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 * check out http://www.gromacs.org for more information.
7 * Copyright (c) 2012,2013, by the GROMACS development team, led by
8 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
9 * others, as listed in the AUTHORS file in the top-level source
10 * directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
51 #include "md_logging.h"
52 #include "md_support.h"
62 #include "mpelogging.h"
68 #include "checkpoint.h"
69 #include "mtop_util.h"
70 #include "sighandler.h"
73 #include "gmx_detect_hardware.h"
74 #include "gmx_omp_nthreads.h"
75 #include "pull_rotation.h"
76 #include "calc_verletbuf.h"
77 #include "../mdlib/nbnxn_search.h"
78 #include "../mdlib/nbnxn_consts.h"
79 #include "gmx_fatal_collective.h"
82 #include "gmx_thread_affinity.h"
95 #include "gpu_utils.h"
96 #include "nbnxn_cuda_data_mgmt.h"
99 gmx_integrator_t *func;
102 /* The array should match the eI array in include/types/enums.h */
103 const gmx_intp_t integrator[eiNR] = { {do_md}, {do_steep}, {do_cg}, {do_md}, {do_md}, {do_nm}, {do_lbfgs}, {do_tpi}, {do_tpi}, {do_md}, {do_md}, {do_md}};
105 gmx_large_int_t deform_init_init_step_tpx;
106 matrix deform_init_box_tpx;
107 #ifdef GMX_THREAD_MPI
108 tMPI_Thread_mutex_t deform_init_box_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
112 #ifdef GMX_THREAD_MPI
113 struct mdrunner_arglist
115 gmx_hw_opt_t *hw_opt;
128 const char *dddlb_opt;
133 const char *nbpu_opt;
134 gmx_large_int_t nsteps_cmdline;
144 const char *deviceOptions;
146 int ret; /* return value */
150 /* The function used for spawning threads. Extracts the mdrunner()
151 arguments from its one argument and calls mdrunner(), after making
153 static void mdrunner_start_fn(void *arg)
155 struct mdrunner_arglist *mda = (struct mdrunner_arglist*)arg;
156 struct mdrunner_arglist mc = *mda; /* copy the arg list to make sure
157 that it's thread-local. This doesn't
158 copy pointed-to items, of course,
159 but those are all const. */
160 t_commrec *cr; /* we need a local version of this */
164 fnm = dup_tfn(mc.nfile, mc.fnm);
166 cr = init_par_threads(mc.cr);
173 mda->ret = mdrunner(mc.hw_opt, fplog, cr, mc.nfile, fnm, mc.oenv,
174 mc.bVerbose, mc.bCompact, mc.nstglobalcomm,
175 mc.ddxyz, mc.dd_node_order, mc.rdd,
176 mc.rconstr, mc.dddlb_opt, mc.dlb_scale,
177 mc.ddcsx, mc.ddcsy, mc.ddcsz,
179 mc.nsteps_cmdline, mc.nstepout, mc.resetstep,
180 mc.nmultisim, mc.repl_ex_nst, mc.repl_ex_nex, mc.repl_ex_seed, mc.pforce,
181 mc.cpt_period, mc.max_hours, mc.deviceOptions, mc.Flags);
184 /* called by mdrunner() to start a specific number of threads (including
185 the main thread) for thread-parallel runs. This in turn calls mdrunner()
187 All options besides nthreads are the same as for mdrunner(). */
188 static t_commrec *mdrunner_start_threads(gmx_hw_opt_t *hw_opt,
189 FILE *fplog, t_commrec *cr, int nfile,
190 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
191 gmx_bool bCompact, int nstglobalcomm,
192 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
193 const char *dddlb_opt, real dlb_scale,
194 const char *ddcsx, const char *ddcsy, const char *ddcsz,
195 const char *nbpu_opt,
196 gmx_large_int_t nsteps_cmdline,
197 int nstepout, int resetstep,
198 int nmultisim, int repl_ex_nst, int repl_ex_nex, int repl_ex_seed,
199 real pforce, real cpt_period, real max_hours,
200 const char *deviceOptions, unsigned long Flags)
203 struct mdrunner_arglist *mda;
204 t_commrec *crn; /* the new commrec */
207 /* first check whether we even need to start tMPI */
208 if (hw_opt->nthreads_tmpi < 2)
213 /* a few small, one-time, almost unavoidable memory leaks: */
215 fnmn = dup_tfn(nfile, fnm);
217 /* fill the data structure to pass as void pointer to thread start fn */
218 mda->hw_opt = hw_opt;
224 mda->bVerbose = bVerbose;
225 mda->bCompact = bCompact;
226 mda->nstglobalcomm = nstglobalcomm;
227 mda->ddxyz[XX] = ddxyz[XX];
228 mda->ddxyz[YY] = ddxyz[YY];
229 mda->ddxyz[ZZ] = ddxyz[ZZ];
230 mda->dd_node_order = dd_node_order;
232 mda->rconstr = rconstr;
233 mda->dddlb_opt = dddlb_opt;
234 mda->dlb_scale = dlb_scale;
238 mda->nbpu_opt = nbpu_opt;
239 mda->nsteps_cmdline = nsteps_cmdline;
240 mda->nstepout = nstepout;
241 mda->resetstep = resetstep;
242 mda->nmultisim = nmultisim;
243 mda->repl_ex_nst = repl_ex_nst;
244 mda->repl_ex_nex = repl_ex_nex;
245 mda->repl_ex_seed = repl_ex_seed;
246 mda->pforce = pforce;
247 mda->cpt_period = cpt_period;
248 mda->max_hours = max_hours;
249 mda->deviceOptions = deviceOptions;
252 /* now spawn new threads that start mdrunner_start_fn(), while
253 the main thread returns, we set thread affinity later */
254 ret = tMPI_Init_fn(TRUE, hw_opt->nthreads_tmpi, TMPI_AFFINITY_NONE,
255 mdrunner_start_fn, (void*)(mda) );
256 if (ret != TMPI_SUCCESS)
261 /* make a new comm_rec to reflect the new situation */
262 crn = init_par_threads(cr);
267 static int get_tmpi_omp_thread_division(const gmx_hw_info_t *hwinfo,
268 const gmx_hw_opt_t *hw_opt,
274 /* There are no separate PME nodes here, as we ensured in
275 * check_and_update_hw_opt that nthreads_tmpi>0 with PME nodes
276 * and a conditional ensures we would not have ended up here.
277 * Note that separate PME nodes might be switched on later.
281 nthreads_tmpi = ngpu;
282 if (nthreads_tot > 0 && nthreads_tot < nthreads_tmpi)
284 nthreads_tmpi = nthreads_tot;
287 else if (hw_opt->nthreads_omp > 0)
289 /* Here we could oversubscribe, when we do, we issue a warning later */
290 nthreads_tmpi = max(1, nthreads_tot/hw_opt->nthreads_omp);
294 /* TODO choose nthreads_omp based on hardware topology
295 when we have a hardware topology detection library */
296 /* In general, when running up to 4 threads, OpenMP should be faster.
297 * Note: on AMD Bulldozer we should avoid running OpenMP over two dies.
298 * On Intel>=Nehalem running OpenMP on a single CPU is always faster,
299 * even on two CPUs it's usually faster (but with many OpenMP threads
300 * it could be faster not to use HT, currently we always use HT).
301 * On Nehalem/Westmere we want to avoid running 16 threads over
302 * two CPUs with HT, so we need a limit<16; thus we use 12.
303 * A reasonable limit for Intel Sandy and Ivy bridge,
304 * not knowing the topology, is 16 threads.
306 const int nthreads_omp_always_faster = 4;
307 const int nthreads_omp_always_faster_Nehalem = 12;
308 const int nthreads_omp_always_faster_SandyBridge = 16;
309 const int first_model_Nehalem = 0x1A;
310 const int first_model_SandyBridge = 0x2A;
311 gmx_bool bIntel_Family6;
314 (gmx_cpuid_vendor(hwinfo->cpuid_info) == GMX_CPUID_VENDOR_INTEL &&
315 gmx_cpuid_family(hwinfo->cpuid_info) == 6);
317 if (nthreads_tot <= nthreads_omp_always_faster ||
319 ((gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_Nehalem && nthreads_tot <= nthreads_omp_always_faster_Nehalem) ||
320 (gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_SandyBridge && nthreads_tot <= nthreads_omp_always_faster_SandyBridge))))
322 /* Use pure OpenMP parallelization */
327 /* Don't use OpenMP parallelization */
328 nthreads_tmpi = nthreads_tot;
332 return nthreads_tmpi;
336 /* Get the number of threads to use for thread-MPI based on how many
337 * were requested, which algorithms we're using,
338 * and how many particles there are.
339 * At the point we have already called check_and_update_hw_opt.
340 * Thus all options should be internally consistent and consistent
341 * with the hardware, except that ntmpi could be larger than #GPU.
343 static int get_nthreads_mpi(const gmx_hw_info_t *hwinfo,
344 gmx_hw_opt_t *hw_opt,
345 t_inputrec *inputrec, gmx_mtop_t *mtop,
349 int nthreads_hw, nthreads_tot_max, nthreads_tmpi, nthreads_new, ngpu;
350 int min_atoms_per_mpi_thread;
355 if (hw_opt->nthreads_tmpi > 0)
357 /* Trivial, return right away */
358 return hw_opt->nthreads_tmpi;
361 nthreads_hw = hwinfo->nthreads_hw_avail;
363 /* How many total (#tMPI*#OpenMP) threads can we start? */
364 if (hw_opt->nthreads_tot > 0)
366 nthreads_tot_max = hw_opt->nthreads_tot;
370 nthreads_tot_max = nthreads_hw;
373 bCanUseGPU = (inputrec->cutoff_scheme == ecutsVERLET && hwinfo->bCanUseGPU);
376 ngpu = hwinfo->gpu_info.ncuda_dev_use;
384 get_tmpi_omp_thread_division(hwinfo, hw_opt, nthreads_tot_max, ngpu);
386 if (inputrec->eI == eiNM || EI_TPI(inputrec->eI))
388 /* Dims/steps are divided over the nodes iso splitting the atoms */
389 min_atoms_per_mpi_thread = 0;
395 min_atoms_per_mpi_thread = MIN_ATOMS_PER_GPU;
399 min_atoms_per_mpi_thread = MIN_ATOMS_PER_MPI_THREAD;
403 /* Check if an algorithm does not support parallel simulation. */
404 if (nthreads_tmpi != 1 &&
405 ( inputrec->eI == eiLBFGS ||
406 inputrec->coulombtype == eelEWALD ) )
410 md_print_warn(cr, fplog, "The integration or electrostatics algorithm doesn't support parallel runs. Using a single thread-MPI thread.\n");
411 if (hw_opt->nthreads_tmpi > nthreads_tmpi)
413 gmx_fatal(FARGS, "You asked for more than 1 thread-MPI thread, but an algorithm doesn't support that");
416 else if (mtop->natoms/nthreads_tmpi < min_atoms_per_mpi_thread)
418 /* the thread number was chosen automatically, but there are too many
419 threads (too few atoms per thread) */
420 nthreads_new = max(1, mtop->natoms/min_atoms_per_mpi_thread);
422 /* Avoid partial use of Hyper-Threading */
423 if (gmx_cpuid_x86_smt(hwinfo->cpuid_info) == GMX_CPUID_X86_SMT_ENABLED &&
424 nthreads_new > nthreads_hw/2 && nthreads_new < nthreads_hw)
426 nthreads_new = nthreads_hw/2;
429 /* Avoid large prime numbers in the thread count */
430 if (nthreads_new >= 6)
432 /* Use only 6,8,10 with additional factors of 2 */
436 while (3*fac*2 <= nthreads_new)
441 nthreads_new = (nthreads_new/fac)*fac;
446 if (nthreads_new == 5)
452 nthreads_tmpi = nthreads_new;
454 fprintf(stderr, "\n");
455 fprintf(stderr, "NOTE: Parallelization is limited by the small number of atoms,\n");
456 fprintf(stderr, " only starting %d thread-MPI threads.\n", nthreads_tmpi);
457 fprintf(stderr, " You can use the -nt and/or -ntmpi option to optimize the number of threads.\n\n");
460 return nthreads_tmpi;
462 #endif /* GMX_THREAD_MPI */
465 /* Environment variable for setting nstlist */
466 static const char* NSTLIST_ENVVAR = "GMX_NSTLIST";
467 /* Try to increase nstlist when using a GPU with nstlist less than this */
468 static const int NSTLIST_GPU_ENOUGH = 20;
469 /* Increase nstlist until the non-bonded cost increases more than this factor */
470 static const float NBNXN_GPU_LIST_OK_FAC = 1.25;
471 /* Don't increase nstlist beyond a non-bonded cost increases of this factor */
472 static const float NBNXN_GPU_LIST_MAX_FAC = 1.40;
474 /* Try to increase nstlist when running on a GPU */
475 static void increase_nstlist(FILE *fp, t_commrec *cr,
476 t_inputrec *ir, const gmx_mtop_t *mtop, matrix box)
479 int nstlist_orig, nstlist_prev;
480 verletbuf_list_setup_t ls;
481 real rlist_inc, rlist_ok, rlist_max, rlist_new, rlist_prev;
484 gmx_bool bBox, bDD, bCont;
485 const char *nstl_fmt = "\nFor optimal performance with a GPU nstlist (now %d) should be larger.\nThe optimum depends on your CPU and GPU resources.\nYou might want to try several nstlist values.\n";
486 const char *vbd_err = "Can not increase nstlist for GPU run because verlet-buffer-drift is not set or used";
487 const char *box_err = "Can not increase nstlist for GPU run because the box is too small";
488 const char *dd_err = "Can not increase nstlist for GPU run because of domain decomposition limitations";
491 /* Number of + nstlist alternative values to try when switching */
492 const int nstl[] = { 20, 25, 40, 50 };
493 #define NNSTL sizeof(nstl)/sizeof(nstl[0])
495 env = getenv(NSTLIST_ENVVAR);
500 fprintf(fp, nstl_fmt, ir->nstlist);
504 if (ir->verletbuf_drift == 0)
506 gmx_fatal(FARGS, "You are using an old tpr file with a GPU, please generate a new tpr file with an up to date version of grompp");
509 if (ir->verletbuf_drift < 0)
513 fprintf(stderr, "%s\n", vbd_err);
517 fprintf(fp, "%s\n", vbd_err);
523 nstlist_orig = ir->nstlist;
526 sprintf(buf, "Getting nstlist from environment variable GMX_NSTLIST=%s", env);
529 fprintf(stderr, "%s\n", buf);
533 fprintf(fp, "%s\n", buf);
535 sscanf(env, "%d", &ir->nstlist);
538 verletbuf_get_list_setup(TRUE, &ls);
540 /* Allow rlist to make the list double the size of the cut-off sphere */
541 rlist_inc = nbnxn_get_rlist_effective_inc(NBNXN_GPU_CLUSTER_SIZE, mtop->natoms/det(box));
542 rlist_ok = (max(ir->rvdw, ir->rcoulomb) + rlist_inc)*pow(NBNXN_GPU_LIST_OK_FAC, 1.0/3.0) - rlist_inc;
543 rlist_max = (max(ir->rvdw, ir->rcoulomb) + rlist_inc)*pow(NBNXN_GPU_LIST_MAX_FAC, 1.0/3.0) - rlist_inc;
546 fprintf(debug, "GPU nstlist tuning: rlist_inc %.3f rlist_max %.3f\n",
547 rlist_inc, rlist_max);
551 nstlist_prev = nstlist_orig;
552 rlist_prev = ir->rlist;
557 ir->nstlist = nstl[i];
560 /* Set the pair-list buffer size in ir */
561 calc_verlet_buffer_size(mtop, det(box), ir, ir->verletbuf_drift, &ls,
564 /* Does rlist fit in the box? */
565 bBox = (sqr(rlist_new) < max_cutoff2(ir->ePBC, box));
567 if (bBox && DOMAINDECOMP(cr))
569 /* Check if rlist fits in the domain decomposition */
570 if (inputrec2nboundeddim(ir) < DIM)
572 gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
574 copy_mat(box, state_tmp.box);
575 bDD = change_dd_cutoff(cr, &state_tmp, ir, rlist_new);
582 if (bBox && bDD && rlist_new <= rlist_max)
584 /* Increase nstlist */
585 nstlist_prev = ir->nstlist;
586 rlist_prev = rlist_new;
587 bCont = (i+1 < NNSTL && rlist_new < rlist_ok);
591 /* Stick with the previous nstlist */
592 ir->nstlist = nstlist_prev;
593 rlist_new = rlist_prev;
605 gmx_warning(!bBox ? box_err : dd_err);
608 fprintf(fp, "\n%s\n", bBox ? box_err : dd_err);
610 ir->nstlist = nstlist_orig;
612 else if (ir->nstlist != nstlist_orig || rlist_new != ir->rlist)
614 sprintf(buf, "Changing nstlist from %d to %d, rlist from %g to %g",
615 nstlist_orig, ir->nstlist,
616 ir->rlist, rlist_new);
619 fprintf(stderr, "%s\n\n", buf);
623 fprintf(fp, "%s\n\n", buf);
625 ir->rlist = rlist_new;
626 ir->rlistlong = rlist_new;
630 static void prepare_verlet_scheme(FILE *fplog,
631 const gmx_hw_info_t *hwinfo,
633 const char *nbpu_opt,
635 const gmx_mtop_t *mtop,
639 /* Here we only check for GPU usage on the MPI master process,
640 * as here we don't know how many GPUs we will use yet.
641 * We check for a GPU on all processes later.
643 *bUseGPU = hwinfo->bCanUseGPU || (getenv("GMX_EMULATE_GPU") != NULL);
645 if (ir->verletbuf_drift > 0)
647 /* Update the Verlet buffer size for the current run setup */
648 verletbuf_list_setup_t ls;
651 /* Here we assume CPU acceleration is on. But as currently
652 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
653 * and 4x2 gives a larger buffer than 4x4, this is ok.
655 verletbuf_get_list_setup(*bUseGPU, &ls);
657 calc_verlet_buffer_size(mtop, det(box), ir,
658 ir->verletbuf_drift, &ls,
660 if (rlist_new != ir->rlist)
664 fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
665 ir->rlist, rlist_new,
666 ls.cluster_size_i, ls.cluster_size_j);
668 ir->rlist = rlist_new;
669 ir->rlistlong = rlist_new;
673 /* With GPU or emulation we should check nstlist for performance */
674 if ((EI_DYNAMICS(ir->eI) &&
676 ir->nstlist < NSTLIST_GPU_ENOUGH) ||
677 getenv(NSTLIST_ENVVAR) != NULL)
679 /* Choose a better nstlist */
680 increase_nstlist(fplog, cr, ir, mtop, box);
684 static void convert_to_verlet_scheme(FILE *fplog,
686 gmx_mtop_t *mtop, real box_vol)
688 char *conv_mesg = "Converting input file with group cut-off scheme to the Verlet cut-off scheme";
690 md_print_warn(NULL, fplog, "%s\n", conv_mesg);
692 ir->cutoff_scheme = ecutsVERLET;
693 ir->verletbuf_drift = 0.005;
695 if (ir->rcoulomb != ir->rvdw)
697 gmx_fatal(FARGS, "The VdW and Coulomb cut-offs are different, whereas the Verlet scheme only supports equal cut-offs");
700 if (ir->vdwtype == evdwUSER || EEL_USER(ir->coulombtype))
702 gmx_fatal(FARGS, "User non-bonded potentials are not (yet) supported with the Verlet scheme");
704 else if (EVDW_SWITCHED(ir->vdwtype) || EEL_SWITCHED(ir->coulombtype))
706 md_print_warn(NULL, fplog, "Converting switched or shifted interactions to a shifted potential (without force shift), this will lead to slightly different interaction potentials");
708 if (EVDW_SWITCHED(ir->vdwtype))
710 ir->vdwtype = evdwCUT;
712 if (EEL_SWITCHED(ir->coulombtype))
714 if (EEL_FULL(ir->coulombtype))
716 /* With full electrostatic only PME can be switched */
717 ir->coulombtype = eelPME;
721 md_print_warn(NULL, fplog, "NOTE: Replacing %s electrostatics with reaction-field with epsilon-rf=inf\n", eel_names[ir->coulombtype]);
722 ir->coulombtype = eelRF;
723 ir->epsilon_rf = 0.0;
727 /* We set the target energy drift to a small number.
728 * Note that this is only for testing. For production the user
729 * should think about this and set the mdp options.
731 ir->verletbuf_drift = 1e-4;
734 if (inputrec2nboundeddim(ir) != 3)
736 gmx_fatal(FARGS, "Can only convert old tpr files to the Verlet cut-off scheme with 3D pbc");
739 if (ir->efep != efepNO || ir->implicit_solvent != eisNO)
741 gmx_fatal(FARGS, "Will not convert old tpr files to the Verlet cut-off scheme with free-energy calculations or implicit solvent");
744 if (EI_DYNAMICS(ir->eI) && !(EI_MD(ir->eI) && ir->etc == etcNO))
746 verletbuf_list_setup_t ls;
748 verletbuf_get_list_setup(FALSE, &ls);
749 calc_verlet_buffer_size(mtop, box_vol, ir, ir->verletbuf_drift, &ls,
754 ir->verletbuf_drift = -1;
755 ir->rlist = 1.05*max(ir->rvdw, ir->rcoulomb);
758 gmx_mtop_remove_chargegroups(mtop);
761 static void check_and_update_hw_opt(gmx_hw_opt_t *hw_opt,
763 gmx_bool bIsSimMaster)
765 gmx_omp_nthreads_read_env(&hw_opt->nthreads_omp, bIsSimMaster);
767 #ifndef GMX_THREAD_MPI
768 if (hw_opt->nthreads_tot > 0)
770 gmx_fatal(FARGS, "Setting the total number of threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
772 if (hw_opt->nthreads_tmpi > 0)
774 gmx_fatal(FARGS, "Setting the number of thread-MPI threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
778 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp_pme <= 0)
780 /* We have the same number of OpenMP threads for PP and PME processes,
781 * thus we can perform several consistency checks.
783 if (hw_opt->nthreads_tmpi > 0 &&
784 hw_opt->nthreads_omp > 0 &&
785 hw_opt->nthreads_tot != hw_opt->nthreads_tmpi*hw_opt->nthreads_omp)
787 gmx_fatal(FARGS, "The total number of threads requested (%d) does not match the thread-MPI threads (%d) times the OpenMP threads (%d) requested",
788 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi, hw_opt->nthreads_omp);
791 if (hw_opt->nthreads_tmpi > 0 &&
792 hw_opt->nthreads_tot % hw_opt->nthreads_tmpi != 0)
794 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of thread-MPI threads requested (%d)",
795 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi);
798 if (hw_opt->nthreads_omp > 0 &&
799 hw_opt->nthreads_tot % hw_opt->nthreads_omp != 0)
801 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of OpenMP threads requested (%d)",
802 hw_opt->nthreads_tot, hw_opt->nthreads_omp);
805 if (hw_opt->nthreads_tmpi > 0 &&
806 hw_opt->nthreads_omp <= 0)
808 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
813 if (hw_opt->nthreads_omp > 1)
815 gmx_fatal(FARGS, "OpenMP threads are requested, but Gromacs was compiled without OpenMP support");
819 if (cutoff_scheme == ecutsGROUP)
821 /* We only have OpenMP support for PME only nodes */
822 if (hw_opt->nthreads_omp > 1)
824 gmx_fatal(FARGS, "OpenMP threads have been requested with cut-off scheme %s, but these are only supported with cut-off scheme %s",
825 ecutscheme_names[cutoff_scheme],
826 ecutscheme_names[ecutsVERLET]);
828 hw_opt->nthreads_omp = 1;
831 if (hw_opt->nthreads_omp_pme > 0 && hw_opt->nthreads_omp <= 0)
833 gmx_fatal(FARGS, "You need to specify -ntomp in addition to -ntomp_pme");
836 if (hw_opt->nthreads_tot == 1)
838 hw_opt->nthreads_tmpi = 1;
840 if (hw_opt->nthreads_omp > 1)
842 gmx_fatal(FARGS, "You requested %d OpenMP threads with %d total threads",
843 hw_opt->nthreads_tmpi, hw_opt->nthreads_tot);
845 hw_opt->nthreads_omp = 1;
848 if (hw_opt->nthreads_omp_pme <= 0 && hw_opt->nthreads_omp > 0)
850 hw_opt->nthreads_omp_pme = hw_opt->nthreads_omp;
855 fprintf(debug, "hw_opt: nt %d ntmpi %d ntomp %d ntomp_pme %d gpu_id '%s'\n",
856 hw_opt->nthreads_tot,
857 hw_opt->nthreads_tmpi,
858 hw_opt->nthreads_omp,
859 hw_opt->nthreads_omp_pme,
860 hw_opt->gpu_id != NULL ? hw_opt->gpu_id : "");
866 /* Override the value in inputrec with value passed on the command line (if any) */
867 static void override_nsteps_cmdline(FILE *fplog,
868 gmx_large_int_t nsteps_cmdline,
872 char sbuf[STEPSTRSIZE];
877 /* override with anything else than the default -2 */
878 if (nsteps_cmdline > -2)
882 ir->nsteps = nsteps_cmdline;
883 if (EI_DYNAMICS(ir->eI))
885 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps, %.3f ps",
886 gmx_step_str(nsteps_cmdline, sbuf),
887 nsteps_cmdline*ir->delta_t);
891 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps",
892 gmx_step_str(nsteps_cmdline, sbuf));
895 md_print_warn(cr, fplog, "%s\n", stmp);
899 /* Data structure set by SIMMASTER which needs to be passed to all nodes
900 * before the other nodes have read the tpx file and called gmx_detect_hardware.
903 int cutoff_scheme; /* The cutoff scheme from inputrec_t */
904 gmx_bool bUseGPU; /* Use GPU or GPU emulation */
907 int mdrunner(gmx_hw_opt_t *hw_opt,
908 FILE *fplog, t_commrec *cr, int nfile,
909 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
910 gmx_bool bCompact, int nstglobalcomm,
911 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
912 const char *dddlb_opt, real dlb_scale,
913 const char *ddcsx, const char *ddcsy, const char *ddcsz,
914 const char *nbpu_opt,
915 gmx_large_int_t nsteps_cmdline, int nstepout, int resetstep,
916 int nmultisim, int repl_ex_nst, int repl_ex_nex,
917 int repl_ex_seed, real pforce, real cpt_period, real max_hours,
918 const char *deviceOptions, unsigned long Flags)
920 gmx_bool bForceUseGPU, bTryUseGPU;
921 double nodetime = 0, realtime;
922 t_inputrec *inputrec;
923 t_state *state = NULL;
925 gmx_ddbox_t ddbox = {0};
926 int npme_major, npme_minor;
929 gmx_mtop_t *mtop = NULL;
930 t_mdatoms *mdatoms = NULL;
931 t_forcerec *fr = NULL;
932 t_fcdata *fcd = NULL;
934 gmx_pme_t *pmedata = NULL;
935 gmx_vsite_t *vsite = NULL;
937 int i, m, nChargePerturbed = -1, status, nalloc;
939 gmx_wallcycle_t wcycle;
940 gmx_bool bReadRNG, bReadEkin;
942 gmx_runtime_t runtime;
944 gmx_large_int_t reset_counters;
945 gmx_edsam_t ed = NULL;
946 t_commrec *cr_old = cr;
947 int nthreads_pme = 1;
949 gmx_membed_t membed = NULL;
950 gmx_hw_info_t *hwinfo = NULL;
951 master_inf_t minf = {-1, FALSE};
953 /* CAUTION: threads may be started later on in this function, so
954 cr doesn't reflect the final parallel state right now */
958 if (Flags & MD_APPENDFILES)
963 bForceUseGPU = (strncmp(nbpu_opt, "gpu", 3) == 0);
964 bTryUseGPU = (strncmp(nbpu_opt, "auto", 4) == 0) || bForceUseGPU;
966 /* Detect hardware, gather information. This is an operation that is
967 * global for this process (MPI rank). */
968 hwinfo = gmx_detect_hardware(fplog, cr,
969 bForceUseGPU, bTryUseGPU, hw_opt->gpu_id);
975 /* Read (nearly) all data required for the simulation */
976 read_tpx_state(ftp2fn(efTPX, nfile, fnm), inputrec, state, NULL, mtop);
978 if (inputrec->cutoff_scheme != ecutsVERLET &&
979 ((Flags & MD_TESTVERLET) || getenv("GMX_VERLET_SCHEME") != NULL))
981 convert_to_verlet_scheme(fplog, inputrec, mtop, det(state->box));
985 minf.cutoff_scheme = inputrec->cutoff_scheme;
986 minf.bUseGPU = FALSE;
988 if (inputrec->cutoff_scheme == ecutsVERLET)
990 prepare_verlet_scheme(fplog, hwinfo, cr, nbpu_opt,
991 inputrec, mtop, state->box,
994 else if (hwinfo->bCanUseGPU)
996 md_print_warn(cr, fplog,
997 "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
998 " To use a GPU, set the mdp option: cutoff-scheme = Verlet\n"
999 " (for quick performance testing you can use the -testverlet option)\n");
1003 gmx_fatal(FARGS, "GPU requested, but can't be used without cutoff-scheme=Verlet");
1007 #ifndef GMX_THREAD_MPI
1010 gmx_bcast_sim(sizeof(minf), &minf, cr);
1013 if (minf.bUseGPU && cr->npmenodes == -1)
1015 /* Don't automatically use PME-only nodes with GPUs */
1019 /* Check for externally set OpenMP affinity and turn off internal
1020 * pinning if any is found. We need to do this check early to tell
1021 * thread-MPI whether it should do pinning when spawning threads.
1022 * TODO: the above no longer holds, we should move these checks down
1024 gmx_omp_check_thread_affinity(fplog, cr, hw_opt);
1026 #ifdef GMX_THREAD_MPI
1027 /* With thread-MPI inputrec is only set here on the master thread */
1031 check_and_update_hw_opt(hw_opt, minf.cutoff_scheme, SIMMASTER(cr));
1033 #ifdef GMX_THREAD_MPI
1034 /* Early check for externally set process affinity. Can't do over all
1035 * MPI processes because hwinfo is not available everywhere, but with
1036 * thread-MPI it's needed as pinning might get turned off which needs
1037 * to be known before starting thread-MPI. */
1038 gmx_check_thread_affinity_set(fplog,
1040 hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1043 #ifdef GMX_THREAD_MPI
1044 if (cr->npmenodes > 0 && hw_opt->nthreads_tmpi <= 0)
1046 gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME nodes");
1050 if (hw_opt->nthreads_omp_pme != hw_opt->nthreads_omp &&
1053 gmx_fatal(FARGS, "You need to explicitly specify the number of PME nodes (-npme) when using different number of OpenMP threads for PP and PME nodes");
1057 #ifdef GMX_THREAD_MPI
1060 /* NOW the threads will be started: */
1061 hw_opt->nthreads_tmpi = get_nthreads_mpi(hwinfo,
1065 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp <= 0)
1067 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
1070 if (hw_opt->nthreads_tmpi > 1)
1072 /* now start the threads. */
1073 cr = mdrunner_start_threads(hw_opt, fplog, cr_old, nfile, fnm,
1074 oenv, bVerbose, bCompact, nstglobalcomm,
1075 ddxyz, dd_node_order, rdd, rconstr,
1076 dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
1078 nsteps_cmdline, nstepout, resetstep, nmultisim,
1079 repl_ex_nst, repl_ex_nex, repl_ex_seed, pforce,
1080 cpt_period, max_hours, deviceOptions,
1082 /* the main thread continues here with a new cr. We don't deallocate
1083 the old cr because other threads may still be reading it. */
1086 gmx_comm("Failed to spawn threads");
1091 /* END OF CAUTION: cr is now reliable */
1093 /* g_membed initialisation *
1094 * Because we change the mtop, init_membed is called before the init_parallel *
1095 * (in case we ever want to make it run in parallel) */
1096 if (opt2bSet("-membed", nfile, fnm))
1100 fprintf(stderr, "Initializing membed");
1102 membed = init_membed(fplog, nfile, fnm, mtop, inputrec, state, cr, &cpt_period);
1107 /* now broadcast everything to the non-master nodes/threads: */
1108 init_parallel(fplog, cr, inputrec, mtop);
1110 /* This check needs to happen after get_nthreads_mpi() */
1111 if (inputrec->cutoff_scheme == ecutsVERLET && (Flags & MD_PARTDEC))
1113 gmx_fatal_collective(FARGS, cr, NULL,
1114 "The Verlet cut-off scheme is not supported with particle decomposition.\n"
1115 "You can achieve the same effect as particle decomposition by running in parallel using only OpenMP threads.");
1120 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
1123 /* now make sure the state is initialized and propagated */
1124 set_state_entries(state, inputrec, cr->nnodes);
1126 /* A parallel command line option consistency check that we can
1127 only do after any threads have started. */
1129 (ddxyz[XX] > 1 || ddxyz[YY] > 1 || ddxyz[ZZ] > 1 || cr->npmenodes > 0))
1132 "The -dd or -npme option request a parallel simulation, "
1134 "but %s was compiled without threads or MPI enabled"
1136 #ifdef GMX_THREAD_MPI
1137 "but the number of threads (option -nt) is 1"
1139 "but %s was not started through mpirun/mpiexec or only one process was requested through mpirun/mpiexec"
1146 if ((Flags & MD_RERUN) &&
1147 (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
1149 gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
1152 if (can_use_allvsall(inputrec, mtop, TRUE, cr, fplog) && PAR(cr))
1154 /* Simple neighbour searching and (also?) all-vs-all loops
1155 * do not work with domain decomposition. */
1156 Flags |= MD_PARTDEC;
1159 if (!EEL_PME(inputrec->coulombtype) || (Flags & MD_PARTDEC))
1161 if (cr->npmenodes > 0)
1163 if (!EEL_PME(inputrec->coulombtype))
1165 gmx_fatal_collective(FARGS, cr, NULL,
1166 "PME nodes are requested, but the system does not use PME electrostatics");
1168 if (Flags & MD_PARTDEC)
1170 gmx_fatal_collective(FARGS, cr, NULL,
1171 "PME nodes are requested, but particle decomposition does not support separate PME nodes");
1179 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1182 /* NMR restraints must be initialized before load_checkpoint,
1183 * since with time averaging the history is added to t_state.
1184 * For proper consistency check we therefore need to extend
1186 * So the PME-only nodes (if present) will also initialize
1187 * the distance restraints.
1191 /* This needs to be called before read_checkpoint to extend the state */
1192 init_disres(fplog, mtop, inputrec, cr, Flags & MD_PARTDEC, fcd, state, repl_ex_nst > 0);
1194 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0)
1196 if (PAR(cr) && !(Flags & MD_PARTDEC))
1198 gmx_fatal(FARGS, "Orientation restraints do not work (yet) with domain decomposition, use particle decomposition (mdrun option -pd)");
1200 /* Orientation restraints */
1203 init_orires(fplog, mtop, state->x, inputrec, cr->ms, &(fcd->orires),
1208 if (DEFORM(*inputrec))
1210 /* Store the deform reference box before reading the checkpoint */
1213 copy_mat(state->box, box);
1217 gmx_bcast(sizeof(box), box, cr);
1219 /* Because we do not have the update struct available yet
1220 * in which the reference values should be stored,
1221 * we store them temporarily in static variables.
1222 * This should be thread safe, since they are only written once
1223 * and with identical values.
1225 #ifdef GMX_THREAD_MPI
1226 tMPI_Thread_mutex_lock(&deform_init_box_mutex);
1228 deform_init_init_step_tpx = inputrec->init_step;
1229 copy_mat(box, deform_init_box_tpx);
1230 #ifdef GMX_THREAD_MPI
1231 tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
1235 if (opt2bSet("-cpi", nfile, fnm))
1237 /* Check if checkpoint file exists before doing continuation.
1238 * This way we can use identical input options for the first and subsequent runs...
1240 if (gmx_fexist_master(opt2fn_master("-cpi", nfile, fnm, cr), cr) )
1242 load_checkpoint(opt2fn_master("-cpi", nfile, fnm, cr), &fplog,
1243 cr, Flags & MD_PARTDEC, ddxyz,
1244 inputrec, state, &bReadRNG, &bReadEkin,
1245 (Flags & MD_APPENDFILES),
1246 (Flags & MD_APPENDFILESSET));
1250 Flags |= MD_READ_RNG;
1254 Flags |= MD_READ_EKIN;
1259 if (((MASTER(cr) || (Flags & MD_SEPPOT)) && (Flags & MD_APPENDFILES))
1260 #ifdef GMX_THREAD_MPI
1261 /* With thread MPI only the master node/thread exists in mdrun.c,
1262 * therefore non-master nodes need to open the "seppot" log file here.
1264 || (!MASTER(cr) && (Flags & MD_SEPPOT))
1268 gmx_log_open(ftp2fn(efLOG, nfile, fnm), cr, !(Flags & MD_SEPPOT),
1272 /* override nsteps with value from cmdline */
1273 override_nsteps_cmdline(fplog, nsteps_cmdline, inputrec, cr);
1277 copy_mat(state->box, box);
1282 gmx_bcast(sizeof(box), box, cr);
1285 /* Essential dynamics */
1286 if (opt2bSet("-ei", nfile, fnm))
1288 /* Open input and output files, allocate space for ED data structure */
1289 ed = ed_open(mtop->natoms, &state->edsamstate, nfile, fnm, Flags, oenv, cr);
1292 if (PAR(cr) && !((Flags & MD_PARTDEC) ||
1293 EI_TPI(inputrec->eI) ||
1294 inputrec->eI == eiNM))
1296 cr->dd = init_domain_decomposition(fplog, cr, Flags, ddxyz, rdd, rconstr,
1297 dddlb_opt, dlb_scale,
1298 ddcsx, ddcsy, ddcsz,
1301 &ddbox, &npme_major, &npme_minor);
1303 make_dd_communicators(fplog, cr, dd_node_order);
1305 /* Set overallocation to avoid frequent reallocation of arrays */
1306 set_over_alloc_dd(TRUE);
1310 /* PME, if used, is done on all nodes with 1D decomposition */
1312 cr->duty = (DUTY_PP | DUTY_PME);
1315 /* NM and TPI perform single node energy calculations in parallel */
1316 if (!(inputrec->eI == eiNM || EI_TPI(inputrec->eI)))
1318 npme_major = cr->nnodes;
1321 if (inputrec->ePBC == epbcSCREW)
1324 "pbc=%s is only implemented with domain decomposition",
1325 epbc_names[inputrec->ePBC]);
1331 /* After possible communicator splitting in make_dd_communicators.
1332 * we can set up the intra/inter node communication.
1334 gmx_setup_nodecomm(fplog, cr);
1337 /* Initialize per-physical-node MPI process/thread ID and counters. */
1338 gmx_init_intranode_counters(cr);
1341 md_print_info(cr, fplog, "Using %d MPI %s\n",
1343 #ifdef GMX_THREAD_MPI
1344 cr->nnodes == 1 ? "thread" : "threads"
1346 cr->nnodes == 1 ? "process" : "processes"
1352 gmx_omp_nthreads_init(fplog, cr,
1353 hwinfo->nthreads_hw_avail,
1354 hw_opt->nthreads_omp,
1355 hw_opt->nthreads_omp_pme,
1356 (cr->duty & DUTY_PP) == 0,
1357 inputrec->cutoff_scheme == ecutsVERLET);
1359 /* check consistency and decide on the number of gpus to use. */
1360 gmx_check_hw_runconf_consistency(fplog, hwinfo, cr, hw_opt->nthreads_tmpi,
1363 /* getting number of PP/PME threads
1364 PME: env variable should be read only on one node to make sure it is
1365 identical everywhere;
1367 /* TODO nthreads_pp is only used for pinning threads.
1368 * This is a temporary solution until we have a hw topology library.
1370 nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
1371 nthreads_pme = gmx_omp_nthreads_get(emntPME);
1373 wcycle = wallcycle_init(fplog, resetstep, cr, nthreads_pp, nthreads_pme);
1377 /* Master synchronizes its value of reset_counters with all nodes
1378 * including PME only nodes */
1379 reset_counters = wcycle_get_reset_counters(wcycle);
1380 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1381 wcycle_set_reset_counters(wcycle, reset_counters);
1385 if (cr->duty & DUTY_PP)
1387 /* For domain decomposition we allocate dynamically
1388 * in dd_partition_system.
1390 if (DOMAINDECOMP(cr))
1392 bcast_state_setup(cr, state);
1398 bcast_state(cr, state, TRUE);
1402 /* Initiate forcerecord */
1404 fr->hwinfo = hwinfo;
1405 init_forcerec(fplog, oenv, fr, fcd, inputrec, mtop, cr, box, FALSE,
1406 opt2fn("-table", nfile, fnm),
1407 opt2fn("-tabletf", nfile, fnm),
1408 opt2fn("-tablep", nfile, fnm),
1409 opt2fn("-tableb", nfile, fnm),
1413 /* version for PCA_NOT_READ_NODE (see md.c) */
1414 /*init_forcerec(fplog,fr,fcd,inputrec,mtop,cr,box,FALSE,
1415 "nofile","nofile","nofile","nofile",FALSE,pforce);
1417 fr->bSepDVDL = ((Flags & MD_SEPPOT) == MD_SEPPOT);
1419 /* Initialize QM-MM */
1422 init_QMMMrec(cr, box, mtop, inputrec, fr);
1425 /* Initialize the mdatoms structure.
1426 * mdatoms is not filled with atom data,
1427 * as this can not be done now with domain decomposition.
1429 mdatoms = init_mdatoms(fplog, mtop, inputrec->efep != efepNO);
1431 if (mdatoms->nPerturbed > 0 && inputrec->cutoff_scheme == ecutsVERLET)
1433 gmx_fatal(FARGS, "The Verlet cut-off scheme does not (yet) support free-energy calculations with perturbed atoms, only perturbed interactions. This will be implemented soon. Use the group scheme for now.");
1436 /* Initialize the virtual site communication */
1437 vsite = init_vsite(mtop, cr, FALSE);
1439 calc_shifts(box, fr->shift_vec);
1441 /* With periodic molecules the charge groups should be whole at start up
1442 * and the virtual sites should not be far from their proper positions.
1444 if (!inputrec->bContinuation && MASTER(cr) &&
1445 !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1447 /* Make molecules whole at start of run */
1448 if (fr->ePBC != epbcNONE)
1450 do_pbc_first_mtop(fplog, inputrec->ePBC, box, mtop, state->x);
1454 /* Correct initial vsite positions are required
1455 * for the initial distribution in the domain decomposition
1456 * and for the initial shell prediction.
1458 construct_vsites_mtop(fplog, vsite, mtop, state->x);
1462 if (EEL_PME(fr->eeltype))
1464 ewaldcoeff = fr->ewaldcoeff;
1465 pmedata = &fr->pmedata;
1474 /* This is a PME only node */
1476 /* We don't need the state */
1479 ewaldcoeff = calc_ewaldcoeff(inputrec->rcoulomb, inputrec->ewald_rtol);
1483 if (hw_opt->thread_affinity != threadaffOFF)
1485 /* Before setting affinity, check whether the affinity has changed
1486 * - which indicates that probably the OpenMP library has changed it
1487 * since we first checked).
1489 gmx_check_thread_affinity_set(fplog, cr,
1490 hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1492 /* Set the CPU affinity */
1493 gmx_set_thread_affinity(fplog, cr, hw_opt, nthreads_pme, hwinfo,
1497 /* Initiate PME if necessary,
1498 * either on all nodes or on dedicated PME nodes only. */
1499 if (EEL_PME(inputrec->coulombtype))
1503 nChargePerturbed = mdatoms->nChargePerturbed;
1505 if (cr->npmenodes > 0)
1507 /* The PME only nodes need to know nChargePerturbed */
1508 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1511 if (cr->duty & DUTY_PME)
1513 status = gmx_pme_init(pmedata, cr, npme_major, npme_minor, inputrec,
1514 mtop ? mtop->natoms : 0, nChargePerturbed,
1515 (Flags & MD_REPRODUCIBLE), nthreads_pme);
1518 gmx_fatal(FARGS, "Error %d initializing PME", status);
1524 if (integrator[inputrec->eI].func == do_md)
1526 /* Turn on signal handling on all nodes */
1528 * (A user signal from the PME nodes (if any)
1529 * is communicated to the PP nodes.
1531 signal_handler_install();
1534 if (cr->duty & DUTY_PP)
1536 if (inputrec->ePull != epullNO)
1538 /* Initialize pull code */
1539 init_pull(fplog, inputrec, nfile, fnm, mtop, cr, oenv, inputrec->fepvals->init_lambda,
1540 EI_DYNAMICS(inputrec->eI) && MASTER(cr), Flags);
1545 /* Initialize enforced rotation code */
1546 init_rot(fplog, inputrec, nfile, fnm, cr, state->x, box, mtop, oenv,
1550 constr = init_constraints(fplog, mtop, inputrec, ed, state, cr);
1552 if (DOMAINDECOMP(cr))
1554 dd_init_bondeds(fplog, cr->dd, mtop, vsite, constr, inputrec,
1555 Flags & MD_DDBONDCHECK, fr->cginfo_mb);
1557 set_dd_parameters(fplog, cr->dd, dlb_scale, inputrec, fr, &ddbox);
1559 setup_dd_grid(fplog, cr->dd);
1562 /* Now do whatever the user wants us to do (how flexible...) */
1563 integrator[inputrec->eI].func(fplog, cr, nfile, fnm,
1564 oenv, bVerbose, bCompact,
1567 nstepout, inputrec, mtop,
1569 mdatoms, nrnb, wcycle, ed, fr,
1570 repl_ex_nst, repl_ex_nex, repl_ex_seed,
1572 cpt_period, max_hours,
1577 if (inputrec->ePull != epullNO)
1579 finish_pull(fplog, inputrec->pull);
1584 finish_rot(inputrec->rot);
1591 gmx_pmeonly(*pmedata, cr, nrnb, wcycle, ewaldcoeff, FALSE, inputrec);
1594 if (EI_DYNAMICS(inputrec->eI) || EI_TPI(inputrec->eI))
1596 /* Some timing stats */
1599 if (runtime.proc == 0)
1601 runtime.proc = runtime.real;
1610 wallcycle_stop(wcycle, ewcRUN);
1612 /* Finish up, write some stuff
1613 * if rerunMD, don't write last frame again
1615 finish_run(fplog, cr, ftp2fn(efSTO, nfile, fnm),
1616 inputrec, nrnb, wcycle, &runtime,
1617 fr != NULL && fr->nbv != NULL && fr->nbv->bUseGPU ?
1618 nbnxn_cuda_get_timings(fr->nbv->cu_nbv) : NULL,
1620 EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
1622 if ((cr->duty & DUTY_PP) && fr->nbv != NULL && fr->nbv->bUseGPU)
1624 char gpu_err_str[STRLEN];
1626 /* free GPU memory and uninitialize GPU (by destroying the context) */
1627 nbnxn_cuda_free(fplog, fr->nbv->cu_nbv);
1629 if (!free_gpu(gpu_err_str))
1631 gmx_warning("On node %d failed to free GPU #%d: %s",
1632 cr->nodeid, get_current_gpu_device_id(), gpu_err_str);
1636 if (opt2bSet("-membed", nfile, fnm))
1641 gmx_hardware_info_free(hwinfo);
1643 /* Does what it says */
1644 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", &runtime);
1646 /* Close logfile already here if we were appending to it */
1647 if (MASTER(cr) && (Flags & MD_APPENDFILES))
1649 gmx_log_close(fplog);
1652 rc = (int)gmx_get_stop_condition();
1654 #ifdef GMX_THREAD_MPI
1655 /* we need to join all threads. The sub-threads join when they
1656 exit this function, but the master thread needs to be told to
1658 if (PAR(cr) && MASTER(cr))