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) 2011,2012,2013, 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.
54 #include "md_logging.h"
55 #include "md_support.h"
58 #include "pull_rotation.h"
71 #include "checkpoint.h"
72 #include "mtop_util.h"
73 #include "sighandler.h"
75 #include "gmx_detect_hardware.h"
76 #include "gmx_omp_nthreads.h"
77 #include "pull_rotation.h"
78 #include "calc_verletbuf.h"
79 #include "gmx_fatal_collective.h"
82 #include "gmx_thread_affinity.h"
84 #include "gromacs/fileio/tpxio.h"
85 #include "gromacs/mdlib/nbnxn_search.h"
86 #include "gromacs/mdlib/nbnxn_consts.h"
87 #include "gromacs/timing/wallcycle.h"
88 #include "gromacs/utility/gmxmpi.h"
89 #include "gromacs/utility/gmxomp.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 tMPI_Thread_mutex_t deform_init_box_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
110 #ifdef GMX_THREAD_MPI
111 struct mdrunner_arglist
126 const char *dddlb_opt;
131 const char *nbpu_opt;
132 gmx_large_int_t nsteps_cmdline;
142 const char *deviceOptions;
144 int ret; /* return value */
148 /* The function used for spawning threads. Extracts the mdrunner()
149 arguments from its one argument and calls mdrunner(), after making
151 static void mdrunner_start_fn(void *arg)
153 struct mdrunner_arglist *mda = (struct mdrunner_arglist*)arg;
154 struct mdrunner_arglist mc = *mda; /* copy the arg list to make sure
155 that it's thread-local. This doesn't
156 copy pointed-to items, of course,
157 but those are all const. */
158 t_commrec *cr; /* we need a local version of this */
162 fnm = dup_tfn(mc.nfile, mc.fnm);
164 cr = reinitialize_commrec_for_this_thread(mc.cr);
171 mda->ret = mdrunner(&mc.hw_opt, fplog, cr, mc.nfile, fnm, mc.oenv,
172 mc.bVerbose, mc.bCompact, mc.nstglobalcomm,
173 mc.ddxyz, mc.dd_node_order, mc.rdd,
174 mc.rconstr, mc.dddlb_opt, mc.dlb_scale,
175 mc.ddcsx, mc.ddcsy, mc.ddcsz,
177 mc.nsteps_cmdline, mc.nstepout, mc.resetstep,
178 mc.nmultisim, mc.repl_ex_nst, mc.repl_ex_nex, mc.repl_ex_seed, mc.pforce,
179 mc.cpt_period, mc.max_hours, mc.deviceOptions, mc.Flags);
182 /* called by mdrunner() to start a specific number of threads (including
183 the main thread) for thread-parallel runs. This in turn calls mdrunner()
185 All options besides nthreads are the same as for mdrunner(). */
186 static t_commrec *mdrunner_start_threads(gmx_hw_opt_t *hw_opt,
187 FILE *fplog, t_commrec *cr, int nfile,
188 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
189 gmx_bool bCompact, int nstglobalcomm,
190 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
191 const char *dddlb_opt, real dlb_scale,
192 const char *ddcsx, const char *ddcsy, const char *ddcsz,
193 const char *nbpu_opt,
194 gmx_large_int_t nsteps_cmdline,
195 int nstepout, int resetstep,
196 int nmultisim, int repl_ex_nst, int repl_ex_nex, int repl_ex_seed,
197 real pforce, real cpt_period, real max_hours,
198 const char *deviceOptions, unsigned long Flags)
201 struct mdrunner_arglist *mda;
202 t_commrec *crn; /* the new commrec */
205 /* first check whether we even need to start tMPI */
206 if (hw_opt->nthreads_tmpi < 2)
211 /* a few small, one-time, almost unavoidable memory leaks: */
213 fnmn = dup_tfn(nfile, fnm);
215 /* fill the data structure to pass as void pointer to thread start fn */
216 /* hw_opt contains pointers, which should all be NULL at this stage */
217 mda->hw_opt = *hw_opt;
223 mda->bVerbose = bVerbose;
224 mda->bCompact = bCompact;
225 mda->nstglobalcomm = nstglobalcomm;
226 mda->ddxyz[XX] = ddxyz[XX];
227 mda->ddxyz[YY] = ddxyz[YY];
228 mda->ddxyz[ZZ] = ddxyz[ZZ];
229 mda->dd_node_order = dd_node_order;
231 mda->rconstr = rconstr;
232 mda->dddlb_opt = dddlb_opt;
233 mda->dlb_scale = dlb_scale;
237 mda->nbpu_opt = nbpu_opt;
238 mda->nsteps_cmdline = nsteps_cmdline;
239 mda->nstepout = nstepout;
240 mda->resetstep = resetstep;
241 mda->nmultisim = nmultisim;
242 mda->repl_ex_nst = repl_ex_nst;
243 mda->repl_ex_nex = repl_ex_nex;
244 mda->repl_ex_seed = repl_ex_seed;
245 mda->pforce = pforce;
246 mda->cpt_period = cpt_period;
247 mda->max_hours = max_hours;
248 mda->deviceOptions = deviceOptions;
251 /* now spawn new threads that start mdrunner_start_fn(), while
252 the main thread returns, we set thread affinity later */
253 ret = tMPI_Init_fn(TRUE, hw_opt->nthreads_tmpi, TMPI_AFFINITY_NONE,
254 mdrunner_start_fn, (void*)(mda) );
255 if (ret != TMPI_SUCCESS)
260 crn = reinitialize_commrec_for_this_thread(cr);
265 static int get_tmpi_omp_thread_division(const gmx_hw_info_t *hwinfo,
266 const gmx_hw_opt_t *hw_opt,
272 /* There are no separate PME nodes here, as we ensured in
273 * check_and_update_hw_opt that nthreads_tmpi>0 with PME nodes
274 * and a conditional ensures we would not have ended up here.
275 * Note that separate PME nodes might be switched on later.
279 nthreads_tmpi = ngpu;
280 if (nthreads_tot > 0 && nthreads_tot < nthreads_tmpi)
282 nthreads_tmpi = nthreads_tot;
285 else if (hw_opt->nthreads_omp > 0)
287 /* Here we could oversubscribe, when we do, we issue a warning later */
288 nthreads_tmpi = max(1, nthreads_tot/hw_opt->nthreads_omp);
292 /* TODO choose nthreads_omp based on hardware topology
293 when we have a hardware topology detection library */
294 /* In general, when running up to 4 threads, OpenMP should be faster.
295 * Note: on AMD Bulldozer we should avoid running OpenMP over two dies.
296 * On Intel>=Nehalem running OpenMP on a single CPU is always faster,
297 * even on two CPUs it's usually faster (but with many OpenMP threads
298 * it could be faster not to use HT, currently we always use HT).
299 * On Nehalem/Westmere we want to avoid running 16 threads over
300 * two CPUs with HT, so we need a limit<16; thus we use 12.
301 * A reasonable limit for Intel Sandy and Ivy bridge,
302 * not knowing the topology, is 16 threads.
304 const int nthreads_omp_always_faster = 4;
305 const int nthreads_omp_always_faster_Nehalem = 12;
306 const int nthreads_omp_always_faster_SandyBridge = 16;
307 const int first_model_Nehalem = 0x1A;
308 const int first_model_SandyBridge = 0x2A;
309 gmx_bool bIntel_Family6;
312 (gmx_cpuid_vendor(hwinfo->cpuid_info) == GMX_CPUID_VENDOR_INTEL &&
313 gmx_cpuid_family(hwinfo->cpuid_info) == 6);
315 if (nthreads_tot <= nthreads_omp_always_faster ||
317 ((gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_Nehalem && nthreads_tot <= nthreads_omp_always_faster_Nehalem) ||
318 (gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_SandyBridge && nthreads_tot <= nthreads_omp_always_faster_SandyBridge))))
320 /* Use pure OpenMP parallelization */
325 /* Don't use OpenMP parallelization */
326 nthreads_tmpi = nthreads_tot;
330 return nthreads_tmpi;
334 /* Get the number of threads to use for thread-MPI based on how many
335 * were requested, which algorithms we're using,
336 * and how many particles there are.
337 * At the point we have already called check_and_update_hw_opt.
338 * Thus all options should be internally consistent and consistent
339 * with the hardware, except that ntmpi could be larger than #GPU.
341 static int get_nthreads_mpi(const gmx_hw_info_t *hwinfo,
342 gmx_hw_opt_t *hw_opt,
343 t_inputrec *inputrec, gmx_mtop_t *mtop,
347 int nthreads_hw, nthreads_tot_max, nthreads_tmpi, nthreads_new, ngpu;
348 int min_atoms_per_mpi_thread;
353 if (hw_opt->nthreads_tmpi > 0)
355 /* Trivial, return right away */
356 return hw_opt->nthreads_tmpi;
359 nthreads_hw = hwinfo->nthreads_hw_avail;
361 /* How many total (#tMPI*#OpenMP) threads can we start? */
362 if (hw_opt->nthreads_tot > 0)
364 nthreads_tot_max = hw_opt->nthreads_tot;
368 nthreads_tot_max = nthreads_hw;
371 bCanUseGPU = (inputrec->cutoff_scheme == ecutsVERLET &&
372 hwinfo->gpu_info.ncuda_dev_compatible > 0);
375 ngpu = hwinfo->gpu_info.ncuda_dev_compatible;
382 if (inputrec->cutoff_scheme == ecutsGROUP)
384 /* We checked this before, but it doesn't hurt to do it once more */
385 assert(hw_opt->nthreads_omp == 1);
389 get_tmpi_omp_thread_division(hwinfo, hw_opt, nthreads_tot_max, ngpu);
391 if (inputrec->eI == eiNM || EI_TPI(inputrec->eI))
393 /* Dims/steps are divided over the nodes iso splitting the atoms */
394 min_atoms_per_mpi_thread = 0;
400 min_atoms_per_mpi_thread = MIN_ATOMS_PER_GPU;
404 min_atoms_per_mpi_thread = MIN_ATOMS_PER_MPI_THREAD;
408 /* Check if an algorithm does not support parallel simulation. */
409 if (nthreads_tmpi != 1 &&
410 ( inputrec->eI == eiLBFGS ||
411 inputrec->coulombtype == eelEWALD ) )
415 md_print_warn(cr, fplog, "The integration or electrostatics algorithm doesn't support parallel runs. Using a single thread-MPI thread.\n");
416 if (hw_opt->nthreads_tmpi > nthreads_tmpi)
418 gmx_fatal(FARGS, "You asked for more than 1 thread-MPI thread, but an algorithm doesn't support that");
421 else if (mtop->natoms/nthreads_tmpi < min_atoms_per_mpi_thread)
423 /* the thread number was chosen automatically, but there are too many
424 threads (too few atoms per thread) */
425 nthreads_new = max(1, mtop->natoms/min_atoms_per_mpi_thread);
427 /* Avoid partial use of Hyper-Threading */
428 if (gmx_cpuid_x86_smt(hwinfo->cpuid_info) == GMX_CPUID_X86_SMT_ENABLED &&
429 nthreads_new > nthreads_hw/2 && nthreads_new < nthreads_hw)
431 nthreads_new = nthreads_hw/2;
434 /* Avoid large prime numbers in the thread count */
435 if (nthreads_new >= 6)
437 /* Use only 6,8,10 with additional factors of 2 */
441 while (3*fac*2 <= nthreads_new)
446 nthreads_new = (nthreads_new/fac)*fac;
451 if (nthreads_new == 5)
457 nthreads_tmpi = nthreads_new;
459 fprintf(stderr, "\n");
460 fprintf(stderr, "NOTE: Parallelization is limited by the small number of atoms,\n");
461 fprintf(stderr, " only starting %d thread-MPI threads.\n", nthreads_tmpi);
462 fprintf(stderr, " You can use the -nt and/or -ntmpi option to optimize the number of threads.\n\n");
465 return nthreads_tmpi;
467 #endif /* GMX_THREAD_MPI */
470 /* Environment variable for setting nstlist */
471 static const char* NSTLIST_ENVVAR = "GMX_NSTLIST";
472 /* Try to increase nstlist when using a GPU with nstlist less than this */
473 static const int NSTLIST_GPU_ENOUGH = 20;
474 /* Increase nstlist until the non-bonded cost increases more than this factor */
475 static const float NBNXN_GPU_LIST_OK_FAC = 1.20;
476 /* Don't increase nstlist beyond a non-bonded cost increases of this factor.
477 * A standard (protein+)water system at 300K with PME ewald_rtol=1e-5
478 * needs 1.28 at rcoulomb=0.9 and 1.24 at rcoulomb=1.0 to get to nstlist=40.
480 static const float NBNXN_GPU_LIST_MAX_FAC = 1.30;
482 /* Try to increase nstlist when running on a GPU */
483 static void increase_nstlist(FILE *fp, t_commrec *cr,
484 t_inputrec *ir, const gmx_mtop_t *mtop, matrix box)
487 int nstlist_orig, nstlist_prev;
488 verletbuf_list_setup_t ls;
489 real rlist_nstlist10, rlist_inc, rlist_ok, rlist_max;
490 real rlist_new, rlist_prev;
493 gmx_bool bBox, bDD, bCont;
494 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";
495 const char *vbd_err = "Can not increase nstlist for GPU run because verlet-buffer-drift is not set or used";
496 const char *box_err = "Can not increase nstlist for GPU run because the box is too small";
497 const char *dd_err = "Can not increase nstlist for GPU run because of domain decomposition limitations";
500 /* Number of + nstlist alternative values to try when switching */
501 const int nstl[] = { 20, 25, 40 };
502 #define NNSTL sizeof(nstl)/sizeof(nstl[0])
504 env = getenv(NSTLIST_ENVVAR);
509 fprintf(fp, nstl_fmt, ir->nstlist);
513 if (ir->verletbuf_drift == 0)
515 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");
518 if (ir->verletbuf_drift < 0)
522 fprintf(stderr, "%s\n", vbd_err);
526 fprintf(fp, "%s\n", vbd_err);
532 nstlist_orig = ir->nstlist;
535 sprintf(buf, "Getting nstlist from environment variable GMX_NSTLIST=%s", env);
538 fprintf(stderr, "%s\n", buf);
542 fprintf(fp, "%s\n", buf);
544 sscanf(env, "%d", &ir->nstlist);
547 verletbuf_get_list_setup(TRUE, &ls);
549 /* Allow rlist to make the list a given factor larger than the list
550 * would be with nstlist=10.
552 nstlist_prev = ir->nstlist;
554 calc_verlet_buffer_size(mtop, det(box), ir, ir->verletbuf_drift, &ls,
555 NULL, &rlist_nstlist10);
556 ir->nstlist = nstlist_prev;
558 /* Determine the pair list size increase due to zero interactions */
559 rlist_inc = nbnxn_get_rlist_effective_inc(NBNXN_GPU_CLUSTER_SIZE, mtop->natoms/det(box));
560 rlist_ok = (rlist_nstlist10 + rlist_inc)*pow(NBNXN_GPU_LIST_OK_FAC, 1.0/3.0) - rlist_inc;
561 rlist_max = (rlist_nstlist10 + rlist_inc)*pow(NBNXN_GPU_LIST_MAX_FAC, 1.0/3.0) - rlist_inc;
564 fprintf(debug, "GPU nstlist tuning: rlist_inc %.3f rlist_max %.3f\n",
565 rlist_inc, rlist_max);
569 nstlist_prev = nstlist_orig;
570 rlist_prev = ir->rlist;
575 ir->nstlist = nstl[i];
578 /* Set the pair-list buffer size in ir */
579 calc_verlet_buffer_size(mtop, det(box), ir, ir->verletbuf_drift, &ls,
582 /* Does rlist fit in the box? */
583 bBox = (sqr(rlist_new) < max_cutoff2(ir->ePBC, box));
585 if (bBox && DOMAINDECOMP(cr))
587 /* Check if rlist fits in the domain decomposition */
588 if (inputrec2nboundeddim(ir) < DIM)
590 gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
592 copy_mat(box, state_tmp.box);
593 bDD = change_dd_cutoff(cr, &state_tmp, ir, rlist_new);
600 if (bBox && bDD && rlist_new <= rlist_max)
602 /* Increase nstlist */
603 nstlist_prev = ir->nstlist;
604 rlist_prev = rlist_new;
605 bCont = (i+1 < NNSTL && rlist_new < rlist_ok);
609 /* Stick with the previous nstlist */
610 ir->nstlist = nstlist_prev;
611 rlist_new = rlist_prev;
623 gmx_warning(!bBox ? box_err : dd_err);
626 fprintf(fp, "\n%s\n", bBox ? box_err : dd_err);
628 ir->nstlist = nstlist_orig;
630 else if (ir->nstlist != nstlist_orig || rlist_new != ir->rlist)
632 sprintf(buf, "Changing nstlist from %d to %d, rlist from %g to %g",
633 nstlist_orig, ir->nstlist,
634 ir->rlist, rlist_new);
637 fprintf(stderr, "%s\n\n", buf);
641 fprintf(fp, "%s\n\n", buf);
643 ir->rlist = rlist_new;
644 ir->rlistlong = rlist_new;
648 static void prepare_verlet_scheme(FILE *fplog,
651 const gmx_mtop_t *mtop,
655 if (ir->verletbuf_drift > 0)
657 /* Update the Verlet buffer size for the current run setup */
658 verletbuf_list_setup_t ls;
661 /* Here we assume CPU acceleration is on. But as currently
662 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
663 * and 4x2 gives a larger buffer than 4x4, this is ok.
665 verletbuf_get_list_setup(bUseGPU, &ls);
667 calc_verlet_buffer_size(mtop, det(box), ir,
668 ir->verletbuf_drift, &ls,
670 if (rlist_new != ir->rlist)
674 fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
675 ir->rlist, rlist_new,
676 ls.cluster_size_i, ls.cluster_size_j);
678 ir->rlist = rlist_new;
679 ir->rlistlong = rlist_new;
683 /* With GPU or emulation we should check nstlist for performance */
684 if ((EI_DYNAMICS(ir->eI) &&
686 ir->nstlist < NSTLIST_GPU_ENOUGH) ||
687 getenv(NSTLIST_ENVVAR) != NULL)
689 /* Choose a better nstlist */
690 increase_nstlist(fplog, cr, ir, mtop, box);
694 static void convert_to_verlet_scheme(FILE *fplog,
696 gmx_mtop_t *mtop, real box_vol)
698 char *conv_mesg = "Converting input file with group cut-off scheme to the Verlet cut-off scheme";
700 md_print_warn(NULL, fplog, "%s\n", conv_mesg);
702 ir->cutoff_scheme = ecutsVERLET;
703 ir->verletbuf_drift = 0.005;
705 if (ir->rcoulomb != ir->rvdw)
707 gmx_fatal(FARGS, "The VdW and Coulomb cut-offs are different, whereas the Verlet scheme only supports equal cut-offs");
710 if (ir->vdwtype == evdwUSER || EEL_USER(ir->coulombtype))
712 gmx_fatal(FARGS, "User non-bonded potentials are not (yet) supported with the Verlet scheme");
714 else if (EVDW_SWITCHED(ir->vdwtype) || EEL_SWITCHED(ir->coulombtype))
716 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");
718 if (EVDW_SWITCHED(ir->vdwtype))
720 ir->vdwtype = evdwCUT;
722 if (EEL_SWITCHED(ir->coulombtype))
724 if (EEL_FULL(ir->coulombtype))
726 /* With full electrostatic only PME can be switched */
727 ir->coulombtype = eelPME;
731 md_print_warn(NULL, fplog, "NOTE: Replacing %s electrostatics with reaction-field with epsilon-rf=inf\n", eel_names[ir->coulombtype]);
732 ir->coulombtype = eelRF;
733 ir->epsilon_rf = 0.0;
737 /* We set the target energy drift to a small number.
738 * Note that this is only for testing. For production the user
739 * should think about this and set the mdp options.
741 ir->verletbuf_drift = 1e-4;
744 if (inputrec2nboundeddim(ir) != 3)
746 gmx_fatal(FARGS, "Can only convert old tpr files to the Verlet cut-off scheme with 3D pbc");
749 if (ir->efep != efepNO || ir->implicit_solvent != eisNO)
751 gmx_fatal(FARGS, "Will not convert old tpr files to the Verlet cut-off scheme with free-energy calculations or implicit solvent");
754 if (EI_DYNAMICS(ir->eI) && !(EI_MD(ir->eI) && ir->etc == etcNO))
756 verletbuf_list_setup_t ls;
758 verletbuf_get_list_setup(FALSE, &ls);
759 calc_verlet_buffer_size(mtop, box_vol, ir, ir->verletbuf_drift, &ls,
764 ir->verletbuf_drift = -1;
765 ir->rlist = 1.05*max(ir->rvdw, ir->rcoulomb);
768 gmx_mtop_remove_chargegroups(mtop);
771 static void print_hw_opt(FILE *fp, const gmx_hw_opt_t *hw_opt)
773 fprintf(fp, "hw_opt: nt %d ntmpi %d ntomp %d ntomp_pme %d gpu_id '%s'\n",
774 hw_opt->nthreads_tot,
775 hw_opt->nthreads_tmpi,
776 hw_opt->nthreads_omp,
777 hw_opt->nthreads_omp_pme,
778 hw_opt->gpu_opt.gpu_id != NULL ? hw_opt->gpu_opt.gpu_id : "");
781 /* Checks we can do when we don't (yet) know the cut-off scheme */
782 static void check_and_update_hw_opt_1(gmx_hw_opt_t *hw_opt,
783 gmx_bool bIsSimMaster)
785 gmx_omp_nthreads_read_env(&hw_opt->nthreads_omp, bIsSimMaster);
787 #ifndef GMX_THREAD_MPI
788 if (hw_opt->nthreads_tot > 0)
790 gmx_fatal(FARGS, "Setting the total number of threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
792 if (hw_opt->nthreads_tmpi > 0)
794 gmx_fatal(FARGS, "Setting the number of thread-MPI threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
799 if (hw_opt->nthreads_omp > 1)
801 gmx_fatal(FARGS, "More than 1 OpenMP thread requested, but Gromacs was compiled without OpenMP support");
803 hw_opt->nthreads_omp = 1;
806 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp_pme <= 0)
808 /* We have the same number of OpenMP threads for PP and PME processes,
809 * thus we can perform several consistency checks.
811 if (hw_opt->nthreads_tmpi > 0 &&
812 hw_opt->nthreads_omp > 0 &&
813 hw_opt->nthreads_tot != hw_opt->nthreads_tmpi*hw_opt->nthreads_omp)
815 gmx_fatal(FARGS, "The total number of threads requested (%d) does not match the thread-MPI threads (%d) times the OpenMP threads (%d) requested",
816 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi, hw_opt->nthreads_omp);
819 if (hw_opt->nthreads_tmpi > 0 &&
820 hw_opt->nthreads_tot % hw_opt->nthreads_tmpi != 0)
822 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of thread-MPI threads requested (%d)",
823 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi);
826 if (hw_opt->nthreads_omp > 0 &&
827 hw_opt->nthreads_tot % hw_opt->nthreads_omp != 0)
829 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of OpenMP threads requested (%d)",
830 hw_opt->nthreads_tot, hw_opt->nthreads_omp);
833 if (hw_opt->nthreads_tmpi > 0 &&
834 hw_opt->nthreads_omp <= 0)
836 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
841 if (hw_opt->nthreads_omp > 1)
843 gmx_fatal(FARGS, "OpenMP threads are requested, but Gromacs was compiled without OpenMP support");
847 if (hw_opt->nthreads_omp_pme > 0 && hw_opt->nthreads_omp <= 0)
849 gmx_fatal(FARGS, "You need to specify -ntomp in addition to -ntomp_pme");
852 if (hw_opt->nthreads_tot == 1)
854 hw_opt->nthreads_tmpi = 1;
856 if (hw_opt->nthreads_omp > 1)
858 gmx_fatal(FARGS, "You requested %d OpenMP threads with %d total threads",
859 hw_opt->nthreads_tmpi, hw_opt->nthreads_tot);
861 hw_opt->nthreads_omp = 1;
864 if (hw_opt->nthreads_omp_pme <= 0 && hw_opt->nthreads_omp > 0)
866 hw_opt->nthreads_omp_pme = hw_opt->nthreads_omp;
869 /* Parse GPU IDs, if provided.
870 * We check consistency with the tMPI thread count later.
872 gmx_parse_gpu_ids(&hw_opt->gpu_opt);
874 #ifdef GMX_THREAD_MPI
875 if (hw_opt->gpu_opt.ncuda_dev_use > 0 && hw_opt->nthreads_tmpi == 0)
877 /* Set the number of MPI threads equal to the number of GPUs */
878 hw_opt->nthreads_tmpi = hw_opt->gpu_opt.ncuda_dev_use;
880 if (hw_opt->nthreads_tot > 0 &&
881 hw_opt->nthreads_tmpi > hw_opt->nthreads_tot)
883 /* We have more GPUs than total threads requested.
884 * We choose to (later) generate a mismatch error,
885 * instead of launching more threads than requested.
887 hw_opt->nthreads_tmpi = hw_opt->nthreads_tot;
894 print_hw_opt(debug, hw_opt);
898 /* Checks we can do when we know the cut-off scheme */
899 static void check_and_update_hw_opt_2(gmx_hw_opt_t *hw_opt,
902 if (cutoff_scheme == ecutsGROUP)
904 /* We only have OpenMP support for PME only nodes */
905 if (hw_opt->nthreads_omp > 1)
907 gmx_fatal(FARGS, "OpenMP threads have been requested with cut-off scheme %s, but these are only supported with cut-off scheme %s",
908 ecutscheme_names[cutoff_scheme],
909 ecutscheme_names[ecutsVERLET]);
911 hw_opt->nthreads_omp = 1;
914 if (hw_opt->nthreads_omp_pme <= 0 && hw_opt->nthreads_omp > 0)
916 hw_opt->nthreads_omp_pme = hw_opt->nthreads_omp;
921 print_hw_opt(debug, hw_opt);
926 /* Override the value in inputrec with value passed on the command line (if any) */
927 static void override_nsteps_cmdline(FILE *fplog,
928 gmx_large_int_t nsteps_cmdline,
932 char sbuf[STEPSTRSIZE];
937 /* override with anything else than the default -2 */
938 if (nsteps_cmdline > -2)
942 ir->nsteps = nsteps_cmdline;
943 if (EI_DYNAMICS(ir->eI))
945 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps, %.3f ps",
946 gmx_step_str(nsteps_cmdline, sbuf),
947 nsteps_cmdline*ir->delta_t);
951 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps",
952 gmx_step_str(nsteps_cmdline, sbuf));
955 md_print_warn(cr, fplog, "%s\n", stmp);
959 /* Frees GPU memory and destroys the CUDA context.
961 * Note that this function needs to be called even if GPUs are not used
962 * in this run because the PME ranks have no knowledge of whether GPUs
963 * are used or not, but all ranks need to enter the barrier below.
965 static void free_gpu_resources(FILE *fplog,
966 const t_forcerec *fr,
969 gmx_bool bIsPPrankUsingGPU;
970 char gpu_err_str[STRLEN];
972 bIsPPrankUsingGPU = (cr->duty & DUTY_PP) && fr->nbv != NULL && fr->nbv->bUseGPU;
974 if (bIsPPrankUsingGPU)
976 /* free nbnxn data in GPU memory */
977 nbnxn_cuda_free(fplog, fr->nbv->cu_nbv);
979 /* With tMPI we need to wait for all ranks to finish deallocation before
980 * destroying the context in free_gpu() as some ranks may be sharing
982 * Note: as only PP ranks need to free GPU resources, so it is safe to
983 * not call the barrier on PME ranks.
985 #ifdef GMX_THREAD_MPI
990 #endif /* GMX_THREAD_MPI */
992 /* uninitialize GPU (by destroying the context) */
993 if (!free_gpu(gpu_err_str))
995 gmx_warning("On node %d failed to free GPU #%d: %s",
996 cr->nodeid, get_current_gpu_device_id(), gpu_err_str);
1001 int mdrunner(gmx_hw_opt_t *hw_opt,
1002 FILE *fplog, t_commrec *cr, int nfile,
1003 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
1004 gmx_bool bCompact, int nstglobalcomm,
1005 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
1006 const char *dddlb_opt, real dlb_scale,
1007 const char *ddcsx, const char *ddcsy, const char *ddcsz,
1008 const char *nbpu_opt,
1009 gmx_large_int_t nsteps_cmdline, int nstepout, int resetstep,
1010 int gmx_unused nmultisim, int repl_ex_nst, int repl_ex_nex,
1011 int repl_ex_seed, real pforce, real cpt_period, real max_hours,
1012 const char *deviceOptions, unsigned long Flags)
1014 gmx_bool bForceUseGPU, bTryUseGPU;
1015 double nodetime = 0, realtime;
1016 t_inputrec *inputrec;
1017 t_state *state = NULL;
1019 gmx_ddbox_t ddbox = {0};
1020 int npme_major, npme_minor;
1023 gmx_mtop_t *mtop = NULL;
1024 t_mdatoms *mdatoms = NULL;
1025 t_forcerec *fr = NULL;
1026 t_fcdata *fcd = NULL;
1027 real ewaldcoeff = 0;
1028 gmx_pme_t *pmedata = NULL;
1029 gmx_vsite_t *vsite = NULL;
1030 gmx_constr_t constr;
1031 int i, m, nChargePerturbed = -1, status, nalloc;
1033 gmx_wallcycle_t wcycle;
1034 gmx_bool bReadRNG, bReadEkin;
1036 gmx_walltime_accounting_t walltime_accounting = NULL;
1038 gmx_large_int_t reset_counters;
1039 gmx_edsam_t ed = NULL;
1040 t_commrec *cr_old = cr;
1041 int nthreads_pme = 1;
1042 int nthreads_pp = 1;
1043 gmx_membed_t membed = NULL;
1044 gmx_hw_info_t *hwinfo = NULL;
1045 /* The master rank decides early on bUseGPU and broadcasts this later */
1046 gmx_bool bUseGPU = FALSE;
1048 /* CAUTION: threads may be started later on in this function, so
1049 cr doesn't reflect the final parallel state right now */
1053 if (Flags & MD_APPENDFILES)
1058 bForceUseGPU = (strncmp(nbpu_opt, "gpu", 3) == 0);
1059 bTryUseGPU = (strncmp(nbpu_opt, "auto", 4) == 0) || bForceUseGPU;
1061 /* Detect hardware, gather information. This is an operation that is
1062 * global for this process (MPI rank). */
1063 hwinfo = gmx_detect_hardware(fplog, cr, bTryUseGPU);
1069 /* Read (nearly) all data required for the simulation */
1070 read_tpx_state(ftp2fn(efTPX, nfile, fnm), inputrec, state, NULL, mtop);
1072 if (inputrec->cutoff_scheme != ecutsVERLET &&
1073 ((Flags & MD_TESTVERLET) || getenv("GMX_VERLET_SCHEME") != NULL))
1075 convert_to_verlet_scheme(fplog, inputrec, mtop, det(state->box));
1078 if (inputrec->cutoff_scheme == ecutsVERLET)
1080 /* Here the master rank decides if all ranks will use GPUs */
1081 bUseGPU = (hwinfo->gpu_info.ncuda_dev_compatible > 0 ||
1082 getenv("GMX_EMULATE_GPU") != NULL);
1084 prepare_verlet_scheme(fplog, cr,
1085 inputrec, mtop, state->box,
1088 else if (hwinfo->gpu_info.ncuda_dev_compatible > 0)
1090 md_print_warn(cr, fplog,
1091 "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
1092 " To use a GPU, set the mdp option: cutoff-scheme = Verlet\n"
1093 " (for quick performance testing you can use the -testverlet option)\n");
1097 gmx_fatal(FARGS, "GPU requested, but can't be used without cutoff-scheme=Verlet");
1100 #ifdef GMX_TARGET_BGQ
1103 md_print_warn(cr, fplog,
1104 "NOTE: There is no SIMD implementation of the group scheme kernels on\n"
1105 " BlueGene/Q. You will observe better performance from using the\n"
1106 " Verlet cut-off scheme.\n");
1111 /* Check for externally set OpenMP affinity and turn off internal
1112 * pinning if any is found. We need to do this check early to tell
1113 * thread-MPI whether it should do pinning when spawning threads.
1114 * TODO: the above no longer holds, we should move these checks down
1116 gmx_omp_check_thread_affinity(fplog, cr, hw_opt);
1118 /* Check and update the hardware options for internal consistency */
1119 check_and_update_hw_opt_1(hw_opt, SIMMASTER(cr));
1123 #ifdef GMX_THREAD_MPI
1124 /* Early check for externally set process affinity.
1125 * With thread-MPI this is needed as pinning might get turned off,
1126 * which needs to be known before starting thread-MPI.
1127 * With thread-MPI hw_opt is processed here on the master rank
1128 * and passed to the other ranks later, so we only do this on master.
1130 gmx_check_thread_affinity_set(fplog,
1132 hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1135 #ifdef GMX_THREAD_MPI
1136 if (cr->npmenodes > 0 && hw_opt->nthreads_tmpi <= 0)
1138 gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME nodes");
1142 if (hw_opt->nthreads_omp_pme != hw_opt->nthreads_omp &&
1145 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");
1149 #ifdef GMX_THREAD_MPI
1152 /* Since the master knows the cut-off scheme, update hw_opt for this.
1153 * This is done later for normal MPI and also once more with tMPI
1154 * for all tMPI ranks.
1156 check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
1158 /* NOW the threads will be started: */
1159 hw_opt->nthreads_tmpi = get_nthreads_mpi(hwinfo,
1163 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp <= 0)
1165 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
1168 if (hw_opt->nthreads_tmpi > 1)
1170 /* now start the threads. */
1171 cr = mdrunner_start_threads(hw_opt, fplog, cr_old, nfile, fnm,
1172 oenv, bVerbose, bCompact, nstglobalcomm,
1173 ddxyz, dd_node_order, rdd, rconstr,
1174 dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
1176 nsteps_cmdline, nstepout, resetstep, nmultisim,
1177 repl_ex_nst, repl_ex_nex, repl_ex_seed, pforce,
1178 cpt_period, max_hours, deviceOptions,
1180 /* the main thread continues here with a new cr. We don't deallocate
1181 the old cr because other threads may still be reading it. */
1184 gmx_comm("Failed to spawn threads");
1189 /* END OF CAUTION: cr is now reliable */
1191 /* g_membed initialisation *
1192 * Because we change the mtop, init_membed is called before the init_parallel *
1193 * (in case we ever want to make it run in parallel) */
1194 if (opt2bSet("-membed", nfile, fnm))
1198 fprintf(stderr, "Initializing membed");
1200 membed = init_membed(fplog, nfile, fnm, mtop, inputrec, state, cr, &cpt_period);
1205 /* now broadcast everything to the non-master nodes/threads: */
1206 init_parallel(cr, inputrec, mtop);
1208 /* This check needs to happen after get_nthreads_mpi() */
1209 if (inputrec->cutoff_scheme == ecutsVERLET && (Flags & MD_PARTDEC))
1211 gmx_fatal_collective(FARGS, cr, NULL,
1212 "The Verlet cut-off scheme is not supported with particle decomposition.\n"
1213 "You can achieve the same effect as particle decomposition by running in parallel using only OpenMP threads.");
1218 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
1221 /* now make sure the state is initialized and propagated */
1222 set_state_entries(state, inputrec, cr->nnodes);
1224 /* A parallel command line option consistency check that we can
1225 only do after any threads have started. */
1227 (ddxyz[XX] > 1 || ddxyz[YY] > 1 || ddxyz[ZZ] > 1 || cr->npmenodes > 0))
1230 "The -dd or -npme option request a parallel simulation, "
1232 "but %s was compiled without threads or MPI enabled"
1234 #ifdef GMX_THREAD_MPI
1235 "but the number of threads (option -nt) is 1"
1237 "but %s was not started through mpirun/mpiexec or only one process was requested through mpirun/mpiexec"
1244 if ((Flags & MD_RERUN) &&
1245 (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
1247 gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
1250 if (can_use_allvsall(inputrec, TRUE, cr, fplog) && PAR(cr))
1252 /* Simple neighbour searching and (also?) all-vs-all loops
1253 * do not work with domain decomposition. */
1254 Flags |= MD_PARTDEC;
1257 if (!EEL_PME(inputrec->coulombtype) || (Flags & MD_PARTDEC))
1259 if (cr->npmenodes > 0)
1261 if (!EEL_PME(inputrec->coulombtype))
1263 gmx_fatal_collective(FARGS, cr, NULL,
1264 "PME nodes are requested, but the system does not use PME electrostatics");
1266 if (Flags & MD_PARTDEC)
1268 gmx_fatal_collective(FARGS, cr, NULL,
1269 "PME nodes are requested, but particle decomposition does not support separate PME nodes");
1277 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1280 /* NMR restraints must be initialized before load_checkpoint,
1281 * since with time averaging the history is added to t_state.
1282 * For proper consistency check we therefore need to extend
1284 * So the PME-only nodes (if present) will also initialize
1285 * the distance restraints.
1289 /* This needs to be called before read_checkpoint to extend the state */
1290 init_disres(fplog, mtop, inputrec, cr, Flags & MD_PARTDEC, fcd, state, repl_ex_nst > 0);
1292 if (gmx_mtop_ftype_count(mtop, F_ORIRES) > 0)
1294 if (PAR(cr) && !(Flags & MD_PARTDEC))
1296 gmx_fatal(FARGS, "Orientation restraints do not work (yet) with domain decomposition, use particle decomposition (mdrun option -pd)");
1298 /* Orientation restraints */
1301 init_orires(fplog, mtop, state->x, inputrec, cr->ms, &(fcd->orires),
1306 if (DEFORM(*inputrec))
1308 /* Store the deform reference box before reading the checkpoint */
1311 copy_mat(state->box, box);
1315 gmx_bcast(sizeof(box), box, cr);
1317 /* Because we do not have the update struct available yet
1318 * in which the reference values should be stored,
1319 * we store them temporarily in static variables.
1320 * This should be thread safe, since they are only written once
1321 * and with identical values.
1323 tMPI_Thread_mutex_lock(&deform_init_box_mutex);
1324 deform_init_init_step_tpx = inputrec->init_step;
1325 copy_mat(box, deform_init_box_tpx);
1326 tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
1329 if (opt2bSet("-cpi", nfile, fnm))
1331 /* Check if checkpoint file exists before doing continuation.
1332 * This way we can use identical input options for the first and subsequent runs...
1334 if (gmx_fexist_master(opt2fn_master("-cpi", nfile, fnm, cr), cr) )
1336 load_checkpoint(opt2fn_master("-cpi", nfile, fnm, cr), &fplog,
1337 cr, Flags & MD_PARTDEC, ddxyz,
1338 inputrec, state, &bReadRNG, &bReadEkin,
1339 (Flags & MD_APPENDFILES),
1340 (Flags & MD_APPENDFILESSET));
1344 Flags |= MD_READ_RNG;
1348 Flags |= MD_READ_EKIN;
1353 if (((MASTER(cr) || (Flags & MD_SEPPOT)) && (Flags & MD_APPENDFILES))
1354 #ifdef GMX_THREAD_MPI
1355 /* With thread MPI only the master node/thread exists in mdrun.c,
1356 * therefore non-master nodes need to open the "seppot" log file here.
1358 || (!MASTER(cr) && (Flags & MD_SEPPOT))
1362 gmx_log_open(ftp2fn(efLOG, nfile, fnm), cr, !(Flags & MD_SEPPOT),
1366 /* override nsteps with value from cmdline */
1367 override_nsteps_cmdline(fplog, nsteps_cmdline, inputrec, cr);
1371 copy_mat(state->box, box);
1376 gmx_bcast(sizeof(box), box, cr);
1379 /* Essential dynamics */
1380 if (opt2bSet("-ei", nfile, fnm))
1382 /* Open input and output files, allocate space for ED data structure */
1383 ed = ed_open(mtop->natoms, &state->edsamstate, nfile, fnm, Flags, oenv, cr);
1386 if (PAR(cr) && !((Flags & MD_PARTDEC) ||
1387 EI_TPI(inputrec->eI) ||
1388 inputrec->eI == eiNM))
1390 cr->dd = init_domain_decomposition(fplog, cr, Flags, ddxyz, rdd, rconstr,
1391 dddlb_opt, dlb_scale,
1392 ddcsx, ddcsy, ddcsz,
1395 &ddbox, &npme_major, &npme_minor);
1397 make_dd_communicators(fplog, cr, dd_node_order);
1399 /* Set overallocation to avoid frequent reallocation of arrays */
1400 set_over_alloc_dd(TRUE);
1404 /* PME, if used, is done on all nodes with 1D decomposition */
1406 cr->duty = (DUTY_PP | DUTY_PME);
1409 /* NM and TPI perform single node energy calculations in parallel */
1410 if (!(inputrec->eI == eiNM || EI_TPI(inputrec->eI)))
1412 npme_major = cr->nnodes;
1415 if (inputrec->ePBC == epbcSCREW)
1418 "pbc=%s is only implemented with domain decomposition",
1419 epbc_names[inputrec->ePBC]);
1425 /* After possible communicator splitting in make_dd_communicators.
1426 * we can set up the intra/inter node communication.
1428 gmx_setup_nodecomm(fplog, cr);
1431 /* Initialize per-physical-node MPI process/thread ID and counters. */
1432 gmx_init_intranode_counters(cr);
1435 md_print_info(cr, fplog, "Using %d MPI %s\n",
1437 #ifdef GMX_THREAD_MPI
1438 cr->nnodes == 1 ? "thread" : "threads"
1440 cr->nnodes == 1 ? "process" : "processes"
1446 /* Check and update hw_opt for the cut-off scheme */
1447 check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
1449 gmx_omp_nthreads_init(fplog, cr,
1450 hwinfo->nthreads_hw_avail,
1451 hw_opt->nthreads_omp,
1452 hw_opt->nthreads_omp_pme,
1453 (cr->duty & DUTY_PP) == 0,
1454 inputrec->cutoff_scheme == ecutsVERLET);
1458 /* The master rank decided on the use of GPUs,
1459 * broadcast this information to all ranks.
1461 gmx_bcast_sim(sizeof(bUseGPU), &bUseGPU, cr);
1466 if (cr->npmenodes == -1)
1468 /* Don't automatically use PME-only nodes with GPUs */
1472 /* Select GPU id's to use */
1473 gmx_select_gpu_ids(fplog, cr, &hwinfo->gpu_info, bForceUseGPU,
1478 /* Ignore (potentially) manually selected GPUs */
1479 hw_opt->gpu_opt.ncuda_dev_use = 0;
1482 /* check consistency of CPU acceleration and number of GPUs selected */
1483 gmx_check_hw_runconf_consistency(fplog, hwinfo, cr, hw_opt, bUseGPU);
1485 if (DOMAINDECOMP(cr))
1487 /* When we share GPUs over ranks, we need to know this for the DLB */
1488 dd_setup_dlb_resource_sharing(cr, hwinfo, hw_opt);
1491 /* getting number of PP/PME threads
1492 PME: env variable should be read only on one node to make sure it is
1493 identical everywhere;
1495 /* TODO nthreads_pp is only used for pinning threads.
1496 * This is a temporary solution until we have a hw topology library.
1498 nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
1499 nthreads_pme = gmx_omp_nthreads_get(emntPME);
1501 wcycle = wallcycle_init(fplog, resetstep, cr, nthreads_pp, nthreads_pme);
1505 /* Master synchronizes its value of reset_counters with all nodes
1506 * including PME only nodes */
1507 reset_counters = wcycle_get_reset_counters(wcycle);
1508 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1509 wcycle_set_reset_counters(wcycle, reset_counters);
1513 if (cr->duty & DUTY_PP)
1515 /* For domain decomposition we allocate dynamically
1516 * in dd_partition_system.
1518 if (DOMAINDECOMP(cr))
1520 bcast_state_setup(cr, state);
1526 bcast_state(cr, state, TRUE);
1530 /* Initiate forcerecord */
1532 fr->hwinfo = hwinfo;
1533 fr->gpu_opt = &hw_opt->gpu_opt;
1534 init_forcerec(fplog, oenv, fr, fcd, inputrec, mtop, cr, box,
1535 opt2fn("-table", nfile, fnm),
1536 opt2fn("-tabletf", nfile, fnm),
1537 opt2fn("-tablep", nfile, fnm),
1538 opt2fn("-tableb", nfile, fnm),
1543 /* version for PCA_NOT_READ_NODE (see md.c) */
1544 /*init_forcerec(fplog,fr,fcd,inputrec,mtop,cr,box,FALSE,
1545 "nofile","nofile","nofile","nofile",FALSE,pforce);
1547 fr->bSepDVDL = ((Flags & MD_SEPPOT) == MD_SEPPOT);
1549 /* Initialize QM-MM */
1552 init_QMMMrec(cr, mtop, inputrec, fr);
1555 /* Initialize the mdatoms structure.
1556 * mdatoms is not filled with atom data,
1557 * as this can not be done now with domain decomposition.
1559 mdatoms = init_mdatoms(fplog, mtop, inputrec->efep != efepNO);
1561 if (mdatoms->nPerturbed > 0 && inputrec->cutoff_scheme == ecutsVERLET)
1563 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.");
1566 /* Initialize the virtual site communication */
1567 vsite = init_vsite(mtop, cr, FALSE);
1569 calc_shifts(box, fr->shift_vec);
1571 /* With periodic molecules the charge groups should be whole at start up
1572 * and the virtual sites should not be far from their proper positions.
1574 if (!inputrec->bContinuation && MASTER(cr) &&
1575 !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1577 /* Make molecules whole at start of run */
1578 if (fr->ePBC != epbcNONE)
1580 do_pbc_first_mtop(fplog, inputrec->ePBC, box, mtop, state->x);
1584 /* Correct initial vsite positions are required
1585 * for the initial distribution in the domain decomposition
1586 * and for the initial shell prediction.
1588 construct_vsites_mtop(vsite, mtop, state->x);
1592 if (EEL_PME(fr->eeltype))
1594 ewaldcoeff = fr->ewaldcoeff;
1595 pmedata = &fr->pmedata;
1604 /* This is a PME only node */
1606 /* We don't need the state */
1609 ewaldcoeff = calc_ewaldcoeff(inputrec->rcoulomb, inputrec->ewald_rtol);
1613 if (hw_opt->thread_affinity != threadaffOFF)
1615 /* Before setting affinity, check whether the affinity has changed
1616 * - which indicates that probably the OpenMP library has changed it
1617 * since we first checked).
1619 gmx_check_thread_affinity_set(fplog, cr,
1620 hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1622 /* Set the CPU affinity */
1623 gmx_set_thread_affinity(fplog, cr, hw_opt, hwinfo);
1626 /* Initiate PME if necessary,
1627 * either on all nodes or on dedicated PME nodes only. */
1628 if (EEL_PME(inputrec->coulombtype))
1632 nChargePerturbed = mdatoms->nChargePerturbed;
1634 if (cr->npmenodes > 0)
1636 /* The PME only nodes need to know nChargePerturbed */
1637 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1640 if (cr->duty & DUTY_PME)
1642 status = gmx_pme_init(pmedata, cr, npme_major, npme_minor, inputrec,
1643 mtop ? mtop->natoms : 0, nChargePerturbed,
1644 (Flags & MD_REPRODUCIBLE), nthreads_pme);
1647 gmx_fatal(FARGS, "Error %d initializing PME", status);
1653 if (integrator[inputrec->eI].func == do_md)
1655 /* Turn on signal handling on all nodes */
1657 * (A user signal from the PME nodes (if any)
1658 * is communicated to the PP nodes.
1660 signal_handler_install();
1663 if (cr->duty & DUTY_PP)
1665 /* Assumes uniform use of the number of OpenMP threads */
1666 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1668 if (inputrec->ePull != epullNO)
1670 /* Initialize pull code */
1671 init_pull(fplog, inputrec, nfile, fnm, mtop, cr, oenv, inputrec->fepvals->init_lambda,
1672 EI_DYNAMICS(inputrec->eI) && MASTER(cr), Flags);
1677 /* Initialize enforced rotation code */
1678 init_rot(fplog, inputrec, nfile, fnm, cr, state->x, box, mtop, oenv,
1682 constr = init_constraints(fplog, mtop, inputrec, ed, state, cr);
1684 if (DOMAINDECOMP(cr))
1686 dd_init_bondeds(fplog, cr->dd, mtop, vsite, inputrec,
1687 Flags & MD_DDBONDCHECK, fr->cginfo_mb);
1689 set_dd_parameters(fplog, cr->dd, dlb_scale, inputrec, &ddbox);
1691 setup_dd_grid(fplog, cr->dd);
1694 /* Now do whatever the user wants us to do (how flexible...) */
1695 integrator[inputrec->eI].func(fplog, cr, nfile, fnm,
1696 oenv, bVerbose, bCompact,
1699 nstepout, inputrec, mtop,
1701 mdatoms, nrnb, wcycle, ed, fr,
1702 repl_ex_nst, repl_ex_nex, repl_ex_seed,
1704 cpt_period, max_hours,
1707 walltime_accounting);
1709 if (inputrec->ePull != epullNO)
1711 finish_pull(inputrec->pull);
1716 finish_rot(inputrec->rot);
1723 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1724 gmx_pmeonly(*pmedata, cr, nrnb, wcycle, walltime_accounting, ewaldcoeff, inputrec);
1727 wallcycle_stop(wcycle, ewcRUN);
1729 /* Finish up, write some stuff
1730 * if rerunMD, don't write last frame again
1732 finish_run(fplog, cr,
1733 inputrec, nrnb, wcycle, walltime_accounting,
1734 fr != NULL && fr->nbv != NULL && fr->nbv->bUseGPU ?
1735 nbnxn_cuda_get_timings(fr->nbv->cu_nbv) : NULL,
1736 EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
1739 /* Free GPU memory and context */
1740 free_gpu_resources(fplog, fr, cr);
1742 if (opt2bSet("-membed", nfile, fnm))
1747 gmx_hardware_info_free(hwinfo);
1749 /* Does what it says */
1750 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", walltime_accounting);
1751 walltime_accounting_destroy(walltime_accounting);
1753 /* Close logfile already here if we were appending to it */
1754 if (MASTER(cr) && (Flags & MD_APPENDFILES))
1756 gmx_log_close(fplog);
1759 rc = (int)gmx_get_stop_condition();
1761 #ifdef GMX_THREAD_MPI
1762 /* we need to join all threads. The sub-threads join when they
1763 exit this function, but the master thread needs to be told to
1765 if (PAR(cr) && MASTER(cr))