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) 2011,2012,2013,2014,2015,2016, 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.
39 * \brief Implements the MD runner routine calling all integrators.
41 * \author David van der Spoel <david.vanderspoel@icm.uu.se>
42 * \ingroup module_mdlib
57 #include "gromacs/commandline/filenm.h"
58 #include "gromacs/domdec/domdec.h"
59 #include "gromacs/domdec/domdec_struct.h"
60 #include "gromacs/essentialdynamics/edsam.h"
61 #include "gromacs/ewald/pme.h"
62 #include "gromacs/fileio/checkpoint.h"
63 #include "gromacs/fileio/oenv.h"
64 #include "gromacs/fileio/tpxio.h"
65 #include "gromacs/gmxlib/md_logging.h"
66 #include "gromacs/gmxlib/network.h"
67 #include "gromacs/gpu_utils/gpu_utils.h"
68 #include "gromacs/hardware/cpuinfo.h"
69 #include "gromacs/hardware/detecthardware.h"
70 #include "gromacs/listed-forces/disre.h"
71 #include "gromacs/listed-forces/orires.h"
72 #include "gromacs/math/calculate-ewald-splitting-coefficient.h"
73 #include "gromacs/math/functions.h"
74 #include "gromacs/math/utilities.h"
75 #include "gromacs/math/vec.h"
76 #include "gromacs/mdlib/calc_verletbuf.h"
77 #include "gromacs/mdlib/constr.h"
78 #include "gromacs/mdlib/force.h"
79 #include "gromacs/mdlib/forcerec.h"
80 #include "gromacs/mdlib/gmx_omp_nthreads.h"
81 #include "gromacs/mdlib/integrator.h"
82 #include "gromacs/mdlib/main.h"
83 #include "gromacs/mdlib/md_support.h"
84 #include "gromacs/mdlib/mdatoms.h"
85 #include "gromacs/mdlib/mdrun.h"
86 #include "gromacs/mdlib/minimize.h"
87 #include "gromacs/mdlib/nbnxn_search.h"
88 #include "gromacs/mdlib/qmmm.h"
89 #include "gromacs/mdlib/sighandler.h"
90 #include "gromacs/mdlib/sim_util.h"
91 #include "gromacs/mdlib/tpi.h"
92 #include "gromacs/mdrunutility/threadaffinity.h"
93 #include "gromacs/mdtypes/commrec.h"
94 #include "gromacs/mdtypes/inputrec.h"
95 #include "gromacs/mdtypes/md_enums.h"
96 #include "gromacs/mdtypes/state.h"
97 #include "gromacs/pbcutil/pbc.h"
98 #include "gromacs/pulling/pull.h"
99 #include "gromacs/pulling/pull_rotation.h"
100 #include "gromacs/timing/wallcycle.h"
101 #include "gromacs/topology/mtop_util.h"
102 #include "gromacs/trajectory/trajectoryframe.h"
103 #include "gromacs/utility/cstringutil.h"
104 #include "gromacs/utility/exceptions.h"
105 #include "gromacs/utility/fatalerror.h"
106 #include "gromacs/utility/gmxassert.h"
107 #include "gromacs/utility/gmxmpi.h"
108 #include "gromacs/utility/pleasecite.h"
109 #include "gromacs/utility/smalloc.h"
115 #include "resource-division.h"
118 #include "corewrap.h"
121 //! First step used in pressure scaling
122 gmx_int64_t deform_init_init_step_tpx
;
123 //! Initial box for pressure scaling
124 matrix deform_init_box_tpx
;
125 //! MPI variable for use in pressure scaling
126 tMPI_Thread_mutex_t deform_init_box_mutex
= TMPI_THREAD_MUTEX_INITIALIZER
;
129 /* The minimum number of atoms per tMPI thread. With fewer atoms than this,
130 * the number of threads will get lowered.
132 #define MIN_ATOMS_PER_MPI_THREAD 90
133 #define MIN_ATOMS_PER_GPU 900
135 struct mdrunner_arglist
142 const gmx_output_env_t
*oenv
;
150 const char *dddlb_opt
;
155 const char *nbpu_opt
;
157 gmx_int64_t nsteps_cmdline
;
172 /* The function used for spawning threads. Extracts the mdrunner()
173 arguments from its one argument and calls mdrunner(), after making
175 static void mdrunner_start_fn(void *arg
)
179 struct mdrunner_arglist
*mda
= (struct mdrunner_arglist
*)arg
;
180 struct mdrunner_arglist mc
= *mda
; /* copy the arg list to make sure
181 that it's thread-local. This doesn't
182 copy pointed-to items, of course,
183 but those are all const. */
184 t_commrec
*cr
; /* we need a local version of this */
188 fnm
= dup_tfn(mc
.nfile
, mc
.fnm
);
190 cr
= reinitialize_commrec_for_this_thread(mc
.cr
);
197 gmx::mdrunner(&mc
.hw_opt
, fplog
, cr
, mc
.nfile
, fnm
, mc
.oenv
,
198 mc
.bVerbose
, mc
.nstglobalcomm
,
199 mc
.ddxyz
, mc
.dd_rank_order
, mc
.npme
, mc
.rdd
,
200 mc
.rconstr
, mc
.dddlb_opt
, mc
.dlb_scale
,
201 mc
.ddcsx
, mc
.ddcsy
, mc
.ddcsz
,
202 mc
.nbpu_opt
, mc
.nstlist_cmdline
,
203 mc
.nsteps_cmdline
, mc
.nstepout
, mc
.resetstep
,
204 mc
.nmultisim
, mc
.repl_ex_nst
, mc
.repl_ex_nex
, mc
.repl_ex_seed
, mc
.pforce
,
205 mc
.cpt_period
, mc
.max_hours
, mc
.imdport
, mc
.Flags
);
207 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR
;
211 /* called by mdrunner() to start a specific number of threads (including
212 the main thread) for thread-parallel runs. This in turn calls mdrunner()
214 All options besides nthreads are the same as for mdrunner(). */
215 static t_commrec
*mdrunner_start_threads(gmx_hw_opt_t
*hw_opt
,
216 FILE *fplog
, t_commrec
*cr
, int nfile
,
217 const t_filenm fnm
[], const gmx_output_env_t
*oenv
, gmx_bool bVerbose
,
219 ivec ddxyz
, int dd_rank_order
, int npme
,
220 real rdd
, real rconstr
,
221 const char *dddlb_opt
, real dlb_scale
,
222 const char *ddcsx
, const char *ddcsy
, const char *ddcsz
,
223 const char *nbpu_opt
, int nstlist_cmdline
,
224 gmx_int64_t nsteps_cmdline
,
225 int nstepout
, int resetstep
,
226 int nmultisim
, int repl_ex_nst
, int repl_ex_nex
, int repl_ex_seed
,
227 real pforce
, real cpt_period
, real max_hours
,
231 struct mdrunner_arglist
*mda
;
232 t_commrec
*crn
; /* the new commrec */
235 /* first check whether we even need to start tMPI */
236 if (hw_opt
->nthreads_tmpi
< 2)
241 /* a few small, one-time, almost unavoidable memory leaks: */
243 fnmn
= dup_tfn(nfile
, fnm
);
245 /* fill the data structure to pass as void pointer to thread start fn */
246 /* hw_opt contains pointers, which should all be NULL at this stage */
247 mda
->hw_opt
= *hw_opt
;
253 mda
->bVerbose
= bVerbose
;
254 mda
->nstglobalcomm
= nstglobalcomm
;
255 mda
->ddxyz
[XX
] = ddxyz
[XX
];
256 mda
->ddxyz
[YY
] = ddxyz
[YY
];
257 mda
->ddxyz
[ZZ
] = ddxyz
[ZZ
];
258 mda
->dd_rank_order
= dd_rank_order
;
261 mda
->rconstr
= rconstr
;
262 mda
->dddlb_opt
= dddlb_opt
;
263 mda
->dlb_scale
= dlb_scale
;
267 mda
->nbpu_opt
= nbpu_opt
;
268 mda
->nstlist_cmdline
= nstlist_cmdline
;
269 mda
->nsteps_cmdline
= nsteps_cmdline
;
270 mda
->nstepout
= nstepout
;
271 mda
->resetstep
= resetstep
;
272 mda
->nmultisim
= nmultisim
;
273 mda
->repl_ex_nst
= repl_ex_nst
;
274 mda
->repl_ex_nex
= repl_ex_nex
;
275 mda
->repl_ex_seed
= repl_ex_seed
;
276 mda
->pforce
= pforce
;
277 mda
->cpt_period
= cpt_period
;
278 mda
->max_hours
= max_hours
;
281 /* now spawn new threads that start mdrunner_start_fn(), while
282 the main thread returns, we set thread affinity later */
283 ret
= tMPI_Init_fn(TRUE
, hw_opt
->nthreads_tmpi
, TMPI_AFFINITY_NONE
,
284 mdrunner_start_fn
, (void*)(mda
) );
285 if (ret
!= TMPI_SUCCESS
)
290 crn
= reinitialize_commrec_for_this_thread(cr
);
294 #endif /* GMX_THREAD_MPI */
297 /*! \brief Cost of non-bonded kernels
299 * We determine the extra cost of the non-bonded kernels compared to
300 * a reference nstlist value of 10 (which is the default in grompp).
302 static const int nbnxnReferenceNstlist
= 10;
303 //! The values to try when switching
304 const int nstlist_try
[] = { 20, 25, 40 };
305 //! Number of elements in the neighborsearch list trials.
306 #define NNSTL sizeof(nstlist_try)/sizeof(nstlist_try[0])
307 /* Increase nstlist until the non-bonded cost increases more than listfac_ok,
308 * but never more than listfac_max.
309 * A standard (protein+)water system at 300K with PME ewald_rtol=1e-5
310 * needs 1.28 at rcoulomb=0.9 and 1.24 at rcoulomb=1.0 to get to nstlist=40.
311 * Note that both CPU and GPU factors are conservative. Performance should
312 * not go down due to this tuning, except with a relatively slow GPU.
313 * On the other hand, at medium/high parallelization or with fast GPUs
314 * nstlist will not be increased enough to reach optimal performance.
316 /* CPU: pair-search is about a factor 1.5 slower than the non-bonded kernel */
317 //! Max OK performance ratio beween force calc and neighbor searching
318 static const float nbnxn_cpu_listfac_ok
= 1.05;
319 //! Too high performance ratio beween force calc and neighbor searching
320 static const float nbnxn_cpu_listfac_max
= 1.09;
321 /* CPU: pair-search is about a factor 2-3 slower than the non-bonded kernel */
322 //! Max OK performance ratio beween force calc and neighbor searching
323 static const float nbnxn_knl_listfac_ok
= 1.22;
324 //! Too high performance ratio beween force calc and neighbor searching
325 static const float nbnxn_knl_listfac_max
= 1.3;
326 /* GPU: pair-search is a factor 1.5-3 slower than the non-bonded kernel */
327 //! Max OK performance ratio beween force calc and neighbor searching
328 static const float nbnxn_gpu_listfac_ok
= 1.20;
329 //! Too high performance ratio beween force calc and neighbor searching
330 static const float nbnxn_gpu_listfac_max
= 1.30;
332 /*! \brief Try to increase nstlist when using the Verlet cut-off scheme */
333 static void increase_nstlist(FILE *fp
, t_commrec
*cr
,
334 t_inputrec
*ir
, int nstlist_cmdline
,
335 const gmx_mtop_t
*mtop
, matrix box
,
336 gmx_bool bGPU
, const gmx::CpuInfo
&cpuinfo
)
338 float listfac_ok
, listfac_max
;
339 int nstlist_orig
, nstlist_prev
;
340 verletbuf_list_setup_t ls
;
341 real rlistWithReferenceNstlist
, rlist_inc
, rlist_ok
, rlist_max
;
342 real rlist_new
, rlist_prev
;
343 size_t nstlist_ind
= 0;
345 gmx_bool bBox
, bDD
, bCont
;
346 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";
347 const char *nve_err
= "Can not increase nstlist because an NVE ensemble is used";
348 const char *vbd_err
= "Can not increase nstlist because verlet-buffer-tolerance is not set or used";
349 const char *box_err
= "Can not increase nstlist because the box is too small";
350 const char *dd_err
= "Can not increase nstlist because of domain decomposition limitations";
353 if (nstlist_cmdline
<= 0)
355 if (ir
->nstlist
== 1)
357 /* The user probably set nstlist=1 for a reason,
358 * don't mess with the settings.
363 if (fp
!= NULL
&& bGPU
&& ir
->nstlist
< nstlist_try
[0])
365 fprintf(fp
, nstl_gpu
, ir
->nstlist
);
368 while (nstlist_ind
< NNSTL
&& ir
->nstlist
>= nstlist_try
[nstlist_ind
])
372 if (nstlist_ind
== NNSTL
)
374 /* There are no larger nstlist value to try */
379 if (EI_MD(ir
->eI
) && ir
->etc
== etcNO
)
383 fprintf(stderr
, "%s\n", nve_err
);
387 fprintf(fp
, "%s\n", nve_err
);
393 if (ir
->verletbuf_tol
== 0 && bGPU
)
395 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");
398 if (ir
->verletbuf_tol
< 0)
402 fprintf(stderr
, "%s\n", vbd_err
);
406 fprintf(fp
, "%s\n", vbd_err
);
414 listfac_ok
= nbnxn_gpu_listfac_ok
;
415 listfac_max
= nbnxn_gpu_listfac_max
;
417 else if (cpuinfo
.feature(gmx::CpuInfo::Feature::X86_Avx512ER
))
419 listfac_ok
= nbnxn_knl_listfac_ok
;
420 listfac_max
= nbnxn_knl_listfac_max
;
424 listfac_ok
= nbnxn_cpu_listfac_ok
;
425 listfac_max
= nbnxn_cpu_listfac_max
;
428 nstlist_orig
= ir
->nstlist
;
429 if (nstlist_cmdline
> 0)
433 sprintf(buf
, "Getting nstlist=%d from command line option",
436 ir
->nstlist
= nstlist_cmdline
;
439 verletbuf_get_list_setup(TRUE
, bGPU
, &ls
);
441 /* Allow rlist to make the list a given factor larger than the list
442 * would be with the reference value for nstlist (10).
444 nstlist_prev
= ir
->nstlist
;
445 ir
->nstlist
= nbnxnReferenceNstlist
;
446 calc_verlet_buffer_size(mtop
, det(box
), ir
, -1, &ls
, NULL
,
447 &rlistWithReferenceNstlist
);
448 ir
->nstlist
= nstlist_prev
;
450 /* Determine the pair list size increase due to zero interactions */
451 rlist_inc
= nbnxn_get_rlist_effective_inc(ls
.cluster_size_j
,
452 mtop
->natoms
/det(box
));
453 rlist_ok
= (rlistWithReferenceNstlist
+ rlist_inc
)*std::cbrt(listfac_ok
) - rlist_inc
;
454 rlist_max
= (rlistWithReferenceNstlist
+ rlist_inc
)*std::cbrt(listfac_max
) - rlist_inc
;
457 fprintf(debug
, "nstlist tuning: rlist_inc %.3f rlist_ok %.3f rlist_max %.3f\n",
458 rlist_inc
, rlist_ok
, rlist_max
);
461 nstlist_prev
= nstlist_orig
;
462 rlist_prev
= ir
->rlist
;
465 if (nstlist_cmdline
<= 0)
467 ir
->nstlist
= nstlist_try
[nstlist_ind
];
470 /* Set the pair-list buffer size in ir */
471 calc_verlet_buffer_size(mtop
, det(box
), ir
, -1, &ls
, NULL
, &rlist_new
);
473 /* Does rlist fit in the box? */
474 bBox
= (gmx::square(rlist_new
) < max_cutoff2(ir
->ePBC
, box
));
476 if (bBox
&& DOMAINDECOMP(cr
))
478 /* Check if rlist fits in the domain decomposition */
479 if (inputrec2nboundeddim(ir
) < DIM
)
481 gmx_incons("Changing nstlist with domain decomposition and unbounded dimensions is not implemented yet");
483 copy_mat(box
, state_tmp
.box
);
484 bDD
= change_dd_cutoff(cr
, &state_tmp
, ir
, rlist_new
);
489 fprintf(debug
, "nstlist %d rlist %.3f bBox %d bDD %d\n",
490 ir
->nstlist
, rlist_new
, bBox
, bDD
);
495 if (nstlist_cmdline
<= 0)
497 if (bBox
&& bDD
&& rlist_new
<= rlist_max
)
499 /* Increase nstlist */
500 nstlist_prev
= ir
->nstlist
;
501 rlist_prev
= rlist_new
;
502 bCont
= (nstlist_ind
+1 < NNSTL
&& rlist_new
< rlist_ok
);
506 /* Stick with the previous nstlist */
507 ir
->nstlist
= nstlist_prev
;
508 rlist_new
= rlist_prev
;
520 gmx_warning(!bBox
? box_err
: dd_err
);
523 fprintf(fp
, "\n%s\n", bBox
? box_err
: dd_err
);
525 ir
->nstlist
= nstlist_orig
;
527 else if (ir
->nstlist
!= nstlist_orig
|| rlist_new
!= ir
->rlist
)
529 sprintf(buf
, "Changing nstlist from %d to %d, rlist from %g to %g",
530 nstlist_orig
, ir
->nstlist
,
531 ir
->rlist
, rlist_new
);
534 fprintf(stderr
, "%s\n\n", buf
);
538 fprintf(fp
, "%s\n\n", buf
);
540 ir
->rlist
= rlist_new
;
544 /*! \brief Initialize variables for Verlet scheme simulation */
545 static void prepare_verlet_scheme(FILE *fplog
,
549 const gmx_mtop_t
*mtop
,
552 const gmx::CpuInfo
&cpuinfo
)
554 /* For NVE simulations, we will retain the initial list buffer */
555 if (EI_DYNAMICS(ir
->eI
) &&
556 ir
->verletbuf_tol
> 0 &&
557 !(EI_MD(ir
->eI
) && ir
->etc
== etcNO
))
559 /* Update the Verlet buffer size for the current run setup */
560 verletbuf_list_setup_t ls
;
563 /* Here we assume SIMD-enabled kernels are being used. But as currently
564 * calc_verlet_buffer_size gives the same results for 4x8 and 4x4
565 * and 4x2 gives a larger buffer than 4x4, this is ok.
567 verletbuf_get_list_setup(TRUE
, bUseGPU
, &ls
);
569 calc_verlet_buffer_size(mtop
, det(box
), ir
, -1, &ls
, NULL
, &rlist_new
);
571 if (rlist_new
!= ir
->rlist
)
575 fprintf(fplog
, "\nChanging rlist from %g to %g for non-bonded %dx%d atom kernels\n\n",
576 ir
->rlist
, rlist_new
,
577 ls
.cluster_size_i
, ls
.cluster_size_j
);
579 ir
->rlist
= rlist_new
;
583 if (nstlist_cmdline
> 0 && (!EI_DYNAMICS(ir
->eI
) || ir
->verletbuf_tol
<= 0))
585 gmx_fatal(FARGS
, "Can not set nstlist without %s",
586 !EI_DYNAMICS(ir
->eI
) ? "dynamics" : "verlet-buffer-tolerance");
589 if (EI_DYNAMICS(ir
->eI
))
591 /* Set or try nstlist values */
592 increase_nstlist(fplog
, cr
, ir
, nstlist_cmdline
, mtop
, box
, bUseGPU
, cpuinfo
);
596 /*! \brief Override the nslist value in inputrec
598 * with value passed on the command line (if any)
600 static void override_nsteps_cmdline(FILE *fplog
,
601 gmx_int64_t nsteps_cmdline
,
608 /* override with anything else than the default -2 */
609 if (nsteps_cmdline
> -2)
611 char sbuf_steps
[STEPSTRSIZE
];
612 char sbuf_msg
[STRLEN
];
614 ir
->nsteps
= nsteps_cmdline
;
615 if (EI_DYNAMICS(ir
->eI
) && nsteps_cmdline
!= -1)
617 sprintf(sbuf_msg
, "Overriding nsteps with value passed on the command line: %s steps, %.3g ps",
618 gmx_step_str(nsteps_cmdline
, sbuf_steps
),
619 fabs(nsteps_cmdline
*ir
->delta_t
));
623 sprintf(sbuf_msg
, "Overriding nsteps with value passed on the command line: %s steps",
624 gmx_step_str(nsteps_cmdline
, sbuf_steps
));
627 md_print_warn(cr
, fplog
, "%s\n", sbuf_msg
);
629 else if (nsteps_cmdline
< -2)
631 gmx_fatal(FARGS
, "Invalid nsteps value passed on the command line: %d",
634 /* Do nothing if nsteps_cmdline == -2 */
640 //! \brief Return the correct integrator function.
641 static integrator_t
*my_integrator(unsigned int ei
)
650 if (!EI_DYNAMICS(ei
))
652 GMX_THROW(APIError("do_md integrator would be called for a non-dynamical integrator"));
667 GMX_THROW(APIError("do_tpi integrator would be called for a non-TPI integrator"));
671 GMX_THROW(NotImplementedError("SD2 integrator has been removed"));
673 GMX_THROW(APIError("Non existing integrator selected"));
677 int mdrunner(gmx_hw_opt_t
*hw_opt
,
678 FILE *fplog
, t_commrec
*cr
, int nfile
,
679 const t_filenm fnm
[], const gmx_output_env_t
*oenv
, gmx_bool bVerbose
,
681 ivec ddxyz
, int dd_rank_order
, int npme
, real rdd
, real rconstr
,
682 const char *dddlb_opt
, real dlb_scale
,
683 const char *ddcsx
, const char *ddcsy
, const char *ddcsz
,
684 const char *nbpu_opt
, int nstlist_cmdline
,
685 gmx_int64_t nsteps_cmdline
, int nstepout
, int resetstep
,
686 int gmx_unused nmultisim
, int repl_ex_nst
, int repl_ex_nex
,
687 int repl_ex_seed
, real pforce
, real cpt_period
, real max_hours
,
688 int imdport
, unsigned long Flags
)
690 gmx_bool bForceUseGPU
, bTryUseGPU
, bRerunMD
;
691 t_inputrec
*inputrec
;
692 t_state
*state
= NULL
;
694 gmx_ddbox_t ddbox
= {0};
695 int npme_major
, npme_minor
;
697 gmx_mtop_t
*mtop
= NULL
;
698 t_mdatoms
*mdatoms
= NULL
;
699 t_forcerec
*fr
= NULL
;
700 t_fcdata
*fcd
= NULL
;
701 real ewaldcoeff_q
= 0;
702 real ewaldcoeff_lj
= 0;
703 struct gmx_pme_t
**pmedata
= NULL
;
704 gmx_vsite_t
*vsite
= NULL
;
706 int nChargePerturbed
= -1, nTypePerturbed
= 0, status
;
707 gmx_wallcycle_t wcycle
;
708 gmx_walltime_accounting_t walltime_accounting
= NULL
;
710 gmx_int64_t reset_counters
;
711 gmx_edsam_t ed
= NULL
;
712 int nthreads_pme
= 1;
713 gmx_membed_t
* membed
= NULL
;
714 gmx_hw_info_t
*hwinfo
= NULL
;
715 /* The master rank decides early on bUseGPU and broadcasts this later */
716 gmx_bool bUseGPU
= FALSE
;
718 /* CAUTION: threads may be started later on in this function, so
719 cr doesn't reflect the final parallel state right now */
723 if (Flags
& MD_APPENDFILES
)
728 bRerunMD
= (Flags
& MD_RERUN
);
729 bForceUseGPU
= (strncmp(nbpu_opt
, "gpu", 3) == 0);
730 bTryUseGPU
= (strncmp(nbpu_opt
, "auto", 4) == 0) || bForceUseGPU
;
732 /* Detect hardware, gather information. This is an operation that is
733 * global for this process (MPI rank). */
734 hwinfo
= gmx_detect_hardware(fplog
, cr
, bTryUseGPU
);
736 gmx_print_detected_hardware(fplog
, cr
, hwinfo
);
740 /* Print references after all software/hardware printing */
741 please_cite(fplog
, "Abraham2015");
742 please_cite(fplog
, "Pall2015");
743 please_cite(fplog
, "Pronk2013");
744 please_cite(fplog
, "Hess2008b");
745 please_cite(fplog
, "Spoel2005a");
746 please_cite(fplog
, "Lindahl2001a");
747 please_cite(fplog
, "Berendsen95a");
753 /* Read (nearly) all data required for the simulation */
754 read_tpx_state(ftp2fn(efTPR
, nfile
, fnm
), inputrec
, state
, mtop
);
756 if (inputrec
->cutoff_scheme
== ecutsVERLET
)
758 /* Here the master rank decides if all ranks will use GPUs */
759 bUseGPU
= (hwinfo
->gpu_info
.n_dev_compatible
> 0 ||
760 getenv("GMX_EMULATE_GPU") != NULL
);
762 /* TODO add GPU kernels for this and replace this check by:
763 * (bUseGPU && (ir->vdwtype == evdwPME &&
764 * ir->ljpme_combination_rule == eljpmeLB))
765 * update the message text and the content of nbnxn_acceleration_supported.
768 !nbnxn_gpu_acceleration_supported(fplog
, cr
, inputrec
, bRerunMD
))
770 /* Fallback message printed by nbnxn_acceleration_supported */
773 gmx_fatal(FARGS
, "GPU acceleration requested, but not supported with the given input settings");
778 prepare_verlet_scheme(fplog
, cr
,
779 inputrec
, nstlist_cmdline
, mtop
, state
->box
,
780 bUseGPU
, *hwinfo
->cpuInfo
);
784 if (nstlist_cmdline
> 0)
786 gmx_fatal(FARGS
, "Can not set nstlist with the group cut-off scheme");
789 if (hwinfo
->gpu_info
.n_dev_compatible
> 0)
791 md_print_warn(cr
, fplog
,
792 "NOTE: GPU(s) found, but the current simulation can not use GPUs\n"
793 " To use a GPU, set the mdp option: cutoff-scheme = Verlet\n");
798 gmx_fatal(FARGS
, "GPU requested, but can't be used without cutoff-scheme=Verlet");
802 md_print_warn(cr
, fplog
,
803 "NOTE: There is no SIMD implementation of the group scheme kernels on\n"
804 " BlueGene/Q. You will observe better performance from using the\n"
805 " Verlet cut-off scheme.\n");
810 /* Check and update the hardware options for internal consistency */
811 check_and_update_hw_opt_1(hw_opt
, cr
, npme
);
813 /* Early check for externally set process affinity. */
814 gmx_check_thread_affinity_set(fplog
, cr
,
815 hw_opt
, hwinfo
->nthreads_hw_avail
, FALSE
);
820 if (npme
> 0 && hw_opt
->nthreads_tmpi
<= 0)
822 gmx_fatal(FARGS
, "You need to explicitly specify the number of MPI threads (-ntmpi) when using separate PME ranks");
825 /* Since the master knows the cut-off scheme, update hw_opt for this.
826 * This is done later for normal MPI and also once more with tMPI
827 * for all tMPI ranks.
829 check_and_update_hw_opt_2(hw_opt
, inputrec
->cutoff_scheme
);
831 /* NOW the threads will be started: */
832 hw_opt
->nthreads_tmpi
= get_nthreads_mpi(hwinfo
,
837 if (hw_opt
->nthreads_tmpi
> 1)
839 t_commrec
*cr_old
= cr
;
840 /* now start the threads. */
841 cr
= mdrunner_start_threads(hw_opt
, fplog
, cr_old
, nfile
, fnm
,
842 oenv
, bVerbose
, nstglobalcomm
,
843 ddxyz
, dd_rank_order
, npme
, rdd
, rconstr
,
844 dddlb_opt
, dlb_scale
, ddcsx
, ddcsy
, ddcsz
,
845 nbpu_opt
, nstlist_cmdline
,
846 nsteps_cmdline
, nstepout
, resetstep
, nmultisim
,
847 repl_ex_nst
, repl_ex_nex
, repl_ex_seed
, pforce
,
848 cpt_period
, max_hours
,
850 /* the main thread continues here with a new cr. We don't deallocate
851 the old cr because other threads may still be reading it. */
854 gmx_comm("Failed to spawn threads");
859 /* END OF CAUTION: cr is now reliable */
863 /* now broadcast everything to the non-master nodes/threads: */
864 init_parallel(cr
, inputrec
, mtop
);
866 /* The master rank decided on the use of GPUs,
867 * broadcast this information to all ranks.
869 gmx_bcast_sim(sizeof(bUseGPU
), &bUseGPU
, cr
);
874 pr_inputrec(fplog
, 0, "Input Parameters", inputrec
, FALSE
);
875 fprintf(fplog
, "\n");
878 /* now make sure the state is initialized and propagated */
879 set_state_entries(state
, inputrec
);
881 /* A parallel command line option consistency check that we can
882 only do after any threads have started. */
884 (ddxyz
[XX
] > 1 || ddxyz
[YY
] > 1 || ddxyz
[ZZ
] > 1 || npme
> 0))
887 "The -dd or -npme option request a parallel simulation, "
889 "but %s was compiled without threads or MPI enabled"
892 "but the number of MPI-threads (option -ntmpi) is not set or is 1"
894 "but %s was not started through mpirun/mpiexec or only one rank was requested through mpirun/mpiexec"
897 , output_env_get_program_display_name(oenv
)
902 (EI_ENERGY_MINIMIZATION(inputrec
->eI
) || eiNM
== inputrec
->eI
))
904 gmx_fatal(FARGS
, "The .mdp file specified an energy mininization or normal mode algorithm, and these are not compatible with mdrun -rerun");
907 if (can_use_allvsall(inputrec
, TRUE
, cr
, fplog
) && DOMAINDECOMP(cr
))
909 gmx_fatal(FARGS
, "All-vs-all loops do not work with domain decomposition, use a single MPI rank");
912 if (!(EEL_PME(inputrec
->coulombtype
) || EVDW_PME(inputrec
->vdwtype
)))
916 gmx_fatal_collective(FARGS
, cr
->mpi_comm_mysim
, MASTER(cr
),
917 "PME-only ranks are requested, but the system does not use PME for electrostatics or LJ");
923 if (bUseGPU
&& npme
< 0)
925 /* With GPUs we don't automatically use PME-only ranks. PME ranks can
926 * improve performance with many threads per GPU, since our OpenMP
927 * scaling is bad, but it's difficult to automate the setup.
935 fcRegisterSteps(inputrec
->nsteps
, inputrec
->init_step
);
939 /* NMR restraints must be initialized before load_checkpoint,
940 * since with time averaging the history is added to t_state.
941 * For proper consistency check we therefore need to extend
943 * So the PME-only nodes (if present) will also initialize
944 * the distance restraints.
948 /* This needs to be called before read_checkpoint to extend the state */
949 init_disres(fplog
, mtop
, inputrec
, cr
, fcd
, state
, repl_ex_nst
> 0);
951 init_orires(fplog
, mtop
, state
->x
, inputrec
, cr
, &(fcd
->orires
),
954 if (inputrecDeform(inputrec
))
956 /* Store the deform reference box before reading the checkpoint */
959 copy_mat(state
->box
, box
);
963 gmx_bcast(sizeof(box
), box
, cr
);
965 /* Because we do not have the update struct available yet
966 * in which the reference values should be stored,
967 * we store them temporarily in static variables.
968 * This should be thread safe, since they are only written once
969 * and with identical values.
971 tMPI_Thread_mutex_lock(&deform_init_box_mutex
);
972 deform_init_init_step_tpx
= inputrec
->init_step
;
973 copy_mat(box
, deform_init_box_tpx
);
974 tMPI_Thread_mutex_unlock(&deform_init_box_mutex
);
977 if (Flags
& MD_STARTFROMCPT
)
979 /* Check if checkpoint file exists before doing continuation.
980 * This way we can use identical input options for the first and subsequent runs...
984 load_checkpoint(opt2fn_master("-cpi", nfile
, fnm
, cr
), &fplog
,
986 inputrec
, state
, &bReadEkin
,
987 (Flags
& MD_APPENDFILES
),
988 (Flags
& MD_APPENDFILESSET
),
989 (Flags
& MD_REPRODUCIBLE
));
993 Flags
|= MD_READ_EKIN
;
997 if (MASTER(cr
) && (Flags
& MD_APPENDFILES
))
999 gmx_log_open(ftp2fn(efLOG
, nfile
, fnm
), cr
,
1003 /* override nsteps with value from cmdline */
1004 override_nsteps_cmdline(fplog
, nsteps_cmdline
, inputrec
, cr
);
1008 copy_mat(state
->box
, box
);
1013 gmx_bcast(sizeof(box
), box
, cr
);
1016 // TODO This should move to do_md(), because it only makes sense
1017 // with dynamical integrators, but there is no test coverage and
1018 // it interacts with constraints, somehow.
1019 /* Essential dynamics */
1020 if (opt2bSet("-ei", nfile
, fnm
))
1022 /* Open input and output files, allocate space for ED data structure */
1023 ed
= ed_open(mtop
->natoms
, &state
->edsamstate
, nfile
, fnm
, Flags
, oenv
, cr
);
1026 if (PAR(cr
) && !(EI_TPI(inputrec
->eI
) ||
1027 inputrec
->eI
== eiNM
))
1029 cr
->dd
= init_domain_decomposition(fplog
, cr
, Flags
, ddxyz
, npme
,
1032 dddlb_opt
, dlb_scale
,
1033 ddcsx
, ddcsy
, ddcsz
,
1036 &ddbox
, &npme_major
, &npme_minor
);
1040 /* PME, if used, is done on all nodes with 1D decomposition */
1042 cr
->duty
= (DUTY_PP
| DUTY_PME
);
1046 if (inputrec
->ePBC
== epbcSCREW
)
1049 "pbc=%s is only implemented with domain decomposition",
1050 epbc_names
[inputrec
->ePBC
]);
1056 /* After possible communicator splitting in make_dd_communicators.
1057 * we can set up the intra/inter node communication.
1059 gmx_setup_nodecomm(fplog
, cr
);
1062 /* Initialize per-physical-node MPI process/thread ID and counters. */
1063 gmx_init_intranode_counters(cr
);
1067 md_print_info(cr
, fplog
,
1068 "This is simulation %d out of %d running as a composite GROMACS\n"
1069 "multi-simulation job. Setup for this simulation:\n\n",
1070 cr
->ms
->sim
, cr
->ms
->nsim
);
1072 md_print_info(cr
, fplog
, "Using %d MPI %s\n",
1075 cr
->nnodes
== 1 ? "thread" : "threads"
1077 cr
->nnodes
== 1 ? "process" : "processes"
1083 /* Check and update hw_opt for the cut-off scheme */
1084 check_and_update_hw_opt_2(hw_opt
, inputrec
->cutoff_scheme
);
1086 /* Check and update hw_opt for the number of MPI ranks */
1087 check_and_update_hw_opt_3(hw_opt
);
1089 gmx_omp_nthreads_init(fplog
, cr
,
1090 hwinfo
->nthreads_hw_avail
,
1091 hw_opt
->nthreads_omp
,
1092 hw_opt
->nthreads_omp_pme
,
1093 (cr
->duty
& DUTY_PP
) == 0,
1094 inputrec
->cutoff_scheme
== ecutsVERLET
);
1097 if (EI_TPI(inputrec
->eI
) &&
1098 inputrec
->cutoff_scheme
== ecutsVERLET
)
1100 gmx_feenableexcept();
1106 /* Select GPU id's to use */
1107 gmx_select_gpu_ids(fplog
, cr
, &hwinfo
->gpu_info
, bForceUseGPU
,
1112 /* Ignore (potentially) manually selected GPUs */
1113 hw_opt
->gpu_opt
.n_dev_use
= 0;
1116 /* check consistency across ranks of things like SIMD
1117 * support and number of GPUs selected */
1118 gmx_check_hw_runconf_consistency(fplog
, hwinfo
, cr
, hw_opt
, bUseGPU
);
1120 /* Now that we know the setup is consistent, check for efficiency */
1121 check_resource_division_efficiency(hwinfo
, hw_opt
, Flags
& MD_NTOMPSET
,
1124 if (DOMAINDECOMP(cr
))
1126 /* When we share GPUs over ranks, we need to know this for the DLB */
1127 dd_setup_dlb_resource_sharing(cr
, hwinfo
, hw_opt
);
1130 /* getting number of PP/PME threads
1131 PME: env variable should be read only on one node to make sure it is
1132 identical everywhere;
1134 nthreads_pme
= gmx_omp_nthreads_get(emntPME
);
1136 wcycle
= wallcycle_init(fplog
, resetstep
, cr
);
1140 /* Master synchronizes its value of reset_counters with all nodes
1141 * including PME only nodes */
1142 reset_counters
= wcycle_get_reset_counters(wcycle
);
1143 gmx_bcast_sim(sizeof(reset_counters
), &reset_counters
, cr
);
1144 wcycle_set_reset_counters(wcycle
, reset_counters
);
1147 // Membrane embedding must be initialized before we call init_forcerec()
1148 if (opt2bSet("-membed", nfile
, fnm
))
1152 fprintf(stderr
, "Initializing membed");
1154 /* Note that membed cannot work in parallel because mtop is
1155 * changed here. Fix this if we ever want to make it run with
1156 * multiple ranks. */
1157 membed
= init_membed(fplog
, nfile
, fnm
, mtop
, inputrec
, state
, cr
, &cpt_period
);
1161 if (cr
->duty
& DUTY_PP
)
1163 bcast_state(cr
, state
);
1165 /* Initiate forcerecord */
1167 fr
->hwinfo
= hwinfo
;
1168 fr
->gpu_opt
= &hw_opt
->gpu_opt
;
1169 init_forcerec(fplog
, fr
, fcd
, inputrec
, mtop
, cr
, box
,
1170 opt2fn("-table", nfile
, fnm
),
1171 opt2fn("-tablep", nfile
, fnm
),
1172 getFilenm("-tableb", nfile
, fnm
),
1177 /* Initialize QM-MM */
1180 init_QMMMrec(cr
, mtop
, inputrec
, fr
);
1183 /* Initialize the mdatoms structure.
1184 * mdatoms is not filled with atom data,
1185 * as this can not be done now with domain decomposition.
1187 mdatoms
= init_mdatoms(fplog
, mtop
, inputrec
->efep
!= efepNO
);
1189 /* Initialize the virtual site communication */
1190 vsite
= init_vsite(mtop
, cr
, FALSE
);
1192 calc_shifts(box
, fr
->shift_vec
);
1194 /* With periodic molecules the charge groups should be whole at start up
1195 * and the virtual sites should not be far from their proper positions.
1197 if (!inputrec
->bContinuation
&& MASTER(cr
) &&
1198 !(inputrec
->ePBC
!= epbcNONE
&& inputrec
->bPeriodicMols
))
1200 /* Make molecules whole at start of run */
1201 if (fr
->ePBC
!= epbcNONE
)
1203 do_pbc_first_mtop(fplog
, inputrec
->ePBC
, box
, mtop
, state
->x
);
1207 /* Correct initial vsite positions are required
1208 * for the initial distribution in the domain decomposition
1209 * and for the initial shell prediction.
1211 construct_vsites_mtop(vsite
, mtop
, state
->x
);
1215 if (EEL_PME(fr
->eeltype
) || EVDW_PME(fr
->vdwtype
))
1217 ewaldcoeff_q
= fr
->ewaldcoeff_q
;
1218 ewaldcoeff_lj
= fr
->ewaldcoeff_lj
;
1219 pmedata
= &fr
->pmedata
;
1228 /* This is a PME only node */
1230 /* We don't need the state */
1233 ewaldcoeff_q
= calc_ewaldcoeff_q(inputrec
->rcoulomb
, inputrec
->ewald_rtol
);
1234 ewaldcoeff_lj
= calc_ewaldcoeff_lj(inputrec
->rvdw
, inputrec
->ewald_rtol_lj
);
1238 if (hw_opt
->thread_affinity
!= threadaffOFF
)
1240 /* Before setting affinity, check whether the affinity has changed
1241 * - which indicates that probably the OpenMP library has changed it
1242 * since we first checked).
1244 gmx_check_thread_affinity_set(fplog
, cr
,
1245 hw_opt
, hwinfo
->nthreads_hw_avail
, TRUE
);
1247 /* Set the CPU affinity */
1248 gmx_set_thread_affinity(fplog
, cr
, hw_opt
, hwinfo
);
1251 /* Initiate PME if necessary,
1252 * either on all nodes or on dedicated PME nodes only. */
1253 if (EEL_PME(inputrec
->coulombtype
) || EVDW_PME(inputrec
->vdwtype
))
1257 nChargePerturbed
= mdatoms
->nChargePerturbed
;
1258 if (EVDW_PME(inputrec
->vdwtype
))
1260 nTypePerturbed
= mdatoms
->nTypePerturbed
;
1263 if (cr
->npmenodes
> 0)
1265 /* The PME only nodes need to know nChargePerturbed(FEP on Q) and nTypePerturbed(FEP on LJ)*/
1266 gmx_bcast_sim(sizeof(nChargePerturbed
), &nChargePerturbed
, cr
);
1267 gmx_bcast_sim(sizeof(nTypePerturbed
), &nTypePerturbed
, cr
);
1270 if (cr
->duty
& DUTY_PME
)
1272 status
= gmx_pme_init(pmedata
, cr
, npme_major
, npme_minor
, inputrec
,
1273 mtop
? mtop
->natoms
: 0, nChargePerturbed
, nTypePerturbed
,
1274 (Flags
& MD_REPRODUCIBLE
), nthreads_pme
);
1277 gmx_fatal(FARGS
, "Error %d initializing PME", status
);
1283 if (EI_DYNAMICS(inputrec
->eI
))
1285 /* Turn on signal handling on all nodes */
1287 * (A user signal from the PME nodes (if any)
1288 * is communicated to the PP nodes.
1290 signal_handler_install();
1293 if (cr
->duty
& DUTY_PP
)
1295 /* Assumes uniform use of the number of OpenMP threads */
1296 walltime_accounting
= walltime_accounting_init(gmx_omp_nthreads_get(emntDefault
));
1298 if (inputrec
->bPull
)
1300 /* Initialize pull code */
1301 inputrec
->pull_work
=
1302 init_pull(fplog
, inputrec
->pull
, inputrec
, nfile
, fnm
,
1303 mtop
, cr
, oenv
, inputrec
->fepvals
->init_lambda
,
1304 EI_DYNAMICS(inputrec
->eI
) && MASTER(cr
), Flags
);
1309 /* Initialize enforced rotation code */
1310 init_rot(fplog
, inputrec
, nfile
, fnm
, cr
, state
->x
, state
->box
, mtop
, oenv
,
1314 constr
= init_constraints(fplog
, mtop
, inputrec
, ed
, state
, cr
);
1316 if (DOMAINDECOMP(cr
))
1318 GMX_RELEASE_ASSERT(fr
, "fr was NULL while cr->duty was DUTY_PP");
1319 /* This call is not included in init_domain_decomposition mainly
1320 * because fr->cginfo_mb is set later.
1322 dd_init_bondeds(fplog
, cr
->dd
, mtop
, vsite
, inputrec
,
1323 Flags
& MD_DDBONDCHECK
, fr
->cginfo_mb
);
1326 /* Now do whatever the user wants us to do (how flexible...) */
1327 my_integrator(inputrec
->eI
) (fplog
, cr
, nfile
, fnm
,
1331 nstepout
, inputrec
, mtop
,
1333 mdatoms
, nrnb
, wcycle
, ed
, fr
,
1334 repl_ex_nst
, repl_ex_nex
, repl_ex_seed
,
1336 cpt_period
, max_hours
,
1339 walltime_accounting
);
1343 finish_rot(inputrec
->rot
);
1346 if (inputrec
->bPull
)
1348 finish_pull(inputrec
->pull_work
);
1354 GMX_RELEASE_ASSERT(pmedata
, "pmedata was NULL while cr->duty was not DUTY_PP");
1356 walltime_accounting
= walltime_accounting_init(gmx_omp_nthreads_get(emntPME
));
1357 gmx_pmeonly(*pmedata
, cr
, nrnb
, wcycle
, walltime_accounting
, ewaldcoeff_q
, ewaldcoeff_lj
, inputrec
);
1360 wallcycle_stop(wcycle
, ewcRUN
);
1362 /* Finish up, write some stuff
1363 * if rerunMD, don't write last frame again
1365 finish_run(fplog
, cr
,
1366 inputrec
, nrnb
, wcycle
, walltime_accounting
,
1367 fr
? fr
->nbv
: NULL
,
1368 EI_DYNAMICS(inputrec
->eI
) && !MULTISIM(cr
));
1371 /* Free GPU memory and context */
1372 free_gpu_resources(fr
, cr
, &hwinfo
->gpu_info
, fr
? fr
->gpu_opt
: NULL
);
1374 if (membed
!= nullptr)
1376 free_membed(membed
);
1379 gmx_hardware_info_free(hwinfo
);
1381 /* Does what it says */
1382 print_date_and_time(fplog
, cr
->nodeid
, "Finished mdrun", gmx_gettime());
1383 walltime_accounting_destroy(walltime_accounting
);
1385 /* Close logfile already here if we were appending to it */
1386 if (MASTER(cr
) && (Flags
& MD_APPENDFILES
))
1388 gmx_log_close(fplog
);
1391 rc
= (int)gmx_get_stop_condition();
1396 /* we need to join all threads. The sub-threads join when they
1397 exit this function, but the master thread needs to be told to
1399 if (PAR(cr
) && MASTER(cr
))