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.
49 #include "gromacs/domdec/domdec.h"
50 #include "gromacs/essentialdynamics/edsam.h"
51 #include "gromacs/ewald/pme.h"
52 #include "gromacs/fileio/tpxio.h"
53 #include "gromacs/gmxpreprocess/calc_verletbuf.h"
54 #include "gromacs/legacyheaders/checkpoint.h"
55 #include "gromacs/legacyheaders/constr.h"
56 #include "gromacs/legacyheaders/disre.h"
57 #include "gromacs/legacyheaders/force.h"
58 #include "gromacs/legacyheaders/gmx_detect_hardware.h"
59 #include "gromacs/legacyheaders/gmx_omp_nthreads.h"
60 #include "gromacs/legacyheaders/gmx_thread_affinity.h"
61 #include "gromacs/legacyheaders/inputrec.h"
62 #include "gromacs/legacyheaders/main.h"
63 #include "gromacs/legacyheaders/md_logging.h"
64 #include "gromacs/legacyheaders/md_support.h"
65 #include "gromacs/legacyheaders/mdatoms.h"
66 #include "gromacs/legacyheaders/mdrun.h"
67 #include "gromacs/legacyheaders/names.h"
68 #include "gromacs/legacyheaders/network.h"
69 #include "gromacs/legacyheaders/oenv.h"
70 #include "gromacs/legacyheaders/orires.h"
71 #include "gromacs/legacyheaders/qmmm.h"
72 #include "gromacs/legacyheaders/sighandler.h"
73 #include "gromacs/legacyheaders/txtdump.h"
74 #include "gromacs/legacyheaders/typedefs.h"
75 #include "gromacs/math/calculate-ewald-splitting-coefficient.h"
76 #include "gromacs/math/vec.h"
77 #include "gromacs/mdlib/nbnxn_consts.h"
78 #include "gromacs/mdlib/nbnxn_search.h"
79 #include "gromacs/pbcutil/pbc.h"
80 #include "gromacs/pulling/pull.h"
81 #include "gromacs/pulling/pull_rotation.h"
82 #include "gromacs/swap/swapcoords.h"
83 #include "gromacs/timing/wallcycle.h"
84 #include "gromacs/topology/mtop_util.h"
85 #include "gromacs/utility/gmxassert.h"
86 #include "gromacs/utility/gmxmpi.h"
87 #include "gromacs/utility/smalloc.h"
97 #include "gromacs/gmxlib/gpu_utils/gpu_utils.h"
100 gmx_integrator_t *func;
103 /* The array should match the eI array in include/types/enums.h */
104 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}};
106 gmx_int64_t deform_init_init_step_tpx;
107 matrix deform_init_box_tpx;
108 tMPI_Thread_mutex_t deform_init_box_mutex = TMPI_THREAD_MUTEX_INITIALIZER;
111 #ifdef GMX_THREAD_MPI
112 /* The minimum number of atoms per tMPI thread. With fewer atoms than this,
113 * the number of threads will get lowered.
115 #define MIN_ATOMS_PER_MPI_THREAD 90
116 #define MIN_ATOMS_PER_GPU 900
118 struct mdrunner_arglist
133 const char *dddlb_opt;
138 const char *nbpu_opt;
140 gmx_int64_t nsteps_cmdline;
150 const char *deviceOptions;
156 /* The function used for spawning threads. Extracts the mdrunner()
157 arguments from its one argument and calls mdrunner(), after making
159 static void mdrunner_start_fn(void *arg)
161 struct mdrunner_arglist *mda = (struct mdrunner_arglist*)arg;
162 struct mdrunner_arglist mc = *mda; /* copy the arg list to make sure
163 that it's thread-local. This doesn't
164 copy pointed-to items, of course,
165 but those are all const. */
166 t_commrec *cr; /* we need a local version of this */
170 fnm = dup_tfn(mc.nfile, mc.fnm);
172 cr = reinitialize_commrec_for_this_thread(mc.cr);
179 mdrunner(&mc.hw_opt, fplog, cr, mc.nfile, fnm, mc.oenv,
180 mc.bVerbose, mc.bCompact, mc.nstglobalcomm,
181 mc.ddxyz, mc.dd_node_order, mc.rdd,
182 mc.rconstr, mc.dddlb_opt, mc.dlb_scale,
183 mc.ddcsx, mc.ddcsy, mc.ddcsz,
184 mc.nbpu_opt, mc.nstlist_cmdline,
185 mc.nsteps_cmdline, mc.nstepout, mc.resetstep,
186 mc.nmultisim, mc.repl_ex_nst, mc.repl_ex_nex, mc.repl_ex_seed, mc.pforce,
187 mc.cpt_period, mc.max_hours, mc.deviceOptions, mc.imdport, mc.Flags);
190 /* called by mdrunner() to start a specific number of threads (including
191 the main thread) for thread-parallel runs. This in turn calls mdrunner()
193 All options besides nthreads are the same as for mdrunner(). */
194 static t_commrec *mdrunner_start_threads(gmx_hw_opt_t *hw_opt,
195 FILE *fplog, t_commrec *cr, int nfile,
196 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
197 gmx_bool bCompact, int nstglobalcomm,
198 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
199 const char *dddlb_opt, real dlb_scale,
200 const char *ddcsx, const char *ddcsy, const char *ddcsz,
201 const char *nbpu_opt, int nstlist_cmdline,
202 gmx_int64_t nsteps_cmdline,
203 int nstepout, int resetstep,
204 int nmultisim, int repl_ex_nst, int repl_ex_nex, int repl_ex_seed,
205 real pforce, real cpt_period, real max_hours,
206 const char *deviceOptions, unsigned long Flags)
209 struct mdrunner_arglist *mda;
210 t_commrec *crn; /* the new commrec */
213 /* first check whether we even need to start tMPI */
214 if (hw_opt->nthreads_tmpi < 2)
219 /* a few small, one-time, almost unavoidable memory leaks: */
221 fnmn = dup_tfn(nfile, fnm);
223 /* fill the data structure to pass as void pointer to thread start fn */
224 /* hw_opt contains pointers, which should all be NULL at this stage */
225 mda->hw_opt = *hw_opt;
231 mda->bVerbose = bVerbose;
232 mda->bCompact = bCompact;
233 mda->nstglobalcomm = nstglobalcomm;
234 mda->ddxyz[XX] = ddxyz[XX];
235 mda->ddxyz[YY] = ddxyz[YY];
236 mda->ddxyz[ZZ] = ddxyz[ZZ];
237 mda->dd_node_order = dd_node_order;
239 mda->rconstr = rconstr;
240 mda->dddlb_opt = dddlb_opt;
241 mda->dlb_scale = dlb_scale;
245 mda->nbpu_opt = nbpu_opt;
246 mda->nstlist_cmdline = nstlist_cmdline;
247 mda->nsteps_cmdline = nsteps_cmdline;
248 mda->nstepout = nstepout;
249 mda->resetstep = resetstep;
250 mda->nmultisim = nmultisim;
251 mda->repl_ex_nst = repl_ex_nst;
252 mda->repl_ex_nex = repl_ex_nex;
253 mda->repl_ex_seed = repl_ex_seed;
254 mda->pforce = pforce;
255 mda->cpt_period = cpt_period;
256 mda->max_hours = max_hours;
257 mda->deviceOptions = deviceOptions;
260 /* now spawn new threads that start mdrunner_start_fn(), while
261 the main thread returns, we set thread affinity later */
262 ret = tMPI_Init_fn(TRUE, hw_opt->nthreads_tmpi, TMPI_AFFINITY_NONE,
263 mdrunner_start_fn, (void*)(mda) );
264 if (ret != TMPI_SUCCESS)
269 crn = reinitialize_commrec_for_this_thread(cr);
274 static int get_tmpi_omp_thread_division(const gmx_hw_info_t *hwinfo,
275 const gmx_hw_opt_t *hw_opt,
281 /* There are no separate PME nodes here, as we ensured in
282 * check_and_update_hw_opt that nthreads_tmpi>0 with PME nodes
283 * and a conditional ensures we would not have ended up here.
284 * Note that separate PME nodes might be switched on later.
288 nthreads_tmpi = ngpu;
289 if (nthreads_tot > 0 && nthreads_tot < nthreads_tmpi)
291 nthreads_tmpi = nthreads_tot;
294 else if (hw_opt->nthreads_omp > 0)
296 /* Here we could oversubscribe, when we do, we issue a warning later */
297 nthreads_tmpi = std::max(1, nthreads_tot/hw_opt->nthreads_omp);
301 /* TODO choose nthreads_omp based on hardware topology
302 when we have a hardware topology detection library */
303 /* In general, when running up to 4 threads, OpenMP should be faster.
304 * Note: on AMD Bulldozer we should avoid running OpenMP over two dies.
305 * On Intel>=Nehalem running OpenMP on a single CPU is always faster,
306 * even on two CPUs it's usually faster (but with many OpenMP threads
307 * it could be faster not to use HT, currently we always use HT).
308 * On Nehalem/Westmere we want to avoid running 16 threads over
309 * two CPUs with HT, so we need a limit<16; thus we use 12.
310 * A reasonable limit for Intel Sandy and Ivy bridge,
311 * not knowing the topology, is 16 threads.
313 const int nthreads_omp_always_faster = 4;
314 const int nthreads_omp_always_faster_Nehalem = 12;
315 const int nthreads_omp_always_faster_SandyBridge = 16;
316 gmx_bool bIntel_Family6;
319 (gmx_cpuid_vendor(hwinfo->cpuid_info) == GMX_CPUID_VENDOR_INTEL &&
320 gmx_cpuid_family(hwinfo->cpuid_info) == 6);
322 if (nthreads_tot <= nthreads_omp_always_faster ||
324 ((gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_Nehalem && nthreads_tot <= nthreads_omp_always_faster_Nehalem) ||
325 (gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_SandyBridge && nthreads_tot <= nthreads_omp_always_faster_SandyBridge))))
327 /* Use pure OpenMP parallelization */
332 /* Don't use OpenMP parallelization */
333 nthreads_tmpi = nthreads_tot;
337 return nthreads_tmpi;
341 /* Get the number of threads to use for thread-MPI based on how many
342 * were requested, which algorithms we're using,
343 * and how many particles there are.
344 * At the point we have already called check_and_update_hw_opt.
345 * Thus all options should be internally consistent and consistent
346 * with the hardware, except that ntmpi could be larger than #GPU.
348 static int get_nthreads_mpi(const gmx_hw_info_t *hwinfo,
349 gmx_hw_opt_t *hw_opt,
350 t_inputrec *inputrec, gmx_mtop_t *mtop,
354 int nthreads_hw, nthreads_tot_max, nthreads_tmpi, nthreads_new, ngpu;
355 int min_atoms_per_mpi_thread;
358 if (hw_opt->nthreads_tmpi > 0)
360 /* Trivial, return right away */
361 return hw_opt->nthreads_tmpi;
364 nthreads_hw = hwinfo->nthreads_hw_avail;
366 /* How many total (#tMPI*#OpenMP) threads can we start? */
367 if (hw_opt->nthreads_tot > 0)
369 nthreads_tot_max = hw_opt->nthreads_tot;
373 nthreads_tot_max = nthreads_hw;
376 bCanUseGPU = (inputrec->cutoff_scheme == ecutsVERLET &&
377 hwinfo->gpu_info.ncuda_dev_compatible > 0);
380 ngpu = hwinfo->gpu_info.ncuda_dev_compatible;
387 if (inputrec->cutoff_scheme == ecutsGROUP)
389 /* We checked this before, but it doesn't hurt to do it once more */
390 assert(hw_opt->nthreads_omp == 1);
394 get_tmpi_omp_thread_division(hwinfo, hw_opt, nthreads_tot_max, ngpu);
396 if (inputrec->eI == eiNM || EI_TPI(inputrec->eI))
398 /* Dims/steps are divided over the nodes iso splitting the atoms */
399 min_atoms_per_mpi_thread = 0;
405 min_atoms_per_mpi_thread = MIN_ATOMS_PER_GPU;
409 min_atoms_per_mpi_thread = MIN_ATOMS_PER_MPI_THREAD;
413 /* Check if an algorithm does not support parallel simulation. */
414 if (nthreads_tmpi != 1 &&
415 ( inputrec->eI == eiLBFGS ||
416 inputrec->coulombtype == eelEWALD ) )
420 md_print_warn(cr, fplog, "The integration or electrostatics algorithm doesn't support parallel runs. Using a single thread-MPI thread.\n");
421 if (hw_opt->nthreads_tmpi > nthreads_tmpi)
423 gmx_fatal(FARGS, "You asked for more than 1 thread-MPI thread, but an algorithm doesn't support that");
426 else if (mtop->natoms/nthreads_tmpi < min_atoms_per_mpi_thread)
428 /* the thread number was chosen automatically, but there are too many
429 threads (too few atoms per thread) */
430 nthreads_new = std::max(1, mtop->natoms/min_atoms_per_mpi_thread);
432 /* Avoid partial use of Hyper-Threading */
433 if (gmx_cpuid_x86_smt(hwinfo->cpuid_info) == GMX_CPUID_X86_SMT_ENABLED &&
434 nthreads_new > nthreads_hw/2 && nthreads_new < nthreads_hw)
436 nthreads_new = nthreads_hw/2;
439 /* Avoid large prime numbers in the thread count */
440 if (nthreads_new >= 6)
442 /* Use only 6,8,10 with additional factors of 2 */
446 while (3*fac*2 <= nthreads_new)
451 nthreads_new = (nthreads_new/fac)*fac;
456 if (nthreads_new == 5)
462 nthreads_tmpi = nthreads_new;
464 fprintf(stderr, "\n");
465 fprintf(stderr, "NOTE: Parallelization is limited by the small number of atoms,\n");
466 fprintf(stderr, " only starting %d thread-MPI threads.\n", nthreads_tmpi);
467 fprintf(stderr, " You can use the -nt and/or -ntmpi option to optimize the number of threads.\n\n");
470 return nthreads_tmpi;
472 #endif /* GMX_THREAD_MPI */
475 /* We determine the extra cost of the non-bonded kernels compared to
476 * a reference nstlist value of 10 (which is the default in grompp).
478 static const int nbnxnReferenceNstlist = 10;
479 /* The values to try when switching */
480 const int nstlist_try[] = { 20, 25, 40 };
481 #define NNSTL sizeof(nstlist_try)/sizeof(nstlist_try[0])
482 /* Increase nstlist until the non-bonded cost increases more than listfac_ok,
483 * but never more than listfac_max.
484 * A standard (protein+)water system at 300K with PME ewald_rtol=1e-5
485 * needs 1.28 at rcoulomb=0.9 and 1.24 at rcoulomb=1.0 to get to nstlist=40.
486 * Note that both CPU and GPU factors are conservative. Performance should
487 * not go down due to this tuning, except with a relatively slow GPU.
488 * On the other hand, at medium/high parallelization or with fast GPUs
489 * nstlist will not be increased enough to reach optimal performance.
491 /* CPU: pair-search is about a factor 1.5 slower than the non-bonded kernel */
492 static const float nbnxn_cpu_listfac_ok = 1.05;
493 static const float nbnxn_cpu_listfac_max = 1.09;
494 /* GPU: pair-search is a factor 1.5-3 slower than the non-bonded kernel */
495 static const float nbnxn_gpu_listfac_ok = 1.20;
496 static const float nbnxn_gpu_listfac_max = 1.30;
498 /* Try to increase nstlist when using the Verlet cut-off scheme */
499 static void increase_nstlist(FILE *fp, t_commrec *cr,
500 t_inputrec *ir, int nstlist_cmdline,
501 const gmx_mtop_t *mtop, matrix box,
504 float listfac_ok, listfac_max;
505 int nstlist_orig, nstlist_prev;
506 verletbuf_list_setup_t ls;
507 real rlistWithReferenceNstlist, rlist_inc, rlist_ok, rlist_max;
508 real rlist_new, rlist_prev;
509 size_t nstlist_ind = 0;
511 gmx_bool bBox, bDD, bCont;
512 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";
513 const char *nve_err = "Can not increase nstlist because an NVE ensemble is used";
514 const char *vbd_err = "Can not increase nstlist because verlet-buffer-tolerance is not set or used";
515 const char *box_err = "Can not increase nstlist because the box is too small";
516 const char *dd_err = "Can not increase nstlist because of domain decomposition limitations";
518 const float oneThird = 1.0f / 3.0f;
520 if (nstlist_cmdline <= 0)
522 if (ir->nstlist == 1)
524 /* The user probably set nstlist=1 for a reason,
525 * don't mess with the settings.
530 if (fp != NULL && bGPU && ir->nstlist < nstlist_try[0])
532 fprintf(fp, nstl_gpu, ir->nstlist);
535 while (nstlist_ind < NNSTL && ir->nstlist >= nstlist_try[nstlist_ind])
539 if (nstlist_ind == NNSTL)
541 /* There are no larger nstlist value to try */
546 if (EI_MD(ir->eI) && ir->etc == etcNO)
550 fprintf(stderr, "%s\n", nve_err);
554 fprintf(fp, "%s\n", nve_err);
560 if (ir->verletbuf_tol == 0 && bGPU)
562 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");
565 if (ir->verletbuf_tol < 0)
569 fprintf(stderr, "%s\n", vbd_err);
573 fprintf(fp, "%s\n", vbd_err);
581 listfac_ok = nbnxn_gpu_listfac_ok;
582 listfac_max = nbnxn_gpu_listfac_max;
586 listfac_ok = nbnxn_cpu_listfac_ok;
587 listfac_max = nbnxn_cpu_listfac_max;
590 nstlist_orig = ir->nstlist;
591 if (nstlist_cmdline > 0)
595 sprintf(buf, "Getting nstlist=%d from command line option",
598 ir->nstlist = nstlist_cmdline;
601 verletbuf_get_list_setup(bGPU, &ls);
603 /* Allow rlist to make the list a given factor larger than the list
604 * would be with the reference value for nstlist (10).
606 nstlist_prev = ir->nstlist;
607 ir->nstlist = nbnxnReferenceNstlist;
608 calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL,
609 &rlistWithReferenceNstlist);
610 ir->nstlist = nstlist_prev;
612 /* Determine the pair list size increase due to zero interactions */
613 rlist_inc = nbnxn_get_rlist_effective_inc(ls.cluster_size_j,
614 mtop->natoms/det(box));
615 rlist_ok = (rlistWithReferenceNstlist + rlist_inc)*pow(listfac_ok, oneThird) - rlist_inc;
616 rlist_max = (rlistWithReferenceNstlist + rlist_inc)*pow(listfac_max, oneThird) - rlist_inc;
619 fprintf(debug, "nstlist tuning: rlist_inc %.3f rlist_ok %.3f rlist_max %.3f\n",
620 rlist_inc, rlist_ok, rlist_max);
623 nstlist_prev = nstlist_orig;
624 rlist_prev = ir->rlist;
627 if (nstlist_cmdline <= 0)
629 ir->nstlist = nstlist_try[nstlist_ind];
632 /* Set the pair-list buffer size in ir */
633 calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
635 /* Does rlist fit in the box? */
636 bBox = (sqr(rlist_new) < max_cutoff2(ir->ePBC, box));
638 if (bBox && DOMAINDECOMP(cr))
640 /* Check if rlist fits in the domain decomposition */
641 if (inputrec2nboundeddim(ir) < DIM)
643 gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
645 copy_mat(box, state_tmp.box);
646 bDD = change_dd_cutoff(cr, &state_tmp, ir, rlist_new);
651 fprintf(debug, "nstlist %d rlist %.3f bBox %d bDD %d\n",
652 ir->nstlist, rlist_new, bBox, bDD);
657 if (nstlist_cmdline <= 0)
659 if (bBox && bDD && rlist_new <= rlist_max)
661 /* Increase nstlist */
662 nstlist_prev = ir->nstlist;
663 rlist_prev = rlist_new;
664 bCont = (nstlist_ind+1 < NNSTL && rlist_new < rlist_ok);
668 /* Stick with the previous nstlist */
669 ir->nstlist = nstlist_prev;
670 rlist_new = rlist_prev;
682 gmx_warning(!bBox ? box_err : dd_err);
685 fprintf(fp, "\n%s\n", bBox ? box_err : dd_err);
687 ir->nstlist = nstlist_orig;
689 else if (ir->nstlist != nstlist_orig || rlist_new != ir->rlist)
691 sprintf(buf, "Changing nstlist from %d to %d, rlist from %g to %g",
692 nstlist_orig, ir->nstlist,
693 ir->rlist, rlist_new);
696 fprintf(stderr, "%s\n\n", buf);
700 fprintf(fp, "%s\n\n", buf);
702 ir->rlist = rlist_new;
703 ir->rlistlong = rlist_new;
707 static void prepare_verlet_scheme(FILE *fplog,
711 const gmx_mtop_t *mtop,
715 /* For NVE simulations, we will retain the initial list buffer */
716 if (ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
718 /* Update the Verlet buffer size for the current run setup */
719 verletbuf_list_setup_t ls;
722 /* Here we assume SIMD-enabled kernels are being used. But as currently
723 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
724 * and 4x2 gives a larger buffer than 4x4, this is ok.
726 verletbuf_get_list_setup(bUseGPU, &ls);
728 calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
730 if (rlist_new != ir->rlist)
734 fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
735 ir->rlist, rlist_new,
736 ls.cluster_size_i, ls.cluster_size_j);
738 ir->rlist = rlist_new;
739 ir->rlistlong = rlist_new;
743 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
745 gmx_fatal(FARGS, "Can not set nstlist without %s",
746 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
749 if (EI_DYNAMICS(ir->eI))
751 /* Set or try nstlist values */
752 increase_nstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, bUseGPU);
756 static void convert_to_verlet_scheme(FILE *fplog,
758 gmx_mtop_t *mtop, real box_vol)
760 const char *conv_mesg = "Converting input file with group cut-off scheme to the Verlet cut-off scheme";
762 md_print_warn(NULL, fplog, "%s\n", conv_mesg);
764 ir->cutoff_scheme = ecutsVERLET;
765 ir->verletbuf_tol = 0.005;
767 if (ir->rcoulomb != ir->rvdw)
769 gmx_fatal(FARGS, "The VdW and Coulomb cut-offs are different, whereas the Verlet scheme only supports equal cut-offs");
772 if (ir->vdwtype == evdwUSER || EEL_USER(ir->coulombtype))
774 gmx_fatal(FARGS, "User non-bonded potentials are not (yet) supported with the Verlet scheme");
776 else if (ir_vdw_switched(ir) || ir_coulomb_switched(ir))
778 if (ir_vdw_switched(ir) && ir->vdw_modifier == eintmodNONE)
780 ir->vdwtype = evdwCUT;
784 case evdwSHIFT: ir->vdw_modifier = eintmodFORCESWITCH; break;
785 case evdwSWITCH: ir->vdw_modifier = eintmodPOTSWITCH; break;
786 default: gmx_fatal(FARGS, "The Verlet scheme does not support Van der Waals interactions of type '%s'", evdw_names[ir->vdwtype]);
789 if (ir_coulomb_switched(ir) && ir->coulomb_modifier == eintmodNONE)
791 if (EEL_FULL(ir->coulombtype))
793 /* With full electrostatic only PME can be switched */
794 ir->coulombtype = eelPME;
795 ir->coulomb_modifier = eintmodPOTSHIFT;
799 md_print_warn(NULL, fplog, "NOTE: Replacing %s electrostatics with reaction-field with epsilon-rf=inf\n", eel_names[ir->coulombtype]);
800 ir->coulombtype = eelRF;
801 ir->epsilon_rf = 0.0;
802 ir->coulomb_modifier = eintmodPOTSHIFT;
806 /* We set the pair energy error tolerance to a small number.
807 * Note that this is only for testing. For production the user
808 * should think about this and set the mdp options.
810 ir->verletbuf_tol = 1e-4;
813 if (inputrec2nboundeddim(ir) != 3)
815 gmx_fatal(FARGS, "Can only convert old tpr files to the Verlet cut-off scheme with 3D pbc");
818 if (ir->efep != efepNO || ir->implicit_solvent != eisNO)
820 gmx_fatal(FARGS, "Will not convert old tpr files to the Verlet cut-off scheme with free-energy calculations or implicit solvent");
823 if (EI_DYNAMICS(ir->eI) && !(EI_MD(ir->eI) && ir->etc == etcNO))
825 verletbuf_list_setup_t ls;
827 verletbuf_get_list_setup(FALSE, &ls);
828 calc_verlet_buffer_size(mtop, box_vol, ir, -1, &ls, NULL, &ir->rlist);
836 rlist_fac = 1 + verlet_buffer_ratio_NVE_T0;
840 rlist_fac = 1 + verlet_buffer_ratio_nodynamics;
842 ir->verletbuf_tol = -1;
843 ir->rlist = rlist_fac*std::max(ir->rvdw, ir->rcoulomb);
846 gmx_mtop_remove_chargegroups(mtop);
849 static void print_hw_opt(FILE *fp, const gmx_hw_opt_t *hw_opt)
851 fprintf(fp, "hw_opt: nt %d ntmpi %d ntomp %d ntomp_pme %d gpu_id '%s'\n",
852 hw_opt->nthreads_tot,
853 hw_opt->nthreads_tmpi,
854 hw_opt->nthreads_omp,
855 hw_opt->nthreads_omp_pme,
856 hw_opt->gpu_opt.gpu_id != NULL ? hw_opt->gpu_opt.gpu_id : "");
859 /* Checks we can do when we don't (yet) know the cut-off scheme */
860 static void check_and_update_hw_opt_1(gmx_hw_opt_t *hw_opt,
861 gmx_bool bIsSimMaster)
863 gmx_omp_nthreads_read_env(&hw_opt->nthreads_omp, bIsSimMaster);
865 #ifndef GMX_THREAD_MPI
866 if (hw_opt->nthreads_tot > 0)
868 gmx_fatal(FARGS, "Setting the total number of threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
870 if (hw_opt->nthreads_tmpi > 0)
872 gmx_fatal(FARGS, "Setting the number of thread-MPI threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
877 if (hw_opt->nthreads_omp > 1)
879 gmx_fatal(FARGS, "More than 1 OpenMP thread requested, but Gromacs was compiled without OpenMP support");
881 hw_opt->nthreads_omp = 1;
884 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp_pme <= 0)
886 /* We have the same number of OpenMP threads for PP and PME processes,
887 * thus we can perform several consistency checks.
889 if (hw_opt->nthreads_tmpi > 0 &&
890 hw_opt->nthreads_omp > 0 &&
891 hw_opt->nthreads_tot != hw_opt->nthreads_tmpi*hw_opt->nthreads_omp)
893 gmx_fatal(FARGS, "The total number of threads requested (%d) does not match the thread-MPI threads (%d) times the OpenMP threads (%d) requested",
894 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi, hw_opt->nthreads_omp);
897 if (hw_opt->nthreads_tmpi > 0 &&
898 hw_opt->nthreads_tot % hw_opt->nthreads_tmpi != 0)
900 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of thread-MPI threads requested (%d)",
901 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi);
904 if (hw_opt->nthreads_omp > 0 &&
905 hw_opt->nthreads_tot % hw_opt->nthreads_omp != 0)
907 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of OpenMP threads requested (%d)",
908 hw_opt->nthreads_tot, hw_opt->nthreads_omp);
911 if (hw_opt->nthreads_tmpi > 0 &&
912 hw_opt->nthreads_omp <= 0)
914 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
919 if (hw_opt->nthreads_omp > 1)
921 gmx_fatal(FARGS, "OpenMP threads are requested, but Gromacs was compiled without OpenMP support");
925 if (hw_opt->nthreads_omp_pme > 0 && hw_opt->nthreads_omp <= 0)
927 gmx_fatal(FARGS, "You need to specify -ntomp in addition to -ntomp_pme");
930 if (hw_opt->nthreads_tot == 1)
932 hw_opt->nthreads_tmpi = 1;
934 if (hw_opt->nthreads_omp > 1)
936 gmx_fatal(FARGS, "You requested %d OpenMP threads with %d total threads",
937 hw_opt->nthreads_tmpi, hw_opt->nthreads_tot);
939 hw_opt->nthreads_omp = 1;
942 if (hw_opt->nthreads_omp_pme <= 0 && hw_opt->nthreads_omp > 0)
944 hw_opt->nthreads_omp_pme = hw_opt->nthreads_omp;
947 /* Parse GPU IDs, if provided.
948 * We check consistency with the tMPI thread count later.
950 gmx_parse_gpu_ids(&hw_opt->gpu_opt);
952 #ifdef GMX_THREAD_MPI
953 if (hw_opt->gpu_opt.ncuda_dev_use > 0 && hw_opt->nthreads_tmpi == 0)
955 /* Set the number of MPI threads equal to the number of GPUs */
956 hw_opt->nthreads_tmpi = hw_opt->gpu_opt.ncuda_dev_use;
958 if (hw_opt->nthreads_tot > 0 &&
959 hw_opt->nthreads_tmpi > hw_opt->nthreads_tot)
961 /* We have more GPUs than total threads requested.
962 * We choose to (later) generate a mismatch error,
963 * instead of launching more threads than requested.
965 hw_opt->nthreads_tmpi = hw_opt->nthreads_tot;
972 print_hw_opt(debug, hw_opt);
976 /* Checks we can do when we know the cut-off scheme */
977 static void check_and_update_hw_opt_2(gmx_hw_opt_t *hw_opt,
980 if (cutoff_scheme == ecutsGROUP)
982 /* We only have OpenMP support for PME only nodes */
983 if (hw_opt->nthreads_omp > 1)
985 gmx_fatal(FARGS, "OpenMP threads have been requested with cut-off scheme %s, but these are only supported with cut-off scheme %s",
986 ecutscheme_names[cutoff_scheme],
987 ecutscheme_names[ecutsVERLET]);
989 hw_opt->nthreads_omp = 1;
992 if (hw_opt->nthreads_omp_pme <= 0 && hw_opt->nthreads_omp > 0)
994 hw_opt->nthreads_omp_pme = hw_opt->nthreads_omp;
999 print_hw_opt(debug, hw_opt);
1004 /* Override the value in inputrec with value passed on the command line (if any) */
1005 static void override_nsteps_cmdline(FILE *fplog,
1006 gmx_int64_t nsteps_cmdline,
1008 const t_commrec *cr)
1010 char sbuf[STEPSTRSIZE];
1015 /* override with anything else than the default -2 */
1016 if (nsteps_cmdline > -2)
1020 ir->nsteps = nsteps_cmdline;
1021 if (EI_DYNAMICS(ir->eI) && nsteps_cmdline != -1)
1023 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
1024 gmx_step_str(nsteps_cmdline, sbuf),
1025 fabs(nsteps_cmdline*ir->delta_t));
1029 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps",
1030 gmx_step_str(nsteps_cmdline, sbuf));
1033 md_print_warn(cr, fplog, "%s\n", stmp);
1037 int mdrunner(gmx_hw_opt_t *hw_opt,
1038 FILE *fplog, t_commrec *cr, int nfile,
1039 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
1040 gmx_bool bCompact, int nstglobalcomm,
1041 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
1042 const char *dddlb_opt, real dlb_scale,
1043 const char *ddcsx, const char *ddcsy, const char *ddcsz,
1044 const char *nbpu_opt, int nstlist_cmdline,
1045 gmx_int64_t nsteps_cmdline, int nstepout, int resetstep,
1046 int gmx_unused nmultisim, int repl_ex_nst, int repl_ex_nex,
1047 int repl_ex_seed, real pforce, real cpt_period, real max_hours,
1048 const char *deviceOptions, int imdport, unsigned long Flags)
1050 gmx_bool bForceUseGPU, bTryUseGPU;
1051 t_inputrec *inputrec;
1052 t_state *state = NULL;
1054 gmx_ddbox_t ddbox = {0};
1055 int npme_major, npme_minor;
1057 gmx_mtop_t *mtop = NULL;
1058 t_mdatoms *mdatoms = NULL;
1059 t_forcerec *fr = NULL;
1060 t_fcdata *fcd = NULL;
1061 real ewaldcoeff_q = 0;
1062 real ewaldcoeff_lj = 0;
1063 gmx_pme_t *pmedata = NULL;
1064 gmx_vsite_t *vsite = NULL;
1065 gmx_constr_t constr;
1066 int nChargePerturbed = -1, nTypePerturbed = 0, status;
1067 gmx_wallcycle_t wcycle;
1069 gmx_walltime_accounting_t walltime_accounting = NULL;
1071 gmx_int64_t reset_counters;
1072 gmx_edsam_t ed = NULL;
1073 int nthreads_pme = 1;
1074 int nthreads_pp = 1;
1075 gmx_membed_t membed = NULL;
1076 gmx_hw_info_t *hwinfo = NULL;
1077 /* The master rank decides early on bUseGPU and broadcasts this later */
1078 gmx_bool bUseGPU = FALSE;
1080 /* CAUTION: threads may be started later on in this function, so
1081 cr doesn't reflect the final parallel state right now */
1085 if (Flags & MD_APPENDFILES)
1090 bForceUseGPU = (strncmp(nbpu_opt, "gpu", 3) == 0);
1091 bTryUseGPU = (strncmp(nbpu_opt, "auto", 4) == 0) || bForceUseGPU;
1093 /* Detect hardware, gather information. This is an operation that is
1094 * global for this process (MPI rank). */
1095 hwinfo = gmx_detect_hardware(fplog, cr, bTryUseGPU);
1101 /* Read (nearly) all data required for the simulation */
1102 read_tpx_state(ftp2fn(efTPR, nfile, fnm), inputrec, state, NULL, mtop);
1104 if (inputrec->cutoff_scheme != ecutsVERLET &&
1105 ((Flags & MD_TESTVERLET) || getenv("GMX_VERLET_SCHEME") != NULL))
1107 convert_to_verlet_scheme(fplog, inputrec, mtop, det(state->box));
1110 if (inputrec->cutoff_scheme == ecutsVERLET)
1112 /* Here the master rank decides if all ranks will use GPUs */
1113 bUseGPU = (hwinfo->gpu_info.ncuda_dev_compatible > 0 ||
1114 getenv("GMX_EMULATE_GPU") != NULL);
1116 /* TODO add GPU kernels for this and replace this check by:
1117 * (bUseGPU && (ir->vdwtype == evdwPME &&
1118 * ir->ljpme_combination_rule == eljpmeLB))
1119 * update the message text and the content of nbnxn_acceleration_supported.
1122 !nbnxn_acceleration_supported(fplog, cr, inputrec, bUseGPU))
1124 /* Fallback message printed by nbnxn_acceleration_supported */
1127 gmx_fatal(FARGS, "GPU acceleration requested, but not supported with the given input settings");
1132 prepare_verlet_scheme(fplog, cr,
1133 inputrec, nstlist_cmdline, mtop, state->box,
1138 if (nstlist_cmdline > 0)
1140 gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
1143 if (hwinfo->gpu_info.ncuda_dev_compatible > 0)
1145 md_print_warn(cr, fplog,
1146 "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
1147 " To use a GPU, set the mdp option: cutoff-scheme = Verlet\n"
1148 " (for quick performance testing you can use the -testverlet option)\n");
1153 gmx_fatal(FARGS, "GPU requested, but can't be used without cutoff-scheme=Verlet");
1156 #ifdef GMX_TARGET_BGQ
1157 md_print_warn(cr, fplog,
1158 "NOTE: There is no SIMD implementation of the group scheme kernels on\n"
1159 " BlueGene/Q. You will observe better performance from using the\n"
1160 " Verlet cut-off scheme.\n");
1164 if (inputrec->eI == eiSD2)
1166 md_print_warn(cr, fplog, "The stochastic dynamics integrator %s is deprecated, since\n"
1167 "it is slower than integrator %s and is slightly less accurate\n"
1168 "with constraints. Use the %s integrator.",
1169 ei_names[inputrec->eI], ei_names[eiSD1], ei_names[eiSD1]);
1173 /* Check and update the hardware options for internal consistency */
1174 check_and_update_hw_opt_1(hw_opt, SIMMASTER(cr));
1176 /* Early check for externally set process affinity. */
1177 gmx_check_thread_affinity_set(fplog, cr,
1178 hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1182 #ifdef GMX_THREAD_MPI
1183 if (cr->npmenodes > 0 && hw_opt->nthreads_tmpi <= 0)
1185 gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME ranks");
1189 if (hw_opt->nthreads_omp_pme != hw_opt->nthreads_omp &&
1192 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");
1196 #ifdef GMX_THREAD_MPI
1199 /* Since the master knows the cut-off scheme, update hw_opt for this.
1200 * This is done later for normal MPI and also once more with tMPI
1201 * for all tMPI ranks.
1203 check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
1205 /* NOW the threads will be started: */
1206 hw_opt->nthreads_tmpi = get_nthreads_mpi(hwinfo,
1210 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp <= 0)
1212 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
1215 if (hw_opt->nthreads_tmpi > 1)
1217 t_commrec *cr_old = cr;
1218 /* now start the threads. */
1219 cr = mdrunner_start_threads(hw_opt, fplog, cr_old, nfile, fnm,
1220 oenv, bVerbose, bCompact, nstglobalcomm,
1221 ddxyz, dd_node_order, rdd, rconstr,
1222 dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
1223 nbpu_opt, nstlist_cmdline,
1224 nsteps_cmdline, nstepout, resetstep, nmultisim,
1225 repl_ex_nst, repl_ex_nex, repl_ex_seed, pforce,
1226 cpt_period, max_hours, deviceOptions,
1228 /* the main thread continues here with a new cr. We don't deallocate
1229 the old cr because other threads may still be reading it. */
1232 gmx_comm("Failed to spawn threads");
1237 /* END OF CAUTION: cr is now reliable */
1239 /* g_membed initialisation *
1240 * Because we change the mtop, init_membed is called before the init_parallel *
1241 * (in case we ever want to make it run in parallel) */
1242 if (opt2bSet("-membed", nfile, fnm))
1246 fprintf(stderr, "Initializing membed");
1248 membed = init_membed(fplog, nfile, fnm, mtop, inputrec, state, cr, &cpt_period);
1253 /* now broadcast everything to the non-master nodes/threads: */
1254 init_parallel(cr, inputrec, mtop);
1258 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
1261 /* now make sure the state is initialized and propagated */
1262 set_state_entries(state, inputrec);
1264 /* A parallel command line option consistency check that we can
1265 only do after any threads have started. */
1267 (ddxyz[XX] > 1 || ddxyz[YY] > 1 || ddxyz[ZZ] > 1 || cr->npmenodes > 0))
1270 "The -dd or -npme option request a parallel simulation, "
1272 "but %s was compiled without threads or MPI enabled"
1274 #ifdef GMX_THREAD_MPI
1275 "but the number of threads (option -nt) is 1"
1277 "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec"
1280 , output_env_get_program_display_name(oenv)
1284 if ((Flags & MD_RERUN) &&
1285 (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
1287 gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
1290 if (can_use_allvsall(inputrec, TRUE, cr, fplog) && DOMAINDECOMP(cr))
1292 gmx_fatal(FARGS, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
1295 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
1297 if (cr->npmenodes > 0)
1299 gmx_fatal_collective(FARGS, cr, NULL,
1300 "PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
1309 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1313 /* NMR restraints must be initialized before load_checkpoint,
1314 * since with time averaging the history is added to t_state.
1315 * For proper consistency check we therefore need to extend
1317 * So the PME-only nodes (if present) will also initialize
1318 * the distance restraints.
1322 /* This needs to be called before read_checkpoint to extend the state */
1323 init_disres(fplog, mtop, inputrec, cr, fcd, state, repl_ex_nst > 0);
1325 init_orires(fplog, mtop, state->x, inputrec, cr, &(fcd->orires),
1328 if (DEFORM(*inputrec))
1330 /* Store the deform reference box before reading the checkpoint */
1333 copy_mat(state->box, box);
1337 gmx_bcast(sizeof(box), box, cr);
1339 /* Because we do not have the update struct available yet
1340 * in which the reference values should be stored,
1341 * we store them temporarily in static variables.
1342 * This should be thread safe, since they are only written once
1343 * and with identical values.
1345 tMPI_Thread_mutex_lock(&deform_init_box_mutex);
1346 deform_init_init_step_tpx = inputrec->init_step;
1347 copy_mat(box, deform_init_box_tpx);
1348 tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
1351 if (opt2bSet("-cpi", nfile, fnm))
1353 /* Check if checkpoint file exists before doing continuation.
1354 * This way we can use identical input options for the first and subsequent runs...
1356 if (gmx_fexist_master(opt2fn_master("-cpi", nfile, fnm, cr), cr) )
1358 load_checkpoint(opt2fn_master("-cpi", nfile, fnm, cr), &fplog,
1360 inputrec, state, &bReadEkin,
1361 (Flags & MD_APPENDFILES),
1362 (Flags & MD_APPENDFILESSET));
1366 Flags |= MD_READ_EKIN;
1371 if (MASTER(cr) && (Flags & MD_APPENDFILES))
1373 gmx_log_open(ftp2fn(efLOG, nfile, fnm), cr,
1377 /* override nsteps with value from cmdline */
1378 override_nsteps_cmdline(fplog, nsteps_cmdline, inputrec, cr);
1382 copy_mat(state->box, box);
1387 gmx_bcast(sizeof(box), box, cr);
1390 /* Essential dynamics */
1391 if (opt2bSet("-ei", nfile, fnm))
1393 /* Open input and output files, allocate space for ED data structure */
1394 ed = ed_open(mtop->natoms, &state->edsamstate, nfile, fnm, Flags, oenv, cr);
1397 if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
1398 inputrec->eI == eiNM))
1400 cr->dd = init_domain_decomposition(fplog, cr, Flags, ddxyz, rdd, rconstr,
1401 dddlb_opt, dlb_scale,
1402 ddcsx, ddcsy, ddcsz,
1405 &ddbox, &npme_major, &npme_minor);
1407 make_dd_communicators(fplog, cr, dd_node_order);
1409 /* Set overallocation to avoid frequent reallocation of arrays */
1410 set_over_alloc_dd(TRUE);
1414 /* PME, if used, is done on all nodes with 1D decomposition */
1416 cr->duty = (DUTY_PP | DUTY_PME);
1420 if (inputrec->ePBC == epbcSCREW)
1423 "pbc=%s is only implemented with domain decomposition",
1424 epbc_names[inputrec->ePBC]);
1430 /* After possible communicator splitting in make_dd_communicators.
1431 * we can set up the intra/inter node communication.
1433 gmx_setup_nodecomm(fplog, cr);
1436 /* Initialize per-physical-node MPI process/thread ID and counters. */
1437 gmx_init_intranode_counters(cr);
1441 md_print_info(cr, fplog,
1442 "This is simulation %d out of %d running as a composite Gromacs\n"
1443 "multi-simulation job. Setup for this simulation:\n\n",
1444 cr->ms->sim, cr->ms->nsim);
1446 md_print_info(cr, fplog, "Using %d MPI %s\n",
1448 #ifdef GMX_THREAD_MPI
1449 cr->nnodes == 1 ? "thread" : "threads"
1451 cr->nnodes == 1 ? "process" : "processes"
1457 /* Check and update hw_opt for the cut-off scheme */
1458 check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
1460 gmx_omp_nthreads_init(fplog, cr,
1461 hwinfo->nthreads_hw_avail,
1462 hw_opt->nthreads_omp,
1463 hw_opt->nthreads_omp_pme,
1464 (cr->duty & DUTY_PP) == 0,
1465 inputrec->cutoff_scheme == ecutsVERLET);
1469 /* The master rank decided on the use of GPUs,
1470 * broadcast this information to all ranks.
1472 gmx_bcast_sim(sizeof(bUseGPU), &bUseGPU, cr);
1477 if (cr->npmenodes == -1)
1479 /* Don't automatically use PME-only nodes with GPUs */
1483 /* Select GPU id's to use */
1484 gmx_select_gpu_ids(fplog, cr, &hwinfo->gpu_info, bForceUseGPU,
1489 /* Ignore (potentially) manually selected GPUs */
1490 hw_opt->gpu_opt.ncuda_dev_use = 0;
1493 /* check consistency across ranks of things like SIMD
1494 * support and number of GPUs selected */
1495 gmx_check_hw_runconf_consistency(fplog, hwinfo, cr, hw_opt, bUseGPU);
1497 if (DOMAINDECOMP(cr))
1499 /* When we share GPUs over ranks, we need to know this for the DLB */
1500 dd_setup_dlb_resource_sharing(cr, hwinfo, hw_opt);
1503 /* getting number of PP/PME threads
1504 PME: env variable should be read only on one node to make sure it is
1505 identical everywhere;
1507 /* TODO nthreads_pp is only used for pinning threads.
1508 * This is a temporary solution until we have a hw topology library.
1510 nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
1511 nthreads_pme = gmx_omp_nthreads_get(emntPME);
1513 wcycle = wallcycle_init(fplog, resetstep, cr, nthreads_pp, nthreads_pme);
1517 /* Master synchronizes its value of reset_counters with all nodes
1518 * including PME only nodes */
1519 reset_counters = wcycle_get_reset_counters(wcycle);
1520 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1521 wcycle_set_reset_counters(wcycle, reset_counters);
1525 if (cr->duty & DUTY_PP)
1527 bcast_state(cr, state);
1529 /* Initiate forcerecord */
1531 fr->hwinfo = hwinfo;
1532 fr->gpu_opt = &hw_opt->gpu_opt;
1533 init_forcerec(fplog, oenv, fr, fcd, inputrec, mtop, cr, box,
1534 opt2fn("-table", nfile, fnm),
1535 opt2fn("-tabletf", nfile, fnm),
1536 opt2fn("-tablep", nfile, fnm),
1537 opt2fn("-tableb", nfile, fnm),
1542 /* version for PCA_NOT_READ_NODE (see md.c) */
1543 /*init_forcerec(fplog,fr,fcd,inputrec,mtop,cr,box,FALSE,
1544 "nofile","nofile","nofile","nofile",FALSE,pforce);
1547 /* Initialize QM-MM */
1550 init_QMMMrec(cr, mtop, inputrec, fr);
1553 /* Initialize the mdatoms structure.
1554 * mdatoms is not filled with atom data,
1555 * as this can not be done now with domain decomposition.
1557 mdatoms = init_mdatoms(fplog, mtop, inputrec->efep != efepNO);
1559 /* Initialize the virtual site communication */
1560 vsite = init_vsite(mtop, cr, FALSE);
1562 calc_shifts(box, fr->shift_vec);
1564 /* With periodic molecules the charge groups should be whole at start up
1565 * and the virtual sites should not be far from their proper positions.
1567 if (!inputrec->bContinuation && MASTER(cr) &&
1568 !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1570 /* Make molecules whole at start of run */
1571 if (fr->ePBC != epbcNONE)
1573 do_pbc_first_mtop(fplog, inputrec->ePBC, box, mtop, state->x);
1577 /* Correct initial vsite positions are required
1578 * for the initial distribution in the domain decomposition
1579 * and for the initial shell prediction.
1581 construct_vsites_mtop(vsite, mtop, state->x);
1585 if (EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype))
1587 ewaldcoeff_q = fr->ewaldcoeff_q;
1588 ewaldcoeff_lj = fr->ewaldcoeff_lj;
1589 pmedata = &fr->pmedata;
1598 /* This is a PME only node */
1600 /* We don't need the state */
1603 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1604 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1608 if (hw_opt->thread_affinity != threadaffOFF)
1610 /* Before setting affinity, check whether the affinity has changed
1611 * - which indicates that probably the OpenMP library has changed it
1612 * since we first checked).
1614 gmx_check_thread_affinity_set(fplog, cr,
1615 hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1617 /* Set the CPU affinity */
1618 gmx_set_thread_affinity(fplog, cr, hw_opt, hwinfo);
1621 /* Initiate PME if necessary,
1622 * either on all nodes or on dedicated PME nodes only. */
1623 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1627 nChargePerturbed = mdatoms->nChargePerturbed;
1628 if (EVDW_PME(inputrec->vdwtype))
1630 nTypePerturbed = mdatoms->nTypePerturbed;
1633 if (cr->npmenodes > 0)
1635 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1636 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1637 gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, 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, nTypePerturbed,
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 if (inputrec->eSwapCoords != eswapNO)
1684 /* Initialize ion swapping code */
1685 init_swapcoords(fplog, bVerbose, inputrec, opt2fn_master("-swap", nfile, fnm, cr),
1686 mtop, state->x, state->box, &state->swapstate, cr, oenv, Flags);
1689 constr = init_constraints(fplog, mtop, inputrec, ed, state, cr);
1691 if (DOMAINDECOMP(cr))
1693 GMX_RELEASE_ASSERT(fr, "fr was NULL while cr->duty was DUTY_PP");
1694 dd_init_bondeds(fplog, cr->dd, mtop, vsite, inputrec,
1695 Flags & MD_DDBONDCHECK, fr->cginfo_mb);
1697 set_dd_parameters(fplog, cr->dd, dlb_scale, inputrec, &ddbox);
1699 setup_dd_grid(fplog, cr->dd);
1702 /* Now do whatever the user wants us to do (how flexible...) */
1703 integrator[inputrec->eI].func(fplog, cr, nfile, fnm,
1704 oenv, bVerbose, bCompact,
1707 nstepout, inputrec, mtop,
1709 mdatoms, nrnb, wcycle, ed, fr,
1710 repl_ex_nst, repl_ex_nex, repl_ex_seed,
1712 cpt_period, max_hours,
1716 walltime_accounting);
1718 if (inputrec->ePull != epullNO)
1720 finish_pull(inputrec->pull);
1725 finish_rot(inputrec->rot);
1731 GMX_RELEASE_ASSERT(pmedata, "pmedata was NULL while cr->duty was not DUTY_PP");
1733 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1734 gmx_pmeonly(*pmedata, cr, nrnb, wcycle, walltime_accounting, ewaldcoeff_q, ewaldcoeff_lj, inputrec);
1737 wallcycle_stop(wcycle, ewcRUN);
1739 /* Finish up, write some stuff
1740 * if rerunMD, don't write last frame again
1742 finish_run(fplog, cr,
1743 inputrec, nrnb, wcycle, walltime_accounting,
1744 fr ? fr->nbv : NULL,
1745 EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
1748 /* Free GPU memory and context */
1749 free_gpu_resources(fr, cr, &hwinfo->gpu_info, fr ? fr->gpu_opt : NULL);
1751 if (opt2bSet("-membed", nfile, fnm))
1756 gmx_hardware_info_free(hwinfo);
1758 /* Does what it says */
1759 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", gmx_gettime());
1760 walltime_accounting_destroy(walltime_accounting);
1762 /* Close logfile already here if we were appending to it */
1763 if (MASTER(cr) && (Flags & MD_APPENDFILES))
1765 gmx_log_close(fplog);
1768 rc = (int)gmx_get_stop_condition();
1770 #ifdef GMX_THREAD_MPI
1771 /* we need to join all threads. The sub-threads join when they
1772 exit this function, but the master thread needs to be told to
1774 if (PAR(cr) && MASTER(cr))