2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013,2014,2015,2016,2017, by the GROMACS development team, led by
5 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
6 * and including many others, as listed in the AUTHORS file in the
7 * top-level source directory and at http://www.gromacs.org.
9 * GROMACS is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public License
11 * as published by the Free Software Foundation; either version 2.1
12 * of the License, or (at your option) any later version.
14 * GROMACS is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with GROMACS; if not, see
21 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
22 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
24 * If you want to redistribute modifications to GROMACS, please
25 * consider that scientific software is very special. Version
26 * control is crucial - bugs must be traceable. We will be happy to
27 * consider code for inclusion in the official distribution, but
28 * derived work must not be called official GROMACS. Details are found
29 * in the README & COPYING files - if they are missing, get the
30 * official version at http://www.gromacs.org.
32 * To help us fund GROMACS development, we humbly ask that you cite
33 * the research papers on the package. Check out http://www.gromacs.org.
37 #include "calc_verletbuf.h"
46 #include "gromacs/ewald/ewald-utils.h"
47 #include "gromacs/math/functions.h"
48 #include "gromacs/math/units.h"
49 #include "gromacs/math/vec.h"
50 #include "gromacs/mdlib/nb_verlet.h"
51 #include "gromacs/mdlib/nbnxn_simd.h"
52 #include "gromacs/mdlib/nbnxn_util.h"
53 #include "gromacs/mdtypes/inputrec.h"
54 #include "gromacs/mdtypes/md_enums.h"
55 #include "gromacs/topology/ifunc.h"
56 #include "gromacs/topology/topology.h"
57 #include "gromacs/utility/fatalerror.h"
58 #include "gromacs/utility/smalloc.h"
60 /* The code in this file estimates a pairlist buffer length
61 * given a target energy drift per atom per picosecond.
62 * This is done by estimating the drift given a buffer length.
63 * Ideally we would like to have a tight overestimate of the drift,
64 * but that can be difficult to achieve.
66 * Significant approximations used:
68 * Uniform particle density. UNDERESTIMATES the drift by rho_global/rho_local.
70 * Interactions don't affect particle motion. OVERESTIMATES the drift on longer
71 * time scales. This approximation probably introduces the largest errors.
73 * Only take one constraint per particle into account: OVERESTIMATES the drift.
75 * For rotating constraints assume the same functional shape for time scales
76 * where the constraints rotate significantly as the exact expression for
77 * short time scales. OVERESTIMATES the drift on long time scales.
79 * For non-linear virtual sites use the mass of the lightest constructing atom
80 * to determine the displacement. OVER/UNDERESTIMATES the drift, depending on
81 * the geometry and masses of constructing atoms.
83 * Note that the formulas for normal atoms and linear virtual sites are exact,
84 * apart from the first two approximations.
86 * Note that apart from the effect of the above approximations, the actual
87 * drift of the total energy of a system can be order of magnitude smaller
88 * due to cancellation of positive and negative drift for different pairs.
92 /* Struct for unique atom type for calculating the energy drift.
93 * The atom displacement depends on mass and constraints.
94 * The energy jump for given distance depend on LJ type and q.
98 atom_nonbonded_kinetic_prop_t prop; /* non-bonded and kinetic atom prop. */
99 int n; /* #atoms of this type in the system */
100 } verletbuf_atomtype_t;
102 // Struct for derivatives of a non-bonded interaction potential
105 real md1; // -V' at the cutoff
106 real d2; // V'' at the cutoff
107 real md3; // -V''' at the cutoff
110 void verletbuf_get_list_setup(bool makeSimdPairList,
111 bool makeGpuPairList,
112 verletbuf_list_setup_t *list_setup)
114 /* When calling this function we often don't know which kernel type we
115 * are going to use. W choose the kernel type with the smallest possible
116 * i- and j-cluster sizes, so we potentially overestimate, but never
117 * underestimate, the buffer drift.
118 * Note that the current buffer estimation code only handles clusters
119 * of size 1, 2 or 4, so for 4x8 or 8x8 we use the estimate for 4x4.
124 /* The GPU kernels split the j-clusters in two halves */
125 list_setup->cluster_size_i = nbnxn_kernel_to_cluster_i_size(nbnxnk8x8x8_GPU);
126 list_setup->cluster_size_j = nbnxn_kernel_to_cluster_j_size(nbnxnk8x8x8_GPU)/2;
132 kernel_type = nbnxnk4x4_PlainC;
134 if (GMX_SIMD && makeSimdPairList)
136 #ifdef GMX_NBNXN_SIMD_2XNN
137 /* We use the smallest cluster size to be on the safe side */
138 kernel_type = nbnxnk4xN_SIMD_2xNN;
140 kernel_type = nbnxnk4xN_SIMD_4xN;
144 list_setup->cluster_size_i = nbnxn_kernel_to_cluster_i_size(kernel_type);
145 list_setup->cluster_size_j = nbnxn_kernel_to_cluster_j_size(kernel_type);
150 atom_nonbonded_kinetic_prop_equal(const atom_nonbonded_kinetic_prop_t *prop1,
151 const atom_nonbonded_kinetic_prop_t *prop2)
153 return (prop1->mass == prop2->mass &&
154 prop1->type == prop2->type &&
155 prop1->q == prop2->q &&
156 prop1->bConstr == prop2->bConstr &&
157 prop1->con_mass == prop2->con_mass &&
158 prop1->con_len == prop2->con_len);
161 static void add_at(verletbuf_atomtype_t **att_p, int *natt_p,
162 const atom_nonbonded_kinetic_prop_t *prop,
165 verletbuf_atomtype_t *att;
170 /* Ignore massless particles */
178 while (i < natt && !atom_nonbonded_kinetic_prop_equal(prop, &att[i].prop))
190 srenew(*att_p, *natt_p);
191 (*att_p)[i].prop = *prop;
192 (*att_p)[i].n = nmol;
196 static void get_vsite_masses(const gmx_moltype_t *moltype,
197 const gmx_ffparams_t *ffparams,
206 /* Check for virtual sites, determine mass from constructing atoms */
207 for (ft = 0; ft < F_NRE; ft++)
211 il = &moltype->ilist[ft];
213 for (i = 0; i < il->nr; i += 1+NRAL(ft))
216 real inv_mass, coeff, m_aj;
219 ip = &ffparams->iparams[il->iatoms[i]];
221 a1 = il->iatoms[i+1];
225 /* Only vsiten can have more than four
226 constructing atoms, so NRAL(ft) <= 5 */
229 const int maxj = NRAL(ft);
233 for (j = 1; j < maxj; j++)
235 cam[j] = moltype->atoms.atom[il->iatoms[i+1+j]].m;
238 cam[j] = vsite_m[il->iatoms[i+1+j]];
242 gmx_fatal(FARGS, "In molecule type '%s' %s construction involves atom %d, which is a virtual site of equal or high complexity. This is not supported.",
244 interaction_function[ft].longname,
245 il->iatoms[i+1+j]+1);
253 vsite_m[a1] = (cam[1]*cam[2])/(cam[2]*gmx::square(1-ip->vsite.a) + cam[1]*gmx::square(ip->vsite.a));
257 vsite_m[a1] = (cam[1]*cam[2]*cam[3])/(cam[2]*cam[3]*gmx::square(1-ip->vsite.a-ip->vsite.b) + cam[1]*cam[3]*gmx::square(ip->vsite.a) + cam[1]*cam[2]*gmx::square(ip->vsite.b));
260 gmx_incons("Invalid vsite type");
263 /* Use the mass of the lightest constructing atom.
264 * This is an approximation.
265 * If the distance of the virtual site to the
266 * constructing atom is less than all distances
267 * between constructing atoms, this is a safe
268 * over-estimate of the displacement of the vsite.
269 * This condition holds for all H mass replacement
270 * vsite constructions, except for SP2/3 groups.
271 * In SP3 groups one H will have a F_VSITE3
272 * construction, so even there the total drift
273 * estimate shouldn't be far off.
275 vsite_m[a1] = cam[1];
276 for (j = 2; j < maxj; j++)
278 vsite_m[a1] = std::min(vsite_m[a1], cam[j]);
291 for (j = 0; j < 3*ffparams->iparams[il->iatoms[i]].vsiten.n; j += 3)
293 aj = il->iatoms[i+j+2];
294 coeff = ffparams->iparams[il->iatoms[i+j]].vsiten.a;
295 if (moltype->atoms.atom[aj].ptype == eptVSite)
301 m_aj = moltype->atoms.atom[aj].m;
305 gmx_incons("The mass of a vsiten constructing atom is <= 0");
307 inv_mass += coeff*coeff/m_aj;
309 vsite_m[a1] = 1/inv_mass;
310 /* Correct for loop increment of i */
311 i += j - 1 - NRAL(ft);
315 fprintf(debug, "atom %4d %-20s mass %6.3f\n",
316 a1, interaction_function[ft].longname, vsite_m[a1]);
323 static void get_verlet_buffer_atomtypes(const gmx_mtop_t *mtop,
324 verletbuf_atomtype_t **att_p,
328 verletbuf_atomtype_t *att;
330 int mb, nmol, ft, i, a1, a2, a3, a;
331 const t_atoms *atoms;
334 atom_nonbonded_kinetic_prop_t *prop;
336 int n_nonlin_vsite_mol;
341 if (n_nonlin_vsite != nullptr)
346 for (mb = 0; mb < mtop->nmolblock; mb++)
348 nmol = mtop->molblock[mb].nmol;
350 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
352 /* Check for constraints, as they affect the kinetic energy.
353 * For virtual sites we need the masses and geometry of
354 * the constructing atoms to determine their velocity distribution.
356 snew(prop, atoms->nr);
357 snew(vsite_m, atoms->nr);
359 for (ft = F_CONSTR; ft <= F_CONSTRNC; ft++)
361 il = &mtop->moltype[mtop->molblock[mb].type].ilist[ft];
363 for (i = 0; i < il->nr; i += 1+NRAL(ft))
365 ip = &mtop->ffparams.iparams[il->iatoms[i]];
366 a1 = il->iatoms[i+1];
367 a2 = il->iatoms[i+2];
368 if (atoms->atom[a2].m > prop[a1].con_mass)
370 prop[a1].con_mass = atoms->atom[a2].m;
371 prop[a1].con_len = ip->constr.dA;
373 if (atoms->atom[a1].m > prop[a2].con_mass)
375 prop[a2].con_mass = atoms->atom[a1].m;
376 prop[a2].con_len = ip->constr.dA;
381 il = &mtop->moltype[mtop->molblock[mb].type].ilist[F_SETTLE];
383 for (i = 0; i < il->nr; i += 1+NRAL(F_SETTLE))
385 ip = &mtop->ffparams.iparams[il->iatoms[i]];
386 a1 = il->iatoms[i+1];
387 a2 = il->iatoms[i+2];
388 a3 = il->iatoms[i+3];
389 /* Usually the mass of a1 (usually oxygen) is larger than a2/a3.
390 * If this is not the case, we overestimate the displacement,
391 * which leads to a larger buffer (ok since this is an exotic case).
393 prop[a1].con_mass = atoms->atom[a2].m;
394 prop[a1].con_len = ip->settle.doh;
396 prop[a2].con_mass = atoms->atom[a1].m;
397 prop[a2].con_len = ip->settle.doh;
399 prop[a3].con_mass = atoms->atom[a1].m;
400 prop[a3].con_len = ip->settle.doh;
403 get_vsite_masses(&mtop->moltype[mtop->molblock[mb].type],
406 &n_nonlin_vsite_mol);
407 if (n_nonlin_vsite != nullptr)
409 *n_nonlin_vsite += nmol*n_nonlin_vsite_mol;
412 for (a = 0; a < atoms->nr; a++)
414 if (atoms->atom[a].ptype == eptVSite)
416 prop[a].mass = vsite_m[a];
420 prop[a].mass = atoms->atom[a].m;
422 prop[a].type = atoms->atom[a].type;
423 prop[a].q = atoms->atom[a].q;
424 /* We consider an atom constrained, #DOF=2, when it is
425 * connected with constraints to (at least one) atom with
426 * a mass of more than 0.4x its own mass. This is not a critical
427 * parameter, since with roughly equal masses the unconstrained
428 * and constrained displacement will not differ much (and both
429 * overestimate the displacement).
431 prop[a].bConstr = (prop[a].con_mass > 0.4*prop[a].mass);
433 add_at(&att, &natt, &prop[a], nmol);
442 for (a = 0; a < natt; a++)
444 fprintf(debug, "type %d: m %5.2f t %d q %6.3f con %d con_m %5.3f con_l %5.3f n %d\n",
445 a, att[a].prop.mass, att[a].prop.type, att[a].prop.q,
446 att[a].prop.bConstr, att[a].prop.con_mass, att[a].prop.con_len,
455 /* This function computes two components of the estimate of the variance
456 * in the displacement of one atom in a system of two constrained atoms.
457 * Returns in sigma2_2d the variance due to rotation of the constrained
458 * atom around the atom to which it constrained.
459 * Returns in sigma2_3d the variance due to displacement of the COM
460 * of the whole system of the two constrained atoms.
462 * Note that we only take a single constraint (the one to the heaviest atom)
463 * into account. If an atom has multiple constraints, this will result in
464 * an overestimate of the displacement, which gives a larger drift and buffer.
466 void constrained_atom_sigma2(real kT_fac,
467 const atom_nonbonded_kinetic_prop_t *prop,
471 /* Here we decompose the motion of a constrained atom into two
472 * components: rotation around the COM and translation of the COM.
475 /* Determine the variance of the arc length for the two rotational DOFs */
476 real massFraction = prop->con_mass/(prop->mass + prop->con_mass);
477 real sigma2_rot = kT_fac*massFraction/prop->mass;
479 /* The distance from the atom to the COM, i.e. the rotational arm */
480 real comDistance = prop->con_len*massFraction;
482 /* The variance relative to the arm */
483 real sigma2_rel = sigma2_rot/gmx::square(comDistance);
485 /* For sigma2_rel << 1 we don't notice the rotational effect and
486 * we have a normal, Gaussian displacement distribution.
487 * For larger sigma2_rel the displacement is much less, in fact it can
488 * not exceed 2*comDistance. We can calculate MSD/arm^2 as:
489 * integral_x=0-inf distance2(x) x/sigma2_rel exp(-x^2/(2 sigma2_rel)) dx
490 * where x is angular displacement and distance2(x) is the distance^2
491 * between points at angle 0 and x:
492 * distance2(x) = (sin(x) - sin(0))^2 + (cos(x) - cos(0))^2
493 * The limiting value of this MSD is 2, which is also the value for
494 * a uniform rotation distribution that would be reached at long time.
495 * The maximum is 2.5695 at sigma2_rel = 4.5119.
496 * We approximate this integral with a rational polynomial with
497 * coefficients from a Taylor expansion. This approximation is an
498 * overestimate for all values of sigma2_rel. Its maximum value
499 * of 2.6491 is reached at sigma2_rel = sqrt(45/2) = 4.7434.
500 * We keep the approximation constant after that.
501 * We use this approximate MSD as the variance for a Gaussian distribution.
503 * NOTE: For any sensible buffer tolerance this will result in a (large)
504 * overestimate of the buffer size, since the Gaussian has a long tail,
505 * whereas the actual distribution can not reach values larger than 2.
507 /* Coeffients obtained from a Taylor expansion */
508 const real a = 1.0/3.0;
509 const real b = 2.0/45.0;
511 /* Our approximation is constant after sigma2_rel = 1/sqrt(b) */
512 sigma2_rel = std::min(sigma2_rel, 1/std::sqrt(b));
514 /* Compute the approximate sigma^2 for 2D motion due to the rotation */
515 *sigma2_2d = gmx::square(comDistance)*
516 sigma2_rel/(1 + a*sigma2_rel + b*gmx::square(sigma2_rel));
518 /* The constrained atom also moves (in 3D) with the COM of both atoms */
519 *sigma2_3d = kT_fac/(prop->mass + prop->con_mass);
522 static void get_atom_sigma2(real kT_fac,
523 const atom_nonbonded_kinetic_prop_t *prop,
529 /* Complicated constraint calculation in a separate function */
530 constrained_atom_sigma2(kT_fac, prop, sigma2_2d, sigma2_3d);
534 /* Unconstrained atom: trivial */
536 *sigma2_3d = kT_fac/prop->mass;
540 static void approx_2dof(real s2, real x, real *shift, real *scale)
542 /* A particle with 1 DOF constrained has 2 DOFs instead of 3.
543 * This code is also used for particles with multiple constraints,
544 * in which case we overestimate the displacement.
545 * The 2DOF distribution is sqrt(pi/2)*erfc(r/(sqrt(2)*s))/(2*s).
546 * We approximate this with scale*Gaussian(s,r+shift),
547 * by matching the distribution value and derivative at x.
548 * This is a tight overestimate for all r>=0 at any s and x.
552 ex = std::exp(-x*x/(2*s2));
553 er = std::erfc(x/std::sqrt(2*s2));
555 *shift = -x + std::sqrt(2*s2/M_PI)*ex/er;
556 *scale = 0.5*M_PI*std::exp(ex*ex/(M_PI*er*er))*er;
559 // Returns an (over)estimate of the energy drift for a single atom pair,
560 // given the kinetic properties, displacement variances and list buffer.
561 static real energyDriftAtomPair(const atom_nonbonded_kinetic_prop_t *prop_i,
562 const atom_nonbonded_kinetic_prop_t *prop_j,
563 real s2, real s2i_2d, real s2j_2d,
565 const pot_derivatives_t *der)
567 // For relatively small arguments erfc() is so small that if will be 0.0
568 // when stored in a float. We set an argument limit of 8 (Erfc(8)=1e-29),
569 // such that we can divide by erfc and have some space left for arithmetic.
570 const real erfc_arg_max = 8.0;
577 if (rsh*rsh > 2*s2*erfc_arg_max*erfc_arg_max)
579 // Below we calculate c_erfc = 0.5*erfc(rsh/sqrt(2*s2))
580 // When rsh/sqrt(2*s2) increases, this erfc will be the first
581 // result that underflows and becomes 0.0. To avoid this,
582 // we set c_exp=0 and c_erfc=0 for large arguments.
583 // This also avoids NaN in approx_2dof().
584 // In any relevant case this has no effect on the results,
585 // since c_exp < 6e-29, so the displacement is completely
586 // negligible for such atom pairs (and an overestimate).
587 // In nearly all use cases, there will be other atom pairs
588 // that contribute much more to the total, so zeroing
589 // this particular contribution has no effect at all.
595 /* For constraints: adapt r and scaling for the Gaussian */
600 approx_2dof(s2i_2d, r_buffer*s2i_2d/s2, &sh, &sc);
608 approx_2dof(s2j_2d, r_buffer*s2j_2d/s2, &sh, &sc);
613 /* Exact contribution of an atom pair with Gaussian displacement
614 * with sigma s to the energy drift for a potential with
615 * derivative -md and second derivative dd at the cut-off.
616 * The only catch is that for potentials that change sign
617 * near the cut-off there could be an unlucky compensation
618 * of positive and negative energy drift.
619 * Such potentials are extremely rare though.
621 * Note that pot has unit energy*length, as the linear
622 * atom density still needs to be put in.
624 c_exp = std::exp(-rsh*rsh/(2*s2))/std::sqrt(2*M_PI);
625 c_erfc = 0.5*std::erfc(rsh/(std::sqrt(2*s2)));
627 real s = std::sqrt(s2);
631 der->md1/2*((rsh2 + s2)*c_erfc - rsh*s*c_exp);
633 der->d2/6*(s*(rsh2 + 2*s2)*c_exp - rsh*(rsh2 + 3*s2)*c_erfc);
635 der->md3/24*((rsh2*rsh2 + 6*rsh2*s2 + 3*s2*s2)*c_erfc - rsh*s*(rsh2 + 5*s2)*c_exp);
637 return pot1 + pot2 + pot3;
640 static real energyDrift(const verletbuf_atomtype_t *att, int natt,
641 const gmx_ffparams_t *ffp,
643 const pot_derivatives_t *ljDisp,
644 const pot_derivatives_t *ljRep,
645 const pot_derivatives_t *elec,
646 real rlj, real rcoulomb,
647 real rlist, real boxvol)
649 double drift_tot = 0;
653 /* No atom displacements: no drift, avoid division by 0 */
657 // Here add up the contribution of all atom pairs in the system to
658 // (estimated) energy drift by looping over all atom type pairs.
659 for (int i = 0; i < natt; i++)
661 // Get the thermal displacement variance for the i-atom type
662 const atom_nonbonded_kinetic_prop_t *prop_i = &att[i].prop;
664 get_atom_sigma2(kT_fac, prop_i, &s2i_2d, &s2i_3d);
666 for (int j = i; j < natt; j++)
668 // Get the thermal displacement variance for the j-atom type
669 const atom_nonbonded_kinetic_prop_t *prop_j = &att[j].prop;
671 get_atom_sigma2(kT_fac, prop_j, &s2j_2d, &s2j_3d);
673 /* Add up the up to four independent variances */
674 real s2 = s2i_2d + s2i_3d + s2j_2d + s2j_3d;
676 // Set -V', V'' and -V''' at the cut-off for LJ */
677 real c6 = ffp->iparams[prop_i->type*ffp->atnr + prop_j->type].lj.c6;
678 real c12 = ffp->iparams[prop_i->type*ffp->atnr + prop_j->type].lj.c12;
679 pot_derivatives_t lj;
680 lj.md1 = c6*ljDisp->md1 + c12*ljRep->md1;
681 lj.d2 = c6*ljDisp->d2 + c12*ljRep->d2;
682 lj.md3 = c6*ljDisp->md3 + c12*ljRep->md3;
684 real pot_lj = energyDriftAtomPair(prop_i, prop_j,
689 // Set -V' and V'' at the cut-off for Coulomb
690 pot_derivatives_t elec_qq;
691 elec_qq.md1 = elec->md1*prop_i->q*prop_j->q;
692 elec_qq.d2 = elec->d2 *prop_i->q*prop_j->q;
695 real pot_q = energyDriftAtomPair(prop_i, prop_j,
700 // Note that attractive and repulsive potentials for individual
701 // pairs can partially cancel.
702 real pot = pot_lj + pot_q;
704 /* Multiply by the number of atom pairs */
707 pot *= (double)att[i].n*(att[i].n - 1)/2;
711 pot *= (double)att[i].n*att[j].n;
713 /* We need the line density to get the energy drift of the system.
714 * The effective average r^2 is close to (rlist+sigma)^2.
716 pot *= 4*M_PI*gmx::square(rlist + std::sqrt(s2))/boxvol;
718 /* Add the unsigned drift to avoid cancellation of errors */
719 drift_tot += std::abs(pot);
726 static real surface_frac(int cluster_size, real particle_distance, real rlist)
730 if (rlist < 0.5*particle_distance)
732 /* We have non overlapping spheres */
736 /* Half the inter-particle distance relative to rlist */
737 d = 0.5*particle_distance/rlist;
739 /* Determine the area of the surface at distance rlist to the closest
740 * particle, relative to surface of a sphere of radius rlist.
741 * The formulas below assume close to cubic cells for the pair search grid,
742 * which the pair search code tries to achieve.
743 * Note that in practice particle distances will not be delta distributed,
744 * but have some spread, often involving shorter distances,
745 * as e.g. O-H bonds in a water molecule. Thus the estimates below will
746 * usually be slightly too high and thus conservative.
748 switch (cluster_size)
751 /* One particle: trivial */
755 /* Two particles: two spheres at fractional distance 2*a */
759 /* We assume a perfect, symmetric tetrahedron geometry.
760 * The surface around a tetrahedron is too complex for a full
761 * analytical solution, so we use a Taylor expansion.
763 area_rel = (1.0 + 1/M_PI*(6*std::acos(1/std::sqrt(3))*d +
764 std::sqrt(3)*d*d*(1.0 +
767 83.0/756.0*d*d*d*d*d*d)));
770 gmx_incons("surface_frac called with unsupported cluster_size");
774 return area_rel/cluster_size;
777 /* Returns the negative of the third derivative of a potential r^-p
778 * with a force-switch function, evaluated at the cut-off rc.
780 static real md3_force_switch(real p, real rswitch, real rc)
782 /* The switched force function is:
783 * p*r^-(p+1) + a*(r - rswitch)^2 + b*(r - rswitch)^3
786 real md3_pot, md3_sw;
788 a = -((p + 4)*rc - (p + 1)*rswitch)/(pow(rc, p+2)*gmx::square(rc-rswitch));
789 b = ((p + 3)*rc - (p + 1)*rswitch)/(pow(rc, p+2)*gmx::power3(rc-rswitch));
791 md3_pot = (p + 2)*(p + 1)*p*pow(rc, p+3);
792 md3_sw = 2*a + 6*b*(rc - rswitch);
794 return md3_pot + md3_sw;
797 void calc_verlet_buffer_size(const gmx_mtop_t *mtop, real boxvol,
798 const t_inputrec *ir,
801 real reference_temperature,
802 const verletbuf_list_setup_t *list_setup,
809 real particle_distance;
810 real nb_clust_frac_pairs_not_in_list_at_cutoff;
812 verletbuf_atomtype_t *att = nullptr;
815 real kT_fac, mass_min;
820 if (!EI_DYNAMICS(ir->eI))
822 gmx_incons("Can only determine the Verlet buffer size for integrators that perform dynamics");
824 if (ir->verletbuf_tol <= 0)
826 gmx_incons("The Verlet buffer tolerance needs to be larger than zero");
829 if (reference_temperature < 0)
831 if (EI_MD(ir->eI) && ir->etc == etcNO)
833 /* This case should be handled outside calc_verlet_buffer_size */
834 gmx_incons("calc_verlet_buffer_size called with an NVE ensemble and reference_temperature < 0");
837 /* We use the maximum temperature with multiple T-coupl groups.
838 * We could use a per particle temperature, but since particles
839 * interact, this might underestimate the buffer size.
841 reference_temperature = 0;
842 for (i = 0; i < ir->opts.ngtc; i++)
844 if (ir->opts.tau_t[i] >= 0)
846 reference_temperature = std::max(reference_temperature,
852 /* Resolution of the buffer size */
855 env = getenv("GMX_VERLET_BUFFER_RES");
858 sscanf(env, "%lf", &resolution);
861 /* In an atom wise pair-list there would be no pairs in the list
862 * beyond the pair-list cut-off.
863 * However, we use a pair-list of groups vs groups of atoms.
864 * For groups of 4 atoms, the parallelism of SSE instructions, only
865 * 10% of the atoms pairs are not in the list just beyond the cut-off.
866 * As this percentage increases slowly compared to the decrease of the
867 * Gaussian displacement distribution over this range, we can simply
868 * reduce the drift by this fraction.
869 * For larger groups, e.g. of 8 atoms, this fraction will be lower,
870 * so then buffer size will be on the conservative (large) side.
872 * Note that the formulas used here do not take into account
873 * cancellation of errors which could occur by missing both
874 * attractive and repulsive interactions.
876 * The only major assumption is homogeneous particle distribution.
877 * For an inhomogeneous system, such as a liquid-vapor system,
878 * the buffer will be underestimated. The actual energy drift
879 * will be higher by the factor: local/homogeneous particle density.
881 * The results of this estimate have been checked againt simulations.
882 * In most cases the real drift differs by less than a factor 2.
885 /* Worst case assumption: HCP packing of particles gives largest distance */
886 particle_distance = std::cbrt(boxvol*std::sqrt(2)/mtop->natoms);
888 get_verlet_buffer_atomtypes(mtop, &att, &natt, n_nonlin_vsite);
889 assert(att != NULL && natt >= 0);
893 fprintf(debug, "particle distance assuming HCP packing: %f nm\n",
895 fprintf(debug, "energy drift atom types: %d\n", natt);
898 pot_derivatives_t ljDisp = { 0, 0, 0 };
899 pot_derivatives_t ljRep = { 0, 0, 0 };
900 real repPow = mtop->ffparams.reppow;
902 if (ir->vdwtype == evdwCUT)
904 real sw_range, md3_pswf;
906 switch (ir->vdw_modifier)
909 case eintmodPOTSHIFT:
910 /* -dV/dr of -r^-6 and r^-reppow */
911 ljDisp.md1 = -6*std::pow(ir->rvdw, -7.0);
912 ljRep.md1 = repPow*std::pow(ir->rvdw, -(repPow + 1));
913 /* The contribution of the higher derivatives is negligible */
915 case eintmodFORCESWITCH:
916 /* At the cut-off: V=V'=V''=0, so we use only V''' */
917 ljDisp.md3 = -md3_force_switch(6.0, ir->rvdw_switch, ir->rvdw);
918 ljRep.md3 = md3_force_switch(repPow, ir->rvdw_switch, ir->rvdw);
920 case eintmodPOTSWITCH:
921 /* At the cut-off: V=V'=V''=0.
922 * V''' is given by the original potential times
923 * the third derivative of the switch function.
925 sw_range = ir->rvdw - ir->rvdw_switch;
926 md3_pswf = 60.0/gmx::power3(sw_range);
928 ljDisp.md3 = -std::pow(ir->rvdw, -6.0 )*md3_pswf;
929 ljRep.md3 = std::pow(ir->rvdw, -repPow)*md3_pswf;
932 gmx_incons("Unimplemented VdW modifier");
935 else if (EVDW_PME(ir->vdwtype))
937 real b = calc_ewaldcoeff_lj(ir->rvdw, ir->ewald_rtol_lj);
943 // -dV/dr of g(br)*r^-6 [where g(x) = exp(-x^2)(1+x^2+x^4/2),
944 // see LJ-PME equations in manual] and r^-reppow
945 ljDisp.md1 = -std::exp(-br2)*(br6 + 3.0*br4 + 6.0*br2 + 6.0)*std::pow(r, -7.0);
946 ljRep.md1 = repPow*pow(r, -(repPow + 1));
947 // The contribution of the higher derivatives is negligible
951 gmx_fatal(FARGS, "Energy drift calculation is only implemented for plain cut-off Lennard-Jones interactions");
954 elfac = ONE_4PI_EPS0/ir->epsilon_r;
956 // Determine the 1st and 2nd derivative for the electostatics
957 pot_derivatives_t elec = { 0, 0, 0 };
959 if (ir->coulombtype == eelCUT || EEL_RF(ir->coulombtype))
963 if (ir->coulombtype == eelCUT)
970 eps_rf = ir->epsilon_rf/ir->epsilon_r;
973 k_rf = (eps_rf - ir->epsilon_r)/( gmx::power3(ir->rcoulomb) * (2*eps_rf + ir->epsilon_r) );
977 /* epsilon_rf = infinity */
978 k_rf = 0.5/gmx::power3(ir->rcoulomb);
984 elec.md1 = elfac*(1.0/gmx::square(ir->rcoulomb) - 2*k_rf*ir->rcoulomb);
986 elec.d2 = elfac*(2.0/gmx::power3(ir->rcoulomb) + 2*k_rf);
988 else if (EEL_PME(ir->coulombtype) || ir->coulombtype == eelEWALD)
992 b = calc_ewaldcoeff_q(ir->rcoulomb, ir->ewald_rtol);
995 elec.md1 = elfac*(b*std::exp(-br*br)*M_2_SQRTPI/rc + std::erfc(br)/(rc*rc));
996 elec.d2 = elfac/(rc*rc)*(2*b*(1 + br*br)*std::exp(-br*br)*M_2_SQRTPI + 2*std::erfc(br)/rc);
1000 gmx_fatal(FARGS, "Energy drift calculation is only implemented for Reaction-Field and Ewald electrostatics");
1003 /* Determine the variance of the atomic displacement
1004 * over list_lifetime steps: kT_fac
1005 * For inertial dynamics (not Brownian dynamics) the mass factor
1006 * is not included in kT_fac, it is added later.
1010 /* Get the displacement distribution from the random component only.
1011 * With accurate integration the systematic (force) displacement
1012 * should be negligible (unless nstlist is extremely large, which
1013 * you wouldn't do anyhow).
1015 kT_fac = 2*BOLTZ*reference_temperature*list_lifetime*ir->delta_t;
1016 if (ir->bd_fric > 0)
1018 /* This is directly sigma^2 of the displacement */
1019 kT_fac /= ir->bd_fric;
1021 /* Set the masses to 1 as kT_fac is the full sigma^2,
1022 * but we divide by m in ener_drift().
1024 for (i = 0; i < natt; i++)
1026 att[i].prop.mass = 1;
1033 /* Per group tau_t is not implemented yet, use the maximum */
1034 tau_t = ir->opts.tau_t[0];
1035 for (i = 1; i < ir->opts.ngtc; i++)
1037 tau_t = std::max(tau_t, ir->opts.tau_t[i]);
1041 /* This kT_fac needs to be divided by the mass to get sigma^2 */
1046 kT_fac = BOLTZ*reference_temperature*gmx::square(list_lifetime*ir->delta_t);
1049 mass_min = att[0].prop.mass;
1050 for (i = 1; i < natt; i++)
1052 mass_min = std::min(mass_min, att[i].prop.mass);
1057 fprintf(debug, "Derivatives of non-bonded potentials at the cut-off:\n");
1058 fprintf(debug, "LJ disp. -V' %9.2e V'' %9.2e -V''' %9.2e\n", ljDisp.md1, ljDisp.d2, ljDisp.md3);
1059 fprintf(debug, "LJ rep. -V' %9.2e V'' %9.2e -V''' %9.2e\n", ljRep.md1, ljRep.d2, ljRep.md3);
1060 fprintf(debug, "Electro. -V' %9.2e V'' %9.2e\n", elec.md1, elec.d2);
1061 fprintf(debug, "sqrt(kT_fac) %f\n", std::sqrt(kT_fac));
1062 fprintf(debug, "mass_min %f\n", mass_min);
1065 /* Search using bisection */
1067 /* The drift will be neglible at 5 times the max sigma */
1068 ib1 = (int)(5*2*std::sqrt(kT_fac/mass_min)/resolution) + 1;
1069 while (ib1 - ib0 > 1)
1073 rl = std::max(ir->rvdw, ir->rcoulomb) + rb;
1075 /* Calculate the average energy drift at the last step
1076 * of the nstlist steps at which the pair-list is used.
1078 drift = energyDrift(att, natt, &mtop->ffparams,
1080 &ljDisp, &ljRep, &elec,
1081 ir->rvdw, ir->rcoulomb,
1084 /* Correct for the fact that we are using a Ni x Nj particle pair list
1085 * and not a 1 x 1 particle pair list. This reduces the drift.
1087 /* We don't have a formula for 8 (yet), use 4 which is conservative */
1088 nb_clust_frac_pairs_not_in_list_at_cutoff =
1089 surface_frac(std::min(list_setup->cluster_size_i, 4),
1090 particle_distance, rl)*
1091 surface_frac(std::min(list_setup->cluster_size_j, 4),
1092 particle_distance, rl);
1093 drift *= nb_clust_frac_pairs_not_in_list_at_cutoff;
1095 /* Convert the drift to drift per unit time per atom */
1096 drift /= nstlist*ir->delta_t*mtop->natoms;
1100 fprintf(debug, "ib %3d %3d %3d rb %.3f %dx%d fac %.3f drift %.1e\n",
1102 list_setup->cluster_size_i, list_setup->cluster_size_j,
1103 nb_clust_frac_pairs_not_in_list_at_cutoff,
1107 if (std::abs(drift) > ir->verletbuf_tol)
1119 *rlist = std::max(ir->rvdw, ir->rcoulomb) + ib1*resolution;