2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
5 * Copyright (c) 2001-2004, The GROMACS development team.
6 * Copyright (c) 2013,2014,2015,2016,2017, by the GROMACS development team, led by
7 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
8 * and including many others, as listed in the AUTHORS file in the
9 * top-level source directory and at http://www.gromacs.org.
11 * GROMACS is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public License
13 * as published by the Free Software Foundation; either version 2.1
14 * of the License, or (at your option) any later version.
16 * GROMACS is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with GROMACS; if not, see
23 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
24 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
26 * If you want to redistribute modifications to GROMACS, please
27 * consider that scientific software is very special. Version
28 * control is crucial - bugs must be traceable. We will be happy to
29 * consider code for inclusion in the official distribution, but
30 * derived work must not be called official GROMACS. Details are found
31 * in the README & COPYING files - if they are missing, get the
32 * official version at http://www.gromacs.org.
34 * To help us fund GROMACS development, we humbly ask that you cite
35 * the research papers on the package. Check out http://www.gromacs.org.
51 #include "gromacs/commandline/filenm.h"
52 #include "gromacs/domdec/domdec.h"
53 #include "gromacs/domdec/domdec_struct.h"
54 #include "gromacs/ewald/ewald.h"
55 #include "gromacs/fileio/filetypes.h"
56 #include "gromacs/gmxlib/network.h"
57 #include "gromacs/gmxlib/nonbonded/nonbonded.h"
58 #include "gromacs/gpu_utils/gpu_utils.h"
59 #include "gromacs/hardware/detecthardware.h"
60 #include "gromacs/listed-forces/manage-threading.h"
61 #include "gromacs/listed-forces/pairs.h"
62 #include "gromacs/math/calculate-ewald-splitting-coefficient.h"
63 #include "gromacs/math/functions.h"
64 #include "gromacs/math/units.h"
65 #include "gromacs/math/utilities.h"
66 #include "gromacs/math/vec.h"
67 #include "gromacs/mdlib/force.h"
68 #include "gromacs/mdlib/forcerec-threading.h"
69 #include "gromacs/mdlib/gmx_omp_nthreads.h"
70 #include "gromacs/mdlib/md_support.h"
71 #include "gromacs/mdlib/nb_verlet.h"
72 #include "gromacs/mdlib/nbnxn_atomdata.h"
73 #include "gromacs/mdlib/nbnxn_gpu_data_mgmt.h"
74 #include "gromacs/mdlib/nbnxn_search.h"
75 #include "gromacs/mdlib/nbnxn_simd.h"
76 #include "gromacs/mdlib/nbnxn_util.h"
77 #include "gromacs/mdlib/ns.h"
78 #include "gromacs/mdlib/qmmm.h"
79 #include "gromacs/mdlib/sim_util.h"
80 #include "gromacs/mdtypes/commrec.h"
81 #include "gromacs/mdtypes/fcdata.h"
82 #include "gromacs/mdtypes/group.h"
83 #include "gromacs/mdtypes/iforceprovider.h"
84 #include "gromacs/mdtypes/inputrec.h"
85 #include "gromacs/mdtypes/md_enums.h"
86 #include "gromacs/pbcutil/ishift.h"
87 #include "gromacs/pbcutil/pbc.h"
88 #include "gromacs/simd/simd.h"
89 #include "gromacs/tables/forcetable.h"
90 #include "gromacs/topology/mtop_util.h"
91 #include "gromacs/trajectory/trajectoryframe.h"
92 #include "gromacs/utility/cstringutil.h"
93 #include "gromacs/utility/exceptions.h"
94 #include "gromacs/utility/fatalerror.h"
95 #include "gromacs/utility/gmxassert.h"
96 #include "gromacs/utility/logger.h"
97 #include "gromacs/utility/pleasecite.h"
98 #include "gromacs/utility/smalloc.h"
99 #include "gromacs/utility/strconvert.h"
101 #include "nbnxn_gpu_jit_support.h"
102 #include "nbnxn_tuning.h"
104 const char *egrp_nm[egNR+1] = {
105 "Coul-SR", "LJ-SR", "Buck-SR",
106 "Coul-14", "LJ-14", nullptr
109 t_forcerec *mk_forcerec(void)
119 static void pr_nbfp(FILE *fp, real *nbfp, gmx_bool bBHAM, int atnr)
123 for (i = 0; (i < atnr); i++)
125 for (j = 0; (j < atnr); j++)
127 fprintf(fp, "%2d - %2d", i, j);
130 fprintf(fp, " a=%10g, b=%10g, c=%10g\n", BHAMA(nbfp, atnr, i, j),
131 BHAMB(nbfp, atnr, i, j), BHAMC(nbfp, atnr, i, j)/6.0);
135 fprintf(fp, " c6=%10g, c12=%10g\n", C6(nbfp, atnr, i, j)/6.0,
136 C12(nbfp, atnr, i, j)/12.0);
143 static real *mk_nbfp(const gmx_ffparams_t *idef, gmx_bool bBHAM)
151 snew(nbfp, 3*atnr*atnr);
152 for (i = k = 0; (i < atnr); i++)
154 for (j = 0; (j < atnr); j++, k++)
156 BHAMA(nbfp, atnr, i, j) = idef->iparams[k].bham.a;
157 BHAMB(nbfp, atnr, i, j) = idef->iparams[k].bham.b;
158 /* nbfp now includes the 6.0 derivative prefactor */
159 BHAMC(nbfp, atnr, i, j) = idef->iparams[k].bham.c*6.0;
165 snew(nbfp, 2*atnr*atnr);
166 for (i = k = 0; (i < atnr); i++)
168 for (j = 0; (j < atnr); j++, k++)
170 /* nbfp now includes the 6.0/12.0 derivative prefactors */
171 C6(nbfp, atnr, i, j) = idef->iparams[k].lj.c6*6.0;
172 C12(nbfp, atnr, i, j) = idef->iparams[k].lj.c12*12.0;
180 static real *make_ljpme_c6grid(const gmx_ffparams_t *idef, t_forcerec *fr)
183 real c6, c6i, c6j, c12i, c12j, epsi, epsj, sigmai, sigmaj;
186 /* For LJ-PME simulations, we correct the energies with the reciprocal space
187 * inside of the cut-off. To do this the non-bonded kernels needs to have
188 * access to the C6-values used on the reciprocal grid in pme.c
192 snew(grid, 2*atnr*atnr);
193 for (i = k = 0; (i < atnr); i++)
195 for (j = 0; (j < atnr); j++, k++)
197 c6i = idef->iparams[i*(atnr+1)].lj.c6;
198 c12i = idef->iparams[i*(atnr+1)].lj.c12;
199 c6j = idef->iparams[j*(atnr+1)].lj.c6;
200 c12j = idef->iparams[j*(atnr+1)].lj.c12;
201 c6 = std::sqrt(c6i * c6j);
202 if (fr->ljpme_combination_rule == eljpmeLB
203 && !gmx_numzero(c6) && !gmx_numzero(c12i) && !gmx_numzero(c12j))
205 sigmai = gmx::sixthroot(c12i / c6i);
206 sigmaj = gmx::sixthroot(c12j / c6j);
207 epsi = c6i * c6i / c12i;
208 epsj = c6j * c6j / c12j;
209 c6 = std::sqrt(epsi * epsj) * gmx::power6(0.5*(sigmai+sigmaj));
211 /* Store the elements at the same relative positions as C6 in nbfp in order
212 * to simplify access in the kernels
214 grid[2*(atnr*i+j)] = c6*6.0;
220 static real *mk_nbfp_combination_rule(const gmx_ffparams_t *idef, int comb_rule)
224 real c6i, c6j, c12i, c12j, epsi, epsj, sigmai, sigmaj;
228 snew(nbfp, 2*atnr*atnr);
229 for (i = 0; i < atnr; ++i)
231 for (j = 0; j < atnr; ++j)
233 c6i = idef->iparams[i*(atnr+1)].lj.c6;
234 c12i = idef->iparams[i*(atnr+1)].lj.c12;
235 c6j = idef->iparams[j*(atnr+1)].lj.c6;
236 c12j = idef->iparams[j*(atnr+1)].lj.c12;
237 c6 = std::sqrt(c6i * c6j);
238 c12 = std::sqrt(c12i * c12j);
239 if (comb_rule == eCOMB_ARITHMETIC
240 && !gmx_numzero(c6) && !gmx_numzero(c12))
242 sigmai = gmx::sixthroot(c12i / c6i);
243 sigmaj = gmx::sixthroot(c12j / c6j);
244 epsi = c6i * c6i / c12i;
245 epsj = c6j * c6j / c12j;
246 c6 = std::sqrt(epsi * epsj) * gmx::power6(0.5*(sigmai+sigmaj));
247 c12 = std::sqrt(epsi * epsj) * gmx::power12(0.5*(sigmai+sigmaj));
249 C6(nbfp, atnr, i, j) = c6*6.0;
250 C12(nbfp, atnr, i, j) = c12*12.0;
256 /* This routine sets fr->solvent_opt to the most common solvent in the
257 * system, e.g. esolSPC or esolTIP4P. It will also mark each charge group in
258 * the fr->solvent_type array with the correct type (or esolNO).
260 * Charge groups that fulfill the conditions but are not identical to the
261 * most common one will be marked as esolNO in the solvent_type array.
263 * TIP3p is identical to SPC for these purposes, so we call it
264 * SPC in the arrays (Apologies to Bill Jorgensen ;-)
266 * NOTE: QM particle should not
267 * become an optimized solvent. Not even if there is only one charge
277 } solvent_parameters_t;
280 check_solvent_cg(const gmx_moltype_t *molt,
283 const unsigned char *qm_grpnr,
284 const t_grps *qm_grps,
286 int *n_solvent_parameters,
287 solvent_parameters_t **solvent_parameters_p,
297 real tmp_charge[4] = { 0.0 }; /* init to zero to make gcc4.8 happy */
298 int tmp_vdwtype[4] = { 0 }; /* init to zero to make gcc4.8 happy */
301 solvent_parameters_t *solvent_parameters;
303 /* We use a list with parameters for each solvent type.
304 * Every time we discover a new molecule that fulfills the basic
305 * conditions for a solvent we compare with the previous entries
306 * in these lists. If the parameters are the same we just increment
307 * the counter for that type, and otherwise we create a new type
308 * based on the current molecule.
310 * Once we've finished going through all molecules we check which
311 * solvent is most common, and mark all those molecules while we
312 * clear the flag on all others.
315 solvent_parameters = *solvent_parameters_p;
317 /* Mark the cg first as non optimized */
320 /* Check if this cg has no exclusions with atoms in other charge groups
321 * and all atoms inside the charge group excluded.
322 * We only have 3 or 4 atom solvent loops.
324 if (GET_CGINFO_EXCL_INTER(cginfo) ||
325 !GET_CGINFO_EXCL_INTRA(cginfo))
330 /* Get the indices of the first atom in this charge group */
331 j0 = molt->cgs.index[cg0];
332 j1 = molt->cgs.index[cg0+1];
334 /* Number of atoms in our molecule */
340 "Moltype '%s': there are %d atoms in this charge group\n",
344 /* Check if it could be an SPC (3 atoms) or TIP4p (4) water,
347 if (nj < 3 || nj > 4)
352 /* Check if we are doing QM on this group */
354 if (qm_grpnr != nullptr)
356 for (j = j0; j < j1 && !qm; j++)
358 qm = (qm_grpnr[j] < qm_grps->nr - 1);
361 /* Cannot use solvent optimization with QM */
367 atom = molt->atoms.atom;
369 /* Still looks like a solvent, time to check parameters */
371 /* If it is perturbed (free energy) we can't use the solvent loops,
372 * so then we just skip to the next molecule.
376 for (j = j0; j < j1 && !perturbed; j++)
378 perturbed = PERTURBED(atom[j]);
386 /* Now it's only a question if the VdW and charge parameters
387 * are OK. Before doing the check we compare and see if they are
388 * identical to a possible previous solvent type.
389 * First we assign the current types and charges.
391 for (j = 0; j < nj; j++)
393 tmp_vdwtype[j] = atom[j0+j].type;
394 tmp_charge[j] = atom[j0+j].q;
397 /* Does it match any previous solvent type? */
398 for (k = 0; k < *n_solvent_parameters; k++)
403 /* We can only match SPC with 3 atoms and TIP4p with 4 atoms */
404 if ( (solvent_parameters[k].model == esolSPC && nj != 3) ||
405 (solvent_parameters[k].model == esolTIP4P && nj != 4) )
410 /* Check that types & charges match for all atoms in molecule */
411 for (j = 0; j < nj && match == TRUE; j++)
413 if (tmp_vdwtype[j] != solvent_parameters[k].vdwtype[j])
417 if (tmp_charge[j] != solvent_parameters[k].charge[j])
424 /* Congratulations! We have a matched solvent.
425 * Flag it with this type for later processing.
428 solvent_parameters[k].count += nmol;
430 /* We are done with this charge group */
435 /* If we get here, we have a tentative new solvent type.
436 * Before we add it we must check that it fulfills the requirements
437 * of the solvent optimized loops. First determine which atoms have
440 for (j = 0; j < nj; j++)
443 tjA = tmp_vdwtype[j];
445 /* Go through all other tpes and see if any have non-zero
446 * VdW parameters when combined with this one.
448 for (k = 0; k < fr->ntype && (has_vdw[j] == FALSE); k++)
450 /* We already checked that the atoms weren't perturbed,
451 * so we only need to check state A now.
455 has_vdw[j] = (has_vdw[j] ||
456 (BHAMA(fr->nbfp, fr->ntype, tjA, k) != 0.0) ||
457 (BHAMB(fr->nbfp, fr->ntype, tjA, k) != 0.0) ||
458 (BHAMC(fr->nbfp, fr->ntype, tjA, k) != 0.0));
463 has_vdw[j] = (has_vdw[j] ||
464 (C6(fr->nbfp, fr->ntype, tjA, k) != 0.0) ||
465 (C12(fr->nbfp, fr->ntype, tjA, k) != 0.0));
470 /* Now we know all we need to make the final check and assignment. */
474 * For this we require thatn all atoms have charge,
475 * the charges on atom 2 & 3 should be the same, and only
476 * atom 1 might have VdW.
478 if (has_vdw[1] == FALSE &&
479 has_vdw[2] == FALSE &&
480 tmp_charge[0] != 0 &&
481 tmp_charge[1] != 0 &&
482 tmp_charge[2] == tmp_charge[1])
484 srenew(solvent_parameters, *n_solvent_parameters+1);
485 solvent_parameters[*n_solvent_parameters].model = esolSPC;
486 solvent_parameters[*n_solvent_parameters].count = nmol;
487 for (k = 0; k < 3; k++)
489 solvent_parameters[*n_solvent_parameters].vdwtype[k] = tmp_vdwtype[k];
490 solvent_parameters[*n_solvent_parameters].charge[k] = tmp_charge[k];
493 *cg_sp = *n_solvent_parameters;
494 (*n_solvent_parameters)++;
499 /* Or could it be a TIP4P?
500 * For this we require thatn atoms 2,3,4 have charge, but not atom 1.
501 * Only atom 1 mght have VdW.
503 if (has_vdw[1] == FALSE &&
504 has_vdw[2] == FALSE &&
505 has_vdw[3] == FALSE &&
506 tmp_charge[0] == 0 &&
507 tmp_charge[1] != 0 &&
508 tmp_charge[2] == tmp_charge[1] &&
511 srenew(solvent_parameters, *n_solvent_parameters+1);
512 solvent_parameters[*n_solvent_parameters].model = esolTIP4P;
513 solvent_parameters[*n_solvent_parameters].count = nmol;
514 for (k = 0; k < 4; k++)
516 solvent_parameters[*n_solvent_parameters].vdwtype[k] = tmp_vdwtype[k];
517 solvent_parameters[*n_solvent_parameters].charge[k] = tmp_charge[k];
520 *cg_sp = *n_solvent_parameters;
521 (*n_solvent_parameters)++;
525 *solvent_parameters_p = solvent_parameters;
529 check_solvent(FILE * fp,
530 const gmx_mtop_t * mtop,
532 cginfo_mb_t *cginfo_mb)
535 const gmx_moltype_t *molt;
536 int mb, mol, cg_mol, at_offset, am, cgm, i, nmol_ch, nmol;
537 int n_solvent_parameters;
538 solvent_parameters_t *solvent_parameters;
544 fprintf(debug, "Going to determine what solvent types we have.\n");
547 n_solvent_parameters = 0;
548 solvent_parameters = nullptr;
549 /* Allocate temporary array for solvent type */
550 snew(cg_sp, mtop->nmolblock);
553 for (mb = 0; mb < mtop->nmolblock; mb++)
555 molt = &mtop->moltype[mtop->molblock[mb].type];
557 /* Here we have to loop over all individual molecules
558 * because we need to check for QMMM particles.
560 snew(cg_sp[mb], cginfo_mb[mb].cg_mod);
561 nmol_ch = cginfo_mb[mb].cg_mod/cgs->nr;
562 nmol = mtop->molblock[mb].nmol/nmol_ch;
563 for (mol = 0; mol < nmol_ch; mol++)
566 am = mol*cgs->index[cgs->nr];
567 for (cg_mol = 0; cg_mol < cgs->nr; cg_mol++)
569 check_solvent_cg(molt, cg_mol, nmol,
570 mtop->groups.grpnr[egcQMMM] ?
571 mtop->groups.grpnr[egcQMMM]+at_offset+am : nullptr,
572 &mtop->groups.grps[egcQMMM],
574 &n_solvent_parameters, &solvent_parameters,
575 cginfo_mb[mb].cginfo[cgm+cg_mol],
576 &cg_sp[mb][cgm+cg_mol]);
579 at_offset += cgs->index[cgs->nr];
582 /* Puh! We finished going through all charge groups.
583 * Now find the most common solvent model.
586 /* Most common solvent this far */
588 for (i = 0; i < n_solvent_parameters; i++)
591 solvent_parameters[i].count > solvent_parameters[bestsp].count)
599 bestsol = solvent_parameters[bestsp].model;
607 for (mb = 0; mb < mtop->nmolblock; mb++)
609 cgs = &mtop->moltype[mtop->molblock[mb].type].cgs;
610 nmol = (mtop->molblock[mb].nmol*cgs->nr)/cginfo_mb[mb].cg_mod;
611 for (i = 0; i < cginfo_mb[mb].cg_mod; i++)
613 if (cg_sp[mb][i] == bestsp)
615 SET_CGINFO_SOLOPT(cginfo_mb[mb].cginfo[i], bestsol);
620 SET_CGINFO_SOLOPT(cginfo_mb[mb].cginfo[i], esolNO);
627 if (bestsol != esolNO && fp != nullptr)
629 fprintf(fp, "\nEnabling %s-like water optimization for %d molecules.\n\n",
631 solvent_parameters[bestsp].count);
634 sfree(solvent_parameters);
635 fr->solvent_opt = bestsol;
639 acNONE = 0, acCONSTRAINT, acSETTLE
642 static cginfo_mb_t *init_cginfo_mb(FILE *fplog, const gmx_mtop_t *mtop,
643 t_forcerec *fr, gmx_bool bNoSolvOpt,
644 gmx_bool *bFEP_NonBonded,
645 gmx_bool *bExcl_IntraCGAll_InterCGNone)
648 const t_blocka *excl;
649 const gmx_moltype_t *molt;
650 const gmx_molblock_t *molb;
651 cginfo_mb_t *cginfo_mb;
654 int cg_offset, a_offset;
655 int mb, m, cg, a0, a1, gid, ai, j, aj, excl_nalloc;
659 gmx_bool bId, *bExcl, bExclIntraAll, bExclInter, bHaveVDW, bHaveQ, bHavePerturbedAtoms;
661 snew(cginfo_mb, mtop->nmolblock);
663 snew(type_VDW, fr->ntype);
664 for (ai = 0; ai < fr->ntype; ai++)
666 type_VDW[ai] = FALSE;
667 for (j = 0; j < fr->ntype; j++)
669 type_VDW[ai] = type_VDW[ai] ||
671 C6(fr->nbfp, fr->ntype, ai, j) != 0 ||
672 C12(fr->nbfp, fr->ntype, ai, j) != 0;
676 *bFEP_NonBonded = FALSE;
677 *bExcl_IntraCGAll_InterCGNone = TRUE;
680 snew(bExcl, excl_nalloc);
683 for (mb = 0; mb < mtop->nmolblock; mb++)
685 molb = &mtop->molblock[mb];
686 molt = &mtop->moltype[molb->type];
690 /* Check if the cginfo is identical for all molecules in this block.
691 * If so, we only need an array of the size of one molecule.
692 * Otherwise we make an array of #mol times #cgs per molecule.
695 for (m = 0; m < molb->nmol; m++)
697 int am = m*cgs->index[cgs->nr];
698 for (cg = 0; cg < cgs->nr; cg++)
701 a1 = cgs->index[cg+1];
702 if (ggrpnr(&mtop->groups, egcENER, a_offset+am+a0) !=
703 ggrpnr(&mtop->groups, egcENER, a_offset +a0))
707 if (mtop->groups.grpnr[egcQMMM] != nullptr)
709 for (ai = a0; ai < a1; ai++)
711 if (mtop->groups.grpnr[egcQMMM][a_offset+am+ai] !=
712 mtop->groups.grpnr[egcQMMM][a_offset +ai])
721 cginfo_mb[mb].cg_start = cg_offset;
722 cginfo_mb[mb].cg_end = cg_offset + molb->nmol*cgs->nr;
723 cginfo_mb[mb].cg_mod = (bId ? 1 : molb->nmol)*cgs->nr;
724 snew(cginfo_mb[mb].cginfo, cginfo_mb[mb].cg_mod);
725 cginfo = cginfo_mb[mb].cginfo;
727 /* Set constraints flags for constrained atoms */
728 snew(a_con, molt->atoms.nr);
729 for (ftype = 0; ftype < F_NRE; ftype++)
731 if (interaction_function[ftype].flags & IF_CONSTRAINT)
736 for (ia = 0; ia < molt->ilist[ftype].nr; ia += 1+nral)
740 for (a = 0; a < nral; a++)
742 a_con[molt->ilist[ftype].iatoms[ia+1+a]] =
743 (ftype == F_SETTLE ? acSETTLE : acCONSTRAINT);
749 for (m = 0; m < (bId ? 1 : molb->nmol); m++)
752 int am = m*cgs->index[cgs->nr];
753 for (cg = 0; cg < cgs->nr; cg++)
756 a1 = cgs->index[cg+1];
758 /* Store the energy group in cginfo */
759 gid = ggrpnr(&mtop->groups, egcENER, a_offset+am+a0);
760 SET_CGINFO_GID(cginfo[cgm+cg], gid);
762 /* Check the intra/inter charge group exclusions */
763 if (a1-a0 > excl_nalloc)
765 excl_nalloc = a1 - a0;
766 srenew(bExcl, excl_nalloc);
768 /* bExclIntraAll: all intra cg interactions excluded
769 * bExclInter: any inter cg interactions excluded
771 bExclIntraAll = TRUE;
775 bHavePerturbedAtoms = FALSE;
776 for (ai = a0; ai < a1; ai++)
778 /* Check VDW and electrostatic interactions */
779 bHaveVDW = bHaveVDW || (type_VDW[molt->atoms.atom[ai].type] ||
780 type_VDW[molt->atoms.atom[ai].typeB]);
781 bHaveQ = bHaveQ || (molt->atoms.atom[ai].q != 0 ||
782 molt->atoms.atom[ai].qB != 0);
784 bHavePerturbedAtoms = bHavePerturbedAtoms || (PERTURBED(molt->atoms.atom[ai]) != 0);
786 /* Clear the exclusion list for atom ai */
787 for (aj = a0; aj < a1; aj++)
789 bExcl[aj-a0] = FALSE;
791 /* Loop over all the exclusions of atom ai */
792 for (j = excl->index[ai]; j < excl->index[ai+1]; j++)
795 if (aj < a0 || aj >= a1)
804 /* Check if ai excludes a0 to a1 */
805 for (aj = a0; aj < a1; aj++)
809 bExclIntraAll = FALSE;
816 SET_CGINFO_CONSTR(cginfo[cgm+cg]);
819 SET_CGINFO_SETTLE(cginfo[cgm+cg]);
827 SET_CGINFO_EXCL_INTRA(cginfo[cgm+cg]);
831 SET_CGINFO_EXCL_INTER(cginfo[cgm+cg]);
833 if (a1 - a0 > MAX_CHARGEGROUP_SIZE)
835 /* The size in cginfo is currently only read with DD */
836 gmx_fatal(FARGS, "A charge group has size %d which is larger than the limit of %d atoms", a1-a0, MAX_CHARGEGROUP_SIZE);
840 SET_CGINFO_HAS_VDW(cginfo[cgm+cg]);
844 SET_CGINFO_HAS_Q(cginfo[cgm+cg]);
846 if (bHavePerturbedAtoms && fr->efep != efepNO)
848 SET_CGINFO_FEP(cginfo[cgm+cg]);
849 *bFEP_NonBonded = TRUE;
851 /* Store the charge group size */
852 SET_CGINFO_NATOMS(cginfo[cgm+cg], a1-a0);
854 if (!bExclIntraAll || bExclInter)
856 *bExcl_IntraCGAll_InterCGNone = FALSE;
863 cg_offset += molb->nmol*cgs->nr;
864 a_offset += molb->nmol*cgs->index[cgs->nr];
868 /* the solvent optimizer is called after the QM is initialized,
869 * because we don't want to have the QM subsystemto become an
873 check_solvent(fplog, mtop, fr, cginfo_mb);
875 if (getenv("GMX_NO_SOLV_OPT"))
879 fprintf(fplog, "Found environment variable GMX_NO_SOLV_OPT.\n"
880 "Disabling all solvent optimization\n");
882 fr->solvent_opt = esolNO;
886 fr->solvent_opt = esolNO;
888 if (!fr->solvent_opt)
890 for (mb = 0; mb < mtop->nmolblock; mb++)
892 for (cg = 0; cg < cginfo_mb[mb].cg_mod; cg++)
894 SET_CGINFO_SOLOPT(cginfo_mb[mb].cginfo[cg], esolNO);
902 static int *cginfo_expand(int nmb, cginfo_mb_t *cgi_mb)
907 ncg = cgi_mb[nmb-1].cg_end;
910 for (cg = 0; cg < ncg; cg++)
912 while (cg >= cgi_mb[mb].cg_end)
917 cgi_mb[mb].cginfo[(cg - cgi_mb[mb].cg_start) % cgi_mb[mb].cg_mod];
923 static void set_chargesum(FILE *log, t_forcerec *fr, const gmx_mtop_t *mtop)
925 /*This now calculates sum for q and c6*/
926 double qsum, q2sum, q, c6sum, c6;
928 const t_atoms *atoms;
933 for (mb = 0; mb < mtop->nmolblock; mb++)
935 nmol = mtop->molblock[mb].nmol;
936 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
937 for (i = 0; i < atoms->nr; i++)
939 q = atoms->atom[i].q;
942 c6 = mtop->ffparams.iparams[atoms->atom[i].type*(mtop->ffparams.atnr+1)].lj.c6;
947 fr->q2sum[0] = q2sum;
948 fr->c6sum[0] = c6sum;
950 if (fr->efep != efepNO)
955 for (mb = 0; mb < mtop->nmolblock; mb++)
957 nmol = mtop->molblock[mb].nmol;
958 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
959 for (i = 0; i < atoms->nr; i++)
961 q = atoms->atom[i].qB;
964 c6 = mtop->ffparams.iparams[atoms->atom[i].typeB*(mtop->ffparams.atnr+1)].lj.c6;
968 fr->q2sum[1] = q2sum;
969 fr->c6sum[1] = c6sum;
974 fr->qsum[1] = fr->qsum[0];
975 fr->q2sum[1] = fr->q2sum[0];
976 fr->c6sum[1] = fr->c6sum[0];
980 if (fr->efep == efepNO)
982 fprintf(log, "System total charge: %.3f\n", fr->qsum[0]);
986 fprintf(log, "System total charge, top. A: %.3f top. B: %.3f\n",
987 fr->qsum[0], fr->qsum[1]);
992 void update_forcerec(t_forcerec *fr, matrix box)
994 if (fr->eeltype == eelGRF)
996 calc_rffac(nullptr, fr->eeltype, fr->epsilon_r, fr->epsilon_rf,
997 fr->rcoulomb, fr->temp, fr->zsquare, box,
998 &fr->kappa, &fr->k_rf, &fr->c_rf);
1002 void set_avcsixtwelve(FILE *fplog, t_forcerec *fr, const gmx_mtop_t *mtop)
1004 const t_atoms *atoms, *atoms_tpi;
1005 const t_blocka *excl;
1006 int mb, nmol, nmolc, i, j, tpi, tpj, j1, j2, k, nexcl, q;
1007 gmx_int64_t npair, npair_ij, tmpi, tmpj;
1008 double csix, ctwelve;
1009 int ntp, *typecount;
1012 real *nbfp_comb = nullptr;
1018 /* For LJ-PME, we want to correct for the difference between the
1019 * actual C6 values and the C6 values used by the LJ-PME based on
1020 * combination rules. */
1022 if (EVDW_PME(fr->vdwtype))
1024 nbfp_comb = mk_nbfp_combination_rule(&mtop->ffparams,
1025 (fr->ljpme_combination_rule == eljpmeLB) ? eCOMB_ARITHMETIC : eCOMB_GEOMETRIC);
1026 for (tpi = 0; tpi < ntp; ++tpi)
1028 for (tpj = 0; tpj < ntp; ++tpj)
1030 C6(nbfp_comb, ntp, tpi, tpj) =
1031 C6(nbfp, ntp, tpi, tpj) - C6(nbfp_comb, ntp, tpi, tpj);
1032 C12(nbfp_comb, ntp, tpi, tpj) = C12(nbfp, ntp, tpi, tpj);
1037 for (q = 0; q < (fr->efep == efepNO ? 1 : 2); q++)
1045 /* Count the types so we avoid natoms^2 operations */
1046 snew(typecount, ntp);
1047 gmx_mtop_count_atomtypes(mtop, q, typecount);
1049 for (tpi = 0; tpi < ntp; tpi++)
1051 for (tpj = tpi; tpj < ntp; tpj++)
1053 tmpi = typecount[tpi];
1054 tmpj = typecount[tpj];
1057 npair_ij = tmpi*tmpj;
1061 npair_ij = tmpi*(tmpi - 1)/2;
1065 /* nbfp now includes the 6.0 derivative prefactor */
1066 csix += npair_ij*BHAMC(nbfp, ntp, tpi, tpj)/6.0;
1070 /* nbfp now includes the 6.0/12.0 derivative prefactors */
1071 csix += npair_ij* C6(nbfp, ntp, tpi, tpj)/6.0;
1072 ctwelve += npair_ij* C12(nbfp, ntp, tpi, tpj)/12.0;
1078 /* Subtract the excluded pairs.
1079 * The main reason for substracting exclusions is that in some cases
1080 * some combinations might never occur and the parameters could have
1081 * any value. These unused values should not influence the dispersion
1084 for (mb = 0; mb < mtop->nmolblock; mb++)
1086 nmol = mtop->molblock[mb].nmol;
1087 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
1088 excl = &mtop->moltype[mtop->molblock[mb].type].excls;
1089 for (i = 0; (i < atoms->nr); i++)
1093 tpi = atoms->atom[i].type;
1097 tpi = atoms->atom[i].typeB;
1099 j1 = excl->index[i];
1100 j2 = excl->index[i+1];
1101 for (j = j1; j < j2; j++)
1108 tpj = atoms->atom[k].type;
1112 tpj = atoms->atom[k].typeB;
1116 /* nbfp now includes the 6.0 derivative prefactor */
1117 csix -= nmol*BHAMC(nbfp, ntp, tpi, tpj)/6.0;
1121 /* nbfp now includes the 6.0/12.0 derivative prefactors */
1122 csix -= nmol*C6 (nbfp, ntp, tpi, tpj)/6.0;
1123 ctwelve -= nmol*C12(nbfp, ntp, tpi, tpj)/12.0;
1133 /* Only correct for the interaction of the test particle
1134 * with the rest of the system.
1137 &mtop->moltype[mtop->molblock[mtop->nmolblock-1].type].atoms;
1140 for (mb = 0; mb < mtop->nmolblock; mb++)
1142 nmol = mtop->molblock[mb].nmol;
1143 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
1144 for (j = 0; j < atoms->nr; j++)
1147 /* Remove the interaction of the test charge group
1150 if (mb == mtop->nmolblock-1)
1154 if (mb == 0 && nmol == 1)
1156 gmx_fatal(FARGS, "Old format tpr with TPI, please generate a new tpr file");
1161 tpj = atoms->atom[j].type;
1165 tpj = atoms->atom[j].typeB;
1167 for (i = 0; i < fr->n_tpi; i++)
1171 tpi = atoms_tpi->atom[i].type;
1175 tpi = atoms_tpi->atom[i].typeB;
1179 /* nbfp now includes the 6.0 derivative prefactor */
1180 csix += nmolc*BHAMC(nbfp, ntp, tpi, tpj)/6.0;
1184 /* nbfp now includes the 6.0/12.0 derivative prefactors */
1185 csix += nmolc*C6 (nbfp, ntp, tpi, tpj)/6.0;
1186 ctwelve += nmolc*C12(nbfp, ntp, tpi, tpj)/12.0;
1193 if (npair - nexcl <= 0 && fplog)
1195 fprintf(fplog, "\nWARNING: There are no atom pairs for dispersion correction\n\n");
1201 csix /= npair - nexcl;
1202 ctwelve /= npair - nexcl;
1206 fprintf(debug, "Counted %d exclusions\n", nexcl);
1207 fprintf(debug, "Average C6 parameter is: %10g\n", (double)csix);
1208 fprintf(debug, "Average C12 parameter is: %10g\n", (double)ctwelve);
1210 fr->avcsix[q] = csix;
1211 fr->avctwelve[q] = ctwelve;
1214 if (EVDW_PME(fr->vdwtype))
1219 if (fplog != nullptr)
1221 if (fr->eDispCorr == edispcAllEner ||
1222 fr->eDispCorr == edispcAllEnerPres)
1224 fprintf(fplog, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
1225 fr->avcsix[0], fr->avctwelve[0]);
1229 fprintf(fplog, "Long Range LJ corr.: <C6> %10.4e\n", fr->avcsix[0]);
1235 static void set_bham_b_max(FILE *fplog, t_forcerec *fr,
1236 const gmx_mtop_t *mtop)
1238 const t_atoms *at1, *at2;
1239 int mt1, mt2, i, j, tpi, tpj, ntypes;
1245 fprintf(fplog, "Determining largest Buckingham b parameter for table\n");
1252 for (mt1 = 0; mt1 < mtop->nmoltype; mt1++)
1254 at1 = &mtop->moltype[mt1].atoms;
1255 for (i = 0; (i < at1->nr); i++)
1257 tpi = at1->atom[i].type;
1260 gmx_fatal(FARGS, "Atomtype[%d] = %d, maximum = %d", i, tpi, ntypes);
1263 for (mt2 = mt1; mt2 < mtop->nmoltype; mt2++)
1265 at2 = &mtop->moltype[mt2].atoms;
1266 for (j = 0; (j < at2->nr); j++)
1268 tpj = at2->atom[j].type;
1271 gmx_fatal(FARGS, "Atomtype[%d] = %d, maximum = %d", j, tpj, ntypes);
1273 b = BHAMB(nbfp, ntypes, tpi, tpj);
1274 if (b > fr->bham_b_max)
1278 if ((b < bmin) || (bmin == -1))
1288 fprintf(fplog, "Buckingham b parameters, min: %g, max: %g\n",
1289 bmin, fr->bham_b_max);
1293 static void make_nbf_tables(FILE *fp,
1294 t_forcerec *fr, real rtab,
1295 const char *tabfn, char *eg1, char *eg2,
1301 if (tabfn == nullptr)
1305 fprintf(debug, "No table file name passed, can not read table, can not do non-bonded interactions\n");
1310 sprintf(buf, "%s", tabfn);
1313 /* Append the two energy group names */
1314 sprintf(buf + strlen(tabfn) - strlen(ftp2ext(efXVG)) - 1, "_%s_%s.%s",
1315 eg1, eg2, ftp2ext(efXVG));
1317 nbl->table_elec_vdw = make_tables(fp, fr, buf, rtab, 0);
1318 /* Copy the contents of the table to separate coulomb and LJ tables too,
1319 * to improve cache performance.
1321 /* For performance reasons we want
1322 * the table data to be aligned to 16-byte. The pointers could be freed
1323 * but currently aren't.
1325 snew(nbl->table_elec, 1);
1326 nbl->table_elec->interaction = GMX_TABLE_INTERACTION_ELEC;
1327 nbl->table_elec->format = nbl->table_elec_vdw->format;
1328 nbl->table_elec->r = nbl->table_elec_vdw->r;
1329 nbl->table_elec->n = nbl->table_elec_vdw->n;
1330 nbl->table_elec->scale = nbl->table_elec_vdw->scale;
1331 nbl->table_elec->formatsize = nbl->table_elec_vdw->formatsize;
1332 nbl->table_elec->ninteractions = 1;
1333 nbl->table_elec->stride = nbl->table_elec->formatsize * nbl->table_elec->ninteractions;
1334 snew_aligned(nbl->table_elec->data, nbl->table_elec->stride*(nbl->table_elec->n+1), 32);
1336 snew(nbl->table_vdw, 1);
1337 nbl->table_vdw->interaction = GMX_TABLE_INTERACTION_VDWREP_VDWDISP;
1338 nbl->table_vdw->format = nbl->table_elec_vdw->format;
1339 nbl->table_vdw->r = nbl->table_elec_vdw->r;
1340 nbl->table_vdw->n = nbl->table_elec_vdw->n;
1341 nbl->table_vdw->scale = nbl->table_elec_vdw->scale;
1342 nbl->table_vdw->formatsize = nbl->table_elec_vdw->formatsize;
1343 nbl->table_vdw->ninteractions = 2;
1344 nbl->table_vdw->stride = nbl->table_vdw->formatsize * nbl->table_vdw->ninteractions;
1345 snew_aligned(nbl->table_vdw->data, nbl->table_vdw->stride*(nbl->table_vdw->n+1), 32);
1347 for (i = 0; i <= nbl->table_elec_vdw->n; i++)
1349 for (j = 0; j < 4; j++)
1351 nbl->table_elec->data[4*i+j] = nbl->table_elec_vdw->data[12*i+j];
1353 for (j = 0; j < 8; j++)
1355 nbl->table_vdw->data[8*i+j] = nbl->table_elec_vdw->data[12*i+4+j];
1360 /*!\brief If there's bonded interactions of type \c ftype1 or \c
1361 * ftype2 present in the topology, build an array of the number of
1362 * interactions present for each bonded interaction index found in the
1365 * \c ftype1 or \c ftype2 may be set to -1 to disable seeking for a
1366 * valid type with that parameter.
1368 * \c count will be reallocated as necessary to fit the largest bonded
1369 * interaction index found, and its current size will be returned in
1370 * \c ncount. It will contain zero for every bonded interaction index
1371 * for which no interactions are present in the topology.
1373 static void count_tables(int ftype1, int ftype2, const gmx_mtop_t *mtop,
1374 int *ncount, int **count)
1376 const gmx_moltype_t *molt;
1378 int mt, ftype, stride, i, j, tabnr;
1380 // Loop over all moleculetypes
1381 for (mt = 0; mt < mtop->nmoltype; mt++)
1383 molt = &mtop->moltype[mt];
1384 // Loop over all interaction types
1385 for (ftype = 0; ftype < F_NRE; ftype++)
1387 // If the current interaction type is one of the types whose tables we're trying to count...
1388 if (ftype == ftype1 || ftype == ftype2)
1390 il = &molt->ilist[ftype];
1391 stride = 1 + NRAL(ftype);
1392 // ... and there are actually some interactions for this type
1393 for (i = 0; i < il->nr; i += stride)
1395 // Find out which table index the user wanted
1396 tabnr = mtop->ffparams.iparams[il->iatoms[i]].tab.table;
1399 gmx_fatal(FARGS, "A bonded table number is smaller than 0: %d\n", tabnr);
1401 // Make room for this index in the data structure
1402 if (tabnr >= *ncount)
1404 srenew(*count, tabnr+1);
1405 for (j = *ncount; j < tabnr+1; j++)
1411 // Record that this table index is used and must have a valid file
1419 /*!\brief If there's bonded interactions of flavour \c tabext and type
1420 * \c ftype1 or \c ftype2 present in the topology, seek them in the
1421 * list of filenames passed to mdrun, and make bonded tables from
1424 * \c ftype1 or \c ftype2 may be set to -1 to disable seeking for a
1425 * valid type with that parameter.
1427 * A fatal error occurs if no matching filename is found.
1429 static bondedtable_t *make_bonded_tables(FILE *fplog,
1430 int ftype1, int ftype2,
1431 const gmx_mtop_t *mtop,
1432 const t_filenm *tabbfnm,
1442 count_tables(ftype1, ftype2, mtop, &ncount, &count);
1444 // Are there any relevant tabulated bond interactions?
1448 for (int i = 0; i < ncount; i++)
1450 // Do any interactions exist that requires this table?
1453 // This pattern enforces the current requirement that
1454 // table filenames end in a characteristic sequence
1455 // before the file type extension, and avoids table 13
1456 // being recognized and used for table 1.
1457 std::string patternToFind = gmx::formatString("_%s%d.%s", tabext, i, ftp2ext(efXVG));
1458 bool madeTable = false;
1459 for (int j = 0; j < tabbfnm->nfiles && !madeTable; ++j)
1461 std::string filename(tabbfnm->fns[j]);
1462 if (gmx::endsWith(filename, patternToFind))
1464 // Finally read the table from the file found
1465 tab[i] = make_bonded_table(fplog, tabbfnm->fns[j], NRAL(ftype1)-2);
1471 bool isPlural = (ftype2 != -1);
1472 gmx_fatal(FARGS, "Tabulated interaction of type '%s%s%s' with index %d cannot be used because no table file whose name matched '%s' was passed via the gmx mdrun -tableb command-line option.",
1473 interaction_function[ftype1].longname,
1474 isPlural ? "' or '" : "",
1475 isPlural ? interaction_function[ftype2].longname : "",
1477 patternToFind.c_str());
1487 void forcerec_set_ranges(t_forcerec *fr,
1488 int ncg_home, int ncg_force,
1490 int natoms_force_constr, int natoms_f_novirsum)
1495 /* fr->ncg_force is unused in the standard code,
1496 * but it can be useful for modified code dealing with charge groups.
1498 fr->ncg_force = ncg_force;
1499 fr->natoms_force = natoms_force;
1500 fr->natoms_force_constr = natoms_force_constr;
1502 if (fr->natoms_force_constr > fr->nalloc_force)
1504 fr->nalloc_force = over_alloc_dd(fr->natoms_force_constr);
1507 if (fr->bF_NoVirSum)
1509 /* TODO: remove this + 1 when padding is properly implemented */
1510 fr->forceBufferNoVirialSummation->resize(natoms_f_novirsum + 1);
1514 static real cutoff_inf(real cutoff)
1518 cutoff = GMX_CUTOFF_INF;
1524 gmx_bool can_use_allvsall(const t_inputrec *ir, gmx_bool bPrintNote, t_commrec *cr, FILE *fp)
1531 ir->rcoulomb == 0 &&
1533 ir->ePBC == epbcNONE &&
1534 ir->vdwtype == evdwCUT &&
1535 ir->coulombtype == eelCUT &&
1536 ir->efep == efepNO &&
1537 (ir->implicit_solvent == eisNO ||
1538 (ir->implicit_solvent == eisGBSA && (ir->gb_algorithm == egbSTILL ||
1539 ir->gb_algorithm == egbHCT ||
1540 ir->gb_algorithm == egbOBC))) &&
1541 getenv("GMX_NO_ALLVSALL") == nullptr
1544 if (bAllvsAll && ir->opts.ngener > 1)
1546 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";
1552 fprintf(fp, "\n%s\n", note);
1558 if (bAllvsAll && fp && MASTER(cr))
1560 fprintf(fp, "\nUsing SIMD all-vs-all kernels.\n\n");
1567 gmx_bool nbnxn_simd_supported(const gmx::MDLogger &mdlog,
1568 const t_inputrec *ir)
1570 if (ir->vdwtype == evdwPME && ir->ljpme_combination_rule == eljpmeLB)
1572 /* LJ PME with LB combination rule does 7 mesh operations.
1573 * This so slow that we don't compile SIMD non-bonded kernels
1575 GMX_LOG(mdlog.warning).asParagraph().appendText("LJ-PME with Lorentz-Berthelot is not supported with SIMD kernels, falling back to plain C kernels");
1583 static void pick_nbnxn_kernel_cpu(const t_inputrec gmx_unused *ir,
1587 *kernel_type = nbnxnk4x4_PlainC;
1588 *ewald_excl = ewaldexclTable;
1592 #ifdef GMX_NBNXN_SIMD_4XN
1593 *kernel_type = nbnxnk4xN_SIMD_4xN;
1595 #ifdef GMX_NBNXN_SIMD_2XNN
1596 *kernel_type = nbnxnk4xN_SIMD_2xNN;
1599 #if defined GMX_NBNXN_SIMD_2XNN && defined GMX_NBNXN_SIMD_4XN
1600 /* We need to choose if we want 2x(N+N) or 4xN kernels.
1601 * Currently this is based on the SIMD acceleration choice,
1602 * but it might be better to decide this at runtime based on CPU.
1604 * 4xN calculates more (zero) interactions, but has less pair-search
1605 * work and much better kernel instruction scheduling.
1607 * Up till now we have only seen that on Intel Sandy/Ivy Bridge,
1608 * which doesn't have FMA, both the analytical and tabulated Ewald
1609 * kernels have similar pair rates for 4x8 and 2x(4+4), so we choose
1610 * 2x(4+4) because it results in significantly fewer pairs.
1611 * For RF, the raw pair rate of the 4x8 kernel is higher than 2x(4+4),
1612 * 10% with HT, 50% without HT. As we currently don't detect the actual
1613 * use of HT, use 4x8 to avoid a potential performance hit.
1614 * On Intel Haswell 4x8 is always faster.
1616 *kernel_type = nbnxnk4xN_SIMD_4xN;
1618 #if !GMX_SIMD_HAVE_FMA
1619 if (EEL_PME_EWALD(ir->coulombtype) ||
1620 EVDW_PME(ir->vdwtype))
1622 /* We have Ewald kernels without FMA (Intel Sandy/Ivy Bridge).
1623 * There are enough instructions to make 2x(4+4) efficient.
1625 *kernel_type = nbnxnk4xN_SIMD_2xNN;
1628 #endif /* GMX_NBNXN_SIMD_2XNN && GMX_NBNXN_SIMD_4XN */
1631 if (getenv("GMX_NBNXN_SIMD_4XN") != nullptr)
1633 #ifdef GMX_NBNXN_SIMD_4XN
1634 *kernel_type = nbnxnk4xN_SIMD_4xN;
1636 gmx_fatal(FARGS, "SIMD 4xN kernels requested, but GROMACS has been compiled without support for these kernels");
1639 if (getenv("GMX_NBNXN_SIMD_2XNN") != nullptr)
1641 #ifdef GMX_NBNXN_SIMD_2XNN
1642 *kernel_type = nbnxnk4xN_SIMD_2xNN;
1644 gmx_fatal(FARGS, "SIMD 2x(N+N) kernels requested, but GROMACS has been compiled without support for these kernels");
1648 /* Analytical Ewald exclusion correction is only an option in
1650 * Since table lookup's don't parallelize with SIMD, analytical
1651 * will probably always be faster for a SIMD width of 8 or more.
1652 * With FMA analytical is sometimes faster for a width if 4 as well.
1653 * On BlueGene/Q, this is faster regardless of precision.
1654 * In single precision, this is faster on Bulldozer.
1655 * On Skylake table is faster in single and double. TODO: Test 5xxx series.
1657 #if ((GMX_SIMD_REAL_WIDTH >= 8 || (GMX_SIMD_REAL_WIDTH >= 4 && GMX_SIMD_HAVE_FMA && !GMX_DOUBLE)) \
1658 && !GMX_SIMD_X86_AVX_512) || GMX_SIMD_IBM_QPX
1659 *ewald_excl = ewaldexclAnalytical;
1661 if (getenv("GMX_NBNXN_EWALD_TABLE") != nullptr)
1663 *ewald_excl = ewaldexclTable;
1665 if (getenv("GMX_NBNXN_EWALD_ANALYTICAL") != nullptr)
1667 *ewald_excl = ewaldexclAnalytical;
1675 const char *lookup_nbnxn_kernel_name(int kernel_type)
1677 const char *returnvalue = nullptr;
1678 switch (kernel_type)
1681 returnvalue = "not set";
1683 case nbnxnk4x4_PlainC:
1684 returnvalue = "plain C";
1686 case nbnxnk4xN_SIMD_4xN:
1687 case nbnxnk4xN_SIMD_2xNN:
1689 returnvalue = "SIMD";
1691 returnvalue = "not available";
1694 case nbnxnk8x8x8_GPU: returnvalue = "GPU"; break;
1695 case nbnxnk8x8x8_PlainC: returnvalue = "plain C"; break;
1699 gmx_fatal(FARGS, "Illegal kernel type selected");
1700 returnvalue = nullptr;
1706 static void pick_nbnxn_kernel(FILE *fp,
1707 const gmx::MDLogger &mdlog,
1708 gmx_bool use_simd_kernels,
1711 const t_inputrec *ir,
1714 gmx_bool bDoNonbonded)
1716 assert(kernel_type);
1718 *kernel_type = nbnxnkNotSet;
1719 *ewald_excl = ewaldexclTable;
1723 *kernel_type = nbnxnk8x8x8_PlainC;
1727 GMX_LOG(mdlog.warning).asParagraph().appendText("Emulating a GPU run on the CPU (slow)");
1732 *kernel_type = nbnxnk8x8x8_GPU;
1735 if (*kernel_type == nbnxnkNotSet)
1737 if (use_simd_kernels &&
1738 nbnxn_simd_supported(mdlog, ir))
1740 pick_nbnxn_kernel_cpu(ir, kernel_type, ewald_excl);
1744 *kernel_type = nbnxnk4x4_PlainC;
1748 if (bDoNonbonded && fp != nullptr)
1750 fprintf(fp, "\nUsing %s %dx%d non-bonded kernels\n\n",
1751 lookup_nbnxn_kernel_name(*kernel_type),
1752 nbnxn_kernel_to_cluster_i_size(*kernel_type),
1753 nbnxn_kernel_to_cluster_j_size(*kernel_type));
1755 if (nbnxnk4x4_PlainC == *kernel_type ||
1756 nbnxnk8x8x8_PlainC == *kernel_type)
1758 GMX_LOG(mdlog.warning).asParagraph().appendTextFormatted(
1759 "WARNING: Using the slow %s kernels. This should\n"
1760 "not happen during routine usage on supported platforms.",
1761 lookup_nbnxn_kernel_name(*kernel_type));
1766 gmx_bool uses_simple_tables(int cutoff_scheme,
1767 nonbonded_verlet_t *nbv,
1770 gmx_bool bUsesSimpleTables = TRUE;
1773 switch (cutoff_scheme)
1776 bUsesSimpleTables = TRUE;
1779 assert(NULL != nbv && NULL != nbv->grp);
1780 grp_index = (group < 0) ? 0 : (nbv->ngrp - 1);
1781 bUsesSimpleTables = nbnxn_kernel_pairlist_simple(nbv->grp[grp_index].kernel_type);
1784 gmx_incons("unimplemented");
1786 return bUsesSimpleTables;
1789 static void init_ewald_f_table(interaction_const_t *ic,
1794 /* Get the Ewald table spacing based on Coulomb and/or LJ
1795 * Ewald coefficients and rtol.
1797 ic->tabq_scale = ewald_spline3_table_scale(ic);
1799 if (ic->cutoff_scheme == ecutsVERLET)
1801 maxr = ic->rcoulomb;
1805 maxr = std::max(ic->rcoulomb, rtab);
1807 ic->tabq_size = static_cast<int>(maxr*ic->tabq_scale) + 2;
1809 sfree_aligned(ic->tabq_coul_FDV0);
1810 sfree_aligned(ic->tabq_coul_F);
1811 sfree_aligned(ic->tabq_coul_V);
1813 sfree_aligned(ic->tabq_vdw_FDV0);
1814 sfree_aligned(ic->tabq_vdw_F);
1815 sfree_aligned(ic->tabq_vdw_V);
1817 if (EEL_PME_EWALD(ic->eeltype))
1819 /* Create the original table data in FDV0 */
1820 snew_aligned(ic->tabq_coul_FDV0, ic->tabq_size*4, 32);
1821 snew_aligned(ic->tabq_coul_F, ic->tabq_size, 32);
1822 snew_aligned(ic->tabq_coul_V, ic->tabq_size, 32);
1823 table_spline3_fill_ewald_lr(ic->tabq_coul_F, ic->tabq_coul_V, ic->tabq_coul_FDV0,
1824 ic->tabq_size, 1/ic->tabq_scale, ic->ewaldcoeff_q, v_q_ewald_lr);
1827 if (EVDW_PME(ic->vdwtype))
1829 snew_aligned(ic->tabq_vdw_FDV0, ic->tabq_size*4, 32);
1830 snew_aligned(ic->tabq_vdw_F, ic->tabq_size, 32);
1831 snew_aligned(ic->tabq_vdw_V, ic->tabq_size, 32);
1832 table_spline3_fill_ewald_lr(ic->tabq_vdw_F, ic->tabq_vdw_V, ic->tabq_vdw_FDV0,
1833 ic->tabq_size, 1/ic->tabq_scale, ic->ewaldcoeff_lj, v_lj_ewald_lr);
1837 void init_interaction_const_tables(FILE *fp,
1838 interaction_const_t *ic,
1841 if (EEL_PME_EWALD(ic->eeltype) || EVDW_PME(ic->vdwtype))
1843 init_ewald_f_table(ic, rtab);
1847 fprintf(fp, "Initialized non-bonded Ewald correction tables, spacing: %.2e size: %d\n\n",
1848 1/ic->tabq_scale, ic->tabq_size);
1853 static void clear_force_switch_constants(shift_consts_t *sc)
1860 static void force_switch_constants(real p,
1864 /* Here we determine the coefficient for shifting the force to zero
1865 * between distance rsw and the cut-off rc.
1866 * For a potential of r^-p, we have force p*r^-(p+1).
1867 * But to save flops we absorb p in the coefficient.
1869 * force/p = r^-(p+1) + c2*r^2 + c3*r^3
1870 * potential = r^-p + c2/3*r^3 + c3/4*r^4 + cpot
1872 sc->c2 = ((p + 1)*rsw - (p + 4)*rc)/(pow(rc, p + 2)*gmx::square(rc - rsw));
1873 sc->c3 = -((p + 1)*rsw - (p + 3)*rc)/(pow(rc, p + 2)*gmx::power3(rc - rsw));
1874 sc->cpot = -pow(rc, -p) + p*sc->c2/3*gmx::power3(rc - rsw) + p*sc->c3/4*gmx::power4(rc - rsw);
1877 static void potential_switch_constants(real rsw, real rc,
1878 switch_consts_t *sc)
1880 /* The switch function is 1 at rsw and 0 at rc.
1881 * The derivative and second derivate are zero at both ends.
1882 * rsw = max(r - r_switch, 0)
1883 * sw = 1 + c3*rsw^3 + c4*rsw^4 + c5*rsw^5
1884 * dsw = 3*c3*rsw^2 + 4*c4*rsw^3 + 5*c5*rsw^4
1885 * force = force*dsw - potential*sw
1888 sc->c3 = -10/gmx::power3(rc - rsw);
1889 sc->c4 = 15/gmx::power4(rc - rsw);
1890 sc->c5 = -6/gmx::power5(rc - rsw);
1893 /*! \brief Construct interaction constants
1895 * This data is used (particularly) by search and force code for
1896 * short-range interactions. Many of these are constant for the whole
1897 * simulation; some are constant only after PME tuning completes.
1900 init_interaction_const(FILE *fp,
1901 interaction_const_t **interaction_const,
1902 const t_forcerec *fr)
1904 interaction_const_t *ic;
1908 ic->cutoff_scheme = fr->cutoff_scheme;
1910 /* Just allocate something so we can free it */
1911 snew_aligned(ic->tabq_coul_FDV0, 16, 32);
1912 snew_aligned(ic->tabq_coul_F, 16, 32);
1913 snew_aligned(ic->tabq_coul_V, 16, 32);
1916 ic->vdwtype = fr->vdwtype;
1917 ic->vdw_modifier = fr->vdw_modifier;
1918 ic->rvdw = fr->rvdw;
1919 ic->rvdw_switch = fr->rvdw_switch;
1920 ic->ewaldcoeff_lj = fr->ewaldcoeff_lj;
1921 ic->ljpme_comb_rule = fr->ljpme_combination_rule;
1922 ic->sh_lj_ewald = 0;
1923 clear_force_switch_constants(&ic->dispersion_shift);
1924 clear_force_switch_constants(&ic->repulsion_shift);
1926 switch (ic->vdw_modifier)
1928 case eintmodPOTSHIFT:
1929 /* Only shift the potential, don't touch the force */
1930 ic->dispersion_shift.cpot = -1.0/gmx::power6(ic->rvdw);
1931 ic->repulsion_shift.cpot = -1.0/gmx::power12(ic->rvdw);
1932 if (EVDW_PME(ic->vdwtype))
1936 crc2 = gmx::square(ic->ewaldcoeff_lj*ic->rvdw);
1937 ic->sh_lj_ewald = (std::exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)/gmx::power6(ic->rvdw);
1940 case eintmodFORCESWITCH:
1941 /* Switch the force, switch and shift the potential */
1942 force_switch_constants(6.0, ic->rvdw_switch, ic->rvdw,
1943 &ic->dispersion_shift);
1944 force_switch_constants(12.0, ic->rvdw_switch, ic->rvdw,
1945 &ic->repulsion_shift);
1947 case eintmodPOTSWITCH:
1948 /* Switch the potential and force */
1949 potential_switch_constants(ic->rvdw_switch, ic->rvdw,
1953 case eintmodEXACTCUTOFF:
1954 /* Nothing to do here */
1957 gmx_incons("unimplemented potential modifier");
1960 ic->sh_invrc6 = -ic->dispersion_shift.cpot;
1962 /* Electrostatics */
1963 ic->eeltype = fr->eeltype;
1964 ic->coulomb_modifier = fr->coulomb_modifier;
1965 ic->rcoulomb = fr->rcoulomb;
1966 ic->epsilon_r = fr->epsilon_r;
1967 ic->epsfac = fr->epsfac;
1968 ic->ewaldcoeff_q = fr->ewaldcoeff_q;
1970 if (fr->coulomb_modifier == eintmodPOTSHIFT)
1972 ic->sh_ewald = std::erfc(ic->ewaldcoeff_q*ic->rcoulomb);
1979 /* Reaction-field */
1980 if (EEL_RF(ic->eeltype))
1982 ic->epsilon_rf = fr->epsilon_rf;
1983 ic->k_rf = fr->k_rf;
1984 ic->c_rf = fr->c_rf;
1988 /* For plain cut-off we might use the reaction-field kernels */
1989 ic->epsilon_rf = ic->epsilon_r;
1991 if (fr->coulomb_modifier == eintmodPOTSHIFT)
1993 ic->c_rf = 1/ic->rcoulomb;
2003 real dispersion_shift;
2005 dispersion_shift = ic->dispersion_shift.cpot;
2006 if (EVDW_PME(ic->vdwtype))
2008 dispersion_shift -= ic->sh_lj_ewald;
2010 fprintf(fp, "Potential shift: LJ r^-12: %.3e r^-6: %.3e",
2011 ic->repulsion_shift.cpot, dispersion_shift);
2013 if (ic->eeltype == eelCUT)
2015 fprintf(fp, ", Coulomb %.e", -ic->c_rf);
2017 else if (EEL_PME(ic->eeltype))
2019 fprintf(fp, ", Ewald %.3e", -ic->sh_ewald);
2024 *interaction_const = ic;
2027 /* TODO deviceInfo should be logically const, but currently
2028 * init_gpu modifies it to set up NVML support. This could
2029 * happen during the detection phase, and deviceInfo could
2030 * the become const. */
2031 static void init_nb_verlet(FILE *fp,
2032 const gmx::MDLogger &mdlog,
2033 nonbonded_verlet_t **nb_verlet,
2034 gmx_bool bFEP_NonBonded,
2035 const t_inputrec *ir,
2036 const t_forcerec *fr,
2037 const t_commrec *cr,
2038 const char *nbpu_opt,
2039 gmx_device_info_t *deviceInfo,
2040 const gmx_mtop_t *mtop,
2043 nonbonded_verlet_t *nbv;
2046 gmx_bool bHybridGPURun = FALSE;
2048 nbnxn_alloc_t *nb_alloc;
2049 nbnxn_free_t *nb_free;
2051 nbv = new nonbonded_verlet_t();
2053 nbv->emulateGpu = (getenv("GMX_EMULATE_GPU") != nullptr);
2054 nbv->bUseGPU = deviceInfo != nullptr;
2056 GMX_RELEASE_ASSERT(!(nbv->emulateGpu && nbv->bUseGPU), "When GPU emulation is active, there cannot be a GPU assignment");
2060 /* Use the assigned GPU. */
2061 init_gpu(mdlog, cr->nodeid, deviceInfo);
2065 nbv->min_ci_balanced = 0;
2067 nbv->ngrp = (DOMAINDECOMP(cr) ? 2 : 1);
2068 for (i = 0; i < nbv->ngrp; i++)
2070 nbv->grp[i].nbl_lists.nnbl = 0;
2071 nbv->grp[i].nbat = nullptr;
2072 nbv->grp[i].kernel_type = nbnxnkNotSet;
2074 if (i == 0) /* local */
2076 pick_nbnxn_kernel(fp, mdlog, fr->use_simd_kernels,
2077 nbv->bUseGPU, nbv->emulateGpu, ir,
2078 &nbv->grp[i].kernel_type,
2079 &nbv->grp[i].ewald_excl,
2082 else /* non-local */
2084 if (nbpu_opt != nullptr && strcmp(nbpu_opt, "gpu_cpu") == 0)
2086 /* Use GPU for local, select a CPU kernel for non-local */
2087 pick_nbnxn_kernel(fp, mdlog, fr->use_simd_kernels,
2089 &nbv->grp[i].kernel_type,
2090 &nbv->grp[i].ewald_excl,
2093 bHybridGPURun = TRUE;
2097 /* Use the same kernel for local and non-local interactions */
2098 nbv->grp[i].kernel_type = nbv->grp[0].kernel_type;
2099 nbv->grp[i].ewald_excl = nbv->grp[0].ewald_excl;
2104 nbv->listParams = std::unique_ptr<NbnxnListParameters>(new NbnxnListParameters(ir->rlist));
2105 setupDynamicPairlistPruning(fp, ir, mtop, box, nbv->bUseGPU, fr->ic,
2106 nbv->listParams.get());
2108 nbnxn_init_search(&nbv->nbs,
2109 DOMAINDECOMP(cr) ? &cr->dd->nc : nullptr,
2110 DOMAINDECOMP(cr) ? domdec_zones(cr->dd) : nullptr,
2112 gmx_omp_nthreads_get(emntPairsearch));
2114 for (i = 0; i < nbv->ngrp; i++)
2116 gpu_set_host_malloc_and_free(nbv->grp[0].kernel_type == nbnxnk8x8x8_GPU,
2117 &nb_alloc, &nb_free);
2119 nbnxn_init_pairlist_set(&nbv->grp[i].nbl_lists,
2120 nbnxn_kernel_pairlist_simple(nbv->grp[i].kernel_type),
2121 /* 8x8x8 "non-simple" lists are ATM always combined */
2122 !nbnxn_kernel_pairlist_simple(nbv->grp[i].kernel_type),
2126 nbv->grp[0].kernel_type != nbv->grp[i].kernel_type)
2128 gmx_bool bSimpleList;
2129 int enbnxninitcombrule;
2131 bSimpleList = nbnxn_kernel_pairlist_simple(nbv->grp[i].kernel_type);
2133 if (fr->vdwtype == evdwCUT &&
2134 (fr->vdw_modifier == eintmodNONE ||
2135 fr->vdw_modifier == eintmodPOTSHIFT) &&
2136 getenv("GMX_NO_LJ_COMB_RULE") == nullptr)
2138 /* Plain LJ cut-off: we can optimize with combination rules */
2139 enbnxninitcombrule = enbnxninitcombruleDETECT;
2141 else if (fr->vdwtype == evdwPME)
2143 /* LJ-PME: we need to use a combination rule for the grid */
2144 if (fr->ljpme_combination_rule == eljpmeGEOM)
2146 enbnxninitcombrule = enbnxninitcombruleGEOM;
2150 enbnxninitcombrule = enbnxninitcombruleLB;
2155 /* We use a full combination matrix: no rule required */
2156 enbnxninitcombrule = enbnxninitcombruleNONE;
2160 snew(nbv->grp[i].nbat, 1);
2161 nbnxn_atomdata_init(fp,
2163 nbv->grp[i].kernel_type,
2165 fr->ntype, fr->nbfp,
2167 bSimpleList ? gmx_omp_nthreads_get(emntNonbonded) : 1,
2172 nbv->grp[i].nbat = nbv->grp[0].nbat;
2178 /* init the NxN GPU data; the last argument tells whether we'll have
2179 * both local and non-local NB calculation on GPU */
2180 nbnxn_gpu_init(&nbv->gpu_nbv,
2183 nbv->listParams.get(),
2186 (nbv->ngrp > 1) && !bHybridGPURun);
2188 /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
2189 * also sharing texture references. To keep the code simple, we don't
2190 * treat texture references as shared resources, but this means that
2191 * the coulomb_tab and nbfp texture refs will get updated by multiple threads.
2192 * Hence, to ensure that the non-bonded kernels don't start before all
2193 * texture binding operations are finished, we need to wait for all ranks
2194 * to arrive here before continuing.
2196 * Note that we could omit this barrier if GPUs are not shared (or
2197 * texture objects are used), but as this is initialization code, there
2198 * is no point in complicating things.
2205 #endif /* GMX_THREAD_MPI */
2207 if ((env = getenv("GMX_NB_MIN_CI")) != nullptr)
2211 nbv->min_ci_balanced = strtol(env, &end, 10);
2212 if (!end || (*end != 0) || nbv->min_ci_balanced < 0)
2214 gmx_fatal(FARGS, "Invalid value passed in GMX_NB_MIN_CI=%s, non-negative integer required", env);
2219 fprintf(debug, "Neighbor-list balancing parameter: %d (passed as env. var.)\n",
2220 nbv->min_ci_balanced);
2225 nbv->min_ci_balanced = nbnxn_gpu_min_ci_balanced(nbv->gpu_nbv);
2228 fprintf(debug, "Neighbor-list balancing parameter: %d (auto-adjusted to the number of GPU multi-processors)\n",
2229 nbv->min_ci_balanced);
2238 gmx_bool usingGpu(nonbonded_verlet_t *nbv)
2240 return nbv != nullptr && nbv->bUseGPU;
2243 void init_forcerec(FILE *fp,
2244 const gmx::MDLogger &mdlog,
2247 const t_inputrec *ir,
2248 const gmx_mtop_t *mtop,
2249 const t_commrec *cr,
2253 const t_filenm *tabbfnm,
2254 const char *nbpu_opt,
2255 gmx_device_info_t *deviceInfo,
2256 gmx_bool bNoSolvOpt,
2259 int i, m, negp_pp, negptable, egi, egj;
2264 gmx_bool bGenericKernelOnly;
2265 gmx_bool needGroupSchemeTables, bSomeNormalNbListsAreInUse;
2266 gmx_bool bFEP_NonBonded;
2267 int *nm_ind, egp_flags;
2269 /* By default we turn SIMD kernels on, but it might be turned off further down... */
2270 fr->use_simd_kernels = TRUE;
2272 fr->bDomDec = DOMAINDECOMP(cr);
2274 if (check_box(ir->ePBC, box))
2276 gmx_fatal(FARGS, check_box(ir->ePBC, box));
2279 /* Test particle insertion ? */
2282 /* Set to the size of the molecule to be inserted (the last one) */
2283 /* Because of old style topologies, we have to use the last cg
2284 * instead of the last molecule type.
2286 cgs = &mtop->moltype[mtop->molblock[mtop->nmolblock-1].type].cgs;
2287 fr->n_tpi = cgs->index[cgs->nr] - cgs->index[cgs->nr-1];
2288 if (fr->n_tpi != mtop->mols.index[mtop->mols.nr] - mtop->mols.index[mtop->mols.nr-1])
2290 gmx_fatal(FARGS, "The molecule to insert can not consist of multiple charge groups.\nMake it a single charge group.");
2298 if (ir->coulombtype == eelRF_NEC_UNSUPPORTED)
2300 gmx_fatal(FARGS, "%s electrostatics is no longer supported",
2301 eel_names[ir->coulombtype]);
2306 gmx_fatal(FARGS, "AdResS simulations are no longer supported");
2308 if (ir->useTwinRange)
2310 gmx_fatal(FARGS, "Twin-range simulations are no longer supported");
2312 /* Copy the user determined parameters */
2313 fr->userint1 = ir->userint1;
2314 fr->userint2 = ir->userint2;
2315 fr->userint3 = ir->userint3;
2316 fr->userint4 = ir->userint4;
2317 fr->userreal1 = ir->userreal1;
2318 fr->userreal2 = ir->userreal2;
2319 fr->userreal3 = ir->userreal3;
2320 fr->userreal4 = ir->userreal4;
2323 fr->fc_stepsize = ir->fc_stepsize;
2326 fr->efep = ir->efep;
2327 fr->sc_alphavdw = ir->fepvals->sc_alpha;
2328 if (ir->fepvals->bScCoul)
2330 fr->sc_alphacoul = ir->fepvals->sc_alpha;
2331 fr->sc_sigma6_min = gmx::power6(ir->fepvals->sc_sigma_min);
2335 fr->sc_alphacoul = 0;
2336 fr->sc_sigma6_min = 0; /* only needed when bScCoul is on */
2338 fr->sc_power = ir->fepvals->sc_power;
2339 fr->sc_r_power = ir->fepvals->sc_r_power;
2340 fr->sc_sigma6_def = gmx::power6(ir->fepvals->sc_sigma);
2342 env = getenv("GMX_SCSIGMA_MIN");
2346 sscanf(env, "%20lf", &dbl);
2347 fr->sc_sigma6_min = gmx::power6(dbl);
2350 fprintf(fp, "Setting the minimum soft core sigma to %g nm\n", dbl);
2354 fr->bNonbonded = TRUE;
2355 if (getenv("GMX_NO_NONBONDED") != nullptr)
2357 /* turn off non-bonded calculations */
2358 fr->bNonbonded = FALSE;
2359 GMX_LOG(mdlog.warning).asParagraph().appendText(
2360 "Found environment variable GMX_NO_NONBONDED.\n"
2361 "Disabling nonbonded calculations.");
2364 bGenericKernelOnly = FALSE;
2366 /* We now check in the NS code whether a particular combination of interactions
2367 * can be used with water optimization, and disable it if that is not the case.
2370 if (getenv("GMX_NB_GENERIC") != nullptr)
2375 "Found environment variable GMX_NB_GENERIC.\n"
2376 "Disabling all interaction-specific nonbonded kernels, will only\n"
2377 "use the slow generic ones in src/gmxlib/nonbonded/nb_generic.c\n\n");
2379 bGenericKernelOnly = TRUE;
2382 if (bGenericKernelOnly == TRUE)
2387 if ( (getenv("GMX_DISABLE_SIMD_KERNELS") != nullptr) || (getenv("GMX_NOOPTIMIZEDKERNELS") != nullptr) )
2389 fr->use_simd_kernels = FALSE;
2393 "\nFound environment variable GMX_DISABLE_SIMD_KERNELS.\n"
2394 "Disabling the usage of any SIMD-specific non-bonded & bonded kernel routines\n"
2395 "(e.g. SSE2/SSE4.1/AVX).\n\n");
2399 fr->bBHAM = (mtop->ffparams.functype[0] == F_BHAM);
2401 /* Check if we can/should do all-vs-all kernels */
2402 fr->bAllvsAll = can_use_allvsall(ir, FALSE, nullptr, nullptr);
2403 fr->AllvsAll_work = nullptr;
2404 fr->AllvsAll_workgb = nullptr;
2406 /* All-vs-all kernels have not been implemented in 4.6 and later.
2407 * See Redmine #1249. */
2410 fr->bAllvsAll = FALSE;
2414 "\nYour simulation settings would have triggered the efficient all-vs-all\n"
2415 "kernels in GROMACS 4.5, but these have not been implemented in GROMACS\n"
2416 "4.6 and 5.x. If performance is important, please use GROMACS 4.5.7\n"
2417 "or try cutoff-scheme = Verlet.\n\n");
2421 /* Neighbour searching stuff */
2422 fr->cutoff_scheme = ir->cutoff_scheme;
2423 fr->bGrid = (ir->ns_type == ensGRID);
2424 fr->ePBC = ir->ePBC;
2426 if (fr->cutoff_scheme == ecutsGROUP)
2428 const char *note = "NOTE: This file uses the deprecated 'group' cutoff_scheme. This will be\n"
2429 "removed in a future release when 'verlet' supports all interaction forms.\n";
2433 fprintf(stderr, "\n%s\n", note);
2437 fprintf(fp, "\n%s\n", note);
2441 /* Determine if we will do PBC for distances in bonded interactions */
2442 if (fr->ePBC == epbcNONE)
2444 fr->bMolPBC = FALSE;
2448 if (!DOMAINDECOMP(cr))
2452 bSHAKE = (ir->eConstrAlg == econtSHAKE &&
2453 (gmx_mtop_ftype_count(mtop, F_CONSTR) > 0 ||
2454 gmx_mtop_ftype_count(mtop, F_CONSTRNC) > 0));
2456 /* The group cut-off scheme and SHAKE assume charge groups
2457 * are whole, but not using molpbc is faster in most cases.
2458 * With intermolecular interactions we need PBC for calculating
2459 * distances between atoms in different molecules.
2461 if ((fr->cutoff_scheme == ecutsGROUP || bSHAKE) &&
2462 !mtop->bIntermolecularInteractions)
2464 fr->bMolPBC = ir->bPeriodicMols;
2466 if (bSHAKE && fr->bMolPBC)
2468 gmx_fatal(FARGS, "SHAKE is not supported with periodic molecules");
2473 /* Not making molecules whole is faster in most cases,
2474 * but With orientation restraints we need whole molecules.
2476 fr->bMolPBC = (fcd->orires.nr == 0);
2478 if (getenv("GMX_USE_GRAPH") != nullptr)
2480 fr->bMolPBC = FALSE;
2483 GMX_LOG(mdlog.warning).asParagraph().appendText("GMX_USE_GRAPH is set, using the graph for bonded interactions");
2486 if (mtop->bIntermolecularInteractions)
2488 GMX_LOG(mdlog.warning).asParagraph().appendText("WARNING: Molecules linked by intermolecular interactions have to reside in the same periodic image, otherwise artifacts will occur!");
2492 GMX_RELEASE_ASSERT(fr->bMolPBC || !mtop->bIntermolecularInteractions, "We need to use PBC within molecules with inter-molecular interactions");
2494 if (bSHAKE && fr->bMolPBC)
2496 gmx_fatal(FARGS, "SHAKE is not properly supported with intermolecular interactions. For short simulations where linked molecules remain in the same periodic image, the environment variable GMX_USE_GRAPH can be used to override this check.\n");
2502 fr->bMolPBC = dd_bonded_molpbc(cr->dd, fr->ePBC);
2505 fr->bGB = (ir->implicit_solvent == eisGBSA);
2507 fr->rc_scaling = ir->refcoord_scaling;
2508 copy_rvec(ir->posres_com, fr->posres_com);
2509 copy_rvec(ir->posres_comB, fr->posres_comB);
2510 fr->rlist = cutoff_inf(ir->rlist);
2511 fr->eeltype = ir->coulombtype;
2512 fr->vdwtype = ir->vdwtype;
2513 fr->ljpme_combination_rule = ir->ljpme_combination_rule;
2515 fr->coulomb_modifier = ir->coulomb_modifier;
2516 fr->vdw_modifier = ir->vdw_modifier;
2518 /* Electrostatics: Translate from interaction-setting-in-mdp-file to kernel interaction format */
2519 switch (fr->eeltype)
2522 fr->nbkernel_elec_interaction = (fr->bGB) ? GMX_NBKERNEL_ELEC_GENERALIZEDBORN : GMX_NBKERNEL_ELEC_COULOMB;
2527 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_REACTIONFIELD;
2531 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_REACTIONFIELD;
2532 fr->coulomb_modifier = eintmodEXACTCUTOFF;
2541 case eelPMEUSERSWITCH:
2542 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_CUBICSPLINETABLE;
2548 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_EWALD;
2552 gmx_fatal(FARGS, "Unsupported electrostatic interaction: %s", eel_names[fr->eeltype]);
2556 /* Vdw: Translate from mdp settings to kernel format */
2557 switch (fr->vdwtype)
2562 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_BUCKINGHAM;
2566 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_LENNARDJONES;
2570 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_LJEWALD;
2576 case evdwENCADSHIFT:
2577 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_CUBICSPLINETABLE;
2581 gmx_fatal(FARGS, "Unsupported vdw interaction: %s", evdw_names[fr->vdwtype]);
2585 /* These start out identical to ir, but might be altered if we e.g. tabulate the interaction in the kernel */
2586 fr->nbkernel_elec_modifier = fr->coulomb_modifier;
2587 fr->nbkernel_vdw_modifier = fr->vdw_modifier;
2589 fr->rvdw = cutoff_inf(ir->rvdw);
2590 fr->rvdw_switch = ir->rvdw_switch;
2591 fr->rcoulomb = cutoff_inf(ir->rcoulomb);
2592 fr->rcoulomb_switch = ir->rcoulomb_switch;
2594 fr->bEwald = EEL_PME_EWALD(fr->eeltype);
2596 fr->reppow = mtop->ffparams.reppow;
2598 if (ir->cutoff_scheme == ecutsGROUP)
2600 fr->bvdwtab = ((fr->vdwtype != evdwCUT || !gmx_within_tol(fr->reppow, 12.0, 10*GMX_DOUBLE_EPS))
2601 && !EVDW_PME(fr->vdwtype));
2602 /* We have special kernels for standard Ewald and PME, but the pme-switch ones are tabulated above */
2603 fr->bcoultab = !(fr->eeltype == eelCUT ||
2604 fr->eeltype == eelEWALD ||
2605 fr->eeltype == eelPME ||
2606 fr->eeltype == eelRF ||
2607 fr->eeltype == eelRF_ZERO);
2609 /* If the user absolutely wants different switch/shift settings for coul/vdw, it is likely
2610 * going to be faster to tabulate the interaction than calling the generic kernel.
2611 * However, if generic kernels have been requested we keep things analytically.
2613 if (fr->nbkernel_elec_modifier == eintmodPOTSWITCH &&
2614 fr->nbkernel_vdw_modifier == eintmodPOTSWITCH &&
2615 bGenericKernelOnly == FALSE)
2617 if ((fr->rcoulomb_switch != fr->rvdw_switch) || (fr->rcoulomb != fr->rvdw))
2619 fr->bcoultab = TRUE;
2620 /* Once we tabulate electrostatics, we can use the switch function for LJ,
2621 * which would otherwise need two tables.
2625 else if ((fr->nbkernel_elec_modifier == eintmodPOTSHIFT && fr->nbkernel_vdw_modifier == eintmodPOTSHIFT) ||
2626 ((fr->nbkernel_elec_interaction == GMX_NBKERNEL_ELEC_REACTIONFIELD &&
2627 fr->nbkernel_elec_modifier == eintmodEXACTCUTOFF &&
2628 (fr->nbkernel_vdw_modifier == eintmodPOTSWITCH || fr->nbkernel_vdw_modifier == eintmodPOTSHIFT))))
2630 if ((fr->rcoulomb != fr->rvdw) && (bGenericKernelOnly == FALSE))
2632 fr->bcoultab = TRUE;
2636 if (fr->nbkernel_elec_modifier == eintmodFORCESWITCH)
2638 fr->bcoultab = TRUE;
2640 if (fr->nbkernel_vdw_modifier == eintmodFORCESWITCH)
2645 if (getenv("GMX_REQUIRE_TABLES"))
2648 fr->bcoultab = TRUE;
2653 fprintf(fp, "Table routines are used for coulomb: %s\n",
2654 gmx::boolToString(fr->bcoultab));
2655 fprintf(fp, "Table routines are used for vdw: %s\n",
2656 gmx::boolToString(fr->bvdwtab));
2659 if (fr->bvdwtab == TRUE)
2661 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_CUBICSPLINETABLE;
2662 fr->nbkernel_vdw_modifier = eintmodNONE;
2664 if (fr->bcoultab == TRUE)
2666 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_CUBICSPLINETABLE;
2667 fr->nbkernel_elec_modifier = eintmodNONE;
2671 if (ir->cutoff_scheme == ecutsVERLET)
2673 if (!gmx_within_tol(fr->reppow, 12.0, 10*GMX_DOUBLE_EPS))
2675 gmx_fatal(FARGS, "Cut-off scheme %S only supports LJ repulsion power 12", ecutscheme_names[ir->cutoff_scheme]);
2677 fr->bvdwtab = FALSE;
2678 fr->bcoultab = FALSE;
2681 /* This now calculates sum for q and C6 */
2682 set_chargesum(fp, fr, mtop);
2684 /* Tables are used for direct ewald sum */
2687 if (EEL_PME(ir->coulombtype))
2691 fprintf(fp, "Will do PME sum in reciprocal space for electrostatic interactions.\n");
2693 if (ir->coulombtype == eelP3M_AD)
2695 please_cite(fp, "Hockney1988");
2696 please_cite(fp, "Ballenegger2012");
2700 please_cite(fp, "Essmann95a");
2703 if (ir->ewald_geometry == eewg3DC)
2705 bool haveNetCharge = (fabs(fr->qsum[0]) > 1e-4 ||
2706 fabs(fr->qsum[1]) > 1e-4);
2709 fprintf(fp, "Using the Ewald3DC correction for systems with a slab geometry%s.\n",
2710 haveNetCharge ? " and net charge" : "");
2712 please_cite(fp, "In-Chul99a");
2715 please_cite(fp, "Ballenegger2009");
2719 fr->ewaldcoeff_q = calc_ewaldcoeff_q(ir->rcoulomb, ir->ewald_rtol);
2720 init_ewald_tab(&(fr->ewald_table), ir, fp);
2723 fprintf(fp, "Using a Gaussian width (1/beta) of %g nm for Ewald\n",
2724 1/fr->ewaldcoeff_q);
2728 if (EVDW_PME(ir->vdwtype))
2732 fprintf(fp, "Will do PME sum in reciprocal space for LJ dispersion interactions.\n");
2734 please_cite(fp, "Essmann95a");
2735 fr->ewaldcoeff_lj = calc_ewaldcoeff_lj(ir->rvdw, ir->ewald_rtol_lj);
2738 fprintf(fp, "Using a Gaussian width (1/beta) of %g nm for LJ Ewald\n",
2739 1/fr->ewaldcoeff_lj);
2743 /* Electrostatics */
2744 fr->epsilon_r = ir->epsilon_r;
2745 fr->epsilon_rf = ir->epsilon_rf;
2746 fr->fudgeQQ = mtop->ffparams.fudgeQQ;
2748 /* Parameters for generalized RF */
2752 if (fr->eeltype == eelGRF)
2754 init_generalized_rf(fp, mtop, ir, fr);
2757 fr->bF_NoVirSum = (EEL_FULL(fr->eeltype) || EVDW_PME(fr->vdwtype) ||
2758 fr->forceProviders->hasForcesWithoutVirialContribution() ||
2759 gmx_mtop_ftype_count(mtop, F_POSRES) > 0 ||
2760 gmx_mtop_ftype_count(mtop, F_FBPOSRES) > 0);
2762 if (fr->bF_NoVirSum)
2764 fr->forceBufferNoVirialSummation = new PaddedRVecVector;
2767 if (fr->cutoff_scheme == ecutsGROUP &&
2768 ncg_mtop(mtop) > fr->cg_nalloc && !DOMAINDECOMP(cr))
2770 /* Count the total number of charge groups */
2771 fr->cg_nalloc = ncg_mtop(mtop);
2772 srenew(fr->cg_cm, fr->cg_nalloc);
2774 if (fr->shift_vec == nullptr)
2776 snew(fr->shift_vec, SHIFTS);
2779 if (fr->fshift == nullptr)
2781 snew(fr->fshift, SHIFTS);
2784 if (fr->nbfp == nullptr)
2786 fr->ntype = mtop->ffparams.atnr;
2787 fr->nbfp = mk_nbfp(&mtop->ffparams, fr->bBHAM);
2788 if (EVDW_PME(fr->vdwtype))
2790 fr->ljpme_c6grid = make_ljpme_c6grid(&mtop->ffparams, fr);
2794 /* Copy the energy group exclusions */
2795 fr->egp_flags = ir->opts.egp_flags;
2797 /* Van der Waals stuff */
2798 if ((fr->vdwtype != evdwCUT) && (fr->vdwtype != evdwUSER) && !fr->bBHAM)
2800 if (fr->rvdw_switch >= fr->rvdw)
2802 gmx_fatal(FARGS, "rvdw_switch (%f) must be < rvdw (%f)",
2803 fr->rvdw_switch, fr->rvdw);
2807 fprintf(fp, "Using %s Lennard-Jones, switch between %g and %g nm\n",
2808 (fr->eeltype == eelSWITCH) ? "switched" : "shifted",
2809 fr->rvdw_switch, fr->rvdw);
2813 if (fr->bBHAM && EVDW_PME(fr->vdwtype))
2815 gmx_fatal(FARGS, "LJ PME not supported with Buckingham");
2818 if (fr->bBHAM && (fr->vdwtype == evdwSHIFT || fr->vdwtype == evdwSWITCH))
2820 gmx_fatal(FARGS, "Switch/shift interaction not supported with Buckingham");
2823 if (fr->bBHAM && fr->cutoff_scheme == ecutsVERLET)
2825 gmx_fatal(FARGS, "Verlet cutoff-scheme is not supported with Buckingham");
2830 fprintf(fp, "Cut-off's: NS: %g Coulomb: %g %s: %g\n",
2831 fr->rlist, fr->rcoulomb, fr->bBHAM ? "BHAM" : "LJ", fr->rvdw);
2834 fr->eDispCorr = ir->eDispCorr;
2835 fr->numAtomsForDispersionCorrection = mtop->natoms;
2836 if (ir->eDispCorr != edispcNO)
2838 set_avcsixtwelve(fp, fr, mtop);
2843 set_bham_b_max(fp, fr, mtop);
2846 fr->gb_epsilon_solvent = ir->gb_epsilon_solvent;
2848 /* Copy the GBSA data (radius, volume and surftens for each
2849 * atomtype) from the topology atomtype section to forcerec.
2851 snew(fr->atype_radius, fr->ntype);
2852 snew(fr->atype_vol, fr->ntype);
2853 snew(fr->atype_surftens, fr->ntype);
2854 snew(fr->atype_gb_radius, fr->ntype);
2855 snew(fr->atype_S_hct, fr->ntype);
2857 if (mtop->atomtypes.nr > 0)
2859 for (i = 0; i < fr->ntype; i++)
2861 fr->atype_radius[i] = mtop->atomtypes.radius[i];
2863 for (i = 0; i < fr->ntype; i++)
2865 fr->atype_vol[i] = mtop->atomtypes.vol[i];
2867 for (i = 0; i < fr->ntype; i++)
2869 fr->atype_surftens[i] = mtop->atomtypes.surftens[i];
2871 for (i = 0; i < fr->ntype; i++)
2873 fr->atype_gb_radius[i] = mtop->atomtypes.gb_radius[i];
2875 for (i = 0; i < fr->ntype; i++)
2877 fr->atype_S_hct[i] = mtop->atomtypes.S_hct[i];
2881 /* Generate the GB table if needed */
2885 fr->gbtabscale = 2000;
2887 fr->gbtabscale = 500;
2891 fr->gbtab = make_gb_table(fr);
2893 init_gb(&fr->born, fr, ir, mtop, ir->gb_algorithm);
2895 /* Copy local gb data (for dd, this is done in dd_partition_system) */
2896 if (!DOMAINDECOMP(cr))
2898 make_local_gb(cr, fr->born, ir->gb_algorithm);
2902 /* Set the charge scaling */
2903 if (fr->epsilon_r != 0)
2905 fr->epsfac = ONE_4PI_EPS0/fr->epsilon_r;
2909 /* eps = 0 is infinite dieletric: no coulomb interactions */
2913 /* Reaction field constants */
2914 if (EEL_RF(fr->eeltype))
2916 calc_rffac(fp, fr->eeltype, fr->epsilon_r, fr->epsilon_rf,
2917 fr->rcoulomb, fr->temp, fr->zsquare, box,
2918 &fr->kappa, &fr->k_rf, &fr->c_rf);
2921 /* Construct tables for the group scheme. A little unnecessary to
2922 * make both vdw and coul tables sometimes, but what the
2923 * heck. Note that both cutoff schemes construct Ewald tables in
2924 * init_interaction_const_tables. */
2925 needGroupSchemeTables = (ir->cutoff_scheme == ecutsGROUP &&
2926 (fr->bcoultab || fr->bvdwtab));
2928 negp_pp = ir->opts.ngener - ir->nwall;
2930 if (!needGroupSchemeTables)
2932 bSomeNormalNbListsAreInUse = TRUE;
2937 bSomeNormalNbListsAreInUse = FALSE;
2938 for (egi = 0; egi < negp_pp; egi++)
2940 for (egj = egi; egj < negp_pp; egj++)
2942 egp_flags = ir->opts.egp_flags[GID(egi, egj, ir->opts.ngener)];
2943 if (!(egp_flags & EGP_EXCL))
2945 if (egp_flags & EGP_TABLE)
2951 bSomeNormalNbListsAreInUse = TRUE;
2956 if (bSomeNormalNbListsAreInUse)
2958 fr->nnblists = negptable + 1;
2962 fr->nnblists = negptable;
2964 if (fr->nnblists > 1)
2966 snew(fr->gid2nblists, ir->opts.ngener*ir->opts.ngener);
2970 snew(fr->nblists, fr->nnblists);
2972 /* This code automatically gives table length tabext without cut-off's,
2973 * in that case grompp should already have checked that we do not need
2974 * normal tables and we only generate tables for 1-4 interactions.
2976 rtab = ir->rlist + ir->tabext;
2978 if (needGroupSchemeTables)
2980 /* make tables for ordinary interactions */
2981 if (bSomeNormalNbListsAreInUse)
2983 make_nbf_tables(fp, fr, rtab, tabfn, nullptr, nullptr, &fr->nblists[0]);
2992 /* Read the special tables for certain energy group pairs */
2993 nm_ind = mtop->groups.grps[egcENER].nm_ind;
2994 for (egi = 0; egi < negp_pp; egi++)
2996 for (egj = egi; egj < negp_pp; egj++)
2998 egp_flags = ir->opts.egp_flags[GID(egi, egj, ir->opts.ngener)];
2999 if ((egp_flags & EGP_TABLE) && !(egp_flags & EGP_EXCL))
3001 if (fr->nnblists > 1)
3003 fr->gid2nblists[GID(egi, egj, ir->opts.ngener)] = m;
3005 /* Read the table file with the two energy groups names appended */
3006 make_nbf_tables(fp, fr, rtab, tabfn,
3007 *mtop->groups.grpname[nm_ind[egi]],
3008 *mtop->groups.grpname[nm_ind[egj]],
3012 else if (fr->nnblists > 1)
3014 fr->gid2nblists[GID(egi, egj, ir->opts.ngener)] = 0;
3021 /* Tables might not be used for the potential modifier
3022 * interactions per se, but we still need them to evaluate
3023 * switch/shift dispersion corrections in this case. */
3024 if (fr->eDispCorr != edispcNO)
3026 fr->dispersionCorrectionTable = makeDispersionCorrectionTable(fp, fr, rtab, tabfn);
3029 /* We want to use unmodified tables for 1-4 coulombic
3030 * interactions, so we must in general have an extra set of
3032 if (gmx_mtop_ftype_count(mtop, F_LJ14) > 0 ||
3033 gmx_mtop_ftype_count(mtop, F_LJC14_Q) > 0 ||
3034 gmx_mtop_ftype_count(mtop, F_LJC_PAIRS_NB) > 0)
3036 fr->pairsTable = make_tables(fp, fr, tabpfn, rtab,
3037 GMX_MAKETABLES_14ONLY);
3041 fr->nwall = ir->nwall;
3042 if (ir->nwall && ir->wall_type == ewtTABLE)
3044 make_wall_tables(fp, ir, tabfn, &mtop->groups, fr);
3049 // Need to catch std::bad_alloc
3050 // TODO Don't need to catch this here, when merging with master branch
3053 fcd->bondtab = make_bonded_tables(fp,
3054 F_TABBONDS, F_TABBONDSNC,
3055 mtop, tabbfnm, "b");
3056 fcd->angletab = make_bonded_tables(fp,
3058 mtop, tabbfnm, "a");
3059 fcd->dihtab = make_bonded_tables(fp,
3061 mtop, tabbfnm, "d");
3063 GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
3069 fprintf(debug, "No fcdata or table file name passed, can not read table, can not do bonded interactions\n");
3073 /* QM/MM initialization if requested
3077 fprintf(stderr, "QM/MM calculation requested.\n");
3080 fr->bQMMM = ir->bQMMM;
3081 fr->qr = mk_QMMMrec();
3083 /* Set all the static charge group info */
3084 fr->cginfo_mb = init_cginfo_mb(fp, mtop, fr, bNoSolvOpt,
3086 &fr->bExcl_IntraCGAll_InterCGNone);
3087 if (DOMAINDECOMP(cr))
3089 fr->cginfo = nullptr;
3093 fr->cginfo = cginfo_expand(mtop->nmolblock, fr->cginfo_mb);
3096 if (!DOMAINDECOMP(cr))
3098 forcerec_set_ranges(fr, ncg_mtop(mtop), ncg_mtop(mtop),
3099 mtop->natoms, mtop->natoms, mtop->natoms);
3102 fr->print_force = print_force;
3105 /* coarse load balancing vars */
3110 /* Initialize neighbor search */
3112 init_ns(fp, cr, fr->ns, fr, mtop);
3114 if (cr->duty & DUTY_PP)
3116 gmx_nonbonded_setup(fr, bGenericKernelOnly);
3119 /* Initialize the thread working data for bonded interactions */
3120 init_bonded_threading(fp, mtop->groups.grps[egcENER].nr,
3121 &fr->bonded_threading);
3123 fr->nthread_ewc = gmx_omp_nthreads_get(emntBonded);
3124 snew(fr->ewc_t, fr->nthread_ewc);
3126 /* fr->ic is used both by verlet and group kernels (to some extent) now */
3127 init_interaction_const(fp, &fr->ic, fr);
3128 init_interaction_const_tables(fp, fr->ic, rtab);
3130 if (fr->cutoff_scheme == ecutsVERLET)
3132 // We checked the cut-offs in grompp, but double-check here.
3133 // We have PME+LJcutoff kernels for rcoulomb>rvdw.
3134 if (EEL_PME_EWALD(ir->coulombtype) && ir->vdwtype == eelCUT)
3136 GMX_RELEASE_ASSERT(ir->rcoulomb >= ir->rvdw, "With Verlet lists and PME we should have rcoulomb>=rvdw");
3140 GMX_RELEASE_ASSERT(ir->rcoulomb == ir->rvdw, "With Verlet lists and no PME rcoulomb and rvdw should be identical");
3143 init_nb_verlet(fp, mdlog, &fr->nbv, bFEP_NonBonded, ir, fr,
3144 cr, nbpu_opt, deviceInfo,
3148 if (ir->eDispCorr != edispcNO)
3150 calc_enervirdiff(fp, ir->eDispCorr, fr);
3154 #define pr_real(fp, r) fprintf(fp, "%s: %e\n",#r, r)
3155 #define pr_int(fp, i) fprintf((fp), "%s: %d\n",#i, i)
3156 #define pr_bool(fp, b) fprintf((fp), "%s: %s\n",#b, gmx::boolToString(b))
3158 void pr_forcerec(FILE *fp, t_forcerec *fr)
3162 pr_real(fp, fr->rlist);
3163 pr_real(fp, fr->rcoulomb);
3164 pr_real(fp, fr->fudgeQQ);
3165 pr_bool(fp, fr->bGrid);
3166 /*pr_int(fp,fr->cg0);
3167 pr_int(fp,fr->hcg);*/
3168 for (i = 0; i < fr->nnblists; i++)
3170 pr_int(fp, fr->nblists[i].table_elec_vdw->n);
3172 pr_real(fp, fr->rcoulomb_switch);
3173 pr_real(fp, fr->rcoulomb);
3178 /* Frees GPU memory and destroys the GPU context.
3180 * Note that this function needs to be called even if GPUs are not used
3181 * in this run because the PME ranks have no knowledge of whether GPUs
3182 * are used or not, but all ranks need to enter the barrier below.
3184 void free_gpu_resources(const t_forcerec *fr,
3185 const t_commrec *cr,
3186 const gmx_device_info_t *deviceInfo)
3188 gmx_bool bIsPPrankUsingGPU;
3189 char gpu_err_str[STRLEN];
3191 bIsPPrankUsingGPU = (cr->duty & DUTY_PP) && fr && fr->nbv && fr->nbv->bUseGPU;
3193 if (bIsPPrankUsingGPU)
3195 /* free nbnxn data in GPU memory */
3196 nbnxn_gpu_free(fr->nbv->gpu_nbv);
3197 /* stop the GPU profiler (only CUDA) */
3201 /* With tMPI we need to wait for all ranks to finish deallocation before
3202 * destroying the CUDA context in free_gpu() as some tMPI ranks may be sharing
3205 * This is not a concern in OpenCL where we use one context per rank which
3206 * is freed in nbnxn_gpu_free().
3208 * Note: it is safe to not call the barrier on the ranks which do not use GPU,
3209 * but it is easier and more futureproof to call it on the whole node.
3212 if (PAR(cr) || MULTISIM(cr))
3214 gmx_barrier_physical_node(cr);
3216 #endif /* GMX_THREAD_MPI */
3218 if (bIsPPrankUsingGPU)
3220 /* uninitialize GPU (by destroying the context) */
3221 if (!free_cuda_gpu(deviceInfo, gpu_err_str))
3223 gmx_warning("On rank %d failed to free GPU #%d: %s",
3224 cr->nodeid, get_current_cuda_gpu_device_id(), gpu_err_str);