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) 2011,2012,2013,2014, 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.
50 #include "gromacs/legacyheaders/typedefs.h"
51 #include "gromacs/legacyheaders/oenv.h"
52 #include "gromacs/legacyheaders/force.h"
53 #include "gromacs/legacyheaders/mdrun.h"
54 #include "gromacs/legacyheaders/md_logging.h"
55 #include "gromacs/legacyheaders/md_support.h"
56 #include "gromacs/legacyheaders/network.h"
57 #include "gromacs/legacyheaders/names.h"
58 #include "gromacs/legacyheaders/disre.h"
59 #include "gromacs/legacyheaders/orires.h"
60 #include "gromacs/legacyheaders/pme.h"
61 #include "gromacs/legacyheaders/mdatoms.h"
64 #include "gromacs/legacyheaders/qmmm.h"
65 #include "gromacs/legacyheaders/domdec.h"
66 #include "gromacs/legacyheaders/coulomb.h"
67 #include "gromacs/legacyheaders/constr.h"
68 #include "gromacs/legacyheaders/mvdata.h"
69 #include "gromacs/legacyheaders/checkpoint.h"
70 #include "gromacs/topology/mtop_util.h"
71 #include "gromacs/legacyheaders/sighandler.h"
72 #include "gromacs/legacyheaders/txtdump.h"
73 #include "gromacs/legacyheaders/gmx_detect_hardware.h"
74 #include "gromacs/legacyheaders/gmx_omp_nthreads.h"
75 #include "gromacs/gmxpreprocess/calc_verletbuf.h"
77 #include "gromacs/legacyheaders/gmx_thread_affinity.h"
78 #include "gromacs/legacyheaders/inputrec.h"
79 #include "gromacs/legacyheaders/main.h"
81 #include "gromacs/essentialdynamics/edsam.h"
82 #include "gromacs/fileio/tpxio.h"
83 #include "gromacs/math/vec.h"
84 #include "gromacs/mdlib/nbnxn_search.h"
85 #include "gromacs/mdlib/nbnxn_consts.h"
86 #include "gromacs/pbcutil/pbc.h"
87 #include "gromacs/pulling/pull.h"
88 #include "gromacs/pulling/pull_rotation.h"
89 #include "gromacs/swap/swapcoords.h"
90 #include "gromacs/timing/wallcycle.h"
91 #include "gromacs/utility/gmxassert.h"
92 #include "gromacs/utility/gmxmpi.h"
93 #include "gromacs/utility/smalloc.h"
99 #include "gromacs/legacyheaders/gpu_utils.h"
102 gmx_integrator_t *func;
105 /* The array should match the eI array in include/types/enums.h */
106 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}};
108 gmx_int64_t deform_init_init_step_tpx;
109 matrix deform_init_box_tpx;
110 tMPI_Thread_mutex_t deform_init_box_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
113 #ifdef GMX_THREAD_MPI
114 /* The minimum number of atoms per tMPI thread. With fewer atoms than this,
115 * the number of threads will get lowered.
117 #define MIN_ATOMS_PER_MPI_THREAD 90
118 #define MIN_ATOMS_PER_GPU 900
120 struct mdrunner_arglist
135 const char *dddlb_opt;
140 const char *nbpu_opt;
142 gmx_int64_t nsteps_cmdline;
152 const char *deviceOptions;
158 /* The function used for spawning threads. Extracts the mdrunner()
159 arguments from its one argument and calls mdrunner(), after making
161 static void mdrunner_start_fn(void *arg)
163 struct mdrunner_arglist *mda = (struct mdrunner_arglist*)arg;
164 struct mdrunner_arglist mc = *mda; /* copy the arg list to make sure
165 that it's thread-local. This doesn't
166 copy pointed-to items, of course,
167 but those are all const. */
168 t_commrec *cr; /* we need a local version of this */
172 fnm = dup_tfn(mc.nfile, mc.fnm);
174 cr = reinitialize_commrec_for_this_thread(mc.cr);
181 mdrunner(&mc.hw_opt, fplog, cr, mc.nfile, fnm, mc.oenv,
182 mc.bVerbose, mc.bCompact, mc.nstglobalcomm,
183 mc.ddxyz, mc.dd_node_order, mc.rdd,
184 mc.rconstr, mc.dddlb_opt, mc.dlb_scale,
185 mc.ddcsx, mc.ddcsy, mc.ddcsz,
186 mc.nbpu_opt, mc.nstlist_cmdline,
187 mc.nsteps_cmdline, mc.nstepout, mc.resetstep,
188 mc.nmultisim, mc.repl_ex_nst, mc.repl_ex_nex, mc.repl_ex_seed, mc.pforce,
189 mc.cpt_period, mc.max_hours, mc.deviceOptions, mc.imdport, mc.Flags);
192 /* called by mdrunner() to start a specific number of threads (including
193 the main thread) for thread-parallel runs. This in turn calls mdrunner()
195 All options besides nthreads are the same as for mdrunner(). */
196 static t_commrec *mdrunner_start_threads(gmx_hw_opt_t *hw_opt,
197 FILE *fplog, t_commrec *cr, int nfile,
198 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
199 gmx_bool bCompact, int nstglobalcomm,
200 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
201 const char *dddlb_opt, real dlb_scale,
202 const char *ddcsx, const char *ddcsy, const char *ddcsz,
203 const char *nbpu_opt, int nstlist_cmdline,
204 gmx_int64_t nsteps_cmdline,
205 int nstepout, int resetstep,
206 int nmultisim, int repl_ex_nst, int repl_ex_nex, int repl_ex_seed,
207 real pforce, real cpt_period, real max_hours,
208 const char *deviceOptions, unsigned long Flags)
211 struct mdrunner_arglist *mda;
212 t_commrec *crn; /* the new commrec */
215 /* first check whether we even need to start tMPI */
216 if (hw_opt->nthreads_tmpi < 2)
221 /* a few small, one-time, almost unavoidable memory leaks: */
223 fnmn = dup_tfn(nfile, fnm);
225 /* fill the data structure to pass as void pointer to thread start fn */
226 /* hw_opt contains pointers, which should all be NULL at this stage */
227 mda->hw_opt = *hw_opt;
233 mda->bVerbose = bVerbose;
234 mda->bCompact = bCompact;
235 mda->nstglobalcomm = nstglobalcomm;
236 mda->ddxyz[XX] = ddxyz[XX];
237 mda->ddxyz[YY] = ddxyz[YY];
238 mda->ddxyz[ZZ] = ddxyz[ZZ];
239 mda->dd_node_order = dd_node_order;
241 mda->rconstr = rconstr;
242 mda->dddlb_opt = dddlb_opt;
243 mda->dlb_scale = dlb_scale;
247 mda->nbpu_opt = nbpu_opt;
248 mda->nstlist_cmdline = nstlist_cmdline;
249 mda->nsteps_cmdline = nsteps_cmdline;
250 mda->nstepout = nstepout;
251 mda->resetstep = resetstep;
252 mda->nmultisim = nmultisim;
253 mda->repl_ex_nst = repl_ex_nst;
254 mda->repl_ex_nex = repl_ex_nex;
255 mda->repl_ex_seed = repl_ex_seed;
256 mda->pforce = pforce;
257 mda->cpt_period = cpt_period;
258 mda->max_hours = max_hours;
259 mda->deviceOptions = deviceOptions;
262 /* now spawn new threads that start mdrunner_start_fn(), while
263 the main thread returns, we set thread affinity later */
264 ret = tMPI_Init_fn(TRUE, hw_opt->nthreads_tmpi, TMPI_AFFINITY_NONE,
265 mdrunner_start_fn, (void*)(mda) );
266 if (ret != TMPI_SUCCESS)
271 crn = reinitialize_commrec_for_this_thread(cr);
276 static int get_tmpi_omp_thread_division(const gmx_hw_info_t *hwinfo,
277 const gmx_hw_opt_t *hw_opt,
283 /* There are no separate PME nodes here, as we ensured in
284 * check_and_update_hw_opt that nthreads_tmpi>0 with PME nodes
285 * and a conditional ensures we would not have ended up here.
286 * Note that separate PME nodes might be switched on later.
290 nthreads_tmpi = ngpu;
291 if (nthreads_tot > 0 && nthreads_tot < nthreads_tmpi)
293 nthreads_tmpi = nthreads_tot;
296 else if (hw_opt->nthreads_omp > 0)
298 /* Here we could oversubscribe, when we do, we issue a warning later */
299 nthreads_tmpi = std::max(1, nthreads_tot/hw_opt->nthreads_omp);
303 /* TODO choose nthreads_omp based on hardware topology
304 when we have a hardware topology detection library */
305 /* In general, when running up to 4 threads, OpenMP should be faster.
306 * Note: on AMD Bulldozer we should avoid running OpenMP over two dies.
307 * On Intel>=Nehalem running OpenMP on a single CPU is always faster,
308 * even on two CPUs it's usually faster (but with many OpenMP threads
309 * it could be faster not to use HT, currently we always use HT).
310 * On Nehalem/Westmere we want to avoid running 16 threads over
311 * two CPUs with HT, so we need a limit<16; thus we use 12.
312 * A reasonable limit for Intel Sandy and Ivy bridge,
313 * not knowing the topology, is 16 threads.
315 const int nthreads_omp_always_faster = 4;
316 const int nthreads_omp_always_faster_Nehalem = 12;
317 const int nthreads_omp_always_faster_SandyBridge = 16;
318 gmx_bool bIntel_Family6;
321 (gmx_cpuid_vendor(hwinfo->cpuid_info) == GMX_CPUID_VENDOR_INTEL &&
322 gmx_cpuid_family(hwinfo->cpuid_info) == 6);
324 if (nthreads_tot <= nthreads_omp_always_faster ||
326 ((gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_Nehalem && nthreads_tot <= nthreads_omp_always_faster_Nehalem) ||
327 (gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_SandyBridge && nthreads_tot <= nthreads_omp_always_faster_SandyBridge))))
329 /* Use pure OpenMP parallelization */
334 /* Don't use OpenMP parallelization */
335 nthreads_tmpi = nthreads_tot;
339 return nthreads_tmpi;
343 /* Get the number of threads to use for thread-MPI based on how many
344 * were requested, which algorithms we're using,
345 * and how many particles there are.
346 * At the point we have already called check_and_update_hw_opt.
347 * Thus all options should be internally consistent and consistent
348 * with the hardware, except that ntmpi could be larger than #GPU.
350 static int get_nthreads_mpi(const gmx_hw_info_t *hwinfo,
351 gmx_hw_opt_t *hw_opt,
352 t_inputrec *inputrec, gmx_mtop_t *mtop,
356 int nthreads_hw, nthreads_tot_max, nthreads_tmpi, nthreads_new, ngpu;
357 int min_atoms_per_mpi_thread;
360 if (hw_opt->nthreads_tmpi > 0)
362 /* Trivial, return right away */
363 return hw_opt->nthreads_tmpi;
366 nthreads_hw = hwinfo->nthreads_hw_avail;
368 /* How many total (#tMPI*#OpenMP) threads can we start? */
369 if (hw_opt->nthreads_tot > 0)
371 nthreads_tot_max = hw_opt->nthreads_tot;
375 nthreads_tot_max = nthreads_hw;
378 bCanUseGPU = (inputrec->cutoff_scheme == ecutsVERLET &&
379 hwinfo->gpu_info.ncuda_dev_compatible > 0);
382 ngpu = hwinfo->gpu_info.ncuda_dev_compatible;
389 if (inputrec->cutoff_scheme == ecutsGROUP)
391 /* We checked this before, but it doesn't hurt to do it once more */
392 assert(hw_opt->nthreads_omp == 1);
396 get_tmpi_omp_thread_division(hwinfo, hw_opt, nthreads_tot_max, ngpu);
398 if (inputrec->eI == eiNM || EI_TPI(inputrec->eI))
400 /* Dims/steps are divided over the nodes iso splitting the atoms */
401 min_atoms_per_mpi_thread = 0;
407 min_atoms_per_mpi_thread = MIN_ATOMS_PER_GPU;
411 min_atoms_per_mpi_thread = MIN_ATOMS_PER_MPI_THREAD;
415 /* Check if an algorithm does not support parallel simulation. */
416 if (nthreads_tmpi != 1 &&
417 ( inputrec->eI == eiLBFGS ||
418 inputrec->coulombtype == eelEWALD ) )
422 md_print_warn(cr, fplog, "The integration or electrostatics algorithm doesn't support parallel runs. Using a single thread-MPI thread.\n");
423 if (hw_opt->nthreads_tmpi > nthreads_tmpi)
425 gmx_fatal(FARGS, "You asked for more than 1 thread-MPI thread, but an algorithm doesn't support that");
428 else if (mtop->natoms/nthreads_tmpi < min_atoms_per_mpi_thread)
430 /* the thread number was chosen automatically, but there are too many
431 threads (too few atoms per thread) */
432 nthreads_new = std::max(1, mtop->natoms/min_atoms_per_mpi_thread);
434 /* Avoid partial use of Hyper-Threading */
435 if (gmx_cpuid_x86_smt(hwinfo->cpuid_info) == GMX_CPUID_X86_SMT_ENABLED &&
436 nthreads_new > nthreads_hw/2 && nthreads_new < nthreads_hw)
438 nthreads_new = nthreads_hw/2;
441 /* Avoid large prime numbers in the thread count */
442 if (nthreads_new >= 6)
444 /* Use only 6,8,10 with additional factors of 2 */
448 while (3*fac*2 <= nthreads_new)
453 nthreads_new = (nthreads_new/fac)*fac;
458 if (nthreads_new == 5)
464 nthreads_tmpi = nthreads_new;
466 fprintf(stderr, "\n");
467 fprintf(stderr, "NOTE: Parallelization is limited by the small number of atoms,\n");
468 fprintf(stderr, " only starting %d thread-MPI threads.\n", nthreads_tmpi);
469 fprintf(stderr, " You can use the -nt and/or -ntmpi option to optimize the number of threads.\n\n");
472 return nthreads_tmpi;
474 #endif /* GMX_THREAD_MPI */
477 /* We determine the extra cost of the non-bonded kernels compared to
478 * a reference nstlist value of 10 (which is the default in grompp).
480 static const int nbnxnReferenceNstlist = 10;
481 /* The values to try when switching */
482 const int nstlist_try[] = { 20, 25, 40 };
483 #define NNSTL sizeof(nstlist_try)/sizeof(nstlist_try[0])
484 /* Increase nstlist until the non-bonded cost increases more than listfac_ok,
485 * but never more than listfac_max.
486 * A standard (protein+)water system at 300K with PME ewald_rtol=1e-5
487 * needs 1.28 at rcoulomb=0.9 and 1.24 at rcoulomb=1.0 to get to nstlist=40.
488 * Note that both CPU and GPU factors are conservative. Performance should
489 * not go down due to this tuning, except with a relatively slow GPU.
490 * On the other hand, at medium/high parallelization or with fast GPUs
491 * nstlist will not be increased enough to reach optimal performance.
493 /* CPU: pair-search is about a factor 1.5 slower than the non-bonded kernel */
494 static const float nbnxn_cpu_listfac_ok = 1.05;
495 static const float nbnxn_cpu_listfac_max = 1.09;
496 /* GPU: pair-search is a factor 1.5-3 slower than the non-bonded kernel */
497 static const float nbnxn_gpu_listfac_ok = 1.20;
498 static const float nbnxn_gpu_listfac_max = 1.30;
500 /* Try to increase nstlist when using the Verlet cut-off scheme */
501 static void increase_nstlist(FILE *fp, t_commrec *cr,
502 t_inputrec *ir, int nstlist_cmdline,
503 const gmx_mtop_t *mtop, matrix box,
506 float listfac_ok, listfac_max;
507 int nstlist_orig, nstlist_prev;
508 verletbuf_list_setup_t ls;
509 real rlistWithReferenceNstlist, rlist_inc, rlist_ok, rlist_max;
510 real rlist_new, rlist_prev;
511 size_t nstlist_ind = 0;
513 gmx_bool bBox, bDD, bCont;
514 const char *nstl_gpu = "\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";
515 const char *nve_err = "Can not increase nstlist because an NVE ensemble is used";
516 const char *vbd_err = "Can not increase nstlist because verlet-buffer-tolerance is not set or used";
517 const char *box_err = "Can not increase nstlist because the box is too small";
518 const char *dd_err = "Can not increase nstlist because of domain decomposition limitations";
520 const float oneThird = 1.0f / 3.0f;
522 if (nstlist_cmdline <= 0)
524 if (ir->nstlist == 1)
526 /* The user probably set nstlist=1 for a reason,
527 * don't mess with the settings.
532 if (fp != NULL && bGPU && ir->nstlist < nstlist_try[0])
534 fprintf(fp, nstl_gpu, ir->nstlist);
537 while (nstlist_ind < NNSTL && ir->nstlist >= nstlist_try[nstlist_ind])
541 if (nstlist_ind == NNSTL)
543 /* There are no larger nstlist value to try */
548 if (EI_MD(ir->eI) && ir->etc == etcNO)
552 fprintf(stderr, "%s\n", nve_err);
556 fprintf(fp, "%s\n", nve_err);
562 if (ir->verletbuf_tol == 0 && bGPU)
564 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");
567 if (ir->verletbuf_tol < 0)
571 fprintf(stderr, "%s\n", vbd_err);
575 fprintf(fp, "%s\n", vbd_err);
583 listfac_ok = nbnxn_gpu_listfac_ok;
584 listfac_max = nbnxn_gpu_listfac_max;
588 listfac_ok = nbnxn_cpu_listfac_ok;
589 listfac_max = nbnxn_cpu_listfac_max;
592 nstlist_orig = ir->nstlist;
593 if (nstlist_cmdline > 0)
597 sprintf(buf, "Getting nstlist=%d from command line option",
600 ir->nstlist = nstlist_cmdline;
603 verletbuf_get_list_setup(bGPU, &ls);
605 /* Allow rlist to make the list a given factor larger than the list
606 * would be with the reference value for nstlist (10).
608 nstlist_prev = ir->nstlist;
609 ir->nstlist = nbnxnReferenceNstlist;
610 calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL,
611 &rlistWithReferenceNstlist);
612 ir->nstlist = nstlist_prev;
614 /* Determine the pair list size increase due to zero interactions */
615 rlist_inc = nbnxn_get_rlist_effective_inc(ls.cluster_size_j,
616 mtop->natoms/det(box));
617 rlist_ok = (rlistWithReferenceNstlist + rlist_inc)*pow(listfac_ok, oneThird) - rlist_inc;
618 rlist_max = (rlistWithReferenceNstlist + rlist_inc)*pow(listfac_max, oneThird) - rlist_inc;
621 fprintf(debug, "nstlist tuning: rlist_inc %.3f rlist_ok %.3f rlist_max %.3f\n",
622 rlist_inc, rlist_ok, rlist_max);
625 nstlist_prev = nstlist_orig;
626 rlist_prev = ir->rlist;
629 if (nstlist_cmdline <= 0)
631 ir->nstlist = nstlist_try[nstlist_ind];
634 /* Set the pair-list buffer size in ir */
635 calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
637 /* Does rlist fit in the box? */
638 bBox = (sqr(rlist_new) < max_cutoff2(ir->ePBC, box));
640 if (bBox && DOMAINDECOMP(cr))
642 /* Check if rlist fits in the domain decomposition */
643 if (inputrec2nboundeddim(ir) < DIM)
645 gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
647 copy_mat(box, state_tmp.box);
648 bDD = change_dd_cutoff(cr, &state_tmp, ir, rlist_new);
653 fprintf(debug, "nstlist %d rlist %.3f bBox %d bDD %d\n",
654 ir->nstlist, rlist_new, bBox, bDD);
659 if (nstlist_cmdline <= 0)
661 if (bBox && bDD && rlist_new <= rlist_max)
663 /* Increase nstlist */
664 nstlist_prev = ir->nstlist;
665 rlist_prev = rlist_new;
666 bCont = (nstlist_ind+1 < NNSTL && rlist_new < rlist_ok);
670 /* Stick with the previous nstlist */
671 ir->nstlist = nstlist_prev;
672 rlist_new = rlist_prev;
684 gmx_warning(!bBox ? box_err : dd_err);
687 fprintf(fp, "\n%s\n", bBox ? box_err : dd_err);
689 ir->nstlist = nstlist_orig;
691 else if (ir->nstlist != nstlist_orig || rlist_new != ir->rlist)
693 sprintf(buf, "Changing nstlist from %d to %d, rlist from %g to %g",
694 nstlist_orig, ir->nstlist,
695 ir->rlist, rlist_new);
698 fprintf(stderr, "%s\n\n", buf);
702 fprintf(fp, "%s\n\n", buf);
704 ir->rlist = rlist_new;
705 ir->rlistlong = rlist_new;
709 static void prepare_verlet_scheme(FILE *fplog,
713 const gmx_mtop_t *mtop,
717 /* For NVE simulations, we will retain the initial list buffer */
718 if (ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
720 /* Update the Verlet buffer size for the current run setup */
721 verletbuf_list_setup_t ls;
724 /* Here we assume SIMD-enabled kernels are being used. But as currently
725 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
726 * and 4x2 gives a larger buffer than 4x4, this is ok.
728 verletbuf_get_list_setup(bUseGPU, &ls);
730 calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
732 if (rlist_new != ir->rlist)
736 fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
737 ir->rlist, rlist_new,
738 ls.cluster_size_i, ls.cluster_size_j);
740 ir->rlist = rlist_new;
741 ir->rlistlong = rlist_new;
745 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
747 gmx_fatal(FARGS, "Can not set nstlist without %s",
748 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
751 if (EI_DYNAMICS(ir->eI))
753 /* Set or try nstlist values */
754 increase_nstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, bUseGPU);
758 static void convert_to_verlet_scheme(FILE *fplog,
760 gmx_mtop_t *mtop, real box_vol)
762 const char *conv_mesg = "Converting input file with group cut-off scheme to the Verlet cut-off scheme";
764 md_print_warn(NULL, fplog, "%s\n", conv_mesg);
766 ir->cutoff_scheme = ecutsVERLET;
767 ir->verletbuf_tol = 0.005;
769 if (ir->rcoulomb != ir->rvdw)
771 gmx_fatal(FARGS, "The VdW and Coulomb cut-offs are different, whereas the Verlet scheme only supports equal cut-offs");
774 if (ir->vdwtype == evdwUSER || EEL_USER(ir->coulombtype))
776 gmx_fatal(FARGS, "User non-bonded potentials are not (yet) supported with the Verlet scheme");
778 else if (ir_vdw_switched(ir) || ir_coulomb_switched(ir))
780 if (ir_vdw_switched(ir) && ir->vdw_modifier == eintmodNONE)
782 ir->vdwtype = evdwCUT;
786 case evdwSHIFT: ir->vdw_modifier = eintmodFORCESWITCH; break;
787 case evdwSWITCH: ir->vdw_modifier = eintmodPOTSWITCH; break;
788 default: gmx_fatal(FARGS, "The Verlet scheme does not support Van der Waals interactions of type '%s'", evdw_names[ir->vdwtype]);
791 if (ir_coulomb_switched(ir) && ir->coulomb_modifier == eintmodNONE)
793 if (EEL_FULL(ir->coulombtype))
795 /* With full electrostatic only PME can be switched */
796 ir->coulombtype = eelPME;
797 ir->coulomb_modifier = eintmodPOTSHIFT;
801 md_print_warn(NULL, fplog, "NOTE: Replacing %s electrostatics with reaction-field with epsilon-rf=inf\n", eel_names[ir->coulombtype]);
802 ir->coulombtype = eelRF;
803 ir->epsilon_rf = 0.0;
804 ir->coulomb_modifier = eintmodPOTSHIFT;
808 /* We set the pair energy error tolerance to a small number.
809 * Note that this is only for testing. For production the user
810 * should think about this and set the mdp options.
812 ir->verletbuf_tol = 1e-4;
815 if (inputrec2nboundeddim(ir) != 3)
817 gmx_fatal(FARGS, "Can only convert old tpr files to the Verlet cut-off scheme with 3D pbc");
820 if (ir->efep != efepNO || ir->implicit_solvent != eisNO)
822 gmx_fatal(FARGS, "Will not convert old tpr files to the Verlet cut-off scheme with free-energy calculations or implicit solvent");
825 if (EI_DYNAMICS(ir->eI) && !(EI_MD(ir->eI) && ir->etc == etcNO))
827 verletbuf_list_setup_t ls;
829 verletbuf_get_list_setup(FALSE, &ls);
830 calc_verlet_buffer_size(mtop, box_vol, ir, -1, &ls, NULL, &ir->rlist);
838 rlist_fac = 1 + verlet_buffer_ratio_NVE_T0;
842 rlist_fac = 1 + verlet_buffer_ratio_nodynamics;
844 ir->verletbuf_tol = -1;
845 ir->rlist = rlist_fac*std::max(ir->rvdw, ir->rcoulomb);
848 gmx_mtop_remove_chargegroups(mtop);
851 static void print_hw_opt(FILE *fp, const gmx_hw_opt_t *hw_opt)
853 fprintf(fp, "hw_opt: nt %d ntmpi %d ntomp %d ntomp_pme %d gpu_id '%s'\n",
854 hw_opt->nthreads_tot,
855 hw_opt->nthreads_tmpi,
856 hw_opt->nthreads_omp,
857 hw_opt->nthreads_omp_pme,
858 hw_opt->gpu_opt.gpu_id != NULL ? hw_opt->gpu_opt.gpu_id : "");
861 /* Checks we can do when we don't (yet) know the cut-off scheme */
862 static void check_and_update_hw_opt_1(gmx_hw_opt_t *hw_opt,
863 gmx_bool bIsSimMaster)
865 gmx_omp_nthreads_read_env(&hw_opt->nthreads_omp, bIsSimMaster);
867 #ifndef GMX_THREAD_MPI
868 if (hw_opt->nthreads_tot > 0)
870 gmx_fatal(FARGS, "Setting the total number of threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
872 if (hw_opt->nthreads_tmpi > 0)
874 gmx_fatal(FARGS, "Setting the number of thread-MPI threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
879 if (hw_opt->nthreads_omp > 1)
881 gmx_fatal(FARGS, "More than 1 OpenMP thread requested, but Gromacs was compiled without OpenMP support");
883 hw_opt->nthreads_omp = 1;
886 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp_pme <= 0)
888 /* We have the same number of OpenMP threads for PP and PME processes,
889 * thus we can perform several consistency checks.
891 if (hw_opt->nthreads_tmpi > 0 &&
892 hw_opt->nthreads_omp > 0 &&
893 hw_opt->nthreads_tot != hw_opt->nthreads_tmpi*hw_opt->nthreads_omp)
895 gmx_fatal(FARGS, "The total number of threads requested (%d) does not match the thread-MPI threads (%d) times the OpenMP threads (%d) requested",
896 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi, hw_opt->nthreads_omp);
899 if (hw_opt->nthreads_tmpi > 0 &&
900 hw_opt->nthreads_tot % hw_opt->nthreads_tmpi != 0)
902 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of thread-MPI threads requested (%d)",
903 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi);
906 if (hw_opt->nthreads_omp > 0 &&
907 hw_opt->nthreads_tot % hw_opt->nthreads_omp != 0)
909 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of OpenMP threads requested (%d)",
910 hw_opt->nthreads_tot, hw_opt->nthreads_omp);
913 if (hw_opt->nthreads_tmpi > 0 &&
914 hw_opt->nthreads_omp <= 0)
916 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
921 if (hw_opt->nthreads_omp > 1)
923 gmx_fatal(FARGS, "OpenMP threads are requested, but Gromacs was compiled without OpenMP support");
927 if (hw_opt->nthreads_omp_pme > 0 && hw_opt->nthreads_omp <= 0)
929 gmx_fatal(FARGS, "You need to specify -ntomp in addition to -ntomp_pme");
932 if (hw_opt->nthreads_tot == 1)
934 hw_opt->nthreads_tmpi = 1;
936 if (hw_opt->nthreads_omp > 1)
938 gmx_fatal(FARGS, "You requested %d OpenMP threads with %d total threads",
939 hw_opt->nthreads_tmpi, hw_opt->nthreads_tot);
941 hw_opt->nthreads_omp = 1;
944 if (hw_opt->nthreads_omp_pme <= 0 && hw_opt->nthreads_omp > 0)
946 hw_opt->nthreads_omp_pme = hw_opt->nthreads_omp;
949 /* Parse GPU IDs, if provided.
950 * We check consistency with the tMPI thread count later.
952 gmx_parse_gpu_ids(&hw_opt->gpu_opt);
954 #ifdef GMX_THREAD_MPI
955 if (hw_opt->gpu_opt.ncuda_dev_use > 0 && hw_opt->nthreads_tmpi == 0)
957 /* Set the number of MPI threads equal to the number of GPUs */
958 hw_opt->nthreads_tmpi = hw_opt->gpu_opt.ncuda_dev_use;
960 if (hw_opt->nthreads_tot > 0 &&
961 hw_opt->nthreads_tmpi > hw_opt->nthreads_tot)
963 /* We have more GPUs than total threads requested.
964 * We choose to (later) generate a mismatch error,
965 * instead of launching more threads than requested.
967 hw_opt->nthreads_tmpi = hw_opt->nthreads_tot;
974 print_hw_opt(debug, hw_opt);
978 /* Checks we can do when we know the cut-off scheme */
979 static void check_and_update_hw_opt_2(gmx_hw_opt_t *hw_opt,
982 if (cutoff_scheme == ecutsGROUP)
984 /* We only have OpenMP support for PME only nodes */
985 if (hw_opt->nthreads_omp > 1)
987 gmx_fatal(FARGS, "OpenMP threads have been requested with cut-off scheme %s, but these are only supported with cut-off scheme %s",
988 ecutscheme_names[cutoff_scheme],
989 ecutscheme_names[ecutsVERLET]);
991 hw_opt->nthreads_omp = 1;
994 if (hw_opt->nthreads_omp_pme <= 0 && hw_opt->nthreads_omp > 0)
996 hw_opt->nthreads_omp_pme = hw_opt->nthreads_omp;
1001 print_hw_opt(debug, hw_opt);
1006 /* Override the value in inputrec with value passed on the command line (if any) */
1007 static void override_nsteps_cmdline(FILE *fplog,
1008 gmx_int64_t nsteps_cmdline,
1010 const t_commrec *cr)
1012 char sbuf[STEPSTRSIZE];
1017 /* override with anything else than the default -2 */
1018 if (nsteps_cmdline > -2)
1022 ir->nsteps = nsteps_cmdline;
1023 if (EI_DYNAMICS(ir->eI))
1025 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps, %.3f ps",
1026 gmx_step_str(nsteps_cmdline, sbuf),
1027 nsteps_cmdline*ir->delta_t);
1031 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps",
1032 gmx_step_str(nsteps_cmdline, sbuf));
1035 md_print_warn(cr, fplog, "%s\n", stmp);
1039 int mdrunner(gmx_hw_opt_t *hw_opt,
1040 FILE *fplog, t_commrec *cr, int nfile,
1041 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
1042 gmx_bool bCompact, int nstglobalcomm,
1043 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
1044 const char *dddlb_opt, real dlb_scale,
1045 const char *ddcsx, const char *ddcsy, const char *ddcsz,
1046 const char *nbpu_opt, int nstlist_cmdline,
1047 gmx_int64_t nsteps_cmdline, int nstepout, int resetstep,
1048 int gmx_unused nmultisim, int repl_ex_nst, int repl_ex_nex,
1049 int repl_ex_seed, real pforce, real cpt_period, real max_hours,
1050 const char *deviceOptions, int imdport, unsigned long Flags)
1052 gmx_bool bForceUseGPU, bTryUseGPU;
1053 t_inputrec *inputrec;
1054 t_state *state = NULL;
1056 gmx_ddbox_t ddbox = {0};
1057 int npme_major, npme_minor;
1059 gmx_mtop_t *mtop = NULL;
1060 t_mdatoms *mdatoms = NULL;
1061 t_forcerec *fr = NULL;
1062 t_fcdata *fcd = NULL;
1063 real ewaldcoeff_q = 0;
1064 real ewaldcoeff_lj = 0;
1065 gmx_pme_t *pmedata = NULL;
1066 gmx_vsite_t *vsite = NULL;
1067 gmx_constr_t constr;
1068 int nChargePerturbed = -1, nTypePerturbed = 0, status;
1069 gmx_wallcycle_t wcycle;
1071 gmx_walltime_accounting_t walltime_accounting = NULL;
1073 gmx_int64_t reset_counters;
1074 gmx_edsam_t ed = NULL;
1075 int nthreads_pme = 1;
1076 int nthreads_pp = 1;
1077 gmx_membed_t membed = NULL;
1078 gmx_hw_info_t *hwinfo = NULL;
1079 /* The master rank decides early on bUseGPU and broadcasts this later */
1080 gmx_bool bUseGPU = FALSE;
1082 /* CAUTION: threads may be started later on in this function, so
1083 cr doesn't reflect the final parallel state right now */
1087 if (Flags & MD_APPENDFILES)
1092 bForceUseGPU = (strncmp(nbpu_opt, "gpu", 3) == 0);
1093 bTryUseGPU = (strncmp(nbpu_opt, "auto", 4) == 0) || bForceUseGPU;
1095 /* Detect hardware, gather information. This is an operation that is
1096 * global for this process (MPI rank). */
1097 hwinfo = gmx_detect_hardware(fplog, cr, bTryUseGPU);
1103 /* Read (nearly) all data required for the simulation */
1104 read_tpx_state(ftp2fn(efTPX, nfile, fnm), inputrec, state, NULL, mtop);
1106 if (inputrec->cutoff_scheme != ecutsVERLET &&
1107 ((Flags & MD_TESTVERLET) || getenv("GMX_VERLET_SCHEME") != NULL))
1109 convert_to_verlet_scheme(fplog, inputrec, mtop, det(state->box));
1112 if (inputrec->cutoff_scheme == ecutsVERLET)
1114 /* Here the master rank decides if all ranks will use GPUs */
1115 bUseGPU = (hwinfo->gpu_info.ncuda_dev_compatible > 0 ||
1116 getenv("GMX_EMULATE_GPU") != NULL);
1118 /* TODO add GPU kernels for this and replace this check by:
1119 * (bUseGPU && (ir->vdwtype == evdwPME &&
1120 * ir->ljpme_combination_rule == eljpmeLB))
1121 * update the message text and the content of nbnxn_acceleration_supported.
1124 !nbnxn_acceleration_supported(fplog, cr, inputrec, bUseGPU))
1126 /* Fallback message printed by nbnxn_acceleration_supported */
1129 gmx_fatal(FARGS, "GPU acceleration requested, but not supported with the given input settings");
1134 prepare_verlet_scheme(fplog, cr,
1135 inputrec, nstlist_cmdline, mtop, state->box,
1140 if (nstlist_cmdline > 0)
1142 gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
1145 if (hwinfo->gpu_info.ncuda_dev_compatible > 0)
1147 md_print_warn(cr, fplog,
1148 "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
1149 " To use a GPU, set the mdp option: cutoff-scheme = Verlet\n"
1150 " (for quick performance testing you can use the -testverlet option)\n");
1155 gmx_fatal(FARGS, "GPU requested, but can't be used without cutoff-scheme=Verlet");
1158 #ifdef GMX_TARGET_BGQ
1159 md_print_warn(cr, fplog,
1160 "NOTE: There is no SIMD implementation of the group scheme kernels on\n"
1161 " BlueGene/Q. You will observe better performance from using the\n"
1162 " Verlet cut-off scheme.\n");
1166 if (inputrec->eI == eiSD2)
1168 md_print_warn(cr, fplog, "The stochastic dynamics integrator %s is deprecated, since\n"
1169 "it is slower than integrator %s and is slightly less accurate\n"
1170 "with constraints. Use the %s integrator.",
1171 ei_names[inputrec->eI], ei_names[eiSD1], ei_names[eiSD1]);
1175 /* Check and update the hardware options for internal consistency */
1176 check_and_update_hw_opt_1(hw_opt, SIMMASTER(cr));
1178 /* Early check for externally set process affinity. */
1179 gmx_check_thread_affinity_set(fplog, cr,
1180 hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1184 #ifdef GMX_THREAD_MPI
1185 if (cr->npmenodes > 0 && hw_opt->nthreads_tmpi <= 0)
1187 gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME ranks");
1191 if (hw_opt->nthreads_omp_pme != hw_opt->nthreads_omp &&
1194 gmx_fatal(FARGS, "You need to explicitly specify the number of PME ranks (-npme) when using different number of OpenMP threads for PP and PME ranks");
1198 #ifdef GMX_THREAD_MPI
1201 /* Since the master knows the cut-off scheme, update hw_opt for this.
1202 * This is done later for normal MPI and also once more with tMPI
1203 * for all tMPI ranks.
1205 check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
1207 /* NOW the threads will be started: */
1208 hw_opt->nthreads_tmpi = get_nthreads_mpi(hwinfo,
1212 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp <= 0)
1214 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
1217 if (hw_opt->nthreads_tmpi > 1)
1219 t_commrec *cr_old = cr;
1220 /* now start the threads. */
1221 cr = mdrunner_start_threads(hw_opt, fplog, cr_old, nfile, fnm,
1222 oenv, bVerbose, bCompact, nstglobalcomm,
1223 ddxyz, dd_node_order, rdd, rconstr,
1224 dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
1225 nbpu_opt, nstlist_cmdline,
1226 nsteps_cmdline, nstepout, resetstep, nmultisim,
1227 repl_ex_nst, repl_ex_nex, repl_ex_seed, pforce,
1228 cpt_period, max_hours, deviceOptions,
1230 /* the main thread continues here with a new cr. We don't deallocate
1231 the old cr because other threads may still be reading it. */
1234 gmx_comm("Failed to spawn threads");
1239 /* END OF CAUTION: cr is now reliable */
1241 /* g_membed initialisation *
1242 * Because we change the mtop, init_membed is called before the init_parallel *
1243 * (in case we ever want to make it run in parallel) */
1244 if (opt2bSet("-membed", nfile, fnm))
1248 fprintf(stderr, "Initializing membed");
1250 membed = init_membed(fplog, nfile, fnm, mtop, inputrec, state, cr, &cpt_period);
1255 /* now broadcast everything to the non-master nodes/threads: */
1256 init_parallel(cr, inputrec, mtop);
1260 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
1263 /* now make sure the state is initialized and propagated */
1264 set_state_entries(state, inputrec);
1266 /* A parallel command line option consistency check that we can
1267 only do after any threads have started. */
1269 (ddxyz[XX] > 1 || ddxyz[YY] > 1 || ddxyz[ZZ] > 1 || cr->npmenodes > 0))
1272 "The -dd or -npme option request a parallel simulation, "
1274 "but %s was compiled without threads or MPI enabled"
1276 #ifdef GMX_THREAD_MPI
1277 "but the number of threads (option -nt) is 1"
1279 "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec"
1282 , output_env_get_program_display_name(oenv)
1286 if ((Flags & MD_RERUN) &&
1287 (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
1289 gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
1292 if (can_use_allvsall(inputrec, TRUE, cr, fplog) && DOMAINDECOMP(cr))
1294 gmx_fatal(FARGS, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
1297 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
1299 if (cr->npmenodes > 0)
1301 gmx_fatal_collective(FARGS, cr, NULL,
1302 "PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
1311 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1315 /* NMR restraints must be initialized before load_checkpoint,
1316 * since with time averaging the history is added to t_state.
1317 * For proper consistency check we therefore need to extend
1319 * So the PME-only nodes (if present) will also initialize
1320 * the distance restraints.
1324 /* This needs to be called before read_checkpoint to extend the state */
1325 init_disres(fplog, mtop, inputrec, cr, fcd, state, repl_ex_nst > 0);
1327 init_orires(fplog, mtop, state->x, inputrec, cr, &(fcd->orires),
1330 if (DEFORM(*inputrec))
1332 /* Store the deform reference box before reading the checkpoint */
1335 copy_mat(state->box, box);
1339 gmx_bcast(sizeof(box), box, cr);
1341 /* Because we do not have the update struct available yet
1342 * in which the reference values should be stored,
1343 * we store them temporarily in static variables.
1344 * This should be thread safe, since they are only written once
1345 * and with identical values.
1347 tMPI_Thread_mutex_lock(&deform_init_box_mutex);
1348 deform_init_init_step_tpx = inputrec->init_step;
1349 copy_mat(box, deform_init_box_tpx);
1350 tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
1353 if (opt2bSet("-cpi", nfile, fnm))
1355 /* Check if checkpoint file exists before doing continuation.
1356 * This way we can use identical input options for the first and subsequent runs...
1358 if (gmx_fexist_master(opt2fn_master("-cpi", nfile, fnm, cr), cr) )
1360 load_checkpoint(opt2fn_master("-cpi", nfile, fnm, cr), &fplog,
1362 inputrec, state, &bReadEkin,
1363 (Flags & MD_APPENDFILES),
1364 (Flags & MD_APPENDFILESSET));
1368 Flags |= MD_READ_EKIN;
1373 if (MASTER(cr) && (Flags & MD_APPENDFILES))
1375 gmx_log_open(ftp2fn(efLOG, nfile, fnm), cr,
1379 /* override nsteps with value from cmdline */
1380 override_nsteps_cmdline(fplog, nsteps_cmdline, inputrec, cr);
1384 copy_mat(state->box, box);
1389 gmx_bcast(sizeof(box), box, cr);
1392 /* Essential dynamics */
1393 if (opt2bSet("-ei", nfile, fnm))
1395 /* Open input and output files, allocate space for ED data structure */
1396 ed = ed_open(mtop->natoms, &state->edsamstate, nfile, fnm, Flags, oenv, cr);
1399 if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
1400 inputrec->eI == eiNM))
1402 cr->dd = init_domain_decomposition(fplog, cr, Flags, ddxyz, rdd, rconstr,
1403 dddlb_opt, dlb_scale,
1404 ddcsx, ddcsy, ddcsz,
1407 &ddbox, &npme_major, &npme_minor);
1409 make_dd_communicators(fplog, cr, dd_node_order);
1411 /* Set overallocation to avoid frequent reallocation of arrays */
1412 set_over_alloc_dd(TRUE);
1416 /* PME, if used, is done on all nodes with 1D decomposition */
1418 cr->duty = (DUTY_PP | DUTY_PME);
1422 if (inputrec->ePBC == epbcSCREW)
1425 "pbc=%s is only implemented with domain decomposition",
1426 epbc_names[inputrec->ePBC]);
1432 /* After possible communicator splitting in make_dd_communicators.
1433 * we can set up the intra/inter node communication.
1435 gmx_setup_nodecomm(fplog, cr);
1438 /* Initialize per-physical-node MPI process/thread ID and counters. */
1439 gmx_init_intranode_counters(cr);
1443 md_print_info(cr, fplog,
1444 "This is simulation %d out of %d running as a composite Gromacs\n"
1445 "multi-simulation job. Setup for this simulation:\n\n",
1446 cr->ms->sim, cr->ms->nsim);
1448 md_print_info(cr, fplog, "Using %d MPI %s\n",
1450 #ifdef GMX_THREAD_MPI
1451 cr->nnodes == 1 ? "thread" : "threads"
1453 cr->nnodes == 1 ? "process" : "processes"
1459 /* Check and update hw_opt for the cut-off scheme */
1460 check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
1462 gmx_omp_nthreads_init(fplog, cr,
1463 hwinfo->nthreads_hw_avail,
1464 hw_opt->nthreads_omp,
1465 hw_opt->nthreads_omp_pme,
1466 (cr->duty & DUTY_PP) == 0,
1467 inputrec->cutoff_scheme == ecutsVERLET);
1471 /* The master rank decided on the use of GPUs,
1472 * broadcast this information to all ranks.
1474 gmx_bcast_sim(sizeof(bUseGPU), &bUseGPU, cr);
1479 if (cr->npmenodes == -1)
1481 /* Don't automatically use PME-only nodes with GPUs */
1485 /* Select GPU id's to use */
1486 gmx_select_gpu_ids(fplog, cr, &hwinfo->gpu_info, bForceUseGPU,
1491 /* Ignore (potentially) manually selected GPUs */
1492 hw_opt->gpu_opt.ncuda_dev_use = 0;
1495 /* check consistency across ranks of things like SIMD
1496 * support and number of GPUs selected */
1497 gmx_check_hw_runconf_consistency(fplog, hwinfo, cr, hw_opt, bUseGPU);
1499 if (DOMAINDECOMP(cr))
1501 /* When we share GPUs over ranks, we need to know this for the DLB */
1502 dd_setup_dlb_resource_sharing(cr, hwinfo, hw_opt);
1505 /* getting number of PP/PME threads
1506 PME: env variable should be read only on one node to make sure it is
1507 identical everywhere;
1509 /* TODO nthreads_pp is only used for pinning threads.
1510 * This is a temporary solution until we have a hw topology library.
1512 nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
1513 nthreads_pme = gmx_omp_nthreads_get(emntPME);
1515 wcycle = wallcycle_init(fplog, resetstep, cr, nthreads_pp, nthreads_pme);
1519 /* Master synchronizes its value of reset_counters with all nodes
1520 * including PME only nodes */
1521 reset_counters = wcycle_get_reset_counters(wcycle);
1522 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1523 wcycle_set_reset_counters(wcycle, reset_counters);
1527 if (cr->duty & DUTY_PP)
1529 bcast_state(cr, state);
1531 /* Initiate forcerecord */
1533 fr->hwinfo = hwinfo;
1534 fr->gpu_opt = &hw_opt->gpu_opt;
1535 init_forcerec(fplog, oenv, fr, fcd, inputrec, mtop, cr, box,
1536 opt2fn("-table", nfile, fnm),
1537 opt2fn("-tabletf", nfile, fnm),
1538 opt2fn("-tablep", nfile, fnm),
1539 opt2fn("-tableb", nfile, fnm),
1544 /* version for PCA_NOT_READ_NODE (see md.c) */
1545 /*init_forcerec(fplog,fr,fcd,inputrec,mtop,cr,box,FALSE,
1546 "nofile","nofile","nofile","nofile",FALSE,pforce);
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 /* Initialize the virtual site communication */
1562 vsite = init_vsite(mtop, cr, FALSE);
1564 calc_shifts(box, fr->shift_vec);
1566 /* With periodic molecules the charge groups should be whole at start up
1567 * and the virtual sites should not be far from their proper positions.
1569 if (!inputrec->bContinuation && MASTER(cr) &&
1570 !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1572 /* Make molecules whole at start of run */
1573 if (fr->ePBC != epbcNONE)
1575 do_pbc_first_mtop(fplog, inputrec->ePBC, box, mtop, state->x);
1579 /* Correct initial vsite positions are required
1580 * for the initial distribution in the domain decomposition
1581 * and for the initial shell prediction.
1583 construct_vsites_mtop(vsite, mtop, state->x);
1587 if (EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype))
1589 ewaldcoeff_q = fr->ewaldcoeff_q;
1590 ewaldcoeff_lj = fr->ewaldcoeff_lj;
1591 pmedata = &fr->pmedata;
1600 /* This is a PME only node */
1602 /* We don't need the state */
1605 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1606 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1610 if (hw_opt->thread_affinity != threadaffOFF)
1612 /* Before setting affinity, check whether the affinity has changed
1613 * - which indicates that probably the OpenMP library has changed it
1614 * since we first checked).
1616 gmx_check_thread_affinity_set(fplog, cr,
1617 hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1619 /* Set the CPU affinity */
1620 gmx_set_thread_affinity(fplog, cr, hw_opt, hwinfo);
1623 /* Initiate PME if necessary,
1624 * either on all nodes or on dedicated PME nodes only. */
1625 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1629 nChargePerturbed = mdatoms->nChargePerturbed;
1630 if (EVDW_PME(inputrec->vdwtype))
1632 nTypePerturbed = mdatoms->nTypePerturbed;
1635 if (cr->npmenodes > 0)
1637 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1638 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1639 gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
1642 if (cr->duty & DUTY_PME)
1644 status = gmx_pme_init(pmedata, cr, npme_major, npme_minor, inputrec,
1645 mtop ? mtop->natoms : 0, nChargePerturbed, nTypePerturbed,
1646 (Flags & MD_REPRODUCIBLE), nthreads_pme);
1649 gmx_fatal(FARGS, "Error %d initializing PME", status);
1655 if (integrator[inputrec->eI].func == do_md)
1657 /* Turn on signal handling on all nodes */
1659 * (A user signal from the PME nodes (if any)
1660 * is communicated to the PP nodes.
1662 signal_handler_install();
1665 if (cr->duty & DUTY_PP)
1667 /* Assumes uniform use of the number of OpenMP threads */
1668 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1670 if (inputrec->ePull != epullNO)
1672 /* Initialize pull code */
1673 init_pull(fplog, inputrec, nfile, fnm, mtop, cr, oenv, inputrec->fepvals->init_lambda,
1674 EI_DYNAMICS(inputrec->eI) && MASTER(cr), Flags);
1679 /* Initialize enforced rotation code */
1680 init_rot(fplog, inputrec, nfile, fnm, cr, state->x, box, mtop, oenv,
1684 if (inputrec->eSwapCoords != eswapNO)
1686 /* Initialize ion swapping code */
1687 init_swapcoords(fplog, bVerbose, inputrec, opt2fn_master("-swap", nfile, fnm, cr),
1688 mtop, state->x, state->box, &state->swapstate, cr, oenv, Flags);
1691 constr = init_constraints(fplog, mtop, inputrec, ed, state, cr);
1693 if (DOMAINDECOMP(cr))
1695 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1696 dd_init_bondeds(fplog, cr->dd, mtop, vsite, inputrec,
1697 Flags & MD_DDBONDCHECK, fr->cginfo_mb);
1699 set_dd_parameters(fplog, cr->dd, dlb_scale, inputrec, &ddbox);
1701 setup_dd_grid(fplog, cr->dd);
1704 /* Now do whatever the user wants us to do (how flexible...) */
1705 integrator[inputrec->eI].func(fplog, cr, nfile, fnm,
1706 oenv, bVerbose, bCompact,
1709 nstepout, inputrec, mtop,
1711 mdatoms, nrnb, wcycle, ed, fr,
1712 repl_ex_nst, repl_ex_nex, repl_ex_seed,
1714 cpt_period, max_hours,
1718 walltime_accounting);
1720 if (inputrec->ePull != epullNO)
1722 finish_pull(inputrec->pull);
1727 finish_rot(inputrec->rot);
1733 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1735 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1736 gmx_pmeonly(*pmedata, cr, nrnb, wcycle, walltime_accounting, ewaldcoeff_q, ewaldcoeff_lj, inputrec);
1739 wallcycle_stop(wcycle, ewcRUN);
1741 /* Finish up, write some stuff
1742 * if rerunMD, don't write last frame again
1744 finish_run(fplog, cr,
1745 inputrec, nrnb, wcycle, walltime_accounting,
1746 fr ? fr->nbv : NULL,
1747 EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
1750 /* Free GPU memory and context */
1751 free_gpu_resources(fr, cr);
1753 if (opt2bSet("-membed", nfile, fnm))
1758 gmx_hardware_info_free(hwinfo);
1760 /* Does what it says */
1761 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1762 walltime_accounting_destroy(walltime_accounting);
1764 /* Close logfile already here if we were appending to it */
1765 if (MASTER(cr) && (Flags & MD_APPENDFILES))
1767 gmx_log_close(fplog);
1770 rc = (int)gmx_get_stop_condition();
1772 #ifdef GMX_THREAD_MPI
1773 /* we need to join all threads. The sub-threads join when they
1774 exit this function, but the master thread needs to be told to
1776 if (PAR(cr) && MASTER(cr))