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.
54 #include "md_logging.h"
55 #include "md_support.h"
68 #include "checkpoint.h"
69 #include "mtop_util.h"
70 #include "sighandler.h"
72 #include "gmx_detect_hardware.h"
73 #include "gmx_omp_nthreads.h"
74 #include "calc_verletbuf.h"
75 #include "gmx_fatal_collective.h"
78 #include "gmx_thread_affinity.h"
81 #include "gromacs/fileio/tpxio.h"
82 #include "gromacs/mdlib/nbnxn_search.h"
83 #include "gromacs/mdlib/nbnxn_consts.h"
84 #include "gromacs/timing/wallcycle.h"
85 #include "gromacs/utility/gmxmpi.h"
86 #include "gromacs/utility/gmxomp.h"
87 #include "gromacs/swap/swapcoords.h"
88 #include "gromacs/essentialdynamics/edsam.h"
89 #include "gromacs/pulling/pull.h"
90 #include "gromacs/pulling/pull_rotation.h"
96 #include "gpu_utils.h"
97 #include "nbnxn_cuda_data_mgmt.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 struct mdrunner_arglist
127 const char *dddlb_opt;
132 const char *nbpu_opt;
134 gmx_int64_t nsteps_cmdline;
144 const char *deviceOptions;
149 /* The function used for spawning threads. Extracts the mdrunner()
150 arguments from its one argument and calls mdrunner(), after making
152 static void mdrunner_start_fn(void *arg)
154 struct mdrunner_arglist *mda = (struct mdrunner_arglist*)arg;
155 struct mdrunner_arglist mc = *mda; /* copy the arg list to make sure
156 that it's thread-local. This doesn't
157 copy pointed-to items, of course,
158 but those are all const. */
159 t_commrec *cr; /* we need a local version of this */
163 fnm = dup_tfn(mc.nfile, mc.fnm);
165 cr = reinitialize_commrec_for_this_thread(mc.cr);
172 mdrunner(&mc.hw_opt, fplog, cr, mc.nfile, fnm, mc.oenv,
173 mc.bVerbose, mc.bCompact, mc.nstglobalcomm,
174 mc.ddxyz, mc.dd_node_order, mc.rdd,
175 mc.rconstr, mc.dddlb_opt, mc.dlb_scale,
176 mc.ddcsx, mc.ddcsy, mc.ddcsz,
177 mc.nbpu_opt, mc.nstlist_cmdline,
178 mc.nsteps_cmdline, mc.nstepout, mc.resetstep,
179 mc.nmultisim, mc.repl_ex_nst, mc.repl_ex_nex, mc.repl_ex_seed, mc.pforce,
180 mc.cpt_period, mc.max_hours, mc.deviceOptions, mc.Flags);
183 /* called by mdrunner() to start a specific number of threads (including
184 the main thread) for thread-parallel runs. This in turn calls mdrunner()
186 All options besides nthreads are the same as for mdrunner(). */
187 static t_commrec *mdrunner_start_threads(gmx_hw_opt_t *hw_opt,
188 FILE *fplog, t_commrec *cr, int nfile,
189 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
190 gmx_bool bCompact, int nstglobalcomm,
191 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
192 const char *dddlb_opt, real dlb_scale,
193 const char *ddcsx, const char *ddcsy, const char *ddcsz,
194 const char *nbpu_opt, int nstlist_cmdline,
195 gmx_int64_t nsteps_cmdline,
196 int nstepout, int resetstep,
197 int nmultisim, int repl_ex_nst, int repl_ex_nex, int repl_ex_seed,
198 real pforce, real cpt_period, real max_hours,
199 const char *deviceOptions, unsigned long Flags)
202 struct mdrunner_arglist *mda;
203 t_commrec *crn; /* the new commrec */
206 /* first check whether we even need to start tMPI */
207 if (hw_opt->nthreads_tmpi < 2)
212 /* a few small, one-time, almost unavoidable memory leaks: */
214 fnmn = dup_tfn(nfile, fnm);
216 /* fill the data structure to pass as void pointer to thread start fn */
217 /* hw_opt contains pointers, which should all be NULL at this stage */
218 mda->hw_opt = *hw_opt;
224 mda->bVerbose = bVerbose;
225 mda->bCompact = bCompact;
226 mda->nstglobalcomm = nstglobalcomm;
227 mda->ddxyz[XX] = ddxyz[XX];
228 mda->ddxyz[YY] = ddxyz[YY];
229 mda->ddxyz[ZZ] = ddxyz[ZZ];
230 mda->dd_node_order = dd_node_order;
232 mda->rconstr = rconstr;
233 mda->dddlb_opt = dddlb_opt;
234 mda->dlb_scale = dlb_scale;
238 mda->nbpu_opt = nbpu_opt;
239 mda->nstlist_cmdline = nstlist_cmdline;
240 mda->nsteps_cmdline = nsteps_cmdline;
241 mda->nstepout = nstepout;
242 mda->resetstep = resetstep;
243 mda->nmultisim = nmultisim;
244 mda->repl_ex_nst = repl_ex_nst;
245 mda->repl_ex_nex = repl_ex_nex;
246 mda->repl_ex_seed = repl_ex_seed;
247 mda->pforce = pforce;
248 mda->cpt_period = cpt_period;
249 mda->max_hours = max_hours;
250 mda->deviceOptions = deviceOptions;
253 /* now spawn new threads that start mdrunner_start_fn(), while
254 the main thread returns, we set thread affinity later */
255 ret = tMPI_Init_fn(TRUE, hw_opt->nthreads_tmpi, TMPI_AFFINITY_NONE,
256 mdrunner_start_fn, (void*)(mda) );
257 if (ret != TMPI_SUCCESS)
262 crn = reinitialize_commrec_for_this_thread(cr);
267 static int get_tmpi_omp_thread_division(const gmx_hw_info_t *hwinfo,
268 const gmx_hw_opt_t *hw_opt,
274 /* There are no separate PME nodes here, as we ensured in
275 * check_and_update_hw_opt that nthreads_tmpi>0 with PME nodes
276 * and a conditional ensures we would not have ended up here.
277 * Note that separate PME nodes might be switched on later.
281 nthreads_tmpi = ngpu;
282 if (nthreads_tot > 0 && nthreads_tot < nthreads_tmpi)
284 nthreads_tmpi = nthreads_tot;
287 else if (hw_opt->nthreads_omp > 0)
289 /* Here we could oversubscribe, when we do, we issue a warning later */
290 nthreads_tmpi = max(1, nthreads_tot/hw_opt->nthreads_omp);
294 /* TODO choose nthreads_omp based on hardware topology
295 when we have a hardware topology detection library */
296 /* In general, when running up to 4 threads, OpenMP should be faster.
297 * Note: on AMD Bulldozer we should avoid running OpenMP over two dies.
298 * On Intel>=Nehalem running OpenMP on a single CPU is always faster,
299 * even on two CPUs it's usually faster (but with many OpenMP threads
300 * it could be faster not to use HT, currently we always use HT).
301 * On Nehalem/Westmere we want to avoid running 16 threads over
302 * two CPUs with HT, so we need a limit<16; thus we use 12.
303 * A reasonable limit for Intel Sandy and Ivy bridge,
304 * not knowing the topology, is 16 threads.
306 const int nthreads_omp_always_faster = 4;
307 const int nthreads_omp_always_faster_Nehalem = 12;
308 const int nthreads_omp_always_faster_SandyBridge = 16;
309 const int first_model_Nehalem = 0x1A;
310 const int first_model_SandyBridge = 0x2A;
311 gmx_bool bIntel_Family6;
314 (gmx_cpuid_vendor(hwinfo->cpuid_info) == GMX_CPUID_VENDOR_INTEL &&
315 gmx_cpuid_family(hwinfo->cpuid_info) == 6);
317 if (nthreads_tot <= nthreads_omp_always_faster ||
319 ((gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_Nehalem && nthreads_tot <= nthreads_omp_always_faster_Nehalem) ||
320 (gmx_cpuid_model(hwinfo->cpuid_info) >= nthreads_omp_always_faster_SandyBridge && nthreads_tot <= nthreads_omp_always_faster_SandyBridge))))
322 /* Use pure OpenMP parallelization */
327 /* Don't use OpenMP parallelization */
328 nthreads_tmpi = nthreads_tot;
332 return nthreads_tmpi;
336 /* Get the number of threads to use for thread-MPI based on how many
337 * were requested, which algorithms we're using,
338 * and how many particles there are.
339 * At the point we have already called check_and_update_hw_opt.
340 * Thus all options should be internally consistent and consistent
341 * with the hardware, except that ntmpi could be larger than #GPU.
343 static int get_nthreads_mpi(const gmx_hw_info_t *hwinfo,
344 gmx_hw_opt_t *hw_opt,
345 t_inputrec *inputrec, gmx_mtop_t *mtop,
349 int nthreads_hw, nthreads_tot_max, nthreads_tmpi, nthreads_new, ngpu;
350 int min_atoms_per_mpi_thread;
355 if (hw_opt->nthreads_tmpi > 0)
357 /* Trivial, return right away */
358 return hw_opt->nthreads_tmpi;
361 nthreads_hw = hwinfo->nthreads_hw_avail;
363 /* How many total (#tMPI*#OpenMP) threads can we start? */
364 if (hw_opt->nthreads_tot > 0)
366 nthreads_tot_max = hw_opt->nthreads_tot;
370 nthreads_tot_max = nthreads_hw;
373 bCanUseGPU = (inputrec->cutoff_scheme == ecutsVERLET &&
374 hwinfo->gpu_info.ncuda_dev_compatible > 0);
377 ngpu = hwinfo->gpu_info.ncuda_dev_compatible;
384 if (inputrec->cutoff_scheme == ecutsGROUP)
386 /* We checked this before, but it doesn't hurt to do it once more */
387 assert(hw_opt->nthreads_omp == 1);
391 get_tmpi_omp_thread_division(hwinfo, hw_opt, nthreads_tot_max, ngpu);
393 if (inputrec->eI == eiNM || EI_TPI(inputrec->eI))
395 /* Dims/steps are divided over the nodes iso splitting the atoms */
396 min_atoms_per_mpi_thread = 0;
402 min_atoms_per_mpi_thread = MIN_ATOMS_PER_GPU;
406 min_atoms_per_mpi_thread = MIN_ATOMS_PER_MPI_THREAD;
410 /* Check if an algorithm does not support parallel simulation. */
411 if (nthreads_tmpi != 1 &&
412 ( inputrec->eI == eiLBFGS ||
413 inputrec->coulombtype == eelEWALD ) )
417 md_print_warn(cr, fplog, "The integration or electrostatics algorithm doesn't support parallel runs. Using a single thread-MPI thread.\n");
418 if (hw_opt->nthreads_tmpi > nthreads_tmpi)
420 gmx_fatal(FARGS, "You asked for more than 1 thread-MPI thread, but an algorithm doesn't support that");
423 else if (mtop->natoms/nthreads_tmpi < min_atoms_per_mpi_thread)
425 /* the thread number was chosen automatically, but there are too many
426 threads (too few atoms per thread) */
427 nthreads_new = max(1, mtop->natoms/min_atoms_per_mpi_thread);
429 /* Avoid partial use of Hyper-Threading */
430 if (gmx_cpuid_x86_smt(hwinfo->cpuid_info) == GMX_CPUID_X86_SMT_ENABLED &&
431 nthreads_new > nthreads_hw/2 && nthreads_new < nthreads_hw)
433 nthreads_new = nthreads_hw/2;
436 /* Avoid large prime numbers in the thread count */
437 if (nthreads_new >= 6)
439 /* Use only 6,8,10 with additional factors of 2 */
443 while (3*fac*2 <= nthreads_new)
448 nthreads_new = (nthreads_new/fac)*fac;
453 if (nthreads_new == 5)
459 nthreads_tmpi = nthreads_new;
461 fprintf(stderr, "\n");
462 fprintf(stderr, "NOTE: Parallelization is limited by the small number of atoms,\n");
463 fprintf(stderr, " only starting %d thread-MPI threads.\n", nthreads_tmpi);
464 fprintf(stderr, " You can use the -nt and/or -ntmpi option to optimize the number of threads.\n\n");
467 return nthreads_tmpi;
469 #endif /* GMX_THREAD_MPI */
472 /* We determine the extra cost of the non-bonded kernels compared to
473 * a reference nstlist value of 10 (which is the default in grompp).
475 static const int nbnxn_reference_nstlist = 10;
476 /* The values to try when switching */
477 const int nstlist_try[] = { 20, 25, 40 };
478 #define NNSTL sizeof(nstlist_try)/sizeof(nstlist_try[0])
479 /* Increase nstlist until the non-bonded cost increases more than listfac_ok,
480 * but never more than listfac_max.
481 * A standard (protein+)water system at 300K with PME ewald_rtol=1e-5
482 * needs 1.28 at rcoulomb=0.9 and 1.24 at rcoulomb=1.0 to get to nstlist=40.
483 * Note that both CPU and GPU factors are conservative. Performance should
484 * not go down due to this tuning, except with a relatively slow GPU.
485 * On the other hand, at medium/high parallelization or with fast GPUs
486 * nstlist will not be increased enough to reach optimal performance.
488 /* CPU: pair-search is about a factor 1.5 slower than the non-bonded kernel */
489 static const float nbnxn_cpu_listfac_ok = 1.05;
490 static const float nbnxn_cpu_listfac_max = 1.09;
491 /* GPU: pair-search is a factor 1.5-3 slower than the non-bonded kernel */
492 static const float nbnxn_gpu_listfac_ok = 1.20;
493 static const float nbnxn_gpu_listfac_max = 1.30;
495 /* Try to increase nstlist when using the Verlet cut-off scheme */
496 static void increase_nstlist(FILE *fp, t_commrec *cr,
497 t_inputrec *ir, int nstlist_cmdline,
498 const gmx_mtop_t *mtop, matrix box,
501 float listfac_ok, listfac_max;
502 int nstlist_orig, nstlist_prev;
503 verletbuf_list_setup_t ls;
504 real rlist_nstlist10, rlist_inc, rlist_ok, rlist_max;
505 real rlist_new, rlist_prev;
508 gmx_bool bBox, bDD, bCont;
509 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";
510 const char *nve_err = "Can not increase nstlist because an NVE ensemble is used";
511 const char *vbd_err = "Can not increase nstlist because verlet-buffer-tolerance is not set or used";
512 const char *box_err = "Can not increase nstlist because the box is too small";
513 const char *dd_err = "Can not increase nstlist because of domain decomposition limitations";
516 if (nstlist_cmdline <= 0)
518 if (fp != NULL && bGPU && ir->nstlist < nstlist_try[0])
520 fprintf(fp, nstl_gpu, ir->nstlist);
523 while (nstlist_ind < NNSTL && ir->nstlist >= nstlist_try[nstlist_ind])
527 if (nstlist_ind == NNSTL)
529 /* There are no larger nstlist value to try */
534 if (EI_MD(ir->eI) && ir->etc == etcNO)
538 fprintf(stderr, "%s\n", nve_err);
542 fprintf(fp, "%s\n", nve_err);
548 if (ir->verletbuf_tol == 0 && bGPU)
550 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");
553 if (ir->verletbuf_tol < 0)
557 fprintf(stderr, "%s\n", vbd_err);
561 fprintf(fp, "%s\n", vbd_err);
569 listfac_ok = nbnxn_gpu_listfac_ok;
570 listfac_max = nbnxn_gpu_listfac_max;
574 listfac_ok = nbnxn_cpu_listfac_ok;
575 listfac_max = nbnxn_cpu_listfac_max;
578 nstlist_orig = ir->nstlist;
579 if (nstlist_cmdline > 0)
583 sprintf(buf, "Getting nstlist=%d from command line option",
586 ir->nstlist = nstlist_cmdline;
589 verletbuf_get_list_setup(bGPU, &ls);
591 /* Allow rlist to make the list a given factor larger than the list
592 * would be with nstlist=10.
594 nstlist_prev = ir->nstlist;
596 calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL,
598 ir->nstlist = nstlist_prev;
600 /* Determine the pair list size increase due to zero interactions */
601 rlist_inc = nbnxn_get_rlist_effective_inc(ls.cluster_size_j,
602 mtop->natoms/det(box));
603 rlist_ok = (rlist_nstlist10 + rlist_inc)*pow(listfac_ok, 1.0/3.0) - rlist_inc;
604 rlist_max = (rlist_nstlist10 + rlist_inc)*pow(listfac_max, 1.0/3.0) - rlist_inc;
607 fprintf(debug, "nstlist tuning: rlist_inc %.3f rlist_ok %.3f rlist_max %.3f\n",
608 rlist_inc, rlist_ok, rlist_max);
611 nstlist_prev = nstlist_orig;
612 rlist_prev = ir->rlist;
615 if (nstlist_cmdline <= 0)
617 ir->nstlist = nstlist_try[nstlist_ind];
620 /* Set the pair-list buffer size in ir */
621 calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
623 /* Does rlist fit in the box? */
624 bBox = (sqr(rlist_new) < max_cutoff2(ir->ePBC, box));
626 if (bBox && DOMAINDECOMP(cr))
628 /* Check if rlist fits in the domain decomposition */
629 if (inputrec2nboundeddim(ir) < DIM)
631 gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
633 copy_mat(box, state_tmp.box);
634 bDD = change_dd_cutoff(cr, &state_tmp, ir, rlist_new);
639 fprintf(debug, "nstlist %d rlist %.3f bBox %d bDD %d\n",
640 ir->nstlist, rlist_new, bBox, bDD);
645 if (nstlist_cmdline <= 0)
647 if (bBox && bDD && rlist_new <= rlist_max)
649 /* Increase nstlist */
650 nstlist_prev = ir->nstlist;
651 rlist_prev = rlist_new;
652 bCont = (nstlist_ind+1 < NNSTL && rlist_new < rlist_ok);
656 /* Stick with the previous nstlist */
657 ir->nstlist = nstlist_prev;
658 rlist_new = rlist_prev;
670 gmx_warning(!bBox ? box_err : dd_err);
673 fprintf(fp, "\n%s\n", bBox ? box_err : dd_err);
675 ir->nstlist = nstlist_orig;
677 else if (ir->nstlist != nstlist_orig || rlist_new != ir->rlist)
679 sprintf(buf, "Changing nstlist from %d to %d, rlist from %g to %g",
680 nstlist_orig, ir->nstlist,
681 ir->rlist, rlist_new);
684 fprintf(stderr, "%s\n\n", buf);
688 fprintf(fp, "%s\n\n", buf);
690 ir->rlist = rlist_new;
691 ir->rlistlong = rlist_new;
695 static void prepare_verlet_scheme(FILE *fplog,
699 const gmx_mtop_t *mtop,
703 /* For NVE simulations, we will retain the initial list buffer */
704 if (ir->verletbuf_tol > 0 && !(EI_MD(ir->eI) && ir->etc == etcNO))
706 /* Update the Verlet buffer size for the current run setup */
707 verletbuf_list_setup_t ls;
710 /* Here we assume SIMD-enabled kernels are being used. But as currently
711 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
712 * and 4x2 gives a larger buffer than 4x4, this is ok.
714 verletbuf_get_list_setup(bUseGPU, &ls);
716 calc_verlet_buffer_size(mtop, det(box), ir, -1, &ls, NULL, &rlist_new);
718 if (rlist_new != ir->rlist)
722 fprintf(fplog, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
723 ir->rlist, rlist_new,
724 ls.cluster_size_i, ls.cluster_size_j);
726 ir->rlist = rlist_new;
727 ir->rlistlong = rlist_new;
731 if (nstlist_cmdline > 0 && (!EI_DYNAMICS(ir->eI) || ir->verletbuf_tol <= 0))
733 gmx_fatal(FARGS, "Can not set nstlist without %s",
734 !EI_DYNAMICS(ir->eI) ? "dynamics" : "verlet-buffer-tolerance");
737 if (EI_DYNAMICS(ir->eI))
739 /* Set or try nstlist values */
740 increase_nstlist(fplog, cr, ir, nstlist_cmdline, mtop, box, bUseGPU);
744 static void convert_to_verlet_scheme(FILE *fplog,
746 gmx_mtop_t *mtop, real box_vol)
748 char *conv_mesg = "Converting input file with group cut-off scheme to the Verlet cut-off scheme";
750 md_print_warn(NULL, fplog, "%s\n", conv_mesg);
752 ir->cutoff_scheme = ecutsVERLET;
753 ir->verletbuf_tol = 0.005;
755 if (ir->rcoulomb != ir->rvdw)
757 gmx_fatal(FARGS, "The VdW and Coulomb cut-offs are different, whereas the Verlet scheme only supports equal cut-offs");
760 if (ir->vdwtype == evdwUSER || EEL_USER(ir->coulombtype))
762 gmx_fatal(FARGS, "User non-bonded potentials are not (yet) supported with the Verlet scheme");
764 else if (ir_vdw_switched(ir) || ir_coulomb_switched(ir))
766 if (ir_vdw_switched(ir) && ir->vdw_modifier == eintmodNONE)
768 ir->vdwtype = evdwCUT;
772 case evdwSHIFT: ir->vdw_modifier = eintmodFORCESWITCH; break;
773 case evdwSWITCH: ir->vdw_modifier = eintmodPOTSWITCH; break;
774 default: gmx_fatal(FARGS, "The Verlet scheme does not support Van der Waals interactions of type '%s'", evdw_names[ir->vdwtype]);
777 if (ir_coulomb_switched(ir) && ir->coulomb_modifier == eintmodNONE)
779 if (EEL_FULL(ir->coulombtype))
781 /* With full electrostatic only PME can be switched */
782 ir->coulombtype = eelPME;
783 ir->coulomb_modifier = eintmodPOTSHIFT;
787 md_print_warn(NULL, fplog, "NOTE: Replacing %s electrostatics with reaction-field with epsilon-rf=inf\n", eel_names[ir->coulombtype]);
788 ir->coulombtype = eelRF;
789 ir->epsilon_rf = 0.0;
790 ir->coulomb_modifier = eintmodPOTSHIFT;
794 /* We set the pair energy error tolerance to a small number.
795 * Note that this is only for testing. For production the user
796 * should think about this and set the mdp options.
798 ir->verletbuf_tol = 1e-4;
801 if (inputrec2nboundeddim(ir) != 3)
803 gmx_fatal(FARGS, "Can only convert old tpr files to the Verlet cut-off scheme with 3D pbc");
806 if (ir->efep != efepNO || ir->implicit_solvent != eisNO)
808 gmx_fatal(FARGS, "Will not convert old tpr files to the Verlet cut-off scheme with free-energy calculations or implicit solvent");
811 if (EI_DYNAMICS(ir->eI) && !(EI_MD(ir->eI) && ir->etc == etcNO))
813 verletbuf_list_setup_t ls;
815 verletbuf_get_list_setup(FALSE, &ls);
816 calc_verlet_buffer_size(mtop, box_vol, ir, -1, &ls, NULL, &ir->rlist);
824 rlist_fac = 1 + verlet_buffer_ratio_NVE_T0;
828 rlist_fac = 1 + verlet_buffer_ratio_nodynamics;
830 ir->verletbuf_tol = -1;
831 ir->rlist = rlist_fac*max(ir->rvdw, ir->rcoulomb);
834 gmx_mtop_remove_chargegroups(mtop);
837 static void print_hw_opt(FILE *fp, const gmx_hw_opt_t *hw_opt)
839 fprintf(fp, "hw_opt: nt %d ntmpi %d ntomp %d ntomp_pme %d gpu_id '%s'\n",
840 hw_opt->nthreads_tot,
841 hw_opt->nthreads_tmpi,
842 hw_opt->nthreads_omp,
843 hw_opt->nthreads_omp_pme,
844 hw_opt->gpu_opt.gpu_id != NULL ? hw_opt->gpu_opt.gpu_id : "");
847 /* Checks we can do when we don't (yet) know the cut-off scheme */
848 static void check_and_update_hw_opt_1(gmx_hw_opt_t *hw_opt,
849 gmx_bool bIsSimMaster)
851 gmx_omp_nthreads_read_env(&hw_opt->nthreads_omp, bIsSimMaster);
853 #ifndef GMX_THREAD_MPI
854 if (hw_opt->nthreads_tot > 0)
856 gmx_fatal(FARGS, "Setting the total number of threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
858 if (hw_opt->nthreads_tmpi > 0)
860 gmx_fatal(FARGS, "Setting the number of thread-MPI threads is only supported with thread-MPI and Gromacs was compiled without thread-MPI");
865 if (hw_opt->nthreads_omp > 1)
867 gmx_fatal(FARGS, "More than 1 OpenMP thread requested, but Gromacs was compiled without OpenMP support");
869 hw_opt->nthreads_omp = 1;
872 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp_pme <= 0)
874 /* We have the same number of OpenMP threads for PP and PME processes,
875 * thus we can perform several consistency checks.
877 if (hw_opt->nthreads_tmpi > 0 &&
878 hw_opt->nthreads_omp > 0 &&
879 hw_opt->nthreads_tot != hw_opt->nthreads_tmpi*hw_opt->nthreads_omp)
881 gmx_fatal(FARGS, "The total number of threads requested (%d) does not match the thread-MPI threads (%d) times the OpenMP threads (%d) requested",
882 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi, hw_opt->nthreads_omp);
885 if (hw_opt->nthreads_tmpi > 0 &&
886 hw_opt->nthreads_tot % hw_opt->nthreads_tmpi != 0)
888 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of thread-MPI threads requested (%d)",
889 hw_opt->nthreads_tot, hw_opt->nthreads_tmpi);
892 if (hw_opt->nthreads_omp > 0 &&
893 hw_opt->nthreads_tot % hw_opt->nthreads_omp != 0)
895 gmx_fatal(FARGS, "The total number of threads requested (%d) is not divisible by the number of OpenMP threads requested (%d)",
896 hw_opt->nthreads_tot, hw_opt->nthreads_omp);
899 if (hw_opt->nthreads_tmpi > 0 &&
900 hw_opt->nthreads_omp <= 0)
902 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
907 if (hw_opt->nthreads_omp > 1)
909 gmx_fatal(FARGS, "OpenMP threads are requested, but Gromacs was compiled without OpenMP support");
913 if (hw_opt->nthreads_omp_pme > 0 && hw_opt->nthreads_omp <= 0)
915 gmx_fatal(FARGS, "You need to specify -ntomp in addition to -ntomp_pme");
918 if (hw_opt->nthreads_tot == 1)
920 hw_opt->nthreads_tmpi = 1;
922 if (hw_opt->nthreads_omp > 1)
924 gmx_fatal(FARGS, "You requested %d OpenMP threads with %d total threads",
925 hw_opt->nthreads_tmpi, hw_opt->nthreads_tot);
927 hw_opt->nthreads_omp = 1;
930 if (hw_opt->nthreads_omp_pme <= 0 && hw_opt->nthreads_omp > 0)
932 hw_opt->nthreads_omp_pme = hw_opt->nthreads_omp;
935 /* Parse GPU IDs, if provided.
936 * We check consistency with the tMPI thread count later.
938 gmx_parse_gpu_ids(&hw_opt->gpu_opt);
940 #ifdef GMX_THREAD_MPI
941 if (hw_opt->gpu_opt.ncuda_dev_use > 0 && hw_opt->nthreads_tmpi == 0)
943 /* Set the number of MPI threads equal to the number of GPUs */
944 hw_opt->nthreads_tmpi = hw_opt->gpu_opt.ncuda_dev_use;
946 if (hw_opt->nthreads_tot > 0 &&
947 hw_opt->nthreads_tmpi > hw_opt->nthreads_tot)
949 /* We have more GPUs than total threads requested.
950 * We choose to (later) generate a mismatch error,
951 * instead of launching more threads than requested.
953 hw_opt->nthreads_tmpi = hw_opt->nthreads_tot;
960 print_hw_opt(debug, hw_opt);
964 /* Checks we can do when we know the cut-off scheme */
965 static void check_and_update_hw_opt_2(gmx_hw_opt_t *hw_opt,
968 if (cutoff_scheme == ecutsGROUP)
970 /* We only have OpenMP support for PME only nodes */
971 if (hw_opt->nthreads_omp > 1)
973 gmx_fatal(FARGS, "OpenMP threads have been requested with cut-off scheme %s, but these are only supported with cut-off scheme %s",
974 ecutscheme_names[cutoff_scheme],
975 ecutscheme_names[ecutsVERLET]);
977 hw_opt->nthreads_omp = 1;
980 if (hw_opt->nthreads_omp_pme <= 0 && hw_opt->nthreads_omp > 0)
982 hw_opt->nthreads_omp_pme = hw_opt->nthreads_omp;
987 print_hw_opt(debug, hw_opt);
992 /* Override the value in inputrec with value passed on the command line (if any) */
993 static void override_nsteps_cmdline(FILE *fplog,
994 gmx_int64_t nsteps_cmdline,
998 char sbuf[STEPSTRSIZE];
1003 /* override with anything else than the default -2 */
1004 if (nsteps_cmdline > -2)
1008 ir->nsteps = nsteps_cmdline;
1009 if (EI_DYNAMICS(ir->eI))
1011 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps, %.3f ps",
1012 gmx_step_str(nsteps_cmdline, sbuf),
1013 nsteps_cmdline*ir->delta_t);
1017 sprintf(stmp, "Overriding nsteps with value passed on the command line: %s steps",
1018 gmx_step_str(nsteps_cmdline, sbuf));
1021 md_print_warn(cr, fplog, "%s\n", stmp);
1025 /* Frees GPU memory and destroys the CUDA context.
1027 * Note that this function needs to be called even if GPUs are not used
1028 * in this run because the PME ranks have no knowledge of whether GPUs
1029 * are used or not, but all ranks need to enter the barrier below.
1031 static void free_gpu_resources(const t_forcerec *fr,
1032 const t_commrec *cr)
1034 gmx_bool bIsPPrankUsingGPU;
1035 char gpu_err_str[STRLEN];
1037 bIsPPrankUsingGPU = (cr->duty & DUTY_PP) && fr->nbv != NULL && fr->nbv->bUseGPU;
1039 if (bIsPPrankUsingGPU)
1041 /* free nbnxn data in GPU memory */
1042 nbnxn_cuda_free(fr->nbv->cu_nbv);
1044 /* With tMPI we need to wait for all ranks to finish deallocation before
1045 * destroying the context in free_gpu() as some ranks may be sharing
1047 * Note: as only PP ranks need to free GPU resources, so it is safe to
1048 * not call the barrier on PME ranks.
1050 #ifdef GMX_THREAD_MPI
1055 #endif /* GMX_THREAD_MPI */
1057 /* uninitialize GPU (by destroying the context) */
1058 if (!free_gpu(gpu_err_str))
1060 gmx_warning("On node %d failed to free GPU #%d: %s",
1061 cr->nodeid, get_current_gpu_device_id(), gpu_err_str);
1066 int mdrunner(gmx_hw_opt_t *hw_opt,
1067 FILE *fplog, t_commrec *cr, int nfile,
1068 const t_filenm fnm[], const output_env_t oenv, gmx_bool bVerbose,
1069 gmx_bool bCompact, int nstglobalcomm,
1070 ivec ddxyz, int dd_node_order, real rdd, real rconstr,
1071 const char *dddlb_opt, real dlb_scale,
1072 const char *ddcsx, const char *ddcsy, const char *ddcsz,
1073 const char *nbpu_opt, int nstlist_cmdline,
1074 gmx_int64_t nsteps_cmdline, int nstepout, int resetstep,
1075 int gmx_unused nmultisim, int repl_ex_nst, int repl_ex_nex,
1076 int repl_ex_seed, real pforce, real cpt_period, real max_hours,
1077 const char *deviceOptions, unsigned long Flags)
1079 gmx_bool bForceUseGPU, bTryUseGPU;
1080 double nodetime = 0, realtime;
1081 t_inputrec *inputrec;
1082 t_state *state = NULL;
1084 gmx_ddbox_t ddbox = {0};
1085 int npme_major, npme_minor;
1088 gmx_mtop_t *mtop = NULL;
1089 t_mdatoms *mdatoms = NULL;
1090 t_forcerec *fr = NULL;
1091 t_fcdata *fcd = NULL;
1092 real ewaldcoeff_q = 0;
1093 real ewaldcoeff_lj = 0;
1094 gmx_pme_t *pmedata = NULL;
1095 gmx_vsite_t *vsite = NULL;
1096 gmx_constr_t constr;
1097 int i, m, nChargePerturbed = -1, nTypePerturbed = 0, status, nalloc;
1099 gmx_wallcycle_t wcycle;
1102 gmx_walltime_accounting_t walltime_accounting = NULL;
1104 gmx_int64_t reset_counters;
1105 gmx_edsam_t ed = NULL;
1106 t_commrec *cr_old = cr;
1107 int nthreads_pme = 1;
1108 int nthreads_pp = 1;
1109 gmx_membed_t membed = NULL;
1110 gmx_hw_info_t *hwinfo = NULL;
1111 /* The master rank decides early on bUseGPU and broadcasts this later */
1112 gmx_bool bUseGPU = FALSE;
1114 /* CAUTION: threads may be started later on in this function, so
1115 cr doesn't reflect the final parallel state right now */
1119 if (Flags & MD_APPENDFILES)
1124 bForceUseGPU = (strncmp(nbpu_opt, "gpu", 3) == 0);
1125 bTryUseGPU = (strncmp(nbpu_opt, "auto", 4) == 0) || bForceUseGPU;
1127 /* Detect hardware, gather information. This is an operation that is
1128 * global for this process (MPI rank). */
1129 hwinfo = gmx_detect_hardware(fplog, cr, bTryUseGPU);
1135 /* Read (nearly) all data required for the simulation */
1136 read_tpx_state(ftp2fn(efTPX, nfile, fnm), inputrec, state, NULL, mtop);
1138 if (inputrec->cutoff_scheme != ecutsVERLET &&
1139 ((Flags & MD_TESTVERLET) || getenv("GMX_VERLET_SCHEME") != NULL))
1141 convert_to_verlet_scheme(fplog, inputrec, mtop, det(state->box));
1144 if (inputrec->cutoff_scheme == ecutsVERLET)
1146 /* Here the master rank decides if all ranks will use GPUs */
1147 bUseGPU = (hwinfo->gpu_info.ncuda_dev_compatible > 0 ||
1148 getenv("GMX_EMULATE_GPU") != NULL);
1150 /* TODO add GPU kernels for this and replace this check by:
1151 * (bUseGPU && (ir->vdwtype == evdwPME &&
1152 * ir->ljpme_combination_rule == eljpmeLB))
1153 * update the message text and the content of nbnxn_acceleration_supported.
1156 !nbnxn_acceleration_supported(fplog, cr, inputrec, bUseGPU))
1158 /* Fallback message printed by nbnxn_acceleration_supported */
1161 gmx_fatal(FARGS, "GPU acceleration requested, but not supported with the given input settings");
1166 prepare_verlet_scheme(fplog, cr,
1167 inputrec, nstlist_cmdline, mtop, state->box,
1172 if (nstlist_cmdline > 0)
1174 gmx_fatal(FARGS, "Can not set nstlist with the group cut-off scheme");
1177 if (hwinfo->gpu_info.ncuda_dev_compatible > 0)
1179 md_print_warn(cr, fplog,
1180 "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
1181 " To use a GPU, set the mdp option: cutoff-scheme = Verlet\n"
1182 " (for quick performance testing you can use the -testverlet option)\n");
1187 gmx_fatal(FARGS, "GPU requested, but can't be used without cutoff-scheme=Verlet");
1190 #ifdef GMX_TARGET_BGQ
1191 md_print_warn(cr, fplog,
1192 "NOTE: There is no SIMD implementation of the group scheme kernels on\n"
1193 " BlueGene/Q. You will observe better performance from using the\n"
1194 " Verlet cut-off scheme.\n");
1199 /* Check for externally set OpenMP affinity and turn off internal
1200 * pinning if any is found. We need to do this check early to tell
1201 * thread-MPI whether it should do pinning when spawning threads.
1202 * TODO: the above no longer holds, we should move these checks down
1204 gmx_omp_check_thread_affinity(fplog, cr, hw_opt);
1206 /* Check and update the hardware options for internal consistency */
1207 check_and_update_hw_opt_1(hw_opt, SIMMASTER(cr));
1211 #ifdef GMX_THREAD_MPI
1212 /* Early check for externally set process affinity.
1213 * With thread-MPI this is needed as pinning might get turned off,
1214 * which needs to be known before starting thread-MPI.
1215 * With thread-MPI hw_opt is processed here on the master rank
1216 * and passed to the other ranks later, so we only do this on master.
1218 gmx_check_thread_affinity_set(fplog,
1220 hw_opt, hwinfo->nthreads_hw_avail, FALSE);
1223 #ifdef GMX_THREAD_MPI
1224 if (cr->npmenodes > 0 && hw_opt->nthreads_tmpi <= 0)
1226 gmx_fatal(FARGS, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME nodes");
1230 if (hw_opt->nthreads_omp_pme != hw_opt->nthreads_omp &&
1233 gmx_fatal(FARGS, "You need to explicitly specify the number of PME nodes (-npme) when using different number of OpenMP threads for PP and PME nodes");
1237 #ifdef GMX_THREAD_MPI
1240 /* Since the master knows the cut-off scheme, update hw_opt for this.
1241 * This is done later for normal MPI and also once more with tMPI
1242 * for all tMPI ranks.
1244 check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
1246 /* NOW the threads will be started: */
1247 hw_opt->nthreads_tmpi = get_nthreads_mpi(hwinfo,
1251 if (hw_opt->nthreads_tot > 0 && hw_opt->nthreads_omp <= 0)
1253 hw_opt->nthreads_omp = hw_opt->nthreads_tot/hw_opt->nthreads_tmpi;
1256 if (hw_opt->nthreads_tmpi > 1)
1258 /* now start the threads. */
1259 cr = mdrunner_start_threads(hw_opt, fplog, cr_old, nfile, fnm,
1260 oenv, bVerbose, bCompact, nstglobalcomm,
1261 ddxyz, dd_node_order, rdd, rconstr,
1262 dddlb_opt, dlb_scale, ddcsx, ddcsy, ddcsz,
1263 nbpu_opt, nstlist_cmdline,
1264 nsteps_cmdline, nstepout, resetstep, nmultisim,
1265 repl_ex_nst, repl_ex_nex, repl_ex_seed, pforce,
1266 cpt_period, max_hours, deviceOptions,
1268 /* the main thread continues here with a new cr. We don't deallocate
1269 the old cr because other threads may still be reading it. */
1272 gmx_comm("Failed to spawn threads");
1277 /* END OF CAUTION: cr is now reliable */
1279 /* g_membed initialisation *
1280 * Because we change the mtop, init_membed is called before the init_parallel *
1281 * (in case we ever want to make it run in parallel) */
1282 if (opt2bSet("-membed", nfile, fnm))
1286 fprintf(stderr, "Initializing membed");
1288 membed = init_membed(fplog, nfile, fnm, mtop, inputrec, state, cr, &cpt_period);
1293 /* now broadcast everything to the non-master nodes/threads: */
1294 init_parallel(cr, inputrec, mtop);
1298 pr_inputrec(fplog, 0, "Input Parameters", inputrec, FALSE);
1301 /* now make sure the state is initialized and propagated */
1302 set_state_entries(state, inputrec);
1304 /* A parallel command line option consistency check that we can
1305 only do after any threads have started. */
1307 (ddxyz[XX] > 1 || ddxyz[YY] > 1 || ddxyz[ZZ] > 1 || cr->npmenodes > 0))
1310 "The -dd or -npme option request a parallel simulation, "
1312 "but %s was compiled without threads or MPI enabled"
1314 #ifdef GMX_THREAD_MPI
1315 "but the number of threads (option -nt) is 1"
1317 "but %s was not started through mpirun/mpiexec or only one process was requested through mpirun/mpiexec"
1324 if ((Flags & MD_RERUN) &&
1325 (EI_ENERGY_MINIMIZATION(inputrec->eI) || eiNM == inputrec->eI))
1327 gmx_fatal(FARGS, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
1330 if (can_use_allvsall(inputrec, TRUE, cr, fplog) && DOMAINDECOMP(cr))
1332 gmx_fatal(FARGS, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
1335 if (!(EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype)))
1337 if (cr->npmenodes > 0)
1339 gmx_fatal_collective(FARGS, cr, NULL,
1340 "PME nodes are requested, but the system does not use PME electrostatics or LJ-PME");
1349 fcRegisterSteps(inputrec->nsteps, inputrec->init_step);
1353 /* NMR restraints must be initialized before load_checkpoint,
1354 * since with time averaging the history is added to t_state.
1355 * For proper consistency check we therefore need to extend
1357 * So the PME-only nodes (if present) will also initialize
1358 * the distance restraints.
1362 /* This needs to be called before read_checkpoint to extend the state */
1363 init_disres(fplog, mtop, inputrec, cr, fcd, state, repl_ex_nst > 0);
1365 init_orires(fplog, mtop, state->x, inputrec, cr, &(fcd->orires),
1368 if (DEFORM(*inputrec))
1370 /* Store the deform reference box before reading the checkpoint */
1373 copy_mat(state->box, box);
1377 gmx_bcast(sizeof(box), box, cr);
1379 /* Because we do not have the update struct available yet
1380 * in which the reference values should be stored,
1381 * we store them temporarily in static variables.
1382 * This should be thread safe, since they are only written once
1383 * and with identical values.
1385 tMPI_Thread_mutex_lock(&deform_init_box_mutex);
1386 deform_init_init_step_tpx = inputrec->init_step;
1387 copy_mat(box, deform_init_box_tpx);
1388 tMPI_Thread_mutex_unlock(&deform_init_box_mutex);
1391 if (opt2bSet("-cpi", nfile, fnm))
1393 /* Check if checkpoint file exists before doing continuation.
1394 * This way we can use identical input options for the first and subsequent runs...
1396 if (gmx_fexist_master(opt2fn_master("-cpi", nfile, fnm, cr), cr) )
1398 load_checkpoint(opt2fn_master("-cpi", nfile, fnm, cr), &fplog,
1400 inputrec, state, &bReadEkin,
1401 (Flags & MD_APPENDFILES),
1402 (Flags & MD_APPENDFILESSET));
1406 Flags |= MD_READ_EKIN;
1411 if (((MASTER(cr) || (Flags & MD_SEPPOT)) && (Flags & MD_APPENDFILES))
1412 #ifdef GMX_THREAD_MPI
1413 /* With thread MPI only the master node/thread exists in mdrun.c,
1414 * therefore non-master nodes need to open the "seppot" log file here.
1416 || (!MASTER(cr) && (Flags & MD_SEPPOT))
1420 gmx_log_open(ftp2fn(efLOG, nfile, fnm), cr, !(Flags & MD_SEPPOT),
1424 /* override nsteps with value from cmdline */
1425 override_nsteps_cmdline(fplog, nsteps_cmdline, inputrec, cr);
1429 copy_mat(state->box, box);
1434 gmx_bcast(sizeof(box), box, cr);
1437 /* Essential dynamics */
1438 if (opt2bSet("-ei", nfile, fnm))
1440 /* Open input and output files, allocate space for ED data structure */
1441 ed = ed_open(mtop->natoms, &state->edsamstate, nfile, fnm, Flags, oenv, cr);
1444 if (PAR(cr) && !(EI_TPI(inputrec->eI) ||
1445 inputrec->eI == eiNM))
1447 cr->dd = init_domain_decomposition(fplog, cr, Flags, ddxyz, rdd, rconstr,
1448 dddlb_opt, dlb_scale,
1449 ddcsx, ddcsy, ddcsz,
1452 &ddbox, &npme_major, &npme_minor);
1454 make_dd_communicators(fplog, cr, dd_node_order);
1456 /* Set overallocation to avoid frequent reallocation of arrays */
1457 set_over_alloc_dd(TRUE);
1461 /* PME, if used, is done on all nodes with 1D decomposition */
1463 cr->duty = (DUTY_PP | DUTY_PME);
1467 if (inputrec->ePBC == epbcSCREW)
1470 "pbc=%s is only implemented with domain decomposition",
1471 epbc_names[inputrec->ePBC]);
1477 /* After possible communicator splitting in make_dd_communicators.
1478 * we can set up the intra/inter node communication.
1480 gmx_setup_nodecomm(fplog, cr);
1483 /* Initialize per-physical-node MPI process/thread ID and counters. */
1484 gmx_init_intranode_counters(cr);
1487 md_print_info(cr, fplog, "Using %d MPI %s\n",
1489 #ifdef GMX_THREAD_MPI
1490 cr->nnodes == 1 ? "thread" : "threads"
1492 cr->nnodes == 1 ? "process" : "processes"
1498 /* Check and update hw_opt for the cut-off scheme */
1499 check_and_update_hw_opt_2(hw_opt, inputrec->cutoff_scheme);
1501 gmx_omp_nthreads_init(fplog, cr,
1502 hwinfo->nthreads_hw_avail,
1503 hw_opt->nthreads_omp,
1504 hw_opt->nthreads_omp_pme,
1505 (cr->duty & DUTY_PP) == 0,
1506 inputrec->cutoff_scheme == ecutsVERLET);
1510 /* The master rank decided on the use of GPUs,
1511 * broadcast this information to all ranks.
1513 gmx_bcast_sim(sizeof(bUseGPU), &bUseGPU, cr);
1518 if (cr->npmenodes == -1)
1520 /* Don't automatically use PME-only nodes with GPUs */
1524 /* Select GPU id's to use */
1525 gmx_select_gpu_ids(fplog, cr, &hwinfo->gpu_info, bForceUseGPU,
1530 /* Ignore (potentially) manually selected GPUs */
1531 hw_opt->gpu_opt.ncuda_dev_use = 0;
1534 /* check consistency across ranks of things like SIMD
1535 * support and number of GPUs selected */
1536 gmx_check_hw_runconf_consistency(fplog, hwinfo, cr, hw_opt, bUseGPU);
1538 if (DOMAINDECOMP(cr))
1540 /* When we share GPUs over ranks, we need to know this for the DLB */
1541 dd_setup_dlb_resource_sharing(cr, hwinfo, hw_opt);
1544 /* getting number of PP/PME threads
1545 PME: env variable should be read only on one node to make sure it is
1546 identical everywhere;
1548 /* TODO nthreads_pp is only used for pinning threads.
1549 * This is a temporary solution until we have a hw topology library.
1551 nthreads_pp = gmx_omp_nthreads_get(emntNonbonded);
1552 nthreads_pme = gmx_omp_nthreads_get(emntPME);
1554 wcycle = wallcycle_init(fplog, resetstep, cr, nthreads_pp, nthreads_pme);
1558 /* Master synchronizes its value of reset_counters with all nodes
1559 * including PME only nodes */
1560 reset_counters = wcycle_get_reset_counters(wcycle);
1561 gmx_bcast_sim(sizeof(reset_counters), &reset_counters, cr);
1562 wcycle_set_reset_counters(wcycle, reset_counters);
1566 if (cr->duty & DUTY_PP)
1568 bcast_state(cr, state);
1570 /* Initiate forcerecord */
1572 fr->hwinfo = hwinfo;
1573 fr->gpu_opt = &hw_opt->gpu_opt;
1574 init_forcerec(fplog, oenv, fr, fcd, inputrec, mtop, cr, box,
1575 opt2fn("-table", nfile, fnm),
1576 opt2fn("-tabletf", nfile, fnm),
1577 opt2fn("-tablep", nfile, fnm),
1578 opt2fn("-tableb", nfile, fnm),
1583 /* version for PCA_NOT_READ_NODE (see md.c) */
1584 /*init_forcerec(fplog,fr,fcd,inputrec,mtop,cr,box,FALSE,
1585 "nofile","nofile","nofile","nofile",FALSE,pforce);
1587 fr->bSepDVDL = ((Flags & MD_SEPPOT) == MD_SEPPOT);
1589 /* Initialize QM-MM */
1592 init_QMMMrec(cr, mtop, inputrec, fr);
1595 /* Initialize the mdatoms structure.
1596 * mdatoms is not filled with atom data,
1597 * as this can not be done now with domain decomposition.
1599 mdatoms = init_mdatoms(fplog, mtop, inputrec->efep != efepNO);
1601 /* Initialize the virtual site communication */
1602 vsite = init_vsite(mtop, cr, FALSE);
1604 calc_shifts(box, fr->shift_vec);
1606 /* With periodic molecules the charge groups should be whole at start up
1607 * and the virtual sites should not be far from their proper positions.
1609 if (!inputrec->bContinuation && MASTER(cr) &&
1610 !(inputrec->ePBC != epbcNONE && inputrec->bPeriodicMols))
1612 /* Make molecules whole at start of run */
1613 if (fr->ePBC != epbcNONE)
1615 do_pbc_first_mtop(fplog, inputrec->ePBC, box, mtop, state->x);
1619 /* Correct initial vsite positions are required
1620 * for the initial distribution in the domain decomposition
1621 * and for the initial shell prediction.
1623 construct_vsites_mtop(vsite, mtop, state->x);
1627 if (EEL_PME(fr->eeltype) || EVDW_PME(fr->vdwtype))
1629 ewaldcoeff_q = fr->ewaldcoeff_q;
1630 ewaldcoeff_lj = fr->ewaldcoeff_lj;
1631 pmedata = &fr->pmedata;
1640 /* This is a PME only node */
1642 /* We don't need the state */
1645 ewaldcoeff_q = calc_ewaldcoeff_q(inputrec->rcoulomb, inputrec->ewald_rtol);
1646 ewaldcoeff_lj = calc_ewaldcoeff_lj(inputrec->rvdw, inputrec->ewald_rtol_lj);
1650 if (hw_opt->thread_affinity != threadaffOFF)
1652 /* Before setting affinity, check whether the affinity has changed
1653 * - which indicates that probably the OpenMP library has changed it
1654 * since we first checked).
1656 gmx_check_thread_affinity_set(fplog, cr,
1657 hw_opt, hwinfo->nthreads_hw_avail, TRUE);
1659 /* Set the CPU affinity */
1660 gmx_set_thread_affinity(fplog, cr, hw_opt, hwinfo);
1663 /* Initiate PME if necessary,
1664 * either on all nodes or on dedicated PME nodes only. */
1665 if (EEL_PME(inputrec->coulombtype) || EVDW_PME(inputrec->vdwtype))
1669 nChargePerturbed = mdatoms->nChargePerturbed;
1670 if (EVDW_PME(inputrec->vdwtype))
1672 nTypePerturbed = mdatoms->nTypePerturbed;
1675 if (cr->npmenodes > 0)
1677 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1678 gmx_bcast_sim(sizeof(nChargePerturbed), &nChargePerturbed, cr);
1679 gmx_bcast_sim(sizeof(nTypePerturbed), &nTypePerturbed, cr);
1682 if (cr->duty & DUTY_PME)
1684 status = gmx_pme_init(pmedata, cr, npme_major, npme_minor, inputrec,
1685 mtop ? mtop->natoms : 0, nChargePerturbed, nTypePerturbed,
1686 (Flags & MD_REPRODUCIBLE), nthreads_pme);
1689 gmx_fatal(FARGS, "Error %d initializing PME", status);
1695 if (integrator[inputrec->eI].func == do_md)
1697 /* Turn on signal handling on all nodes */
1699 * (A user signal from the PME nodes (if any)
1700 * is communicated to the PP nodes.
1702 signal_handler_install();
1705 if (cr->duty & DUTY_PP)
1707 /* Assumes uniform use of the number of OpenMP threads */
1708 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntDefault));
1710 if (inputrec->ePull != epullNO)
1712 /* Initialize pull code */
1713 init_pull(fplog, inputrec, nfile, fnm, mtop, cr, oenv, inputrec->fepvals->init_lambda,
1714 EI_DYNAMICS(inputrec->eI) && MASTER(cr), Flags);
1719 /* Initialize enforced rotation code */
1720 init_rot(fplog, inputrec, nfile, fnm, cr, state->x, box, mtop, oenv,
1724 if (inputrec->eSwapCoords != eswapNO)
1726 /* Initialize ion swapping code */
1727 init_swapcoords(fplog, bVerbose, inputrec, opt2fn_master("-swap", nfile, fnm, cr),
1728 mtop, state->x, state->box, &state->swapstate, cr, oenv, Flags);
1731 constr = init_constraints(fplog, mtop, inputrec, ed, state, cr);
1733 if (DOMAINDECOMP(cr))
1735 dd_init_bondeds(fplog, cr->dd, mtop, vsite, inputrec,
1736 Flags & MD_DDBONDCHECK, fr->cginfo_mb);
1738 set_dd_parameters(fplog, cr->dd, dlb_scale, inputrec, &ddbox);
1740 setup_dd_grid(fplog, cr->dd);
1743 /* Now do whatever the user wants us to do (how flexible...) */
1744 integrator[inputrec->eI].func(fplog, cr, nfile, fnm,
1745 oenv, bVerbose, bCompact,
1748 nstepout, inputrec, mtop,
1750 mdatoms, nrnb, wcycle, ed, fr,
1751 repl_ex_nst, repl_ex_nex, repl_ex_seed,
1753 cpt_period, max_hours,
1756 walltime_accounting);
1758 if (inputrec->ePull != epullNO)
1760 finish_pull(inputrec->pull);
1765 finish_rot(inputrec->rot);
1772 walltime_accounting = walltime_accounting_init(gmx_omp_nthreads_get(emntPME));
1773 gmx_pmeonly(*pmedata, cr, nrnb, wcycle, walltime_accounting, ewaldcoeff_q, ewaldcoeff_lj, inputrec);
1776 wallcycle_stop(wcycle, ewcRUN);
1778 /* Finish up, write some stuff
1779 * if rerunMD, don't write last frame again
1781 finish_run(fplog, cr,
1782 inputrec, nrnb, wcycle, walltime_accounting,
1783 fr != NULL && fr->nbv != NULL && fr->nbv->bUseGPU ?
1784 nbnxn_cuda_get_timings(fr->nbv->cu_nbv) : NULL,
1785 EI_DYNAMICS(inputrec->eI) && !MULTISIM(cr));
1788 /* Free GPU memory and context */
1789 free_gpu_resources(fr, cr);
1791 if (opt2bSet("-membed", nfile, fnm))
1796 gmx_hardware_info_free(hwinfo);
1798 /* Does what it says */
1799 print_date_and_time(fplog, cr->nodeid, "Finished mdrun", walltime_accounting);
1800 walltime_accounting_destroy(walltime_accounting);
1802 /* Close logfile already here if we were appending to it */
1803 if (MASTER(cr) && (Flags & MD_APPENDFILES))
1805 gmx_log_close(fplog);
1808 rc = (int)gmx_get_stop_condition();
1810 #ifdef GMX_THREAD_MPI
1811 /* we need to join all threads. The sub-threads join when they
1812 exit this function, but the master thread needs to be told to
1814 if (PAR(cr) && MASTER(cr))