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, 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.
46 #include "types/commrec.h"
48 #include "gromacs/math/utilities.h"
52 #include "gmx_fatal.h"
53 #include "gmx_fatal_collective.h"
57 #include "nonbonded.h"
66 #include "md_support.h"
67 #include "md_logging.h"
71 #include "mtop_util.h"
72 #include "nbnxn_simd.h"
73 #include "nbnxn_search.h"
74 #include "nbnxn_atomdata.h"
75 #include "nbnxn_consts.h"
76 #include "gmx_omp_nthreads.h"
77 #include "gmx_detect_hardware.h"
81 /* MSVC definition for __cpuid() */
85 #include "types/nbnxn_cuda_types_ext.h"
86 #include "gpu_utils.h"
87 #include "nbnxn_cuda_data_mgmt.h"
88 #include "pmalloc_cuda.h"
90 t_forcerec *mk_forcerec(void)
100 static void pr_nbfp(FILE *fp, real *nbfp, gmx_bool bBHAM, int atnr)
104 for (i = 0; (i < atnr); i++)
106 for (j = 0; (j < atnr); j++)
108 fprintf(fp, "%2d - %2d", i, j);
111 fprintf(fp, " a=%10g, b=%10g, c=%10g\n", BHAMA(nbfp, atnr, i, j),
112 BHAMB(nbfp, atnr, i, j), BHAMC(nbfp, atnr, i, j)/6.0);
116 fprintf(fp, " c6=%10g, c12=%10g\n", C6(nbfp, atnr, i, j)/6.0,
117 C12(nbfp, atnr, i, j)/12.0);
124 static real *mk_nbfp(const gmx_ffparams_t *idef, gmx_bool bBHAM)
132 snew(nbfp, 3*atnr*atnr);
133 for (i = k = 0; (i < atnr); i++)
135 for (j = 0; (j < atnr); j++, k++)
137 BHAMA(nbfp, atnr, i, j) = idef->iparams[k].bham.a;
138 BHAMB(nbfp, atnr, i, j) = idef->iparams[k].bham.b;
139 /* nbfp now includes the 6.0 derivative prefactor */
140 BHAMC(nbfp, atnr, i, j) = idef->iparams[k].bham.c*6.0;
146 snew(nbfp, 2*atnr*atnr);
147 for (i = k = 0; (i < atnr); i++)
149 for (j = 0; (j < atnr); j++, k++)
151 /* nbfp now includes the 6.0/12.0 derivative prefactors */
152 C6(nbfp, atnr, i, j) = idef->iparams[k].lj.c6*6.0;
153 C12(nbfp, atnr, i, j) = idef->iparams[k].lj.c12*12.0;
161 static real *make_ljpme_c6grid(const gmx_ffparams_t *idef, t_forcerec *fr)
164 real c6, c6i, c6j, c12i, c12j, epsi, epsj, sigmai, sigmaj;
167 /* For LJ-PME simulations, we correct the energies with the reciprocal space
168 * inside of the cut-off. To do this the non-bonded kernels needs to have
169 * access to the C6-values used on the reciprocal grid in pme.c
173 snew(grid, 2*atnr*atnr);
174 for (i = k = 0; (i < atnr); i++)
176 for (j = 0; (j < atnr); j++, k++)
178 c6i = idef->iparams[i*(atnr+1)].lj.c6;
179 c12i = idef->iparams[i*(atnr+1)].lj.c12;
180 c6j = idef->iparams[j*(atnr+1)].lj.c6;
181 c12j = idef->iparams[j*(atnr+1)].lj.c12;
182 c6 = sqrt(c6i * c6j);
183 if (fr->ljpme_combination_rule == eljpmeLB
184 && !gmx_numzero(c6) && !gmx_numzero(c12i) && !gmx_numzero(c12j))
186 sigmai = pow(c12i / c6i, 1.0/6.0);
187 sigmaj = pow(c12j / c6j, 1.0/6.0);
188 epsi = c6i * c6i / c12i;
189 epsj = c6j * c6j / c12j;
190 c6 = sqrt(epsi * epsj) * pow(0.5*(sigmai+sigmaj), 6);
192 /* Store the elements at the same relative positions as C6 in nbfp in order
193 * to simplify access in the kernels
195 grid[2*(atnr*i+j)] = c6*6.0;
201 static real *mk_nbfp_combination_rule(const gmx_ffparams_t *idef, int comb_rule)
205 real c6i, c6j, c12i, c12j, epsi, epsj, sigmai, sigmaj;
209 snew(nbfp, 2*atnr*atnr);
210 for (i = 0; i < atnr; ++i)
212 for (j = 0; j < atnr; ++j)
214 c6i = idef->iparams[i*(atnr+1)].lj.c6;
215 c12i = idef->iparams[i*(atnr+1)].lj.c12;
216 c6j = idef->iparams[j*(atnr+1)].lj.c6;
217 c12j = idef->iparams[j*(atnr+1)].lj.c12;
218 c6 = sqrt(c6i * c6j);
219 c12 = sqrt(c12i * c12j);
220 if (comb_rule == eCOMB_ARITHMETIC
221 && !gmx_numzero(c6) && !gmx_numzero(c12))
223 sigmai = pow(c12i / c6i, 1.0/6.0);
224 sigmaj = pow(c12j / c6j, 1.0/6.0);
225 epsi = c6i * c6i / c12i;
226 epsj = c6j * c6j / c12j;
227 c6 = epsi * epsj * pow(0.5*(sigmai+sigmaj), 6);
228 c12 = epsi * epsj * pow(0.5*(sigmai+sigmaj), 12);
230 C6(nbfp, atnr, i, j) = c6*6.0;
231 C12(nbfp, atnr, i, j) = c12*12.0;
237 /* This routine sets fr->solvent_opt to the most common solvent in the
238 * system, e.g. esolSPC or esolTIP4P. It will also mark each charge group in
239 * the fr->solvent_type array with the correct type (or esolNO).
241 * Charge groups that fulfill the conditions but are not identical to the
242 * most common one will be marked as esolNO in the solvent_type array.
244 * TIP3p is identical to SPC for these purposes, so we call it
245 * SPC in the arrays (Apologies to Bill Jorgensen ;-)
247 * NOTE: QM particle should not
248 * become an optimized solvent. Not even if there is only one charge
258 } solvent_parameters_t;
261 check_solvent_cg(const gmx_moltype_t *molt,
264 const unsigned char *qm_grpnr,
265 const t_grps *qm_grps,
267 int *n_solvent_parameters,
268 solvent_parameters_t **solvent_parameters_p,
272 const t_blocka *excl;
279 real tmp_charge[4] = { 0.0 }; /* init to zero to make gcc4.8 happy */
280 int tmp_vdwtype[4] = { 0 }; /* init to zero to make gcc4.8 happy */
283 solvent_parameters_t *solvent_parameters;
285 /* We use a list with parameters for each solvent type.
286 * Every time we discover a new molecule that fulfills the basic
287 * conditions for a solvent we compare with the previous entries
288 * in these lists. If the parameters are the same we just increment
289 * the counter for that type, and otherwise we create a new type
290 * based on the current molecule.
292 * Once we've finished going through all molecules we check which
293 * solvent is most common, and mark all those molecules while we
294 * clear the flag on all others.
297 solvent_parameters = *solvent_parameters_p;
299 /* Mark the cg first as non optimized */
302 /* Check if this cg has no exclusions with atoms in other charge groups
303 * and all atoms inside the charge group excluded.
304 * We only have 3 or 4 atom solvent loops.
306 if (GET_CGINFO_EXCL_INTER(cginfo) ||
307 !GET_CGINFO_EXCL_INTRA(cginfo))
312 /* Get the indices of the first atom in this charge group */
313 j0 = molt->cgs.index[cg0];
314 j1 = molt->cgs.index[cg0+1];
316 /* Number of atoms in our molecule */
322 "Moltype '%s': there are %d atoms in this charge group\n",
326 /* Check if it could be an SPC (3 atoms) or TIP4p (4) water,
329 if (nj < 3 || nj > 4)
334 /* Check if we are doing QM on this group */
336 if (qm_grpnr != NULL)
338 for (j = j0; j < j1 && !qm; j++)
340 qm = (qm_grpnr[j] < qm_grps->nr - 1);
343 /* Cannot use solvent optimization with QM */
349 atom = molt->atoms.atom;
351 /* Still looks like a solvent, time to check parameters */
353 /* If it is perturbed (free energy) we can't use the solvent loops,
354 * so then we just skip to the next molecule.
358 for (j = j0; j < j1 && !perturbed; j++)
360 perturbed = PERTURBED(atom[j]);
368 /* Now it's only a question if the VdW and charge parameters
369 * are OK. Before doing the check we compare and see if they are
370 * identical to a possible previous solvent type.
371 * First we assign the current types and charges.
373 for (j = 0; j < nj; j++)
375 tmp_vdwtype[j] = atom[j0+j].type;
376 tmp_charge[j] = atom[j0+j].q;
379 /* Does it match any previous solvent type? */
380 for (k = 0; k < *n_solvent_parameters; k++)
385 /* We can only match SPC with 3 atoms and TIP4p with 4 atoms */
386 if ( (solvent_parameters[k].model == esolSPC && nj != 3) ||
387 (solvent_parameters[k].model == esolTIP4P && nj != 4) )
392 /* Check that types & charges match for all atoms in molecule */
393 for (j = 0; j < nj && match == TRUE; j++)
395 if (tmp_vdwtype[j] != solvent_parameters[k].vdwtype[j])
399 if (tmp_charge[j] != solvent_parameters[k].charge[j])
406 /* Congratulations! We have a matched solvent.
407 * Flag it with this type for later processing.
410 solvent_parameters[k].count += nmol;
412 /* We are done with this charge group */
417 /* If we get here, we have a tentative new solvent type.
418 * Before we add it we must check that it fulfills the requirements
419 * of the solvent optimized loops. First determine which atoms have
422 for (j = 0; j < nj; j++)
425 tjA = tmp_vdwtype[j];
427 /* Go through all other tpes and see if any have non-zero
428 * VdW parameters when combined with this one.
430 for (k = 0; k < fr->ntype && (has_vdw[j] == FALSE); k++)
432 /* We already checked that the atoms weren't perturbed,
433 * so we only need to check state A now.
437 has_vdw[j] = (has_vdw[j] ||
438 (BHAMA(fr->nbfp, fr->ntype, tjA, k) != 0.0) ||
439 (BHAMB(fr->nbfp, fr->ntype, tjA, k) != 0.0) ||
440 (BHAMC(fr->nbfp, fr->ntype, tjA, k) != 0.0));
445 has_vdw[j] = (has_vdw[j] ||
446 (C6(fr->nbfp, fr->ntype, tjA, k) != 0.0) ||
447 (C12(fr->nbfp, fr->ntype, tjA, k) != 0.0));
452 /* Now we know all we need to make the final check and assignment. */
456 * For this we require thatn all atoms have charge,
457 * the charges on atom 2 & 3 should be the same, and only
458 * atom 1 might have VdW.
460 if (has_vdw[1] == FALSE &&
461 has_vdw[2] == FALSE &&
462 tmp_charge[0] != 0 &&
463 tmp_charge[1] != 0 &&
464 tmp_charge[2] == tmp_charge[1])
466 srenew(solvent_parameters, *n_solvent_parameters+1);
467 solvent_parameters[*n_solvent_parameters].model = esolSPC;
468 solvent_parameters[*n_solvent_parameters].count = nmol;
469 for (k = 0; k < 3; k++)
471 solvent_parameters[*n_solvent_parameters].vdwtype[k] = tmp_vdwtype[k];
472 solvent_parameters[*n_solvent_parameters].charge[k] = tmp_charge[k];
475 *cg_sp = *n_solvent_parameters;
476 (*n_solvent_parameters)++;
481 /* Or could it be a TIP4P?
482 * For this we require thatn atoms 2,3,4 have charge, but not atom 1.
483 * Only atom 1 mght have VdW.
485 if (has_vdw[1] == FALSE &&
486 has_vdw[2] == FALSE &&
487 has_vdw[3] == FALSE &&
488 tmp_charge[0] == 0 &&
489 tmp_charge[1] != 0 &&
490 tmp_charge[2] == tmp_charge[1] &&
493 srenew(solvent_parameters, *n_solvent_parameters+1);
494 solvent_parameters[*n_solvent_parameters].model = esolTIP4P;
495 solvent_parameters[*n_solvent_parameters].count = nmol;
496 for (k = 0; k < 4; k++)
498 solvent_parameters[*n_solvent_parameters].vdwtype[k] = tmp_vdwtype[k];
499 solvent_parameters[*n_solvent_parameters].charge[k] = tmp_charge[k];
502 *cg_sp = *n_solvent_parameters;
503 (*n_solvent_parameters)++;
507 *solvent_parameters_p = solvent_parameters;
511 check_solvent(FILE * fp,
512 const gmx_mtop_t * mtop,
514 cginfo_mb_t *cginfo_mb)
517 const t_block * mols;
518 const gmx_moltype_t *molt;
519 int mb, mol, cg_mol, at_offset, cg_offset, am, cgm, i, nmol_ch, nmol;
520 int n_solvent_parameters;
521 solvent_parameters_t *solvent_parameters;
527 fprintf(debug, "Going to determine what solvent types we have.\n");
532 n_solvent_parameters = 0;
533 solvent_parameters = NULL;
534 /* Allocate temporary array for solvent type */
535 snew(cg_sp, mtop->nmolblock);
539 for (mb = 0; mb < mtop->nmolblock; mb++)
541 molt = &mtop->moltype[mtop->molblock[mb].type];
543 /* Here we have to loop over all individual molecules
544 * because we need to check for QMMM particles.
546 snew(cg_sp[mb], cginfo_mb[mb].cg_mod);
547 nmol_ch = cginfo_mb[mb].cg_mod/cgs->nr;
548 nmol = mtop->molblock[mb].nmol/nmol_ch;
549 for (mol = 0; mol < nmol_ch; mol++)
552 am = mol*cgs->index[cgs->nr];
553 for (cg_mol = 0; cg_mol < cgs->nr; cg_mol++)
555 check_solvent_cg(molt, cg_mol, nmol,
556 mtop->groups.grpnr[egcQMMM] ?
557 mtop->groups.grpnr[egcQMMM]+at_offset+am : 0,
558 &mtop->groups.grps[egcQMMM],
560 &n_solvent_parameters, &solvent_parameters,
561 cginfo_mb[mb].cginfo[cgm+cg_mol],
562 &cg_sp[mb][cgm+cg_mol]);
565 cg_offset += cgs->nr;
566 at_offset += cgs->index[cgs->nr];
569 /* Puh! We finished going through all charge groups.
570 * Now find the most common solvent model.
573 /* Most common solvent this far */
575 for (i = 0; i < n_solvent_parameters; i++)
578 solvent_parameters[i].count > solvent_parameters[bestsp].count)
586 bestsol = solvent_parameters[bestsp].model;
593 #ifdef DISABLE_WATER_NLIST
598 for (mb = 0; mb < mtop->nmolblock; mb++)
600 cgs = &mtop->moltype[mtop->molblock[mb].type].cgs;
601 nmol = (mtop->molblock[mb].nmol*cgs->nr)/cginfo_mb[mb].cg_mod;
602 for (i = 0; i < cginfo_mb[mb].cg_mod; i++)
604 if (cg_sp[mb][i] == bestsp)
606 SET_CGINFO_SOLOPT(cginfo_mb[mb].cginfo[i], bestsol);
611 SET_CGINFO_SOLOPT(cginfo_mb[mb].cginfo[i], esolNO);
618 if (bestsol != esolNO && fp != NULL)
620 fprintf(fp, "\nEnabling %s-like water optimization for %d molecules.\n\n",
622 solvent_parameters[bestsp].count);
625 sfree(solvent_parameters);
626 fr->solvent_opt = bestsol;
630 acNONE = 0, acCONSTRAINT, acSETTLE
633 static cginfo_mb_t *init_cginfo_mb(FILE *fplog, const gmx_mtop_t *mtop,
634 t_forcerec *fr, gmx_bool bNoSolvOpt,
635 gmx_bool *bFEP_NonBonded,
636 gmx_bool *bExcl_IntraCGAll_InterCGNone)
639 const t_blocka *excl;
640 const gmx_moltype_t *molt;
641 const gmx_molblock_t *molb;
642 cginfo_mb_t *cginfo_mb;
645 int cg_offset, a_offset, cgm, am;
646 int mb, m, ncg_tot, cg, a0, a1, gid, ai, j, aj, excl_nalloc;
650 gmx_bool bId, *bExcl, bExclIntraAll, bExclInter, bHaveVDW, bHaveQ, bHavePerturbedAtoms;
652 ncg_tot = ncg_mtop(mtop);
653 snew(cginfo_mb, mtop->nmolblock);
655 snew(type_VDW, fr->ntype);
656 for (ai = 0; ai < fr->ntype; ai++)
658 type_VDW[ai] = FALSE;
659 for (j = 0; j < fr->ntype; j++)
661 type_VDW[ai] = type_VDW[ai] ||
663 C6(fr->nbfp, fr->ntype, ai, j) != 0 ||
664 C12(fr->nbfp, fr->ntype, ai, j) != 0;
668 *bFEP_NonBonded = FALSE;
669 *bExcl_IntraCGAll_InterCGNone = TRUE;
672 snew(bExcl, excl_nalloc);
675 for (mb = 0; mb < mtop->nmolblock; mb++)
677 molb = &mtop->molblock[mb];
678 molt = &mtop->moltype[molb->type];
682 /* Check if the cginfo is identical for all molecules in this block.
683 * If so, we only need an array of the size of one molecule.
684 * Otherwise we make an array of #mol times #cgs per molecule.
688 for (m = 0; m < molb->nmol; m++)
690 am = m*cgs->index[cgs->nr];
691 for (cg = 0; cg < cgs->nr; cg++)
694 a1 = cgs->index[cg+1];
695 if (ggrpnr(&mtop->groups, egcENER, a_offset+am+a0) !=
696 ggrpnr(&mtop->groups, egcENER, a_offset +a0))
700 if (mtop->groups.grpnr[egcQMMM] != NULL)
702 for (ai = a0; ai < a1; ai++)
704 if (mtop->groups.grpnr[egcQMMM][a_offset+am+ai] !=
705 mtop->groups.grpnr[egcQMMM][a_offset +ai])
714 cginfo_mb[mb].cg_start = cg_offset;
715 cginfo_mb[mb].cg_end = cg_offset + molb->nmol*cgs->nr;
716 cginfo_mb[mb].cg_mod = (bId ? 1 : molb->nmol)*cgs->nr;
717 snew(cginfo_mb[mb].cginfo, cginfo_mb[mb].cg_mod);
718 cginfo = cginfo_mb[mb].cginfo;
720 /* Set constraints flags for constrained atoms */
721 snew(a_con, molt->atoms.nr);
722 for (ftype = 0; ftype < F_NRE; ftype++)
724 if (interaction_function[ftype].flags & IF_CONSTRAINT)
729 for (ia = 0; ia < molt->ilist[ftype].nr; ia += 1+nral)
733 for (a = 0; a < nral; a++)
735 a_con[molt->ilist[ftype].iatoms[ia+1+a]] =
736 (ftype == F_SETTLE ? acSETTLE : acCONSTRAINT);
742 for (m = 0; m < (bId ? 1 : molb->nmol); m++)
745 am = m*cgs->index[cgs->nr];
746 for (cg = 0; cg < cgs->nr; cg++)
749 a1 = cgs->index[cg+1];
751 /* Store the energy group in cginfo */
752 gid = ggrpnr(&mtop->groups, egcENER, a_offset+am+a0);
753 SET_CGINFO_GID(cginfo[cgm+cg], gid);
755 /* Check the intra/inter charge group exclusions */
756 if (a1-a0 > excl_nalloc)
758 excl_nalloc = a1 - a0;
759 srenew(bExcl, excl_nalloc);
761 /* bExclIntraAll: all intra cg interactions excluded
762 * bExclInter: any inter cg interactions excluded
764 bExclIntraAll = TRUE;
768 bHavePerturbedAtoms = FALSE;
769 for (ai = a0; ai < a1; ai++)
771 /* Check VDW and electrostatic interactions */
772 bHaveVDW = bHaveVDW || (type_VDW[molt->atoms.atom[ai].type] ||
773 type_VDW[molt->atoms.atom[ai].typeB]);
774 bHaveQ = bHaveQ || (molt->atoms.atom[ai].q != 0 ||
775 molt->atoms.atom[ai].qB != 0);
777 bHavePerturbedAtoms = bHavePerturbedAtoms || (PERTURBED(molt->atoms.atom[ai]) != 0);
779 /* Clear the exclusion list for atom ai */
780 for (aj = a0; aj < a1; aj++)
782 bExcl[aj-a0] = FALSE;
784 /* Loop over all the exclusions of atom ai */
785 for (j = excl->index[ai]; j < excl->index[ai+1]; j++)
788 if (aj < a0 || aj >= a1)
797 /* Check if ai excludes a0 to a1 */
798 for (aj = a0; aj < a1; aj++)
802 bExclIntraAll = FALSE;
809 SET_CGINFO_CONSTR(cginfo[cgm+cg]);
812 SET_CGINFO_SETTLE(cginfo[cgm+cg]);
820 SET_CGINFO_EXCL_INTRA(cginfo[cgm+cg]);
824 SET_CGINFO_EXCL_INTER(cginfo[cgm+cg]);
826 if (a1 - a0 > MAX_CHARGEGROUP_SIZE)
828 /* The size in cginfo is currently only read with DD */
829 gmx_fatal(FARGS, "A charge group has size %d which is larger than the limit of %d atoms", a1-a0, MAX_CHARGEGROUP_SIZE);
833 SET_CGINFO_HAS_VDW(cginfo[cgm+cg]);
837 SET_CGINFO_HAS_Q(cginfo[cgm+cg]);
839 if (bHavePerturbedAtoms && fr->efep != efepNO)
841 SET_CGINFO_FEP(cginfo[cgm+cg]);
842 *bFEP_NonBonded = TRUE;
844 /* Store the charge group size */
845 SET_CGINFO_NATOMS(cginfo[cgm+cg], a1-a0);
847 if (!bExclIntraAll || bExclInter)
849 *bExcl_IntraCGAll_InterCGNone = FALSE;
856 cg_offset += molb->nmol*cgs->nr;
857 a_offset += molb->nmol*cgs->index[cgs->nr];
861 /* the solvent optimizer is called after the QM is initialized,
862 * because we don't want to have the QM subsystemto become an
866 check_solvent(fplog, mtop, fr, cginfo_mb);
868 if (getenv("GMX_NO_SOLV_OPT"))
872 fprintf(fplog, "Found environment variable GMX_NO_SOLV_OPT.\n"
873 "Disabling all solvent optimization\n");
875 fr->solvent_opt = esolNO;
879 fr->solvent_opt = esolNO;
881 if (!fr->solvent_opt)
883 for (mb = 0; mb < mtop->nmolblock; mb++)
885 for (cg = 0; cg < cginfo_mb[mb].cg_mod; cg++)
887 SET_CGINFO_SOLOPT(cginfo_mb[mb].cginfo[cg], esolNO);
895 static int *cginfo_expand(int nmb, cginfo_mb_t *cgi_mb)
900 ncg = cgi_mb[nmb-1].cg_end;
903 for (cg = 0; cg < ncg; cg++)
905 while (cg >= cgi_mb[mb].cg_end)
910 cgi_mb[mb].cginfo[(cg - cgi_mb[mb].cg_start) % cgi_mb[mb].cg_mod];
916 static void set_chargesum(FILE *log, t_forcerec *fr, const gmx_mtop_t *mtop)
918 /*This now calculates sum for q and c6*/
919 double qsum, q2sum, q, c6sum, c6;
921 const t_atoms *atoms;
926 for (mb = 0; mb < mtop->nmolblock; mb++)
928 nmol = mtop->molblock[mb].nmol;
929 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
930 for (i = 0; i < atoms->nr; i++)
932 q = atoms->atom[i].q;
935 c6 = mtop->ffparams.iparams[atoms->atom[i].type*(mtop->ffparams.atnr+1)].lj.c6;
940 fr->q2sum[0] = q2sum;
941 fr->c6sum[0] = c6sum;
943 if (fr->efep != efepNO)
948 for (mb = 0; mb < mtop->nmolblock; mb++)
950 nmol = mtop->molblock[mb].nmol;
951 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
952 for (i = 0; i < atoms->nr; i++)
954 q = atoms->atom[i].qB;
957 c6 = mtop->ffparams.iparams[atoms->atom[i].typeB*(mtop->ffparams.atnr+1)].lj.c6;
961 fr->q2sum[1] = q2sum;
962 fr->c6sum[1] = c6sum;
967 fr->qsum[1] = fr->qsum[0];
968 fr->q2sum[1] = fr->q2sum[0];
969 fr->c6sum[1] = fr->c6sum[0];
973 if (fr->efep == efepNO)
975 fprintf(log, "System total charge: %.3f\n", fr->qsum[0]);
979 fprintf(log, "System total charge, top. A: %.3f top. B: %.3f\n",
980 fr->qsum[0], fr->qsum[1]);
985 void update_forcerec(t_forcerec *fr, matrix box)
987 if (fr->eeltype == eelGRF)
989 calc_rffac(NULL, fr->eeltype, fr->epsilon_r, fr->epsilon_rf,
990 fr->rcoulomb, fr->temp, fr->zsquare, box,
991 &fr->kappa, &fr->k_rf, &fr->c_rf);
995 void set_avcsixtwelve(FILE *fplog, t_forcerec *fr, const gmx_mtop_t *mtop)
997 const t_atoms *atoms, *atoms_tpi;
998 const t_blocka *excl;
999 int mb, nmol, nmolc, i, j, tpi, tpj, j1, j2, k, n, nexcl, q;
1000 gmx_int64_t npair, npair_ij, tmpi, tmpj;
1001 double csix, ctwelve;
1002 int ntp, *typecount;
1005 real *nbfp_comb = NULL;
1011 /* For LJ-PME, we want to correct for the difference between the
1012 * actual C6 values and the C6 values used by the LJ-PME based on
1013 * combination rules. */
1015 if (EVDW_PME(fr->vdwtype))
1017 nbfp_comb = mk_nbfp_combination_rule(&mtop->ffparams,
1018 (fr->ljpme_combination_rule == eljpmeLB) ? eCOMB_ARITHMETIC : eCOMB_GEOMETRIC);
1019 for (tpi = 0; tpi < ntp; ++tpi)
1021 for (tpj = 0; tpj < ntp; ++tpj)
1023 C6(nbfp_comb, ntp, tpi, tpj) =
1024 C6(nbfp, ntp, tpi, tpj) - C6(nbfp_comb, ntp, tpi, tpj);
1025 C12(nbfp_comb, ntp, tpi, tpj) = C12(nbfp, ntp, tpi, tpj);
1030 for (q = 0; q < (fr->efep == efepNO ? 1 : 2); q++)
1038 /* Count the types so we avoid natoms^2 operations */
1039 snew(typecount, ntp);
1040 gmx_mtop_count_atomtypes(mtop, q, typecount);
1042 for (tpi = 0; tpi < ntp; tpi++)
1044 for (tpj = tpi; tpj < ntp; tpj++)
1046 tmpi = typecount[tpi];
1047 tmpj = typecount[tpj];
1050 npair_ij = tmpi*tmpj;
1054 npair_ij = tmpi*(tmpi - 1)/2;
1058 /* nbfp now includes the 6.0 derivative prefactor */
1059 csix += npair_ij*BHAMC(nbfp, ntp, tpi, tpj)/6.0;
1063 /* nbfp now includes the 6.0/12.0 derivative prefactors */
1064 csix += npair_ij* C6(nbfp, ntp, tpi, tpj)/6.0;
1065 ctwelve += npair_ij* C12(nbfp, ntp, tpi, tpj)/12.0;
1071 /* Subtract the excluded pairs.
1072 * The main reason for substracting exclusions is that in some cases
1073 * some combinations might never occur and the parameters could have
1074 * any value. These unused values should not influence the dispersion
1077 for (mb = 0; mb < mtop->nmolblock; mb++)
1079 nmol = mtop->molblock[mb].nmol;
1080 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
1081 excl = &mtop->moltype[mtop->molblock[mb].type].excls;
1082 for (i = 0; (i < atoms->nr); i++)
1086 tpi = atoms->atom[i].type;
1090 tpi = atoms->atom[i].typeB;
1092 j1 = excl->index[i];
1093 j2 = excl->index[i+1];
1094 for (j = j1; j < j2; j++)
1101 tpj = atoms->atom[k].type;
1105 tpj = atoms->atom[k].typeB;
1109 /* nbfp now includes the 6.0 derivative prefactor */
1110 csix -= nmol*BHAMC(nbfp, ntp, tpi, tpj)/6.0;
1114 /* nbfp now includes the 6.0/12.0 derivative prefactors */
1115 csix -= nmol*C6 (nbfp, ntp, tpi, tpj)/6.0;
1116 ctwelve -= nmol*C12(nbfp, ntp, tpi, tpj)/12.0;
1126 /* Only correct for the interaction of the test particle
1127 * with the rest of the system.
1130 &mtop->moltype[mtop->molblock[mtop->nmolblock-1].type].atoms;
1133 for (mb = 0; mb < mtop->nmolblock; mb++)
1135 nmol = mtop->molblock[mb].nmol;
1136 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
1137 for (j = 0; j < atoms->nr; j++)
1140 /* Remove the interaction of the test charge group
1143 if (mb == mtop->nmolblock-1)
1147 if (mb == 0 && nmol == 1)
1149 gmx_fatal(FARGS, "Old format tpr with TPI, please generate a new tpr file");
1154 tpj = atoms->atom[j].type;
1158 tpj = atoms->atom[j].typeB;
1160 for (i = 0; i < fr->n_tpi; i++)
1164 tpi = atoms_tpi->atom[i].type;
1168 tpi = atoms_tpi->atom[i].typeB;
1172 /* nbfp now includes the 6.0 derivative prefactor */
1173 csix += nmolc*BHAMC(nbfp, ntp, tpi, tpj)/6.0;
1177 /* nbfp now includes the 6.0/12.0 derivative prefactors */
1178 csix += nmolc*C6 (nbfp, ntp, tpi, tpj)/6.0;
1179 ctwelve += nmolc*C12(nbfp, ntp, tpi, tpj)/12.0;
1186 if (npair - nexcl <= 0 && fplog)
1188 fprintf(fplog, "\nWARNING: There are no atom pairs for dispersion correction\n\n");
1194 csix /= npair - nexcl;
1195 ctwelve /= npair - nexcl;
1199 fprintf(debug, "Counted %d exclusions\n", nexcl);
1200 fprintf(debug, "Average C6 parameter is: %10g\n", (double)csix);
1201 fprintf(debug, "Average C12 parameter is: %10g\n", (double)ctwelve);
1203 fr->avcsix[q] = csix;
1204 fr->avctwelve[q] = ctwelve;
1207 if (EVDW_PME(fr->vdwtype))
1214 if (fr->eDispCorr == edispcAllEner ||
1215 fr->eDispCorr == edispcAllEnerPres)
1217 fprintf(fplog, "Long Range LJ corr.: <C6> %10.4e, <C12> %10.4e\n",
1218 fr->avcsix[0], fr->avctwelve[0]);
1222 fprintf(fplog, "Long Range LJ corr.: <C6> %10.4e\n", fr->avcsix[0]);
1228 static void set_bham_b_max(FILE *fplog, t_forcerec *fr,
1229 const gmx_mtop_t *mtop)
1231 const t_atoms *at1, *at2;
1232 int mt1, mt2, i, j, tpi, tpj, ntypes;
1238 fprintf(fplog, "Determining largest Buckingham b parameter for table\n");
1245 for (mt1 = 0; mt1 < mtop->nmoltype; mt1++)
1247 at1 = &mtop->moltype[mt1].atoms;
1248 for (i = 0; (i < at1->nr); i++)
1250 tpi = at1->atom[i].type;
1253 gmx_fatal(FARGS, "Atomtype[%d] = %d, maximum = %d", i, tpi, ntypes);
1256 for (mt2 = mt1; mt2 < mtop->nmoltype; mt2++)
1258 at2 = &mtop->moltype[mt2].atoms;
1259 for (j = 0; (j < at2->nr); j++)
1261 tpj = at2->atom[j].type;
1264 gmx_fatal(FARGS, "Atomtype[%d] = %d, maximum = %d", j, tpj, ntypes);
1266 b = BHAMB(nbfp, ntypes, tpi, tpj);
1267 if (b > fr->bham_b_max)
1271 if ((b < bmin) || (bmin == -1))
1281 fprintf(fplog, "Buckingham b parameters, min: %g, max: %g\n",
1282 bmin, fr->bham_b_max);
1286 static void make_nbf_tables(FILE *fp, const output_env_t oenv,
1287 t_forcerec *fr, real rtab,
1288 const t_commrec *cr,
1289 const char *tabfn, char *eg1, char *eg2,
1299 fprintf(debug, "No table file name passed, can not read table, can not do non-bonded interactions\n");
1304 sprintf(buf, "%s", tabfn);
1307 /* Append the two energy group names */
1308 sprintf(buf + strlen(tabfn) - strlen(ftp2ext(efXVG)) - 1, "_%s_%s.%s",
1309 eg1, eg2, ftp2ext(efXVG));
1311 nbl->table_elec_vdw = make_tables(fp, oenv, fr, MASTER(cr), buf, rtab, 0);
1312 /* Copy the contents of the table to separate coulomb and LJ tables too,
1313 * to improve cache performance.
1315 /* For performance reasons we want
1316 * the table data to be aligned to 16-byte. The pointers could be freed
1317 * but currently aren't.
1319 nbl->table_elec.interaction = GMX_TABLE_INTERACTION_ELEC;
1320 nbl->table_elec.format = nbl->table_elec_vdw.format;
1321 nbl->table_elec.r = nbl->table_elec_vdw.r;
1322 nbl->table_elec.n = nbl->table_elec_vdw.n;
1323 nbl->table_elec.scale = nbl->table_elec_vdw.scale;
1324 nbl->table_elec.scale_exp = nbl->table_elec_vdw.scale_exp;
1325 nbl->table_elec.formatsize = nbl->table_elec_vdw.formatsize;
1326 nbl->table_elec.ninteractions = 1;
1327 nbl->table_elec.stride = nbl->table_elec.formatsize * nbl->table_elec.ninteractions;
1328 snew_aligned(nbl->table_elec.data, nbl->table_elec.stride*(nbl->table_elec.n+1), 32);
1330 nbl->table_vdw.interaction = GMX_TABLE_INTERACTION_VDWREP_VDWDISP;
1331 nbl->table_vdw.format = nbl->table_elec_vdw.format;
1332 nbl->table_vdw.r = nbl->table_elec_vdw.r;
1333 nbl->table_vdw.n = nbl->table_elec_vdw.n;
1334 nbl->table_vdw.scale = nbl->table_elec_vdw.scale;
1335 nbl->table_vdw.scale_exp = nbl->table_elec_vdw.scale_exp;
1336 nbl->table_vdw.formatsize = nbl->table_elec_vdw.formatsize;
1337 nbl->table_vdw.ninteractions = 2;
1338 nbl->table_vdw.stride = nbl->table_vdw.formatsize * nbl->table_vdw.ninteractions;
1339 snew_aligned(nbl->table_vdw.data, nbl->table_vdw.stride*(nbl->table_vdw.n+1), 32);
1341 for (i = 0; i <= nbl->table_elec_vdw.n; i++)
1343 for (j = 0; j < 4; j++)
1345 nbl->table_elec.data[4*i+j] = nbl->table_elec_vdw.data[12*i+j];
1347 for (j = 0; j < 8; j++)
1349 nbl->table_vdw.data[8*i+j] = nbl->table_elec_vdw.data[12*i+4+j];
1354 static void count_tables(int ftype1, int ftype2, const gmx_mtop_t *mtop,
1355 int *ncount, int **count)
1357 const gmx_moltype_t *molt;
1359 int mt, ftype, stride, i, j, tabnr;
1361 for (mt = 0; mt < mtop->nmoltype; mt++)
1363 molt = &mtop->moltype[mt];
1364 for (ftype = 0; ftype < F_NRE; ftype++)
1366 if (ftype == ftype1 || ftype == ftype2)
1368 il = &molt->ilist[ftype];
1369 stride = 1 + NRAL(ftype);
1370 for (i = 0; i < il->nr; i += stride)
1372 tabnr = mtop->ffparams.iparams[il->iatoms[i]].tab.table;
1375 gmx_fatal(FARGS, "A bonded table number is smaller than 0: %d\n", tabnr);
1377 if (tabnr >= *ncount)
1379 srenew(*count, tabnr+1);
1380 for (j = *ncount; j < tabnr+1; j++)
1393 static bondedtable_t *make_bonded_tables(FILE *fplog,
1394 int ftype1, int ftype2,
1395 const gmx_mtop_t *mtop,
1396 const char *basefn, const char *tabext)
1398 int i, ncount, *count;
1406 count_tables(ftype1, ftype2, mtop, &ncount, &count);
1411 for (i = 0; i < ncount; i++)
1415 sprintf(tabfn, "%s", basefn);
1416 sprintf(tabfn + strlen(basefn) - strlen(ftp2ext(efXVG)) - 1, "_%s%d.%s",
1417 tabext, i, ftp2ext(efXVG));
1418 tab[i] = make_bonded_table(fplog, tabfn, NRAL(ftype1)-2);
1427 void forcerec_set_ranges(t_forcerec *fr,
1428 int ncg_home, int ncg_force,
1430 int natoms_force_constr, int natoms_f_novirsum)
1435 /* fr->ncg_force is unused in the standard code,
1436 * but it can be useful for modified code dealing with charge groups.
1438 fr->ncg_force = ncg_force;
1439 fr->natoms_force = natoms_force;
1440 fr->natoms_force_constr = natoms_force_constr;
1442 if (fr->natoms_force_constr > fr->nalloc_force)
1444 fr->nalloc_force = over_alloc_dd(fr->natoms_force_constr);
1448 srenew(fr->f_twin, fr->nalloc_force);
1452 if (fr->bF_NoVirSum)
1454 fr->f_novirsum_n = natoms_f_novirsum;
1455 if (fr->f_novirsum_n > fr->f_novirsum_nalloc)
1457 fr->f_novirsum_nalloc = over_alloc_dd(fr->f_novirsum_n);
1458 srenew(fr->f_novirsum_alloc, fr->f_novirsum_nalloc);
1463 fr->f_novirsum_n = 0;
1467 static real cutoff_inf(real cutoff)
1471 cutoff = GMX_CUTOFF_INF;
1477 static void make_adress_tf_tables(FILE *fp, const output_env_t oenv,
1478 t_forcerec *fr, const t_inputrec *ir,
1479 const char *tabfn, const gmx_mtop_t *mtop,
1487 gmx_fatal(FARGS, "No thermoforce table file given. Use -tabletf to specify a file\n");
1491 snew(fr->atf_tabs, ir->adress->n_tf_grps);
1493 sprintf(buf, "%s", tabfn);
1494 for (i = 0; i < ir->adress->n_tf_grps; i++)
1496 j = ir->adress->tf_table_index[i]; /* get energy group index */
1497 sprintf(buf + strlen(tabfn) - strlen(ftp2ext(efXVG)) - 1, "tf_%s.%s",
1498 *(mtop->groups.grpname[mtop->groups.grps[egcENER].nm_ind[j]]), ftp2ext(efXVG));
1501 fprintf(fp, "loading tf table for energygrp index %d from %s\n", ir->adress->tf_table_index[i], buf);
1503 fr->atf_tabs[i] = make_atf_table(fp, oenv, fr, buf, box);
1508 gmx_bool can_use_allvsall(const t_inputrec *ir, gmx_bool bPrintNote, t_commrec *cr, FILE *fp)
1515 ir->rcoulomb == 0 &&
1517 ir->ePBC == epbcNONE &&
1518 ir->vdwtype == evdwCUT &&
1519 ir->coulombtype == eelCUT &&
1520 ir->efep == efepNO &&
1521 (ir->implicit_solvent == eisNO ||
1522 (ir->implicit_solvent == eisGBSA && (ir->gb_algorithm == egbSTILL ||
1523 ir->gb_algorithm == egbHCT ||
1524 ir->gb_algorithm == egbOBC))) &&
1525 getenv("GMX_NO_ALLVSALL") == NULL
1528 if (bAllvsAll && ir->opts.ngener > 1)
1530 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";
1536 fprintf(stderr, "\n%s\n", note);
1540 fprintf(fp, "\n%s\n", note);
1546 if (bAllvsAll && fp && MASTER(cr))
1548 fprintf(fp, "\nUsing SIMD all-vs-all kernels.\n\n");
1555 static void init_forcerec_f_threads(t_forcerec *fr, int nenergrp)
1559 /* These thread local data structures are used for bondeds only */
1560 fr->nthreads = gmx_omp_nthreads_get(emntBonded);
1562 if (fr->nthreads > 1)
1564 snew(fr->f_t, fr->nthreads);
1565 /* Thread 0 uses the global force and energy arrays */
1566 for (t = 1; t < fr->nthreads; t++)
1568 fr->f_t[t].f = NULL;
1569 fr->f_t[t].f_nalloc = 0;
1570 snew(fr->f_t[t].fshift, SHIFTS);
1571 fr->f_t[t].grpp.nener = nenergrp*nenergrp;
1572 for (i = 0; i < egNR; i++)
1574 snew(fr->f_t[t].grpp.ener[i], fr->f_t[t].grpp.nener);
1581 gmx_bool nbnxn_acceleration_supported(FILE *fplog,
1582 const t_commrec *cr,
1583 const t_inputrec *ir,
1586 if (!bGPU && (ir->vdwtype == evdwPME && ir->ljpme_combination_rule == eljpmeLB))
1588 md_print_warn(cr, fplog, "LJ-PME with Lorentz-Berthelot is not supported with %s, falling back to %s\n",
1589 bGPU ? "GPUs" : "SIMD kernels",
1590 bGPU ? "CPU only" : "plain-C kernels");
1598 static void pick_nbnxn_kernel_cpu(const t_inputrec gmx_unused *ir,
1602 *kernel_type = nbnxnk4x4_PlainC;
1603 *ewald_excl = ewaldexclTable;
1605 #ifdef GMX_NBNXN_SIMD
1607 #ifdef GMX_NBNXN_SIMD_4XN
1608 *kernel_type = nbnxnk4xN_SIMD_4xN;
1610 #ifdef GMX_NBNXN_SIMD_2XNN
1611 *kernel_type = nbnxnk4xN_SIMD_2xNN;
1614 #if defined GMX_NBNXN_SIMD_2XNN && defined GMX_NBNXN_SIMD_4XN
1615 /* We need to choose if we want 2x(N+N) or 4xN kernels.
1616 * Currently this is based on the SIMD acceleration choice,
1617 * but it might be better to decide this at runtime based on CPU.
1619 * 4xN calculates more (zero) interactions, but has less pair-search
1620 * work and much better kernel instruction scheduling.
1622 * Up till now we have only seen that on Intel Sandy/Ivy Bridge,
1623 * which doesn't have FMA, both the analytical and tabulated Ewald
1624 * kernels have similar pair rates for 4x8 and 2x(4+4), so we choose
1625 * 2x(4+4) because it results in significantly fewer pairs.
1626 * For RF, the raw pair rate of the 4x8 kernel is higher than 2x(4+4),
1627 * 10% with HT, 50% without HT. As we currently don't detect the actual
1628 * use of HT, use 4x8 to avoid a potential performance hit.
1629 * On Intel Haswell 4x8 is always faster.
1631 *kernel_type = nbnxnk4xN_SIMD_4xN;
1633 #ifndef GMX_SIMD_HAVE_FMA
1634 if (EEL_PME_EWALD(ir->coulombtype) ||
1635 EVDW_PME(ir->vdwtype))
1637 /* We have Ewald kernels without FMA (Intel Sandy/Ivy Bridge).
1638 * There are enough instructions to make 2x(4+4) efficient.
1640 *kernel_type = nbnxnk4xN_SIMD_2xNN;
1643 #endif /* GMX_NBNXN_SIMD_2XNN && GMX_NBNXN_SIMD_4XN */
1646 if (getenv("GMX_NBNXN_SIMD_4XN") != NULL)
1648 #ifdef GMX_NBNXN_SIMD_4XN
1649 *kernel_type = nbnxnk4xN_SIMD_4xN;
1651 gmx_fatal(FARGS, "SIMD 4xN kernels requested, but Gromacs has been compiled without support for these kernels");
1654 if (getenv("GMX_NBNXN_SIMD_2XNN") != NULL)
1656 #ifdef GMX_NBNXN_SIMD_2XNN
1657 *kernel_type = nbnxnk4xN_SIMD_2xNN;
1659 gmx_fatal(FARGS, "SIMD 2x(N+N) kernels requested, but Gromacs has been compiled without support for these kernels");
1663 /* Analytical Ewald exclusion correction is only an option in
1665 * Since table lookup's don't parallelize with SIMD, analytical
1666 * will probably always be faster for a SIMD width of 8 or more.
1667 * With FMA analytical is sometimes faster for a width if 4 as well.
1668 * On BlueGene/Q, this is faster regardless of precision.
1669 * In single precision, this is faster on Bulldozer.
1671 #if GMX_SIMD_REAL_WIDTH >= 8 || \
1672 (GMX_SIMD_REAL_WIDTH >= 4 && defined GMX_SIMD_HAVE_FMA && !defined GMX_DOUBLE) || \
1673 defined GMX_SIMD_IBM_QPX
1674 *ewald_excl = ewaldexclAnalytical;
1676 if (getenv("GMX_NBNXN_EWALD_TABLE") != NULL)
1678 *ewald_excl = ewaldexclTable;
1680 if (getenv("GMX_NBNXN_EWALD_ANALYTICAL") != NULL)
1682 *ewald_excl = ewaldexclAnalytical;
1686 #endif /* GMX_NBNXN_SIMD */
1690 const char *lookup_nbnxn_kernel_name(int kernel_type)
1692 const char *returnvalue = NULL;
1693 switch (kernel_type)
1696 returnvalue = "not set";
1698 case nbnxnk4x4_PlainC:
1699 returnvalue = "plain C";
1701 case nbnxnk4xN_SIMD_4xN:
1702 case nbnxnk4xN_SIMD_2xNN:
1703 #ifdef GMX_NBNXN_SIMD
1704 #if defined GMX_SIMD_X86_SSE2
1705 returnvalue = "SSE2";
1706 #elif defined GMX_SIMD_X86_SSE4_1
1707 returnvalue = "SSE4.1";
1708 #elif defined GMX_SIMD_X86_AVX_128_FMA
1709 returnvalue = "AVX_128_FMA";
1710 #elif defined GMX_SIMD_X86_AVX_256
1711 returnvalue = "AVX_256";
1712 #elif defined GMX_SIMD_X86_AVX2_256
1713 returnvalue = "AVX2_256";
1715 returnvalue = "SIMD";
1717 #else /* GMX_NBNXN_SIMD */
1718 returnvalue = "not available";
1719 #endif /* GMX_NBNXN_SIMD */
1721 case nbnxnk8x8x8_CUDA: returnvalue = "CUDA"; break;
1722 case nbnxnk8x8x8_PlainC: returnvalue = "plain C"; break;
1726 gmx_fatal(FARGS, "Illegal kernel type selected");
1733 static void pick_nbnxn_kernel(FILE *fp,
1734 const t_commrec *cr,
1735 gmx_bool use_simd_kernels,
1737 gmx_bool bEmulateGPU,
1738 const t_inputrec *ir,
1741 gmx_bool bDoNonbonded)
1743 assert(kernel_type);
1745 *kernel_type = nbnxnkNotSet;
1746 *ewald_excl = ewaldexclTable;
1750 *kernel_type = nbnxnk8x8x8_PlainC;
1754 md_print_warn(cr, fp, "Emulating a GPU run on the CPU (slow)");
1759 *kernel_type = nbnxnk8x8x8_CUDA;
1762 if (*kernel_type == nbnxnkNotSet)
1764 /* LJ PME with LB combination rule does 7 mesh operations.
1765 * This so slow that we don't compile SIMD non-bonded kernels for that.
1767 if (use_simd_kernels &&
1768 nbnxn_acceleration_supported(fp, cr, ir, FALSE))
1770 pick_nbnxn_kernel_cpu(ir, kernel_type, ewald_excl);
1774 *kernel_type = nbnxnk4x4_PlainC;
1778 if (bDoNonbonded && fp != NULL)
1780 fprintf(fp, "\nUsing %s %dx%d non-bonded kernels\n\n",
1781 lookup_nbnxn_kernel_name(*kernel_type),
1782 nbnxn_kernel_pairlist_simple(*kernel_type) ? NBNXN_CPU_CLUSTER_I_SIZE : NBNXN_GPU_CLUSTER_SIZE,
1783 nbnxn_kernel_to_cj_size(*kernel_type));
1787 static void pick_nbnxn_resources(const t_commrec *cr,
1788 const gmx_hw_info_t *hwinfo,
1789 gmx_bool bDoNonbonded,
1791 gmx_bool *bEmulateGPU,
1792 const gmx_gpu_opt_t *gpu_opt)
1794 gmx_bool bEmulateGPUEnvVarSet;
1795 char gpu_err_str[STRLEN];
1799 bEmulateGPUEnvVarSet = (getenv("GMX_EMULATE_GPU") != NULL);
1801 /* Run GPU emulation mode if GMX_EMULATE_GPU is defined. Because
1802 * GPUs (currently) only handle non-bonded calculations, we will
1803 * automatically switch to emulation if non-bonded calculations are
1804 * turned off via GMX_NO_NONBONDED - this is the simple and elegant
1805 * way to turn off GPU initialization, data movement, and cleanup.
1807 * GPU emulation can be useful to assess the performance one can expect by
1808 * adding GPU(s) to the machine. The conditional below allows this even
1809 * if mdrun is compiled without GPU acceleration support.
1810 * Note that you should freezing the system as otherwise it will explode.
1812 *bEmulateGPU = (bEmulateGPUEnvVarSet ||
1814 gpu_opt->ncuda_dev_use > 0));
1816 /* Enable GPU mode when GPUs are available or no GPU emulation is requested.
1818 if (gpu_opt->ncuda_dev_use > 0 && !(*bEmulateGPU))
1820 /* Each PP node will use the intra-node id-th device from the
1821 * list of detected/selected GPUs. */
1822 if (!init_gpu(cr->rank_pp_intranode, gpu_err_str,
1823 &hwinfo->gpu_info, gpu_opt))
1825 /* At this point the init should never fail as we made sure that
1826 * we have all the GPUs we need. If it still does, we'll bail. */
1827 gmx_fatal(FARGS, "On node %d failed to initialize GPU #%d: %s",
1829 get_gpu_device_id(&hwinfo->gpu_info, gpu_opt,
1830 cr->rank_pp_intranode),
1834 /* Here we actually turn on hardware GPU acceleration */
1839 gmx_bool uses_simple_tables(int cutoff_scheme,
1840 nonbonded_verlet_t *nbv,
1843 gmx_bool bUsesSimpleTables = TRUE;
1846 switch (cutoff_scheme)
1849 bUsesSimpleTables = TRUE;
1852 assert(NULL != nbv && NULL != nbv->grp);
1853 grp_index = (group < 0) ? 0 : (nbv->ngrp - 1);
1854 bUsesSimpleTables = nbnxn_kernel_pairlist_simple(nbv->grp[grp_index].kernel_type);
1857 gmx_incons("unimplemented");
1859 return bUsesSimpleTables;
1862 static void init_ewald_f_table(interaction_const_t *ic,
1863 gmx_bool bUsesSimpleTables,
1868 if (bUsesSimpleTables)
1870 /* With a spacing of 0.0005 we are at the force summation accuracy
1871 * for the SSE kernels for "normal" atomistic simulations.
1873 ic->tabq_scale = ewald_spline3_table_scale(ic->ewaldcoeff_q,
1876 maxr = (rtab > ic->rcoulomb) ? rtab : ic->rcoulomb;
1877 ic->tabq_size = (int)(maxr*ic->tabq_scale) + 2;
1881 ic->tabq_size = GPU_EWALD_COULOMB_FORCE_TABLE_SIZE;
1882 /* Subtract 2 iso 1 to avoid access out of range due to rounding */
1883 ic->tabq_scale = (ic->tabq_size - 2)/ic->rcoulomb;
1886 sfree_aligned(ic->tabq_coul_FDV0);
1887 sfree_aligned(ic->tabq_coul_F);
1888 sfree_aligned(ic->tabq_coul_V);
1890 /* Create the original table data in FDV0 */
1891 snew_aligned(ic->tabq_coul_FDV0, ic->tabq_size*4, 32);
1892 snew_aligned(ic->tabq_coul_F, ic->tabq_size, 32);
1893 snew_aligned(ic->tabq_coul_V, ic->tabq_size, 32);
1894 table_spline3_fill_ewald_lr(ic->tabq_coul_F, ic->tabq_coul_V, ic->tabq_coul_FDV0,
1895 ic->tabq_size, 1/ic->tabq_scale, ic->ewaldcoeff_q);
1898 void init_interaction_const_tables(FILE *fp,
1899 interaction_const_t *ic,
1900 gmx_bool bUsesSimpleTables,
1905 if (ic->eeltype == eelEWALD || EEL_PME(ic->eeltype))
1907 init_ewald_f_table(ic, bUsesSimpleTables, rtab);
1911 fprintf(fp, "Initialized non-bonded Ewald correction tables, spacing: %.2e size: %d\n\n",
1912 1/ic->tabq_scale, ic->tabq_size);
1917 static void clear_force_switch_constants(shift_consts_t *sc)
1924 static void force_switch_constants(real p,
1928 /* Here we determine the coefficient for shifting the force to zero
1929 * between distance rsw and the cut-off rc.
1930 * For a potential of r^-p, we have force p*r^-(p+1).
1931 * But to save flops we absorb p in the coefficient.
1933 * force/p = r^-(p+1) + c2*r^2 + c3*r^3
1934 * potential = r^-p + c2/3*r^3 + c3/4*r^4 + cpot
1936 sc->c2 = ((p + 1)*rsw - (p + 4)*rc)/(pow(rc, p + 2)*pow(rc - rsw, 2));
1937 sc->c3 = -((p + 1)*rsw - (p + 3)*rc)/(pow(rc, p + 2)*pow(rc - rsw, 3));
1938 sc->cpot = -pow(rc, -p) + p*sc->c2/3*pow(rc - rsw, 3) + p*sc->c3/4*pow(rc - rsw, 4);
1941 static void potential_switch_constants(real rsw, real rc,
1942 switch_consts_t *sc)
1944 /* The switch function is 1 at rsw and 0 at rc.
1945 * The derivative and second derivate are zero at both ends.
1946 * rsw = max(r - r_switch, 0)
1947 * sw = 1 + c3*rsw^3 + c4*rsw^4 + c5*rsw^5
1948 * dsw = 3*c3*rsw^2 + 4*c4*rsw^3 + 5*c5*rsw^4
1949 * force = force*dsw - potential*sw
1952 sc->c3 = -10*pow(rc - rsw, -3);
1953 sc->c4 = 15*pow(rc - rsw, -4);
1954 sc->c5 = -6*pow(rc - rsw, -5);
1958 init_interaction_const(FILE *fp,
1959 const t_commrec gmx_unused *cr,
1960 interaction_const_t **interaction_const,
1961 const t_forcerec *fr,
1964 interaction_const_t *ic;
1965 gmx_bool bUsesSimpleTables = TRUE;
1969 /* Just allocate something so we can free it */
1970 snew_aligned(ic->tabq_coul_FDV0, 16, 32);
1971 snew_aligned(ic->tabq_coul_F, 16, 32);
1972 snew_aligned(ic->tabq_coul_V, 16, 32);
1974 ic->rlist = fr->rlist;
1975 ic->rlistlong = fr->rlistlong;
1978 ic->vdwtype = fr->vdwtype;
1979 ic->vdw_modifier = fr->vdw_modifier;
1980 ic->rvdw = fr->rvdw;
1981 ic->rvdw_switch = fr->rvdw_switch;
1982 ic->ewaldcoeff_lj = fr->ewaldcoeff_lj;
1983 ic->ljpme_comb_rule = fr->ljpme_combination_rule;
1984 ic->sh_lj_ewald = 0;
1985 clear_force_switch_constants(&ic->dispersion_shift);
1986 clear_force_switch_constants(&ic->repulsion_shift);
1988 switch (ic->vdw_modifier)
1990 case eintmodPOTSHIFT:
1991 /* Only shift the potential, don't touch the force */
1992 ic->dispersion_shift.cpot = -pow(ic->rvdw, -6.0);
1993 ic->repulsion_shift.cpot = -pow(ic->rvdw, -12.0);
1994 if (EVDW_PME(ic->vdwtype))
1998 crc2 = sqr(ic->ewaldcoeff_lj*ic->rvdw);
1999 ic->sh_lj_ewald = (exp(-crc2)*(1 + crc2 + 0.5*crc2*crc2) - 1)*pow(ic->rvdw, -6.0);
2002 case eintmodFORCESWITCH:
2003 /* Switch the force, switch and shift the potential */
2004 force_switch_constants(6.0, ic->rvdw_switch, ic->rvdw,
2005 &ic->dispersion_shift);
2006 force_switch_constants(12.0, ic->rvdw_switch, ic->rvdw,
2007 &ic->repulsion_shift);
2009 case eintmodPOTSWITCH:
2010 /* Switch the potential and force */
2011 potential_switch_constants(ic->rvdw_switch, ic->rvdw,
2015 case eintmodEXACTCUTOFF:
2016 /* Nothing to do here */
2019 gmx_incons("unimplemented potential modifier");
2022 ic->sh_invrc6 = -ic->dispersion_shift.cpot;
2024 /* Electrostatics */
2025 ic->eeltype = fr->eeltype;
2026 ic->coulomb_modifier = fr->coulomb_modifier;
2027 ic->rcoulomb = fr->rcoulomb;
2028 ic->epsilon_r = fr->epsilon_r;
2029 ic->epsfac = fr->epsfac;
2030 ic->ewaldcoeff_q = fr->ewaldcoeff_q;
2032 if (fr->coulomb_modifier == eintmodPOTSHIFT)
2034 ic->sh_ewald = gmx_erfc(ic->ewaldcoeff_q*ic->rcoulomb);
2041 /* Reaction-field */
2042 if (EEL_RF(ic->eeltype))
2044 ic->epsilon_rf = fr->epsilon_rf;
2045 ic->k_rf = fr->k_rf;
2046 ic->c_rf = fr->c_rf;
2050 /* For plain cut-off we might use the reaction-field kernels */
2051 ic->epsilon_rf = ic->epsilon_r;
2053 if (fr->coulomb_modifier == eintmodPOTSHIFT)
2055 ic->c_rf = 1/ic->rcoulomb;
2065 real dispersion_shift;
2067 dispersion_shift = ic->dispersion_shift.cpot;
2068 if (EVDW_PME(ic->vdwtype))
2070 dispersion_shift -= ic->sh_lj_ewald;
2072 fprintf(fp, "Potential shift: LJ r^-12: %.3e r^-6: %.3e",
2073 ic->repulsion_shift.cpot, dispersion_shift);
2075 if (ic->eeltype == eelCUT)
2077 fprintf(fp, ", Coulomb %.e", -ic->c_rf);
2079 else if (EEL_PME(ic->eeltype))
2081 fprintf(fp, ", Ewald %.3e", -ic->sh_ewald);
2086 *interaction_const = ic;
2088 if (fr->nbv != NULL && fr->nbv->bUseGPU)
2090 nbnxn_cuda_init_const(fr->nbv->cu_nbv, ic, fr->nbv->grp);
2092 /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
2093 * also sharing texture references. To keep the code simple, we don't
2094 * treat texture references as shared resources, but this means that
2095 * the coulomb_tab and nbfp texture refs will get updated by multiple threads.
2096 * Hence, to ensure that the non-bonded kernels don't start before all
2097 * texture binding operations are finished, we need to wait for all ranks
2098 * to arrive here before continuing.
2100 * Note that we could omit this barrier if GPUs are not shared (or
2101 * texture objects are used), but as this is initialization code, there
2102 * is not point in complicating things.
2104 #ifdef GMX_THREAD_MPI
2109 #endif /* GMX_THREAD_MPI */
2112 bUsesSimpleTables = uses_simple_tables(fr->cutoff_scheme, fr->nbv, -1);
2113 init_interaction_const_tables(fp, ic, bUsesSimpleTables, rtab);
2116 static void init_nb_verlet(FILE *fp,
2117 nonbonded_verlet_t **nb_verlet,
2118 gmx_bool bFEP_NonBonded,
2119 const t_inputrec *ir,
2120 const t_forcerec *fr,
2121 const t_commrec *cr,
2122 const char *nbpu_opt)
2124 nonbonded_verlet_t *nbv;
2127 gmx_bool bEmulateGPU, bHybridGPURun = FALSE;
2129 nbnxn_alloc_t *nb_alloc;
2130 nbnxn_free_t *nb_free;
2134 pick_nbnxn_resources(cr, fr->hwinfo,
2142 nbv->ngrp = (DOMAINDECOMP(cr) ? 2 : 1);
2143 for (i = 0; i < nbv->ngrp; i++)
2145 nbv->grp[i].nbl_lists.nnbl = 0;
2146 nbv->grp[i].nbat = NULL;
2147 nbv->grp[i].kernel_type = nbnxnkNotSet;
2149 if (i == 0) /* local */
2151 pick_nbnxn_kernel(fp, cr, fr->use_simd_kernels,
2152 nbv->bUseGPU, bEmulateGPU, ir,
2153 &nbv->grp[i].kernel_type,
2154 &nbv->grp[i].ewald_excl,
2157 else /* non-local */
2159 if (nbpu_opt != NULL && strcmp(nbpu_opt, "gpu_cpu") == 0)
2161 /* Use GPU for local, select a CPU kernel for non-local */
2162 pick_nbnxn_kernel(fp, cr, fr->use_simd_kernels,
2164 &nbv->grp[i].kernel_type,
2165 &nbv->grp[i].ewald_excl,
2168 bHybridGPURun = TRUE;
2172 /* Use the same kernel for local and non-local interactions */
2173 nbv->grp[i].kernel_type = nbv->grp[0].kernel_type;
2174 nbv->grp[i].ewald_excl = nbv->grp[0].ewald_excl;
2181 /* init the NxN GPU data; the last argument tells whether we'll have
2182 * both local and non-local NB calculation on GPU */
2183 nbnxn_cuda_init(fp, &nbv->cu_nbv,
2184 &fr->hwinfo->gpu_info, fr->gpu_opt,
2185 cr->rank_pp_intranode,
2186 (nbv->ngrp > 1) && !bHybridGPURun);
2188 if ((env = getenv("GMX_NB_MIN_CI")) != NULL)
2192 nbv->min_ci_balanced = strtol(env, &end, 10);
2193 if (!end || (*end != 0) || nbv->min_ci_balanced <= 0)
2195 gmx_fatal(FARGS, "Invalid value passed in GMX_NB_MIN_CI=%s, positive integer required", env);
2200 fprintf(debug, "Neighbor-list balancing parameter: %d (passed as env. var.)\n",
2201 nbv->min_ci_balanced);
2206 nbv->min_ci_balanced = nbnxn_cuda_min_ci_balanced(nbv->cu_nbv);
2209 fprintf(debug, "Neighbor-list balancing parameter: %d (auto-adjusted to the number of GPU multi-processors)\n",
2210 nbv->min_ci_balanced);
2216 nbv->min_ci_balanced = 0;
2221 nbnxn_init_search(&nbv->nbs,
2222 DOMAINDECOMP(cr) ? &cr->dd->nc : NULL,
2223 DOMAINDECOMP(cr) ? domdec_zones(cr->dd) : NULL,
2225 gmx_omp_nthreads_get(emntNonbonded));
2227 for (i = 0; i < nbv->ngrp; i++)
2229 if (nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
2231 nb_alloc = &pmalloc;
2240 nbnxn_init_pairlist_set(&nbv->grp[i].nbl_lists,
2241 nbnxn_kernel_pairlist_simple(nbv->grp[i].kernel_type),
2242 /* 8x8x8 "non-simple" lists are ATM always combined */
2243 !nbnxn_kernel_pairlist_simple(nbv->grp[i].kernel_type),
2247 nbv->grp[0].kernel_type != nbv->grp[i].kernel_type)
2249 gmx_bool bSimpleList;
2250 int enbnxninitcombrule;
2252 bSimpleList = nbnxn_kernel_pairlist_simple(nbv->grp[i].kernel_type);
2254 if (bSimpleList && (fr->vdwtype == evdwCUT && (fr->vdw_modifier == eintmodNONE || fr->vdw_modifier == eintmodPOTSHIFT)))
2256 /* Plain LJ cut-off: we can optimize with combination rules */
2257 enbnxninitcombrule = enbnxninitcombruleDETECT;
2259 else if (fr->vdwtype == evdwPME)
2261 /* LJ-PME: we need to use a combination rule for the grid */
2262 if (fr->ljpme_combination_rule == eljpmeGEOM)
2264 enbnxninitcombrule = enbnxninitcombruleGEOM;
2268 enbnxninitcombrule = enbnxninitcombruleLB;
2273 /* We use a full combination matrix: no rule required */
2274 enbnxninitcombrule = enbnxninitcombruleNONE;
2278 snew(nbv->grp[i].nbat, 1);
2279 nbnxn_atomdata_init(fp,
2281 nbv->grp[i].kernel_type,
2283 fr->ntype, fr->nbfp,
2285 bSimpleList ? gmx_omp_nthreads_get(emntNonbonded) : 1,
2290 nbv->grp[i].nbat = nbv->grp[0].nbat;
2295 void init_forcerec(FILE *fp,
2296 const output_env_t oenv,
2299 const t_inputrec *ir,
2300 const gmx_mtop_t *mtop,
2301 const t_commrec *cr,
2307 const char *nbpu_opt,
2308 gmx_bool bNoSolvOpt,
2311 int i, j, m, natoms, ngrp, negp_pp, negptable, egi, egj;
2316 gmx_bool bGenericKernelOnly;
2317 gmx_bool bMakeTables, bMakeSeparate14Table, bSomeNormalNbListsAreInUse;
2318 gmx_bool bFEP_NonBonded;
2320 int *nm_ind, egp_flags;
2322 if (fr->hwinfo == NULL)
2324 /* Detect hardware, gather information.
2325 * In mdrun, hwinfo has already been set before calling init_forcerec.
2326 * Here we ignore GPUs, as tools will not use them anyhow.
2328 fr->hwinfo = gmx_detect_hardware(fp, cr, FALSE);
2331 /* By default we turn SIMD kernels on, but it might be turned off further down... */
2332 fr->use_simd_kernels = TRUE;
2334 fr->bDomDec = DOMAINDECOMP(cr);
2336 natoms = mtop->natoms;
2338 if (check_box(ir->ePBC, box))
2340 gmx_fatal(FARGS, check_box(ir->ePBC, box));
2343 /* Test particle insertion ? */
2346 /* Set to the size of the molecule to be inserted (the last one) */
2347 /* Because of old style topologies, we have to use the last cg
2348 * instead of the last molecule type.
2350 cgs = &mtop->moltype[mtop->molblock[mtop->nmolblock-1].type].cgs;
2351 fr->n_tpi = cgs->index[cgs->nr] - cgs->index[cgs->nr-1];
2352 if (fr->n_tpi != mtop->mols.index[mtop->mols.nr] - mtop->mols.index[mtop->mols.nr-1])
2354 gmx_fatal(FARGS, "The molecule to insert can not consist of multiple charge groups.\nMake it a single charge group.");
2362 /* Copy AdResS parameters */
2365 fr->adress_type = ir->adress->type;
2366 fr->adress_const_wf = ir->adress->const_wf;
2367 fr->adress_ex_width = ir->adress->ex_width;
2368 fr->adress_hy_width = ir->adress->hy_width;
2369 fr->adress_icor = ir->adress->icor;
2370 fr->adress_site = ir->adress->site;
2371 fr->adress_ex_forcecap = ir->adress->ex_forcecap;
2372 fr->adress_do_hybridpairs = ir->adress->do_hybridpairs;
2375 snew(fr->adress_group_explicit, ir->adress->n_energy_grps);
2376 for (i = 0; i < ir->adress->n_energy_grps; i++)
2378 fr->adress_group_explicit[i] = ir->adress->group_explicit[i];
2381 fr->n_adress_tf_grps = ir->adress->n_tf_grps;
2382 snew(fr->adress_tf_table_index, fr->n_adress_tf_grps);
2383 for (i = 0; i < fr->n_adress_tf_grps; i++)
2385 fr->adress_tf_table_index[i] = ir->adress->tf_table_index[i];
2387 copy_rvec(ir->adress->refs, fr->adress_refs);
2391 fr->adress_type = eAdressOff;
2392 fr->adress_do_hybridpairs = FALSE;
2395 /* Copy the user determined parameters */
2396 fr->userint1 = ir->userint1;
2397 fr->userint2 = ir->userint2;
2398 fr->userint3 = ir->userint3;
2399 fr->userint4 = ir->userint4;
2400 fr->userreal1 = ir->userreal1;
2401 fr->userreal2 = ir->userreal2;
2402 fr->userreal3 = ir->userreal3;
2403 fr->userreal4 = ir->userreal4;
2406 fr->fc_stepsize = ir->fc_stepsize;
2409 fr->efep = ir->efep;
2410 fr->sc_alphavdw = ir->fepvals->sc_alpha;
2411 if (ir->fepvals->bScCoul)
2413 fr->sc_alphacoul = ir->fepvals->sc_alpha;
2414 fr->sc_sigma6_min = pow(ir->fepvals->sc_sigma_min, 6);
2418 fr->sc_alphacoul = 0;
2419 fr->sc_sigma6_min = 0; /* only needed when bScCoul is on */
2421 fr->sc_power = ir->fepvals->sc_power;
2422 fr->sc_r_power = ir->fepvals->sc_r_power;
2423 fr->sc_sigma6_def = pow(ir->fepvals->sc_sigma, 6);
2425 env = getenv("GMX_SCSIGMA_MIN");
2429 sscanf(env, "%lf", &dbl);
2430 fr->sc_sigma6_min = pow(dbl, 6);
2433 fprintf(fp, "Setting the minimum soft core sigma to %g nm\n", dbl);
2437 fr->bNonbonded = TRUE;
2438 if (getenv("GMX_NO_NONBONDED") != NULL)
2440 /* turn off non-bonded calculations */
2441 fr->bNonbonded = FALSE;
2442 md_print_warn(cr, fp,
2443 "Found environment variable GMX_NO_NONBONDED.\n"
2444 "Disabling nonbonded calculations.\n");
2447 bGenericKernelOnly = FALSE;
2449 /* We now check in the NS code whether a particular combination of interactions
2450 * can be used with water optimization, and disable it if that is not the case.
2453 if (getenv("GMX_NB_GENERIC") != NULL)
2458 "Found environment variable GMX_NB_GENERIC.\n"
2459 "Disabling all interaction-specific nonbonded kernels, will only\n"
2460 "use the slow generic ones in src/gmxlib/nonbonded/nb_generic.c\n\n");
2462 bGenericKernelOnly = TRUE;
2465 if (bGenericKernelOnly == TRUE)
2470 if ( (getenv("GMX_DISABLE_SIMD_KERNELS") != NULL) || (getenv("GMX_NOOPTIMIZEDKERNELS") != NULL) )
2472 fr->use_simd_kernels = FALSE;
2476 "\nFound environment variable GMX_DISABLE_SIMD_KERNELS.\n"
2477 "Disabling the usage of any SIMD-specific kernel routines (e.g. SSE2/SSE4.1/AVX).\n\n");
2481 fr->bBHAM = (mtop->ffparams.functype[0] == F_BHAM);
2483 /* Check if we can/should do all-vs-all kernels */
2484 fr->bAllvsAll = can_use_allvsall(ir, FALSE, NULL, NULL);
2485 fr->AllvsAll_work = NULL;
2486 fr->AllvsAll_workgb = NULL;
2488 /* All-vs-all kernels have not been implemented in 4.6, and
2489 * the SIMD group kernels are also buggy in this case. Non-SIMD
2490 * group kernels are OK. See Redmine #1249. */
2493 fr->bAllvsAll = FALSE;
2494 fr->use_simd_kernels = FALSE;
2498 "\nYour simulation settings would have triggered the efficient all-vs-all\n"
2499 "kernels in GROMACS 4.5, but these have not been implemented in GROMACS\n"
2500 "4.6. Also, we can't use the accelerated SIMD kernels here because\n"
2501 "of an unfixed bug. The reference C kernels are correct, though, so\n"
2502 "we are proceeding by disabling all CPU architecture-specific\n"
2503 "(e.g. SSE2/SSE4/AVX) routines. If performance is important, please\n"
2504 "use GROMACS 4.5.7 or try cutoff-scheme = Verlet.\n\n");
2508 /* Neighbour searching stuff */
2509 fr->cutoff_scheme = ir->cutoff_scheme;
2510 fr->bGrid = (ir->ns_type == ensGRID);
2511 fr->ePBC = ir->ePBC;
2513 if (fr->cutoff_scheme == ecutsGROUP)
2515 const char *note = "NOTE: This file uses the deprecated 'group' cutoff_scheme. This will be\n"
2516 "removed in a future release when 'verlet' supports all interaction forms.\n";
2520 fprintf(stderr, "\n%s\n", note);
2524 fprintf(fp, "\n%s\n", note);
2528 /* Determine if we will do PBC for distances in bonded interactions */
2529 if (fr->ePBC == epbcNONE)
2531 fr->bMolPBC = FALSE;
2535 if (!DOMAINDECOMP(cr))
2537 /* The group cut-off scheme and SHAKE assume charge groups
2538 * are whole, but not using molpbc is faster in most cases.
2540 if (fr->cutoff_scheme == ecutsGROUP ||
2541 (ir->eConstrAlg == econtSHAKE &&
2542 (gmx_mtop_ftype_count(mtop, F_CONSTR) > 0 ||
2543 gmx_mtop_ftype_count(mtop, F_CONSTRNC) > 0)))
2545 fr->bMolPBC = ir->bPeriodicMols;
2550 if (getenv("GMX_USE_GRAPH") != NULL)
2552 fr->bMolPBC = FALSE;
2555 fprintf(fp, "\nGMX_MOLPBC is set, using the graph for bonded interactions\n\n");
2562 fr->bMolPBC = dd_bonded_molpbc(cr->dd, fr->ePBC);
2565 fr->bGB = (ir->implicit_solvent == eisGBSA);
2567 fr->rc_scaling = ir->refcoord_scaling;
2568 copy_rvec(ir->posres_com, fr->posres_com);
2569 copy_rvec(ir->posres_comB, fr->posres_comB);
2570 fr->rlist = cutoff_inf(ir->rlist);
2571 fr->rlistlong = cutoff_inf(ir->rlistlong);
2572 fr->eeltype = ir->coulombtype;
2573 fr->vdwtype = ir->vdwtype;
2574 fr->ljpme_combination_rule = ir->ljpme_combination_rule;
2576 fr->coulomb_modifier = ir->coulomb_modifier;
2577 fr->vdw_modifier = ir->vdw_modifier;
2579 /* Electrostatics: Translate from interaction-setting-in-mdp-file to kernel interaction format */
2580 switch (fr->eeltype)
2583 fr->nbkernel_elec_interaction = (fr->bGB) ? GMX_NBKERNEL_ELEC_GENERALIZEDBORN : GMX_NBKERNEL_ELEC_COULOMB;
2589 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_REACTIONFIELD;
2593 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_REACTIONFIELD;
2594 fr->coulomb_modifier = eintmodEXACTCUTOFF;
2603 case eelPMEUSERSWITCH:
2604 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_CUBICSPLINETABLE;
2609 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_EWALD;
2613 gmx_fatal(FARGS, "Unsupported electrostatic interaction: %s", eel_names[fr->eeltype]);
2617 /* Vdw: Translate from mdp settings to kernel format */
2618 switch (fr->vdwtype)
2623 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_BUCKINGHAM;
2627 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_LENNARDJONES;
2631 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_LJEWALD;
2637 case evdwENCADSHIFT:
2638 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_CUBICSPLINETABLE;
2642 gmx_fatal(FARGS, "Unsupported vdw interaction: %s", evdw_names[fr->vdwtype]);
2646 /* These start out identical to ir, but might be altered if we e.g. tabulate the interaction in the kernel */
2647 fr->nbkernel_elec_modifier = fr->coulomb_modifier;
2648 fr->nbkernel_vdw_modifier = fr->vdw_modifier;
2650 fr->bTwinRange = fr->rlistlong > fr->rlist;
2651 fr->bEwald = (EEL_PME(fr->eeltype) || fr->eeltype == eelEWALD);
2653 fr->reppow = mtop->ffparams.reppow;
2655 if (ir->cutoff_scheme == ecutsGROUP)
2657 fr->bvdwtab = ((fr->vdwtype != evdwCUT || !gmx_within_tol(fr->reppow, 12.0, 10*GMX_DOUBLE_EPS))
2658 && !EVDW_PME(fr->vdwtype));
2659 /* We have special kernels for standard Ewald and PME, but the pme-switch ones are tabulated above */
2660 fr->bcoultab = !(fr->eeltype == eelCUT ||
2661 fr->eeltype == eelEWALD ||
2662 fr->eeltype == eelPME ||
2663 fr->eeltype == eelRF ||
2664 fr->eeltype == eelRF_ZERO);
2666 /* If the user absolutely wants different switch/shift settings for coul/vdw, it is likely
2667 * going to be faster to tabulate the interaction than calling the generic kernel.
2669 if (fr->nbkernel_elec_modifier == eintmodPOTSWITCH && fr->nbkernel_vdw_modifier == eintmodPOTSWITCH)
2671 if ((fr->rcoulomb_switch != fr->rvdw_switch) || (fr->rcoulomb != fr->rvdw))
2673 fr->bcoultab = TRUE;
2676 else if ((fr->nbkernel_elec_modifier == eintmodPOTSHIFT && fr->nbkernel_vdw_modifier == eintmodPOTSHIFT) ||
2677 ((fr->nbkernel_elec_interaction == GMX_NBKERNEL_ELEC_REACTIONFIELD &&
2678 fr->nbkernel_elec_modifier == eintmodEXACTCUTOFF &&
2679 (fr->nbkernel_vdw_modifier == eintmodPOTSWITCH || fr->nbkernel_vdw_modifier == eintmodPOTSHIFT))))
2681 if (fr->rcoulomb != fr->rvdw)
2683 fr->bcoultab = TRUE;
2687 if (getenv("GMX_REQUIRE_TABLES"))
2690 fr->bcoultab = TRUE;
2695 fprintf(fp, "Table routines are used for coulomb: %s\n", bool_names[fr->bcoultab]);
2696 fprintf(fp, "Table routines are used for vdw: %s\n", bool_names[fr->bvdwtab ]);
2699 if (fr->bvdwtab == TRUE)
2701 fr->nbkernel_vdw_interaction = GMX_NBKERNEL_VDW_CUBICSPLINETABLE;
2702 fr->nbkernel_vdw_modifier = eintmodNONE;
2704 if (fr->bcoultab == TRUE)
2706 fr->nbkernel_elec_interaction = GMX_NBKERNEL_ELEC_CUBICSPLINETABLE;
2707 fr->nbkernel_elec_modifier = eintmodNONE;
2711 if (ir->cutoff_scheme == ecutsVERLET)
2713 if (!gmx_within_tol(fr->reppow, 12.0, 10*GMX_DOUBLE_EPS))
2715 gmx_fatal(FARGS, "Cut-off scheme %S only supports LJ repulsion power 12", ecutscheme_names[ir->cutoff_scheme]);
2717 fr->bvdwtab = FALSE;
2718 fr->bcoultab = FALSE;
2721 /* Tables are used for direct ewald sum */
2724 if (EEL_PME(ir->coulombtype))
2728 fprintf(fp, "Will do PME sum in reciprocal space for electrostatic interactions.\n");
2730 if (ir->coulombtype == eelP3M_AD)
2732 please_cite(fp, "Hockney1988");
2733 please_cite(fp, "Ballenegger2012");
2737 please_cite(fp, "Essmann95a");
2740 if (ir->ewald_geometry == eewg3DC)
2744 fprintf(fp, "Using the Ewald3DC correction for systems with a slab geometry.\n");
2746 please_cite(fp, "In-Chul99a");
2749 fr->ewaldcoeff_q = calc_ewaldcoeff_q(ir->rcoulomb, ir->ewald_rtol);
2750 init_ewald_tab(&(fr->ewald_table), ir, fp);
2753 fprintf(fp, "Using a Gaussian width (1/beta) of %g nm for Ewald\n",
2754 1/fr->ewaldcoeff_q);
2758 if (EVDW_PME(ir->vdwtype))
2762 fprintf(fp, "Will do PME sum in reciprocal space for LJ dispersion interactions.\n");
2764 please_cite(fp, "Essmann95a");
2765 fr->ewaldcoeff_lj = calc_ewaldcoeff_lj(ir->rvdw, ir->ewald_rtol_lj);
2768 fprintf(fp, "Using a Gaussian width (1/beta) of %g nm for LJ Ewald\n",
2769 1/fr->ewaldcoeff_lj);
2773 /* Electrostatics */
2774 fr->epsilon_r = ir->epsilon_r;
2775 fr->epsilon_rf = ir->epsilon_rf;
2776 fr->fudgeQQ = mtop->ffparams.fudgeQQ;
2777 fr->rcoulomb_switch = ir->rcoulomb_switch;
2778 fr->rcoulomb = cutoff_inf(ir->rcoulomb);
2780 /* Parameters for generalized RF */
2784 if (fr->eeltype == eelGRF)
2786 init_generalized_rf(fp, mtop, ir, fr);
2789 fr->bF_NoVirSum = (EEL_FULL(fr->eeltype) || EVDW_PME(fr->vdwtype) ||
2790 gmx_mtop_ftype_count(mtop, F_POSRES) > 0 ||
2791 gmx_mtop_ftype_count(mtop, F_FBPOSRES) > 0 ||
2792 IR_ELEC_FIELD(*ir) ||
2793 (fr->adress_icor != eAdressICOff)
2796 if (fr->cutoff_scheme == ecutsGROUP &&
2797 ncg_mtop(mtop) > fr->cg_nalloc && !DOMAINDECOMP(cr))
2799 /* Count the total number of charge groups */
2800 fr->cg_nalloc = ncg_mtop(mtop);
2801 srenew(fr->cg_cm, fr->cg_nalloc);
2803 if (fr->shift_vec == NULL)
2805 snew(fr->shift_vec, SHIFTS);
2808 if (fr->fshift == NULL)
2810 snew(fr->fshift, SHIFTS);
2813 if (fr->nbfp == NULL)
2815 fr->ntype = mtop->ffparams.atnr;
2816 fr->nbfp = mk_nbfp(&mtop->ffparams, fr->bBHAM);
2817 if (EVDW_PME(fr->vdwtype))
2819 fr->ljpme_c6grid = make_ljpme_c6grid(&mtop->ffparams, fr);
2823 /* Copy the energy group exclusions */
2824 fr->egp_flags = ir->opts.egp_flags;
2826 /* Van der Waals stuff */
2827 fr->rvdw = cutoff_inf(ir->rvdw);
2828 fr->rvdw_switch = ir->rvdw_switch;
2829 if ((fr->vdwtype != evdwCUT) && (fr->vdwtype != evdwUSER) && !fr->bBHAM)
2831 if (fr->rvdw_switch >= fr->rvdw)
2833 gmx_fatal(FARGS, "rvdw_switch (%f) must be < rvdw (%f)",
2834 fr->rvdw_switch, fr->rvdw);
2838 fprintf(fp, "Using %s Lennard-Jones, switch between %g and %g nm\n",
2839 (fr->eeltype == eelSWITCH) ? "switched" : "shifted",
2840 fr->rvdw_switch, fr->rvdw);
2844 if (fr->bBHAM && EVDW_PME(fr->vdwtype))
2846 gmx_fatal(FARGS, "LJ PME not supported with Buckingham");
2849 if (fr->bBHAM && (fr->vdwtype == evdwSHIFT || fr->vdwtype == evdwSWITCH))
2851 gmx_fatal(FARGS, "Switch/shift interaction not supported with Buckingham");
2856 fprintf(fp, "Cut-off's: NS: %g Coulomb: %g %s: %g\n",
2857 fr->rlist, fr->rcoulomb, fr->bBHAM ? "BHAM" : "LJ", fr->rvdw);
2860 fr->eDispCorr = ir->eDispCorr;
2861 if (ir->eDispCorr != edispcNO)
2863 set_avcsixtwelve(fp, fr, mtop);
2868 set_bham_b_max(fp, fr, mtop);
2871 fr->gb_epsilon_solvent = ir->gb_epsilon_solvent;
2873 /* Copy the GBSA data (radius, volume and surftens for each
2874 * atomtype) from the topology atomtype section to forcerec.
2876 snew(fr->atype_radius, fr->ntype);
2877 snew(fr->atype_vol, fr->ntype);
2878 snew(fr->atype_surftens, fr->ntype);
2879 snew(fr->atype_gb_radius, fr->ntype);
2880 snew(fr->atype_S_hct, fr->ntype);
2882 if (mtop->atomtypes.nr > 0)
2884 for (i = 0; i < fr->ntype; i++)
2886 fr->atype_radius[i] = mtop->atomtypes.radius[i];
2888 for (i = 0; i < fr->ntype; i++)
2890 fr->atype_vol[i] = mtop->atomtypes.vol[i];
2892 for (i = 0; i < fr->ntype; i++)
2894 fr->atype_surftens[i] = mtop->atomtypes.surftens[i];
2896 for (i = 0; i < fr->ntype; i++)
2898 fr->atype_gb_radius[i] = mtop->atomtypes.gb_radius[i];
2900 for (i = 0; i < fr->ntype; i++)
2902 fr->atype_S_hct[i] = mtop->atomtypes.S_hct[i];
2906 /* Generate the GB table if needed */
2910 fr->gbtabscale = 2000;
2912 fr->gbtabscale = 500;
2916 fr->gbtab = make_gb_table(oenv, fr);
2918 init_gb(&fr->born, fr, ir, mtop, ir->gb_algorithm);
2920 /* Copy local gb data (for dd, this is done in dd_partition_system) */
2921 if (!DOMAINDECOMP(cr))
2923 make_local_gb(cr, fr->born, ir->gb_algorithm);
2927 /* Set the charge scaling */
2928 if (fr->epsilon_r != 0)
2930 fr->epsfac = ONE_4PI_EPS0/fr->epsilon_r;
2934 /* eps = 0 is infinite dieletric: no coulomb interactions */
2938 /* Reaction field constants */
2939 if (EEL_RF(fr->eeltype))
2941 calc_rffac(fp, fr->eeltype, fr->epsilon_r, fr->epsilon_rf,
2942 fr->rcoulomb, fr->temp, fr->zsquare, box,
2943 &fr->kappa, &fr->k_rf, &fr->c_rf);
2946 /*This now calculates sum for q and c6*/
2947 set_chargesum(fp, fr, mtop);
2949 /* if we are using LR electrostatics, and they are tabulated,
2950 * the tables will contain modified coulomb interactions.
2951 * Since we want to use the non-shifted ones for 1-4
2952 * coulombic interactions, we must have an extra set of tables.
2955 /* Construct tables.
2956 * A little unnecessary to make both vdw and coul tables sometimes,
2957 * but what the heck... */
2959 bMakeTables = fr->bcoultab || fr->bvdwtab || fr->bEwald ||
2960 (ir->eDispCorr != edispcNO && ir_vdw_switched(ir));
2962 bMakeSeparate14Table = ((!bMakeTables || fr->eeltype != eelCUT || fr->vdwtype != evdwCUT ||
2963 fr->bBHAM || fr->bEwald) &&
2964 (gmx_mtop_ftype_count(mtop, F_LJ14) > 0 ||
2965 gmx_mtop_ftype_count(mtop, F_LJC14_Q) > 0 ||
2966 gmx_mtop_ftype_count(mtop, F_LJC_PAIRS_NB) > 0));
2968 negp_pp = ir->opts.ngener - ir->nwall;
2972 bSomeNormalNbListsAreInUse = TRUE;
2977 bSomeNormalNbListsAreInUse = (ir->eDispCorr != edispcNO);
2978 for (egi = 0; egi < negp_pp; egi++)
2980 for (egj = egi; egj < negp_pp; egj++)
2982 egp_flags = ir->opts.egp_flags[GID(egi, egj, ir->opts.ngener)];
2983 if (!(egp_flags & EGP_EXCL))
2985 if (egp_flags & EGP_TABLE)
2991 bSomeNormalNbListsAreInUse = TRUE;
2996 if (bSomeNormalNbListsAreInUse)
2998 fr->nnblists = negptable + 1;
3002 fr->nnblists = negptable;
3004 if (fr->nnblists > 1)
3006 snew(fr->gid2nblists, ir->opts.ngener*ir->opts.ngener);
3015 snew(fr->nblists, fr->nnblists);
3017 /* This code automatically gives table length tabext without cut-off's,
3018 * in that case grompp should already have checked that we do not need
3019 * normal tables and we only generate tables for 1-4 interactions.
3021 rtab = ir->rlistlong + ir->tabext;
3025 /* make tables for ordinary interactions */
3026 if (bSomeNormalNbListsAreInUse)
3028 make_nbf_tables(fp, oenv, fr, rtab, cr, tabfn, NULL, NULL, &fr->nblists[0]);
3031 make_nbf_tables(fp, oenv, fr, rtab, cr, tabfn, NULL, NULL, &fr->nblists[fr->nnblists/2]);
3033 if (!bMakeSeparate14Table)
3035 fr->tab14 = fr->nblists[0].table_elec_vdw;
3045 /* Read the special tables for certain energy group pairs */
3046 nm_ind = mtop->groups.grps[egcENER].nm_ind;
3047 for (egi = 0; egi < negp_pp; egi++)
3049 for (egj = egi; egj < negp_pp; egj++)
3051 egp_flags = ir->opts.egp_flags[GID(egi, egj, ir->opts.ngener)];
3052 if ((egp_flags & EGP_TABLE) && !(egp_flags & EGP_EXCL))
3054 nbl = &(fr->nblists[m]);
3055 if (fr->nnblists > 1)
3057 fr->gid2nblists[GID(egi, egj, ir->opts.ngener)] = m;
3059 /* Read the table file with the two energy groups names appended */
3060 make_nbf_tables(fp, oenv, fr, rtab, cr, tabfn,
3061 *mtop->groups.grpname[nm_ind[egi]],
3062 *mtop->groups.grpname[nm_ind[egj]],
3066 make_nbf_tables(fp, oenv, fr, rtab, cr, tabfn,
3067 *mtop->groups.grpname[nm_ind[egi]],
3068 *mtop->groups.grpname[nm_ind[egj]],
3069 &fr->nblists[fr->nnblists/2+m]);
3073 else if (fr->nnblists > 1)
3075 fr->gid2nblists[GID(egi, egj, ir->opts.ngener)] = 0;
3081 if (bMakeSeparate14Table)
3083 /* generate extra tables with plain Coulomb for 1-4 interactions only */
3084 fr->tab14 = make_tables(fp, oenv, fr, MASTER(cr), tabpfn, rtab,
3085 GMX_MAKETABLES_14ONLY);
3088 /* Read AdResS Thermo Force table if needed */
3089 if (fr->adress_icor == eAdressICThermoForce)
3091 /* old todo replace */
3093 if (ir->adress->n_tf_grps > 0)
3095 make_adress_tf_tables(fp, oenv, fr, ir, tabfn, mtop, box);
3100 /* load the default table */
3101 snew(fr->atf_tabs, 1);
3102 fr->atf_tabs[DEFAULT_TF_TABLE] = make_atf_table(fp, oenv, fr, tabafn, box);
3107 fr->nwall = ir->nwall;
3108 if (ir->nwall && ir->wall_type == ewtTABLE)
3110 make_wall_tables(fp, oenv, ir, tabfn, &mtop->groups, fr);
3115 fcd->bondtab = make_bonded_tables(fp,
3116 F_TABBONDS, F_TABBONDSNC,
3118 fcd->angletab = make_bonded_tables(fp,
3121 fcd->dihtab = make_bonded_tables(fp,
3129 fprintf(debug, "No fcdata or table file name passed, can not read table, can not do bonded interactions\n");
3133 /* QM/MM initialization if requested
3137 fprintf(stderr, "QM/MM calculation requested.\n");
3140 fr->bQMMM = ir->bQMMM;
3141 fr->qr = mk_QMMMrec();
3143 /* Set all the static charge group info */
3144 fr->cginfo_mb = init_cginfo_mb(fp, mtop, fr, bNoSolvOpt,
3146 &fr->bExcl_IntraCGAll_InterCGNone);
3147 if (DOMAINDECOMP(cr))
3153 fr->cginfo = cginfo_expand(mtop->nmolblock, fr->cginfo_mb);
3156 if (!DOMAINDECOMP(cr))
3158 forcerec_set_ranges(fr, ncg_mtop(mtop), ncg_mtop(mtop),
3159 mtop->natoms, mtop->natoms, mtop->natoms);
3162 fr->print_force = print_force;
3165 /* coarse load balancing vars */
3170 /* Initialize neighbor search */
3171 init_ns(fp, cr, &fr->ns, fr, mtop);
3173 if (cr->duty & DUTY_PP)
3175 gmx_nonbonded_setup(fr, bGenericKernelOnly);
3179 gmx_setup_adress_kernels(fp,bGenericKernelOnly);
3184 /* Initialize the thread working data for bonded interactions */
3185 init_forcerec_f_threads(fr, mtop->groups.grps[egcENER].nr);
3187 snew(fr->excl_load, fr->nthreads+1);
3189 if (fr->cutoff_scheme == ecutsVERLET)
3191 if (ir->rcoulomb != ir->rvdw)
3193 gmx_fatal(FARGS, "With Verlet lists rcoulomb and rvdw should be identical");
3196 init_nb_verlet(fp, &fr->nbv, bFEP_NonBonded, ir, fr, cr, nbpu_opt);
3199 /* fr->ic is used both by verlet and group kernels (to some extent) now */
3200 init_interaction_const(fp, cr, &fr->ic, fr, rtab);
3202 if (ir->eDispCorr != edispcNO)
3204 calc_enervirdiff(fp, ir->eDispCorr, fr);
3208 #define pr_real(fp, r) fprintf(fp, "%s: %e\n",#r, r)
3209 #define pr_int(fp, i) fprintf((fp), "%s: %d\n",#i, i)
3210 #define pr_bool(fp, b) fprintf((fp), "%s: %s\n",#b, bool_names[b])
3212 void pr_forcerec(FILE *fp, t_forcerec *fr)
3216 pr_real(fp, fr->rlist);
3217 pr_real(fp, fr->rcoulomb);
3218 pr_real(fp, fr->fudgeQQ);
3219 pr_bool(fp, fr->bGrid);
3220 pr_bool(fp, fr->bTwinRange);
3221 /*pr_int(fp,fr->cg0);
3222 pr_int(fp,fr->hcg);*/
3223 for (i = 0; i < fr->nnblists; i++)
3225 pr_int(fp, fr->nblists[i].table_elec_vdw.n);
3227 pr_real(fp, fr->rcoulomb_switch);
3228 pr_real(fp, fr->rcoulomb);
3233 void forcerec_set_excl_load(t_forcerec *fr,
3234 const gmx_localtop_t *top)
3237 int t, i, j, ntot, n, ntarget;
3239 ind = top->excls.index;
3243 for (i = 0; i < top->excls.nr; i++)
3245 for (j = ind[i]; j < ind[i+1]; j++)
3254 fr->excl_load[0] = 0;
3257 for (t = 1; t <= fr->nthreads; t++)
3259 ntarget = (ntot*t)/fr->nthreads;
3260 while (i < top->excls.nr && n < ntarget)
3262 for (j = ind[i]; j < ind[i+1]; j++)
3271 fr->excl_load[t] = i;