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 * check out http://www.gromacs.org for more information.
7 * Copyright (c) 2012,2013, by the GROMACS development team, led by
8 * David van der Spoel, Berk Hess, Erik Lindahl, and including many
9 * others, as listed in the AUTHORS file in the top-level source
10 * directory and at http://www.gromacs.org.
12 * GROMACS is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU Lesser General Public License
14 * as published by the Free Software Foundation; either version 2.1
15 * of the License, or (at your option) any later version.
17 * GROMACS is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Lesser General Public License for more details.
22 * You should have received a copy of the GNU Lesser General Public
23 * License along with GROMACS; if not, see
24 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
27 * If you want to redistribute modifications to GROMACS, please
28 * consider that scientific software is very special. Version
29 * control is crucial - bugs must be traceable. We will be happy to
30 * consider code for inclusion in the official distribution, but
31 * derived work must not be called official GROMACS. Details are found
32 * in the README & COPYING files - if they are missing, get the
33 * official version at http://www.gromacs.org.
35 * To help us fund GROMACS development, we humbly ask that you cite
36 * the research papers on the package. Check out http://www.gromacs.org.
43 #include <catamount/dclock.h>
49 #ifdef HAVE_SYS_TIME_H
53 #include "visibility.h"
63 #include "chargegroup.h"
86 #include "pull_rotation.h"
87 #include "gmx_random.h"
88 #include "mpelogging.h"
91 #include "gmx_wallcycle.h"
93 #include "nbnxn_atomdata.h"
94 #include "nbnxn_search.h"
95 #include "nbnxn_kernels/nbnxn_kernel_ref.h"
96 #include "nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn.h"
97 #include "nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn.h"
98 #include "nbnxn_kernels/nbnxn_kernel_gpu_ref.h"
103 #ifdef GMX_THREAD_MPI
110 #include "nbnxn_cuda_data_mgmt.h"
111 #include "nbnxn_cuda/nbnxn_cuda.h"
114 typedef struct gmx_timeprint {
119 /* Portable version of ctime_r implemented in src/gmxlib/string2.c, but we do not want it declared in public installed headers */
122 gmx_ctime_r(const time_t *clock, char *buf, int n);
128 #ifdef HAVE_GETTIMEOFDAY
132 gettimeofday(&t, NULL);
134 seconds = (double) t.tv_sec + 1e-6*(double)t.tv_usec;
140 seconds = time(NULL);
147 #define difftime(end, start) ((double)(end)-(double)(start))
149 void print_time(FILE *out, gmx_runtime_t *runtime, gmx_large_int_t step,
150 t_inputrec *ir, t_commrec *cr)
153 char timebuf[STRLEN];
157 #ifndef GMX_THREAD_MPI
163 fprintf(out, "step %s", gmx_step_str(step, buf));
164 if ((step >= ir->nstlist))
166 runtime->last = gmx_gettime();
167 dt = difftime(runtime->last, runtime->real);
168 runtime->time_per_step = dt/(step - ir->init_step + 1);
170 dt = (ir->nsteps + ir->init_step - step)*runtime->time_per_step;
176 finish = (time_t) (runtime->last + dt);
177 gmx_ctime_r(&finish, timebuf, STRLEN);
178 sprintf(buf, "%s", timebuf);
179 buf[strlen(buf)-1] = '\0';
180 fprintf(out, ", will finish %s", buf);
184 fprintf(out, ", remaining runtime: %5d s ", (int)dt);
189 fprintf(out, " performance: %.1f ns/day ",
190 ir->delta_t/1000*24*60*60/runtime->time_per_step);
193 #ifndef GMX_THREAD_MPI
207 static double set_proctime(gmx_runtime_t *runtime)
213 prev = runtime->proc;
214 runtime->proc = dclock();
216 diff = runtime->proc - prev;
220 prev = runtime->proc;
221 runtime->proc = clock();
223 diff = (double)(runtime->proc - prev)/(double)CLOCKS_PER_SEC;
227 /* The counter has probably looped, ignore this data */
234 void runtime_start(gmx_runtime_t *runtime)
236 runtime->real = gmx_gettime();
238 set_proctime(runtime);
239 runtime->realtime = 0;
240 runtime->proctime = 0;
242 runtime->time_per_step = 0;
245 void runtime_end(gmx_runtime_t *runtime)
251 runtime->proctime += set_proctime(runtime);
252 runtime->realtime = now - runtime->real;
256 void runtime_upd_proc(gmx_runtime_t *runtime)
258 runtime->proctime += set_proctime(runtime);
261 void print_date_and_time(FILE *fplog, int nodeid, const char *title,
262 const gmx_runtime_t *runtime)
265 char timebuf[STRLEN];
266 char time_string[STRLEN];
273 tmptime = (time_t) runtime->real;
274 gmx_ctime_r(&tmptime, timebuf, STRLEN);
278 tmptime = (time_t) gmx_gettime();
279 gmx_ctime_r(&tmptime, timebuf, STRLEN);
281 for (i = 0; timebuf[i] >= ' '; i++)
283 time_string[i] = timebuf[i];
285 time_string[i] = '\0';
287 fprintf(fplog, "%s on node %d %s\n", title, nodeid, time_string);
291 static void sum_forces(int start, int end, rvec f[], rvec flr[])
297 pr_rvecs(debug, 0, "fsr", f+start, end-start);
298 pr_rvecs(debug, 0, "flr", flr+start, end-start);
300 for (i = start; (i < end); i++)
302 rvec_inc(f[i], flr[i]);
307 * calc_f_el calculates forces due to an electric field.
309 * force is kJ mol^-1 nm^-1 = e * kJ mol^-1 nm^-1 / e
311 * Et[] contains the parameters for the time dependent
312 * part of the field (not yet used).
313 * Ex[] contains the parameters for
314 * the spatial dependent part of the field. You can have cool periodic
315 * fields in principle, but only a constant field is supported
317 * The function should return the energy due to the electric field
318 * (if any) but for now returns 0.
321 * There can be problems with the virial.
322 * Since the field is not self-consistent this is unavoidable.
323 * For neutral molecules the virial is correct within this approximation.
324 * For neutral systems with many charged molecules the error is small.
325 * But for systems with a net charge or a few charged molecules
326 * the error can be significant when the field is high.
327 * Solution: implement a self-consitent electric field into PME.
329 static void calc_f_el(FILE *fp, int start, int homenr,
330 real charge[], rvec x[], rvec f[],
331 t_cosines Ex[], t_cosines Et[], double t)
337 for (m = 0; (m < DIM); m++)
344 Ext[m] = cos(Et[m].a[0]*(t-t0))*exp(-sqr(t-t0)/(2.0*sqr(Et[m].a[2])));
348 Ext[m] = cos(Et[m].a[0]*t);
357 /* Convert the field strength from V/nm to MD-units */
358 Ext[m] *= Ex[m].a[0]*FIELDFAC;
359 for (i = start; (i < start+homenr); i++)
361 f[i][m] += charge[i]*Ext[m];
371 fprintf(fp, "%10g %10g %10g %10g #FIELD\n", t,
372 Ext[XX]/FIELDFAC, Ext[YY]/FIELDFAC, Ext[ZZ]/FIELDFAC);
376 static void calc_virial(FILE *fplog, int start, int homenr, rvec x[], rvec f[],
377 tensor vir_part, t_graph *graph, matrix box,
378 t_nrnb *nrnb, const t_forcerec *fr, int ePBC)
383 /* The short-range virial from surrounding boxes */
385 calc_vir(fplog, SHIFTS, fr->shift_vec, fr->fshift, vir_part, ePBC == epbcSCREW, box);
386 inc_nrnb(nrnb, eNR_VIRIAL, SHIFTS);
388 /* Calculate partial virial, for local atoms only, based on short range.
389 * Total virial is computed in global_stat, called from do_md
391 f_calc_vir(fplog, start, start+homenr, x, f, vir_part, graph, box);
392 inc_nrnb(nrnb, eNR_VIRIAL, homenr);
394 /* Add position restraint contribution */
395 for (i = 0; i < DIM; i++)
397 vir_part[i][i] += fr->vir_diag_posres[i];
400 /* Add wall contribution */
401 for (i = 0; i < DIM; i++)
403 vir_part[i][ZZ] += fr->vir_wall_z[i];
408 pr_rvecs(debug, 0, "vir_part", vir_part, DIM);
412 static void posres_wrapper(FILE *fplog,
418 matrix box, rvec x[],
420 gmx_enerdata_t *enerd,
428 /* Position restraints always require full pbc */
429 set_pbc(&pbc, ir->ePBC, box);
431 v = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
432 top->idef.iparams_posres,
433 (const rvec*)x, fr->f_novirsum, fr->vir_diag_posres,
434 ir->ePBC == epbcNONE ? NULL : &pbc,
435 lambda[efptRESTRAINT], &dvdl,
436 fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
439 fprintf(fplog, sepdvdlformat,
440 interaction_function[F_POSRES].longname, v, dvdl);
442 enerd->term[F_POSRES] += v;
443 /* If just the force constant changes, the FEP term is linear,
444 * but if k changes, it is not.
446 enerd->dvdl_nonlin[efptRESTRAINT] += dvdl;
447 inc_nrnb(nrnb, eNR_POSRES, top->idef.il[F_POSRES].nr/2);
449 if ((ir->fepvals->n_lambda > 0) && (flags & GMX_FORCE_DHDL))
451 for (i = 0; i < enerd->n_lambda; i++)
453 real dvdl_dum, lambda_dum;
455 lambda_dum = (i == 0 ? lambda[efptRESTRAINT] : ir->fepvals->all_lambda[efptRESTRAINT][i-1]);
456 v = posres(top->idef.il[F_POSRES].nr, top->idef.il[F_POSRES].iatoms,
457 top->idef.iparams_posres,
458 (const rvec*)x, NULL, NULL,
459 ir->ePBC == epbcNONE ? NULL : &pbc, lambda_dum, &dvdl,
460 fr->rc_scaling, fr->ePBC, fr->posres_com, fr->posres_comB);
461 enerd->enerpart_lambda[i] += v;
466 static void pull_potential_wrapper(FILE *fplog,
470 matrix box, rvec x[],
474 gmx_enerdata_t *enerd,
481 /* Calculate the center of mass forces, this requires communication,
482 * which is why pull_potential is called close to other communication.
483 * The virial contribution is calculated directly,
484 * which is why we call pull_potential after calc_virial.
486 set_pbc(&pbc, ir->ePBC, box);
488 enerd->term[F_COM_PULL] +=
489 pull_potential(ir->ePull, ir->pull, mdatoms, &pbc,
490 cr, t, lambda[efptRESTRAINT], x, f, vir_force, &dvdl);
493 fprintf(fplog, sepdvdlformat, "Com pull", enerd->term[F_COM_PULL], dvdl);
495 enerd->dvdl_lin[efptRESTRAINT] += dvdl;
498 static void pme_receive_force_ener(FILE *fplog,
501 gmx_wallcycle_t wcycle,
502 gmx_enerdata_t *enerd,
506 float cycles_ppdpme, cycles_seppme;
508 cycles_ppdpme = wallcycle_stop(wcycle, ewcPPDURINGPME);
509 dd_cycles_add(cr->dd, cycles_ppdpme, ddCyclPPduringPME);
511 /* In case of node-splitting, the PP nodes receive the long-range
512 * forces, virial and energy from the PME nodes here.
514 wallcycle_start(wcycle, ewcPP_PMEWAITRECVF);
516 gmx_pme_receive_f(cr, fr->f_novirsum, fr->vir_el_recip, &e, &dvdl,
520 fprintf(fplog, sepdvdlformat, "PME mesh", e, dvdl);
522 enerd->term[F_COUL_RECIP] += e;
523 enerd->dvdl_lin[efptCOUL] += dvdl;
526 dd_cycles_add(cr->dd, cycles_seppme, ddCyclPME);
528 wallcycle_stop(wcycle, ewcPP_PMEWAITRECVF);
531 static void print_large_forces(FILE *fp, t_mdatoms *md, t_commrec *cr,
532 gmx_large_int_t step, real pforce, rvec *x, rvec *f)
536 char buf[STEPSTRSIZE];
539 for (i = md->start; i < md->start+md->homenr; i++)
542 /* We also catch NAN, if the compiler does not optimize this away. */
543 if (fn2 >= pf2 || fn2 != fn2)
545 fprintf(fp, "step %s atom %6d x %8.3f %8.3f %8.3f force %12.5e\n",
546 gmx_step_str(step, buf),
547 ddglatnr(cr->dd, i), x[i][XX], x[i][YY], x[i][ZZ], sqrt(fn2));
552 static void post_process_forces(FILE *fplog,
554 gmx_large_int_t step,
555 t_nrnb *nrnb, gmx_wallcycle_t wcycle,
557 matrix box, rvec x[],
562 t_forcerec *fr, gmx_vsite_t *vsite,
569 /* Spread the mesh force on virtual sites to the other particles...
570 * This is parallellized. MPI communication is performed
571 * if the constructing atoms aren't local.
573 wallcycle_start(wcycle, ewcVSITESPREAD);
574 spread_vsite_f(fplog, vsite, x, fr->f_novirsum, NULL,
575 (flags & GMX_FORCE_VIRIAL), fr->vir_el_recip,
577 &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
578 wallcycle_stop(wcycle, ewcVSITESPREAD);
580 if (flags & GMX_FORCE_VIRIAL)
582 /* Now add the forces, this is local */
585 sum_forces(0, fr->f_novirsum_n, f, fr->f_novirsum);
589 sum_forces(mdatoms->start, mdatoms->start+mdatoms->homenr,
592 if (EEL_FULL(fr->eeltype))
594 /* Add the mesh contribution to the virial */
595 m_add(vir_force, fr->vir_el_recip, vir_force);
599 pr_rvecs(debug, 0, "vir_force", vir_force, DIM);
604 if (fr->print_force >= 0)
606 print_large_forces(stderr, mdatoms, cr, step, fr->print_force, x, f);
610 static void do_nb_verlet(t_forcerec *fr,
611 interaction_const_t *ic,
612 gmx_enerdata_t *enerd,
613 int flags, int ilocality,
616 gmx_wallcycle_t wcycle)
618 int nnbl, kernel_type, enr_nbnxn_kernel_ljc, enr_nbnxn_kernel_lj;
620 nonbonded_verlet_group_t *nbvg;
623 if (!(flags & GMX_FORCE_NONBONDED))
625 /* skip non-bonded calculation */
629 nbvg = &fr->nbv->grp[ilocality];
631 /* CUDA kernel launch overhead is already timed separately */
632 if (fr->cutoff_scheme != ecutsVERLET)
634 gmx_incons("Invalid cut-off scheme passed!");
637 bCUDA = (nbvg->kernel_type == nbnxnk8x8x8_CUDA);
641 wallcycle_sub_start(wcycle, ewcsNONBONDED);
643 switch (nbvg->kernel_type)
645 case nbnxnk4x4_PlainC:
646 nbnxn_kernel_ref(&nbvg->nbl_lists,
652 enerd->grpp.ener[egCOULSR],
654 enerd->grpp.ener[egBHAMSR] :
655 enerd->grpp.ener[egLJSR]);
658 case nbnxnk4xN_SIMD_4xN:
659 nbnxn_kernel_simd_4xn(&nbvg->nbl_lists,
666 enerd->grpp.ener[egCOULSR],
668 enerd->grpp.ener[egBHAMSR] :
669 enerd->grpp.ener[egLJSR]);
671 case nbnxnk4xN_SIMD_2xNN:
672 nbnxn_kernel_simd_2xnn(&nbvg->nbl_lists,
679 enerd->grpp.ener[egCOULSR],
681 enerd->grpp.ener[egBHAMSR] :
682 enerd->grpp.ener[egLJSR]);
685 case nbnxnk8x8x8_CUDA:
686 nbnxn_cuda_launch_kernel(fr->nbv->cu_nbv, nbvg->nbat, flags, ilocality);
689 case nbnxnk8x8x8_PlainC:
690 nbnxn_kernel_gpu_ref(nbvg->nbl_lists.nbl[0],
695 nbvg->nbat->out[0].f,
697 enerd->grpp.ener[egCOULSR],
699 enerd->grpp.ener[egBHAMSR] :
700 enerd->grpp.ener[egLJSR]);
704 gmx_incons("Invalid nonbonded kernel type passed!");
709 wallcycle_sub_stop(wcycle, ewcsNONBONDED);
712 if (EEL_RF(ic->eeltype) || ic->eeltype == eelCUT)
714 enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_RF;
716 else if ((!bCUDA && nbvg->ewald_excl == ewaldexclAnalytical) ||
717 (bCUDA && nbnxn_cuda_is_kernel_ewald_analytical(fr->nbv->cu_nbv)))
719 enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_EWALD;
723 enr_nbnxn_kernel_ljc = eNR_NBNXN_LJ_TAB;
725 enr_nbnxn_kernel_lj = eNR_NBNXN_LJ;
726 if (flags & GMX_FORCE_ENERGY)
728 /* In eNR_??? the nbnxn F+E kernels are always the F kernel + 1 */
729 enr_nbnxn_kernel_ljc += 1;
730 enr_nbnxn_kernel_lj += 1;
733 inc_nrnb(nrnb, enr_nbnxn_kernel_ljc,
734 nbvg->nbl_lists.natpair_ljq);
735 inc_nrnb(nrnb, enr_nbnxn_kernel_lj,
736 nbvg->nbl_lists.natpair_lj);
737 inc_nrnb(nrnb, enr_nbnxn_kernel_ljc-eNR_NBNXN_LJ_RF+eNR_NBNXN_RF,
738 nbvg->nbl_lists.natpair_q);
741 void do_force_cutsVERLET(FILE *fplog, t_commrec *cr,
742 t_inputrec *inputrec,
743 gmx_large_int_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
746 gmx_groups_t *groups,
747 matrix box, rvec x[], history_t *hist,
751 gmx_enerdata_t *enerd, t_fcdata *fcd,
752 real *lambda, t_graph *graph,
753 t_forcerec *fr, interaction_const_t *ic,
754 gmx_vsite_t *vsite, rvec mu_tot,
755 double t, FILE *field, gmx_edsam_t ed,
763 gmx_bool bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
764 gmx_bool bDoLongRange, bDoForces, bSepLRF, bUseGPU, bUseOrEmulGPU;
765 gmx_bool bDiffKernels = FALSE;
767 rvec vzero, box_diag;
769 float cycles_pme, cycles_force;
770 nonbonded_verlet_t *nbv;
774 nb_kernel_type = fr->nbv->grp[0].kernel_type;
776 start = mdatoms->start;
777 homenr = mdatoms->homenr;
779 bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
781 clear_mat(vir_force);
784 if (DOMAINDECOMP(cr))
786 cg1 = cr->dd->ncg_tot;
797 bStateChanged = (flags & GMX_FORCE_STATECHANGED);
798 bNS = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
799 bFillGrid = (bNS && bStateChanged);
800 bCalcCGCM = (bFillGrid && !DOMAINDECOMP(cr));
801 bDoLongRange = (fr->bTwinRange && bNS && (flags & GMX_FORCE_DO_LR));
802 bDoForces = (flags & GMX_FORCE_FORCES);
803 bSepLRF = (bDoLongRange && bDoForces && (flags & GMX_FORCE_SEPLRF));
804 bUseGPU = fr->nbv->bUseGPU;
805 bUseOrEmulGPU = bUseGPU || (nbv->grp[0].kernel_type == nbnxnk8x8x8_PlainC);
809 update_forcerec(fplog, fr, box);
811 if (NEED_MUTOT(*inputrec))
813 /* Calculate total (local) dipole moment in a temporary common array.
814 * This makes it possible to sum them over nodes faster.
816 calc_mu(start, homenr,
817 x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
822 if (fr->ePBC != epbcNONE)
824 /* Compute shift vectors every step,
825 * because of pressure coupling or box deformation!
827 if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
829 calc_shifts(box, fr->shift_vec);
834 put_atoms_in_box_omp(fr->ePBC, box, homenr, x);
835 inc_nrnb(nrnb, eNR_SHIFTX, homenr);
837 else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
839 unshift_self(graph, box, x);
843 nbnxn_atomdata_copy_shiftvec(flags & GMX_FORCE_DYNAMICBOX,
844 fr->shift_vec, nbv->grp[0].nbat);
847 if (!(cr->duty & DUTY_PME))
849 /* Send particle coordinates to the pme nodes.
850 * Since this is only implemented for domain decomposition
851 * and domain decomposition does not use the graph,
852 * we do not need to worry about shifting.
855 wallcycle_start(wcycle, ewcPP_PMESENDX);
856 GMX_MPE_LOG(ev_send_coordinates_start);
858 bBS = (inputrec->nwall == 2);
862 svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
865 gmx_pme_send_x(cr, bBS ? boxs : box, x,
866 mdatoms->nChargePerturbed, lambda[efptCOUL],
867 (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)), step);
869 GMX_MPE_LOG(ev_send_coordinates_finish);
870 wallcycle_stop(wcycle, ewcPP_PMESENDX);
874 /* do gridding for pair search */
877 if (graph && bStateChanged)
879 /* Calculate intramolecular shift vectors to make molecules whole */
880 mk_mshift(fplog, graph, fr->ePBC, box, x);
884 box_diag[XX] = box[XX][XX];
885 box_diag[YY] = box[YY][YY];
886 box_diag[ZZ] = box[ZZ][ZZ];
888 wallcycle_start(wcycle, ewcNS);
891 wallcycle_sub_start(wcycle, ewcsNBS_GRID_LOCAL);
892 nbnxn_put_on_grid(nbv->nbs, fr->ePBC, box,
894 0, mdatoms->homenr, -1, fr->cginfo, x,
896 nbv->grp[eintLocal].kernel_type,
897 nbv->grp[eintLocal].nbat);
898 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_LOCAL);
902 wallcycle_sub_start(wcycle, ewcsNBS_GRID_NONLOCAL);
903 nbnxn_put_on_grid_nonlocal(nbv->nbs, domdec_zones(cr->dd),
905 nbv->grp[eintNonlocal].kernel_type,
906 nbv->grp[eintNonlocal].nbat);
907 wallcycle_sub_stop(wcycle, ewcsNBS_GRID_NONLOCAL);
910 if (nbv->ngrp == 1 ||
911 nbv->grp[eintNonlocal].nbat == nbv->grp[eintLocal].nbat)
913 nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatAll,
914 nbv->nbs, mdatoms, fr->cginfo);
918 nbnxn_atomdata_set(nbv->grp[eintLocal].nbat, eatLocal,
919 nbv->nbs, mdatoms, fr->cginfo);
920 nbnxn_atomdata_set(nbv->grp[eintNonlocal].nbat, eatAll,
921 nbv->nbs, mdatoms, fr->cginfo);
923 wallcycle_stop(wcycle, ewcNS);
926 /* initialize the GPU atom data and copy shift vector */
931 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
932 nbnxn_cuda_init_atomdata(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
933 wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
936 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
937 nbnxn_cuda_upload_shiftvec(nbv->cu_nbv, nbv->grp[eintLocal].nbat);
938 wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
941 /* do local pair search */
944 wallcycle_start_nocount(wcycle, ewcNS);
945 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_LOCAL);
946 nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintLocal].nbat,
949 nbv->min_ci_balanced,
950 &nbv->grp[eintLocal].nbl_lists,
952 nbv->grp[eintLocal].kernel_type,
954 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_LOCAL);
958 /* initialize local pair-list on the GPU */
959 nbnxn_cuda_init_pairlist(nbv->cu_nbv,
960 nbv->grp[eintLocal].nbl_lists.nbl[0],
963 wallcycle_stop(wcycle, ewcNS);
967 wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
968 wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
969 nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, FALSE, x,
970 nbv->grp[eintLocal].nbat);
971 wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
972 wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
977 wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
978 /* launch local nonbonded F on GPU */
979 do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFNo,
981 wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
984 /* Communicate coordinates and sum dipole if necessary +
985 do non-local pair search */
986 if (DOMAINDECOMP(cr))
988 bDiffKernels = (nbv->grp[eintNonlocal].kernel_type !=
989 nbv->grp[eintLocal].kernel_type);
993 /* With GPU+CPU non-bonded calculations we need to copy
994 * the local coordinates to the non-local nbat struct
995 * (in CPU format) as the non-local kernel call also
996 * calculates the local - non-local interactions.
998 wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
999 wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1000 nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatLocal, TRUE, x,
1001 nbv->grp[eintNonlocal].nbat);
1002 wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1003 wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1008 wallcycle_start_nocount(wcycle, ewcNS);
1009 wallcycle_sub_start(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1013 nbnxn_grid_add_simple(nbv->nbs, nbv->grp[eintNonlocal].nbat);
1016 nbnxn_make_pairlist(nbv->nbs, nbv->grp[eintNonlocal].nbat,
1019 nbv->min_ci_balanced,
1020 &nbv->grp[eintNonlocal].nbl_lists,
1022 nbv->grp[eintNonlocal].kernel_type,
1025 wallcycle_sub_stop(wcycle, ewcsNBS_SEARCH_NONLOCAL);
1027 if (nbv->grp[eintNonlocal].kernel_type == nbnxnk8x8x8_CUDA)
1029 /* initialize non-local pair-list on the GPU */
1030 nbnxn_cuda_init_pairlist(nbv->cu_nbv,
1031 nbv->grp[eintNonlocal].nbl_lists.nbl[0],
1034 wallcycle_stop(wcycle, ewcNS);
1038 wallcycle_start(wcycle, ewcMOVEX);
1039 dd_move_x(cr->dd, box, x);
1041 /* When we don't need the total dipole we sum it in global_stat */
1042 if (bStateChanged && NEED_MUTOT(*inputrec))
1044 gmx_sumd(2*DIM, mu, cr);
1046 wallcycle_stop(wcycle, ewcMOVEX);
1048 wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1049 wallcycle_sub_start(wcycle, ewcsNB_X_BUF_OPS);
1050 nbnxn_atomdata_copy_x_to_nbat_x(nbv->nbs, eatNonlocal, FALSE, x,
1051 nbv->grp[eintNonlocal].nbat);
1052 wallcycle_sub_stop(wcycle, ewcsNB_X_BUF_OPS);
1053 cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1056 if (bUseGPU && !bDiffKernels)
1058 wallcycle_start(wcycle, ewcLAUNCH_GPU_NB);
1059 /* launch non-local nonbonded F on GPU */
1060 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFNo,
1062 cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1068 /* launch D2H copy-back F */
1069 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1070 if (DOMAINDECOMP(cr) && !bDiffKernels)
1072 nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintNonlocal].nbat,
1073 flags, eatNonlocal);
1075 nbnxn_cuda_launch_cpyback(nbv->cu_nbv, nbv->grp[eintLocal].nbat,
1077 cycles_force += wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1080 if (bStateChanged && NEED_MUTOT(*inputrec))
1084 gmx_sumd(2*DIM, mu, cr);
1087 for (i = 0; i < 2; i++)
1089 for (j = 0; j < DIM; j++)
1091 fr->mu_tot[i][j] = mu[i*DIM + j];
1095 if (fr->efep == efepNO)
1097 copy_rvec(fr->mu_tot[0], mu_tot);
1101 for (j = 0; j < DIM; j++)
1104 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] +
1105 lambda[efptCOUL]*fr->mu_tot[1][j];
1109 /* Reset energies */
1110 reset_enerdata(&(inputrec->opts), fr, bNS, enerd, MASTER(cr));
1111 clear_rvecs(SHIFTS, fr->fshift);
1113 if (DOMAINDECOMP(cr))
1115 if (!(cr->duty & DUTY_PME))
1117 wallcycle_start(wcycle, ewcPPDURINGPME);
1118 dd_force_flop_start(cr->dd, nrnb);
1124 /* Enforced rotation has its own cycle counter that starts after the collective
1125 * coordinates have been communicated. It is added to ddCyclF to allow
1126 * for proper load-balancing */
1127 wallcycle_start(wcycle, ewcROT);
1128 do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1129 wallcycle_stop(wcycle, ewcROT);
1132 /* Start the force cycle counter.
1133 * This counter is stopped in do_forcelow_level.
1134 * No parallel communication should occur while this counter is running,
1135 * since that will interfere with the dynamic load balancing.
1137 wallcycle_start(wcycle, ewcFORCE);
1140 /* Reset forces for which the virial is calculated separately:
1141 * PME/Ewald forces if necessary */
1142 if (fr->bF_NoVirSum)
1144 if (flags & GMX_FORCE_VIRIAL)
1146 fr->f_novirsum = fr->f_novirsum_alloc;
1147 GMX_BARRIER(cr->mpi_comm_mygroup);
1150 clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1154 clear_rvecs(homenr, fr->f_novirsum+start);
1156 GMX_BARRIER(cr->mpi_comm_mygroup);
1160 /* We are not calculating the pressure so we do not need
1161 * a separate array for forces that do not contribute
1168 /* Clear the short- and long-range forces */
1169 clear_rvecs(fr->natoms_force_constr, f);
1170 if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1172 clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1175 clear_rvec(fr->vir_diag_posres);
1177 GMX_BARRIER(cr->mpi_comm_mygroup);
1180 if (inputrec->ePull == epullCONSTRAINT)
1182 clear_pull_forces(inputrec->pull);
1185 /* We calculate the non-bonded forces, when done on the CPU, here.
1186 * We do this before calling do_force_lowlevel, as in there bondeds
1187 * forces are calculated before PME, which does communication.
1188 * With this order, non-bonded and bonded force calculation imbalance
1189 * can be balanced out by the domain decomposition load balancing.
1194 /* Maybe we should move this into do_force_lowlevel */
1195 do_nb_verlet(fr, ic, enerd, flags, eintLocal, enbvClearFYes,
1199 if (!bUseOrEmulGPU || bDiffKernels)
1203 if (DOMAINDECOMP(cr))
1205 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal,
1206 bDiffKernels ? enbvClearFYes : enbvClearFNo,
1216 aloc = eintNonlocal;
1219 /* Add all the non-bonded force to the normal force array.
1220 * This can be split into a local a non-local part when overlapping
1221 * communication with calculation with domain decomposition.
1223 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1224 wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1225 wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1226 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatAll, nbv->grp[aloc].nbat, f);
1227 wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1228 cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1229 wallcycle_start_nocount(wcycle, ewcFORCE);
1231 /* if there are multiple fshift output buffers reduce them */
1232 if ((flags & GMX_FORCE_VIRIAL) &&
1233 nbv->grp[aloc].nbl_lists.nnbl > 1)
1235 nbnxn_atomdata_add_nbat_fshift_to_fshift(nbv->grp[aloc].nbat,
1240 /* update QMMMrec, if necessary */
1243 update_QMMMrec(cr, fr, x, mdatoms, box, top);
1246 if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1248 posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1249 f, enerd, lambda, fr);
1252 /* Compute the bonded and non-bonded energies and optionally forces */
1253 do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1254 cr, nrnb, wcycle, mdatoms, &(inputrec->opts),
1255 x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, mtop, top, fr->born,
1256 &(top->atomtypes), bBornRadii, box,
1257 inputrec->fepvals, lambda, graph, &(top->excls), fr->mu_tot,
1258 flags, &cycles_pme);
1262 if (do_per_step(step, inputrec->nstcalclr))
1264 /* Add the long range forces to the short range forces */
1265 for (i = 0; i < fr->natoms_force_constr; i++)
1267 rvec_add(fr->f_twin[i], f[i], f[i]);
1272 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1273 GMX_BARRIER(cr->mpi_comm_mygroup);
1277 do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1280 if (bUseOrEmulGPU && !bDiffKernels)
1282 /* wait for non-local forces (or calculate in emulation mode) */
1283 if (DOMAINDECOMP(cr))
1287 wallcycle_start(wcycle, ewcWAIT_GPU_NB_NL);
1288 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1289 nbv->grp[eintNonlocal].nbat,
1291 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1293 cycles_force += wallcycle_stop(wcycle, ewcWAIT_GPU_NB_NL);
1297 wallcycle_start_nocount(wcycle, ewcFORCE);
1298 do_nb_verlet(fr, ic, enerd, flags, eintNonlocal, enbvClearFYes,
1300 cycles_force += wallcycle_stop(wcycle, ewcFORCE);
1302 wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1303 wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1304 /* skip the reduction if there was no non-local work to do */
1305 if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1307 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatNonlocal,
1308 nbv->grp[eintNonlocal].nbat, f);
1310 wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1311 cycles_force += wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1317 /* Communicate the forces */
1320 wallcycle_start(wcycle, ewcMOVEF);
1321 if (DOMAINDECOMP(cr))
1323 dd_move_f(cr->dd, f, fr->fshift);
1324 /* Do we need to communicate the separate force array
1325 * for terms that do not contribute to the single sum virial?
1326 * Position restraints and electric fields do not introduce
1327 * inter-cg forces, only full electrostatics methods do.
1328 * When we do not calculate the virial, fr->f_novirsum = f,
1329 * so we have already communicated these forces.
1331 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1332 (flags & GMX_FORCE_VIRIAL))
1334 dd_move_f(cr->dd, fr->f_novirsum, NULL);
1338 /* We should not update the shift forces here,
1339 * since f_twin is already included in f.
1341 dd_move_f(cr->dd, fr->f_twin, NULL);
1344 wallcycle_stop(wcycle, ewcMOVEF);
1350 /* wait for local forces (or calculate in emulation mode) */
1353 wallcycle_start(wcycle, ewcWAIT_GPU_NB_L);
1354 nbnxn_cuda_wait_gpu(nbv->cu_nbv,
1355 nbv->grp[eintLocal].nbat,
1357 enerd->grpp.ener[egLJSR], enerd->grpp.ener[egCOULSR],
1359 wallcycle_stop(wcycle, ewcWAIT_GPU_NB_L);
1361 /* now clear the GPU outputs while we finish the step on the CPU */
1363 wallcycle_start_nocount(wcycle, ewcLAUNCH_GPU_NB);
1364 nbnxn_cuda_clear_outputs(nbv->cu_nbv, flags);
1365 wallcycle_stop(wcycle, ewcLAUNCH_GPU_NB);
1369 wallcycle_start_nocount(wcycle, ewcFORCE);
1370 do_nb_verlet(fr, ic, enerd, flags, eintLocal,
1371 DOMAINDECOMP(cr) ? enbvClearFNo : enbvClearFYes,
1373 wallcycle_stop(wcycle, ewcFORCE);
1375 wallcycle_start(wcycle, ewcNB_XF_BUF_OPS);
1376 wallcycle_sub_start(wcycle, ewcsNB_F_BUF_OPS);
1377 if (nbv->grp[eintLocal].nbl_lists.nbl[0]->nsci > 0)
1379 /* skip the reduction if there was no non-local work to do */
1380 nbnxn_atomdata_add_nbat_f_to_f(nbv->nbs, eatLocal,
1381 nbv->grp[eintLocal].nbat, f);
1383 wallcycle_sub_stop(wcycle, ewcsNB_F_BUF_OPS);
1384 wallcycle_stop(wcycle, ewcNB_XF_BUF_OPS);
1387 if (DOMAINDECOMP(cr))
1389 dd_force_flop_stop(cr->dd, nrnb);
1392 dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1398 if (IR_ELEC_FIELD(*inputrec))
1400 /* Compute forces due to electric field */
1401 calc_f_el(MASTER(cr) ? field : NULL,
1402 start, homenr, mdatoms->chargeA, x, fr->f_novirsum,
1403 inputrec->ex, inputrec->et, t);
1406 /* If we have NoVirSum forces, but we do not calculate the virial,
1407 * we sum fr->f_novirum=f later.
1409 if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1411 wallcycle_start(wcycle, ewcVSITESPREAD);
1412 spread_vsite_f(fplog, vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1413 &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1414 wallcycle_stop(wcycle, ewcVSITESPREAD);
1418 wallcycle_start(wcycle, ewcVSITESPREAD);
1419 spread_vsite_f(fplog, vsite, x, fr->f_twin, NULL, FALSE, NULL,
1421 &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1422 wallcycle_stop(wcycle, ewcVSITESPREAD);
1426 if (flags & GMX_FORCE_VIRIAL)
1428 /* Calculation of the virial must be done after vsites! */
1429 calc_virial(fplog, mdatoms->start, mdatoms->homenr, x, f,
1430 vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1434 if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1436 pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1437 f, vir_force, mdatoms, enerd, lambda, t);
1440 /* Add the forces from enforced rotation potentials (if any) */
1443 wallcycle_start(wcycle, ewcROTadd);
1444 enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1445 wallcycle_stop(wcycle, ewcROTadd);
1448 if (PAR(cr) && !(cr->duty & DUTY_PME))
1450 /* In case of node-splitting, the PP nodes receive the long-range
1451 * forces, virial and energy from the PME nodes here.
1453 pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1458 post_process_forces(fplog, cr, step, nrnb, wcycle,
1459 top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1463 /* Sum the potential energy terms from group contributions */
1464 sum_epot(&(inputrec->opts), &(enerd->grpp), enerd->term);
1467 void do_force_cutsGROUP(FILE *fplog, t_commrec *cr,
1468 t_inputrec *inputrec,
1469 gmx_large_int_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1470 gmx_localtop_t *top,
1472 gmx_groups_t *groups,
1473 matrix box, rvec x[], history_t *hist,
1477 gmx_enerdata_t *enerd, t_fcdata *fcd,
1478 real *lambda, t_graph *graph,
1479 t_forcerec *fr, gmx_vsite_t *vsite, rvec mu_tot,
1480 double t, FILE *field, gmx_edsam_t ed,
1481 gmx_bool bBornRadii,
1487 gmx_bool bSepDVDL, bStateChanged, bNS, bFillGrid, bCalcCGCM, bBS;
1488 gmx_bool bDoLongRangeNS, bDoForces, bDoPotential, bSepLRF;
1489 gmx_bool bDoAdressWF;
1491 rvec vzero, box_diag;
1492 real e, v, dvdlambda[efptNR];
1494 float cycles_pme, cycles_force;
1496 start = mdatoms->start;
1497 homenr = mdatoms->homenr;
1499 bSepDVDL = (fr->bSepDVDL && do_per_step(step, inputrec->nstlog));
1501 clear_mat(vir_force);
1505 pd_cg_range(cr, &cg0, &cg1);
1510 if (DOMAINDECOMP(cr))
1512 cg1 = cr->dd->ncg_tot;
1524 bStateChanged = (flags & GMX_FORCE_STATECHANGED);
1525 bNS = (flags & GMX_FORCE_NS) && (fr->bAllvsAll == FALSE);
1526 /* Should we update the long-range neighborlists at this step? */
1527 bDoLongRangeNS = fr->bTwinRange && bNS;
1528 /* Should we perform the long-range nonbonded evaluation inside the neighborsearching? */
1529 bFillGrid = (bNS && bStateChanged);
1530 bCalcCGCM = (bFillGrid && !DOMAINDECOMP(cr));
1531 bDoForces = (flags & GMX_FORCE_FORCES);
1532 bDoPotential = (flags & GMX_FORCE_ENERGY);
1533 bSepLRF = ((inputrec->nstcalclr > 1) && bDoForces &&
1534 (flags & GMX_FORCE_SEPLRF) && (flags & GMX_FORCE_DO_LR));
1536 /* should probably move this to the forcerec since it doesn't change */
1537 bDoAdressWF = ((fr->adress_type != eAdressOff));
1541 update_forcerec(fplog, fr, box);
1543 if (NEED_MUTOT(*inputrec))
1545 /* Calculate total (local) dipole moment in a temporary common array.
1546 * This makes it possible to sum them over nodes faster.
1548 calc_mu(start, homenr,
1549 x, mdatoms->chargeA, mdatoms->chargeB, mdatoms->nChargePerturbed,
1554 if (fr->ePBC != epbcNONE)
1556 /* Compute shift vectors every step,
1557 * because of pressure coupling or box deformation!
1559 if ((flags & GMX_FORCE_DYNAMICBOX) && bStateChanged)
1561 calc_shifts(box, fr->shift_vec);
1566 put_charge_groups_in_box(fplog, cg0, cg1, fr->ePBC, box,
1567 &(top->cgs), x, fr->cg_cm);
1568 inc_nrnb(nrnb, eNR_CGCM, homenr);
1569 inc_nrnb(nrnb, eNR_RESETX, cg1-cg0);
1571 else if (EI_ENERGY_MINIMIZATION(inputrec->eI) && graph)
1573 unshift_self(graph, box, x);
1578 calc_cgcm(fplog, cg0, cg1, &(top->cgs), x, fr->cg_cm);
1579 inc_nrnb(nrnb, eNR_CGCM, homenr);
1586 move_cgcm(fplog, cr, fr->cg_cm);
1590 pr_rvecs(debug, 0, "cgcm", fr->cg_cm, top->cgs.nr);
1595 if (!(cr->duty & DUTY_PME))
1597 /* Send particle coordinates to the pme nodes.
1598 * Since this is only implemented for domain decomposition
1599 * and domain decomposition does not use the graph,
1600 * we do not need to worry about shifting.
1603 wallcycle_start(wcycle, ewcPP_PMESENDX);
1604 GMX_MPE_LOG(ev_send_coordinates_start);
1606 bBS = (inputrec->nwall == 2);
1609 copy_mat(box, boxs);
1610 svmul(inputrec->wall_ewald_zfac, boxs[ZZ], boxs[ZZ]);
1613 gmx_pme_send_x(cr, bBS ? boxs : box, x,
1614 mdatoms->nChargePerturbed, lambda[efptCOUL],
1615 (flags & (GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY)), step);
1617 GMX_MPE_LOG(ev_send_coordinates_finish);
1618 wallcycle_stop(wcycle, ewcPP_PMESENDX);
1620 #endif /* GMX_MPI */
1622 /* Communicate coordinates and sum dipole if necessary */
1625 wallcycle_start(wcycle, ewcMOVEX);
1626 if (DOMAINDECOMP(cr))
1628 dd_move_x(cr->dd, box, x);
1632 move_x(fplog, cr, GMX_LEFT, GMX_RIGHT, x, nrnb);
1634 wallcycle_stop(wcycle, ewcMOVEX);
1637 /* update adress weight beforehand */
1638 if (bStateChanged && bDoAdressWF)
1640 /* need pbc for adress weight calculation with pbc_dx */
1641 set_pbc(&pbc, inputrec->ePBC, box);
1642 if (fr->adress_site == eAdressSITEcog)
1644 update_adress_weights_cog(top->idef.iparams, top->idef.il, x, fr, mdatoms,
1645 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1647 else if (fr->adress_site == eAdressSITEcom)
1649 update_adress_weights_com(fplog, cg0, cg1, &(top->cgs), x, fr, mdatoms,
1650 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1652 else if (fr->adress_site == eAdressSITEatomatom)
1654 update_adress_weights_atom_per_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1655 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1659 update_adress_weights_atom(cg0, cg1, &(top->cgs), x, fr, mdatoms,
1660 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1664 if (NEED_MUTOT(*inputrec))
1671 gmx_sumd(2*DIM, mu, cr);
1673 for (i = 0; i < 2; i++)
1675 for (j = 0; j < DIM; j++)
1677 fr->mu_tot[i][j] = mu[i*DIM + j];
1681 if (fr->efep == efepNO)
1683 copy_rvec(fr->mu_tot[0], mu_tot);
1687 for (j = 0; j < DIM; j++)
1690 (1.0 - lambda[efptCOUL])*fr->mu_tot[0][j] + lambda[efptCOUL]*fr->mu_tot[1][j];
1695 /* Reset energies */
1696 reset_enerdata(&(inputrec->opts), fr, bNS, enerd, MASTER(cr));
1697 clear_rvecs(SHIFTS, fr->fshift);
1701 wallcycle_start(wcycle, ewcNS);
1703 if (graph && bStateChanged)
1705 /* Calculate intramolecular shift vectors to make molecules whole */
1706 mk_mshift(fplog, graph, fr->ePBC, box, x);
1709 /* Do the actual neighbour searching */
1710 ns(fplog, fr, x, box,
1711 groups, &(inputrec->opts), top, mdatoms,
1712 cr, nrnb, lambda, dvdlambda, &enerd->grpp, bFillGrid,
1715 wallcycle_stop(wcycle, ewcNS);
1718 if (inputrec->implicit_solvent && bNS)
1720 make_gb_nblist(cr, inputrec->gb_algorithm, inputrec->rlist,
1721 x, box, fr, &top->idef, graph, fr->born);
1724 if (DOMAINDECOMP(cr))
1726 if (!(cr->duty & DUTY_PME))
1728 wallcycle_start(wcycle, ewcPPDURINGPME);
1729 dd_force_flop_start(cr->dd, nrnb);
1735 /* Enforced rotation has its own cycle counter that starts after the collective
1736 * coordinates have been communicated. It is added to ddCyclF to allow
1737 * for proper load-balancing */
1738 wallcycle_start(wcycle, ewcROT);
1739 do_rotation(cr, inputrec, box, x, t, step, wcycle, bNS);
1740 wallcycle_stop(wcycle, ewcROT);
1743 /* Start the force cycle counter.
1744 * This counter is stopped in do_forcelow_level.
1745 * No parallel communication should occur while this counter is running,
1746 * since that will interfere with the dynamic load balancing.
1748 wallcycle_start(wcycle, ewcFORCE);
1752 /* Reset forces for which the virial is calculated separately:
1753 * PME/Ewald forces if necessary */
1754 if (fr->bF_NoVirSum)
1756 if (flags & GMX_FORCE_VIRIAL)
1758 fr->f_novirsum = fr->f_novirsum_alloc;
1759 GMX_BARRIER(cr->mpi_comm_mygroup);
1762 clear_rvecs(fr->f_novirsum_n, fr->f_novirsum);
1766 clear_rvecs(homenr, fr->f_novirsum+start);
1768 GMX_BARRIER(cr->mpi_comm_mygroup);
1772 /* We are not calculating the pressure so we do not need
1773 * a separate array for forces that do not contribute
1780 /* Clear the short- and long-range forces */
1781 clear_rvecs(fr->natoms_force_constr, f);
1782 if (bSepLRF && do_per_step(step, inputrec->nstcalclr))
1784 clear_rvecs(fr->natoms_force_constr, fr->f_twin);
1787 clear_rvec(fr->vir_diag_posres);
1789 GMX_BARRIER(cr->mpi_comm_mygroup);
1791 if (inputrec->ePull == epullCONSTRAINT)
1793 clear_pull_forces(inputrec->pull);
1796 /* update QMMMrec, if necessary */
1799 update_QMMMrec(cr, fr, x, mdatoms, box, top);
1802 if ((flags & GMX_FORCE_BONDED) && top->idef.il[F_POSRES].nr > 0)
1804 posres_wrapper(fplog, flags, bSepDVDL, inputrec, nrnb, top, box, x,
1805 f, enerd, lambda, fr);
1808 /* Compute the bonded and non-bonded energies and optionally forces */
1809 do_force_lowlevel(fplog, step, fr, inputrec, &(top->idef),
1810 cr, nrnb, wcycle, mdatoms, &(inputrec->opts),
1811 x, hist, f, bSepLRF ? fr->f_twin : f, enerd, fcd, mtop, top, fr->born,
1812 &(top->atomtypes), bBornRadii, box,
1813 inputrec->fepvals, lambda,
1814 graph, &(top->excls), fr->mu_tot,
1820 if (do_per_step(step, inputrec->nstcalclr))
1822 /* Add the long range forces to the short range forces */
1823 for (i = 0; i < fr->natoms_force_constr; i++)
1825 rvec_add(fr->f_twin[i], f[i], f[i]);
1830 cycles_force = wallcycle_stop(wcycle, ewcFORCE);
1831 GMX_BARRIER(cr->mpi_comm_mygroup);
1835 do_flood(cr, inputrec, x, f, ed, box, step, bNS);
1838 if (DOMAINDECOMP(cr))
1840 dd_force_flop_stop(cr->dd, nrnb);
1843 dd_cycles_add(cr->dd, cycles_force-cycles_pme, ddCyclF);
1849 if (IR_ELEC_FIELD(*inputrec))
1851 /* Compute forces due to electric field */
1852 calc_f_el(MASTER(cr) ? field : NULL,
1853 start, homenr, mdatoms->chargeA, x, fr->f_novirsum,
1854 inputrec->ex, inputrec->et, t);
1857 if (bDoAdressWF && fr->adress_icor == eAdressICThermoForce)
1859 /* Compute thermodynamic force in hybrid AdResS region */
1860 adress_thermo_force(start, homenr, &(top->cgs), x, fr->f_novirsum, fr, mdatoms,
1861 inputrec->ePBC == epbcNONE ? NULL : &pbc);
1864 /* Communicate the forces */
1867 wallcycle_start(wcycle, ewcMOVEF);
1868 if (DOMAINDECOMP(cr))
1870 dd_move_f(cr->dd, f, fr->fshift);
1871 /* Do we need to communicate the separate force array
1872 * for terms that do not contribute to the single sum virial?
1873 * Position restraints and electric fields do not introduce
1874 * inter-cg forces, only full electrostatics methods do.
1875 * When we do not calculate the virial, fr->f_novirsum = f,
1876 * so we have already communicated these forces.
1878 if (EEL_FULL(fr->eeltype) && cr->dd->n_intercg_excl &&
1879 (flags & GMX_FORCE_VIRIAL))
1881 dd_move_f(cr->dd, fr->f_novirsum, NULL);
1885 /* We should not update the shift forces here,
1886 * since f_twin is already included in f.
1888 dd_move_f(cr->dd, fr->f_twin, NULL);
1893 pd_move_f(cr, f, nrnb);
1896 pd_move_f(cr, fr->f_twin, nrnb);
1899 wallcycle_stop(wcycle, ewcMOVEF);
1902 /* If we have NoVirSum forces, but we do not calculate the virial,
1903 * we sum fr->f_novirum=f later.
1905 if (vsite && !(fr->bF_NoVirSum && !(flags & GMX_FORCE_VIRIAL)))
1907 wallcycle_start(wcycle, ewcVSITESPREAD);
1908 spread_vsite_f(fplog, vsite, x, f, fr->fshift, FALSE, NULL, nrnb,
1909 &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1910 wallcycle_stop(wcycle, ewcVSITESPREAD);
1914 wallcycle_start(wcycle, ewcVSITESPREAD);
1915 spread_vsite_f(fplog, vsite, x, fr->f_twin, NULL, FALSE, NULL,
1917 &top->idef, fr->ePBC, fr->bMolPBC, graph, box, cr);
1918 wallcycle_stop(wcycle, ewcVSITESPREAD);
1922 if (flags & GMX_FORCE_VIRIAL)
1924 /* Calculation of the virial must be done after vsites! */
1925 calc_virial(fplog, mdatoms->start, mdatoms->homenr, x, f,
1926 vir_force, graph, box, nrnb, fr, inputrec->ePBC);
1930 if (inputrec->ePull == epullUMBRELLA || inputrec->ePull == epullCONST_F)
1932 pull_potential_wrapper(fplog, bSepDVDL, cr, inputrec, box, x,
1933 f, vir_force, mdatoms, enerd, lambda, t);
1936 /* Add the forces from enforced rotation potentials (if any) */
1939 wallcycle_start(wcycle, ewcROTadd);
1940 enerd->term[F_COM_PULL] += add_rot_forces(inputrec->rot, f, cr, step, t);
1941 wallcycle_stop(wcycle, ewcROTadd);
1944 if (PAR(cr) && !(cr->duty & DUTY_PME))
1946 /* In case of node-splitting, the PP nodes receive the long-range
1947 * forces, virial and energy from the PME nodes here.
1949 pme_receive_force_ener(fplog, bSepDVDL, cr, wcycle, enerd, fr);
1954 post_process_forces(fplog, cr, step, nrnb, wcycle,
1955 top, box, x, f, vir_force, mdatoms, graph, fr, vsite,
1959 /* Sum the potential energy terms from group contributions */
1960 sum_epot(&(inputrec->opts), &(enerd->grpp), enerd->term);
1963 void do_force(FILE *fplog, t_commrec *cr,
1964 t_inputrec *inputrec,
1965 gmx_large_int_t step, t_nrnb *nrnb, gmx_wallcycle_t wcycle,
1966 gmx_localtop_t *top,
1968 gmx_groups_t *groups,
1969 matrix box, rvec x[], history_t *hist,
1973 gmx_enerdata_t *enerd, t_fcdata *fcd,
1974 real *lambda, t_graph *graph,
1976 gmx_vsite_t *vsite, rvec mu_tot,
1977 double t, FILE *field, gmx_edsam_t ed,
1978 gmx_bool bBornRadii,
1981 /* modify force flag if not doing nonbonded */
1982 if (!fr->bNonbonded)
1984 flags &= ~GMX_FORCE_NONBONDED;
1987 switch (inputrec->cutoff_scheme)
1990 do_force_cutsVERLET(fplog, cr, inputrec,
2006 do_force_cutsGROUP(fplog, cr, inputrec,
2021 gmx_incons("Invalid cut-off scheme passed!");
2026 void do_constrain_first(FILE *fplog, gmx_constr_t constr,
2027 t_inputrec *ir, t_mdatoms *md,
2028 t_state *state, rvec *f,
2029 t_graph *graph, t_commrec *cr, t_nrnb *nrnb,
2030 t_forcerec *fr, gmx_localtop_t *top, tensor shake_vir)
2032 int i, m, start, end;
2033 gmx_large_int_t step;
2034 real dt = ir->delta_t;
2038 snew(savex, state->natoms);
2041 end = md->homenr + start;
2045 fprintf(debug, "vcm: start=%d, homenr=%d, end=%d\n",
2046 start, md->homenr, end);
2048 /* Do a first constrain to reset particles... */
2049 step = ir->init_step;
2052 char buf[STEPSTRSIZE];
2053 fprintf(fplog, "\nConstraining the starting coordinates (step %s)\n",
2054 gmx_step_str(step, buf));
2058 /* constrain the current position */
2059 constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2060 ir, NULL, cr, step, 0, md,
2061 state->x, state->x, NULL,
2062 fr->bMolPBC, state->box,
2063 state->lambda[efptBONDED], &dvdl_dum,
2064 NULL, NULL, nrnb, econqCoord,
2065 ir->epc == epcMTTK, state->veta, state->veta);
2068 /* constrain the inital velocity, and save it */
2069 /* also may be useful if we need the ekin from the halfstep for velocity verlet */
2070 /* might not yet treat veta correctly */
2071 constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2072 ir, NULL, cr, step, 0, md,
2073 state->x, state->v, state->v,
2074 fr->bMolPBC, state->box,
2075 state->lambda[efptBONDED], &dvdl_dum,
2076 NULL, NULL, nrnb, econqVeloc,
2077 ir->epc == epcMTTK, state->veta, state->veta);
2079 /* constrain the inital velocities at t-dt/2 */
2080 if (EI_STATE_VELOCITY(ir->eI) && ir->eI != eiVV)
2082 for (i = start; (i < end); i++)
2084 for (m = 0; (m < DIM); m++)
2086 /* Reverse the velocity */
2087 state->v[i][m] = -state->v[i][m];
2088 /* Store the position at t-dt in buf */
2089 savex[i][m] = state->x[i][m] + dt*state->v[i][m];
2092 /* Shake the positions at t=-dt with the positions at t=0
2093 * as reference coordinates.
2097 char buf[STEPSTRSIZE];
2098 fprintf(fplog, "\nConstraining the coordinates at t0-dt (step %s)\n",
2099 gmx_step_str(step, buf));
2102 constrain(NULL, TRUE, FALSE, constr, &(top->idef),
2103 ir, NULL, cr, step, -1, md,
2104 state->x, savex, NULL,
2105 fr->bMolPBC, state->box,
2106 state->lambda[efptBONDED], &dvdl_dum,
2107 state->v, NULL, nrnb, econqCoord,
2108 ir->epc == epcMTTK, state->veta, state->veta);
2110 for (i = start; i < end; i++)
2112 for (m = 0; m < DIM; m++)
2114 /* Re-reverse the velocities */
2115 state->v[i][m] = -state->v[i][m];
2122 void calc_enervirdiff(FILE *fplog, int eDispCorr, t_forcerec *fr)
2124 double eners[2], virs[2], enersum, virsum, y0, f, g, h;
2125 double r0, r1, r, rc3, rc9, ea, eb, ec, pa, pb, pc, pd;
2126 double invscale, invscale2, invscale3;
2127 int ri0, ri1, ri, i, offstart, offset;
2128 real scale, *vdwtab, tabfactor, tmp;
2130 fr->enershiftsix = 0;
2131 fr->enershifttwelve = 0;
2132 fr->enerdiffsix = 0;
2133 fr->enerdifftwelve = 0;
2135 fr->virdifftwelve = 0;
2137 if (eDispCorr != edispcNO)
2139 for (i = 0; i < 2; i++)
2144 if ((fr->vdwtype == evdwSWITCH) || (fr->vdwtype == evdwSHIFT))
2146 if (fr->rvdw_switch == 0)
2149 "With dispersion correction rvdw-switch can not be zero "
2150 "for vdw-type = %s", evdw_names[fr->vdwtype]);
2153 scale = fr->nblists[0].table_elec_vdw.scale;
2154 vdwtab = fr->nblists[0].table_vdw.data;
2156 /* Round the cut-offs to exact table values for precision */
2157 ri0 = floor(fr->rvdw_switch*scale);
2158 ri1 = ceil(fr->rvdw*scale);
2164 if (fr->vdwtype == evdwSHIFT)
2166 /* Determine the constant energy shift below rvdw_switch.
2167 * Table has a scale factor since we have scaled it down to compensate
2168 * for scaling-up c6/c12 with the derivative factors to save flops in analytical kernels.
2170 fr->enershiftsix = (real)(-1.0/(rc3*rc3)) - 6.0*vdwtab[8*ri0];
2171 fr->enershifttwelve = (real)( 1.0/(rc9*rc3)) - 12.0*vdwtab[8*ri0 + 4];
2173 /* Add the constant part from 0 to rvdw_switch.
2174 * This integration from 0 to rvdw_switch overcounts the number
2175 * of interactions by 1, as it also counts the self interaction.
2176 * We will correct for this later.
2178 eners[0] += 4.0*M_PI*fr->enershiftsix*rc3/3.0;
2179 eners[1] += 4.0*M_PI*fr->enershifttwelve*rc3/3.0;
2181 invscale = 1.0/(scale);
2182 invscale2 = invscale*invscale;
2183 invscale3 = invscale*invscale2;
2185 /* following summation derived from cubic spline definition,
2186 Numerical Recipies in C, second edition, p. 113-116. Exact
2187 for the cubic spline. We first calculate the negative of
2188 the energy from rvdw to rvdw_switch, assuming that g(r)=1,
2189 and then add the more standard, abrupt cutoff correction to
2190 that result, yielding the long-range correction for a
2191 switched function. We perform both the pressure and energy
2192 loops at the same time for simplicity, as the computational
2195 for (i = 0; i < 2; i++)
2197 enersum = 0.0; virsum = 0.0;
2201 /* Since the dispersion table has been scaled down a factor 6.0 and the repulsion
2202 * a factor 12.0 to compensate for the c6/c12 parameters inside nbfp[] being scaled
2203 * up (to save flops in kernels), we need to correct for this.
2212 for (ri = ri0; ri < ri1; ri++)
2216 eb = 2.0*invscale2*r;
2220 pb = 3.0*invscale2*r;
2221 pc = 3.0*invscale*r*r;
2224 /* this "8" is from the packing in the vdwtab array - perhaps should be #define'ed? */
2225 offset = 8*ri + offstart;
2226 y0 = vdwtab[offset];
2227 f = vdwtab[offset+1];
2228 g = vdwtab[offset+2];
2229 h = vdwtab[offset+3];
2231 enersum += y0*(ea/3 + eb/2 + ec) + f*(ea/4 + eb/3 + ec/2) + g*(ea/5 + eb/4 + ec/3) + h*(ea/6 + eb/5 + ec/4);
2232 virsum += f*(pa/4 + pb/3 + pc/2 + pd) + 2*g*(pa/5 + pb/4 + pc/3 + pd/2) + 3*h*(pa/6 + pb/5 + pc/4 + pd/3);
2235 enersum *= 4.0*M_PI*tabfactor;
2236 virsum *= 4.0*M_PI*tabfactor;
2237 eners[i] -= enersum;
2241 /* now add the correction for rvdw_switch to infinity */
2242 eners[0] += -4.0*M_PI/(3.0*rc3);
2243 eners[1] += 4.0*M_PI/(9.0*rc9);
2244 virs[0] += 8.0*M_PI/rc3;
2245 virs[1] += -16.0*M_PI/(3.0*rc9);
2247 else if ((fr->vdwtype == evdwCUT) || (fr->vdwtype == evdwUSER))
2249 if (fr->vdwtype == evdwUSER && fplog)
2252 "WARNING: using dispersion correction with user tables\n");
2254 rc3 = fr->rvdw*fr->rvdw*fr->rvdw;
2256 /* Contribution beyond the cut-off */
2257 eners[0] += -4.0*M_PI/(3.0*rc3);
2258 eners[1] += 4.0*M_PI/(9.0*rc9);
2259 if (fr->vdw_modifier == eintmodPOTSHIFT)
2261 /* Contribution within the cut-off */
2262 eners[0] += -4.0*M_PI/(3.0*rc3);
2263 eners[1] += 4.0*M_PI/(3.0*rc9);
2265 /* Contribution beyond the cut-off */
2266 virs[0] += 8.0*M_PI/rc3;
2267 virs[1] += -16.0*M_PI/(3.0*rc9);
2272 "Dispersion correction is not implemented for vdw-type = %s",
2273 evdw_names[fr->vdwtype]);
2275 fr->enerdiffsix = eners[0];
2276 fr->enerdifftwelve = eners[1];
2277 /* The 0.5 is due to the Gromacs definition of the virial */
2278 fr->virdiffsix = 0.5*virs[0];
2279 fr->virdifftwelve = 0.5*virs[1];
2283 void calc_dispcorr(FILE *fplog, t_inputrec *ir, t_forcerec *fr,
2284 gmx_large_int_t step, int natoms,
2285 matrix box, real lambda, tensor pres, tensor virial,
2286 real *prescorr, real *enercorr, real *dvdlcorr)
2288 gmx_bool bCorrAll, bCorrPres;
2289 real dvdlambda, invvol, dens, ninter, avcsix, avctwelve, enerdiff, svir = 0, spres = 0;
2299 if (ir->eDispCorr != edispcNO)
2301 bCorrAll = (ir->eDispCorr == edispcAllEner ||
2302 ir->eDispCorr == edispcAllEnerPres);
2303 bCorrPres = (ir->eDispCorr == edispcEnerPres ||
2304 ir->eDispCorr == edispcAllEnerPres);
2306 invvol = 1/det(box);
2309 /* Only correct for the interactions with the inserted molecule */
2310 dens = (natoms - fr->n_tpi)*invvol;
2315 dens = natoms*invvol;
2316 ninter = 0.5*natoms;
2319 if (ir->efep == efepNO)
2321 avcsix = fr->avcsix[0];
2322 avctwelve = fr->avctwelve[0];
2326 avcsix = (1 - lambda)*fr->avcsix[0] + lambda*fr->avcsix[1];
2327 avctwelve = (1 - lambda)*fr->avctwelve[0] + lambda*fr->avctwelve[1];
2330 enerdiff = ninter*(dens*fr->enerdiffsix - fr->enershiftsix);
2331 *enercorr += avcsix*enerdiff;
2333 if (ir->efep != efepNO)
2335 dvdlambda += (fr->avcsix[1] - fr->avcsix[0])*enerdiff;
2339 enerdiff = ninter*(dens*fr->enerdifftwelve - fr->enershifttwelve);
2340 *enercorr += avctwelve*enerdiff;
2341 if (fr->efep != efepNO)
2343 dvdlambda += (fr->avctwelve[1] - fr->avctwelve[0])*enerdiff;
2349 svir = ninter*dens*avcsix*fr->virdiffsix/3.0;
2350 if (ir->eDispCorr == edispcAllEnerPres)
2352 svir += ninter*dens*avctwelve*fr->virdifftwelve/3.0;
2354 /* The factor 2 is because of the Gromacs virial definition */
2355 spres = -2.0*invvol*svir*PRESFAC;
2357 for (m = 0; m < DIM; m++)
2359 virial[m][m] += svir;
2360 pres[m][m] += spres;
2365 /* Can't currently control when it prints, for now, just print when degugging */
2370 fprintf(debug, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
2376 "Long Range LJ corr.: Epot %10g, Pres: %10g, Vir: %10g\n",
2377 *enercorr, spres, svir);
2381 fprintf(debug, "Long Range LJ corr.: Epot %10g\n", *enercorr);
2385 if (fr->bSepDVDL && do_per_step(step, ir->nstlog))
2387 fprintf(fplog, sepdvdlformat, "Dispersion correction",
2388 *enercorr, dvdlambda);
2390 if (fr->efep != efepNO)
2392 *dvdlcorr += dvdlambda;
2397 void do_pbc_first(FILE *fplog, matrix box, t_forcerec *fr,
2398 t_graph *graph, rvec x[])
2402 fprintf(fplog, "Removing pbc first time\n");
2404 calc_shifts(box, fr->shift_vec);
2407 mk_mshift(fplog, graph, fr->ePBC, box, x);
2410 p_graph(debug, "do_pbc_first 1", graph);
2412 shift_self(graph, box, x);
2413 /* By doing an extra mk_mshift the molecules that are broken
2414 * because they were e.g. imported from another software
2415 * will be made whole again. Such are the healing powers
2418 mk_mshift(fplog, graph, fr->ePBC, box, x);
2421 p_graph(debug, "do_pbc_first 2", graph);
2426 fprintf(fplog, "Done rmpbc\n");
2430 static void low_do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2431 gmx_mtop_t *mtop, rvec x[],
2436 gmx_molblock_t *molb;
2438 if (bFirst && fplog)
2440 fprintf(fplog, "Removing pbc first time\n");
2445 for (mb = 0; mb < mtop->nmolblock; mb++)
2447 molb = &mtop->molblock[mb];
2448 if (molb->natoms_mol == 1 ||
2449 (!bFirst && mtop->moltype[molb->type].cgs.nr == 1))
2451 /* Just one atom or charge group in the molecule, no PBC required */
2452 as += molb->nmol*molb->natoms_mol;
2456 /* Pass NULL iso fplog to avoid graph prints for each molecule type */
2457 mk_graph_ilist(NULL, mtop->moltype[molb->type].ilist,
2458 0, molb->natoms_mol, FALSE, FALSE, graph);
2460 for (mol = 0; mol < molb->nmol; mol++)
2462 mk_mshift(fplog, graph, ePBC, box, x+as);
2464 shift_self(graph, box, x+as);
2465 /* The molecule is whole now.
2466 * We don't need the second mk_mshift call as in do_pbc_first,
2467 * since we no longer need this graph.
2470 as += molb->natoms_mol;
2478 void do_pbc_first_mtop(FILE *fplog, int ePBC, matrix box,
2479 gmx_mtop_t *mtop, rvec x[])
2481 low_do_pbc_mtop(fplog, ePBC, box, mtop, x, TRUE);
2484 void do_pbc_mtop(FILE *fplog, int ePBC, matrix box,
2485 gmx_mtop_t *mtop, rvec x[])
2487 low_do_pbc_mtop(fplog, ePBC, box, mtop, x, FALSE);
2490 void finish_run(FILE *fplog, t_commrec *cr, const char *confout,
2491 t_inputrec *inputrec,
2492 t_nrnb nrnb[], gmx_wallcycle_t wcycle,
2493 gmx_runtime_t *runtime,
2494 wallclock_gpu_t *gputimes,
2496 gmx_bool bWriteStat)
2499 t_nrnb *nrnb_tot = NULL;
2503 wallcycle_sum(cr, wcycle);
2509 MPI_Allreduce(nrnb->n, nrnb_tot->n, eNRNB, MPI_DOUBLE, MPI_SUM,
2510 cr->mpi_comm_mysim);
2518 #if defined(GMX_MPI) && !defined(GMX_THREAD_MPI)
2521 /* reduce nodetime over all MPI processes in the current simulation */
2523 MPI_Allreduce(&runtime->proctime, &sum, 1, MPI_DOUBLE, MPI_SUM,
2524 cr->mpi_comm_mysim);
2525 runtime->proctime = sum;
2531 print_flop(fplog, nrnb_tot, &nbfs, &mflop);
2538 if ((cr->duty & DUTY_PP) && DOMAINDECOMP(cr))
2540 print_dd_statistics(cr, inputrec, fplog);
2552 snew(nrnb_all, cr->nnodes);
2553 nrnb_all[0] = *nrnb;
2554 for (s = 1; s < cr->nnodes; s++)
2556 MPI_Recv(nrnb_all[s].n, eNRNB, MPI_DOUBLE, s, 0,
2557 cr->mpi_comm_mysim, &stat);
2559 pr_load(fplog, cr, nrnb_all);
2564 MPI_Send(nrnb->n, eNRNB, MPI_DOUBLE, MASTERRANK(cr), 0,
2565 cr->mpi_comm_mysim);
2572 wallcycle_print(fplog, cr->nnodes, cr->npmenodes, runtime->realtime,
2575 if (EI_DYNAMICS(inputrec->eI))
2577 delta_t = inputrec->delta_t;
2586 print_perf(fplog, runtime->proctime, runtime->realtime,
2587 cr->nnodes-cr->npmenodes,
2588 runtime->nsteps_done, delta_t, nbfs, mflop,
2593 print_perf(stderr, runtime->proctime, runtime->realtime,
2594 cr->nnodes-cr->npmenodes,
2595 runtime->nsteps_done, delta_t, nbfs, mflop,
2601 extern void initialize_lambdas(FILE *fplog, t_inputrec *ir, int *fep_state, real *lambda, double *lam0)
2603 /* this function works, but could probably use a logic rewrite to keep all the different
2604 types of efep straight. */
2607 t_lambda *fep = ir->fepvals;
2609 if ((ir->efep == efepNO) && (ir->bSimTemp == FALSE))
2611 for (i = 0; i < efptNR; i++)
2623 *fep_state = fep->init_fep_state; /* this might overwrite the checkpoint
2624 if checkpoint is set -- a kludge is in for now
2626 for (i = 0; i < efptNR; i++)
2628 /* overwrite lambda state with init_lambda for now for backwards compatibility */
2629 if (fep->init_lambda >= 0) /* if it's -1, it was never initializd */
2631 lambda[i] = fep->init_lambda;
2634 lam0[i] = lambda[i];
2639 lambda[i] = fep->all_lambda[i][*fep_state];
2642 lam0[i] = lambda[i];
2648 /* need to rescale control temperatures to match current state */
2649 for (i = 0; i < ir->opts.ngtc; i++)
2651 if (ir->opts.ref_t[i] > 0)
2653 ir->opts.ref_t[i] = ir->simtempvals->temperatures[*fep_state];
2659 /* Send to the log the information on the current lambdas */
2662 fprintf(fplog, "Initial vector of lambda components:[ ");
2663 for (i = 0; i < efptNR; i++)
2665 fprintf(fplog, "%10.4f ", lambda[i]);
2667 fprintf(fplog, "]\n");
2673 void init_md(FILE *fplog,
2674 t_commrec *cr, t_inputrec *ir, const output_env_t oenv,
2675 double *t, double *t0,
2676 real *lambda, int *fep_state, double *lam0,
2677 t_nrnb *nrnb, gmx_mtop_t *mtop,
2679 int nfile, const t_filenm fnm[],
2680 gmx_mdoutf_t **outf, t_mdebin **mdebin,
2681 tensor force_vir, tensor shake_vir, rvec mu_tot,
2682 gmx_bool *bSimAnn, t_vcm **vcm, t_state *state, unsigned long Flags)
2687 /* Initial values */
2688 *t = *t0 = ir->init_t;
2691 for (i = 0; i < ir->opts.ngtc; i++)
2693 /* set bSimAnn if any group is being annealed */
2694 if (ir->opts.annealing[i] != eannNO)
2701 update_annealing_target_temp(&(ir->opts), ir->init_t);
2704 /* Initialize lambda variables */
2705 initialize_lambdas(fplog, ir, fep_state, lambda, lam0);
2709 *upd = init_update(fplog, ir);
2715 *vcm = init_vcm(fplog, &mtop->groups, ir);
2718 if (EI_DYNAMICS(ir->eI) && !(Flags & MD_APPENDFILES))
2720 if (ir->etc == etcBERENDSEN)
2722 please_cite(fplog, "Berendsen84a");
2724 if (ir->etc == etcVRESCALE)
2726 please_cite(fplog, "Bussi2007a");
2734 *outf = init_mdoutf(nfile, fnm, Flags, cr, ir, oenv);
2736 *mdebin = init_mdebin((Flags & MD_APPENDFILES) ? NULL : (*outf)->fp_ene,
2737 mtop, ir, (*outf)->fp_dhdl);
2742 please_cite(fplog, "Fritsch12");
2743 please_cite(fplog, "Junghans10");
2745 /* Initiate variables */
2746 clear_mat(force_vir);
2747 clear_mat(shake_vir);