1 /* -*- mode: c; tab-width: 4; indent-tabs-mode: nil; c-basic-offset: 4; c-file-style: "stroustrup"; -*-
4 * This source code is part of
8 * GROningen MAchine for Chemical Simulations
11 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
12 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
13 * Copyright (c) 2001-2004, The GROMACS development team,
14 * check out http://www.gromacs.org for more information.
16 * This program is free software; you can redistribute it and/or
17 * modify it under the terms of the GNU General Public License
18 * as published by the Free Software Foundation; either version 2
19 * of the License, or (at your option) any later version.
21 * If you want to redistribute modifications, please consider that
22 * scientific software is very special. Version control is crucial -
23 * bugs must be traceable. We will be happy to consider code for
24 * inclusion in the official distribution, but derived work must not
25 * be called official GROMACS. Details are found in the README & COPYING
26 * files - if they are missing, get the official version at www.gromacs.org.
28 * To help us fund GROMACS development, we humbly ask that you cite
29 * the papers on the package - you can find them in the top README file.
31 * For more info, check our website at http://www.gromacs.org
34 * GROwing Monsters And Cloning Shrimps
50 #include "gmx_fatal.h"
51 #include "gmx_fatal_collective.h"
55 #include "nonbonded.h"
64 #include "md_support.h"
65 #include "md_logging.h"
70 #include "mtop_util.h"
71 #include "nbnxn_search.h"
72 #include "nbnxn_atomdata.h"
73 #include "nbnxn_consts.h"
75 #include "gmx_omp_nthreads.h"
76 #include "gmx_detect_hardware.h"
79 /* MSVC definition for __cpuid() */
83 #include "types/nbnxn_cuda_types_ext.h"
84 #include "gpu_utils.h"
85 #include "nbnxn_cuda_data_mgmt.h"
86 #include "pmalloc_cuda.h"
88 t_forcerec *mk_forcerec(void)
98 static void pr_nbfp(FILE *fp, real *nbfp, gmx_bool bBHAM, int atnr)
102 for (i = 0; (i < atnr); i++)
104 for (j = 0; (j < atnr); j++)
106 fprintf(fp, "%2d - %2d", i, j);
109 fprintf(fp, " a=%10g, b=%10g, c=%10g\n", BHAMA(nbfp, atnr, i, j),
110 BHAMB(nbfp, atnr, i, j), BHAMC(nbfp, atnr, i, j)/6.0);
114 fprintf(fp, " c6=%10g, c12=%10g\n", C6(nbfp, atnr, i, j)/6.0,
115 C12(nbfp, atnr, i, j)/12.0);
122 static real *mk_nbfp(const gmx_ffparams_t *idef, gmx_bool bBHAM)
130 snew(nbfp, 3*atnr*atnr);
131 for (i = k = 0; (i < atnr); i++)
133 for (j = 0; (j < atnr); j++, k++)
135 BHAMA(nbfp, atnr, i, j) = idef->iparams[k].bham.a;
136 BHAMB(nbfp, atnr, i, j) = idef->iparams[k].bham.b;
137 /* nbfp now includes the 6.0 derivative prefactor */
138 BHAMC(nbfp, atnr, i, j) = idef->iparams[k].bham.c*6.0;
144 snew(nbfp, 2*atnr*atnr);
145 for (i = k = 0; (i < atnr); i++)
147 for (j = 0; (j < atnr); j++, k++)
149 /* nbfp now includes the 6.0/12.0 derivative prefactors */
150 C6(nbfp, atnr, i, j) = idef->iparams[k].lj.c6*6.0;
151 C12(nbfp, atnr, i, j) = idef->iparams[k].lj.c12*12.0;
159 /* This routine sets fr->solvent_opt to the most common solvent in the
160 * system, e.g. esolSPC or esolTIP4P. It will also mark each charge group in
161 * the fr->solvent_type array with the correct type (or esolNO).
163 * Charge groups that fulfill the conditions but are not identical to the
164 * most common one will be marked as esolNO in the solvent_type array.
166 * TIP3p is identical to SPC for these purposes, so we call it
167 * SPC in the arrays (Apologies to Bill Jorgensen ;-)
169 * NOTE: QM particle should not
170 * become an optimized solvent. Not even if there is only one charge
180 } solvent_parameters_t;
183 check_solvent_cg(const gmx_moltype_t *molt,
186 const unsigned char *qm_grpnr,
187 const t_grps *qm_grps,
189 int *n_solvent_parameters,
190 solvent_parameters_t **solvent_parameters_p,
194 const t_blocka * excl;
205 solvent_parameters_t *solvent_parameters;
207 /* We use a list with parameters for each solvent type.
208 * Every time we discover a new molecule that fulfills the basic
209 * conditions for a solvent we compare with the previous entries
210 * in these lists. If the parameters are the same we just increment
211 * the counter for that type, and otherwise we create a new type
212 * based on the current molecule.
214 * Once we've finished going through all molecules we check which
215 * solvent is most common, and mark all those molecules while we
216 * clear the flag on all others.
219 solvent_parameters = *solvent_parameters_p;
221 /* Mark the cg first as non optimized */
224 /* Check if this cg has no exclusions with atoms in other charge groups
225 * and all atoms inside the charge group excluded.
226 * We only have 3 or 4 atom solvent loops.
228 if (GET_CGINFO_EXCL_INTER(cginfo) ||
229 !GET_CGINFO_EXCL_INTRA(cginfo))
234 /* Get the indices of the first atom in this charge group */
235 j0 = molt->cgs.index[cg0];
236 j1 = molt->cgs.index[cg0+1];
238 /* Number of atoms in our molecule */
244 "Moltype '%s': there are %d atoms in this charge group\n",
248 /* Check if it could be an SPC (3 atoms) or TIP4p (4) water,
251 if (nj < 3 || nj > 4)
256 /* Check if we are doing QM on this group */
258 if (qm_grpnr != NULL)
260 for (j = j0; j < j1 && !qm; j++)
262 qm = (qm_grpnr[j] < qm_grps->nr - 1);
265 /* Cannot use solvent optimization with QM */
271 atom = molt->atoms.atom;
273 /* Still looks like a solvent, time to check parameters */
275 /* If it is perturbed (free energy) we can't use the solvent loops,
276 * so then we just skip to the next molecule.
280 for (j = j0; j < j1 && !perturbed; j++)
282 perturbed = PERTURBED(atom[j]);
290 /* Now it's only a question if the VdW and charge parameters
291 * are OK. Before doing the check we compare and see if they are
292 * identical to a possible previous solvent type.
293 * First we assign the current types and charges.
295 for (j = 0; j < nj; j++)
297 tmp_vdwtype[j] = atom[j0+j].type;
298 tmp_charge[j] = atom[j0+j].q;
301 /* Does it match any previous solvent type? */
302 for (k = 0; k < *n_solvent_parameters; k++)
307 /* We can only match SPC with 3 atoms and TIP4p with 4 atoms */
308 if ( (solvent_parameters[k].model == esolSPC && nj != 3) ||
309 (solvent_parameters[k].model == esolTIP4P && nj != 4) )
314 /* Check that types & charges match for all atoms in molecule */
315 for (j = 0; j < nj && match == TRUE; j++)
317 if (tmp_vdwtype[j] != solvent_parameters[k].vdwtype[j])
321 if (tmp_charge[j] != solvent_parameters[k].charge[j])
328 /* Congratulations! We have a matched solvent.
329 * Flag it with this type for later processing.
332 solvent_parameters[k].count += nmol;
334 /* We are done with this charge group */
339 /* If we get here, we have a tentative new solvent type.
340 * Before we add it we must check that it fulfills the requirements
341 * of the solvent optimized loops. First determine which atoms have
344 for (j = 0; j < nj; j++)
347 tjA = tmp_vdwtype[j];
349 /* Go through all other tpes and see if any have non-zero
350 * VdW parameters when combined with this one.
352 for (k = 0; k < fr->ntype && (has_vdw[j] == FALSE); k++)
354 /* We already checked that the atoms weren't perturbed,
355 * so we only need to check state A now.
359 has_vdw[j] = (has_vdw[j] ||
360 (BHAMA(fr->nbfp, fr->ntype, tjA, k) != 0.0) ||
361 (BHAMB(fr->nbfp, fr->ntype, tjA, k) != 0.0) ||
362 (BHAMC(fr->nbfp, fr->ntype, tjA, k) != 0.0));
367 has_vdw[j] = (has_vdw[j] ||
368 (C6(fr->nbfp, fr->ntype, tjA, k) != 0.0) ||
369 (C12(fr->nbfp, fr->ntype, tjA, k) != 0.0));
374 /* Now we know all we need to make the final check and assignment. */
378 * For this we require thatn all atoms have charge,
379 * the charges on atom 2 & 3 should be the same, and only
380 * atom 1 might have VdW.
382 if (has_vdw[1] == FALSE &&
383 has_vdw[2] == FALSE &&
384 tmp_charge[0] != 0 &&
385 tmp_charge[1] != 0 &&
386 tmp_charge[2] == tmp_charge[1])
388 srenew(solvent_parameters, *n_solvent_parameters+1);
389 solvent_parameters[*n_solvent_parameters].model = esolSPC;
390 solvent_parameters[*n_solvent_parameters].count = nmol;
391 for (k = 0; k < 3; k++)
393 solvent_parameters[*n_solvent_parameters].vdwtype[k] = tmp_vdwtype[k];
394 solvent_parameters[*n_solvent_parameters].charge[k] = tmp_charge[k];
397 *cg_sp = *n_solvent_parameters;
398 (*n_solvent_parameters)++;
403 /* Or could it be a TIP4P?
404 * For this we require thatn atoms 2,3,4 have charge, but not atom 1.
405 * Only atom 1 mght have VdW.
407 if (has_vdw[1] == FALSE &&
408 has_vdw[2] == FALSE &&
409 has_vdw[3] == FALSE &&
410 tmp_charge[0] == 0 &&
411 tmp_charge[1] != 0 &&
412 tmp_charge[2] == tmp_charge[1] &&
415 srenew(solvent_parameters, *n_solvent_parameters+1);
416 solvent_parameters[*n_solvent_parameters].model = esolTIP4P;
417 solvent_parameters[*n_solvent_parameters].count = nmol;
418 for (k = 0; k < 4; k++)
420 solvent_parameters[*n_solvent_parameters].vdwtype[k] = tmp_vdwtype[k];
421 solvent_parameters[*n_solvent_parameters].charge[k] = tmp_charge[k];
424 *cg_sp = *n_solvent_parameters;
425 (*n_solvent_parameters)++;
429 *solvent_parameters_p = solvent_parameters;
433 check_solvent(FILE * fp,
434 const gmx_mtop_t * mtop,
436 cginfo_mb_t *cginfo_mb)
439 const t_block * mols;
440 const gmx_moltype_t *molt;
441 int mb, mol, cg_mol, at_offset, cg_offset, am, cgm, i, nmol_ch, nmol;
442 int n_solvent_parameters;
443 solvent_parameters_t *solvent_parameters;
449 fprintf(debug, "Going to determine what solvent types we have.\n");
454 n_solvent_parameters = 0;
455 solvent_parameters = NULL;
456 /* Allocate temporary array for solvent type */
457 snew(cg_sp, mtop->nmolblock);
461 for (mb = 0; mb < mtop->nmolblock; mb++)
463 molt = &mtop->moltype[mtop->molblock[mb].type];
465 /* Here we have to loop over all individual molecules
466 * because we need to check for QMMM particles.
468 snew(cg_sp[mb], cginfo_mb[mb].cg_mod);
469 nmol_ch = cginfo_mb[mb].cg_mod/cgs->nr;
470 nmol = mtop->molblock[mb].nmol/nmol_ch;
471 for (mol = 0; mol < nmol_ch; mol++)
474 am = mol*cgs->index[cgs->nr];
475 for (cg_mol = 0; cg_mol < cgs->nr; cg_mol++)
477 check_solvent_cg(molt, cg_mol, nmol,
478 mtop->groups.grpnr[egcQMMM] ?
479 mtop->groups.grpnr[egcQMMM]+at_offset+am : 0,
480 &mtop->groups.grps[egcQMMM],
482 &n_solvent_parameters, &solvent_parameters,
483 cginfo_mb[mb].cginfo[cgm+cg_mol],
484 &cg_sp[mb][cgm+cg_mol]);
487 cg_offset += cgs->nr;
488 at_offset += cgs->index[cgs->nr];
491 /* Puh! We finished going through all charge groups.
492 * Now find the most common solvent model.
495 /* Most common solvent this far */
497 for (i = 0; i < n_solvent_parameters; i++)
500 solvent_parameters[i].count > solvent_parameters[bestsp].count)
508 bestsol = solvent_parameters[bestsp].model;
515 #ifdef DISABLE_WATER_NLIST
520 for (mb = 0; mb < mtop->nmolblock; mb++)
522 cgs = &mtop->moltype[mtop->molblock[mb].type].cgs;
523 nmol = (mtop->molblock[mb].nmol*cgs->nr)/cginfo_mb[mb].cg_mod;
524 for (i = 0; i < cginfo_mb[mb].cg_mod; i++)
526 if (cg_sp[mb][i] == bestsp)
528 SET_CGINFO_SOLOPT(cginfo_mb[mb].cginfo[i], bestsol);
533 SET_CGINFO_SOLOPT(cginfo_mb[mb].cginfo[i], esolNO);
540 if (bestsol != esolNO && fp != NULL)
542 fprintf(fp, "\nEnabling %s-like water optimization for %d molecules.\n\n",
544 solvent_parameters[bestsp].count);
547 sfree(solvent_parameters);
548 fr->solvent_opt = bestsol;
552 acNONE = 0, acCONSTRAINT, acSETTLE
555 static cginfo_mb_t *init_cginfo_mb(FILE *fplog, const gmx_mtop_t *mtop,
556 t_forcerec *fr, gmx_bool bNoSolvOpt,
557 gmx_bool *bExcl_IntraCGAll_InterCGNone)
560 const t_blocka *excl;
561 const gmx_moltype_t *molt;
562 const gmx_molblock_t *molb;
563 cginfo_mb_t *cginfo_mb;
566 int cg_offset, a_offset, cgm, am;
567 int mb, m, ncg_tot, cg, a0, a1, gid, ai, j, aj, excl_nalloc;
571 gmx_bool bId, *bExcl, bExclIntraAll, bExclInter, bHaveVDW, bHaveQ;
573 ncg_tot = ncg_mtop(mtop);
574 snew(cginfo_mb, mtop->nmolblock);
576 snew(type_VDW, fr->ntype);
577 for (ai = 0; ai < fr->ntype; ai++)
579 type_VDW[ai] = FALSE;
580 for (j = 0; j < fr->ntype; j++)
582 type_VDW[ai] = type_VDW[ai] ||
584 C6(fr->nbfp, fr->ntype, ai, j) != 0 ||
585 C12(fr->nbfp, fr->ntype, ai, j) != 0;
589 *bExcl_IntraCGAll_InterCGNone = TRUE;
592 snew(bExcl, excl_nalloc);
595 for (mb = 0; mb < mtop->nmolblock; mb++)
597 molb = &mtop->molblock[mb];
598 molt = &mtop->moltype[molb->type];
602 /* Check if the cginfo is identical for all molecules in this block.
603 * If so, we only need an array of the size of one molecule.
604 * Otherwise we make an array of #mol times #cgs per molecule.
608 for (m = 0; m < molb->nmol; m++)
610 am = m*cgs->index[cgs->nr];
611 for (cg = 0; cg < cgs->nr; cg++)
614 a1 = cgs->index[cg+1];
615 if (ggrpnr(&mtop->groups, egcENER, a_offset+am+a0) !=
616 ggrpnr(&mtop->groups, egcENER, a_offset +a0))
620 if (mtop->groups.grpnr[egcQMMM] != NULL)
622 for (ai = a0; ai < a1; ai++)
624 if (mtop->groups.grpnr[egcQMMM][a_offset+am+ai] !=
625 mtop->groups.grpnr[egcQMMM][a_offset +ai])
634 cginfo_mb[mb].cg_start = cg_offset;
635 cginfo_mb[mb].cg_end = cg_offset + molb->nmol*cgs->nr;
636 cginfo_mb[mb].cg_mod = (bId ? 1 : molb->nmol)*cgs->nr;
637 snew(cginfo_mb[mb].cginfo, cginfo_mb[mb].cg_mod);
638 cginfo = cginfo_mb[mb].cginfo;
640 /* Set constraints flags for constrained atoms */
641 snew(a_con, molt->atoms.nr);
642 for (ftype = 0; ftype < F_NRE; ftype++)
644 if (interaction_function[ftype].flags & IF_CONSTRAINT)
649 for (ia = 0; ia < molt->ilist[ftype].nr; ia += 1+nral)
653 for (a = 0; a < nral; a++)
655 a_con[molt->ilist[ftype].iatoms[ia+1+a]] =
656 (ftype == F_SETTLE ? acSETTLE : acCONSTRAINT);
662 for (m = 0; m < (bId ? 1 : molb->nmol); m++)
665 am = m*cgs->index[cgs->nr];
666 for (cg = 0; cg < cgs->nr; cg++)
669 a1 = cgs->index[cg+1];
671 /* Store the energy group in cginfo */
672 gid = ggrpnr(&mtop->groups, egcENER, a_offset+am+a0);
673 SET_CGINFO_GID(cginfo[cgm+cg], gid);
675 /* Check the intra/inter charge group exclusions */
676 if (a1-a0 > excl_nalloc)
678 excl_nalloc = a1 - a0;
679 srenew(bExcl, excl_nalloc);
681 /* bExclIntraAll: all intra cg interactions excluded
682 * bExclInter: any inter cg interactions excluded
684 bExclIntraAll = TRUE;
688 for (ai = a0; ai < a1; ai++)
690 /* Check VDW and electrostatic interactions */
691 bHaveVDW = bHaveVDW || (type_VDW[molt->atoms.atom[ai].type] ||
692 type_VDW[molt->atoms.atom[ai].typeB]);
693 bHaveQ = bHaveQ || (molt->atoms.atom[ai].q != 0 ||
694 molt->atoms.atom[ai].qB != 0);
696 /* Clear the exclusion list for atom ai */
697 for (aj = a0; aj < a1; aj++)
699 bExcl[aj-a0] = FALSE;
701 /* Loop over all the exclusions of atom ai */
702 for (j = excl->index[ai]; j < excl->index[ai+1]; j++)
705 if (aj < a0 || aj >= a1)
714 /* Check if ai excludes a0 to a1 */
715 for (aj = a0; aj < a1; aj++)
719 bExclIntraAll = FALSE;
726 SET_CGINFO_CONSTR(cginfo[cgm+cg]);
729 SET_CGINFO_SETTLE(cginfo[cgm+cg]);
737 SET_CGINFO_EXCL_INTRA(cginfo[cgm+cg]);
741 SET_CGINFO_EXCL_INTER(cginfo[cgm+cg]);
743 if (a1 - a0 > MAX_CHARGEGROUP_SIZE)
745 /* The size in cginfo is currently only read with DD */
746 gmx_fatal(FARGS, "A charge group has size %d which is larger than the limit of %d atoms", a1-a0, MAX_CHARGEGROUP_SIZE);
750 SET_CGINFO_HAS_VDW(cginfo[cgm+cg]);
754 SET_CGINFO_HAS_Q(cginfo[cgm+cg]);
756 /* Store the charge group size */
757 SET_CGINFO_NATOMS(cginfo[cgm+cg], a1-a0);
759 if (!bExclIntraAll || bExclInter)
761 *bExcl_IntraCGAll_InterCGNone = FALSE;
768 cg_offset += molb->nmol*cgs->nr;
769 a_offset += molb->nmol*cgs->index[cgs->nr];
773 /* the solvent optimizer is called after the QM is initialized,
774 * because we don't want to have the QM subsystemto become an
778 check_solvent(fplog, mtop, fr, cginfo_mb);
780 if (getenv("GMX_NO_SOLV_OPT"))
784 fprintf(fplog, "Found environment variable GMX_NO_SOLV_OPT.\n"
785 "Disabling all solvent optimization\n");
787 fr->solvent_opt = esolNO;
791 fr->solvent_opt = esolNO;
793 if (!fr->solvent_opt)
795 for (mb = 0; mb < mtop->nmolblock; mb++)
797 for (cg = 0; cg < cginfo_mb[mb].cg_mod; cg++)
799 SET_CGINFO_SOLOPT(cginfo_mb[mb].cginfo[cg], esolNO);
807 static int *cginfo_expand(int nmb, cginfo_mb_t *cgi_mb)
812 ncg = cgi_mb[nmb-1].cg_end;
815 for (cg = 0; cg < ncg; cg++)
817 while (cg >= cgi_mb[mb].cg_end)
822 cgi_mb[mb].cginfo[(cg - cgi_mb[mb].cg_start) % cgi_mb[mb].cg_mod];
828 static void set_chargesum(FILE *log, t_forcerec *fr, const gmx_mtop_t *mtop)
830 double qsum, q2sum, q;
832 const t_atoms *atoms;
836 for (mb = 0; mb < mtop->nmolblock; mb++)
838 nmol = mtop->molblock[mb].nmol;
839 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
840 for (i = 0; i < atoms->nr; i++)
842 q = atoms->atom[i].q;
848 fr->q2sum[0] = q2sum;
849 if (fr->efep != efepNO)
853 for (mb = 0; mb < mtop->nmolblock; mb++)
855 nmol = mtop->molblock[mb].nmol;
856 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
857 for (i = 0; i < atoms->nr; i++)
859 q = atoms->atom[i].qB;
864 fr->q2sum[1] = q2sum;
869 fr->qsum[1] = fr->qsum[0];
870 fr->q2sum[1] = fr->q2sum[0];
874 if (fr->efep == efepNO)
876 fprintf(log, "System total charge: %.3f\n", fr->qsum[0]);
880 fprintf(log, "System total charge, top. A: %.3f top. B: %.3f\n",
881 fr->qsum[0], fr->qsum[1]);
886 void update_forcerec(t_forcerec *fr, matrix box)
888 if (fr->eeltype == eelGRF)
890 calc_rffac(NULL, fr->eeltype, fr->epsilon_r, fr->epsilon_rf,
891 fr->rcoulomb, fr->temp, fr->zsquare, box,
892 &fr->kappa, &fr->k_rf, &fr->c_rf);
896 void set_avcsixtwelve(FILE *fplog, t_forcerec *fr, const gmx_mtop_t *mtop)
898 const t_atoms *atoms, *atoms_tpi;
899 const t_blocka *excl;
900 int mb, nmol, nmolc, i, j, tpi, tpj, j1, j2, k, n, nexcl, q;
901 #if (defined SIZEOF_LONG_LONG_INT) && (SIZEOF_LONG_LONG_INT >= 8)
902 long long int npair, npair_ij, tmpi, tmpj;
904 double npair, npair_ij, tmpi, tmpj;
906 double csix, ctwelve;
915 for (q = 0; q < (fr->efep == efepNO ? 1 : 2); q++)
923 /* Count the types so we avoid natoms^2 operations */
924 snew(typecount, ntp);
925 for (mb = 0; mb < mtop->nmolblock; mb++)
927 nmol = mtop->molblock[mb].nmol;
928 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
929 for (i = 0; i < atoms->nr; i++)
933 tpi = atoms->atom[i].type;
937 tpi = atoms->atom[i].typeB;
939 typecount[tpi] += nmol;
942 for (tpi = 0; tpi < ntp; tpi++)
944 for (tpj = tpi; tpj < ntp; tpj++)
946 tmpi = typecount[tpi];
947 tmpj = typecount[tpj];
950 npair_ij = tmpi*tmpj;
954 npair_ij = tmpi*(tmpi - 1)/2;
958 /* nbfp now includes the 6.0 derivative prefactor */
959 csix += npair_ij*BHAMC(nbfp, ntp, tpi, tpj)/6.0;
963 /* nbfp now includes the 6.0/12.0 derivative prefactors */
964 csix += npair_ij* C6(nbfp, ntp, tpi, tpj)/6.0;
965 ctwelve += npair_ij* C12(nbfp, ntp, tpi, tpj)/12.0;
971 /* Subtract the excluded pairs.
972 * The main reason for substracting exclusions is that in some cases
973 * some combinations might never occur and the parameters could have
974 * any value. These unused values should not influence the dispersion
977 for (mb = 0; mb < mtop->nmolblock; mb++)
979 nmol = mtop->molblock[mb].nmol;
980 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
981 excl = &mtop->moltype[mtop->molblock[mb].type].excls;
982 for (i = 0; (i < atoms->nr); i++)
986 tpi = atoms->atom[i].type;
990 tpi = atoms->atom[i].typeB;
993 j2 = excl->index[i+1];
994 for (j = j1; j < j2; j++)
1001 tpj = atoms->atom[k].type;
1005 tpj = atoms->atom[k].typeB;
1009 /* nbfp now includes the 6.0 derivative prefactor */
1010 csix -= nmol*BHAMC(nbfp, ntp, tpi, tpj)/6.0;
1014 /* nbfp now includes the 6.0/12.0 derivative prefactors */
1015 csix -= nmol*C6 (nbfp, ntp, tpi, tpj)/6.0;
1016 ctwelve -= nmol*C12(nbfp, ntp, tpi, tpj)/12.0;
1026 /* Only correct for the interaction of the test particle
1027 * with the rest of the system.
1030 &mtop->moltype[mtop->molblock[mtop->nmolblock-1].type].atoms;
1033 for (mb = 0; mb < mtop->nmolblock; mb++)
1035 nmol = mtop->molblock[mb].nmol;
1036 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
1037 for (j = 0; j < atoms->nr; j++)
1040 /* Remove the interaction of the test charge group
1043 if (mb == mtop->nmolblock-1)
1047 if (mb == 0 && nmol == 1)
1049 gmx_fatal(FARGS, "Old format tpr with TPI, please generate a new tpr file");
1054 tpj = atoms->atom[j].type;
1058 tpj = atoms->atom[j].typeB;
1060 for (i = 0; i < fr->n_tpi; i++)
1064 tpi = atoms_tpi->atom[i].type;
1068 tpi = atoms_tpi->atom[i].typeB;
1072 /* nbfp now includes the 6.0 derivative prefactor */
1073 csix += nmolc*BHAMC(nbfp, ntp, tpi, tpj)/6.0;
1077 /* nbfp now includes the 6.0/12.0 derivative prefactors */
1078 csix += nmolc*C6 (nbfp, ntp, tpi, tpj)/6.0;
1079 ctwelve += nmolc*C12(nbfp, ntp, tpi, tpj)/12.0;
1086 if (npair - nexcl <= 0 && fplog)
1088 fprintf(fplog, "\nWARNING: There are no atom pairs for dispersion correction\n\n");
1094 csix /= npair - nexcl;
1095 ctwelve /= npair - nexcl;
1099 fprintf(debug, "Counted %d exclusions\n", nexcl);
1100 fprintf(debug, "Average C6 parameter is: %10g\n", (double)csix);
1101 fprintf(debug, "Average C12 parameter is: %10g\n", (double)ctwelve);
1103 fr->avcsix[q] = csix;
1104 fr->avctwelve[q] = ctwelve;
1108 if (fr->eDispCorr == edispcAllEner ||
1109 fr->eDispCorr == edispcAllEnerPres)
1111 fprintf(fplog, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
1112 fr->avcsix[0], fr->avctwelve[0]);
1116 fprintf(fplog, "Long Range LJ corr.: <C6> %10.4e\n", fr->avcsix[0]);
1122 static void set_bham_b_max(FILE *fplog, t_forcerec *fr,
1123 const gmx_mtop_t *mtop)
1125 const t_atoms *at1, *at2;
1126 int mt1, mt2, i, j, tpi, tpj, ntypes;
1132 fprintf(fplog, "Determining largest Buckingham b parameter for table\n");
1139 for (mt1 = 0; mt1 < mtop->nmoltype; mt1++)
1141 at1 = &mtop->moltype[mt1].atoms;
1142 for (i = 0; (i < at1->nr); i++)
1144 tpi = at1->atom[i].type;
1147 gmx_fatal(FARGS, "Atomtype[%d] = %d, maximum = %d", i, tpi, ntypes);
1150 for (mt2 = mt1; mt2 < mtop->nmoltype; mt2++)
1152 at2 = &mtop->moltype[mt2].atoms;
1153 for (j = 0; (j < at2->nr); j++)
1155 tpj = at2->atom[j].type;
1158 gmx_fatal(FARGS, "Atomtype[%d] = %d, maximum = %d", j, tpj, ntypes);
1160 b = BHAMB(nbfp, ntypes, tpi, tpj);
1161 if (b > fr->bham_b_max)
1165 if ((b < bmin) || (bmin == -1))
1175 fprintf(fplog, "Buckingham b parameters, min: %g, max: %g\n",
1176 bmin, fr->bham_b_max);
1180 static void make_nbf_tables(FILE *fp, const output_env_t oenv,
1181 t_forcerec *fr, real rtab,
1182 const t_commrec *cr,
1183 const char *tabfn, char *eg1, char *eg2,
1193 fprintf(debug, "No table file name passed, can not read table, can not do non-bonded interactions\n");
1198 sprintf(buf, "%s", tabfn);
1201 /* Append the two energy group names */
1202 sprintf(buf + strlen(tabfn) - strlen(ftp2ext(efXVG)) - 1, "_%s_%s.%s",
1203 eg1, eg2, ftp2ext(efXVG));
1205 nbl->table_elec_vdw = make_tables(fp, oenv, fr, MASTER(cr), buf, rtab, 0);
1206 /* Copy the contents of the table to separate coulomb and LJ tables too,
1207 * to improve cache performance.
1209 /* For performance reasons we want
1210 * the table data to be aligned to 16-byte. The pointers could be freed
1211 * but currently aren't.
1213 nbl->table_elec.interaction = GMX_TABLE_INTERACTION_ELEC;
1214 nbl->table_elec.format = nbl->table_elec_vdw.format;
1215 nbl->table_elec.r = nbl->table_elec_vdw.r;
1216 nbl->table_elec.n = nbl->table_elec_vdw.n;
1217 nbl->table_elec.scale = nbl->table_elec_vdw.scale;
1218 nbl->table_elec.scale_exp = nbl->table_elec_vdw.scale_exp;
1219 nbl->table_elec.formatsize = nbl->table_elec_vdw.formatsize;
1220 nbl->table_elec.ninteractions = 1;
1221 nbl->table_elec.stride = nbl->table_elec.formatsize * nbl->table_elec.ninteractions;
1222 snew_aligned(nbl->table_elec.data, nbl->table_elec.stride*(nbl->table_elec.n+1), 32);
1224 nbl->table_vdw.interaction = GMX_TABLE_INTERACTION_VDWREP_VDWDISP;
1225 nbl->table_vdw.format = nbl->table_elec_vdw.format;
1226 nbl->table_vdw.r = nbl->table_elec_vdw.r;
1227 nbl->table_vdw.n = nbl->table_elec_vdw.n;
1228 nbl->table_vdw.scale = nbl->table_elec_vdw.scale;
1229 nbl->table_vdw.scale_exp = nbl->table_elec_vdw.scale_exp;
1230 nbl->table_vdw.formatsize = nbl->table_elec_vdw.formatsize;
1231 nbl->table_vdw.ninteractions = 2;
1232 nbl->table_vdw.stride = nbl->table_vdw.formatsize * nbl->table_vdw.ninteractions;
1233 snew_aligned(nbl->table_vdw.data, nbl->table_vdw.stride*(nbl->table_vdw.n+1), 32);
1235 for (i = 0; i <= nbl->table_elec_vdw.n; i++)
1237 for (j = 0; j < 4; j++)
1239 nbl->table_elec.data[4*i+j] = nbl->table_elec_vdw.data[12*i+j];
1241 for (j = 0; j < 8; j++)
1243 nbl->table_vdw.data[8*i+j] = nbl->table_elec_vdw.data[12*i+4+j];
1248 static void count_tables(int ftype1, int ftype2, const gmx_mtop_t *mtop,
1249 int *ncount, int **count)
1251 const gmx_moltype_t *molt;
1253 int mt, ftype, stride, i, j, tabnr;
1255 for (mt = 0; mt < mtop->nmoltype; mt++)
1257 molt = &mtop->moltype[mt];
1258 for (ftype = 0; ftype < F_NRE; ftype++)
1260 if (ftype == ftype1 || ftype == ftype2)
1262 il = &molt->ilist[ftype];
1263 stride = 1 + NRAL(ftype);
1264 for (i = 0; i < il->nr; i += stride)
1266 tabnr = mtop->ffparams.iparams[il->iatoms[i]].tab.table;
1269 gmx_fatal(FARGS, "A bonded table number is smaller than 0: %d\n", tabnr);
1271 if (tabnr >= *ncount)
1273 srenew(*count, tabnr+1);
1274 for (j = *ncount; j < tabnr+1; j++)
1287 static bondedtable_t *make_bonded_tables(FILE *fplog,
1288 int ftype1, int ftype2,
1289 const gmx_mtop_t *mtop,
1290 const char *basefn, const char *tabext)
1292 int i, ncount, *count;
1300 count_tables(ftype1, ftype2, mtop, &ncount, &count);
1305 for (i = 0; i < ncount; i++)
1309 sprintf(tabfn, "%s", basefn);
1310 sprintf(tabfn + strlen(basefn) - strlen(ftp2ext(efXVG)) - 1, "_%s%d.%s",
1311 tabext, i, ftp2ext(efXVG));
1312 tab[i] = make_bonded_table(fplog, tabfn, NRAL(ftype1)-2);
1321 void forcerec_set_ranges(t_forcerec *fr,
1322 int ncg_home, int ncg_force,
1324 int natoms_force_constr, int natoms_f_novirsum)
1329 /* fr->ncg_force is unused in the standard code,
1330 * but it can be useful for modified code dealing with charge groups.
1332 fr->ncg_force = ncg_force;
1333 fr->natoms_force = natoms_force;
1334 fr->natoms_force_constr = natoms_force_constr;
1336 if (fr->natoms_force_constr > fr->nalloc_force)
1338 fr->nalloc_force = over_alloc_dd(fr->natoms_force_constr);
1342 srenew(fr->f_twin, fr->nalloc_force);
1346 if (fr->bF_NoVirSum)
1348 fr->f_novirsum_n = natoms_f_novirsum;
1349 if (fr->f_novirsum_n > fr->f_novirsum_nalloc)
1351 fr->f_novirsum_nalloc = over_alloc_dd(fr->f_novirsum_n);
1352 srenew(fr->f_novirsum_alloc, fr->f_novirsum_nalloc);
1357 fr->f_novirsum_n = 0;
1361 static real cutoff_inf(real cutoff)
1365 cutoff = GMX_CUTOFF_INF;
1371 static void make_adress_tf_tables(FILE *fp, const output_env_t oenv,
1372 t_forcerec *fr, const t_inputrec *ir,
1373 const char *tabfn, const gmx_mtop_t *mtop,
1381 gmx_fatal(FARGS, "No thermoforce table file given. Use -tabletf to specify a file\n");
1385 snew(fr->atf_tabs, ir->adress->n_tf_grps);
1387 sprintf(buf, "%s", tabfn);
1388 for (i = 0; i < ir->adress->n_tf_grps; i++)
1390 j = ir->adress->tf_table_index[i]; /* get energy group index */
1391 sprintf(buf + strlen(tabfn) - strlen(ftp2ext(efXVG)) - 1, "tf_%s.%s",
1392 *(mtop->groups.grpname[mtop->groups.grps[egcENER].nm_ind[j]]), ftp2ext(efXVG));
1395 fprintf(fp, "loading tf table for energygrp index %d from %s\n", ir->adress->tf_table_index[i], buf);
1397 fr->atf_tabs[i] = make_atf_table(fp, oenv, fr, buf, box);
1402 gmx_bool can_use_allvsall(const t_inputrec *ir, gmx_bool bPrintNote, t_commrec *cr, FILE *fp)
1409 ir->rcoulomb == 0 &&
1411 ir->ePBC == epbcNONE &&
1412 ir->vdwtype == evdwCUT &&
1413 ir->coulombtype == eelCUT &&
1414 ir->efep == efepNO &&
1415 (ir->implicit_solvent == eisNO ||
1416 (ir->implicit_solvent == eisGBSA && (ir->gb_algorithm == egbSTILL ||
1417 ir->gb_algorithm == egbHCT ||
1418 ir->gb_algorithm == egbOBC))) &&
1419 getenv("GMX_NO_ALLVSALL") == NULL
1422 if (bAllvsAll && ir->opts.ngener > 1)
1424 const char *note = "NOTE: Can not use all-vs-all force loops, because there are multiple energy monitor groups; you might get significantly higher performance when using only a single energy monitor group.\n";
1430 fprintf(stderr, "\n%s\n", note);
1434 fprintf(fp, "\n%s\n", note);
1440 if (bAllvsAll && fp && MASTER(cr))
1442 fprintf(fp, "\nUsing accelerated all-vs-all kernels.\n\n");
1449 static void init_forcerec_f_threads(t_forcerec *fr, int nenergrp)
1453 /* These thread local data structures are used for bondeds only */
1454 fr->nthreads = gmx_omp_nthreads_get(emntBonded);
1456 if (fr->nthreads > 1)
1458 snew(fr->f_t, fr->nthreads);
1459 /* Thread 0 uses the global force and energy arrays */
1460 for (t = 1; t < fr->nthreads; t++)
1462 fr->f_t[t].f = NULL;
1463 fr->f_t[t].f_nalloc = 0;
1464 snew(fr->f_t[t].fshift, SHIFTS);
1465 fr->f_t[t].grpp.nener = nenergrp*nenergrp;
1466 for (i = 0; i < egNR; i++)
1468 snew(fr->f_t[t].grpp.ener[i], fr->f_t[t].grpp.nener);
1475 static void pick_nbnxn_kernel_cpu(const t_inputrec gmx_unused *ir,
1479 *kernel_type = nbnxnk4x4_PlainC;
1480 *ewald_excl = ewaldexclTable;
1482 #ifdef GMX_NBNXN_SIMD
1484 #ifdef GMX_NBNXN_SIMD_4XN
1485 *kernel_type = nbnxnk4xN_SIMD_4xN;
1487 #ifdef GMX_NBNXN_SIMD_2XNN
1488 /* We expect the 2xNN kernels to be faster in most cases */
1489 *kernel_type = nbnxnk4xN_SIMD_2xNN;
1492 #if defined GMX_NBNXN_SIMD_4XN && defined GMX_X86_AVX_256
1493 if (EEL_RF(ir->coulombtype) || ir->coulombtype == eelCUT)
1495 /* The raw pair rate of the 4x8 kernel is higher than 2x(4+4),
1496 * 10% with HT, 50% without HT, but extra zeros interactions
1497 * can compensate. As we currently don't detect the actual use
1498 * of HT, switch to 4x8 to avoid a potential performance hit.
1500 *kernel_type = nbnxnk4xN_SIMD_4xN;
1503 if (getenv("GMX_NBNXN_SIMD_4XN") != NULL)
1505 #ifdef GMX_NBNXN_SIMD_4XN
1506 *kernel_type = nbnxnk4xN_SIMD_4xN;
1508 gmx_fatal(FARGS, "SIMD 4xN kernels requested, but Gromacs has been compiled without support for these kernels");
1511 if (getenv("GMX_NBNXN_SIMD_2XNN") != NULL)
1513 #ifdef GMX_NBNXN_SIMD_2XNN
1514 *kernel_type = nbnxnk4xN_SIMD_2xNN;
1516 gmx_fatal(FARGS, "SIMD 2x(N+N) kernels requested, but Gromacs has been compiled without support for these kernels");
1520 /* Analytical Ewald exclusion correction is only an option in
1521 * the SIMD kernel. On BlueGene/Q, this is faster regardless
1522 * of precision. In single precision, this is faster on
1523 * Bulldozer, and slightly faster on Sandy Bridge.
1525 #if ((defined GMX_X86_AVX_128_FMA || defined GMX_X86_AVX_256) && !defined GMX_DOUBLE) || (defined GMX_CPU_ACCELERATION_IBM_QPX)
1526 *ewald_excl = ewaldexclAnalytical;
1528 if (getenv("GMX_NBNXN_EWALD_TABLE") != NULL)
1530 *ewald_excl = ewaldexclTable;
1532 if (getenv("GMX_NBNXN_EWALD_ANALYTICAL") != NULL)
1534 *ewald_excl = ewaldexclAnalytical;
1538 #endif /* GMX_NBNXN_SIMD */
1542 const char *lookup_nbnxn_kernel_name(int kernel_type)
1544 const char *returnvalue = NULL;
1545 switch (kernel_type)
1548 returnvalue = "not set";
1550 case nbnxnk4x4_PlainC:
1551 returnvalue = "plain C";
1553 case nbnxnk4xN_SIMD_4xN:
1554 case nbnxnk4xN_SIMD_2xNN:
1555 #ifdef GMX_NBNXN_SIMD
1557 /* We have x86 SSE2 compatible SIMD */
1558 #ifdef GMX_X86_AVX_128_FMA
1559 returnvalue = "AVX-128-FMA";
1561 #if defined GMX_X86_AVX_256 || defined __AVX__
1562 /* x86 SIMD intrinsics can be converted to SSE or AVX depending
1563 * on compiler flags. As we use nearly identical intrinsics,
1564 * compiling for AVX without an AVX macros effectively results
1566 * For gcc we check for __AVX__
1567 * At least a check for icc should be added (if there is a macro)
1569 #if defined GMX_X86_AVX_256 && !defined GMX_NBNXN_HALF_WIDTH_SIMD
1570 returnvalue = "AVX-256";
1572 returnvalue = "AVX-128";
1575 #ifdef GMX_X86_SSE4_1
1576 returnvalue = "SSE4.1";
1578 returnvalue = "SSE2";
1582 #else /* GMX_X86_SSE2 */
1583 /* not GMX_X86_SSE2, but other SIMD */
1584 returnvalue = "SIMD";
1585 #endif /* GMX_X86_SSE2 */
1586 #else /* GMX_NBNXN_SIMD */
1587 returnvalue = "not available";
1588 #endif /* GMX_NBNXN_SIMD */
1590 case nbnxnk8x8x8_CUDA: returnvalue = "CUDA"; break;
1591 case nbnxnk8x8x8_PlainC: returnvalue = "plain C"; break;
1595 gmx_fatal(FARGS, "Illegal kernel type selected");
1602 static void pick_nbnxn_kernel(FILE *fp,
1603 const t_commrec *cr,
1604 gmx_bool use_cpu_acceleration,
1606 gmx_bool bEmulateGPU,
1607 const t_inputrec *ir,
1610 gmx_bool bDoNonbonded)
1612 assert(kernel_type);
1614 *kernel_type = nbnxnkNotSet;
1615 *ewald_excl = ewaldexclTable;
1619 *kernel_type = nbnxnk8x8x8_PlainC;
1623 md_print_warn(cr, fp, "Emulating a GPU run on the CPU (slow)");
1628 *kernel_type = nbnxnk8x8x8_CUDA;
1631 if (*kernel_type == nbnxnkNotSet)
1633 if (use_cpu_acceleration)
1635 pick_nbnxn_kernel_cpu(ir, kernel_type, ewald_excl);
1639 *kernel_type = nbnxnk4x4_PlainC;
1643 if (bDoNonbonded && fp != NULL)
1645 fprintf(fp, "\nUsing %s %dx%d non-bonded kernels\n\n",
1646 lookup_nbnxn_kernel_name(*kernel_type),
1647 nbnxn_kernel_pairlist_simple(*kernel_type) ? NBNXN_CPU_CLUSTER_I_SIZE : NBNXN_GPU_CLUSTER_SIZE,
1648 nbnxn_kernel_to_cj_size(*kernel_type));
1652 static void pick_nbnxn_resources(const t_commrec *cr,
1653 const gmx_hw_info_t *hwinfo,
1654 gmx_bool bDoNonbonded,
1656 gmx_bool *bEmulateGPU,
1657 const gmx_gpu_opt_t *gpu_opt)
1659 gmx_bool bEmulateGPUEnvVarSet;
1660 char gpu_err_str[STRLEN];
1664 bEmulateGPUEnvVarSet = (getenv("GMX_EMULATE_GPU") != NULL);
1666 /* Run GPU emulation mode if GMX_EMULATE_GPU is defined. Because
1667 * GPUs (currently) only handle non-bonded calculations, we will
1668 * automatically switch to emulation if non-bonded calculations are
1669 * turned off via GMX_NO_NONBONDED - this is the simple and elegant
1670 * way to turn off GPU initialization, data movement, and cleanup.
1672 * GPU emulation can be useful to assess the performance one can expect by
1673 * adding GPU(s) to the machine. The conditional below allows this even
1674 * if mdrun is compiled without GPU acceleration support.
1675 * Note that you should freezing the system as otherwise it will explode.
1677 *bEmulateGPU = (bEmulateGPUEnvVarSet ||
1679 gpu_opt->ncuda_dev_use > 0));
1681 /* Enable GPU mode when GPUs are available or no GPU emulation is requested.
1683 if (gpu_opt->ncuda_dev_use > 0 && !(*bEmulateGPU))
1685 /* Each PP node will use the intra-node id-th device from the
1686 * list of detected/selected GPUs. */
1687 if (!init_gpu(cr->rank_pp_intranode, gpu_err_str,
1688 &hwinfo->gpu_info, gpu_opt))
1690 /* At this point the init should never fail as we made sure that
1691 * we have all the GPUs we need. If it still does, we'll bail. */
1692 gmx_fatal(FARGS, "On node %d failed to initialize GPU #%d: %s",
1694 get_gpu_device_id(&hwinfo->gpu_info, gpu_opt,
1695 cr->rank_pp_intranode),
1699 /* Here we actually turn on hardware GPU acceleration */
1704 gmx_bool uses_simple_tables(int cutoff_scheme,
1705 nonbonded_verlet_t *nbv,
1708 gmx_bool bUsesSimpleTables = TRUE;
1711 switch (cutoff_scheme)
1714 bUsesSimpleTables = TRUE;
1717 assert(NULL != nbv && NULL != nbv->grp);
1718 grp_index = (group < 0) ? 0 : (nbv->ngrp - 1);
1719 bUsesSimpleTables = nbnxn_kernel_pairlist_simple(nbv->grp[grp_index].kernel_type);
1722 gmx_incons("unimplemented");
1724 return bUsesSimpleTables;
1727 static void init_ewald_f_table(interaction_const_t *ic,
1728 gmx_bool bUsesSimpleTables,
1733 if (bUsesSimpleTables)
1735 /* With a spacing of 0.0005 we are at the force summation accuracy
1736 * for the SSE kernels for "normal" atomistic simulations.
1738 ic->tabq_scale = ewald_spline3_table_scale(ic->ewaldcoeff,
1741 maxr = (rtab > ic->rcoulomb) ? rtab : ic->rcoulomb;
1742 ic->tabq_size = (int)(maxr*ic->tabq_scale) + 2;
1746 ic->tabq_size = GPU_EWALD_COULOMB_FORCE_TABLE_SIZE;
1747 /* Subtract 2 iso 1 to avoid access out of range due to rounding */
1748 ic->tabq_scale = (ic->tabq_size - 2)/ic->rcoulomb;
1751 sfree_aligned(ic->tabq_coul_FDV0);
1752 sfree_aligned(ic->tabq_coul_F);
1753 sfree_aligned(ic->tabq_coul_V);
1755 /* Create the original table data in FDV0 */
1756 snew_aligned(ic->tabq_coul_FDV0, ic->tabq_size*4, 32);
1757 snew_aligned(ic->tabq_coul_F, ic->tabq_size, 32);
1758 snew_aligned(ic->tabq_coul_V, ic->tabq_size, 32);
1759 table_spline3_fill_ewald_lr(ic->tabq_coul_F, ic->tabq_coul_V, ic->tabq_coul_FDV0,
1760 ic->tabq_size, 1/ic->tabq_scale, ic->ewaldcoeff);
1763 void init_interaction_const_tables(FILE *fp,
1764 interaction_const_t *ic,
1765 gmx_bool bUsesSimpleTables,
1770 if (ic->eeltype == eelEWALD || EEL_PME(ic->eeltype))
1772 init_ewald_f_table(ic, bUsesSimpleTables, rtab);
1776 fprintf(fp, "Initialized non-bonded Ewald correction tables, spacing: %.2e size: %d\n\n",
1777 1/ic->tabq_scale, ic->tabq_size);
1782 static void init_interaction_const(FILE *fp,
1783 const t_commrec *cr,
1784 interaction_const_t **interaction_const,
1785 const t_forcerec *fr,
1788 interaction_const_t *ic;
1789 gmx_bool bUsesSimpleTables = TRUE;
1793 /* Just allocate something so we can free it */
1794 snew_aligned(ic->tabq_coul_FDV0, 16, 32);
1795 snew_aligned(ic->tabq_coul_F, 16, 32);
1796 snew_aligned(ic->tabq_coul_V, 16, 32);
1798 ic->rlist = fr->rlist;
1799 ic->rlistlong = fr->rlistlong;
1802 ic->rvdw = fr->rvdw;
1803 if (fr->vdw_modifier == eintmodPOTSHIFT)
1805 ic->sh_invrc6 = pow(ic->rvdw, -6.0);
1812 /* Electrostatics */
1813 ic->eeltype = fr->eeltype;
1814 ic->rcoulomb = fr->rcoulomb;
1815 ic->epsilon_r = fr->epsilon_r;
1816 ic->epsfac = fr->epsfac;
1819 ic->ewaldcoeff = fr->ewaldcoeff;
1820 if (fr->coulomb_modifier == eintmodPOTSHIFT)
1822 ic->sh_ewald = gmx_erfc(ic->ewaldcoeff*ic->rcoulomb);
1829 /* Reaction-field */
1830 if (EEL_RF(ic->eeltype))
1832 ic->epsilon_rf = fr->epsilon_rf;
1833 ic->k_rf = fr->k_rf;
1834 ic->c_rf = fr->c_rf;
1838 /* For plain cut-off we might use the reaction-field kernels */
1839 ic->epsilon_rf = ic->epsilon_r;
1841 if (fr->coulomb_modifier == eintmodPOTSHIFT)
1843 ic->c_rf = 1/ic->rcoulomb;
1853 fprintf(fp, "Potential shift: LJ r^-12: %.3f r^-6 %.3f",
1854 sqr(ic->sh_invrc6), ic->sh_invrc6);
1855 if (ic->eeltype == eelCUT)
1857 fprintf(fp, ", Coulomb %.3f", ic->c_rf);
1859 else if (EEL_PME(ic->eeltype))
1861 fprintf(fp, ", Ewald %.3e", ic->sh_ewald);
1866 *interaction_const = ic;
1868 if (fr->nbv != NULL && fr->nbv->bUseGPU)
1870 nbnxn_cuda_init_const(fr->nbv->cu_nbv, ic, fr->nbv->grp);
1872 /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
1873 * also sharing texture references. To keep the code simple, we don't
1874 * treat texture references as shared resources, but this means that
1875 * the coulomb_tab and nbfp texture refs will get updated by multiple threads.
1876 * Hence, to ensure that the non-bonded kernels don't start before all
1877 * texture binding operations are finished, we need to wait for all ranks
1878 * to arrive here before continuing.
1880 * Note that we could omit this barrier if GPUs are not shared (or
1881 * texture objects are used), but as this is initialization code, there
1882 * is not point in complicating things.
1884 #ifdef GMX_THREAD_MPI
1889 #endif /* GMX_THREAD_MPI */
1892 bUsesSimpleTables = uses_simple_tables(fr->cutoff_scheme, fr->nbv, -1);
1893 init_interaction_const_tables(fp, ic, bUsesSimpleTables, rtab);
1896 static void init_nb_verlet(FILE *fp,
1897 nonbonded_verlet_t **nb_verlet,
1898 const t_inputrec *ir,
1899 const t_forcerec *fr,
1900 const t_commrec *cr,
1901 const char *nbpu_opt)
1903 nonbonded_verlet_t *nbv;
1906 gmx_bool bEmulateGPU, bHybridGPURun = FALSE;
1908 nbnxn_alloc_t *nb_alloc;
1909 nbnxn_free_t *nb_free;
1913 pick_nbnxn_resources(cr, fr->hwinfo,
1921 nbv->ngrp = (DOMAINDECOMP(cr) ? 2 : 1);
1922 for (i = 0; i < nbv->ngrp; i++)
1924 nbv->grp[i].nbl_lists.nnbl = 0;
1925 nbv->grp[i].nbat = NULL;
1926 nbv->grp[i].kernel_type = nbnxnkNotSet;
1928 if (i == 0) /* local */
1930 pick_nbnxn_kernel(fp, cr, fr->use_cpu_acceleration,
1931 nbv->bUseGPU, bEmulateGPU, ir,
1932 &nbv->grp[i].kernel_type,
1933 &nbv->grp[i].ewald_excl,
1936 else /* non-local */
1938 if (nbpu_opt != NULL && strcmp(nbpu_opt, "gpu_cpu") == 0)
1940 /* Use GPU for local, select a CPU kernel for non-local */
1941 pick_nbnxn_kernel(fp, cr, fr->use_cpu_acceleration,
1943 &nbv->grp[i].kernel_type,
1944 &nbv->grp[i].ewald_excl,
1947 bHybridGPURun = TRUE;
1951 /* Use the same kernel for local and non-local interactions */
1952 nbv->grp[i].kernel_type = nbv->grp[0].kernel_type;
1953 nbv->grp[i].ewald_excl = nbv->grp[0].ewald_excl;
1960 /* init the NxN GPU data; the last argument tells whether we'll have
1961 * both local and non-local NB calculation on GPU */
1962 nbnxn_cuda_init(fp, &nbv->cu_nbv,
1963 &fr->hwinfo->gpu_info, fr->gpu_opt,
1964 cr->rank_pp_intranode,
1965 (nbv->ngrp > 1) && !bHybridGPURun);
1967 if ((env = getenv("GMX_NB_MIN_CI")) != NULL)
1971 nbv->min_ci_balanced = strtol(env, &end, 10);
1972 if (!end || (*end != 0) || nbv->min_ci_balanced <= 0)
1974 gmx_fatal(FARGS, "Invalid value passed in GMX_NB_MIN_CI=%s, positive integer required", env);
1979 fprintf(debug, "Neighbor-list balancing parameter: %d (passed as env. var.)\n",
1980 nbv->min_ci_balanced);
1985 nbv->min_ci_balanced = nbnxn_cuda_min_ci_balanced(nbv->cu_nbv);
1988 fprintf(debug, "Neighbor-list balancing parameter: %d (auto-adjusted to the number of GPU multi-processors)\n",
1989 nbv->min_ci_balanced);
1995 nbv->min_ci_balanced = 0;
2000 nbnxn_init_search(&nbv->nbs,
2001 DOMAINDECOMP(cr) ? &cr->dd->nc : NULL,
2002 DOMAINDECOMP(cr) ? domdec_zones(cr->dd) : NULL,
2003 gmx_omp_nthreads_get(emntNonbonded));
2005 for (i = 0; i < nbv->ngrp; i++)
2007 if (nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
2009 nb_alloc = &pmalloc;
2018 nbnxn_init_pairlist_set(&nbv->grp[i].nbl_lists,
2019 nbnxn_kernel_pairlist_simple(nbv->grp[i].kernel_type),
2020 /* 8x8x8 "non-simple" lists are ATM always combined */
2021 !nbnxn_kernel_pairlist_simple(nbv->grp[i].kernel_type),
2025 nbv->grp[0].kernel_type != nbv->grp[i].kernel_type)
2027 snew(nbv->grp[i].nbat, 1);
2028 nbnxn_atomdata_init(fp,
2030 nbv->grp[i].kernel_type,
2031 fr->ntype, fr->nbfp,
2033 nbnxn_kernel_pairlist_simple(nbv->grp[i].kernel_type) ? gmx_omp_nthreads_get(emntNonbonded) : 1,
2038 nbv->grp[i].nbat = nbv->grp[0].nbat;
2043 void init_forcerec(FILE *fp,
2044 const output_env_t oenv,
2047 const t_inputrec *ir,
2048 const gmx_mtop_t *mtop,
2049 const t_commrec *cr,
2055 const char *nbpu_opt,
2056 gmx_bool bNoSolvOpt,
2059 int i, j, m, natoms, ngrp, negp_pp, negptable, egi, egj;
2064 gmx_bool bGenericKernelOnly;
2065 gmx_bool bTab, bSep14tab, bNormalnblists;
2067 int *nm_ind, egp_flags;
2069 if (fr->hwinfo == NULL)
2071 /* Detect hardware, gather information.
2072 * In mdrun, hwinfo has already been set before calling init_forcerec.
2073 * Here we ignore GPUs, as tools will not use them anyhow.
2075 fr->hwinfo = gmx_detect_hardware(fp, cr, FALSE);
2078 /* By default we turn acceleration on, but it might be turned off further down... */
2079 fr->use_cpu_acceleration = TRUE;
2081 fr->bDomDec = DOMAINDECOMP(cr);
2083 natoms = mtop->natoms;
2085 if (check_box(ir->ePBC, box))
2087 gmx_fatal(FARGS, check_box(ir->ePBC, box));
2090 /* Test particle insertion ? */
2093 /* Set to the size of the molecule to be inserted (the last one) */
2094 /* Because of old style topologies, we have to use the last cg
2095 * instead of the last molecule type.
2097 cgs = &mtop->moltype[mtop->molblock[mtop->nmolblock-1].type].cgs;
2098 fr->n_tpi = cgs->index[cgs->nr] - cgs->index[cgs->nr-1];
2099 if (fr->n_tpi != mtop->mols.index[mtop->mols.nr] - mtop->mols.index[mtop->mols.nr-1])
2101 gmx_fatal(FARGS, "The molecule to insert can not consist of multiple charge groups.\nMake it a single charge group.");
2109 /* Copy AdResS parameters */
2112 fr->adress_type = ir->adress->type;
2113 fr->adress_const_wf = ir->adress->const_wf;
2114 fr->adress_ex_width = ir->adress->ex_width;
2115 fr->adress_hy_width = ir->adress->hy_width;
2116 fr->adress_icor = ir->adress->icor;
2117 fr->adress_site = ir->adress->site;
2118 fr->adress_ex_forcecap = ir->adress->ex_forcecap;
2119 fr->adress_do_hybridpairs = ir->adress->do_hybridpairs;
2122 snew(fr->adress_group_explicit, ir->adress->n_energy_grps);
2123 for (i = 0; i < ir->adress->n_energy_grps; i++)
2125 fr->adress_group_explicit[i] = ir->adress->group_explicit[i];
2128 fr->n_adress_tf_grps = ir->adress->n_tf_grps;
2129 snew(fr->adress_tf_table_index, fr->n_adress_tf_grps);
2130 for (i = 0; i < fr->n_adress_tf_grps; i++)
2132 fr->adress_tf_table_index[i] = ir->adress->tf_table_index[i];
2134 copy_rvec(ir->adress->refs, fr->adress_refs);
2138 fr->adress_type = eAdressOff;
2139 fr->adress_do_hybridpairs = FALSE;
2142 /* Copy the user determined parameters */
2143 fr->userint1 = ir->userint1;
2144 fr->userint2 = ir->userint2;
2145 fr->userint3 = ir->userint3;
2146 fr->userint4 = ir->userint4;
2147 fr->userreal1 = ir->userreal1;
2148 fr->userreal2 = ir->userreal2;
2149 fr->userreal3 = ir->userreal3;
2150 fr->userreal4 = ir->userreal4;
2153 fr->fc_stepsize = ir->fc_stepsize;
2156 fr->efep = ir->efep;
2157 fr->sc_alphavdw = ir->fepvals->sc_alpha;
2158 if (ir->fepvals->bScCoul)
2160 fr->sc_alphacoul = ir->fepvals->sc_alpha;
2161 fr->sc_sigma6_min = pow(ir->fepvals->sc_sigma_min, 6);
2165 fr->sc_alphacoul = 0;
2166 fr->sc_sigma6_min = 0; /* only needed when bScCoul is on */
2168 fr->sc_power = ir->fepvals->sc_power;
2169 fr->sc_r_power = ir->fepvals->sc_r_power;
2170 fr->sc_sigma6_def = pow(ir->fepvals->sc_sigma, 6);
2172 env = getenv("GMX_SCSIGMA_MIN");
2176 sscanf(env, "%lf", &dbl);
2177 fr->sc_sigma6_min = pow(dbl, 6);
2180 fprintf(fp, "Setting the minimum soft core sigma to %g nm\n", dbl);
2184 fr->bNonbonded = TRUE;
2185 if (getenv("GMX_NO_NONBONDED") != NULL)
2187 /* turn off non-bonded calculations */
2188 fr->bNonbonded = FALSE;
2189 md_print_warn(cr, fp,
2190 "Found environment variable GMX_NO_NONBONDED.\n"
2191 "Disabling nonbonded calculations.\n");
2194 bGenericKernelOnly = FALSE;
2196 /* We now check in the NS code whether a particular combination of interactions
2197 * can be used with water optimization, and disable it if that is not the case.
2200 if (getenv("GMX_NB_GENERIC") != NULL)
2205 "Found environment variable GMX_NB_GENERIC.\n"
2206 "Disabling all interaction-specific nonbonded kernels, will only\n"
2207 "use the slow generic ones in src/gmxlib/nonbonded/nb_generic.c\n\n");
2209 bGenericKernelOnly = TRUE;
2212 if (bGenericKernelOnly == TRUE)
2217 if ( (getenv("GMX_DISABLE_CPU_ACCELERATION") != NULL) || (getenv("GMX_NOOPTIMIZEDKERNELS") != NULL) )
2219 fr->use_cpu_acceleration = FALSE;
2223 "\nFound environment variable GMX_DISABLE_CPU_ACCELERATION.\n"
2224 "Disabling all CPU architecture-specific (e.g. SSE2/SSE4/AVX) routines.\n\n");
2228 fr->bBHAM = (mtop->ffparams.functype[0] == F_BHAM);
2230 /* Check if we can/should do all-vs-all kernels */
2231 fr->bAllvsAll = can_use_allvsall(ir, FALSE, NULL, NULL);
2232 fr->AllvsAll_work = NULL;
2233 fr->AllvsAll_workgb = NULL;
2235 /* All-vs-all kernels have not been implemented in 4.6, and
2236 * the SIMD group kernels are also buggy in this case. Non-accelerated
2237 * group kernels are OK. See Redmine #1249. */
2240 fr->bAllvsAll = FALSE;
2241 fr->use_cpu_acceleration = FALSE;
2245 "\nYour simulation settings would have triggered the efficient all-vs-all\n"
2246 "kernels in GROMACS 4.5, but these have not been implemented in GROMACS\n"
2247 "4.6. Also, we can't use the accelerated SIMD kernels here because\n"
2248 "of an unfixed bug. The reference C kernels are correct, though, so\n"
2249 "we are proceeding by disabling all CPU architecture-specific\n"
2250 "(e.g. SSE2/SSE4/AVX) routines. If performance is important, please\n"
2251 "use GROMACS 4.5.7 or try cutoff-scheme = Verlet.\n\n");
2255 /* Neighbour searching stuff */
2256 fr->cutoff_scheme = ir->cutoff_scheme;
2257 fr->bGrid = (ir->ns_type == ensGRID);
2258 fr->ePBC = ir->ePBC;
2260 /* Determine if we will do PBC for distances in bonded interactions */
2261 if (fr->ePBC == epbcNONE)
2263 fr->bMolPBC = FALSE;
2267 if (!DOMAINDECOMP(cr))
2269 /* The group cut-off scheme and SHAKE assume charge groups
2270 * are whole, but not using molpbc is faster in most cases.
2272 if (fr->cutoff_scheme == ecutsGROUP ||
2273 (ir->eConstrAlg == econtSHAKE &&
2274 (gmx_mtop_ftype_count(mtop, F_CONSTR) > 0 ||
2275 gmx_mtop_ftype_count(mtop, F_CONSTRNC) > 0)))
2277 fr->bMolPBC = ir->bPeriodicMols;
2282 if (getenv("GMX_USE_GRAPH") != NULL)
2284 fr->bMolPBC = FALSE;
2287 fprintf(fp, "\nGMX_MOLPBC is set, using the graph for bonded interactions\n\n");
2294 fr->bMolPBC = dd_bonded_molpbc(cr->dd, fr->ePBC);
2297 fr->bGB = (ir->implicit_solvent == eisGBSA);
2299 fr->rc_scaling = ir->refcoord_scaling;
2300 copy_rvec(ir->posres_com, fr->posres_com);
2301 copy_rvec(ir->posres_comB, fr->posres_comB);
2302 fr->rlist = cutoff_inf(ir->rlist);
2303 fr->rlistlong = cutoff_inf(ir->rlistlong);
2304 fr->eeltype = ir->coulombtype;
2305 fr->vdwtype = ir->vdwtype;
2307 fr->coulomb_modifier = ir->coulomb_modifier;
2308 fr->vdw_modifier = ir->vdw_modifier;
2310 /* Electrostatics: Translate from interaction-setting-in-mdp-file to kernel interaction format */
2311 switch (fr->eeltype)
2314 fr->nbkernel_elec_interaction = (fr->bGB) ? GMX_NBKERNEL_ELEC_GENERALIZEDBORN : GMX_NBKERNEL_ELEC_COULOMB;
2320 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_REACTIONFIELD;
2324 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_REACTIONFIELD;
2325 fr->coulomb_modifier = eintmodEXACTCUTOFF;
2334 case eelPMEUSERSWITCH:
2335 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_CUBICSPLINETABLE;
2340 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_EWALD;
2344 gmx_fatal(FARGS, "Unsupported electrostatic interaction: %s", eel_names[fr->eeltype]);
2348 /* Vdw: Translate from mdp settings to kernel format */
2349 switch (fr->vdwtype)
2354 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_BUCKINGHAM;
2358 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_LENNARDJONES;
2365 case evdwENCADSHIFT:
2366 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_CUBICSPLINETABLE;
2370 gmx_fatal(FARGS, "Unsupported vdw interaction: %s", evdw_names[fr->vdwtype]);
2374 /* These start out identical to ir, but might be altered if we e.g. tabulate the interaction in the kernel */
2375 fr->nbkernel_elec_modifier = fr->coulomb_modifier;
2376 fr->nbkernel_vdw_modifier = fr->vdw_modifier;
2378 fr->bTwinRange = fr->rlistlong > fr->rlist;
2379 fr->bEwald = (EEL_PME(fr->eeltype) || fr->eeltype == eelEWALD);
2381 fr->reppow = mtop->ffparams.reppow;
2383 if (ir->cutoff_scheme == ecutsGROUP)
2385 fr->bvdwtab = (fr->vdwtype != evdwCUT ||
2386 !gmx_within_tol(fr->reppow, 12.0, 10*GMX_DOUBLE_EPS));
2387 /* We have special kernels for standard Ewald and PME, but the pme-switch ones are tabulated above */
2388 fr->bcoultab = !(fr->eeltype == eelCUT ||
2389 fr->eeltype == eelEWALD ||
2390 fr->eeltype == eelPME ||
2391 fr->eeltype == eelRF ||
2392 fr->eeltype == eelRF_ZERO);
2394 /* If the user absolutely wants different switch/shift settings for coul/vdw, it is likely
2395 * going to be faster to tabulate the interaction than calling the generic kernel.
2397 if (fr->nbkernel_elec_modifier == eintmodPOTSWITCH && fr->nbkernel_vdw_modifier == eintmodPOTSWITCH)
2399 if ((fr->rcoulomb_switch != fr->rvdw_switch) || (fr->rcoulomb != fr->rvdw))
2401 fr->bcoultab = TRUE;
2404 else if ((fr->nbkernel_elec_modifier == eintmodPOTSHIFT && fr->nbkernel_vdw_modifier == eintmodPOTSHIFT) ||
2405 ((fr->nbkernel_elec_interaction == GMX_NBKERNEL_ELEC_REACTIONFIELD &&
2406 fr->nbkernel_elec_modifier == eintmodEXACTCUTOFF &&
2407 (fr->nbkernel_vdw_modifier == eintmodPOTSWITCH || fr->nbkernel_vdw_modifier == eintmodPOTSHIFT))))
2409 if (fr->rcoulomb != fr->rvdw)
2411 fr->bcoultab = TRUE;
2415 if (getenv("GMX_REQUIRE_TABLES"))
2418 fr->bcoultab = TRUE;
2423 fprintf(fp, "Table routines are used for coulomb: %s\n", bool_names[fr->bcoultab]);
2424 fprintf(fp, "Table routines are used for vdw: %s\n", bool_names[fr->bvdwtab ]);
2427 if (fr->bvdwtab == TRUE)
2429 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_CUBICSPLINETABLE;
2430 fr->nbkernel_vdw_modifier = eintmodNONE;
2432 if (fr->bcoultab == TRUE)
2434 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_CUBICSPLINETABLE;
2435 fr->nbkernel_elec_modifier = eintmodNONE;
2439 if (ir->cutoff_scheme == ecutsVERLET)
2441 if (!gmx_within_tol(fr->reppow, 12.0, 10*GMX_DOUBLE_EPS))
2443 gmx_fatal(FARGS, "Cut-off scheme %S only supports LJ repulsion power 12", ecutscheme_names[ir->cutoff_scheme]);
2445 fr->bvdwtab = FALSE;
2446 fr->bcoultab = FALSE;
2449 /* Tables are used for direct ewald sum */
2452 if (EEL_PME(ir->coulombtype))
2456 fprintf(fp, "Will do PME sum in reciprocal space.\n");
2458 if (ir->coulombtype == eelP3M_AD)
2460 please_cite(fp, "Hockney1988");
2461 please_cite(fp, "Ballenegger2012");
2465 please_cite(fp, "Essmann95a");
2468 if (ir->ewald_geometry == eewg3DC)
2472 fprintf(fp, "Using the Ewald3DC correction for systems with a slab geometry.\n");
2474 please_cite(fp, "In-Chul99a");
2477 fr->ewaldcoeff = calc_ewaldcoeff(ir->rcoulomb, ir->ewald_rtol);
2478 init_ewald_tab(&(fr->ewald_table), ir, fp);
2481 fprintf(fp, "Using a Gaussian width (1/beta) of %g nm for Ewald\n",
2486 /* Electrostatics */
2487 fr->epsilon_r = ir->epsilon_r;
2488 fr->epsilon_rf = ir->epsilon_rf;
2489 fr->fudgeQQ = mtop->ffparams.fudgeQQ;
2490 fr->rcoulomb_switch = ir->rcoulomb_switch;
2491 fr->rcoulomb = cutoff_inf(ir->rcoulomb);
2493 /* Parameters for generalized RF */
2497 if (fr->eeltype == eelGRF)
2499 init_generalized_rf(fp, mtop, ir, fr);
2502 fr->bF_NoVirSum = (EEL_FULL(fr->eeltype) ||
2503 gmx_mtop_ftype_count(mtop, F_POSRES) > 0 ||
2504 gmx_mtop_ftype_count(mtop, F_FBPOSRES) > 0 ||
2505 IR_ELEC_FIELD(*ir) ||
2506 (fr->adress_icor != eAdressICOff)
2509 if (fr->cutoff_scheme == ecutsGROUP &&
2510 ncg_mtop(mtop) > fr->cg_nalloc && !DOMAINDECOMP(cr))
2512 /* Count the total number of charge groups */
2513 fr->cg_nalloc = ncg_mtop(mtop);
2514 srenew(fr->cg_cm, fr->cg_nalloc);
2516 if (fr->shift_vec == NULL)
2518 snew(fr->shift_vec, SHIFTS);
2521 if (fr->fshift == NULL)
2523 snew(fr->fshift, SHIFTS);
2526 if (fr->nbfp == NULL)
2528 fr->ntype = mtop->ffparams.atnr;
2529 fr->nbfp = mk_nbfp(&mtop->ffparams, fr->bBHAM);
2532 /* Copy the energy group exclusions */
2533 fr->egp_flags = ir->opts.egp_flags;
2535 /* Van der Waals stuff */
2536 fr->rvdw = cutoff_inf(ir->rvdw);
2537 fr->rvdw_switch = ir->rvdw_switch;
2538 if ((fr->vdwtype != evdwCUT) && (fr->vdwtype != evdwUSER) && !fr->bBHAM)
2540 if (fr->rvdw_switch >= fr->rvdw)
2542 gmx_fatal(FARGS, "rvdw_switch (%f) must be < rvdw (%f)",
2543 fr->rvdw_switch, fr->rvdw);
2547 fprintf(fp, "Using %s Lennard-Jones, switch between %g and %g nm\n",
2548 (fr->eeltype == eelSWITCH) ? "switched" : "shifted",
2549 fr->rvdw_switch, fr->rvdw);
2553 if (fr->bBHAM && (fr->vdwtype == evdwSHIFT || fr->vdwtype == evdwSWITCH))
2555 gmx_fatal(FARGS, "Switch/shift interaction not supported with Buckingham");
2560 fprintf(fp, "Cut-off's: NS: %g Coulomb: %g %s: %g\n",
2561 fr->rlist, fr->rcoulomb, fr->bBHAM ? "BHAM" : "LJ", fr->rvdw);
2564 fr->eDispCorr = ir->eDispCorr;
2565 if (ir->eDispCorr != edispcNO)
2567 set_avcsixtwelve(fp, fr, mtop);
2572 set_bham_b_max(fp, fr, mtop);
2575 fr->gb_epsilon_solvent = ir->gb_epsilon_solvent;
2577 /* Copy the GBSA data (radius, volume and surftens for each
2578 * atomtype) from the topology atomtype section to forcerec.
2580 snew(fr->atype_radius, fr->ntype);
2581 snew(fr->atype_vol, fr->ntype);
2582 snew(fr->atype_surftens, fr->ntype);
2583 snew(fr->atype_gb_radius, fr->ntype);
2584 snew(fr->atype_S_hct, fr->ntype);
2586 if (mtop->atomtypes.nr > 0)
2588 for (i = 0; i < fr->ntype; i++)
2590 fr->atype_radius[i] = mtop->atomtypes.radius[i];
2592 for (i = 0; i < fr->ntype; i++)
2594 fr->atype_vol[i] = mtop->atomtypes.vol[i];
2596 for (i = 0; i < fr->ntype; i++)
2598 fr->atype_surftens[i] = mtop->atomtypes.surftens[i];
2600 for (i = 0; i < fr->ntype; i++)
2602 fr->atype_gb_radius[i] = mtop->atomtypes.gb_radius[i];
2604 for (i = 0; i < fr->ntype; i++)
2606 fr->atype_S_hct[i] = mtop->atomtypes.S_hct[i];
2610 /* Generate the GB table if needed */
2614 fr->gbtabscale = 2000;
2616 fr->gbtabscale = 500;
2620 fr->gbtab = make_gb_table(oenv, fr);
2622 init_gb(&fr->born, cr, fr, ir, mtop, ir->gb_algorithm);
2624 /* Copy local gb data (for dd, this is done in dd_partition_system) */
2625 if (!DOMAINDECOMP(cr))
2627 make_local_gb(cr, fr->born, ir->gb_algorithm);
2631 /* Set the charge scaling */
2632 if (fr->epsilon_r != 0)
2634 fr->epsfac = ONE_4PI_EPS0/fr->epsilon_r;
2638 /* eps = 0 is infinite dieletric: no coulomb interactions */
2642 /* Reaction field constants */
2643 if (EEL_RF(fr->eeltype))
2645 calc_rffac(fp, fr->eeltype, fr->epsilon_r, fr->epsilon_rf,
2646 fr->rcoulomb, fr->temp, fr->zsquare, box,
2647 &fr->kappa, &fr->k_rf, &fr->c_rf);
2650 set_chargesum(fp, fr, mtop);
2652 /* if we are using LR electrostatics, and they are tabulated,
2653 * the tables will contain modified coulomb interactions.
2654 * Since we want to use the non-shifted ones for 1-4
2655 * coulombic interactions, we must have an extra set of tables.
2658 /* Construct tables.
2659 * A little unnecessary to make both vdw and coul tables sometimes,
2660 * but what the heck... */
2662 bTab = fr->bcoultab || fr->bvdwtab || fr->bEwald;
2664 bSep14tab = ((!bTab || fr->eeltype != eelCUT || fr->vdwtype != evdwCUT ||
2665 fr->bBHAM || fr->bEwald) &&
2666 (gmx_mtop_ftype_count(mtop, F_LJ14) > 0 ||
2667 gmx_mtop_ftype_count(mtop, F_LJC14_Q) > 0 ||
2668 gmx_mtop_ftype_count(mtop, F_LJC_PAIRS_NB) > 0));
2670 negp_pp = ir->opts.ngener - ir->nwall;
2674 bNormalnblists = TRUE;
2679 bNormalnblists = (ir->eDispCorr != edispcNO);
2680 for (egi = 0; egi < negp_pp; egi++)
2682 for (egj = egi; egj < negp_pp; egj++)
2684 egp_flags = ir->opts.egp_flags[GID(egi, egj, ir->opts.ngener)];
2685 if (!(egp_flags & EGP_EXCL))
2687 if (egp_flags & EGP_TABLE)
2693 bNormalnblists = TRUE;
2700 fr->nnblists = negptable + 1;
2704 fr->nnblists = negptable;
2706 if (fr->nnblists > 1)
2708 snew(fr->gid2nblists, ir->opts.ngener*ir->opts.ngener);
2717 snew(fr->nblists, fr->nnblists);
2719 /* This code automatically gives table length tabext without cut-off's,
2720 * in that case grompp should already have checked that we do not need
2721 * normal tables and we only generate tables for 1-4 interactions.
2723 rtab = ir->rlistlong + ir->tabext;
2727 /* make tables for ordinary interactions */
2730 make_nbf_tables(fp, oenv, fr, rtab, cr, tabfn, NULL, NULL, &fr->nblists[0]);
2733 make_nbf_tables(fp, oenv, fr, rtab, cr, tabfn, NULL, NULL, &fr->nblists[fr->nnblists/2]);
2737 fr->tab14 = fr->nblists[0].table_elec_vdw;
2747 /* Read the special tables for certain energy group pairs */
2748 nm_ind = mtop->groups.grps[egcENER].nm_ind;
2749 for (egi = 0; egi < negp_pp; egi++)
2751 for (egj = egi; egj < negp_pp; egj++)
2753 egp_flags = ir->opts.egp_flags[GID(egi, egj, ir->opts.ngener)];
2754 if ((egp_flags & EGP_TABLE) && !(egp_flags & EGP_EXCL))
2756 nbl = &(fr->nblists[m]);
2757 if (fr->nnblists > 1)
2759 fr->gid2nblists[GID(egi, egj, ir->opts.ngener)] = m;
2761 /* Read the table file with the two energy groups names appended */
2762 make_nbf_tables(fp, oenv, fr, rtab, cr, tabfn,
2763 *mtop->groups.grpname[nm_ind[egi]],
2764 *mtop->groups.grpname[nm_ind[egj]],
2768 make_nbf_tables(fp, oenv, fr, rtab, cr, tabfn,
2769 *mtop->groups.grpname[nm_ind[egi]],
2770 *mtop->groups.grpname[nm_ind[egj]],
2771 &fr->nblists[fr->nnblists/2+m]);
2775 else if (fr->nnblists > 1)
2777 fr->gid2nblists[GID(egi, egj, ir->opts.ngener)] = 0;
2785 /* generate extra tables with plain Coulomb for 1-4 interactions only */
2786 fr->tab14 = make_tables(fp, oenv, fr, MASTER(cr), tabpfn, rtab,
2787 GMX_MAKETABLES_14ONLY);
2790 /* Read AdResS Thermo Force table if needed */
2791 if (fr->adress_icor == eAdressICThermoForce)
2793 /* old todo replace */
2795 if (ir->adress->n_tf_grps > 0)
2797 make_adress_tf_tables(fp, oenv, fr, ir, tabfn, mtop, box);
2802 /* load the default table */
2803 snew(fr->atf_tabs, 1);
2804 fr->atf_tabs[DEFAULT_TF_TABLE] = make_atf_table(fp, oenv, fr, tabafn, box);
2809 fr->nwall = ir->nwall;
2810 if (ir->nwall && ir->wall_type == ewtTABLE)
2812 make_wall_tables(fp, oenv, ir, tabfn, &mtop->groups, fr);
2817 fcd->bondtab = make_bonded_tables(fp,
2818 F_TABBONDS, F_TABBONDSNC,
2820 fcd->angletab = make_bonded_tables(fp,
2823 fcd->dihtab = make_bonded_tables(fp,
2831 fprintf(debug, "No fcdata or table file name passed, can not read table, can not do bonded interactions\n");
2835 /* QM/MM initialization if requested
2839 fprintf(stderr, "QM/MM calculation requested.\n");
2842 fr->bQMMM = ir->bQMMM;
2843 fr->qr = mk_QMMMrec();
2845 /* Set all the static charge group info */
2846 fr->cginfo_mb = init_cginfo_mb(fp, mtop, fr, bNoSolvOpt,
2847 &fr->bExcl_IntraCGAll_InterCGNone);
2848 if (DOMAINDECOMP(cr))
2854 fr->cginfo = cginfo_expand(mtop->nmolblock, fr->cginfo_mb);
2857 if (!DOMAINDECOMP(cr))
2859 /* When using particle decomposition, the effect of the second argument,
2860 * which sets fr->hcg, is corrected later in do_md and init_em.
2862 forcerec_set_ranges(fr, ncg_mtop(mtop), ncg_mtop(mtop),
2863 mtop->natoms, mtop->natoms, mtop->natoms);
2866 fr->print_force = print_force;
2869 /* coarse load balancing vars */
2874 /* Initialize neighbor search */
2875 init_ns(fp, cr, &fr->ns, fr, mtop);
2877 if (cr->duty & DUTY_PP)
2879 gmx_nonbonded_setup(fr, bGenericKernelOnly);
2883 gmx_setup_adress_kernels(fp,bGenericKernelOnly);
2888 /* Initialize the thread working data for bonded interactions */
2889 init_forcerec_f_threads(fr, mtop->groups.grps[egcENER].nr);
2891 snew(fr->excl_load, fr->nthreads+1);
2893 if (fr->cutoff_scheme == ecutsVERLET)
2895 if (ir->rcoulomb != ir->rvdw)
2897 gmx_fatal(FARGS, "With Verlet lists rcoulomb and rvdw should be identical");
2900 init_nb_verlet(fp, &fr->nbv, ir, fr, cr, nbpu_opt);
2903 /* fr->ic is used both by verlet and group kernels (to some extent) now */
2904 init_interaction_const(fp, cr, &fr->ic, fr, rtab);
2906 if (ir->eDispCorr != edispcNO)
2908 calc_enervirdiff(fp, ir->eDispCorr, fr);
2912 #define pr_real(fp, r) fprintf(fp, "%s: %e\n",#r, r)
2913 #define pr_int(fp, i) fprintf((fp), "%s: %d\n",#i, i)
2914 #define pr_bool(fp, b) fprintf((fp), "%s: %s\n",#b, bool_names[b])
2916 void pr_forcerec(FILE *fp, t_forcerec *fr)
2920 pr_real(fp, fr->rlist);
2921 pr_real(fp, fr->rcoulomb);
2922 pr_real(fp, fr->fudgeQQ);
2923 pr_bool(fp, fr->bGrid);
2924 pr_bool(fp, fr->bTwinRange);
2925 /*pr_int(fp,fr->cg0);
2926 pr_int(fp,fr->hcg);*/
2927 for (i = 0; i < fr->nnblists; i++)
2929 pr_int(fp, fr->nblists[i].table_elec_vdw.n);
2931 pr_real(fp, fr->rcoulomb_switch);
2932 pr_real(fp, fr->rcoulomb);
2937 void forcerec_set_excl_load(t_forcerec *fr,
2938 const gmx_localtop_t *top, const t_commrec *cr)
2941 int t, i, j, ntot, n, ntarget;
2943 if (cr != NULL && PARTDECOMP(cr))
2945 /* No OpenMP with particle decomposition */
2953 ind = top->excls.index;
2957 for (i = 0; i < top->excls.nr; i++)
2959 for (j = ind[i]; j < ind[i+1]; j++)
2968 fr->excl_load[0] = 0;
2971 for (t = 1; t <= fr->nthreads; t++)
2973 ntarget = (ntot*t)/fr->nthreads;
2974 while (i < top->excls.nr && n < ntarget)
2976 for (j = ind[i]; j < ind[i+1]; j++)
2985 fr->excl_load[t] = i;