2 * This file is part of the GROMACS molecular simulation package.
4 * Copyright (c) 2012,2013, 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.
41 #include <sys/types.h>
46 #include "gmx_fatal.h"
50 #include "calc_verletbuf.h"
51 #include "../mdlib/nbnxn_consts.h"
54 /* The include below sets the SIMD instruction type (precision+width)
55 * for all nbnxn SIMD search and non-bonded kernel code.
57 #ifdef GMX_NBNXN_HALF_WIDTH_SIMD
58 #define GMX_USE_HALF_WIDTH_SIMD_HERE
60 #include "gromacs/simd/macros.h"
64 /* The code in this file estimates a pairlist buffer length
65 * given a target energy drift per atom per picosecond.
66 * This is done by estimating the drift given a buffer length.
67 * Ideally we would like to have a tight overestimate of the drift,
68 * but that can be difficult to achieve.
70 * Significant approximations used:
72 * Uniform particle density. UNDERESTIMATES the drift by rho_global/rho_local.
74 * Interactions don't affect particle motion. OVERESTIMATES the drift on longer
75 * time scales. This approximation probably introduces the largest errors.
77 * Only take one constraint per particle into account: OVERESTIMATES the drift.
79 * For rotating constraints assume the same functional shape for time scales
80 * where the constraints rotate significantly as the exact expression for
81 * short time scales. OVERESTIMATES the drift on long time scales.
83 * For non-linear virtual sites use the mass of the lightest constructing atom
84 * to determine the displacement. OVER/UNDERESTIMATES the drift, depending on
85 * the geometry and masses of constructing atoms.
87 * Note that the formulas for normal atoms and linear virtual sites are exact,
88 * apart from the first two approximations.
90 * Note that apart from the effect of the above approximations, the actual
91 * drift of the total energy of a system can be order of magnitude smaller
92 * due to cancellation of positive and negative drift for different pairs.
96 /* Struct for unique atom type for calculating the energy drift.
97 * The atom displacement depends on mass and constraints.
98 * The energy jump for given distance depend on LJ type and q.
102 real mass; /* mass */
103 int type; /* type (used for LJ parameters) */
105 gmx_bool bConstr; /* constrained, if TRUE, use #DOF=2 iso 3 */
106 real con_mass; /* mass of heaviest atom connected by constraints */
107 real con_len; /* constraint length to the heaviest atom */
108 } atom_nonbonded_kinetic_prop_t;
110 /* Struct for unique atom type for calculating the energy drift.
111 * The atom displacement depends on mass and constraints.
112 * The energy jump for given distance depend on LJ type and q.
116 atom_nonbonded_kinetic_prop_t prop; /* non-bonded and kinetic atom prop. */
117 int n; /* #atoms of this type in the system */
118 } verletbuf_atomtype_t;
120 void verletbuf_get_list_setup(gmx_bool bGPU,
121 verletbuf_list_setup_t *list_setup)
123 list_setup->cluster_size_i = NBNXN_CPU_CLUSTER_I_SIZE;
127 list_setup->cluster_size_j = NBNXN_GPU_CLUSTER_SIZE;
131 #ifndef GMX_NBNXN_SIMD
132 list_setup->cluster_size_j = NBNXN_CPU_CLUSTER_I_SIZE;
134 list_setup->cluster_size_j = GMX_SIMD_WIDTH_HERE;
135 #ifdef GMX_NBNXN_SIMD_2XNN
136 /* We assume the smallest cluster size to be on the safe side */
137 list_setup->cluster_size_j /= 2;
144 atom_nonbonded_kinetic_prop_equal(const atom_nonbonded_kinetic_prop_t *prop1,
145 const atom_nonbonded_kinetic_prop_t *prop2)
147 return (prop1->mass == prop2->mass &&
148 prop1->type == prop2->type &&
149 prop1->q == prop2->q &&
150 prop1->bConstr == prop2->bConstr &&
151 prop1->con_mass == prop2->con_mass &&
152 prop1->con_len == prop2->con_len);
155 static void add_at(verletbuf_atomtype_t **att_p, int *natt_p,
156 const atom_nonbonded_kinetic_prop_t *prop,
159 verletbuf_atomtype_t *att;
164 /* Ignore massless particles */
172 while (i < natt && !atom_nonbonded_kinetic_prop_equal(prop, &att[i].prop))
184 srenew(*att_p, *natt_p);
185 (*att_p)[i].prop = *prop;
186 (*att_p)[i].n = nmol;
190 static void get_vsite_masses(const gmx_moltype_t *moltype,
191 const gmx_ffparams_t *ffparams,
200 /* Check for virtual sites, determine mass from constructing atoms */
201 for (ft = 0; ft < F_NRE; ft++)
205 il = &moltype->ilist[ft];
207 for (i = 0; i < il->nr; i += 1+NRAL(ft))
210 real cam[5], inv_mass, m_aj;
211 int a1, j, aj, coeff;
213 ip = &ffparams->iparams[il->iatoms[i]];
215 a1 = il->iatoms[i+1];
219 for (j = 1; j < NRAL(ft); j++)
221 cam[j] = moltype->atoms.atom[il->iatoms[i+1+j]].m;
224 cam[j] = vsite_m[il->iatoms[i+1+j]];
228 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.",
230 interaction_function[ft].longname,
231 il->iatoms[i+1+j]+1);
240 vsite_m[a1] = (cam[1]*cam[2])/(cam[2]*sqr(1-ip->vsite.a) + cam[1]*sqr(ip->vsite.a));
244 vsite_m[a1] = (cam[1]*cam[2]*cam[3])/(cam[2]*cam[3]*sqr(1-ip->vsite.a-ip->vsite.b) + cam[1]*cam[3]*sqr(ip->vsite.a) + cam[1]*cam[2]*sqr(ip->vsite.b));
249 for (j = 0; j < 3*ip->vsiten.n; j += 3)
251 aj = il->iatoms[i+j+2];
252 coeff = ip[il->iatoms[i+j]].vsiten.a;
253 if (moltype->atoms.atom[aj].ptype == eptVSite)
259 m_aj = moltype->atoms.atom[aj].m;
263 gmx_incons("The mass of a vsiten constructing atom is <= 0");
265 inv_mass += coeff*coeff/m_aj;
267 vsite_m[a1] = 1/inv_mass;
270 /* Use the mass of the lightest constructing atom.
271 * This is an approximation.
272 * If the distance of the virtual site to the
273 * constructing atom is less than all distances
274 * between constructing atoms, this is a safe
275 * over-estimate of the displacement of the vsite.
276 * This condition holds for all H mass replacement
277 * vsite constructions, except for SP2/3 groups.
278 * In SP3 groups one H will have a F_VSITE3
279 * construction, so even there the total drift
280 * estimate shouldn't be far off.
283 vsite_m[a1] = cam[1];
284 for (j = 2; j < NRAL(ft); j++)
286 vsite_m[a1] = min(vsite_m[a1], cam[j]);
293 fprintf(debug, "atom %4d %-20s mass %6.3f\n",
294 a1, interaction_function[ft].longname, vsite_m[a1]);
301 static void get_verlet_buffer_atomtypes(const gmx_mtop_t *mtop,
302 verletbuf_atomtype_t **att_p,
306 verletbuf_atomtype_t *att;
308 int mb, nmol, ft, i, a1, a2, a3, a;
309 const t_atoms *atoms;
312 atom_nonbonded_kinetic_prop_t *prop;
314 int n_nonlin_vsite_mol;
319 if (n_nonlin_vsite != NULL)
324 for (mb = 0; mb < mtop->nmolblock; mb++)
326 nmol = mtop->molblock[mb].nmol;
328 atoms = &mtop->moltype[mtop->molblock[mb].type].atoms;
330 /* Check for constraints, as they affect the kinetic energy.
331 * For virtual sites we need the masses and geometry of
332 * the constructing atoms to determine their velocity distribution.
334 snew(prop, atoms->nr);
335 snew(vsite_m, atoms->nr);
337 for (ft = F_CONSTR; ft <= F_CONSTRNC; ft++)
339 il = &mtop->moltype[mtop->molblock[mb].type].ilist[ft];
341 for (i = 0; i < il->nr; i += 1+NRAL(ft))
343 ip = &mtop->ffparams.iparams[il->iatoms[i]];
344 a1 = il->iatoms[i+1];
345 a2 = il->iatoms[i+2];
346 if (atoms->atom[a2].m > prop[a1].con_mass)
348 prop[a1].con_mass = atoms->atom[a2].m;
349 prop[a1].con_len = ip->constr.dA;
351 if (atoms->atom[a1].m > prop[a2].con_mass)
353 prop[a2].con_mass = atoms->atom[a1].m;
354 prop[a2].con_len = ip->constr.dA;
359 il = &mtop->moltype[mtop->molblock[mb].type].ilist[F_SETTLE];
361 for (i = 0; i < il->nr; i += 1+NRAL(F_SETTLE))
363 ip = &mtop->ffparams.iparams[il->iatoms[i]];
364 a1 = il->iatoms[i+1];
365 a2 = il->iatoms[i+2];
366 a3 = il->iatoms[i+3];
367 /* Usually the mass of a1 (usually oxygen) is larger than a2/a3.
368 * If this is not the case, we overestimate the displacement,
369 * which leads to a larger buffer (ok since this is an exotic case).
371 prop[a1].con_mass = atoms->atom[a2].m;
372 prop[a1].con_len = ip->settle.doh;
374 prop[a2].con_mass = atoms->atom[a1].m;
375 prop[a2].con_len = ip->settle.doh;
377 prop[a3].con_mass = atoms->atom[a1].m;
378 prop[a3].con_len = ip->settle.doh;
381 get_vsite_masses(&mtop->moltype[mtop->molblock[mb].type],
384 &n_nonlin_vsite_mol);
385 if (n_nonlin_vsite != NULL)
387 *n_nonlin_vsite += nmol*n_nonlin_vsite_mol;
390 for (a = 0; a < atoms->nr; a++)
392 if (atoms->atom[a].ptype == eptVSite)
394 prop[a].mass = vsite_m[a];
398 prop[a].mass = atoms->atom[a].m;
400 prop[a].type = atoms->atom[a].type;
401 prop[a].q = atoms->atom[a].q;
402 /* We consider an atom constrained, #DOF=2, when it is
403 * connected with constraints to (at least one) atom with
404 * a mass of more than 0.4x its own mass. This is not a critical
405 * parameter, since with roughly equal masses the unconstrained
406 * and constrained displacement will not differ much (and both
407 * overestimate the displacement).
409 prop[a].bConstr = (prop[a].con_mass > 0.4*prop[a].mass);
411 add_at(&att, &natt, &prop[a], nmol);
420 for (a = 0; a < natt; a++)
422 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",
423 a, att[a].prop.mass, att[a].prop.type, att[a].prop.q,
424 att[a].prop.bConstr, att[a].prop.con_mass, att[a].prop.con_len,
433 /* This function computes two components of the estimate of the variance
434 * in the displacement of one atom in a system of two constrained atoms.
435 * Returns in sigma2_2d the variance due to rotation of the constrained
436 * atom around the atom to which it constrained.
437 * Returns in sigma2_3d the variance due to displacement of the COM
438 * of the whole system of the two constrained atoms.
440 * Note that we only take a single constraint (the one to the heaviest atom)
441 * into account. If an atom has multiple constraints, this will result in
442 * an overestimate of the displacement, which gives a larger drift and buffer.
444 static void constrained_atom_sigma2(real kT_fac,
445 const atom_nonbonded_kinetic_prop_t *prop,
454 /* Here we decompose the motion of a constrained atom into two
455 * components: rotation around the COM and translation of the COM.
458 /* Determine the variance for the displacement of the rotational mode */
459 sigma2_rot = kT_fac/(prop->mass*(prop->mass + prop->con_mass)/prop->con_mass);
461 /* The distance from the atom to the COM, i.e. the rotational arm */
462 com_dist = prop->con_len*prop->con_mass/(prop->mass + prop->con_mass);
464 /* The variance relative to the arm */
465 sigma2_rel = sigma2_rot/(com_dist*com_dist);
466 /* At 6 the scaling formula has slope 0,
467 * so we keep sigma2_2d constant after that.
471 /* A constrained atom rotates around the atom it is constrained to.
472 * This results in a smaller linear displacement than for a free atom.
473 * For a perfectly circular displacement, this lowers the displacement
474 * by: 1/arcsin(arc_length)
475 * and arcsin(x) = 1 + x^2/6 + ...
476 * For sigma2_rel<<1 the displacement distribution is erfc
477 * (exact formula is provided below). For larger sigma, it is clear
478 * that the displacement can't be larger than 2*com_dist.
479 * It turns out that the distribution becomes nearly uniform.
480 * For intermediate sigma2_rel, scaling down sigma with the third
481 * order expansion of arcsin with argument sigma_rel turns out
482 * to give a very good approximation of the distribution and variance.
483 * Even for larger values, the variance is only slightly overestimated.
484 * Note that the most relevant displacements are in the long tail.
485 * This rotation approximation always overestimates the tail (which
486 * runs to infinity, whereas it should be <= 2*com_dist).
487 * Thus we always overestimate the drift and the buffer size.
489 scale = 1/(1 + sigma2_rel/6);
490 *sigma2_2d = sigma2_rot*scale*scale;
494 /* sigma_2d is set to the maximum given by the scaling above.
495 * For large sigma2 the real displacement distribution is close
496 * to uniform over -2*con_len to 2*com_dist.
497 * Our erfc with sigma_2d=sqrt(1.5)*com_dist (which means the sigma
498 * of the erfc output distribution is con_dist) overestimates
499 * the variance and additionally has a long tail. This means
500 * we have a (safe) overestimation of the drift.
502 *sigma2_2d = 1.5*com_dist*com_dist;
505 /* The constrained atom also moves (in 3D) with the COM of both atoms */
506 *sigma2_3d = kT_fac/(prop->mass + prop->con_mass);
509 static void get_atom_sigma2(real kT_fac,
510 const atom_nonbonded_kinetic_prop_t *prop,
516 /* Complicated constraint calculation in a separate function */
517 constrained_atom_sigma2(kT_fac, prop, sigma2_2d, sigma2_3d);
521 /* Unconstrained atom: trivial */
523 *sigma2_3d = kT_fac/prop->mass;
527 static void approx_2dof(real s2, real x, real *shift, real *scale)
529 /* A particle with 1 DOF constrained has 2 DOFs instead of 3.
530 * This code is also used for particles with multiple constraints,
531 * in which case we overestimate the displacement.
532 * The 2DOF distribution is sqrt(pi/2)*erfc(r/(sqrt(2)*s))/(2*s).
533 * We approximate this with scale*Gaussian(s,r+shift),
534 * by matching the distribution value and derivative at x.
535 * This is a tight overestimate for all r>=0 at any s and x.
539 ex = exp(-x*x/(2*s2));
540 er = gmx_erfc(x/sqrt(2*s2));
542 *shift = -x + sqrt(2*s2/M_PI)*ex/er;
543 *scale = 0.5*M_PI*exp(ex*ex/(M_PI*er*er))*er;
546 static real ener_drift(const verletbuf_atomtype_t *att, int natt,
547 const gmx_ffparams_t *ffp,
549 real md_ljd, real md_ljr, real md_el, real dd_el,
551 real rlist, real boxvol)
553 double drift_tot, pot1, pot2, pot;
555 real s2i_2d, s2i_3d, s2j_2d, s2j_3d, s2, s;
559 double c_exp, c_erfc;
563 /* Loop over the different atom type pairs */
564 for (i = 0; i < natt; i++)
566 get_atom_sigma2(kT_fac, &att[i].prop, &s2i_2d, &s2i_3d);
567 ti = att[i].prop.type;
569 for (j = i; j < natt; j++)
571 get_atom_sigma2(kT_fac, &att[j].prop, &s2j_2d, &s2j_3d);
572 tj = att[j].prop.type;
574 /* Add up the up to four independent variances */
575 s2 = s2i_2d + s2i_3d + s2j_2d + s2j_3d;
577 /* Note that attractive and repulsive potentials for individual
578 * pairs will partially cancel.
580 /* -dV/dr at the cut-off for LJ + Coulomb */
582 md_ljd*ffp->iparams[ti*ffp->atnr+tj].lj.c6 +
583 md_ljr*ffp->iparams[ti*ffp->atnr+tj].lj.c12 +
584 md_el*att[i].prop.q*att[j].prop.q;
586 /* d2V/dr2 at the cut-off for Coulomb, we neglect LJ */
587 dd = dd_el*att[i].prop.q*att[j].prop.q;
591 /* For constraints: adapt r and scaling for the Gaussian */
592 if (att[i].prop.bConstr)
596 approx_2dof(s2i_2d, r_buffer*s2i_2d/s2, &sh, &sc);
600 if (att[j].prop.bConstr)
604 approx_2dof(s2j_2d, r_buffer*s2j_2d/s2, &sh, &sc);
609 /* Exact contribution of an atom pair with Gaussian displacement
610 * with sigma s to the energy drift for a potential with
611 * derivative -md and second derivative dd at the cut-off.
612 * The only catch is that for potentials that change sign
613 * near the cut-off there could be an unlucky compensation
614 * of positive and negative energy drift.
615 * Such potentials are extremely rare though.
617 * Note that pot has unit energy*length, as the linear
618 * atom density still needs to be put in.
620 c_exp = exp(-rsh*rsh/(2*s2))/sqrt(2*M_PI);
621 c_erfc = 0.5*gmx_erfc(rsh/(sqrt(2*s2)));
625 md/2*((rsh*rsh + s2)*c_erfc - rsh*s*c_exp);
627 dd/6*(s*(rsh*rsh + 2*s2)*c_exp - rsh*(rsh*rsh + 3*s2)*c_erfc);
632 fprintf(debug, "n %d %d d s %.3f %.3f %.3f %.3f con %d md %8.1e dd %8.1e pot1 %8.1e pot2 %8.1e pot %8.1e\n",
634 sqrt(s2i_2d), sqrt(s2i_3d),
635 sqrt(s2j_2d), sqrt(s2j_3d),
636 att[i].prop.bConstr+att[j].prop.bConstr,
637 md, dd, pot1, pot2, pot);
640 /* Multiply by the number of atom pairs */
643 pot *= (double)att[i].n*(att[i].n - 1)/2;
647 pot *= (double)att[i].n*att[j].n;
649 /* We need the line density to get the energy drift of the system.
650 * The effective average r^2 is close to (rlist+sigma)^2.
652 pot *= 4*M_PI*sqr(rlist + s)/boxvol;
654 /* Add the unsigned drift to avoid cancellation of errors */
655 drift_tot += fabs(pot);
662 static real surface_frac(int cluster_size, real particle_distance, real rlist)
666 if (rlist < 0.5*particle_distance)
668 /* We have non overlapping spheres */
672 /* Half the inter-particle distance relative to rlist */
673 d = 0.5*particle_distance/rlist;
675 /* Determine the area of the surface at distance rlist to the closest
676 * particle, relative to surface of a sphere of radius rlist.
677 * The formulas below assume close to cubic cells for the pair search grid,
678 * which the pair search code tries to achieve.
679 * Note that in practice particle distances will not be delta distributed,
680 * but have some spread, often involving shorter distances,
681 * as e.g. O-H bonds in a water molecule. Thus the estimates below will
682 * usually be slightly too high and thus conservative.
684 switch (cluster_size)
687 /* One particle: trivial */
691 /* Two particles: two spheres at fractional distance 2*a */
695 /* We assume a perfect, symmetric tetrahedron geometry.
696 * The surface around a tetrahedron is too complex for a full
697 * analytical solution, so we use a Taylor expansion.
699 area_rel = (1.0 + 1/M_PI*(6*acos(1/sqrt(3))*d +
703 83.0/756.0*d*d*d*d*d*d)));
706 gmx_incons("surface_frac called with unsupported cluster_size");
710 return area_rel/cluster_size;
713 void calc_verlet_buffer_size(const gmx_mtop_t *mtop, real boxvol,
714 const t_inputrec *ir,
715 const verletbuf_list_setup_t *list_setup,
722 real particle_distance;
723 real nb_clust_frac_pairs_not_in_list_at_cutoff;
725 verletbuf_atomtype_t *att = NULL;
728 real md_ljd, md_ljr, md_el, dd_el;
730 real kT_fac, mass_min;
735 /* Resolution of the buffer size */
738 env = getenv("GMX_VERLET_BUFFER_RES");
741 sscanf(env, "%lf", &resolution);
744 /* In an atom wise pair-list there would be no pairs in the list
745 * beyond the pair-list cut-off.
746 * However, we use a pair-list of groups vs groups of atoms.
747 * For groups of 4 atoms, the parallelism of SSE instructions, only
748 * 10% of the atoms pairs are not in the list just beyond the cut-off.
749 * As this percentage increases slowly compared to the decrease of the
750 * Gaussian displacement distribution over this range, we can simply
751 * reduce the drift by this fraction.
752 * For larger groups, e.g. of 8 atoms, this fraction will be lower,
753 * so then buffer size will be on the conservative (large) side.
755 * Note that the formulas used here do not take into account
756 * cancellation of errors which could occur by missing both
757 * attractive and repulsive interactions.
759 * The only major assumption is homogeneous particle distribution.
760 * For an inhomogeneous system, such as a liquid-vapor system,
761 * the buffer will be underestimated. The actual energy drift
762 * will be higher by the factor: local/homogeneous particle density.
764 * The results of this estimate have been checked againt simulations.
765 * In most cases the real drift differs by less than a factor 2.
768 /* Worst case assumption: HCP packing of particles gives largest distance */
769 particle_distance = pow(boxvol*sqrt(2)/mtop->natoms, 1.0/3.0);
771 get_verlet_buffer_atomtypes(mtop, &att, &natt, n_nonlin_vsite);
772 assert(att != NULL && natt >= 0);
776 fprintf(debug, "particle distance assuming HCP packing: %f nm\n",
778 fprintf(debug, "energy drift atom types: %d\n", natt);
781 reppow = mtop->ffparams.reppow;
784 if (ir->vdwtype == evdwCUT)
786 /* -dV/dr of -r^-6 and r^-repporw */
787 md_ljd = -6*pow(ir->rvdw, -7.0);
788 md_ljr = reppow*pow(ir->rvdw, -(reppow+1));
789 /* The contribution of the second derivative is negligible */
791 else if (EVDW_PME(ir->vdwtype))
793 real b, r, br, br2, br4, br6;
794 b = calc_ewaldcoeff_lj(ir->rvdw, ir->ewald_rtol_lj);
800 /* -dV/dr of g(br)*r^-6 [where g(x) = exp(-x^2)(1+x^2+x^4/2), see LJ-PME equations in manual] and r^-reppow */
801 md_ljd = -exp(-br2)*(br6 + 3.0*br4 + 6.0*br2 + 6.0)*pow(r, -7.0);
802 md_ljr = reppow*pow(r, -(reppow+1));
803 /* The contribution of the second derivative is negligible */
807 gmx_fatal(FARGS, "Energy drift calculation is only implemented for plain cut-off Lennard-Jones interactions");
810 elfac = ONE_4PI_EPS0/ir->epsilon_r;
812 /* Determine md=-dV/dr and dd=d^2V/dr^2 */
815 if (ir->coulombtype == eelCUT || EEL_RF(ir->coulombtype))
819 if (ir->coulombtype == eelCUT)
826 eps_rf = ir->epsilon_rf/ir->epsilon_r;
829 k_rf = pow(ir->rcoulomb, -3.0)*(eps_rf - ir->epsilon_r)/(2*eps_rf + ir->epsilon_r);
833 /* epsilon_rf = infinity */
834 k_rf = 0.5*pow(ir->rcoulomb, -3.0);
840 md_el = elfac*(pow(ir->rcoulomb, -2.0) - 2*k_rf*ir->rcoulomb);
842 dd_el = elfac*(2*pow(ir->rcoulomb, -3.0) + 2*k_rf);
844 else if (EEL_PME(ir->coulombtype) || ir->coulombtype == eelEWALD)
848 b = calc_ewaldcoeff_q(ir->rcoulomb, ir->ewald_rtol);
851 md_el = elfac*(b*exp(-br*br)*M_2_SQRTPI/rc + gmx_erfc(br)/(rc*rc));
852 dd_el = elfac/(rc*rc)*(2*b*(1 + br*br)*exp(-br*br)*M_2_SQRTPI + 2*gmx_erfc(br)/rc);
856 gmx_fatal(FARGS, "Energy drift calculation is only implemented for Reaction-Field and Ewald electrostatics");
859 /* Determine the variance of the atomic displacement
860 * over nstlist-1 steps: kT_fac
861 * For inertial dynamics (not Brownian dynamics) the mass factor
862 * is not included in kT_fac, it is added later.
866 /* Get the displacement distribution from the random component only.
867 * With accurate integration the systematic (force) displacement
868 * should be negligible (unless nstlist is extremely large, which
869 * you wouldn't do anyhow).
871 kT_fac = 2*BOLTZ*ir->opts.ref_t[0]*(ir->nstlist-1)*ir->delta_t;
874 /* This is directly sigma^2 of the displacement */
875 kT_fac /= ir->bd_fric;
877 /* Set the masses to 1 as kT_fac is the full sigma^2,
878 * but we divide by m in ener_drift().
880 for (i = 0; i < natt; i++)
882 att[i].prop.mass = 1;
889 /* Per group tau_t is not implemented yet, use the maximum */
890 tau_t = ir->opts.tau_t[0];
891 for (i = 1; i < ir->opts.ngtc; i++)
893 tau_t = max(tau_t, ir->opts.tau_t[i]);
897 /* This kT_fac needs to be divided by the mass to get sigma^2 */
902 kT_fac = BOLTZ*ir->opts.ref_t[0]*sqr((ir->nstlist-1)*ir->delta_t);
905 mass_min = att[0].prop.mass;
906 for (i = 1; i < natt; i++)
908 mass_min = min(mass_min, att[i].prop.mass);
913 fprintf(debug, "md_ljd %e md_ljr %e\n", md_ljd, md_ljr);
914 fprintf(debug, "md_el %e dd_el %e\n", md_el, dd_el);
915 fprintf(debug, "sqrt(kT_fac) %f\n", sqrt(kT_fac));
916 fprintf(debug, "mass_min %f\n", mass_min);
919 /* Search using bisection */
921 /* The drift will be neglible at 5 times the max sigma */
922 ib1 = (int)(5*2*sqrt(kT_fac/mass_min)/resolution) + 1;
923 while (ib1 - ib0 > 1)
927 rl = max(ir->rvdw, ir->rcoulomb) + rb;
929 /* Calculate the average energy drift at the last step
930 * of the nstlist steps at which the pair-list is used.
932 drift = ener_drift(att, natt, &mtop->ffparams,
934 md_ljd, md_ljr, md_el, dd_el, rb,
937 /* Correct for the fact that we are using a Ni x Nj particle pair list
938 * and not a 1 x 1 particle pair list. This reduces the drift.
940 /* We don't have a formula for 8 (yet), use 4 which is conservative */
941 nb_clust_frac_pairs_not_in_list_at_cutoff =
942 surface_frac(min(list_setup->cluster_size_i, 4),
943 particle_distance, rl)*
944 surface_frac(min(list_setup->cluster_size_j, 4),
945 particle_distance, rl);
946 drift *= nb_clust_frac_pairs_not_in_list_at_cutoff;
948 /* Convert the drift to drift per unit time per atom */
949 drift /= ir->nstlist*ir->delta_t*mtop->natoms;
953 fprintf(debug, "ib %3d %3d %3d rb %.3f %dx%d fac %.3f drift %f\n",
955 list_setup->cluster_size_i, list_setup->cluster_size_j,
956 nb_clust_frac_pairs_not_in_list_at_cutoff,
960 if (fabs(drift) > ir->verletbuf_tol)
972 *rlist = max(ir->rvdw, ir->rcoulomb) + ib1*resolution;