesh force on virtual sites to the other particles...
280 * This is parallellized. MPI communication is performed
281 * if the constructing atoms aren't local.
283 matrix virial = { { 0 } };
284 spread_vsite_f(vsite, x, fDirectVir, nullptr,
285 stepWork.computeVirial, virial,
287 &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
288 forceWithVirial.addVirialContribution(virial);
291 if (stepWork.computeVirial)
293 /* Now add the forces, this is local */
294 sum_forces(f, forceWithVirial.force_);
296 /* Add the direct virial contributions */
297 GMX_ASSERT(forceWithVirial.computeVirial_, "forceWithVirial should request virial computation when we request the virial");
298 m_add(vir_force, forceWithVirial.getVirial(), vir_force);
302 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
307 if (fr->print_force >= 0)
309 print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
313 static void do_nb_verlet(t_forcerec *fr,
314 const interaction_const_t *ic,
315 gmx_enerdata_t *enerd,
316 const StepWorkload &stepWork,
317 const Nbnxm::InteractionLocality ilocality,
321 gmx_wallcycle_t wcycle)
323 if (!stepWork.computeNonbondedForces)
325 /* skip non-bonded calculation */
329 nonbonded_verlet_t *nbv = fr->nbv.get();
331 /* GPU kernel launch overhead is already timed separately */
332 if (fr->cutoff_scheme != ecutsVERLET)
334 gmx_incons("Invalid cut-off scheme passed!");
339 /* When dynamic pair-list pruning is requested, we need to prune
340 * at nstlistPrune steps.
342 if (nbv->isDynamicPruningStepCpu(step))
344 /* Prune the pair-list beyond fr->ic->rlistPrune using
345 * the current coordinates of the atoms.
347 wallcycle_sub_start(wcycle, ewcsNONBONDED_PRUNING);
348 nbv->dispatchPruneKernelCpu(ilocality, fr->shift_vec);
349 wallcycle_sub_stop(wcycle, ewcsNONBONDED_PRUNING);
353 nbv->dispatchNonbondedKernel(ilocality, *ic, stepWork, clearF, *fr, enerd, nrnb);
356 static inline void clear_rvecs_omp(int n, rvec v[])
358 int nth = gmx_omp_nthreads_get_simple_rvec_task(emntDefault, n);
360 /* Note that we would like to avoid this conditional by putting it
361 * into the omp pragma instead, but then we still take the full
362 * omp parallel for overhead (at least with gcc5).
366 for (int i = 0; i < n; i++)
373 #pragma omp parallel for num_threads(nth) schedule(static)
374 for (int i = 0; i < n; i++)
381 /*! \brief Return an estimate of the average kinetic energy or 0 when unreliable
383 * \param groupOptions Group options, containing T-coupling options
385 static real averageKineticEnergyEstimate(const t_grpopts &groupOptions)
387 real nrdfCoupled = 0;
388 real nrdfUncoupled = 0;
389 real kineticEnergy = 0;
390 for (int g = 0; g < groupOptions.ngtc; g++)
392 if (groupOptions.tau_t[g] >= 0)
394 nrdfCoupled += groupOptions.nrdf[g];
395 kineticEnergy += groupOptions.nrdf[g]*0.5*groupOptions.ref_t[g]*BOLTZ;
399 nrdfUncoupled += groupOptions.nrdf[g];
403 /* This conditional with > also catches nrdf=0 */
404 if (nrdfCoupled > nrdfUncoupled)
406 return kineticEnergy*(nrdfCoupled + nrdfUncoupled)/nrdfCoupled;
414 /*! \brief This routine checks that the potential energy is finite.
416 * Always checks that the potential energy is finite. If step equals
417 * inputrec.init_step also checks that the magnitude of the potential energy
418 * is reasonable. Terminates with a fatal error when a check fails.
419 * Note that passing this check does not guarantee finite forces,
420 * since those use slightly different arithmetics. But in most cases
421 * there is just a narrow coordinate range where forces are not finite
422 * and energies are finite.
424 * \param[in] step The step number, used for checking and printing
425 * \param[in] enerd The energy data; the non-bonded group energies need to be added to enerd.term[F_EPOT] before calling this routine
426 * \param[in] inputrec The input record
428 static void checkPotentialEnergyValidity(int64_t step,
429 const gmx_enerdata_t &enerd,
430 const t_inputrec &inputrec)
432 /* Threshold valid for comparing absolute potential energy against
433 * the kinetic energy. Normally one should not consider absolute
434 * potential energy values, but with a factor of one million
435 * we should never get false positives.
437 constexpr real c_thresholdFactor = 1e6;
439 bool energyIsNotFinite = !std::isfinite(enerd.term[F_EPOT]);
440 real averageKineticEnergy = 0;
441 /* We only check for large potential energy at the initial step,
442 * because that is by far the most likely step for this too occur
443 * and because computing the average kinetic energy is not free.
444 * Note: nstcalcenergy >> 1 often does not allow to catch large energies
445 * before they become NaN.
447 if (step == inputrec.init_step && EI_DYNAMICS(inputrec.eI))
449 averageKineticEnergy = averageKineticEnergyEstimate(inputrec.opts);
452 if (energyIsNotFinite || (averageKineticEnergy > 0 &&
453 enerd.term[F_EPOT] > c_thresholdFactor*averageKineticEnergy))
455 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.",
458 energyIsNotFinite ? "not finite" : "extremely high",
460 enerd.term[F_COUL_SR],
461 energyIsNotFinite ? "non-finite" : "very high",
462 energyIsNotFinite ? " or Nan" : "");
466 /*! \brief Return true if there are special forces computed this step.
468 * The conditionals exactly correspond to those in computeSpecialForces().
471 haveSpecialForces(const t_inputrec &inputrec,
472 const gmx::ForceProviders &forceProviders,
473 const pull_t *pull_work,
474 const bool computeForces,
479 ((computeForces && forceProviders.hasForceProvider()) || // forceProviders
480 (inputrec.bPull && pull_have_potential(pull_work)) || // pull
481 inputrec.bRot || // enforced rotation
482 (ed != nullptr) || // flooding
483 (inputrec.bIMD && computeForces)); // IMD
486 /*! \brief Compute forces and/or energies for special algorithms
488 * The intention is to collect all calls to algorithms that compute
489 * forces on local atoms only and that do not contribute to the local
490 * virial sum (but add their virial contribution separately).
491 * Eventually these should likely all become ForceProviders.
492 * Within this function the intention is to have algorithms that do
493 * global communication at the end, so global barriers within the MD loop
494 * are as close together as possible.
496 * \param[in] fplog The log file
497 * \param[in] cr The communication record
498 * \param[in] inputrec The input record
499 * \param[in] awh The Awh module (nullptr if none in use).
500 * \param[in] enforcedRotation Enforced rotation module.
501 * \param[in] imdSession The IMD session
502 * \param[in] pull_work The pull work structure.
503 * \param[in] step The current MD step
504 * \param[in] t The current time
505 * \param[in,out] wcycle Wallcycle accounting struct
506 * \param[in,out] forceProviders Pointer to a list of force providers
507 * \param[in] box The unit cell
508 * \param[in] x The coordinates
509 * \param[in] mdatoms Per atom properties
510 * \param[in] lambda Array of free-energy lambda values
511 * \param[in] stepWork Step schedule flags
512 * \param[in,out] forceWithVirial Force and virial buffers
513 * \param[in,out] enerd Energy buffer
514 * \param[in,out] ed Essential dynamics pointer
515 * \param[in] didNeighborSearch Tells if we did neighbor searching this step, used for ED sampling
517 * \todo Remove didNeighborSearch, which is used incorrectly.
518 * \todo Convert all other algorithms called here to ForceProviders.
521 computeSpecialForces(FILE *fplog,
523 const t_inputrec *inputrec,
525 gmx_enfrot *enforcedRotation,
526 gmx::ImdSession *imdSession,
530 gmx_wallcycle_t wcycle,
531 gmx::ForceProviders *forceProviders,
533 gmx::ArrayRef<const gmx::RVec> x,
534 const t_mdatoms *mdatoms,
536 const StepWorkload &stepWork,
537 gmx::ForceWithVirial *forceWithVirial,
538 gmx_enerdata_t *enerd,
540 bool didNeighborSearch)
542 /* NOTE: Currently all ForceProviders only provide forces.
543 * When they also provide energies, remove this conditional.
545 if (stepWork.computeForces)
547 gmx::ForceProviderInput forceProviderInput(x, *mdatoms, t, box, *cr);
548 gmx::ForceProviderOutput forceProviderOutput(forceWithVirial, enerd);
550 /* Collect forces from modules */
551 forceProviders->calculateForces(forceProviderInput, &forceProviderOutput);
554 if (inputrec->bPull && pull_have_potential(pull_work))
556 pull_potential_wrapper(cr, inputrec, box, x,
558 mdatoms, enerd, pull_work, lambda, t,
563 enerd->term[F_COM_PULL] +=
564 awh->applyBiasForcesAndUpdateBias(inputrec->ePBC, *mdatoms, box,
566 t, step, wcycle, fplog);
570 rvec *f = as_rvec_array(forceWithVirial->force_.data());
572 /* Add the forces from enforced rotation potentials (if any) */
575 wallcycle_start(wcycle, ewcROTadd);
576 enerd->term[F_COM_PULL] += add_rot_forces(enforcedRotation, f, cr, step, t);
577 wallcycle_stop(wcycle, ewcROTadd);
582 /* Note that since init_edsam() is called after the initialization
583 * of forcerec, edsam doesn't request the noVirSum force buffer.
584 * Thus if no other algorithm (e.g. PME) requires it, the forces
585 * here will contribute to the virial.
587 do_flood(cr, inputrec, as_rvec_array(x.data()), f, ed, box, step, didNeighborSearch);
590 /* Add forces from interactive molecular dynamics (IMD), if any */
591 if (inputrec->bIMD && stepWork.computeForces)
593 imdSession->applyForces(f);
597 /*! \brief Makes PME flags from StepWorkload data.
599 * \param[in] stepWork Step schedule flags
602 static int makePmeFlags(const StepWorkload &stepWork)
604 return GMX_PME_SPREAD | GMX_PME_SOLVE |
605 (stepWork.computeVirial ? GMX_PME_CALC_ENER_VIR : 0) |
606 (stepWork.computeEnergy ? GMX_PME_CALC_ENER_VIR : 0) |
607 (stepWork.computeForces ? GMX_PME_CALC_F : 0);
610 /*! \brief Launch the prepare_step and spread stages of PME GPU.
612 * \param[in] pmedata The PME structure
613 * \param[in] box The box matrix
614 * \param[in] stepWork Step schedule flags
615 * \param[in] pmeFlags PME flags
616 * \param[in] wcycle The wallcycle structure
618 static inline void launchPmeGpuSpread(gmx_pme_t *pmedata,
620 const StepWorkload &stepWork,
622 gmx_wallcycle_t wcycle)
624 pme_gpu_prepare_computation(pmedata, stepWork.haveDynamicBox, box, wcycle, pmeFlags, stepWork.useGpuPmeFReduction);
625 pme_gpu_launch_spread(pmedata, wcycle);
628 /*! \brief Launch the FFT and gather stages of PME GPU
630 * This function only implements setting the output forces (no accumulation).
632 * \param[in] pmedata The PME structure
633 * \param[in] wcycle The wallcycle structure
635 static void launchPmeGpuFftAndGather(gmx_pme_t *pmedata,
636 gmx_wallcycle_t wcycle)
638 pme_gpu_launch_complex_transforms(pmedata, wcycle);
639 pme_gpu_launch_gather(pmedata, wcycle, PmeForceOutputHandling::Set);
643 * Polling wait for either of the PME or nonbonded GPU tasks.
645 * Instead of a static order in waiting for GPU tasks, this function
646 * polls checking which of the two tasks completes first, and does the
647 * associated force buffer reduction overlapped with the other task.
648 * By doing that, unlike static scheduling order, it can always overlap
649 * one of the reductions, regardless of the GPU task completion order.
651 * \param[in] nbv Nonbonded verlet structure
652 * \param[in,out] pmedata PME module data
653 * \param[in,out] forceOutputs Output buffer for the forces and virial
654 * \param[in,out] enerd Energy data structure results are reduced into
655 * \param[in] stepWork Step schedule flags
656 * \param[in] pmeFlags PME flags
657 * \param[in] wcycle The wallcycle structure
659 static void alternatePmeNbGpuWaitReduce(nonbonded_verlet_t *nbv,
661 gmx::ForceOutputs *forceOutputs,
662 gmx_enerdata_t *enerd,
663 const StepWorkload &stepWork,
665 gmx_wallcycle_t wcycle)
667 bool isPmeGpuDone = false;
668 bool isNbGpuDone = false;
672 gmx::ForceWithShiftForces &forceWithShiftForces = forceOutputs->forceWithShiftForces();
673 gmx::ForceWithVirial &forceWithVirial = forceOutputs->forceWithVirial();
675 gmx::ArrayRef<const gmx::RVec> pmeGpuForces;
677 while (!isPmeGpuDone || !isNbGpuDone)
681 GpuTaskCompletion completionType = (isNbGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
682 isPmeGpuDone = pme_gpu_try_finish_task(pmedata, pmeFlags, wcycle, &forceWithVirial, enerd, completionType);
687 GpuTaskCompletion completionType = (isPmeGpuDone) ? GpuTaskCompletion::Wait : GpuTaskCompletion::Check;
688 isNbGpuDone = Nbnxm::gpu_try_finish_task(nbv->gpu_nbv,
690 Nbnxm::AtomLocality::Local,
691 enerd->grpp.ener[egLJSR].data(),
692 enerd->grpp.ener[egCOULSR].data(),
693 forceWithShiftForces.shiftForces(), completionType, wcycle);
697 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local,
698 forceWithShiftForces.force());
704 /*! \brief Set up the different force buffers; also does clearing.
706 * \param[in] fr force record pointer
707 * \param[in] pull_work The pull work object.
708 * \param[in] inputrec input record
709 * \param[in] force force array
710 * \param[in] stepWork Step schedule flags
711 * \param[out] wcycle wallcycle recording structure
713 * \returns Cleared force output structure
716 setupForceOutputs(t_forcerec *fr,
718 const t_inputrec &inputrec,
719 gmx::ArrayRefWithPadding<gmx::RVec> force,
720 const StepWorkload &stepWork,
721 gmx_wallcycle_t wcycle)
723 wallcycle_sub_start(wcycle, ewcsCLEAR_FORCE_BUFFER);
725 /* NOTE: We assume fr->shiftForces is all zeros here */
726 gmx::ForceWithShiftForces forceWithShiftForces(force, stepWork.computeVirial, fr->shiftForces);
728 if (stepWork.computeForces)
730 /* Clear the short- and long-range forces */
731 clear_rvecs_omp(fr->natoms_force_constr,
732 as_rvec_array(forceWithShiftForces.force().data()));
735 /* If we need to compute the virial, we might need a separate
736 * force buffer for algorithms for which the virial is calculated
737 * directly, such as PME. Otherwise, forceWithVirial uses the
738 * the same force (f in legacy calls) buffer as other algorithms.
740 const bool useSeparateForceWithVirialBuffer = (stepWork.computeForces &&
741 (stepWork.computeVirial && fr->haveDirectVirialContributions));
742 /* forceWithVirial uses the local atom range only */
743 gmx::ForceWithVirial forceWithVirial(useSeparateForceWithVirialBuffer ?
744 fr->forceBufferForDirectVirialContributions : force.unpaddedArrayRef(),
745 stepWork.computeVirial);
747 if (useSeparateForceWithVirialBuffer)
749 /* TODO: update comment
750 * We only compute forces on local atoms. Note that vsites can
751 * spread to non-local atoms, but that part of the buffer is
752 * cleared separately in the vsite spreading code.
754 clear_rvecs_omp(forceWithVirial.force_.size(), as_rvec_array(forceWithVirial.force_.data()));
757 if (inputrec.bPull && pull_have_constraint(pull_work))
759 clear_pull_forces(pull_work);
762 wallcycle_sub_stop(wcycle, ewcsCLEAR_FORCE_BUFFER);
764 return ForceOutputs(forceWithShiftForces, forceWithVirial);
768 /*! \brief Set up flags that have the lifetime of the domain indicating what type of work is there to compute.
770 static DomainLifetimeWorkload
771 setupDomainLifetimeWorkload(const t_inputrec &inputrec,
772 const t_forcerec &fr,
773 const pull_t *pull_work,
777 const t_mdatoms &mdatoms,
778 const StepWorkload &stepWork)
780 DomainLifetimeWorkload domainWork;
781 // Note that haveSpecialForces is constant over the whole run
782 domainWork.haveSpecialForces = haveSpecialForces(inputrec, *fr.forceProviders, pull_work, stepWork.computeForces, ed);
783 domainWork.haveCpuBondedWork = haveCpuBondeds(fr);
784 domainWork.haveGpuBondedWork = ((fr.gpuBonded != nullptr) && fr.gpuBonded->haveInteractions());
785 domainWork.haveRestraintsWork = havePositionRestraints(idef, fcd);
786 domainWork.haveCpuListedForceWork = haveCpuListedForces(fr, idef, fcd);
787 // Note that haveFreeEnergyWork is constant over the whole run
788 domainWork.haveFreeEnergyWork = (fr.efep != efepNO && mdatoms.nPerturbed != 0);
792 /*! \brief Set up force flag stuct from the force bitmask.
794 * \param[in] legacyFlags Force bitmask flags used to construct the new flags
795 * \param[in] isNonbondedOn Global override, if false forces to turn off all nonbonded calculation.
796 * \param[in] simulationWork Simulation workload description.
797 * \param[in] rankHasPmeDuty If this rank computes PME.
799 * \returns New Stepworkload description.
802 setupStepWorkload(const int legacyFlags,
803 const bool isNonbondedOn,
804 const SimulationWorkload &simulationWork,
805 const bool rankHasPmeDuty)
808 flags.stateChanged = ((legacyFlags & GMX_FORCE_STATECHANGED) != 0);
809 flags.haveDynamicBox = ((legacyFlags & GMX_FORCE_DYNAMICBOX) != 0);
810 flags.doNeighborSearch = ((legacyFlags & GMX_FORCE_NS) != 0);
811 flags.computeVirial = ((legacyFlags & GMX_FORCE_VIRIAL) != 0);
812 flags.computeEnergy = ((legacyFlags & GMX_FORCE_ENERGY) != 0);
813 flags.computeForces = ((legacyFlags & GMX_FORCE_FORCES) != 0);
814 flags.computeListedForces = ((legacyFlags & GMX_FORCE_LISTED) != 0);
815 flags.computeNonbondedForces = ((legacyFlags & GMX_FORCE_NONBONDED) != 0) && isNonbondedOn;
816 flags.computeDhdl = ((legacyFlags & GMX_FORCE_DHDL) != 0);
818 if (simulationWork.useGpuBufferOps)
820 GMX_ASSERT(simulationWork.useGpuNonbonded, "Can only offload buffer ops if nonbonded computation is also offloaded");
822 flags.useGpuXBufferOps = simulationWork.useGpuBufferOps;
823 // on virial steps the CPU reduction path is taken
824 // TODO: remove flags.computeEnergy, ref #3128
825 flags.useGpuFBufferOps = simulationWork.useGpuBufferOps && !(flags.computeVirial || flags.computeEnergy);
826 flags.useGpuPmeFReduction = flags.useGpuFBufferOps && (simulationWork.usePmeGpu && rankHasPmeDuty);
832 /* \brief Launch end-of-step GPU tasks: buffer clearing and rolling pruning.
834 * TODO: eliminate the \p useGpuNonbonded and \p useGpuNonbonded when these are
835 * incorporated in DomainLifetimeWorkload.
838 launchGpuEndOfStepTasks(nonbonded_verlet_t *nbv,
839 gmx::GpuBonded *gpuBonded,
841 gmx_enerdata_t *enerd,
842 const gmx::MdrunScheduleWorkload &runScheduleWork,
843 bool useGpuNonbonded,
846 gmx_wallcycle_t wcycle)
850 /* Launch pruning before buffer clearing because the API overhead of the
851 * clear kernel launches can leave the GPU idle while it could be running
854 if (nbv->isDynamicPruningStepGpu(step))
856 nbv->dispatchPruneKernelGpu(step);
859 /* now clear the GPU outputs while we finish the step on the CPU */
860 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
861 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
862 Nbnxm::gpu_clear_outputs(nbv->gpu_nbv, runScheduleWork.stepWork.computeVirial);
863 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
864 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
869 pme_gpu_reinit_computation(pmedata, wcycle);
872 if (runScheduleWork.domainWork.haveGpuBondedWork && runScheduleWork.stepWork.computeEnergy)
874 // in principle this should be included in the DD balancing region,
875 // but generally it is infrequent so we'll omit it for the sake of
877 gpuBonded->waitAccumulateEnergyTerms(enerd);
879 gpuBonded->clearEnergies();
884 void do_force(FILE *fplog,
886 const gmx_multisim_t *ms,
887 const t_inputrec *inputrec,
889 gmx_enfrot *enforcedRotation,
890 gmx::ImdSession *imdSession,
894 gmx_wallcycle_t wcycle,
895 const gmx_localtop_t *top,
897 gmx::ArrayRefWithPadding<gmx::RVec> x,
899 gmx::ArrayRefWithPadding<gmx::RVec> force,
901 const t_mdatoms *mdatoms,
902 gmx_enerdata_t *enerd,
904 gmx::ArrayRef<real> lambda,
907 gmx::MdrunScheduleWorkload *runScheduleWork,
908 const gmx_vsite_t *vsite,
913 const DDBalanceRegionHandler &ddBalanceRegionHandler)
917 nonbonded_verlet_t *nbv = fr->nbv.get();
918 interaction_const_t *ic = fr->ic;
919 gmx::StatePropagatorDataGpu *stateGpu = fr->stateGpu;
921 // TODO remove the code below when the legacy flags are not in use anymore
922 /* modify force flag if not doing nonbonded */
925 legacyFlags &= ~GMX_FORCE_NONBONDED;
928 const SimulationWorkload &simulationWork = runScheduleWork->simulationWork;
931 runScheduleWork->stepWork = setupStepWorkload(legacyFlags, fr->bNonbonded,
932 simulationWork, thisRankHasDuty(cr, DUTY_PME));
933 const StepWorkload &stepWork = runScheduleWork->stepWork;
936 const bool useGpuPmeOnThisRank = simulationWork.usePmeGpu && thisRankHasDuty(cr, DUTY_PME);
937 const int pmeFlags = makePmeFlags(stepWork);
939 // Switches on whether to use GPU for position and force buffer operations
940 // TODO consider all possible combinations of triggers, and how to combine optimally in each case.
941 const BufferOpsUseGpu useGpuXBufOps = stepWork.useGpuXBufferOps ? BufferOpsUseGpu::True : BufferOpsUseGpu::False;
942 // GPU Force buffer ops are disabled on virial steps, because the virial calc is not yet ported to GPU
943 const BufferOpsUseGpu useGpuFBufOps = stepWork.useGpuFBufferOps ? BufferOpsUseGpu::True : BufferOpsUseGpu::False;
945 /* At a search step we need to start the first balancing region
946 * somewhere early inside the step after communication during domain
947 * decomposition (and not during the previous step as usual).
949 if (stepWork.doNeighborSearch)
951 ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::yes);
955 const int homenr = mdatoms->homenr;
957 clear_mat(vir_force);
959 if (stepWork.stateChanged)
961 if (inputrecNeedMutot(inputrec))
963 /* Calculate total (local) dipole moment in a temporary common array.
964 * This makes it possible to sum them over nodes faster.
966 calc_mu(start, homenr,
967 x.unpaddedArrayRef(), mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
972 if (fr->ePBC != epbcNONE)
974 /* Compute shift vectors every step,
975 * because of pressure coupling or box deformation!
977 if (stepWork.haveDynamicBox && stepWork.stateChanged)
979 calc_shifts(box, fr->shift_vec);
982 const bool fillGrid = (stepWork.doNeighborSearch && stepWork.stateChanged);
983 const bool calcCGCM = (fillGrid && !DOMAINDECOMP(cr));
986 put_atoms_in_box_omp(fr->ePBC, box, x.unpaddedArrayRef().subArray(0, homenr), gmx_omp_nthreads_get(emntDefault));
987 inc_nrnb(nrnb, eNR_SHIFTX, homenr);
989 else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
991 unshift_self(graph, box, as_rvec_array(x.unpaddedArrayRef().data()));
995 nbnxn_atomdata_copy_shiftvec(stepWork.haveDynamicBox,
996 fr->shift_vec, nbv->nbat.get());
999 if (!thisRankHasDuty(cr, DUTY_PME))
1001 /* Send particle coordinates to the pme nodes.
1002 * Since this is only implemented for domain decomposition
1003 * and domain decomposition does not use the graph,
1004 * we do not need to worry about shifting.
1006 gmx_pme_send_coordinates(cr, box, as_rvec_array(x.unpaddedArrayRef().data()),
1007 lambda[efptCOUL], lambda[efptVDW],
1008 (stepWork.computeVirial || stepWork.computeEnergy),
1011 #endif /* GMX_MPI */
1013 // Coordinates on the device are needed if PME or BufferOps are offloaded.
1014 // The local coordinates can be copied right away.
1015 // NOTE: Consider moving this copy to right after they are updated and constrained,
1016 // if the later is not offloaded.
1017 if (useGpuPmeOnThisRank || useGpuXBufOps == BufferOpsUseGpu::True)
1019 if (stepWork.doNeighborSearch)
1021 stateGpu->reinit(mdatoms->homenr, cr->dd != nullptr ? dd_numAtomsZones(*cr->dd) : mdatoms->homenr);
1022 if (useGpuPmeOnThisRank)
1024 // TODO: This should be moved into PME setup function ( pme_gpu_prepare_computation(...) )
1025 pme_gpu_set_device_x(fr->pmedata, stateGpu->getCoordinates());
1028 stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), gmx::StatePropagatorDataGpu::AtomLocality::Local);
1031 if (useGpuPmeOnThisRank)
1033 launchPmeGpuSpread(fr->pmedata, box, stepWork, pmeFlags, wcycle);
1036 /* do gridding for pair search */
1037 if (stepWork.doNeighborSearch)
1039 if (graph && stepWork.stateChanged)
1041 /* Calculate intramolecular shift vectors to make molecules whole */
1042 mk_mshift(fplog, graph, fr->ePBC, box, as_rvec_array(x.unpaddedArrayRef().data()));
1046 // - vzero is constant, do we need to pass it?
1047 // - box_diag should be passed directly to nbnxn_put_on_grid
1053 box_diag[XX] = box[XX][XX];
1054 box_diag[YY] = box[YY][YY];
1055 box_diag[ZZ] = box[ZZ][ZZ];
1057 wallcycle_start(wcycle, ewcNS);
1058 if (!DOMAINDECOMP(cr))
1060 wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
1061 nbnxn_put_on_grid(nbv, box,
1063 nullptr, 0, mdatoms->homenr, -1,
1064 fr->cginfo, x.unpaddedArrayRef(),
1066 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
1070 wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
1071 nbnxn_put_on_grid_nonlocal(nbv, domdec_zones(cr->dd),
1072 fr->cginfo, x.unpaddedArrayRef());
1073 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
1076 nbv->setAtomProperties(*mdatoms, fr->cginfo);
1078 wallcycle_stop(wcycle, ewcNS);
1080 /* initialize the GPU nbnxm atom data and bonded data structures */
1081 if (simulationWork.useGpuNonbonded)
1083 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1085 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1086 Nbnxm::gpu_init_atomdata(nbv->gpu_nbv, nbv->nbat.get());
1087 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1091 /* Now we put all atoms on the grid, we can assign bonded
1092 * interactions to the GPU, where the grid order is
1093 * needed. Also the xq, f and fshift device buffers have
1094 * been reallocated if needed, so the bonded code can
1095 * learn about them. */
1096 // TODO the xq, f, and fshift buffers are now shared
1097 // resources, so they should be maintained by a
1098 // higher-level object than the nb module.
1099 fr->gpuBonded->updateInteractionListsAndDeviceBuffers(nbv->getGridIndices(),
1101 Nbnxm::gpu_get_xq(nbv->gpu_nbv),
1102 Nbnxm::gpu_get_f(nbv->gpu_nbv),
1103 Nbnxm::gpu_get_fshift(nbv->gpu_nbv));
1105 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1109 if (stepWork.doNeighborSearch)
1111 // Need to run after the GPU-offload bonded interaction lists
1112 // are set up to be able to determine whether there is bonded work.
1113 runScheduleWork->domainWork =
1114 setupDomainLifetimeWorkload(*inputrec,
1123 wallcycle_start_nocount(wcycle, ewcNS);
1124 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
1125 /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1126 nbv->constructPairlist(Nbnxm::InteractionLocality::Local,
1127 &top->excls, step, nrnb);
1129 nbv->setupGpuShortRangeWork(fr->gpuBonded, Nbnxm::InteractionLocality::Local);
1131 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
1132 wallcycle_stop(wcycle, ewcNS);
1134 if (useGpuXBufOps == BufferOpsUseGpu::True)
1136 nbv->atomdata_init_copy_x_to_nbat_x_gpu();
1138 // For force buffer ops, we use the below conditon rather than
1139 // useGpuFBufOps to ensure that init is performed even if this
1140 // NS step is also a virial step (on which f buf ops are deactivated).
1141 if (simulationWork.useGpuBufferOps && simulationWork.useGpuNonbonded && (GMX_GPU == GMX_GPU_CUDA))
1143 nbv->atomdata_init_add_nbat_f_to_f_gpu();
1146 else if (!EI_TPI(inputrec->eI))
1148 if (useGpuXBufOps == BufferOpsUseGpu::True)
1150 nbv->convertCoordinatesGpu(Nbnxm::AtomLocality::Local, false,
1151 stateGpu->getCoordinates());
1155 nbv->convertCoordinates(Nbnxm::AtomLocality::Local, false,
1156 x.unpaddedArrayRef());
1160 const gmx::DomainLifetimeWorkload &domainWork = runScheduleWork->domainWork;
1162 if (simulationWork.useGpuNonbonded)
1164 ddBalanceRegionHandler.openBeforeForceComputationGpu();
1166 wallcycle_start(wcycle, ewcLAUNCH_GPU);
1168 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1169 Nbnxm::gpu_upload_shiftvec(nbv->gpu_nbv, nbv->nbat.get());
1170 if (stepWork.doNeighborSearch || (useGpuXBufOps == BufferOpsUseGpu::False))
1172 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1173 Nbnxm::AtomLocality::Local);
1175 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1176 // with X buffer ops offloaded to the GPU on all but the search steps
1178 // bonded work not split into separate local and non-local, so with DD
1179 // we can only launch the kernel after non-local coordinates have been received.
1180 if (domainWork.haveGpuBondedWork && !havePPDomainDecomposition(cr))
1182 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1183 fr->gpuBonded->launchKernel(fr, stepWork, box);
1184 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1187 /* launch local nonbonded work on GPU */
1188 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1189 do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::Local, enbvClearFNo,
1190 step, nrnb, wcycle);
1191 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1192 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1195 if (useGpuPmeOnThisRank)
1197 // In PME GPU and mixed mode we launch FFT / gather after the
1198 // X copy/transform to allow overlap as well as after the GPU NB
1199 // launch to avoid FFT launch overhead hijacking the CPU and delaying
1200 // the nonbonded kernel.
1201 launchPmeGpuFftAndGather(fr->pmedata, wcycle);
1204 // TODO Update this comment when introducing SimulationWorkload
1206 // The conditions for gpuHaloExchange e.g. using GPU buffer
1207 // operations were checked before construction, so here we can
1208 // just use it and assert upon any conditions.
1209 gmx::GpuHaloExchange *gpuHaloExchange = (havePPDomainDecomposition(cr) ? cr->dd->gpuHaloExchange.get() : nullptr);
1210 const bool ddUsesGpuDirectCommunication = (gpuHaloExchange != nullptr);
1211 GMX_ASSERT(!ddUsesGpuDirectCommunication || (useGpuXBufOps == BufferOpsUseGpu::True),
1212 "Must use coordinate buffer ops with GPU halo exchange");
1214 /* Communicate coordinates and sum dipole if necessary +
1215 do non-local pair search */
1216 if (havePPDomainDecomposition(cr))
1218 if (stepWork.doNeighborSearch)
1220 // TODO: fuse this branch with the above large stepWork.doNeighborSearch block
1221 wallcycle_start_nocount(wcycle, ewcNS);
1222 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1223 /* Note that with a GPU the launch overhead of the list transfer is not timed separately */
1224 nbv->constructPairlist(Nbnxm::InteractionLocality::NonLocal,
1225 &top->excls, step, nrnb);
1227 nbv->setupGpuShortRangeWork(fr->gpuBonded, Nbnxm::InteractionLocality::NonLocal);
1228 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1229 wallcycle_stop(wcycle, ewcNS);
1230 if (ddUsesGpuDirectCommunication)
1232 gpuHaloExchange->reinitHalo(stateGpu->getCoordinates(), stateGpu->getForces());
1237 if (ddUsesGpuDirectCommunication)
1239 // The following must be called after local setCoordinates (which records an event
1240 // when the coordinate data has been copied to the device).
1241 gpuHaloExchange->communicateHaloCoordinates(box);
1243 if (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork)
1245 //non-local part of coordinate buffer must be copied back to host for CPU work
1246 stateGpu->copyCoordinatesFromGpu(x.unpaddedArrayRef(), gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1251 dd_move_x(cr->dd, box, x.unpaddedArrayRef(), wcycle);
1254 if (useGpuXBufOps == BufferOpsUseGpu::True)
1256 // The condition here was (pme != nullptr && pme_gpu_get_device_x(fr->pmedata) != nullptr)
1257 if (!useGpuPmeOnThisRank && !ddUsesGpuDirectCommunication)
1259 stateGpu->copyCoordinatesToGpu(x.unpaddedArrayRef(), gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1261 nbv->convertCoordinatesGpu(Nbnxm::AtomLocality::NonLocal, false,
1262 stateGpu->getCoordinates());
1266 nbv->convertCoordinates(Nbnxm::AtomLocality::NonLocal, false,
1267 x.unpaddedArrayRef());
1272 if (simulationWork.useGpuNonbonded)
1274 wallcycle_start(wcycle, ewcLAUNCH_GPU);
1276 if (stepWork.doNeighborSearch || (useGpuXBufOps == BufferOpsUseGpu::False))
1278 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1279 Nbnxm::gpu_copy_xq_to_gpu(nbv->gpu_nbv, nbv->nbat.get(),
1280 Nbnxm::AtomLocality::NonLocal);
1281 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1284 if (domainWork.haveGpuBondedWork)
1286 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_BONDED);
1287 fr->gpuBonded->launchKernel(fr, stepWork, box);
1288 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_BONDED);
1291 /* launch non-local nonbonded tasks on GPU */
1292 wallcycle_sub_start(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1293 do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1294 step, nrnb, wcycle);
1295 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1297 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1301 if (simulationWork.useGpuNonbonded)
1303 /* launch D2H copy-back F */
1304 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU);
1305 wallcycle_sub_start_nocount(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1307 bool copyBackNbForce = (useGpuFBufOps == BufferOpsUseGpu::False);
1309 if (havePPDomainDecomposition(cr))
1311 Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1312 stepWork, Nbnxm::AtomLocality::NonLocal, copyBackNbForce);
1314 Nbnxm::gpu_launch_cpyback(nbv->gpu_nbv, nbv->nbat.get(),
1315 stepWork, Nbnxm::AtomLocality::Local, copyBackNbForce);
1316 wallcycle_sub_stop(wcycle, ewcsLAUNCH_GPU_NONBONDED);
1318 if (domainWork.haveGpuBondedWork && stepWork.computeEnergy)
1320 fr->gpuBonded->launchEnergyTransfer();
1322 wallcycle_stop(wcycle, ewcLAUNCH_GPU);
1325 if (stepWork.stateChanged && inputrecNeedMutot(inputrec))
1329 gmx_sumd(2*DIM, mu, cr);
1331 ddBalanceRegionHandler.reopenRegionCpu();
1334 for (i = 0; i < 2; i++)
1336 for (j = 0; j < DIM; j++)
1338 fr->mu_tot[i][j] = mu[i*DIM + j];
1342 if (mdatoms->nChargePerturbed == 0)
1344 copy_rvec(fr->mu_tot[0], mu_tot);
1348 for (j = 0; j < DIM; j++)
1351 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1352 lambda[efptCOUL]*fr->mu_tot[1][j];
1356 /* Reset energies */
1357 reset_enerdata(enerd);
1358 /* Clear the shift forces */
1359 // TODO: This should be linked to the shift force buffer in use, or cleared before use instead
1360 for (gmx::RVec &elem : fr->shiftForces)
1362 elem = { 0.0_real, 0.0_real, 0.0_real };
1365 if (DOMAINDECOMP(cr) && !thisRankHasDuty(cr, DUTY_PME))
1367 wallcycle_start(wcycle, ewcPPDURINGPME);
1368 dd_force_flop_start(cr->dd, nrnb);
1373 wallcycle_start(wcycle, ewcROT);
1374 do_rotation(cr, enforcedRotation, box, as_rvec_array(x.unpaddedArrayRef().data()), t, step, stepWork.doNeighborSearch);
1375 wallcycle_stop(wcycle, ewcROT);
1378 /* Start the force cycle counter.
1379 * Note that a different counter is used for dynamic load balancing.
1381 wallcycle_start(wcycle, ewcFORCE);
1383 // Set up and clear force outputs.
1384 // We use std::move to keep the compiler happy, it has no effect.
1385 ForceOutputs forceOut = setupForceOutputs(fr, pull_work, *inputrec, std::move(force), stepWork, wcycle);
1387 /* We calculate the non-bonded forces, when done on the CPU, here.
1388 * We do this before calling do_force_lowlevel, because in that
1389 * function, the listed forces are calculated before PME, which
1390 * does communication. With this order, non-bonded and listed
1391 * force calculation imbalance can be balanced out by the domain
1392 * decomposition load balancing.
1395 const bool useOrEmulateGpuNb = simulationWork.useGpuNonbonded || fr->nbv->emulateGpu();
1397 if (!useOrEmulateGpuNb)
1399 do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::Local, enbvClearFYes,
1400 step, nrnb, wcycle);
1403 if (fr->efep != efepNO)
1405 /* Calculate the local and non-local free energy interactions here.
1406 * Happens here on the CPU both with and without GPU.
1408 nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::Local,
1409 fr, as_rvec_array(x.unpaddedArrayRef().data()), &forceOut.forceWithShiftForces(), *mdatoms,
1410 inputrec->fepvals, lambda.data(),
1411 enerd, stepWork, nrnb);
1413 if (havePPDomainDecomposition(cr))
1415 nbv->dispatchFreeEnergyKernel(Nbnxm::InteractionLocality::NonLocal,
1416 fr, as_rvec_array(x.unpaddedArrayRef().data()), &forceOut.forceWithShiftForces(), *mdatoms,
1417 inputrec->fepvals, lambda.data(),
1418 enerd, stepWork, nrnb);
1422 if (!useOrEmulateGpuNb)
1424 if (havePPDomainDecomposition(cr))
1426 do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::NonLocal, enbvClearFNo,
1427 step, nrnb, wcycle);
1430 if (stepWork.computeForces)
1432 /* Add all the non-bonded force to the normal force array.
1433 * This can be split into a local and a non-local part when overlapping
1434 * communication with calculation with domain decomposition.
1436 wallcycle_stop(wcycle, ewcFORCE);
1437 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::All, forceOut.forceWithShiftForces().force());
1438 wallcycle_start_nocount(wcycle, ewcFORCE);
1441 /* If there are multiple fshift output buffers we need to reduce them */
1442 if (stepWork.computeVirial)
1444 /* This is not in a subcounter because it takes a
1445 negligible and constant-sized amount of time */
1446 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1447 forceOut.forceWithShiftForces().shiftForces());
1451 /* update QMMMrec, if necessary */
1454 update_QMMMrec(cr, fr, as_rvec_array(x.unpaddedArrayRef().data()), mdatoms, box);
1457 // TODO Force flags should include haveFreeEnergyWork for this domain
1458 if (ddUsesGpuDirectCommunication &&
1459 (domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork))
1461 /* Wait for non-local coordinate data to be copied from device */
1462 nbv->wait_nonlocal_x_copy_D2H_done();
1464 /* Compute the bonded and non-bonded energies and optionally forces */
1465 do_force_lowlevel(fr, inputrec, &(top->idef),
1466 cr, ms, nrnb, wcycle, mdatoms,
1467 x, hist, &forceOut, enerd, fcd,
1468 box, lambda.data(), graph, fr->mu_tot,
1470 ddBalanceRegionHandler);
1472 wallcycle_stop(wcycle, ewcFORCE);
1474 computeSpecialForces(fplog, cr, inputrec, awh, enforcedRotation,
1475 imdSession, pull_work, step, t, wcycle,
1476 fr->forceProviders, box, x.unpaddedArrayRef(), mdatoms, lambda.data(),
1477 stepWork, &forceOut.forceWithVirial(), enerd,
1478 ed, stepWork.doNeighborSearch);
1481 // Will store the amount of cycles spent waiting for the GPU that
1482 // will be later used in the DLB accounting.
1483 float cycles_wait_gpu = 0;
1484 if (useOrEmulateGpuNb)
1486 auto &forceWithShiftForces = forceOut.forceWithShiftForces();
1488 /* wait for non-local forces (or calculate in emulation mode) */
1489 if (havePPDomainDecomposition(cr))
1491 if (simulationWork.useGpuNonbonded)
1493 cycles_wait_gpu += Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1494 stepWork, Nbnxm::AtomLocality::NonLocal,
1495 enerd->grpp.ener[egLJSR].data(),
1496 enerd->grpp.ener[egCOULSR].data(),
1497 forceWithShiftForces.shiftForces(),
1502 wallcycle_start_nocount(wcycle, ewcFORCE);
1503 do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::NonLocal, enbvClearFYes,
1504 step, nrnb, wcycle);
1505 wallcycle_stop(wcycle, ewcFORCE);
1508 if (useGpuFBufOps == BufferOpsUseGpu::True)
1510 gmx::FixedCapacityVector<GpuEventSynchronizer*, 1> dependencyList;
1512 // TODO: move this into DomainLifetimeWorkload, including the second part of the condition
1513 // The bonded and free energy CPU tasks can have non-local force contributions
1514 // which are a dependency for the GPU force reduction.
1515 bool haveNonLocalForceContribInCpuBuffer = domainWork.haveCpuBondedWork || domainWork.haveFreeEnergyWork;
1517 if (haveNonLocalForceContribInCpuBuffer)
1519 stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1520 dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(gmx::StatePropagatorDataGpu::AtomLocality::NonLocal,
1521 useGpuFBufOps == BufferOpsUseGpu::True));
1524 nbv->atomdata_add_nbat_f_to_f_gpu(Nbnxm::AtomLocality::NonLocal,
1525 stateGpu->getForces(),
1526 pme_gpu_get_device_f(fr->pmedata),
1528 false, haveNonLocalForceContribInCpuBuffer);
1529 stateGpu->copyForcesFromGpu(forceOut.forceWithShiftForces().force(), gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1533 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::NonLocal,
1534 forceWithShiftForces.force());
1538 if (fr->nbv->emulateGpu() && stepWork.computeVirial)
1540 nbnxn_atomdata_add_nbat_fshift_to_fshift(*nbv->nbat,
1541 forceWithShiftForces.shiftForces());
1546 const bool useGpuForcesHaloExchange = ddUsesGpuDirectCommunication && (useGpuFBufOps == BufferOpsUseGpu::True);
1547 const bool useCpuPmeFReduction = thisRankHasDuty(cr, DUTY_PME) && !stepWork.useGpuPmeFReduction;
1548 // TODO: move this into DomainLifetimeWorkload, including the second part of the condition
1549 const bool haveCpuLocalForces = (domainWork.haveSpecialForces || domainWork.haveCpuListedForceWork || useCpuPmeFReduction ||
1550 (fr->efep != efepNO));
1552 if (havePPDomainDecomposition(cr))
1554 /* We are done with the CPU compute.
1555 * We will now communicate the non-local forces.
1556 * If we use a GPU this will overlap with GPU work, so in that case
1557 * we do not close the DD force balancing region here.
1559 ddBalanceRegionHandler.closeAfterForceComputationCpu();
1561 if (stepWork.computeForces)
1564 if (useGpuForcesHaloExchange)
1566 if (haveCpuLocalForces)
1568 stateGpu->copyForcesToGpu(forceOut.forceWithShiftForces().force(), gmx::StatePropagatorDataGpu::AtomLocality::Local);
1570 gpuHaloExchange->communicateHaloForces(haveCpuLocalForces);
1574 if (useGpuFBufOps == BufferOpsUseGpu::True)
1576 stateGpu->waitForcesReadyOnHost(gmx::StatePropagatorDataGpu::AtomLocality::NonLocal);
1578 dd_move_f(cr->dd, &forceOut.forceWithShiftForces(), wcycle);
1584 // With both nonbonded and PME offloaded a GPU on the same rank, we use
1585 // an alternating wait/reduction scheme.
1586 bool alternateGpuWait = (!c_disableAlternatingWait && useGpuPmeOnThisRank && simulationWork.useGpuNonbonded && !DOMAINDECOMP(cr) &&
1587 (useGpuFBufOps == BufferOpsUseGpu::False));
1588 if (alternateGpuWait)
1590 alternatePmeNbGpuWaitReduce(fr->nbv.get(), fr->pmedata, &forceOut, enerd,
1591 stepWork, pmeFlags, wcycle);
1594 if (!alternateGpuWait && useGpuPmeOnThisRank)
1596 pme_gpu_wait_and_reduce(fr->pmedata, pmeFlags, wcycle, &forceOut.forceWithVirial(), enerd);
1599 /* Wait for local GPU NB outputs on the non-alternating wait path */
1600 if (!alternateGpuWait && simulationWork.useGpuNonbonded)
1602 /* Measured overhead on CUDA and OpenCL with(out) GPU sharing
1603 * is between 0.5 and 1.5 Mcycles. So 2 MCycles is an overestimate,
1604 * but even with a step of 0.1 ms the difference is less than 1%
1607 const float gpuWaitApiOverheadMargin = 2e6F; /* cycles */
1608 const float waitCycles =
1609 Nbnxm::gpu_wait_finish_task(nbv->gpu_nbv,
1610 stepWork, Nbnxm::AtomLocality::Local,
1611 enerd->grpp.ener[egLJSR].data(),
1612 enerd->grpp.ener[egCOULSR].data(),
1613 forceOut.forceWithShiftForces().shiftForces(),
1616 if (ddBalanceRegionHandler.useBalancingRegion())
1618 DdBalanceRegionWaitedForGpu waitedForGpu = DdBalanceRegionWaitedForGpu::yes;
1619 if (stepWork.computeForces && waitCycles <= gpuWaitApiOverheadMargin)
1621 /* We measured few cycles, it could be that the kernel
1622 * and transfer finished earlier and there was no actual
1623 * wait time, only API call overhead.
1624 * Then the actual time could be anywhere between 0 and
1625 * cycles_wait_est. We will use half of cycles_wait_est.
1627 waitedForGpu = DdBalanceRegionWaitedForGpu::no;
1629 ddBalanceRegionHandler.closeAfterForceComputationGpu(cycles_wait_gpu, waitedForGpu);
1633 if (fr->nbv->emulateGpu())
1635 // NOTE: emulation kernel is not included in the balancing region,
1636 // but emulation mode does not target performance anyway
1637 wallcycle_start_nocount(wcycle, ewcFORCE);
1638 do_nb_verlet(fr, ic, enerd, stepWork, Nbnxm::InteractionLocality::Local,
1639 DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1640 step, nrnb, wcycle);
1641 wallcycle_stop(wcycle, ewcFORCE);
1644 /* Do the nonbonded GPU (or emulation) force buffer reduction
1645 * on the non-alternating path. */
1646 if (useOrEmulateGpuNb && !alternateGpuWait)
1648 gmx::FixedCapacityVector<GpuEventSynchronizer*, 2> dependencyList;
1650 if (stepWork.useGpuPmeFReduction)
1652 dependencyList.push_back(pme_gpu_get_f_ready_synchronizer(fr->pmedata));
1655 gmx::ArrayRef<gmx::RVec> forceWithShift = forceOut.forceWithShiftForces().force();
1657 if (useGpuFBufOps == BufferOpsUseGpu::True)
1659 // Flag to specify whether the CPU force buffer has contributions to
1660 // local atoms. This depends on whether there are CPU-based force tasks
1661 // or when DD is active the halo exchange has resulted in contributions
1662 // from the non-local part.
1663 const bool haveLocalForceContribInCpuBuffer = (haveCpuLocalForces || havePPDomainDecomposition(cr));
1665 // TODO: move these steps as early as possible:
1666 // - CPU f H2D should be as soon as all CPU-side forces are done
1667 // - wait for force reduction does not need to block host (at least not here, it's sufficient to wait
1668 // before the next CPU task that consumes the forces: vsite spread or update)
1669 // - copy is not perfomed if GPU force halo exchange is active, because it would overwrite the result
1670 // of the halo exchange. In that case the copy is instead performed above, before the exchange.
1671 // These should be unified.
1672 if (haveLocalForceContribInCpuBuffer && !useGpuForcesHaloExchange)
1674 stateGpu->copyForcesToGpu(forceWithShift, gmx::StatePropagatorDataGpu::AtomLocality::Local);
1675 dependencyList.push_back(stateGpu->getForcesReadyOnDeviceEvent(gmx::StatePropagatorDataGpu::AtomLocality::Local,
1676 useGpuFBufOps == BufferOpsUseGpu::True));
1678 if (useGpuForcesHaloExchange)
1680 // Add a stream synchronization to satisfy a dependency
1681 // for the local buffer ops on the result of GPU halo
1682 // exchange, which operates in the non-local stream and
1683 // writes to to local parf og the force buffer.
1685 // TODO improve this through use of an event - see Redmine #3093
1686 // push the event into the dependencyList
1687 nbv->stream_local_wait_for_nonlocal();
1689 nbv->atomdata_add_nbat_f_to_f_gpu(Nbnxm::AtomLocality::Local,
1690 stateGpu->getForces(),
1691 pme_gpu_get_device_f(fr->pmedata),
1693 stepWork.useGpuPmeFReduction, haveLocalForceContribInCpuBuffer);
1694 stateGpu->copyForcesFromGpu(forceWithShift, gmx::StatePropagatorDataGpu::AtomLocality::Local);
1695 stateGpu->waitForcesReadyOnHost(gmx::StatePropagatorDataGpu::AtomLocality::Local);
1699 nbv->atomdata_add_nbat_f_to_f(Nbnxm::AtomLocality::Local, forceWithShift);
1704 launchGpuEndOfStepTasks(nbv, fr->gpuBonded, fr->pmedata, enerd,
1706 simulationWork.useGpuNonbonded, useGpuPmeOnThisRank,
1710 if (DOMAINDECOMP(cr))
1712 dd_force_flop_stop(cr->dd, nrnb);
1715 if (stepWork.computeForces)
1717 rvec *f = as_rvec_array(forceOut.forceWithShiftForces().force().data());
1719 /* If we have NoVirSum forces, but we do not calculate the virial,
1720 * we sum fr->f_novirsum=forceOut.f later.
1722 if (vsite && !(fr->haveDirectVirialContributions && !stepWork.computeVirial))
1724 rvec *fshift = as_rvec_array(forceOut.forceWithShiftForces().shiftForces().data());
1725 spread_vsite_f(vsite, as_rvec_array(x.unpaddedArrayRef().data()), f, fshift, FALSE, nullptr, nrnb,
1726 &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr, wcycle);
1729 if (stepWork.computeVirial)
1731 /* Calculation of the virial must be done after vsites! */
1732 calc_virial(0, mdatoms->homenr, as_rvec_array(x.unpaddedArrayRef().data()),
1733 forceOut.forceWithShiftForces(),
1734 vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1738 if (PAR(cr) && !thisRankHasDuty(cr, DUTY_PME))
1740 /* In case of node-splitting, the PP nodes receive the long-range
1741 * forces, virial and energy from the PME nodes here.
1743 pme_receive_force_ener(cr, &forceOut.forceWithVirial(), enerd, wcycle);
1746 if (stepWork.computeForces)
1748 post_process_forces(cr, step, nrnb, wcycle,
1749 top, box, as_rvec_array(x.unpaddedArrayRef().data()), &forceOut,
1750 vir_force, mdatoms, graph, fr, vsite,
1754 if (stepWork.computeEnergy)
1756 /* Sum the potential energy terms from group contributions */
1757 sum_epot(&(enerd->grpp), enerd->term);
1759 if (!EI_TPI(inputrec->eI))
1761 checkPotentialEnergyValidity(step, *enerd, *inputrec);
1765 /* In case we don't have constraints and are using GPUs, the next balancing
1766 * region starts here.
1767 * Some "special" work at the end of do_force_cuts?, such as vsite spread,
1768 * virial calculation and COM pulling, is not thus not included in
1769 * the balance timing, which is ok as most tasks do communication.
1771 ddBalanceRegionHandler.openBeforeForceComputationCpu(DdAllowBalanceRegionReopen::no);