From: Mark Abraham <mark.j.abraham@gmail.com>
Date: Fri, 19 Feb 2016 09:18:47 +0000 (+0100)
Subject: Merge branch release-5-1
X-Git-Url: http://biod.pnpi.spb.ru/gitweb/?a=commitdiff_plain;h=bcb1cd438273af954d00f342d22318aac079d6b7;p=alexxy%2Fgromacs.git

Merge branch release-5-1

Change-Id: I8935d1351656d6b0888e923c8316fce5d5c27cde
---

bcb1cd438273af954d00f342d22318aac079d6b7
diff --cc src/gromacs/gmxpreprocess/toppush.cpp
index 10bd4a0218,0000000000..9bc3a53aec
mode 100644,000000..100644
--- a/src/gromacs/gmxpreprocess/toppush.cpp
+++ b/src/gromacs/gmxpreprocess/toppush.cpp
@@@ -1,2749 -1,0 +1,2741 @@@
 +/*
 + * This file is part of the GROMACS molecular simulation package.
 + *
 + * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
 + * Copyright (c) 2001-2004, The GROMACS development team.
 + * Copyright (c) 2013,2014,2015,2016, by the GROMACS development team, led by
 + * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
 + * and including many others, as listed in the AUTHORS file in the
 + * top-level source directory and at http://www.gromacs.org.
 + *
 + * GROMACS is free software; you can redistribute it and/or
 + * modify it under the terms of the GNU Lesser General Public License
 + * as published by the Free Software Foundation; either version 2.1
 + * of the License, or (at your option) any later version.
 + *
 + * GROMACS is distributed in the hope that it will be useful,
 + * but WITHOUT ANY WARRANTY; without even the implied warranty of
 + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 + * Lesser General Public License for more details.
 + *
 + * You should have received a copy of the GNU Lesser General Public
 + * License along with GROMACS; if not, see
 + * http://www.gnu.org/licenses, or write to the Free Software Foundation,
 + * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
 + *
 + * If you want to redistribute modifications to GROMACS, please
 + * consider that scientific software is very special. Version
 + * control is crucial - bugs must be traceable. We will be happy to
 + * consider code for inclusion in the official distribution, but
 + * derived work must not be called official GROMACS. Details are found
 + * in the README & COPYING files - if they are missing, get the
 + * official version at http://www.gromacs.org.
 + *
 + * To help us fund GROMACS development, we humbly ask that you cite
 + * the research papers on the package. Check out http://www.gromacs.org.
 + */
 +#include "gmxpre.h"
 +
 +#include "toppush.h"
 +
 +#include <ctype.h>
 +#include <stdlib.h>
 +
 +#include <cmath>
 +
 +#include <algorithm>
 +
 +#include "gromacs/fileio/warninp.h"
 +#include "gromacs/gmxpreprocess/gpp_atomtype.h"
 +#include "gromacs/gmxpreprocess/gpp_bond_atomtype.h"
 +#include "gromacs/gmxpreprocess/notset.h"
 +#include "gromacs/gmxpreprocess/readir.h"
 +#include "gromacs/gmxpreprocess/topdirs.h"
 +#include "gromacs/gmxpreprocess/toputil.h"
 +#include "gromacs/mdtypes/md_enums.h"
 +#include "gromacs/topology/ifunc.h"
 +#include "gromacs/topology/symtab.h"
 +#include "gromacs/utility/cstringutil.h"
 +#include "gromacs/utility/fatalerror.h"
 +#include "gromacs/utility/gmxassert.h"
 +#include "gromacs/utility/smalloc.h"
 +
 +void generate_nbparams(int comb, int ftype, t_params *plist, gpp_atomtype_t atype,
 +                       warninp_t wi)
 +{
 +    int   i, j, k = -1, nf;
 +    int   nr, nrfp;
 +    real  c, bi, bj, ci, cj, ci0, ci1, ci2, cj0, cj1, cj2;
 +    char  errbuf[256];
 +
 +    /* Lean mean shortcuts */
 +    nr   = get_atomtype_ntypes(atype);
 +    nrfp = NRFP(ftype);
 +    snew(plist->param, nr*nr);
 +    plist->nr = nr*nr;
 +
 +    /* Fill the matrix with force parameters */
 +    switch (ftype)
 +    {
 +        case F_LJ:
 +            switch (comb)
 +            {
 +                case eCOMB_GEOMETRIC:
 +                    /* Gromos rules */
 +                    for (i = k = 0; (i < nr); i++)
 +                    {
 +                        for (j = 0; (j < nr); j++, k++)
 +                        {
 +                            for (nf = 0; (nf < nrfp); nf++)
 +                            {
 +                                ci = get_atomtype_nbparam(i, nf, atype);
 +                                cj = get_atomtype_nbparam(j, nf, atype);
 +                                c  = std::sqrt(ci * cj);
 +                                plist->param[k].c[nf]      = c;
 +                            }
 +                        }
 +                    }
 +                    break;
 +
 +                case eCOMB_ARITHMETIC:
 +                    /* c0 and c1 are sigma and epsilon */
 +                    for (i = k = 0; (i < nr); i++)
 +                    {
 +                        for (j = 0; (j < nr); j++, k++)
 +                        {
 +                            ci0                  = get_atomtype_nbparam(i, 0, atype);
 +                            cj0                  = get_atomtype_nbparam(j, 0, atype);
 +                            ci1                  = get_atomtype_nbparam(i, 1, atype);
 +                            cj1                  = get_atomtype_nbparam(j, 1, atype);
 +                            plist->param[k].c[0] = (fabs(ci0) + fabs(cj0))*0.5;
 +                            /* Negative sigma signals that c6 should be set to zero later,
 +                             * so we need to propagate that through the combination rules.
 +                             */
 +                            if (ci0 < 0 || cj0 < 0)
 +                            {
 +                                plist->param[k].c[0] *= -1;
 +                            }
 +                            plist->param[k].c[1] = std::sqrt(ci1*cj1);
 +                        }
 +                    }
 +
 +                    break;
 +                case eCOMB_GEOM_SIG_EPS:
 +                    /* c0 and c1 are sigma and epsilon */
 +                    for (i = k = 0; (i < nr); i++)
 +                    {
 +                        for (j = 0; (j < nr); j++, k++)
 +                        {
 +                            ci0                  = get_atomtype_nbparam(i, 0, atype);
 +                            cj0                  = get_atomtype_nbparam(j, 0, atype);
 +                            ci1                  = get_atomtype_nbparam(i, 1, atype);
 +                            cj1                  = get_atomtype_nbparam(j, 1, atype);
 +                            plist->param[k].c[0] = std::sqrt(fabs(ci0*cj0));
 +                            /* Negative sigma signals that c6 should be set to zero later,
 +                             * so we need to propagate that through the combination rules.
 +                             */
 +                            if (ci0 < 0 || cj0 < 0)
 +                            {
 +                                plist->param[k].c[0] *= -1;
 +                            }
 +                            plist->param[k].c[1] = std::sqrt(ci1*cj1);
 +                        }
 +                    }
 +
 +                    break;
 +                default:
 +                    gmx_fatal(FARGS, "No such combination rule %d", comb);
 +            }
 +            if (plist->nr != k)
 +            {
 +                gmx_incons("Topology processing, generate nb parameters");
 +            }
 +            break;
 +
 +        case F_BHAM:
 +            /* Buckingham rules */
 +            for (i = k = 0; (i < nr); i++)
 +            {
 +                for (j = 0; (j < nr); j++, k++)
 +                {
 +                    ci0                  = get_atomtype_nbparam(i, 0, atype);
 +                    cj0                  = get_atomtype_nbparam(j, 0, atype);
 +                    ci2                  = get_atomtype_nbparam(i, 2, atype);
 +                    cj2                  = get_atomtype_nbparam(j, 2, atype);
 +                    bi                   = get_atomtype_nbparam(i, 1, atype);
 +                    bj                   = get_atomtype_nbparam(j, 1, atype);
 +                    plist->param[k].c[0] = std::sqrt(ci0 * cj0);
 +                    if ((bi == 0) || (bj == 0))
 +                    {
 +                        plist->param[k].c[1] = 0;
 +                    }
 +                    else
 +                    {
 +                        plist->param[k].c[1] = 2.0/(1/bi+1/bj);
 +                    }
 +                    plist->param[k].c[2] = std::sqrt(ci2 * cj2);
 +                }
 +            }
 +
 +            break;
 +        default:
 +            sprintf(errbuf, "Invalid nonbonded type %s",
 +                    interaction_function[ftype].longname);
 +            warning_error(wi, errbuf);
 +    }
 +}
 +
 +static void realloc_nb_params(gpp_atomtype_t at,
 +                              t_nbparam ***nbparam, t_nbparam ***pair)
 +{
 +    /* Add space in the non-bonded parameters matrix */
 +    int atnr = get_atomtype_ntypes(at);
 +    srenew(*nbparam, atnr);
 +    snew((*nbparam)[atnr-1], atnr);
 +    if (pair)
 +    {
 +        srenew(*pair, atnr);
 +        snew((*pair)[atnr-1], atnr);
 +    }
 +}
 +
 +static void copy_B_from_A(int ftype, double *c)
 +{
 +    int nrfpA, nrfpB, i;
 +
 +    nrfpA = NRFPA(ftype);
 +    nrfpB = NRFPB(ftype);
 +
 +    /* Copy the B parameters from the first nrfpB A parameters */
 +    for (i = 0; (i < nrfpB); i++)
 +    {
 +        c[nrfpA+i] = c[i];
 +    }
 +}
 +
 +void push_at (t_symtab *symtab, gpp_atomtype_t at, t_bond_atomtype bat,
 +              char *line, int nb_funct,
 +              t_nbparam ***nbparam, t_nbparam ***pair,
 +              warninp_t wi)
 +{
 +    typedef struct {
 +        const char *entry;
 +        int         ptype;
 +    } t_xlate;
 +    t_xlate    xl[eptNR] = {
 +        { "A",   eptAtom },
 +        { "N",   eptNucleus },
 +        { "S",   eptShell },
 +        { "B",   eptBond },
 +        { "V",   eptVSite },
 +    };
 +
 +    int        nr, i, nfields, j, pt, nfp0 = -1;
 +    int        batype_nr, nread;
 +    char       type[STRLEN], btype[STRLEN], ptype[STRLEN];
 +    double     m, q;
 +    double     c[MAXFORCEPARAM];
 +    double     radius, vol, surftens, gb_radius, S_hct;
 +    char       tmpfield[12][100]; /* Max 12 fields of width 100 */
 +    char       errbuf[256];
 +    t_atom    *atom;
 +    t_param   *param;
 +    int        atomnr;
 +    gmx_bool   have_atomic_number;
 +    gmx_bool   have_bonded_type;
 +
 +    snew(atom, 1);
 +    snew(param, 1);
 +
 +    /* First assign input line to temporary array */
 +    nfields = sscanf(line, "%s%s%s%s%s%s%s%s%s%s%s%s",
 +                     tmpfield[0], tmpfield[1], tmpfield[2], tmpfield[3], tmpfield[4], tmpfield[5],
 +                     tmpfield[6], tmpfield[7], tmpfield[8], tmpfield[9], tmpfield[10], tmpfield[11]);
 +
 +    /* Comments on optional fields in the atomtypes section:
 +     *
 +     * The force field format is getting a bit old. For OPLS-AA we needed
 +     * to add a special bonded atomtype, and for Gerrit Groenhofs QM/MM stuff
 +     * we also needed the atomic numbers.
 +     * To avoid making all old or user-generated force fields unusable we
 +     * have introduced both these quantities as optional columns, and do some
 +     * acrobatics to check whether they are present or not.
 +     * This will all look much nicer when we switch to XML... sigh.
 +     *
 +     * Field 0 (mandatory) is the nonbonded type name. (string)
 +     * Field 1 (optional)  is the bonded type (string)
 +     * Field 2 (optional)  is the atomic number (int)
 +     * Field 3 (mandatory) is the mass (numerical)
 +     * Field 4 (mandatory) is the charge (numerical)
 +     * Field 5 (mandatory) is the particle type (single character)
 +     * This is followed by a number of nonbonded parameters.
 +     *
 +     * The safest way to identify the format is the particle type field.
 +     *
 +     * So, here is what we do:
 +     *
 +     * A. Read in the first six fields as strings
 +     * B. If field 3 (starting from 0) is a single char, we have neither
 +     *    bonded_type or atomic numbers.
 +     * C. If field 5 is a single char we have both.
 +     * D. If field 4 is a single char we check field 1. If this begins with
 +     *    an alphabetical character we have bonded types, otherwise atomic numbers.
 +     */
 +
 +    if (nfields < 6)
 +    {
 +        too_few(wi);
 +        return;
 +    }
 +
 +    if ( (strlen(tmpfield[5]) == 1) && isalpha(tmpfield[5][0]) )
 +    {
 +        have_bonded_type   = TRUE;
 +        have_atomic_number = TRUE;
 +    }
 +    else if ( (strlen(tmpfield[3]) == 1) && isalpha(tmpfield[3][0]) )
 +    {
 +        have_bonded_type   = FALSE;
 +        have_atomic_number = FALSE;
 +    }
 +    else
 +    {
 +        have_bonded_type   = ( isalpha(tmpfield[1][0]) != 0 );
 +        have_atomic_number = !have_bonded_type;
 +    }
 +
 +    /* optional fields */
 +    surftens  = -1;
 +    vol       = -1;
 +    radius    = -1;
 +    gb_radius = -1;
 +    atomnr    = -1;
 +    S_hct     = -1;
 +
 +    switch (nb_funct)
 +    {
 +
 +        case F_LJ:
 +            nfp0 = 2;
 +
 +            if (have_atomic_number)
 +            {
 +                if (have_bonded_type)
 +                {
 +                    nread = sscanf(line, "%s%s%d%lf%lf%s%lf%lf%lf%lf%lf%lf",
 +                                   type, btype, &atomnr, &m, &q, ptype, &c[0], &c[1],
 +                                   &radius, &vol, &surftens, &gb_radius);
 +                    if (nread < 8)
 +                    {
 +                        too_few(wi);
 +                        return;
 +                    }
 +                }
 +                else
 +                {
 +                    /* have_atomic_number && !have_bonded_type */
 +                    nread = sscanf(line, "%s%d%lf%lf%s%lf%lf%lf%lf%lf%lf",
 +                                   type, &atomnr, &m, &q, ptype, &c[0], &c[1],
 +                                   &radius, &vol, &surftens, &gb_radius);
 +                    if (nread < 7)
 +                    {
 +                        too_few(wi);
 +                        return;
 +                    }
 +                }
 +            }
 +            else
 +            {
 +                if (have_bonded_type)
 +                {
 +                    /* !have_atomic_number && have_bonded_type */
 +                    nread = sscanf(line, "%s%s%lf%lf%s%lf%lf%lf%lf%lf%lf",
 +                                   type, btype, &m, &q, ptype, &c[0], &c[1],
 +                                   &radius, &vol, &surftens, &gb_radius);
 +                    if (nread < 7)
 +                    {
 +                        too_few(wi);
 +                        return;
 +                    }
 +                }
 +                else
 +                {
 +                    /* !have_atomic_number && !have_bonded_type */
 +                    nread = sscanf(line, "%s%lf%lf%s%lf%lf%lf%lf%lf%lf",
 +                                   type, &m, &q, ptype, &c[0], &c[1],
 +                                   &radius, &vol, &surftens, &gb_radius);
 +                    if (nread < 6)
 +                    {
 +                        too_few(wi);
 +                        return;
 +                    }
 +                }
 +            }
 +
 +            if (!have_bonded_type)
 +            {
 +                strcpy(btype, type);
 +            }
 +
 +            if (!have_atomic_number)
 +            {
 +                atomnr = -1;
 +            }
 +
 +            break;
 +
 +        case F_BHAM:
 +            nfp0 = 3;
 +
 +            if (have_atomic_number)
 +            {
 +                if (have_bonded_type)
 +                {
 +                    nread = sscanf(line, "%s%s%d%lf%lf%s%lf%lf%lf%lf%lf%lf%lf",
 +                                   type, btype, &atomnr, &m, &q, ptype, &c[0], &c[1], &c[2],
 +                                   &radius, &vol, &surftens, &gb_radius);
 +                    if (nread < 9)
 +                    {
 +                        too_few(wi);
 +                        return;
 +                    }
 +                }
 +                else
 +                {
 +                    /* have_atomic_number && !have_bonded_type */
 +                    nread = sscanf(line, "%s%d%lf%lf%s%lf%lf%lf%lf%lf%lf%lf",
 +                                   type, &atomnr, &m, &q, ptype, &c[0], &c[1], &c[2],
 +                                   &radius, &vol, &surftens, &gb_radius);
 +                    if (nread < 8)
 +                    {
 +                        too_few(wi);
 +                        return;
 +                    }
 +                }
 +            }
 +            else
 +            {
 +                if (have_bonded_type)
 +                {
 +                    /* !have_atomic_number && have_bonded_type */
 +                    nread = sscanf(line, "%s%s%lf%lf%s%lf%lf%lf%lf%lf%lf%lf",
 +                                   type, btype, &m, &q, ptype, &c[0], &c[1], &c[2],
 +                                   &radius, &vol, &surftens, &gb_radius);
 +                    if (nread < 8)
 +                    {
 +                        too_few(wi);
 +                        return;
 +                    }
 +                }
 +                else
 +                {
 +                    /* !have_atomic_number && !have_bonded_type */
 +                    nread = sscanf(line, "%s%lf%lf%s%lf%lf%lf%lf%lf%lf%lf",
 +                                   type, &m, &q, ptype, &c[0], &c[1], &c[2],
 +                                   &radius, &vol, &surftens, &gb_radius);
 +                    if (nread < 7)
 +                    {
 +                        too_few(wi);
 +                        return;
 +                    }
 +                }
 +            }
 +
 +            if (!have_bonded_type)
 +            {
 +                strcpy(btype, type);
 +            }
 +
 +            if (!have_atomic_number)
 +            {
 +                atomnr = -1;
 +            }
 +
 +            break;
 +
 +        default:
 +            gmx_fatal(FARGS, "Invalid function type %d in push_at %s %d", nb_funct,
 +                      __FILE__, __LINE__);
 +    }
 +    for (j = nfp0; (j < MAXFORCEPARAM); j++)
 +    {
 +        c[j] = 0.0;
 +    }
 +
 +    if (strlen(type) == 1 && isdigit(type[0]))
 +    {
 +        gmx_fatal(FARGS, "Atom type names can't be single digits.");
 +    }
 +
 +    if (strlen(btype) == 1 && isdigit(btype[0]))
 +    {
 +        gmx_fatal(FARGS, "Bond atom type names can't be single digits.");
 +    }
 +
 +    /* Hack to read old topologies */
 +    if (gmx_strcasecmp(ptype, "D") == 0)
 +    {
 +        sprintf(ptype, "V");
 +    }
 +    for (j = 0; (j < eptNR); j++)
 +    {
 +        if (gmx_strcasecmp(ptype, xl[j].entry) == 0)
 +        {
 +            break;
 +        }
 +    }
 +    if (j == eptNR)
 +    {
 +        gmx_fatal(FARGS, "Invalid particle type %s on line %s",
 +                  ptype, line);
 +    }
 +    pt = xl[j].ptype;
 +    if (debug)
 +    {
 +        fprintf(debug, "ptype: %s\n", ptype_str[pt]);
 +    }
 +
 +    atom->q     = q;
 +    atom->m     = m;
 +    atom->ptype = pt;
 +    for (i = 0; (i < MAXFORCEPARAM); i++)
 +    {
 +        param->c[i] = c[i];
 +    }
 +
 +    if ((batype_nr = get_bond_atomtype_type(btype, bat)) == NOTSET)
 +    {
 +        add_bond_atomtype(bat, symtab, btype);
 +    }
 +    batype_nr = get_bond_atomtype_type(btype, bat);
 +
 +    if ((nr = get_atomtype_type(type, at)) != NOTSET)
 +    {
 +        sprintf(errbuf, "Overriding atomtype %s", type);
 +        warning(wi, errbuf);
 +        if ((nr = set_atomtype(nr, at, symtab, atom, type, param, batype_nr,
 +                               radius, vol, surftens, atomnr, gb_radius, S_hct)) == NOTSET)
 +        {
 +            gmx_fatal(FARGS, "Replacing atomtype %s failed", type);
 +        }
 +    }
 +    else if ((add_atomtype(at, symtab, atom, type, param,
 +                           batype_nr, radius, vol,
 +                           surftens, atomnr, gb_radius, S_hct)) == NOTSET)
 +    {
 +        gmx_fatal(FARGS, "Adding atomtype %s failed", type);
 +    }
 +    else
 +    {
 +        /* Add space in the non-bonded parameters matrix */
 +        realloc_nb_params(at, nbparam, pair);
 +    }
 +    sfree(atom);
 +    sfree(param);
 +}
 +
 +static void push_bondtype(t_params     *       bt,
 +                          t_param     *        b,
 +                          int                  nral,
 +                          int                  ftype,
 +                          gmx_bool             bAllowRepeat,
 +                          char     *           line,
 +                          warninp_t            wi)
 +{
 +    int      i, j;
 +    gmx_bool bTest, bFound, bCont, bId;
 +    int      nr   = bt->nr;
 +    int      nrfp = NRFP(ftype);
 +    char     errbuf[256];
 +
 +    /* If bAllowRepeat is TRUE, we allow multiple entries as long as they
 +       are on directly _adjacent_ lines.
 +     */
 +
 +    /* First check if our atomtypes are _identical_ (not reversed) to the previous
 +       entry. If they are not identical we search for earlier duplicates. If they are
 +       we can skip it, since we already searched for the first line
 +       in this group.
 +     */
 +
 +    bFound = FALSE;
 +    bCont  = FALSE;
 +
 +    if (bAllowRepeat && nr > 1)
 +    {
 +        for (j = 0, bCont = TRUE; (j < nral); j++)
 +        {
 +            bCont = bCont && (b->a[j] == bt->param[nr-2].a[j]);
 +        }
 +    }
 +
 +    /* Search for earlier duplicates if this entry was not a continuation
 +       from the previous line.
 +     */
 +    if (!bCont)
 +    {
 +        bFound = FALSE;
 +        for (i = 0; (i < nr); i++)
 +        {
 +            bTest = TRUE;
 +            for (j = 0; (j < nral); j++)
 +            {
 +                bTest = (bTest && (b->a[j] == bt->param[i].a[j]));
 +            }
 +            if (!bTest)
 +            {
 +                bTest = TRUE;
 +                for (j = 0; (j < nral); j++)
 +                {
 +                    bTest = (bTest && (b->a[nral-1-j] == bt->param[i].a[j]));
 +                }
 +            }
 +            if (bTest)
 +            {
 +                if (!bFound)
 +                {
 +                    bId = TRUE;
 +                    for (j = 0; (j < nrfp); j++)
 +                    {
 +                        bId = bId && (bt->param[i].c[j] == b->c[j]);
 +                    }
 +                    if (!bId)
 +                    {
 +                        sprintf(errbuf, "Overriding %s parameters.%s",
 +                                interaction_function[ftype].longname,
 +                                (ftype == F_PDIHS) ?
 +                                "\nUse dihedraltype 9 to allow several multiplicity terms. Only consecutive lines are combined. Non-consective lines overwrite each other."
 +                                : "");
 +                        warning(wi, errbuf);
 +                        fprintf(stderr, "  old:                                         ");
 +                        for (j = 0; (j < nrfp); j++)
 +                        {
 +                            fprintf(stderr, " %g", bt->param[i].c[j]);
 +                        }
 +                        fprintf(stderr, " \n  new: %s\n\n", line);
 +                    }
 +                }
 +                /* Overwrite it! */
 +                for (j = 0; (j < nrfp); j++)
 +                {
 +                    bt->param[i].c[j] = b->c[j];
 +                }
 +                bFound = TRUE;
 +            }
 +        }
 +    }
 +    if (!bFound)
 +    {
 +        /* alloc */
 +        pr_alloc (2, bt);
 +
 +        /* fill the arrays up and down */
 +        memcpy(bt->param[bt->nr].c,  b->c, sizeof(b->c));
 +        memcpy(bt->param[bt->nr].a,  b->a, sizeof(b->a));
 +        memcpy(bt->param[bt->nr+1].c, b->c, sizeof(b->c));
 +
 +        /* The definitions of linear angles depend on the order of atoms,
 +         * that means that for atoms i-j-k, with certain parameter a, the
 +         * corresponding k-j-i angle will have parameter 1-a.
 +         */
 +        if (ftype == F_LINEAR_ANGLES)
 +        {
 +            bt->param[bt->nr+1].c[0] = 1-bt->param[bt->nr+1].c[0];
 +            bt->param[bt->nr+1].c[2] = 1-bt->param[bt->nr+1].c[2];
 +        }
 +
 +        for (j = 0; (j < nral); j++)
 +        {
 +            bt->param[bt->nr+1].a[j] = b->a[nral-1-j];
 +        }
 +
 +        bt->nr += 2;
 +    }
 +}
 +
 +void push_bt(directive d, t_params bt[], int nral,
 +             gpp_atomtype_t at,
 +             t_bond_atomtype bat, char *line,
 +             warninp_t wi)
 +{
 +    const char *formal[MAXATOMLIST+1] = {
 +        "%s",
 +        "%s%s",
 +        "%s%s%s",
 +        "%s%s%s%s",
 +        "%s%s%s%s%s",
 +        "%s%s%s%s%s%s",
 +        "%s%s%s%s%s%s%s"
 +    };
 +    const char *formnl[MAXATOMLIST+1] = {
 +        "%*s",
 +        "%*s%*s",
 +        "%*s%*s%*s",
 +        "%*s%*s%*s%*s",
 +        "%*s%*s%*s%*s%*s",
 +        "%*s%*s%*s%*s%*s%*s",
 +        "%*s%*s%*s%*s%*s%*s%*s"
 +    };
 +    const char *formlf = "%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf";
 +    int         i, ft, ftype, nn, nrfp, nrfpA;
 +    char        f1[STRLEN];
 +    char        alc[MAXATOMLIST+1][20];
 +    /* One force parameter more, so we can check if we read too many */
 +    double      c[MAXFORCEPARAM+1];
 +    t_param     p;
 +    char        errbuf[256];
 +
 +    if ((bat && at) || (!bat && !at))
 +    {
 +        gmx_incons("You should pass either bat or at to push_bt");
 +    }
 +
 +    /* Make format string (nral ints+functype) */
 +    if ((nn = sscanf(line, formal[nral],
 +                     alc[0], alc[1], alc[2], alc[3], alc[4], alc[5])) != nral+1)
 +    {
 +        sprintf(errbuf, "Not enough atomtypes (%d instead of %d)", nn-1, nral);
 +        warning_error(wi, errbuf);
 +        return;
 +    }
 +
 +    ft    = strtol(alc[nral], NULL, 10);
 +    ftype = ifunc_index(d, ft);
 +    nrfp  = NRFP(ftype);
 +    nrfpA = interaction_function[ftype].nrfpA;
 +    strcpy(f1, formnl[nral]);
 +    strcat(f1, formlf);
 +    if ((nn = sscanf(line, f1, &c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6], &c[7], &c[8], &c[9], &c[10], &c[11], &c[12]))
 +        != nrfp)
 +    {
 +        if (nn == nrfpA)
 +        {
 +            /* Copy the B-state from the A-state */
 +            copy_B_from_A(ftype, c);
 +        }
 +        else
 +        {
 +            if (nn < nrfpA)
 +            {
 +                warning_error(wi, "Not enough parameters");
 +            }
 +            else if (nn > nrfpA && nn < nrfp)
 +            {
 +                warning_error(wi, "Too many parameters or not enough parameters for topology B");
 +            }
 +            else if (nn > nrfp)
 +            {
 +                warning_error(wi, "Too many parameters");
 +            }
 +            for (i = nn; (i < nrfp); i++)
 +            {
 +                c[i] = 0.0;
 +            }
 +        }
 +    }
 +    for (i = 0; (i < nral); i++)
 +    {
 +        if (at && ((p.a[i] = get_atomtype_type(alc[i], at)) == NOTSET))
 +        {
 +            gmx_fatal(FARGS, "Unknown atomtype %s\n", alc[i]);
 +        }
 +        else if (bat && ((p.a[i] = get_bond_atomtype_type(alc[i], bat)) == NOTSET))
 +        {
 +            gmx_fatal(FARGS, "Unknown bond_atomtype %s\n", alc[i]);
 +        }
 +    }
 +    for (i = 0; (i < nrfp); i++)
 +    {
 +        p.c[i] = c[i];
 +    }
 +    push_bondtype (&(bt[ftype]), &p, nral, ftype, FALSE, line, wi);
 +}
 +
 +
 +void push_dihedraltype(directive d, t_params bt[],
 +                       t_bond_atomtype bat, char *line,
 +                       warninp_t wi)
 +{
 +    const char  *formal[MAXATOMLIST+1] = {
 +        "%s",
 +        "%s%s",
 +        "%s%s%s",
 +        "%s%s%s%s",
 +        "%s%s%s%s%s",
 +        "%s%s%s%s%s%s",
 +        "%s%s%s%s%s%s%s"
 +    };
 +    const char  *formnl[MAXATOMLIST+1] = {
 +        "%*s",
 +        "%*s%*s",
 +        "%*s%*s%*s",
 +        "%*s%*s%*s%*s",
 +        "%*s%*s%*s%*s%*s",
 +        "%*s%*s%*s%*s%*s%*s",
 +        "%*s%*s%*s%*s%*s%*s%*s"
 +    };
 +    const char  *formlf[MAXFORCEPARAM] = {
 +        "%lf",
 +        "%lf%lf",
 +        "%lf%lf%lf",
 +        "%lf%lf%lf%lf",
 +        "%lf%lf%lf%lf%lf",
 +        "%lf%lf%lf%lf%lf%lf",
 +        "%lf%lf%lf%lf%lf%lf%lf",
 +        "%lf%lf%lf%lf%lf%lf%lf%lf",
 +        "%lf%lf%lf%lf%lf%lf%lf%lf%lf",
 +        "%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf",
 +        "%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf",
 +        "%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf",
 +    };
 +    int          i, ft, ftype, nn, nrfp, nrfpA, nral;
 +    char         f1[STRLEN];
 +    char         alc[MAXATOMLIST+1][20];
 +    double       c[MAXFORCEPARAM];
 +    t_param      p;
 +    gmx_bool     bAllowRepeat;
 +    char         errbuf[256];
 +
 +    /* This routine accepts dihedraltypes defined from either 2 or 4 atoms.
 +     *
 +     * We first check for 2 atoms with the 3th column being an integer
 +     * defining the type. If this isn't the case, we try it with 4 atoms
 +     * and the 5th column defining the dihedral type.
 +     */
 +    nn = sscanf(line, formal[4], alc[0], alc[1], alc[2], alc[3], alc[4]);
 +    if (nn >= 3 && strlen(alc[2]) == 1 && isdigit(alc[2][0]))
 +    {
 +        nral  = 2;
 +        ft    = strtol(alc[nral], NULL, 10);
 +        /* Move atom types around a bit and use 'X' for wildcard atoms
 +         * to create a 4-atom dihedral definition with arbitrary atoms in
 +         * position 1 and 4.
 +         */
 +        if (alc[2][0] == '2')
 +        {
 +            /* improper - the two atomtypes are 1,4. Use wildcards for 2,3 */
 +            strcpy(alc[3], alc[1]);
 +            sprintf(alc[2], "X");
 +            sprintf(alc[1], "X");
 +            /* alc[0] stays put */
 +        }
 +        else
 +        {
 +            /* proper - the two atomtypes are 2,3. Use wildcards for 1,4 */
 +            sprintf(alc[3], "X");
 +            strcpy(alc[2], alc[1]);
 +            strcpy(alc[1], alc[0]);
 +            sprintf(alc[0], "X");
 +        }
 +    }
 +    else if (nn == 5 && strlen(alc[4]) == 1 && isdigit(alc[4][0]))
 +    {
 +        nral  = 4;
 +        ft    = strtol(alc[nral], NULL, 10);
 +    }
 +    else
 +    {
 +        sprintf(errbuf, "Incorrect number of atomtypes for dihedral (%d instead of 2 or 4)", nn-1);
 +        warning_error(wi, errbuf);
 +        return;
 +    }
 +
 +    if (ft == 9)
 +    {
 +        /* Previously, we have always overwritten parameters if e.g. a torsion
 +           with the same atomtypes occurs on multiple lines. However, CHARMM and
 +           some other force fields specify multiple dihedrals over some bonds,
 +           including cosines with multiplicity 6 and somethimes even higher.
 +           Thus, they cannot be represented with Ryckaert-Bellemans terms.
 +           To add support for these force fields, Dihedral type 9 is identical to
 +           normal proper dihedrals, but repeated entries are allowed.
 +         */
 +        bAllowRepeat = TRUE;
 +        ft           = 1;
 +    }
 +    else
 +    {
 +        bAllowRepeat = FALSE;
 +    }
 +
 +
 +    ftype = ifunc_index(d, ft);
 +    nrfp  = NRFP(ftype);
 +    nrfpA = interaction_function[ftype].nrfpA;
 +
 +    strcpy(f1, formnl[nral]);
 +    strcat(f1, formlf[nrfp-1]);
 +
 +    /* Check number of parameters given */
 +    if ((nn = sscanf(line, f1, &c[0], &c[1], &c[2], &c[3], &c[4], &c[5], &c[6], &c[7], &c[8], &c[9], &c[10], &c[11]))
 +        != nrfp)
 +    {
 +        if (nn == nrfpA)
 +        {
 +            /* Copy the B-state from the A-state */
 +            copy_B_from_A(ftype, c);
 +        }
 +        else
 +        {
 +            if (nn < nrfpA)
 +            {
 +                warning_error(wi, "Not enough parameters");
 +            }
 +            else if (nn > nrfpA && nn < nrfp)
 +            {
 +                warning_error(wi, "Too many parameters or not enough parameters for topology B");
 +            }
 +            else if (nn > nrfp)
 +            {
 +                warning_error(wi, "Too many parameters");
 +            }
 +            for (i = nn; (i < nrfp); i++)
 +            {
 +                c[i] = 0.0;
 +            }
 +        }
 +    }
 +
 +    for (i = 0; (i < 4); i++)
 +    {
 +        if (!strcmp(alc[i], "X"))
 +        {
 +            p.a[i] = -1;
 +        }
 +        else
 +        {
 +            if ((p.a[i] = get_bond_atomtype_type(alc[i], bat)) == NOTSET)
 +            {
 +                gmx_fatal(FARGS, "Unknown bond_atomtype %s", alc[i]);
 +            }
 +        }
 +    }
 +    for (i = 0; (i < nrfp); i++)
 +    {
 +        p.c[i] = c[i];
 +    }
 +    /* Always use 4 atoms here, since we created two wildcard atoms
 +     * if there wasn't of them 4 already.
 +     */
 +    push_bondtype (&(bt[ftype]), &p, 4, ftype, bAllowRepeat, line, wi);
 +}
 +
 +
 +void push_nbt(directive d, t_nbparam **nbt, gpp_atomtype_t atype,
 +              char *pline, int nb_funct,
 +              warninp_t wi)
 +{
 +    /* swap the atoms */
 +    const char *form3 = "%*s%*s%*s%lf%lf%lf";
 +    const char *form4 = "%*s%*s%*s%lf%lf%lf%lf";
 +    const char *form5 = "%*s%*s%*s%lf%lf%lf%lf%lf";
 +    char        a0[80], a1[80];
 +    int         i, f, n, ftype, nrfp;
 +    double      c[4], dum;
 +    real        cr[4];
 +    int         ai, aj;
 +    t_nbparam  *nbp;
 +    gmx_bool    bId;
 +    char        errbuf[256];
 +
 +    if (sscanf (pline, "%s%s%d", a0, a1, &f) != 3)
 +    {
 +        too_few(wi);
 +        return;
 +    }
 +
 +    ftype = ifunc_index(d, f);
 +
 +    if (ftype != nb_funct)
 +    {
 +        sprintf(errbuf, "Trying to add %s while the default nonbond type is %s",
 +                interaction_function[ftype].longname,
 +                interaction_function[nb_funct].longname);
 +        warning_error(wi, errbuf);
 +        return;
 +    }
 +
 +    /* Get the force parameters */
 +    nrfp = NRFP(ftype);
 +    if (ftype == F_LJ14)
 +    {
 +        n = sscanf(pline, form4, &c[0], &c[1], &c[2], &c[3]);
 +        if (n < 2)
 +        {
 +            too_few(wi);
 +            return;
 +        }
 +        /* When the B topology parameters are not set,
 +         * copy them from topology A
 +         */
 +        GMX_ASSERT(nrfp <= 4, "LJ-14 cannot have more than 4 parameters");
 +        for (i = n; i < nrfp; i++)
 +        {
 +            c[i] = c[i-2];
 +        }
 +    }
 +    else if (ftype == F_LJC14_Q)
 +    {
 +        n = sscanf(pline, form5, &c[0], &c[1], &c[2], &c[3], &dum);
 +        if (n != 4)
 +        {
 +            incorrect_n_param(wi);
 +            return;
 +        }
 +    }
 +    else if (nrfp == 2)
 +    {
 +        if (sscanf(pline, form3, &c[0], &c[1], &dum) != 2)
 +        {
 +            incorrect_n_param(wi);
 +            return;
 +        }
 +    }
 +    else if (nrfp == 3)
 +    {
 +        if (sscanf(pline, form4, &c[0], &c[1], &c[2], &dum) != 3)
 +        {
 +            incorrect_n_param(wi);
 +            return;
 +        }
 +    }
 +    else
 +    {
 +        gmx_fatal(FARGS, "Number of force parameters for nonbonded interactions is %d"
 +                  " in file %s, line %d", nrfp, __FILE__, __LINE__);
 +    }
 +    for (i = 0; (i < nrfp); i++)
 +    {
 +        cr[i] = c[i];
 +    }
 +
 +    /* Put the parameters in the matrix */
 +    if ((ai = get_atomtype_type (a0, atype)) == NOTSET)
 +    {
 +        gmx_fatal(FARGS, "Atomtype %s not found", a0);
 +    }
 +    if ((aj = get_atomtype_type (a1, atype)) == NOTSET)
 +    {
 +        gmx_fatal(FARGS, "Atomtype %s not found", a1);
 +    }
 +    nbp = &(nbt[std::max(ai, aj)][std::min(ai, aj)]);
 +
 +    if (nbp->bSet)
 +    {
 +        bId = TRUE;
 +        for (i = 0; i < nrfp; i++)
 +        {
 +            bId = bId && (nbp->c[i] == cr[i]);
 +        }
 +        if (!bId)
 +        {
 +            sprintf(errbuf, "Overriding non-bonded parameters,");
 +            warning(wi, errbuf);
 +            fprintf(stderr, "  old:");
 +            for (i = 0; i < nrfp; i++)
 +            {
 +                fprintf(stderr, " %g", nbp->c[i]);
 +            }
 +            fprintf(stderr, " new\n%s\n", pline);
 +        }
 +    }
 +    nbp->bSet = TRUE;
 +    for (i = 0; i < nrfp; i++)
 +    {
 +        nbp->c[i] = cr[i];
 +    }
 +}
 +
 +void
 +push_gb_params (gpp_atomtype_t at, char *line,
 +                warninp_t wi)
 +{
 +    int    atype;
 +    double radius, vol, surftens, gb_radius, S_hct;
 +    char   atypename[STRLEN];
 +    char   errbuf[STRLEN];
 +
 +    if ( (sscanf(line, "%s%lf%lf%lf%lf%lf", atypename, &radius, &vol, &surftens, &gb_radius, &S_hct)) != 6)
 +    {
 +        sprintf(errbuf, "Too few gb parameters for type %s\n", atypename);
 +        warning(wi, errbuf);
 +    }
 +
 +    /* Search for atomtype */
 +    atype = get_atomtype_type(atypename, at);
 +
 +    if (atype == NOTSET)
 +    {
 +        printf("Couldn't find topology match for atomtype %s\n", atypename);
 +        abort();
 +    }
 +
 +    set_atomtype_gbparam(at, atype, radius, vol, surftens, gb_radius, S_hct);
 +}
 +
 +void
 +push_cmaptype(directive d, t_params bt[], int nral, gpp_atomtype_t at,
 +              t_bond_atomtype bat, char *line,
 +              warninp_t wi)
 +{
-     const char  *formal = "%s%s%s%s%s%s%s%s";
++    const char  *formal = "%s%s%s%s%s%s%s%s%n";
 +
 +    int          i, ft, ftype, nn, nrfp, nrfpA, nrfpB;
-     int          start;
++    int          start, nchar_consumed;
 +    int          nxcmap, nycmap, ncmap, read_cmap, sl, nct;
 +    char         s[20], alc[MAXATOMLIST+2][20];
 +    t_param      p;
 +    char         errbuf[256];
 +
 +    /* Keep the compiler happy */
 +    read_cmap = 0;
 +    start     = 0;
 +
-     if ((nn = sscanf(line, formal, alc[0], alc[1], alc[2], alc[3], alc[4], alc[5], alc[6], alc[7])) != nral+3)
++    GMX_ASSERT(nral == 5, "CMAP requires 5 atoms per interaction");
++
++    /* Here we can only check for < 8 */
++    if ((nn = sscanf(line, formal, alc[0], alc[1], alc[2], alc[3], alc[4], alc[5], alc[6], alc[7], &nchar_consumed)) < nral+3)
 +    {
 +        sprintf(errbuf, "Incorrect number of atomtypes for cmap (%d instead of 5)", nn-1);
 +        warning_error(wi, errbuf);
 +        return;
 +    }
- 
-     /* Compute an offset for each line where the cmap parameters start
-      * ie. where the atom types and grid spacing information ends
-      */
-     for (i = 0; i < nn; i++)
-     {
-         start += (int)strlen(alc[i]);
-     }
- 
-     /* There are nn-1 spaces between the atom types and the grid spacing info in the cmap.itp file */
-     /* start is the position on the line where we start to read the actual cmap grid data from the itp file */
-     start = start + nn -1;
++    start += nchar_consumed;
 +
 +    ft     = strtol(alc[nral], NULL, 10);
 +    nxcmap = strtol(alc[nral+1], NULL, 10);
 +    nycmap = strtol(alc[nral+2], NULL, 10);
 +
 +    /* Check for equal grid spacing in x and y dims */
 +    if (nxcmap != nycmap)
 +    {
 +        gmx_fatal(FARGS, "Not the same grid spacing in x and y for cmap grid: x=%d, y=%d", nxcmap, nycmap);
 +    }
 +
 +    ncmap  = nxcmap*nycmap;
 +    ftype  = ifunc_index(d, ft);
 +    nrfpA  = strtol(alc[6], NULL, 10)*strtol(alc[6], NULL, 10);
 +    nrfpB  = strtol(alc[7], NULL, 10)*strtol(alc[7], NULL, 10);
 +    nrfp   = nrfpA+nrfpB;
 +
 +    /* Allocate memory for the CMAP grid */
 +    bt[F_CMAP].ncmap += nrfp;
 +    srenew(bt[F_CMAP].cmap, bt[F_CMAP].ncmap);
 +
 +    /* Read in CMAP parameters */
 +    sl = 0;
 +    for (i = 0; i < ncmap; i++)
 +    {
 +        while (isspace(*(line+start+sl)))
 +        {
 +            sl++;
 +        }
 +        nn  = sscanf(line+start+sl, " %s ", s);
 +        sl += strlen(s);
 +        bt[F_CMAP].cmap[i+(bt[F_CMAP].ncmap)-nrfp] = strtod(s, NULL);
 +
 +        if (nn == 1)
 +        {
 +            read_cmap++;
 +        }
 +        else
 +        {
 +            gmx_fatal(FARGS, "Error in reading cmap parameter for angle %s %s %s %s %s", alc[0], alc[1], alc[2], alc[3], alc[4]);
 +        }
 +
 +    }
 +
 +    /* Check do that we got the number of parameters we expected */
 +    if (read_cmap == nrfpA)
 +    {
 +        for (i = 0; i < ncmap; i++)
 +        {
 +            bt[F_CMAP].cmap[i+ncmap] = bt[F_CMAP].cmap[i];
 +        }
 +    }
 +    else
 +    {
 +        if (read_cmap < nrfpA)
 +        {
 +            warning_error(wi, "Not enough cmap parameters");
 +        }
 +        else if (read_cmap > nrfpA && read_cmap < nrfp)
 +        {
 +            warning_error(wi, "Too many cmap parameters or not enough parameters for topology B");
 +        }
 +        else if (read_cmap > nrfp)
 +        {
 +            warning_error(wi, "Too many cmap parameters");
 +        }
 +    }
 +
 +
 +    /* Set grid spacing and the number of grids (we assume these numbers to be the same for all grids
 +     * so we can safely assign them each time
 +     */
 +    bt[F_CMAP].grid_spacing = nxcmap;            /* Or nycmap, they need to be equal */
 +    bt[F_CMAP].nc           = bt[F_CMAP].nc + 1; /* Since we are incrementing here, we need to subtract later, see (*****) */
 +    nct                     = (nral+1) * bt[F_CMAP].nc;
 +
 +    /* Allocate memory for the cmap_types information */
 +    srenew(bt[F_CMAP].cmap_types, nct);
 +
 +    for (i = 0; (i < nral); i++)
 +    {
 +        if (at && ((p.a[i] = get_bond_atomtype_type(alc[i], bat)) == NOTSET))
 +        {
 +            gmx_fatal(FARGS, "Unknown atomtype %s\n", alc[i]);
 +        }
 +        else if (bat && ((p.a[i] = get_bond_atomtype_type(alc[i], bat)) == NOTSET))
 +        {
 +            gmx_fatal(FARGS, "Unknown bond_atomtype %s\n", alc[i]);
 +        }
 +
 +        /* Assign a grid number to each cmap_type */
 +        bt[F_CMAP].cmap_types[bt[F_CMAP].nct++] = get_bond_atomtype_type(alc[i], bat);
 +    }
 +
 +    /* Assign a type number to this cmap */
 +    bt[F_CMAP].cmap_types[bt[F_CMAP].nct++] = bt[F_CMAP].nc-1; /* Since we inremented earlier, we need to subtrac here, to get the types right (****) */
 +
 +    /* Check for the correct number of atoms (again) */
 +    if (bt[F_CMAP].nct != nct)
 +    {
 +        gmx_fatal(FARGS, "Incorrect number of atom types (%d) in cmap type %d\n", nct, bt[F_CMAP].nc);
 +    }
 +
 +    /* Is this correct?? */
 +    for (i = 0; i < MAXFORCEPARAM; i++)
 +    {
 +        p.c[i] = NOTSET;
 +    }
 +
 +    /* Push the bond to the bondlist */
 +    push_bondtype (&(bt[ftype]), &p, nral, ftype, FALSE, line, wi);
 +}
 +
 +
 +static void push_atom_now(t_symtab *symtab, t_atoms *at, int atomnr,
 +                          int atomicnumber,
 +                          int type, char *ctype, int ptype,
 +                          char *resnumberic,
 +                          char *resname, char *name, real m0, real q0,
 +                          int typeB, char *ctypeB, real mB, real qB)
 +{
 +    int           j, resind = 0, resnr;
 +    unsigned char ric;
 +    int           nr = at->nr;
 +
 +    if (((nr == 0) && (atomnr != 1)) || (nr && (atomnr != at->nr+1)))
 +    {
 +        gmx_fatal(FARGS, "Atoms in the .top are not numbered consecutively from 1 (rather, atomnr = %d, while at->nr = %d)", atomnr, at->nr);
 +    }
 +
 +    j = strlen(resnumberic) - 1;
 +    if (isdigit(resnumberic[j]))
 +    {
 +        ric = ' ';
 +    }
 +    else
 +    {
 +        ric = resnumberic[j];
 +        if (j == 0 || !isdigit(resnumberic[j-1]))
 +        {
 +            gmx_fatal(FARGS, "Invalid residue number '%s' for atom %d",
 +                      resnumberic, atomnr);
 +        }
 +    }
 +    resnr = strtol(resnumberic, NULL, 10);
 +
 +    if (nr > 0)
 +    {
 +        resind = at->atom[nr-1].resind;
 +    }
 +    if (nr == 0 || strcmp(resname, *at->resinfo[resind].name) != 0 ||
 +        resnr != at->resinfo[resind].nr ||
 +        ric   != at->resinfo[resind].ic)
 +    {
 +        if (nr == 0)
 +        {
 +            resind = 0;
 +        }
 +        else
 +        {
 +            resind++;
 +        }
 +        at->nres = resind + 1;
 +        srenew(at->resinfo, at->nres);
 +        at->resinfo[resind].name = put_symtab(symtab, resname);
 +        at->resinfo[resind].nr   = resnr;
 +        at->resinfo[resind].ic   = ric;
 +    }
 +    else
 +    {
 +        resind = at->atom[at->nr-1].resind;
 +    }
 +
 +    /* New atom instance
 +     * get new space for arrays
 +     */
 +    srenew(at->atom, nr+1);
 +    srenew(at->atomname, nr+1);
 +    srenew(at->atomtype, nr+1);
 +    srenew(at->atomtypeB, nr+1);
 +
 +    /* fill the list */
 +    at->atom[nr].type  = type;
 +    at->atom[nr].ptype = ptype;
 +    at->atom[nr].q     = q0;
 +    at->atom[nr].m     = m0;
 +    at->atom[nr].typeB = typeB;
 +    at->atom[nr].qB    = qB;
 +    at->atom[nr].mB    = mB;
 +
 +    at->atom[nr].resind     = resind;
 +    at->atom[nr].atomnumber = atomicnumber;
 +    at->atomname[nr]        = put_symtab(symtab, name);
 +    at->atomtype[nr]        = put_symtab(symtab, ctype);
 +    at->atomtypeB[nr]       = put_symtab(symtab, ctypeB);
 +    at->nr++;
 +}
 +
 +void push_cg(t_block *block, int *lastindex, int index, int a)
 +{
 +    if (debug)
 +    {
 +        fprintf (debug, "Index %d, Atom %d\n", index, a);
 +    }
 +
 +    if (((block->nr) && (*lastindex != index)) || (!block->nr))
 +    {
 +        /* add a new block */
 +        block->nr++;
 +        srenew(block->index, block->nr+1);
 +    }
 +    block->index[block->nr] = a + 1;
 +    *lastindex              = index;
 +}
 +
 +void push_atom(t_symtab *symtab, t_block *cgs,
 +               t_atoms *at, gpp_atomtype_t atype, char *line, int *lastcg,
 +               warninp_t wi)
 +{
 +    int           nr, ptype;
 +    int           cgnumber, atomnr, type, typeB, nscan;
 +    char          id[STRLEN], ctype[STRLEN], ctypeB[STRLEN],
 +                  resnumberic[STRLEN], resname[STRLEN], name[STRLEN], check[STRLEN];
 +    double        m, q, mb, qb;
 +    real          m0, q0, mB, qB;
 +
 +    /* Make a shortcut for writing in this molecule  */
 +    nr = at->nr;
 +
 +    /* Fixed parameters */
 +    if (sscanf(line, "%s%s%s%s%s%d",
 +               id, ctype, resnumberic, resname, name, &cgnumber) != 6)
 +    {
 +        too_few(wi);
 +        return;
 +    }
 +    sscanf(id, "%d", &atomnr);
 +    if ((type  = get_atomtype_type(ctype, atype)) == NOTSET)
 +    {
 +        gmx_fatal(FARGS, "Atomtype %s not found", ctype);
 +    }
 +    ptype = get_atomtype_ptype(type, atype);
 +
 +    /* Set default from type */
 +    q0    = get_atomtype_qA(type, atype);
 +    m0    = get_atomtype_massA(type, atype);
 +    typeB = type;
 +    qB    = q0;
 +    mB    = m0;
 +
 +    /* Optional parameters */
 +    nscan = sscanf(line, "%*s%*s%*s%*s%*s%*s%lf%lf%s%lf%lf%s",
 +                   &q, &m, ctypeB, &qb, &mb, check);
 +
 +    /* Nasty switch that falls thru all the way down! */
 +    if (nscan > 0)
 +    {
 +        q0 = qB = q;
 +        if (nscan > 1)
 +        {
 +            m0 = mB = m;
 +            if (nscan > 2)
 +            {
 +                if ((typeB = get_atomtype_type(ctypeB, atype)) == NOTSET)
 +                {
 +                    gmx_fatal(FARGS, "Atomtype %s not found", ctypeB);
 +                }
 +                qB = get_atomtype_qA(typeB, atype);
 +                mB = get_atomtype_massA(typeB, atype);
 +                if (nscan > 3)
 +                {
 +                    qB = qb;
 +                    if (nscan > 4)
 +                    {
 +                        mB = mb;
 +                        if (nscan > 5)
 +                        {
 +                            warning_error(wi, "Too many parameters");
 +                        }
 +                    }
 +                }
 +            }
 +        }
 +    }
 +    if (debug)
 +    {
 +        fprintf(debug, "mB=%g, qB=%g, typeB=%d\n", mB, qB, typeB);
 +    }
 +
 +    push_cg(cgs, lastcg, cgnumber, nr);
 +
 +    push_atom_now(symtab, at, atomnr, get_atomtype_atomnumber(type, atype),
 +                  type, ctype, ptype, resnumberic,
 +                  resname, name, m0, q0, typeB,
 +                  typeB == type ? ctype : ctypeB, mB, qB);
 +}
 +
 +void push_molt(t_symtab *symtab, int *nmol, t_molinfo **mol, char *line,
 +               warninp_t wi)
 +{
 +    char       type[STRLEN];
 +    int        nrexcl, i;
 +    t_molinfo *newmol;
 +
 +    if ((sscanf(line, "%s%d", type, &nrexcl)) != 2)
 +    {
 +        warning_error(wi, "Expected a molecule type name and nrexcl");
 +    }
 +
 +    /* Test if this moleculetype overwrites another */
 +    i    = 0;
 +    while (i < *nmol)
 +    {
 +        if (strcmp(*((*mol)[i].name), type) == 0)
 +        {
 +            gmx_fatal(FARGS, "moleculetype %s is redefined", type);
 +        }
 +        i++;
 +    }
 +
 +    (*nmol)++;
 +    srenew(*mol, *nmol);
 +    newmol = &((*mol)[*nmol-1]);
 +    init_molinfo(newmol);
 +
 +    /* Fill in the values */
 +    newmol->name     = put_symtab(symtab, type);
 +    newmol->nrexcl   = nrexcl;
 +    newmol->excl_set = FALSE;
 +}
 +
 +static gmx_bool default_nb_params(int ftype, t_params bt[], t_atoms *at,
 +                                  t_param *p, int c_start, gmx_bool bB, gmx_bool bGenPairs)
 +{
 +    int          i, j, ti, tj, ntype;
 +    gmx_bool     bFound;
 +    t_param     *pi    = NULL;
 +    int          nr    = bt[ftype].nr;
 +    int          nral  = NRAL(ftype);
 +    int          nrfp  = interaction_function[ftype].nrfpA;
 +    int          nrfpB = interaction_function[ftype].nrfpB;
 +
 +    if ((!bB && nrfp == 0) || (bB && nrfpB == 0))
 +    {
 +        return TRUE;
 +    }
 +
 +    bFound = FALSE;
 +    if (bGenPairs)
 +    {
 +        /* First test the generated-pair position to save
 +         * time when we have 1000*1000 entries for e.g. OPLS...
 +         */
 +        ntype = static_cast<int>(std::sqrt(static_cast<double>(nr)));
 +        GMX_ASSERT(ntype * ntype == nr, "Number of pairs of generated non-bonded parameters should be a perfect square");
 +        if (bB)
 +        {
 +            ti = at->atom[p->a[0]].typeB;
 +            tj = at->atom[p->a[1]].typeB;
 +        }
 +        else
 +        {
 +            ti = at->atom[p->a[0]].type;
 +            tj = at->atom[p->a[1]].type;
 +        }
 +        pi     = &(bt[ftype].param[ntype*ti+tj]);
 +        bFound = ((ti == pi->a[0]) && (tj == pi->a[1]));
 +    }
 +
 +    /* Search explicitly if we didnt find it */
 +    if (!bFound)
 +    {
 +        for (i = 0; ((i < nr) && !bFound); i++)
 +        {
 +            pi = &(bt[ftype].param[i]);
 +            if (bB)
 +            {
 +                for (j = 0; ((j < nral) &&
 +                             (at->atom[p->a[j]].typeB == pi->a[j])); j++)
 +                {
 +                    ;
 +                }
 +            }
 +            else
 +            {
 +                for (j = 0; ((j < nral) &&
 +                             (at->atom[p->a[j]].type == pi->a[j])); j++)
 +                {
 +                    ;
 +                }
 +            }
 +            bFound = (j == nral);
 +        }
 +    }
 +
 +    if (bFound)
 +    {
 +        if (bB)
 +        {
 +            if (nrfp+nrfpB > MAXFORCEPARAM)
 +            {
 +                gmx_incons("Too many force parameters");
 +            }
 +            for (j = c_start; (j < nrfpB); j++)
 +            {
 +                p->c[nrfp+j] = pi->c[j];
 +            }
 +        }
 +        else
 +        {
 +            for (j = c_start; (j < nrfp); j++)
 +            {
 +                p->c[j] = pi->c[j];
 +            }
 +        }
 +    }
 +    else
 +    {
 +        for (j = c_start; (j < nrfp); j++)
 +        {
 +            p->c[j] = 0.0;
 +        }
 +    }
 +    return bFound;
 +}
 +
 +static gmx_bool default_cmap_params(t_params bondtype[],
 +                                    t_atoms *at, gpp_atomtype_t atype,
 +                                    t_param *p, gmx_bool bB,
 +                                    int *cmap_type, int *nparam_def)
 +{
 +    int      i, nparam_found;
 +    int      ct;
 +    gmx_bool bFound = FALSE;
 +
 +    nparam_found = 0;
 +    ct           = 0;
 +
 +    /* Match the current cmap angle against the list of cmap_types */
 +    for (i = 0; i < bondtype[F_CMAP].nct && !bFound; i += 6)
 +    {
 +        if (bB)
 +        {
 +
 +        }
 +        else
 +        {
 +            if (
 +                (get_atomtype_batype(at->atom[p->a[0]].type, atype) == bondtype[F_CMAP].cmap_types[i])   &&
 +                (get_atomtype_batype(at->atom[p->a[1]].type, atype) == bondtype[F_CMAP].cmap_types[i+1]) &&
 +                (get_atomtype_batype(at->atom[p->a[2]].type, atype) == bondtype[F_CMAP].cmap_types[i+2]) &&
 +                (get_atomtype_batype(at->atom[p->a[3]].type, atype) == bondtype[F_CMAP].cmap_types[i+3]) &&
 +                (get_atomtype_batype(at->atom[p->a[4]].type, atype) == bondtype[F_CMAP].cmap_types[i+4]))
 +            {
 +                /* Found cmap torsion */
 +                bFound       = TRUE;
 +                ct           = bondtype[F_CMAP].cmap_types[i+5];
 +                nparam_found = 1;
 +            }
 +        }
 +    }
 +
 +    /* If we did not find a matching type for this cmap torsion */
 +    if (!bFound)
 +    {
 +        gmx_fatal(FARGS, "Unknown cmap torsion between atoms %d %d %d %d %d\n",
 +                  p->a[0]+1, p->a[1]+1, p->a[2]+1, p->a[3]+1, p->a[4]+1);
 +    }
 +
 +    *nparam_def = nparam_found;
 +    *cmap_type  = ct;
 +
 +    return bFound;
 +}
 +
 +static gmx_bool default_params(int ftype, t_params bt[],
 +                               t_atoms *at, gpp_atomtype_t atype,
 +                               t_param *p, gmx_bool bB,
 +                               t_param **param_def,
 +                               int *nparam_def)
 +{
 +    int          i, j, nparam_found;
 +    gmx_bool     bFound, bSame;
 +    t_param     *pi    = NULL;
 +    t_param     *pj    = NULL;
 +    int          nr    = bt[ftype].nr;
 +    int          nral  = NRAL(ftype);
 +    int          nrfpA = interaction_function[ftype].nrfpA;
 +    int          nrfpB = interaction_function[ftype].nrfpB;
 +
 +    if ((!bB && nrfpA == 0) || (bB && nrfpB == 0))
 +    {
 +        return TRUE;
 +    }
 +
 +
 +    /* We allow wildcards now. The first type (with or without wildcards) that
 +     * fits is used, so you should probably put the wildcarded bondtypes
 +     * at the end of each section.
 +     */
 +    bFound       = FALSE;
 +    nparam_found = 0;
 +    /* OPLS uses 1000s of dihedraltypes, so in order to speed up the scanning we have a
 +     * special case for this. Check for B state outside loop to speed it up.
 +     */
 +    if (ftype == F_PDIHS || ftype == F_RBDIHS || ftype == F_IDIHS || ftype == F_PIDIHS)
 +    {
 +        if (bB)
 +        {
 +            for (i = 0; ((i < nr) && !bFound); i++)
 +            {
 +                pi     = &(bt[ftype].param[i]);
 +                bFound =
 +                    (
 +                        ((pi->ai() == -1) || (get_atomtype_batype(at->atom[p->ai()].typeB, atype) == pi->ai())) &&
 +                        ((pi->aj() == -1) || (get_atomtype_batype(at->atom[p->aj()].typeB, atype) == pi->aj())) &&
 +                        ((pi->ak() == -1) || (get_atomtype_batype(at->atom[p->ak()].typeB, atype) == pi->ak())) &&
 +                        ((pi->al() == -1) || (get_atomtype_batype(at->atom[p->al()].typeB, atype) == pi->al()))
 +                    );
 +            }
 +        }
 +        else
 +        {
 +            /* State A */
 +            for (i = 0; ((i < nr) && !bFound); i++)
 +            {
 +                pi     = &(bt[ftype].param[i]);
 +                bFound =
 +                    (
 +                        ((pi->ai() == -1) || (get_atomtype_batype(at->atom[p->ai()].type, atype) == pi->ai())) &&
 +                        ((pi->aj() == -1) || (get_atomtype_batype(at->atom[p->aj()].type, atype) == pi->aj())) &&
 +                        ((pi->ak() == -1) || (get_atomtype_batype(at->atom[p->ak()].type, atype) == pi->ak())) &&
 +                        ((pi->al() == -1) || (get_atomtype_batype(at->atom[p->al()].type, atype) == pi->al()))
 +                    );
 +            }
 +        }
 +        /* Find additional matches for this dihedral - necessary for ftype==9 which is used e.g. for charmm.
 +         * The rules in that case is that additional matches HAVE to be on adjacent lines!
 +         */
 +        if (bFound == TRUE)
 +        {
 +            nparam_found++;
 +            bSame = TRUE;
 +            /* Continue from current i value */
 +            for (j = i+1; j < nr && bSame; j += 2)
 +            {
 +                pj    = &(bt[ftype].param[j]);
 +                bSame = (pi->ai() == pj->ai() && pi->aj() == pj->aj() && pi->ak() == pj->ak() && pi->al() == pj->al());
 +                if (bSame)
 +                {
 +                    nparam_found++;
 +                }
 +                /* nparam_found will be increased as long as the numbers match */
 +            }
 +        }
 +    }
 +    else   /* Not a dihedral */
 +    {
 +        for (i = 0; ((i < nr) && !bFound); i++)
 +        {
 +            pi = &(bt[ftype].param[i]);
 +            if (bB)
 +            {
 +                for (j = 0; ((j < nral) &&
 +                             (get_atomtype_batype(at->atom[p->a[j]].typeB, atype) == pi->a[j])); j++)
 +                {
 +                    ;
 +                }
 +            }
 +            else
 +            {
 +                for (j = 0; ((j < nral) &&
 +                             (get_atomtype_batype(at->atom[p->a[j]].type, atype) == pi->a[j])); j++)
 +                {
 +                    ;
 +                }
 +            }
 +            bFound = (j == nral);
 +        }
 +        if (bFound)
 +        {
 +            nparam_found = 1;
 +        }
 +    }
 +
 +    *param_def  = pi;
 +    *nparam_def = nparam_found;
 +
 +    return bFound;
 +}
 +
 +
 +
 +void push_bond(directive d, t_params bondtype[], t_params bond[],
 +               t_atoms *at, gpp_atomtype_t atype, char *line,
 +               gmx_bool bBonded, gmx_bool bGenPairs, real fudgeQQ,
 +               gmx_bool bZero, gmx_bool *bWarn_copy_A_B,
 +               warninp_t wi)
 +{
 +    const char  *aaformat[MAXATOMLIST] = {
 +        "%d%d",
 +        "%d%d%d",
 +        "%d%d%d%d",
 +        "%d%d%d%d%d",
 +        "%d%d%d%d%d%d",
 +        "%d%d%d%d%d%d%d"
 +    };
 +    const char  *asformat[MAXATOMLIST] = {
 +        "%*s%*s",
 +        "%*s%*s%*s",
 +        "%*s%*s%*s%*s",
 +        "%*s%*s%*s%*s%*s",
 +        "%*s%*s%*s%*s%*s%*s",
 +        "%*s%*s%*s%*s%*s%*s%*s"
 +    };
 +    const char  *ccformat = "%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf%lf";
 +    int          nr, i, j, nral, nral_fmt, nread, ftype;
 +    char         format[STRLEN];
 +    /* One force parameter more, so we can check if we read too many */
 +    double       cc[MAXFORCEPARAM+1];
 +    int          aa[MAXATOMLIST+1];
 +    t_param      param, *param_defA, *param_defB;
 +    gmx_bool     bFoundA = FALSE, bFoundB = FALSE, bDef, bPert, bSwapParity = FALSE;
 +    int          nparam_defA, nparam_defB;
 +    char         errbuf[256];
 +
 +    nparam_defA = nparam_defB = 0;
 +
 +    ftype = ifunc_index(d, 1);
 +    nral  = NRAL(ftype);
 +    for (j = 0; j < MAXATOMLIST; j++)
 +    {
 +        aa[j] = NOTSET;
 +    }
 +    bDef = (NRFP(ftype) > 0);
 +
 +    if (ftype == F_SETTLE)
 +    {
 +        /* SETTLE acts on 3 atoms, but the topology format only specifies
 +         * the first atom (for historical reasons).
 +         */
 +        nral_fmt = 1;
 +    }
 +    else
 +    {
 +        nral_fmt = nral;
 +    }
 +
 +    nread = sscanf(line, aaformat[nral_fmt-1],
 +                   &aa[0], &aa[1], &aa[2], &aa[3], &aa[4], &aa[5]);
 +
 +    if (ftype == F_SETTLE)
 +    {
 +        aa[3] = aa[1];
 +        aa[1] = aa[0] + 1;
 +        aa[2] = aa[0] + 2;
 +    }
 +
 +    if (nread < nral_fmt)
 +    {
 +        too_few(wi);
 +        return;
 +    }
 +    else if (nread > nral_fmt)
 +    {
 +        /* this is a hack to allow for virtual sites with swapped parity */
 +        bSwapParity = (aa[nral] < 0);
 +        if (bSwapParity)
 +        {
 +            aa[nral] = -aa[nral];
 +        }
 +        ftype = ifunc_index(d, aa[nral]);
 +        if (bSwapParity)
 +        {
 +            switch (ftype)
 +            {
 +                case F_VSITE3FAD:
 +                case F_VSITE3OUT:
 +                    break;
 +                default:
 +                    gmx_fatal(FARGS, "Negative function types only allowed for %s and %s",
 +                              interaction_function[F_VSITE3FAD].longname,
 +                              interaction_function[F_VSITE3OUT].longname);
 +            }
 +        }
 +    }
 +
 +
 +    /* Check for double atoms and atoms out of bounds */
 +    for (i = 0; (i < nral); i++)
 +    {
 +        if (aa[i] < 1 || aa[i] > at->nr)
 +        {
 +            gmx_fatal(FARGS, "[ file %s, line %d ]:\n"
 +                      "Atom index (%d) in %s out of bounds (1-%d).\n"
 +                      "This probably means that you have inserted topology section \"%s\"\n"
 +                      "in a part belonging to a different molecule than you intended to.\n"
 +                      "In that case move the \"%s\" section to the right molecule.",
 +                      get_warning_file(wi), get_warning_line(wi),
 +                      aa[i], dir2str(d), at->nr, dir2str(d), dir2str(d));
 +        }
 +        for (j = i+1; (j < nral); j++)
 +        {
 +            if (aa[i] == aa[j])
 +            {
 +                sprintf(errbuf, "Duplicate atom index (%d) in %s", aa[i], dir2str(d));
 +                warning(wi, errbuf);
 +            }
 +        }
 +    }
 +
 +    /* default force parameters  */
 +    for (j = 0; (j < MAXATOMLIST); j++)
 +    {
 +        param.a[j] = aa[j]-1;
 +    }
 +    for (j = 0; (j < MAXFORCEPARAM); j++)
 +    {
 +        param.c[j] = 0.0;
 +    }
 +
 +    /* Get force params for normal and free energy perturbation
 +     * studies, as determined by types!
 +     */
 +
 +    if (bBonded)
 +    {
 +        bFoundA = default_params(ftype, bondtype, at, atype, &param, FALSE, &param_defA, &nparam_defA);
 +        if (bFoundA)
 +        {
 +            /* Copy the A-state and B-state default parameters. */
 +            GMX_ASSERT(NRFPA(ftype)+NRFPB(ftype) <= MAXFORCEPARAM, "Bonded interactions may have at most 12 parameters");
 +            for (j = 0; (j < NRFPA(ftype)+NRFPB(ftype)); j++)
 +            {
 +                param.c[j] = param_defA->c[j];
 +            }
 +        }
 +        bFoundB = default_params(ftype, bondtype, at, atype, &param, TRUE, &param_defB, &nparam_defB);
 +        if (bFoundB)
 +        {
 +            /* Copy only the B-state default parameters */
 +            for (j = NRFPA(ftype); (j < NRFP(ftype)); j++)
 +            {
 +                param.c[j] = param_defB->c[j];
 +            }
 +        }
 +    }
 +    else if (ftype == F_LJ14)
 +    {
 +        bFoundA = default_nb_params(ftype, bondtype, at, &param, 0, FALSE, bGenPairs);
 +        bFoundB = default_nb_params(ftype, bondtype, at, &param, 0, TRUE, bGenPairs);
 +    }
 +    else if (ftype == F_LJC14_Q)
 +    {
 +        param.c[0] = fudgeQQ;
 +        /* Fill in the A-state charges as default parameters */
 +        param.c[1] = at->atom[param.a[0]].q;
 +        param.c[2] = at->atom[param.a[1]].q;
 +        /* The default LJ parameters are the standard 1-4 parameters */
 +        bFoundA = default_nb_params(F_LJ14, bondtype, at, &param, 3, FALSE, bGenPairs);
 +        bFoundB = TRUE;
 +    }
 +    else if (ftype == F_LJC_PAIRS_NB)
 +    {
 +        /* Defaults are not supported here */
 +        bFoundA = FALSE;
 +        bFoundB = TRUE;
 +    }
 +    else
 +    {
 +        gmx_incons("Unknown function type in push_bond");
 +    }
 +
 +    if (nread > nral_fmt)
 +    {
 +        /* Manually specified parameters - in this case we discard multiple torsion info! */
 +
 +        strcpy(format, asformat[nral_fmt-1]);
 +        strcat(format, ccformat);
 +
 +        nread = sscanf(line, format, &cc[0], &cc[1], &cc[2], &cc[3], &cc[4], &cc[5],
 +                       &cc[6], &cc[7], &cc[8], &cc[9], &cc[10], &cc[11], &cc[12]);
 +
 +        if ((nread == NRFPA(ftype)) && (NRFPB(ftype) != 0))
 +        {
 +            /* We only have to issue a warning if these atoms are perturbed! */
 +            bPert = FALSE;
 +            for (j = 0; (j < nral); j++)
 +            {
 +                bPert = bPert || PERTURBED(at->atom[param.a[j]]);
 +            }
 +
 +            if (bPert && *bWarn_copy_A_B)
 +            {
 +                sprintf(errbuf,
 +                        "Some parameters for bonded interaction involving perturbed atoms are specified explicitly in state A, but not B - copying A to B");
 +                warning(wi, errbuf);
 +                *bWarn_copy_A_B = FALSE;
 +            }
 +
 +            /* If only the A parameters were specified, copy them to the B state */
 +            /* The B-state parameters correspond to the first nrfpB
 +             * A-state parameters.
 +             */
 +            for (j = 0; (j < NRFPB(ftype)); j++)
 +            {
 +                cc[nread++] = cc[j];
 +            }
 +        }
 +
 +        /* If nread was 0 or EOF, no parameters were read => use defaults.
 +         * If nread was nrfpA we copied above so nread=nrfp.
 +         * If nread was nrfp we are cool.
 +         * For F_LJC14_Q we allow supplying fudgeQQ only.
 +         * Anything else is an error!
 +         */
 +        if ((nread != 0) && (nread != EOF) && (nread != NRFP(ftype)) &&
 +            !(ftype == F_LJC14_Q && nread == 1))
 +        {
 +            gmx_fatal(FARGS, "Incorrect number of parameters - found %d, expected %d or %d for %s.",
 +                      nread, NRFPA(ftype), NRFP(ftype),
 +                      interaction_function[ftype].longname);
 +        }
 +
 +        for (j = 0; (j < nread); j++)
 +        {
 +            param.c[j] = cc[j];
 +        }
 +
 +        /* Check whether we have to use the defaults */
 +        if (nread == NRFP(ftype))
 +        {
 +            bDef = FALSE;
 +        }
 +    }
 +    else
 +    {
 +        nread = 0;
 +    }
 +    /* nread now holds the number of force parameters read! */
 +
 +    if (bDef)
 +    {
 +        /* Use defaults */
 +        /* When we have multiple terms it would be very dangerous to allow perturbations to a different atom type! */
 +        if (ftype == F_PDIHS)
 +        {
 +            if ((nparam_defA != nparam_defB) || ((nparam_defA > 1 || nparam_defB > 1) && (param_defA != param_defB)))
 +            {
 +                sprintf(errbuf,
 +                        "Cannot automatically perturb a torsion with multiple terms to different form.\n"
 +                        "Please specify perturbed parameters manually for this torsion in your topology!");
 +                warning_error(wi, errbuf);
 +            }
 +        }
 +
 +        if (nread > 0 && nread < NRFPA(ftype))
 +        {
 +            /* Issue an error, do not use defaults */
 +            sprintf(errbuf, "Not enough parameters, there should be at least %d (or 0 for defaults)", NRFPA(ftype));
 +            warning_error(wi, errbuf);
 +        }
 +
 +        if (nread == 0 || nread == EOF)
 +        {
 +            if (!bFoundA)
 +            {
 +                if (interaction_function[ftype].flags & IF_VSITE)
 +                {
 +                    /* set them to NOTSET, will be calculated later */
 +                    for (j = 0; (j < MAXFORCEPARAM); j++)
 +                    {
 +                        param.c[j] = NOTSET;
 +                    }
 +
 +                    if (bSwapParity)
 +                    {
 +                        param.c1() = -1; /* flag to swap parity of vsite construction */
 +                    }
 +                }
 +                else
 +                {
 +                    if (bZero)
 +                    {
 +                        fprintf(stderr, "NOTE: No default %s types, using zeroes\n",
 +                                interaction_function[ftype].longname);
 +                    }
 +                    else
 +                    {
 +                        sprintf(errbuf, "No default %s types", interaction_function[ftype].longname);
 +                        warning_error(wi, errbuf);
 +                    }
 +                }
 +            }
 +            else
 +            {
 +                if (bSwapParity)
 +                {
 +                    switch (ftype)
 +                    {
 +                        case F_VSITE3FAD:
 +                            param.c0() = 360 - param.c0();
 +                            break;
 +                        case F_VSITE3OUT:
 +                            param.c2() = -param.c2();
 +                            break;
 +                    }
 +                }
 +            }
 +            if (!bFoundB)
 +            {
 +                /* We only have to issue a warning if these atoms are perturbed! */
 +                bPert = FALSE;
 +                for (j = 0; (j < nral); j++)
 +                {
 +                    bPert = bPert || PERTURBED(at->atom[param.a[j]]);
 +                }
 +
 +                if (bPert)
 +                {
 +                    sprintf(errbuf, "No default %s types for perturbed atoms, "
 +                            "using normal values", interaction_function[ftype].longname);
 +                    warning(wi, errbuf);
 +                }
 +            }
 +        }
 +    }
 +
 +    if ((ftype == F_PDIHS || ftype == F_ANGRES || ftype == F_ANGRESZ)
 +        && param.c[5] != param.c[2])
 +    {
 +        gmx_fatal(FARGS, "[ file %s, line %d ]:\n"
 +                  "             %s multiplicity can not be perturbed %f!=%f",
 +                  get_warning_file(wi), get_warning_line(wi),
 +                  interaction_function[ftype].longname,
 +                  param.c[2], param.c[5]);
 +    }
 +
 +    if (IS_TABULATED(ftype) && param.c[2] != param.c[0])
 +    {
 +        gmx_fatal(FARGS, "[ file %s, line %d ]:\n"
 +                  "             %s table number can not be perturbed %d!=%d",
 +                  get_warning_file(wi), get_warning_line(wi),
 +                  interaction_function[ftype].longname,
 +                  (int)(param.c[0]+0.5), (int)(param.c[2]+0.5));
 +    }
 +
 +    /* Dont add R-B dihedrals where all parameters are zero (no interaction) */
 +    if (ftype == F_RBDIHS)
 +    {
 +        nr = 0;
 +        for (i = 0; i < NRFP(ftype); i++)
 +        {
 +            if (param.c[i] != 0)
 +            {
 +                nr++;
 +            }
 +        }
 +        if (nr == 0)
 +        {
 +            return;
 +        }
 +    }
 +
 +    /* Put the values in the appropriate arrays */
 +    add_param_to_list (&bond[ftype], &param);
 +
 +    /* Push additional torsions from FF for ftype==9 if we have them.
 +     * We have already checked that the A/B states do not differ in this case,
 +     * so we do not have to double-check that again, or the vsite stuff.
 +     * In addition, those torsions cannot be automatically perturbed.
 +     */
 +    if (bDef && ftype == F_PDIHS)
 +    {
 +        for (i = 1; i < nparam_defA; i++)
 +        {
 +            /* Advance pointer! */
 +            param_defA += 2;
 +            for (j = 0; (j < NRFPA(ftype)+NRFPB(ftype)); j++)
 +            {
 +                param.c[j] = param_defA->c[j];
 +            }
 +            /* And push the next term for this torsion */
 +            add_param_to_list (&bond[ftype], &param);
 +        }
 +    }
 +}
 +
 +void push_cmap(directive d, t_params bondtype[], t_params bond[],
 +               t_atoms *at, gpp_atomtype_t atype, char *line,
 +               warninp_t wi)
 +{
 +    const char *aaformat[MAXATOMLIST+1] =
 +    {
 +        "%d",
 +        "%d%d",
 +        "%d%d%d",
 +        "%d%d%d%d",
 +        "%d%d%d%d%d",
 +        "%d%d%d%d%d%d",
 +        "%d%d%d%d%d%d%d"
 +    };
 +
 +    int      i, j, ftype, nral, nread, ncmap_params;
 +    int      cmap_type;
 +    int      aa[MAXATOMLIST+1];
 +    char     errbuf[256];
 +    gmx_bool bFound;
 +    t_param  param;
 +
 +    ftype        = ifunc_index(d, 1);
 +    nral         = NRAL(ftype);
 +    ncmap_params = 0;
 +
 +    nread = sscanf(line, aaformat[nral-1],
 +                   &aa[0], &aa[1], &aa[2], &aa[3], &aa[4], &aa[5]);
 +
 +    if (nread < nral)
 +    {
 +        too_few(wi);
 +        return;
 +    }
 +    else if (nread == nral)
 +    {
 +        ftype = ifunc_index(d, 1);
 +    }
 +
 +    /* Check for double atoms and atoms out of bounds */
 +    for (i = 0; i < nral; i++)
 +    {
 +        if (aa[i] < 1 || aa[i] > at->nr)
 +        {
 +            gmx_fatal(FARGS, "[ file %s, line %d ]:\n"
 +                      "Atom index (%d) in %s out of bounds (1-%d).\n"
 +                      "This probably means that you have inserted topology section \"%s\"\n"
 +                      "in a part belonging to a different molecule than you intended to.\n"
 +                      "In that case move the \"%s\" section to the right molecule.",
 +                      get_warning_file(wi), get_warning_line(wi),
 +                      aa[i], dir2str(d), at->nr, dir2str(d), dir2str(d));
 +        }
 +
 +        for (j = i+1; (j < nral); j++)
 +        {
 +            if (aa[i] == aa[j])
 +            {
 +                sprintf(errbuf, "Duplicate atom index (%d) in %s", aa[i], dir2str(d));
 +                warning(wi, errbuf);
 +            }
 +        }
 +    }
 +
 +    /* default force parameters  */
 +    for (j = 0; (j < MAXATOMLIST); j++)
 +    {
 +        param.a[j] = aa[j]-1;
 +    }
 +    for (j = 0; (j < MAXFORCEPARAM); j++)
 +    {
 +        param.c[j] = 0.0;
 +    }
 +
 +    /* Get the cmap type for this cmap angle */
 +    bFound = default_cmap_params(bondtype, at, atype, &param, FALSE, &cmap_type, &ncmap_params);
 +
 +    /* We want exactly one parameter (the cmap type in state A (currently no state B) back */
 +    if (bFound && ncmap_params == 1)
 +    {
 +        /* Put the values in the appropriate arrays */
 +        param.c[0] = cmap_type;
 +        add_param_to_list(&bond[ftype], &param);
 +    }
 +    else
 +    {
 +        /* This is essentially the same check as in default_cmap_params() done one more time */
 +        gmx_fatal(FARGS, "Unable to assign a cmap type to torsion %d %d %d %d and %d\n",
 +                  param.a[0]+1, param.a[1]+1, param.a[2]+1, param.a[3]+1, param.a[4]+1);
 +    }
 +}
 +
 +
 +
 +void push_vsitesn(directive d, t_params bond[],
 +                  t_atoms *at, char *line,
 +                  warninp_t wi)
 +{
 +    char   *ptr;
 +    int     type, ftype, j, n, ret, nj, a;
 +    int    *atc    = NULL;
 +    double *weight = NULL, weight_tot;
 +    t_param param;
 +
 +    /* default force parameters  */
 +    for (j = 0; (j < MAXATOMLIST); j++)
 +    {
 +        param.a[j] = NOTSET;
 +    }
 +    for (j = 0; (j < MAXFORCEPARAM); j++)
 +    {
 +        param.c[j] = 0.0;
 +    }
 +
 +    ptr  = line;
 +    ret  = sscanf(ptr, "%d%n", &a, &n);
 +    ptr += n;
 +    if (ret == 0)
 +    {
 +        gmx_fatal(FARGS, "[ file %s, line %d ]:\n"
 +                  "             Expected an atom index in section \"%s\"",
 +                  get_warning_file(wi), get_warning_line(wi),
 +                  dir2str(d));
 +    }
 +
 +    param.a[0] = a - 1;
 +
 +    sscanf(ptr, "%d%n", &type, &n);
 +    ptr  += n;
 +    ftype = ifunc_index(d, type);
 +
 +    weight_tot = 0;
 +    nj         = 0;
 +    do
 +    {
 +        ret  = sscanf(ptr, "%d%n", &a, &n);
 +        ptr += n;
 +        if (ret > 0)
 +        {
 +            if (nj % 20 == 0)
 +            {
 +                srenew(atc, nj+20);
 +                srenew(weight, nj+20);
 +            }
 +            atc[nj] = a - 1;
 +            switch (type)
 +            {
 +                case 1:
 +                    weight[nj] = 1;
 +                    break;
 +                case 2:
 +                    /* Here we use the A-state mass as a parameter.
 +                     * Note that the B-state mass has no influence.
 +                     */
 +                    weight[nj] = at->atom[atc[nj]].m;
 +                    break;
 +                case 3:
 +                    weight[nj] = -1;
 +                    ret        = sscanf(ptr, "%lf%n", &(weight[nj]), &n);
 +                    ptr       += n;
 +                    if (weight[nj] < 0)
 +                    {
 +                        gmx_fatal(FARGS, "[ file %s, line %d ]:\n"
 +                                  "             No weight or negative weight found for vsiten constructing atom %d (atom index %d)",
 +                                  get_warning_file(wi), get_warning_line(wi),
 +                                  nj+1, atc[nj]+1);
 +                    }
 +                    break;
 +                default:
 +                    gmx_fatal(FARGS, "Unknown vsiten type %d", type);
 +            }
 +            weight_tot += weight[nj];
 +            nj++;
 +        }
 +    }
 +    while (ret > 0);
 +
 +    if (nj == 0)
 +    {
 +        gmx_fatal(FARGS, "[ file %s, line %d ]:\n"
 +                  "             Expected more than one atom index in section \"%s\"",
 +                  get_warning_file(wi), get_warning_line(wi),
 +                  dir2str(d));
 +    }
 +
 +    if (weight_tot == 0)
 +    {
 +        gmx_fatal(FARGS, "[ file %s, line %d ]:\n"
 +                  "             The total mass of the construting atoms is zero",
 +                  get_warning_file(wi), get_warning_line(wi));
 +    }
 +
 +    for (j = 0; j < nj; j++)
 +    {
 +        param.a[1] = atc[j];
 +        param.c[0] = nj;
 +        param.c[1] = weight[j]/weight_tot;
 +        /* Put the values in the appropriate arrays */
 +        add_param_to_list (&bond[ftype], &param);
 +    }
 +
 +    sfree(atc);
 +    sfree(weight);
 +}
 +
 +void push_mol(int nrmols, t_molinfo mols[], char *pline, int *whichmol,
 +              int *nrcopies,
 +              warninp_t wi)
 +{
 +    char type[STRLEN];
 +
 +    if (sscanf(pline, "%s%d", type, nrcopies) != 2)
 +    {
 +        too_few(wi);
 +        return;
 +    }
 +
 +    /* Search moleculename.
 +     * Here we originally only did case insensitive matching. But because
 +     * some PDB files can have many chains and use case to generate more
 +     * chain-identifiers, which in turn end up in our moleculetype name,
 +     * we added support for case-sensitivity.
 +     */
 +    int nrcs    = 0;
 +    int nrci    = 0;
 +    int matchci = -1;
 +    int matchcs = -1;
 +    for (int i = 0; i < nrmols; i++)
 +    {
 +        if (strcmp(type, *(mols[i].name)) == 0)
 +        {
 +            nrcs++;
 +            matchcs = i;
 +        }
 +        if (gmx_strcasecmp(type, *(mols[i].name)) == 0)
 +        {
 +            nrci++;
 +            matchci = i;
 +        }
 +    }
 +
 +    if (nrcs == 1)
 +    {
 +        // select the case sensitive match
 +        *whichmol = matchcs;
 +    }
 +    else
 +    {
 +        // avoid matching case-insensitive when we have multiple matches
 +        if (nrci > 1)
 +        {
 +            gmx_fatal(FARGS, "For moleculetype '%s' in [ system ] %d case insensitive matches, but %d case sensitive matches were found. Check the case of the characters in the moleculetypes.", type, nrci, nrcs);
 +        }
 +        if (nrci == 1)
 +        {
 +            // select the unique case insensitive match
 +            *whichmol = matchci;
 +        }
 +        else
 +        {
 +            gmx_fatal(FARGS, "No such moleculetype %s", type);
 +        }
 +    }
 +}
 +
 +void init_block2(t_block2 *b2, int natom)
 +{
 +    int i;
 +
 +    b2->nr = natom;
 +    snew(b2->nra, b2->nr);
 +    snew(b2->a, b2->nr);
 +    for (i = 0; (i < b2->nr); i++)
 +    {
 +        b2->a[i] = NULL;
 +    }
 +}
 +
 +void done_block2(t_block2 *b2)
 +{
 +    int i;
 +
 +    if (b2->nr)
 +    {
 +        for (i = 0; (i < b2->nr); i++)
 +        {
 +            sfree(b2->a[i]);
 +        }
 +        sfree(b2->a);
 +        sfree(b2->nra);
 +        b2->nr = 0;
 +    }
 +}
 +
 +void push_excl(char *line, t_block2 *b2)
 +{
 +    int  i, j;
 +    int  n;
 +    char base[STRLEN], format[STRLEN];
 +
 +    if (sscanf(line, "%d", &i) == 0)
 +    {
 +        return;
 +    }
 +
 +    if ((1 <= i) && (i <= b2->nr))
 +    {
 +        i--;
 +    }
 +    else
 +    {
 +        if (debug)
 +        {
 +            fprintf(debug, "Unbound atom %d\n", i-1);
 +        }
 +        return;
 +    }
 +    strcpy(base, "%*d");
 +    do
 +    {
 +        strcpy(format, base);
 +        strcat(format, "%d");
 +        n = sscanf(line, format, &j);
 +        if (n == 1)
 +        {
 +            if ((1 <= j) && (j <= b2->nr))
 +            {
 +                j--;
 +                srenew(b2->a[i], ++(b2->nra[i]));
 +                b2->a[i][b2->nra[i]-1] = j;
 +                /* also add the reverse exclusion! */
 +                srenew(b2->a[j], ++(b2->nra[j]));
 +                b2->a[j][b2->nra[j]-1] = i;
 +                strcat(base, "%*d");
 +            }
 +            else
 +            {
 +                gmx_fatal(FARGS, "Invalid Atomnr j: %d, b2->nr: %d\n", j, b2->nr);
 +            }
 +        }
 +    }
 +    while (n == 1);
 +}
 +
 +void b_to_b2(t_blocka *b, t_block2 *b2)
 +{
 +    int     i;
 +    int     j, a;
 +
 +    for (i = 0; (i < b->nr); i++)
 +    {
 +        for (j = b->index[i]; (j < b->index[i+1]); j++)
 +        {
 +            a = b->a[j];
 +            srenew(b2->a[i], ++b2->nra[i]);
 +            b2->a[i][b2->nra[i]-1] = a;
 +        }
 +    }
 +}
 +
 +void b2_to_b(t_block2 *b2, t_blocka *b)
 +{
 +    int     i, nra;
 +    int     j;
 +
 +    nra = 0;
 +    for (i = 0; (i < b2->nr); i++)
 +    {
 +        b->index[i] = nra;
 +        for (j = 0; (j < b2->nra[i]); j++)
 +        {
 +            b->a[nra+j] = b2->a[i][j];
 +        }
 +        nra += b2->nra[i];
 +    }
 +    /* terminate list */
 +    b->index[i] = nra;
 +}
 +
 +static int icomp(const void *v1, const void *v2)
 +{
 +    return (*((int *) v1))-(*((int *) v2));
 +}
 +
 +void merge_excl(t_blocka *excl, t_block2 *b2)
 +{
 +    int     i, k;
 +    int     j;
 +    int     nra;
 +
 +    if (!b2->nr)
 +    {
 +        return;
 +    }
 +    else if (b2->nr != excl->nr)
 +    {
 +        gmx_fatal(FARGS, "DEATH HORROR: b2->nr = %d, while excl->nr = %d",
 +                  b2->nr, excl->nr);
 +    }
 +    else if (debug)
 +    {
 +        fprintf(debug, "Entering merge_excl\n");
 +    }
 +
 +    /* First copy all entries from excl to b2 */
 +    b_to_b2(excl, b2);
 +
 +    /* Count and sort the exclusions */
 +    nra = 0;
 +    for (i = 0; (i < b2->nr); i++)
 +    {
 +        if (b2->nra[i] > 0)
 +        {
 +            /* remove double entries */
 +            qsort(b2->a[i], (size_t)b2->nra[i], (size_t)sizeof(b2->a[i][0]), icomp);
 +            k = 1;
 +            for (j = 1; (j < b2->nra[i]); j++)
 +            {
 +                if (b2->a[i][j] != b2->a[i][k-1])
 +                {
 +                    b2->a[i][k] = b2->a[i][j];
 +                    k++;
 +                }
 +            }
 +            b2->nra[i] = k;
 +            nra       += b2->nra[i];
 +        }
 +    }
 +    excl->nra = nra;
 +    srenew(excl->a, excl->nra);
 +
 +    b2_to_b(b2, excl);
 +}
 +
 +int add_atomtype_decoupled(t_symtab *symtab, gpp_atomtype_t at,
 +                           t_nbparam ***nbparam, t_nbparam ***pair)
 +{
 +    t_atom  atom;
 +    t_param param;
 +    int     i, nr;
 +
 +    /* Define an atom type with all parameters set to zero (no interactions) */
 +    atom.q = 0.0;
 +    atom.m = 0.0;
 +    /* Type for decoupled atoms could be anything,
 +     * this should be changed automatically later when required.
 +     */
 +    atom.ptype = eptAtom;
 +    for (i = 0; (i < MAXFORCEPARAM); i++)
 +    {
 +        param.c[i] = 0.0;
 +    }
 +
 +    nr = add_atomtype(at, symtab, &atom, "decoupled", &param, -1, 0.0, 0.0, 0.0, 0, 0, 0);
 +
 +    /* Add space in the non-bonded parameters matrix */
 +    realloc_nb_params(at, nbparam, pair);
 +
 +    return nr;
 +}
 +
 +static void convert_pairs_to_pairsQ(t_params *plist,
 +                                    real fudgeQQ, t_atoms *atoms)
 +{
 +    t_param *paramp1, *paramp2, *paramnew;
 +    int      i, j, p1nr, p2nr, p2newnr;
 +
 +    /* Add the pair list to the pairQ list */
 +    p1nr    = plist[F_LJ14].nr;
 +    p2nr    = plist[F_LJC14_Q].nr;
 +    p2newnr = p1nr + p2nr;
 +    snew(paramnew, p2newnr);
 +
 +    paramp1             = plist[F_LJ14].param;
 +    paramp2             = plist[F_LJC14_Q].param;
 +
 +    /* Fill in the new F_LJC14_Q array with the old one. NOTE:
 +       it may be possible to just ADD the converted F_LJ14 array
 +       to the old F_LJC14_Q array, but since we have to create
 +       a new sized memory structure, better just to deep copy it all.
 +     */
 +
 +    for (i = 0; i < p2nr; i++)
 +    {
 +        /* Copy over parameters */
 +        for (j = 0; j < 5; j++) /* entries are 0-4 for F_LJC14_Q */
 +        {
 +            paramnew[i].c[j] = paramp2[i].c[j];
 +        }
 +
 +        /* copy over atoms */
 +        for (j = 0; j < 2; j++)
 +        {
 +            paramnew[i].a[j] = paramp2[i].a[j];
 +        }
 +    }
 +
 +    for (i = p2nr; i < p2newnr; i++)
 +    {
 +        j             = i-p2nr;
 +
 +        /* Copy over parameters */
 +        paramnew[i].c[0] = fudgeQQ;
 +        paramnew[i].c[1] = atoms->atom[paramp1[j].a[0]].q;
 +        paramnew[i].c[2] = atoms->atom[paramp1[j].a[1]].q;
 +        paramnew[i].c[3] = paramp1[j].c[0];
 +        paramnew[i].c[4] = paramp1[j].c[1];
 +
 +        /* copy over atoms */
 +        paramnew[i].a[0] = paramp1[j].a[0];
 +        paramnew[i].a[1] = paramp1[j].a[1];
 +    }
 +
 +    /* free the old pairlists */
 +    sfree(plist[F_LJC14_Q].param);
 +    sfree(plist[F_LJ14].param);
 +
 +    /* now assign the new data to the F_LJC14_Q structure */
 +    plist[F_LJC14_Q].param   = paramnew;
 +    plist[F_LJC14_Q].nr      = p2newnr;
 +
 +    /* Empty the LJ14 pairlist */
 +    plist[F_LJ14].nr    = 0;
 +    plist[F_LJ14].param = NULL;
 +}
 +
 +static void generate_LJCpairsNB(t_molinfo *mol, int nb_funct, t_params *nbp)
 +{
 +    int       n, ntype, i, j, k;
 +    t_atom   *atom;
 +    t_blocka *excl;
 +    gmx_bool  bExcl;
 +    t_param   param;
 +
 +    n    = mol->atoms.nr;
 +    atom = mol->atoms.atom;
 +
 +    ntype = static_cast<int>(std::sqrt(static_cast<double>(nbp->nr)));
 +    GMX_ASSERT(ntype * ntype == nbp->nr, "Number of pairs of generated non-bonded parameters should be a perfect square");
 +
 +    for (i = 0; i < MAXATOMLIST; i++)
 +    {
 +        param.a[i] = NOTSET;
 +    }
 +    for (i = 0; i < MAXFORCEPARAM; i++)
 +    {
 +        param.c[i] = NOTSET;
 +    }
 +
 +    /* Add a pair interaction for all non-excluded atom pairs */
 +    excl = &mol->excls;
 +    for (i = 0; i < n; i++)
 +    {
 +        for (j = i+1; j < n; j++)
 +        {
 +            bExcl = FALSE;
 +            for (k = excl->index[i]; k < excl->index[i+1]; k++)
 +            {
 +                if (excl->a[k] == j)
 +                {
 +                    bExcl = TRUE;
 +                }
 +            }
 +            if (!bExcl)
 +            {
 +                if (nb_funct != F_LJ)
 +                {
 +                    gmx_fatal(FARGS, "Can only generate non-bonded pair interactions for Van der Waals type Lennard-Jones");
 +                }
 +                param.a[0] = i;
 +                param.a[1] = j;
 +                param.c[0] = atom[i].q;
 +                param.c[1] = atom[j].q;
 +                param.c[2] = nbp->param[ntype*atom[i].type+atom[j].type].c[0];
 +                param.c[3] = nbp->param[ntype*atom[i].type+atom[j].type].c[1];
 +                add_param_to_list(&mol->plist[F_LJC_PAIRS_NB], &param);
 +            }
 +        }
 +    }
 +}
 +
 +static void set_excl_all(t_blocka *excl)
 +{
 +    int nat, i, j, k;
 +
 +    /* Get rid of the current exclusions and exclude all atom pairs */
 +    nat       = excl->nr;
 +    excl->nra = nat*nat;
 +    srenew(excl->a, excl->nra);
 +    k = 0;
 +    for (i = 0; i < nat; i++)
 +    {
 +        excl->index[i] = k;
 +        for (j = 0; j < nat; j++)
 +        {
 +            excl->a[k++] = j;
 +        }
 +    }
 +    excl->index[nat] = k;
 +}
 +
 +static void decouple_atoms(t_atoms *atoms, int atomtype_decouple,
 +                           int couple_lam0, int couple_lam1,
 +                           const char *mol_name)
 +{
 +    int i;
 +
 +    for (i = 0; i < atoms->nr; i++)
 +    {
 +        t_atom *atom;
 +
 +        atom = &atoms->atom[i];
 +
 +        if (atom->qB != atom->q || atom->typeB != atom->type)
 +        {
 +            gmx_fatal(FARGS, "Atom %d in molecule type '%s' has different A and B state charges and/or atom types set in the topology file as well as through the mdp option '%s'. You can not use both these methods simultaneously.",
 +                      i + 1, mol_name, "couple-moltype");
 +        }
 +
 +        if (couple_lam0 == ecouplamNONE || couple_lam0 == ecouplamVDW)
 +        {
 +            atom->q     = 0.0;
 +        }
 +        if (couple_lam0 == ecouplamNONE || couple_lam0 == ecouplamQ)
 +        {
 +            atom->type  = atomtype_decouple;
 +        }
 +        if (couple_lam1 == ecouplamNONE || couple_lam1 == ecouplamVDW)
 +        {
 +            atom->qB    = 0.0;
 +        }
 +        if (couple_lam1 == ecouplamNONE || couple_lam1 == ecouplamQ)
 +        {
 +            atom->typeB = atomtype_decouple;
 +        }
 +    }
 +}
 +
 +void convert_moltype_couple(t_molinfo *mol, int atomtype_decouple, real fudgeQQ,
 +                            int couple_lam0, int couple_lam1,
 +                            gmx_bool bCoupleIntra, int nb_funct, t_params *nbp)
 +{
 +    convert_pairs_to_pairsQ(mol->plist, fudgeQQ, &mol->atoms);
 +    if (!bCoupleIntra)
 +    {
 +        generate_LJCpairsNB(mol, nb_funct, nbp);
 +        set_excl_all(&mol->excls);
 +    }
 +    decouple_atoms(&mol->atoms, atomtype_decouple, couple_lam0, couple_lam1,
 +                   *mol->name);
 +}
diff --cc src/gromacs/mdlib/nbnxn_search.cpp
index 7cf2114f38,0000000000..14bdadede1
mode 100644,000000..100644
--- a/src/gromacs/mdlib/nbnxn_search.cpp
+++ b/src/gromacs/mdlib/nbnxn_search.cpp
@@@ -1,4173 -1,0 +1,4181 @@@
 +/*
 + * This file is part of the GROMACS molecular simulation package.
 + *
 + * Copyright (c) 2012,2013,2014,2015,2016, by the GROMACS development team, led by
 + * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
 + * and including many others, as listed in the AUTHORS file in the
 + * top-level source directory and at http://www.gromacs.org.
 + *
 + * GROMACS is free software; you can redistribute it and/or
 + * modify it under the terms of the GNU Lesser General Public License
 + * as published by the Free Software Foundation; either version 2.1
 + * of the License, or (at your option) any later version.
 + *
 + * GROMACS is distributed in the hope that it will be useful,
 + * but WITHOUT ANY WARRANTY; without even the implied warranty of
 + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 + * Lesser General Public License for more details.
 + *
 + * You should have received a copy of the GNU Lesser General Public
 + * License along with GROMACS; if not, see
 + * http://www.gnu.org/licenses, or write to the Free Software Foundation,
 + * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
 + *
 + * If you want to redistribute modifications to GROMACS, please
 + * consider that scientific software is very special. Version
 + * control is crucial - bugs must be traceable. We will be happy to
 + * consider code for inclusion in the official distribution, but
 + * derived work must not be called official GROMACS. Details are found
 + * in the README & COPYING files - if they are missing, get the
 + * official version at http://www.gromacs.org.
 + *
 + * To help us fund GROMACS development, we humbly ask that you cite
 + * the research papers on the package. Check out http://www.gromacs.org.
 + */
 +
 +#include "gmxpre.h"
 +
 +#include "nbnxn_search.h"
 +
 +#include "config.h"
 +
 +#include <assert.h>
 +#include <string.h>
 +
 +#include <cmath>
 +
 +#include <algorithm>
 +
 +#include "gromacs/domdec/domdec_struct.h"
 +#include "gromacs/gmxlib/nrnb.h"
 +#include "gromacs/math/functions.h"
 +#include "gromacs/math/utilities.h"
 +#include "gromacs/math/vec.h"
 +#include "gromacs/mdlib/gmx_omp_nthreads.h"
 +#include "gromacs/mdlib/nb_verlet.h"
 +#include "gromacs/mdlib/nbnxn_atomdata.h"
 +#include "gromacs/mdlib/nbnxn_consts.h"
 +#include "gromacs/mdlib/nbnxn_grid.h"
 +#include "gromacs/mdlib/nbnxn_internal.h"
 +#include "gromacs/mdlib/nbnxn_simd.h"
 +#include "gromacs/mdlib/nbnxn_util.h"
 +#include "gromacs/mdlib/ns.h"
 +#include "gromacs/mdtypes/group.h"
 +#include "gromacs/mdtypes/md_enums.h"
 +#include "gromacs/pbcutil/ishift.h"
 +#include "gromacs/pbcutil/pbc.h"
 +#include "gromacs/simd/simd.h"
 +#include "gromacs/simd/vector_operations.h"
 +#include "gromacs/utility/exceptions.h"
 +#include "gromacs/utility/fatalerror.h"
 +#include "gromacs/utility/smalloc.h"
 +
 +using namespace gmx; // TODO: Remove when this file is moved into gmx namespace
 +
 +
 +/* We shift the i-particles backward for PBC.
 + * This leads to more conditionals than shifting forward.
 + * We do this to get more balanced pair lists.
 + */
 +static const bool pbc_shift_backward = true;
 +
 +
 +static void nbs_cycle_clear(nbnxn_cycle_t *cc)
 +{
 +    for (int i = 0; i < enbsCCnr; i++)
 +    {
 +        cc[i].count = 0;
 +        cc[i].c     = 0;
 +    }
 +}
 +
 +static double Mcyc_av(const nbnxn_cycle_t *cc)
 +{
 +    return (double)cc->c*1e-6/cc->count;
 +}
 +
 +static void nbs_cycle_print(FILE *fp, const nbnxn_search_t nbs)
 +{
 +    fprintf(fp, "\n");
 +    fprintf(fp, "ns %4d grid %4.1f search %4.1f red.f %5.3f",
 +            nbs->cc[enbsCCgrid].count,
 +            Mcyc_av(&nbs->cc[enbsCCgrid]),
 +            Mcyc_av(&nbs->cc[enbsCCsearch]),
 +            Mcyc_av(&nbs->cc[enbsCCreducef]));
 +
 +    if (nbs->nthread_max > 1)
 +    {
 +        if (nbs->cc[enbsCCcombine].count > 0)
 +        {
 +            fprintf(fp, " comb %5.2f",
 +                    Mcyc_av(&nbs->cc[enbsCCcombine]));
 +        }
 +        fprintf(fp, " s. th");
 +        for (int t = 0; t < nbs->nthread_max; t++)
 +        {
 +            fprintf(fp, " %4.1f",
 +                    Mcyc_av(&nbs->work[t].cc[enbsCCsearch]));
 +        }
 +    }
 +    fprintf(fp, "\n");
 +}
 +
 +static gmx_inline int ci_to_cj(int ci, int na_cj_2log)
 +{
 +    switch (na_cj_2log)
 +    {
 +        case 2: return ci;     break;
 +        case 3: return (ci>>1); break;
 +        case 1: return (ci<<1); break;
 +    }
 +
 +    return 0;
 +}
 +
 +#if GMX_SIMD
 +
 +/* Returns the j-cluster index corresponding to the i-cluster index */
 +template<int cj_size> static gmx_inline int ci_to_cj(int ci)
 +{
 +    if (cj_size == 2)
 +    {
 +        return ci << 1;
 +    }
 +    else if (cj_size == 4)
 +    {
 +        return ci;
 +    }
 +    else if (cj_size == 8)
 +    {
 +        return ci >> 1;
 +    }
 +    else
 +    {
 +        GMX_ASSERT(false, "Only j-cluster sizes 2, 4, 8 are implemented");
 +        return -1;
 +    }
 +}
 +
 +/* Returns the index in the coordinate array corresponding to the i-cluster index */
 +template<int cj_size> static gmx_inline int x_ind_ci(int ci)
 +{
 +    if (cj_size <= 4)
 +    {
 +        /* Coordinates are stored packed in groups of 4 */
 +        return ci*STRIDE_P4;
 +    }
 +    else if (cj_size == 8)
 +    {
 +        /* Coordinates packed in 8, i-cluster size is half the packing width */
 +        return (ci >> 1)*STRIDE_P8 + (ci & 1)*(c_packX8 >> 1);
 +    }
 +    else
 +    {
 +        GMX_ASSERT(false, "Only j-cluster sizes 2, 4, 8 are implemented");
 +        return -1;
 +    }
 +}
 +
 +/* Returns the index in the coordinate array corresponding to the j-cluster index */
 +template<int cj_size> static gmx_inline int x_ind_cj(int cj)
 +{
 +    if (cj_size == 2)
 +    {
 +        /* Coordinates are stored packed in groups of 4 */
 +        return (cj >> 1)*STRIDE_P4 + (cj & 1)*(c_packX4 >> 1);
 +    }
 +    else if (cj_size <= 4)
 +    {
 +        /* Coordinates are stored packed in groups of 4 */
 +        return cj*STRIDE_P4;
 +    }
 +    else if (cj_size == 8)
 +    {
 +        /* Coordinates are stored packed in groups of 8 */
 +        return cj*STRIDE_P8;
 +    }
 +    else
 +    {
 +        GMX_ASSERT(false, "Only j-cluster sizes 2, 4, 8 are implemented");
 +        return -1;
 +    }
 +}
 +
 +/* The 6 functions below are only introduced to make the code more readable */
 +
 +static gmx_inline int ci_to_cj_simd_4xn(int ci)
 +{
 +    return ci_to_cj<GMX_SIMD_REAL_WIDTH>(ci);
 +}
 +
 +static gmx_inline int x_ind_ci_simd_4xn(int ci)
 +{
 +    return x_ind_ci<GMX_SIMD_REAL_WIDTH>(ci);
 +}
 +
 +static gmx_inline int x_ind_cj_simd_4xn(int cj)
 +{
 +    return x_ind_cj<GMX_SIMD_REAL_WIDTH>(cj);
 +}
 +
 +static gmx_inline int ci_to_cj_simd_2xnn(int ci)
 +{
 +    return ci_to_cj<GMX_SIMD_REAL_WIDTH/2>(ci);
 +}
 +
 +static gmx_inline int x_ind_ci_simd_2xnn(int ci)
 +{
 +    return x_ind_ci<GMX_SIMD_REAL_WIDTH/2>(ci);
 +}
 +
 +static gmx_inline int x_ind_cj_simd_2xnn(int cj)
 +{
 +    return x_ind_cj<GMX_SIMD_REAL_WIDTH/2>(cj);
 +}
 +
 +#endif // GMX_SIMD
 +
 +gmx_bool nbnxn_kernel_pairlist_simple(int nb_kernel_type)
 +{
 +    if (nb_kernel_type == nbnxnkNotSet)
 +    {
 +        gmx_fatal(FARGS, "Non-bonded kernel type not set for Verlet-style pair-list.");
 +    }
 +
 +    switch (nb_kernel_type)
 +    {
 +        case nbnxnk8x8x8_GPU:
 +        case nbnxnk8x8x8_PlainC:
 +            return FALSE;
 +
 +        case nbnxnk4x4_PlainC:
 +        case nbnxnk4xN_SIMD_4xN:
 +        case nbnxnk4xN_SIMD_2xNN:
 +            return TRUE;
 +
 +        default:
 +            gmx_incons("Invalid nonbonded kernel type passed!");
 +            return FALSE;
 +    }
 +}
 +
 +/* Initializes a single nbnxn_pairlist_t data structure */
 +static void nbnxn_init_pairlist_fep(t_nblist *nl)
 +{
 +    nl->type        = GMX_NBLIST_INTERACTION_FREE_ENERGY;
 +    nl->igeometry   = GMX_NBLIST_GEOMETRY_PARTICLE_PARTICLE;
 +    /* The interaction functions are set in the free energy kernel fuction */
 +    nl->ivdw        = -1;
 +    nl->ivdwmod     = -1;
 +    nl->ielec       = -1;
 +    nl->ielecmod    = -1;
 +
 +    nl->maxnri      = 0;
 +    nl->maxnrj      = 0;
 +    nl->nri         = 0;
 +    nl->nrj         = 0;
 +    nl->iinr        = NULL;
 +    nl->gid         = NULL;
 +    nl->shift       = NULL;
 +    nl->jindex      = NULL;
 +    nl->jjnr        = NULL;
 +    nl->excl_fep    = NULL;
 +
 +}
 +
 +void nbnxn_init_search(nbnxn_search_t           * nbs_ptr,
 +                       ivec                      *n_dd_cells,
 +                       struct gmx_domdec_zones_t *zones,
 +                       gmx_bool                   bFEP,
 +                       int                        nthread_max)
 +{
 +    nbnxn_search_t nbs;
 +    int            ngrid;
 +
 +    snew(nbs, 1);
 +    *nbs_ptr = nbs;
 +
 +    nbs->bFEP   = bFEP;
 +
 +    nbs->DomDec = (n_dd_cells != NULL);
 +
 +    clear_ivec(nbs->dd_dim);
 +    ngrid = 1;
 +    if (nbs->DomDec)
 +    {
 +        nbs->zones = zones;
 +
 +        for (int d = 0; d < DIM; d++)
 +        {
 +            if ((*n_dd_cells)[d] > 1)
 +            {
 +                nbs->dd_dim[d] = 1;
 +                /* Each grid matches a DD zone */
 +                ngrid *= 2;
 +            }
 +        }
 +    }
 +
 +    nbnxn_grids_init(nbs, ngrid);
 +
 +    nbs->cell        = NULL;
 +    nbs->cell_nalloc = 0;
 +    nbs->a           = NULL;
 +    nbs->a_nalloc    = 0;
 +
 +    nbs->nthread_max = nthread_max;
 +
 +    /* Initialize the work data structures for each thread */
 +    snew(nbs->work, nbs->nthread_max);
 +    for (int t = 0; t < nbs->nthread_max; t++)
 +    {
 +        nbs->work[t].cxy_na           = NULL;
 +        nbs->work[t].cxy_na_nalloc    = 0;
 +        nbs->work[t].sort_work        = NULL;
 +        nbs->work[t].sort_work_nalloc = 0;
 +
 +        snew(nbs->work[t].nbl_fep, 1);
 +        nbnxn_init_pairlist_fep(nbs->work[t].nbl_fep);
 +    }
 +
 +    /* Initialize detailed nbsearch cycle counting */
 +    nbs->print_cycles = (getenv("GMX_NBNXN_CYCLE") != 0);
 +    nbs->search_count = 0;
 +    nbs_cycle_clear(nbs->cc);
 +    for (int t = 0; t < nbs->nthread_max; t++)
 +    {
 +        nbs_cycle_clear(nbs->work[t].cc);
 +    }
 +}
 +
 +static void init_buffer_flags(nbnxn_buffer_flags_t *flags,
 +                              int                   natoms)
 +{
 +    flags->nflag = (natoms + NBNXN_BUFFERFLAG_SIZE - 1)/NBNXN_BUFFERFLAG_SIZE;
 +    if (flags->nflag > flags->flag_nalloc)
 +    {
 +        flags->flag_nalloc = over_alloc_large(flags->nflag);
 +        srenew(flags->flag, flags->flag_nalloc);
 +    }
 +    for (int b = 0; b < flags->nflag; b++)
 +    {
 +        bitmask_clear(&(flags->flag[b]));
 +    }
 +}
 +
 +/* Determines the cell range along one dimension that
 + * the bounding box b0 - b1 sees.
 + */
 +static void get_cell_range(real b0, real b1,
 +                           int nc, real c0, real s, real invs,
 +                           real d2, real r2, int *cf, int *cl)
 +{
 +    *cf = std::max(static_cast<int>((b0 - c0)*invs), 0);
 +
 +    while (*cf > 0 && d2 + gmx::square((b0 - c0) - (*cf-1+1)*s) < r2)
 +    {
 +        (*cf)--;
 +    }
 +
 +    *cl = std::min(static_cast<int>((b1 - c0)*invs), nc-1);
 +    while (*cl < nc-1 && d2 + gmx::square((*cl+1)*s - (b1 - c0)) < r2)
 +    {
 +        (*cl)++;
 +    }
 +}
 +
 +/* Reference code calculating the distance^2 between two bounding boxes */
 +static float box_dist2(float bx0, float bx1, float by0,
 +                       float by1, float bz0, float bz1,
 +                       const nbnxn_bb_t *bb)
 +{
 +    float d2;
 +    float dl, dh, dm, dm0;
 +
 +    d2 = 0;
 +
 +    dl  = bx0 - bb->upper[BB_X];
 +    dh  = bb->lower[BB_X] - bx1;
 +    dm  = std::max(dl, dh);
 +    dm0 = std::max(dm, 0.0f);
 +    d2 += dm0*dm0;
 +
 +    dl  = by0 - bb->upper[BB_Y];
 +    dh  = bb->lower[BB_Y] - by1;
 +    dm  = std::max(dl, dh);
 +    dm0 = std::max(dm, 0.0f);
 +    d2 += dm0*dm0;
 +
 +    dl  = bz0 - bb->upper[BB_Z];
 +    dh  = bb->lower[BB_Z] - bz1;
 +    dm  = std::max(dl, dh);
 +    dm0 = std::max(dm, 0.0f);
 +    d2 += dm0*dm0;
 +
 +    return d2;
 +}
 +
 +/* Plain C code calculating the distance^2 between two bounding boxes */
 +static float subc_bb_dist2(int si, const nbnxn_bb_t *bb_i_ci,
 +                           int csj, const nbnxn_bb_t *bb_j_all)
 +{
 +    const nbnxn_bb_t *bb_i, *bb_j;
 +    float             d2;
 +    float             dl, dh, dm, dm0;
 +
 +    bb_i = bb_i_ci  +  si;
 +    bb_j = bb_j_all + csj;
 +
 +    d2 = 0;
 +
 +    dl  = bb_i->lower[BB_X] - bb_j->upper[BB_X];
 +    dh  = bb_j->lower[BB_X] - bb_i->upper[BB_X];
 +    dm  = std::max(dl, dh);
 +    dm0 = std::max(dm, 0.0f);
 +    d2 += dm0*dm0;
 +
 +    dl  = bb_i->lower[BB_Y] - bb_j->upper[BB_Y];
 +    dh  = bb_j->lower[BB_Y] - bb_i->upper[BB_Y];
 +    dm  = std::max(dl, dh);
 +    dm0 = std::max(dm, 0.0f);
 +    d2 += dm0*dm0;
 +
 +    dl  = bb_i->lower[BB_Z] - bb_j->upper[BB_Z];
 +    dh  = bb_j->lower[BB_Z] - bb_i->upper[BB_Z];
 +    dm  = std::max(dl, dh);
 +    dm0 = std::max(dm, 0.0f);
 +    d2 += dm0*dm0;
 +
 +    return d2;
 +}
 +
 +#if NBNXN_SEARCH_BB_SIMD4
 +
 +/* 4-wide SIMD code for bb distance for bb format xyz0 */
 +static float subc_bb_dist2_simd4(int si, const nbnxn_bb_t *bb_i_ci,
 +                                 int csj, const nbnxn_bb_t *bb_j_all)
 +{
 +    // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
 +    using namespace gmx;
 +
 +    Simd4Float bb_i_S0, bb_i_S1;
 +    Simd4Float bb_j_S0, bb_j_S1;
 +    Simd4Float dl_S;
 +    Simd4Float dh_S;
 +    Simd4Float dm_S;
 +    Simd4Float dm0_S;
 +
 +    bb_i_S0 = load4(&bb_i_ci[si].lower[0]);
 +    bb_i_S1 = load4(&bb_i_ci[si].upper[0]);
 +    bb_j_S0 = load4(&bb_j_all[csj].lower[0]);
 +    bb_j_S1 = load4(&bb_j_all[csj].upper[0]);
 +
 +    dl_S    = bb_i_S0 - bb_j_S1;
 +    dh_S    = bb_j_S0 - bb_i_S1;
 +
 +    dm_S    = max(dl_S, dh_S);
 +    dm0_S   = max(dm_S, simd4SetZeroF());
 +
 +    return dotProduct(dm0_S, dm0_S);
 +}
 +
 +/* Calculate bb bounding distances of bb_i[si,...,si+3] and store them in d2 */
 +#define SUBC_BB_DIST2_SIMD4_XXXX_INNER(si, bb_i, d2) \
 +    {                                                \
 +        int               shi;                                  \
 +                                                 \
 +        Simd4Float        dx_0, dy_0, dz_0;                    \
 +        Simd4Float        dx_1, dy_1, dz_1;                    \
 +                                                 \
 +        Simd4Float        mx, my, mz;                          \
 +        Simd4Float        m0x, m0y, m0z;                       \
 +                                                 \
 +        Simd4Float        d2x, d2y, d2z;                       \
 +        Simd4Float        d2s, d2t;                            \
 +                                                 \
 +        shi = si*NNBSBB_D*DIM;                       \
 +                                                 \
 +        xi_l = load4(bb_i+shi+0*STRIDE_PBB);   \
 +        yi_l = load4(bb_i+shi+1*STRIDE_PBB);   \
 +        zi_l = load4(bb_i+shi+2*STRIDE_PBB);   \
 +        xi_h = load4(bb_i+shi+3*STRIDE_PBB);   \
 +        yi_h = load4(bb_i+shi+4*STRIDE_PBB);   \
 +        zi_h = load4(bb_i+shi+5*STRIDE_PBB);   \
 +                                                 \
 +        dx_0 = xi_l - xj_h;                 \
 +        dy_0 = yi_l - yj_h;                 \
 +        dz_0 = zi_l - zj_h;                 \
 +                                                 \
 +        dx_1 = xj_l - xi_h;                 \
 +        dy_1 = yj_l - yi_h;                 \
 +        dz_1 = zj_l - zi_h;                 \
 +                                                 \
 +        mx   = max(dx_0, dx_1);                 \
 +        my   = max(dy_0, dy_1);                 \
 +        mz   = max(dz_0, dz_1);                 \
 +                                                 \
 +        m0x  = max(mx, zero);                   \
 +        m0y  = max(my, zero);                   \
 +        m0z  = max(mz, zero);                   \
 +                                                 \
 +        d2x  = m0x * m0x;                   \
 +        d2y  = m0y * m0y;                   \
 +        d2z  = m0z * m0z;                   \
 +                                                 \
 +        d2s  = d2x + d2y;                   \
 +        d2t  = d2s + d2z;                   \
 +                                                 \
 +        store4(d2+si, d2t);                      \
 +    }
 +
 +/* 4-wide SIMD code for nsi bb distances for bb format xxxxyyyyzzzz */
 +static void subc_bb_dist2_simd4_xxxx(const float *bb_j,
 +                                     int nsi, const float *bb_i,
 +                                     float *d2)
 +{
 +    // TODO: During SIMDv2 transition only some archs use namespace (remove when done)
 +    using namespace gmx;
 +
 +    Simd4Float xj_l, yj_l, zj_l;
 +    Simd4Float xj_h, yj_h, zj_h;
 +    Simd4Float xi_l, yi_l, zi_l;
 +    Simd4Float xi_h, yi_h, zi_h;
 +
 +    Simd4Float zero;
 +
 +    zero = setZero();
 +
 +    xj_l = Simd4Float(bb_j[0*STRIDE_PBB]);
 +    yj_l = Simd4Float(bb_j[1*STRIDE_PBB]);
 +    zj_l = Simd4Float(bb_j[2*STRIDE_PBB]);
 +    xj_h = Simd4Float(bb_j[3*STRIDE_PBB]);
 +    yj_h = Simd4Float(bb_j[4*STRIDE_PBB]);
 +    zj_h = Simd4Float(bb_j[5*STRIDE_PBB]);
 +
 +    /* Here we "loop" over si (0,STRIDE_PBB) from 0 to nsi with step STRIDE_PBB.
 +     * But as we know the number of iterations is 1 or 2, we unroll manually.
 +     */
 +    SUBC_BB_DIST2_SIMD4_XXXX_INNER(0, bb_i, d2);
 +    if (STRIDE_PBB < nsi)
 +    {
 +        SUBC_BB_DIST2_SIMD4_XXXX_INNER(STRIDE_PBB, bb_i, d2);
 +    }
 +}
 +
 +#endif /* NBNXN_SEARCH_BB_SIMD4 */
 +
 +
 +/* Returns if any atom pair from two clusters is within distance sqrt(rl2) */
 +static gmx_inline gmx_bool
 +clusterpair_in_range(const nbnxn_list_work_t *work,
 +                     int si,
 +                     int csj, int stride, const real *x_j,
 +                     real rl2)
 +{
 +#if !NBNXN_SEARCH_BB_SIMD4
 +
 +    /* Plain C version.
 +     * All coordinates are stored as xyzxyz...
 +     */
 +
 +    const real *x_i = work->x_ci;
 +
 +    for (int i = 0; i < c_nbnxnGpuClusterSize; i++)
 +    {
 +        int i0 = (si*c_nbnxnGpuClusterSize + i)*DIM;
 +        for (int j = 0; j < c_nbnxnGpuClusterSize; j++)
 +        {
 +            int  j0 = (csj*c_nbnxnGpuClusterSize + j)*stride;
 +
 +            real d2 = gmx::square(x_i[i0  ] - x_j[j0  ]) + gmx::square(x_i[i0+1] - x_j[j0+1]) + gmx::square(x_i[i0+2] - x_j[j0+2]);
 +
 +            if (d2 < rl2)
 +            {
 +                return TRUE;
 +            }
 +        }
 +    }
 +
 +    return FALSE;
 +
 +#else /* !NBNXN_SEARCH_BB_SIMD4 */
 +
 +    /* 4-wide SIMD version.
 +     * A cluster is hard-coded to 8 atoms.
 +     * The coordinates x_i are stored as xxxxyyyy..., x_j is stored xyzxyz...
 +     * Using 8-wide AVX is not faster on Intel Sandy Bridge.
 +     */
 +    assert(c_nbnxnGpuClusterSize == 8);
 +
 +    Simd4Real   rc2_S      = Simd4Real(rl2);
 +
 +    const real *x_i        = work->x_ci_simd;
 +
 +    int         dim_stride = c_nbnxnGpuClusterSize*DIM;
 +    Simd4Real   ix_S0      = load4(x_i + si*dim_stride + 0*GMX_SIMD4_WIDTH);
 +    Simd4Real   iy_S0      = load4(x_i + si*dim_stride + 1*GMX_SIMD4_WIDTH);
 +    Simd4Real   iz_S0      = load4(x_i + si*dim_stride + 2*GMX_SIMD4_WIDTH);
 +    Simd4Real   ix_S1      = load4(x_i + si*dim_stride + 3*GMX_SIMD4_WIDTH);
 +    Simd4Real   iy_S1      = load4(x_i + si*dim_stride + 4*GMX_SIMD4_WIDTH);
 +    Simd4Real   iz_S1      = load4(x_i + si*dim_stride + 5*GMX_SIMD4_WIDTH);
 +
 +    /* We loop from the outer to the inner particles to maximize
 +     * the chance that we find a pair in range quickly and return.
 +     */
 +    int j0 = csj*c_nbnxnGpuClusterSize;
 +    int j1 = j0 + c_nbnxnGpuClusterSize - 1;
 +    while (j0 < j1)
 +    {
 +        Simd4Real jx0_S, jy0_S, jz0_S;
 +        Simd4Real jx1_S, jy1_S, jz1_S;
 +
 +        Simd4Real dx_S0, dy_S0, dz_S0;
 +        Simd4Real dx_S1, dy_S1, dz_S1;
 +        Simd4Real dx_S2, dy_S2, dz_S2;
 +        Simd4Real dx_S3, dy_S3, dz_S3;
 +
 +        Simd4Real rsq_S0;
 +        Simd4Real rsq_S1;
 +        Simd4Real rsq_S2;
 +        Simd4Real rsq_S3;
 +
 +        Simd4Bool wco_S0;
 +        Simd4Bool wco_S1;
 +        Simd4Bool wco_S2;
 +        Simd4Bool wco_S3;
 +        Simd4Bool wco_any_S01, wco_any_S23, wco_any_S;
 +
 +        jx0_S = Simd4Real(x_j[j0*stride+0]);
 +        jy0_S = Simd4Real(x_j[j0*stride+1]);
 +        jz0_S = Simd4Real(x_j[j0*stride+2]);
 +
 +        jx1_S = Simd4Real(x_j[j1*stride+0]);
 +        jy1_S = Simd4Real(x_j[j1*stride+1]);
 +        jz1_S = Simd4Real(x_j[j1*stride+2]);
 +
 +        /* Calculate distance */
 +        dx_S0            = ix_S0 - jx0_S;
 +        dy_S0            = iy_S0 - jy0_S;
 +        dz_S0            = iz_S0 - jz0_S;
 +        dx_S1            = ix_S1 - jx0_S;
 +        dy_S1            = iy_S1 - jy0_S;
 +        dz_S1            = iz_S1 - jz0_S;
 +        dx_S2            = ix_S0 - jx1_S;
 +        dy_S2            = iy_S0 - jy1_S;
 +        dz_S2            = iz_S0 - jz1_S;
 +        dx_S3            = ix_S1 - jx1_S;
 +        dy_S3            = iy_S1 - jy1_S;
 +        dz_S3            = iz_S1 - jz1_S;
 +
 +        /* rsq = dx*dx+dy*dy+dz*dz */
 +        rsq_S0           = norm2(dx_S0, dy_S0, dz_S0);
 +        rsq_S1           = norm2(dx_S1, dy_S1, dz_S1);
 +        rsq_S2           = norm2(dx_S2, dy_S2, dz_S2);
 +        rsq_S3           = norm2(dx_S3, dy_S3, dz_S3);
 +
 +        wco_S0           = (rsq_S0 < rc2_S);
 +        wco_S1           = (rsq_S1 < rc2_S);
 +        wco_S2           = (rsq_S2 < rc2_S);
 +        wco_S3           = (rsq_S3 < rc2_S);
 +
 +        wco_any_S01      = wco_S0 || wco_S1;
 +        wco_any_S23      = wco_S2 || wco_S3;
 +        wco_any_S        = wco_any_S01 || wco_any_S23;
 +
 +        if (anyTrue(wco_any_S))
 +        {
 +            return TRUE;
 +        }
 +
 +        j0++;
 +        j1--;
 +    }
 +
 +    return FALSE;
 +
 +#endif /* !NBNXN_SEARCH_BB_SIMD4 */
 +}
 +
 +/* Returns the j sub-cell for index cj_ind */
 +static int nbl_cj(const nbnxn_pairlist_t *nbl, int cj_ind)
 +{
 +    return nbl->cj4[cj_ind/c_nbnxnGpuJgroupSize].cj[cj_ind & (c_nbnxnGpuJgroupSize - 1)];
 +}
 +
 +/* Returns the i-interaction mask of the j sub-cell for index cj_ind */
 +static unsigned int nbl_imask0(const nbnxn_pairlist_t *nbl, int cj_ind)
 +{
 +    return nbl->cj4[cj_ind/c_nbnxnGpuJgroupSize].imei[0].imask;
 +}
 +
 +/* Ensures there is enough space for extra extra exclusion masks */
 +static void check_excl_space(nbnxn_pairlist_t *nbl, int extra)
 +{
 +    if (nbl->nexcl+extra > nbl->excl_nalloc)
 +    {
 +        nbl->excl_nalloc = over_alloc_small(nbl->nexcl+extra);
 +        nbnxn_realloc_void((void **)&nbl->excl,
 +                           nbl->nexcl*sizeof(*nbl->excl),
 +                           nbl->excl_nalloc*sizeof(*nbl->excl),
 +                           nbl->alloc, nbl->free);
 +    }
 +}
 +
 +/* Ensures there is enough space for ncell extra j-cells in the list */
 +static void check_cell_list_space_simple(nbnxn_pairlist_t *nbl,
 +                                         int               ncell)
 +{
 +    int cj_max;
 +
 +    cj_max = nbl->ncj + ncell;
 +
 +    if (cj_max > nbl->cj_nalloc)
 +    {
 +        nbl->cj_nalloc = over_alloc_small(cj_max);
 +        nbnxn_realloc_void((void **)&nbl->cj,
 +                           nbl->ncj*sizeof(*nbl->cj),
 +                           nbl->cj_nalloc*sizeof(*nbl->cj),
 +                           nbl->alloc, nbl->free);
 +    }
 +}
 +
 +/* Ensures there is enough space for ncell extra j-clusters in the list */
 +static void check_cell_list_space_supersub(nbnxn_pairlist_t *nbl,
 +                                           int               ncell)
 +{
 +    int ncj4_max, w;
 +
 +    /* We can have maximally nsupercell*c_gpuNumClusterPerCell sj lists */
 +    /* We can store 4 j-subcell - i-supercell pairs in one struct.
 +     * since we round down, we need one extra entry.
 +     */
 +    ncj4_max = ((nbl->work->cj_ind + ncell*c_gpuNumClusterPerCell + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize);
 +
 +    if (ncj4_max > nbl->cj4_nalloc)
 +    {
 +        nbl->cj4_nalloc = over_alloc_small(ncj4_max);
 +        nbnxn_realloc_void((void **)&nbl->cj4,
 +                           nbl->work->cj4_init*sizeof(*nbl->cj4),
 +                           nbl->cj4_nalloc*sizeof(*nbl->cj4),
 +                           nbl->alloc, nbl->free);
 +    }
 +
 +    if (ncj4_max > nbl->work->cj4_init)
 +    {
 +        for (int j4 = nbl->work->cj4_init; j4 < ncj4_max; j4++)
 +        {
 +            /* No i-subcells and no excl's in the list initially */
 +            for (w = 0; w < c_nbnxnGpuClusterpairSplit; w++)
 +            {
 +                nbl->cj4[j4].imei[w].imask    = 0U;
 +                nbl->cj4[j4].imei[w].excl_ind = 0;
 +
 +            }
 +        }
 +        nbl->work->cj4_init = ncj4_max;
 +    }
 +}
 +
 +/* Set all excl masks for one GPU warp no exclusions */
 +static void set_no_excls(nbnxn_excl_t *excl)
 +{
 +    for (int t = 0; t < c_nbnxnGpuExclSize; t++)
 +    {
 +        /* Turn all interaction bits on */
 +        excl->pair[t] = NBNXN_INTERACTION_MASK_ALL;
 +    }
 +}
 +
 +/* Initializes a single nbnxn_pairlist_t data structure */
 +static void nbnxn_init_pairlist(nbnxn_pairlist_t *nbl,
 +                                gmx_bool          bSimple,
 +                                nbnxn_alloc_t    *alloc,
 +                                nbnxn_free_t     *free)
 +{
 +    if (alloc == NULL)
 +    {
 +        nbl->alloc = nbnxn_alloc_aligned;
 +    }
 +    else
 +    {
 +        nbl->alloc = alloc;
 +    }
 +    if (free == NULL)
 +    {
 +        nbl->free = nbnxn_free_aligned;
 +    }
 +    else
 +    {
 +        nbl->free = free;
 +    }
 +
 +    nbl->bSimple     = bSimple;
 +    nbl->na_sc       = 0;
 +    nbl->na_ci       = 0;
 +    nbl->na_cj       = 0;
 +    nbl->nci         = 0;
 +    nbl->ci          = NULL;
 +    nbl->ci_nalloc   = 0;
 +    nbl->nsci        = 0;
 +    nbl->sci         = NULL;
 +    nbl->sci_nalloc  = 0;
 +    nbl->ncj         = 0;
 +    nbl->cj          = NULL;
 +    nbl->cj_nalloc   = 0;
 +    nbl->ncj4        = 0;
 +    /* We need one element extra in sj, so alloc initially with 1 */
 +    nbl->cj4_nalloc  = 0;
 +    nbl->cj4         = NULL;
 +    nbl->nci_tot     = 0;
 +
 +    if (!nbl->bSimple)
 +    {
 +        GMX_ASSERT(c_nbnxnGpuNumClusterPerSupercluster == c_gpuNumClusterPerCell, "The search code assumes that the a super-cluster matches a search grid cell");
 +
 +        GMX_ASSERT(sizeof(nbl->cj4[0].imei[0].imask)*8 >= c_nbnxnGpuJgroupSize*c_gpuNumClusterPerCell, "The i super-cluster cluster interaction mask does not contain a sufficient number of bits");
 +        GMX_ASSERT(sizeof(nbl->excl[0])*8 >= c_nbnxnGpuJgroupSize*c_gpuNumClusterPerCell, "The GPU exclusion mask does not contain a sufficient number of bits");
 +
 +        nbl->excl        = NULL;
 +        nbl->excl_nalloc = 0;
 +        nbl->nexcl       = 0;
 +        check_excl_space(nbl, 1);
 +        nbl->nexcl       = 1;
 +        set_no_excls(&nbl->excl[0]);
 +    }
 +
 +    snew(nbl->work, 1);
 +    if (nbl->bSimple)
 +    {
 +        snew_aligned(nbl->work->bb_ci, 1, NBNXN_SEARCH_BB_MEM_ALIGN);
 +    }
 +    else
 +    {
 +#if NBNXN_BBXXXX
 +        snew_aligned(nbl->work->pbb_ci, c_gpuNumClusterPerCell/STRIDE_PBB*NNBSBB_XXXX, NBNXN_SEARCH_BB_MEM_ALIGN);
 +#else
 +        snew_aligned(nbl->work->bb_ci, c_gpuNumClusterPerCell, NBNXN_SEARCH_BB_MEM_ALIGN);
 +#endif
 +    }
 +    int gpu_clusterpair_nc = c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize*DIM;
 +    snew(nbl->work->x_ci, gpu_clusterpair_nc);
 +#if GMX_SIMD
 +    snew_aligned(nbl->work->x_ci_simd,
 +                 std::max(NBNXN_CPU_CLUSTER_I_SIZE*DIM*GMX_SIMD_REAL_WIDTH,
 +                          gpu_clusterpair_nc),
 +                 GMX_SIMD_REAL_WIDTH);
 +#endif
 +    snew_aligned(nbl->work->d2, c_gpuNumClusterPerCell, NBNXN_SEARCH_BB_MEM_ALIGN);
 +
 +    nbl->work->sort            = NULL;
 +    nbl->work->sort_nalloc     = 0;
 +    nbl->work->sci_sort        = NULL;
 +    nbl->work->sci_sort_nalloc = 0;
 +}
 +
 +void nbnxn_init_pairlist_set(nbnxn_pairlist_set_t *nbl_list,
 +                             gmx_bool bSimple, gmx_bool bCombined,
 +                             nbnxn_alloc_t *alloc,
 +                             nbnxn_free_t  *free)
 +{
 +    nbl_list->bSimple   = bSimple;
 +    nbl_list->bCombined = bCombined;
 +
 +    nbl_list->nnbl = gmx_omp_nthreads_get(emntNonbonded);
 +
 +    if (!nbl_list->bCombined &&
 +        nbl_list->nnbl > NBNXN_BUFFERFLAG_MAX_THREADS)
 +    {
 +        gmx_fatal(FARGS, "%d OpenMP threads were requested. Since the non-bonded force buffer reduction is prohibitively slow with more than %d threads, we do not allow this. Use %d or less OpenMP threads.",
 +                  nbl_list->nnbl, NBNXN_BUFFERFLAG_MAX_THREADS, NBNXN_BUFFERFLAG_MAX_THREADS);
 +    }
 +
 +    snew(nbl_list->nbl, nbl_list->nnbl);
 +    snew(nbl_list->nbl_fep, nbl_list->nnbl);
 +    /* Execute in order to avoid memory interleaving between threads */
 +#pragma omp parallel for num_threads(nbl_list->nnbl) schedule(static)
 +    for (int i = 0; i < nbl_list->nnbl; i++)
 +    {
 +        try
 +        {
 +            /* Allocate the nblist data structure locally on each thread
 +             * to optimize memory access for NUMA architectures.
 +             */
 +            snew(nbl_list->nbl[i], 1);
 +
 +            /* Only list 0 is used on the GPU, use normal allocation for i>0 */
 +            if (i == 0)
 +            {
 +                nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, alloc, free);
 +            }
 +            else
 +            {
 +                nbnxn_init_pairlist(nbl_list->nbl[i], nbl_list->bSimple, NULL, NULL);
 +            }
 +
 +            snew(nbl_list->nbl_fep[i], 1);
 +            nbnxn_init_pairlist_fep(nbl_list->nbl_fep[i]);
 +        }
 +        GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 +    }
 +}
 +
 +/* Print statistics of a pair list, used for debug output */
 +static void print_nblist_statistics_simple(FILE *fp, const nbnxn_pairlist_t *nbl,
 +                                           const nbnxn_search_t nbs, real rl)
 +{
 +    const nbnxn_grid_t *grid;
 +    int                 cs[SHIFTS];
 +    int                 npexcl;
 +
 +    /* This code only produces correct statistics with domain decomposition */
 +    grid = &nbs->grid[0];
 +
 +    fprintf(fp, "nbl nci %d ncj %d\n",
 +            nbl->nci, nbl->ncj);
 +    fprintf(fp, "nbl na_sc %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
 +            nbl->na_sc, rl, nbl->ncj, nbl->ncj/(double)grid->nc,
 +            nbl->ncj/(double)grid->nc*grid->na_sc,
 +            nbl->ncj/(double)grid->nc*grid->na_sc/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nc*grid->na_sc/(grid->size[XX]*grid->size[YY]*grid->size[ZZ])));
 +
 +    fprintf(fp, "nbl average j cell list length %.1f\n",
 +            0.25*nbl->ncj/(double)std::max(nbl->nci, 1));
 +
 +    for (int s = 0; s < SHIFTS; s++)
 +    {
 +        cs[s] = 0;
 +    }
 +    npexcl = 0;
 +    for (int i = 0; i < nbl->nci; i++)
 +    {
 +        cs[nbl->ci[i].shift & NBNXN_CI_SHIFT] +=
 +            nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start;
 +
 +        int j = nbl->ci[i].cj_ind_start;
 +        while (j < nbl->ci[i].cj_ind_end &&
 +               nbl->cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
 +        {
 +            npexcl++;
 +            j++;
 +        }
 +    }
 +    fprintf(fp, "nbl cell pairs, total: %d excl: %d %.1f%%\n",
 +            nbl->ncj, npexcl, 100*npexcl/(double)std::max(nbl->ncj, 1));
 +    for (int s = 0; s < SHIFTS; s++)
 +    {
 +        if (cs[s] > 0)
 +        {
 +            fprintf(fp, "nbl shift %2d ncj %3d\n", s, cs[s]);
 +        }
 +    }
 +}
 +
 +/* Print statistics of a pair lists, used for debug output */
 +static void print_nblist_statistics_supersub(FILE *fp, const nbnxn_pairlist_t *nbl,
 +                                             const nbnxn_search_t nbs, real rl)
 +{
 +    const nbnxn_grid_t *grid;
 +    int                 b;
 +    int                 c[c_gpuNumClusterPerCell + 1];
 +    double              sum_nsp, sum_nsp2;
 +    int                 nsp_max;
 +
 +    /* This code only produces correct statistics with domain decomposition */
 +    grid = &nbs->grid[0];
 +
 +    fprintf(fp, "nbl nsci %d ncj4 %d nsi %d excl4 %d\n",
 +            nbl->nsci, nbl->ncj4, nbl->nci_tot, nbl->nexcl);
 +    fprintf(fp, "nbl na_c %d rl %g ncp %d per cell %.1f atoms %.1f ratio %.2f\n",
 +            nbl->na_ci, rl, nbl->nci_tot, nbl->nci_tot/(double)grid->nsubc_tot,
 +            nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c,
 +            nbl->nci_tot/(double)grid->nsubc_tot*grid->na_c/(0.5*4.0/3.0*M_PI*rl*rl*rl*grid->nsubc_tot*grid->na_c/(grid->size[XX]*grid->size[YY]*grid->size[ZZ])));
 +
 +    sum_nsp  = 0;
 +    sum_nsp2 = 0;
 +    nsp_max  = 0;
 +    for (int si = 0; si <= c_gpuNumClusterPerCell; si++)
 +    {
 +        c[si] = 0;
 +    }
 +    for (int i = 0; i < nbl->nsci; i++)
 +    {
 +        int nsp;
 +
 +        nsp = 0;
 +        for (int j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
 +        {
 +            for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
 +            {
 +                b = 0;
 +                for (int si = 0; si < c_gpuNumClusterPerCell; si++)
 +                {
 +                    if (nbl->cj4[j4].imei[0].imask & (1U << (j*c_gpuNumClusterPerCell + si)))
 +                    {
 +                        b++;
 +                    }
 +                }
 +                nsp += b;
 +                c[b]++;
 +            }
 +        }
 +        sum_nsp  += nsp;
 +        sum_nsp2 += nsp*nsp;
 +        nsp_max   = std::max(nsp_max, nsp);
 +    }
 +    if (nbl->nsci > 0)
 +    {
 +        sum_nsp  /= nbl->nsci;
 +        sum_nsp2 /= nbl->nsci;
 +    }
 +    fprintf(fp, "nbl #cluster-pairs: av %.1f stddev %.1f max %d\n",
 +            sum_nsp, std::sqrt(sum_nsp2 - sum_nsp*sum_nsp), nsp_max);
 +
 +    if (nbl->ncj4 > 0)
 +    {
 +        for (b = 0; b <= c_gpuNumClusterPerCell; b++)
 +        {
 +            fprintf(fp, "nbl j-list #i-subcell %d %7d %4.1f\n",
 +                    b, c[b],
 +                    100.0*c[b]/(double)(nbl->ncj4*c_nbnxnGpuJgroupSize));
 +        }
 +    }
 +}
 +
 +/* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp */
 +static void low_get_nbl_exclusions(nbnxn_pairlist_t *nbl, int cj4,
 +                                   int warp, nbnxn_excl_t **excl)
 +{
 +    if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
 +    {
 +        /* No exclusions set, make a new list entry */
 +        nbl->cj4[cj4].imei[warp].excl_ind = nbl->nexcl;
 +        nbl->nexcl++;
 +        *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
 +        set_no_excls(*excl);
 +    }
 +    else
 +    {
 +        /* We already have some exclusions, new ones can be added to the list */
 +        *excl = &nbl->excl[nbl->cj4[cj4].imei[warp].excl_ind];
 +    }
 +}
 +
 +/* Returns a pointer to the exclusion mask for cj4-unit cj4, warp warp,
 + * generates a new element and allocates extra memory, if necessary.
 + */
 +static void get_nbl_exclusions_1(nbnxn_pairlist_t *nbl, int cj4,
 +                                 int warp, nbnxn_excl_t **excl)
 +{
 +    if (nbl->cj4[cj4].imei[warp].excl_ind == 0)
 +    {
 +        /* We need to make a new list entry, check if we have space */
 +        check_excl_space(nbl, 1);
 +    }
 +    low_get_nbl_exclusions(nbl, cj4, warp, excl);
 +}
 +
 +/* Returns pointers to the exclusion masks for cj4-unit cj4 for both warps,
 + * generates a new element and allocates extra memory, if necessary.
 + */
 +static void get_nbl_exclusions_2(nbnxn_pairlist_t *nbl, int cj4,
 +                                 nbnxn_excl_t **excl_w0,
 +                                 nbnxn_excl_t **excl_w1)
 +{
 +    /* Check for space we might need */
 +    check_excl_space(nbl, 2);
 +
 +    low_get_nbl_exclusions(nbl, cj4, 0, excl_w0);
 +    low_get_nbl_exclusions(nbl, cj4, 1, excl_w1);
 +}
 +
 +/* Sets the self exclusions i=j and pair exclusions i>j */
 +static void set_self_and_newton_excls_supersub(nbnxn_pairlist_t *nbl,
 +                                               int cj4_ind, int sj_offset,
 +                                               int i_cluster_in_cell)
 +{
 +    nbnxn_excl_t *excl[c_nbnxnGpuClusterpairSplit];
 +
 +    /* Here we only set the set self and double pair exclusions */
 +
 +    assert(c_nbnxnGpuClusterpairSplit == 2);
 +
 +    get_nbl_exclusions_2(nbl, cj4_ind, &excl[0], &excl[1]);
 +
 +    /* Only minor < major bits set */
 +    for (int ej = 0; ej < nbl->na_ci; ej++)
 +    {
 +        int w = (ej>>2);
 +        for (int ei = ej; ei < nbl->na_ci; ei++)
 +        {
 +            excl[w]->pair[(ej & (c_nbnxnGpuJgroupSize-1))*nbl->na_ci + ei] &=
 +                ~(1U << (sj_offset*c_gpuNumClusterPerCell + i_cluster_in_cell));
 +        }
 +    }
 +}
 +
 +/* Returns a diagonal or off-diagonal interaction mask for plain C lists */
 +static unsigned int get_imask(gmx_bool rdiag, int ci, int cj)
 +{
 +    return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
 +}
 +
 +/* Returns a diagonal or off-diagonal interaction mask for cj-size=2 */
 +static unsigned int get_imask_simd_j2(gmx_bool rdiag, int ci, int cj)
 +{
 +    return (rdiag && ci*2 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_0 :
 +            (rdiag && ci*2+1 == cj ? NBNXN_INTERACTION_MASK_DIAG_J2_1 :
 +             NBNXN_INTERACTION_MASK_ALL));
 +}
 +
 +/* Returns a diagonal or off-diagonal interaction mask for cj-size=4 */
 +static unsigned int get_imask_simd_j4(gmx_bool rdiag, int ci, int cj)
 +{
 +    return (rdiag && ci == cj ? NBNXN_INTERACTION_MASK_DIAG : NBNXN_INTERACTION_MASK_ALL);
 +}
 +
 +/* Returns a diagonal or off-diagonal interaction mask for cj-size=8 */
 +static unsigned int get_imask_simd_j8(gmx_bool rdiag, int ci, int cj)
 +{
 +    return (rdiag && ci == cj*2 ? NBNXN_INTERACTION_MASK_DIAG_J8_0 :
 +            (rdiag && ci == cj*2+1 ? NBNXN_INTERACTION_MASK_DIAG_J8_1 :
 +             NBNXN_INTERACTION_MASK_ALL));
 +}
 +
 +#if GMX_SIMD
 +#if GMX_SIMD_REAL_WIDTH == 2
 +#define get_imask_simd_4xn  get_imask_simd_j2
 +#endif
 +#if GMX_SIMD_REAL_WIDTH == 4
 +#define get_imask_simd_4xn  get_imask_simd_j4
 +#endif
 +#if GMX_SIMD_REAL_WIDTH == 8
 +#define get_imask_simd_4xn  get_imask_simd_j8
 +#define get_imask_simd_2xnn get_imask_simd_j4
 +#endif
 +#if GMX_SIMD_REAL_WIDTH == 16
 +#define get_imask_simd_2xnn get_imask_simd_j8
 +#endif
 +#endif
 +
 +/* Plain C code for making a pair list of cell ci vs cell cjf-cjl.
 + * Checks bounding box distances and possibly atom pair distances.
 + */
 +static void make_cluster_list_simple(const nbnxn_grid_t *gridj,
 +                                     nbnxn_pairlist_t *nbl,
 +                                     int ci, int cjf, int cjl,
 +                                     gmx_bool remove_sub_diag,
 +                                     const real *x_j,
 +                                     real rl2, float rbb2,
 +                                     int *ndistc)
 +{
 +    const nbnxn_bb_t        *bb_ci;
 +    const real              *x_ci;
 +
 +    gmx_bool                 InRange;
 +    real                     d2;
 +    int                      cjf_gl, cjl_gl;
 +
 +    bb_ci = nbl->work->bb_ci;
 +    x_ci  = nbl->work->x_ci;
 +
 +    InRange = FALSE;
 +    while (!InRange && cjf <= cjl)
 +    {
 +        d2       = subc_bb_dist2(0, bb_ci, cjf, gridj->bb);
 +        *ndistc += 2;
 +
 +        /* Check if the distance is within the distance where
 +         * we use only the bounding box distance rbb,
 +         * or within the cut-off and there is at least one atom pair
 +         * within the cut-off.
 +         */
 +        if (d2 < rbb2)
 +        {
 +            InRange = TRUE;
 +        }
 +        else if (d2 < rl2)
 +        {
 +            cjf_gl = gridj->cell0 + cjf;
 +            for (int i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
 +            {
 +                for (int j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
 +                {
 +                    InRange = InRange ||
 +                        (gmx::square(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
 +                         gmx::square(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
 +                         gmx::square(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjf_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
 +                }
 +            }
 +            *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
 +        }
 +        if (!InRange)
 +        {
 +            cjf++;
 +        }
 +    }
 +    if (!InRange)
 +    {
 +        return;
 +    }
 +
 +    InRange = FALSE;
 +    while (!InRange && cjl > cjf)
 +    {
 +        d2       = subc_bb_dist2(0, bb_ci, cjl, gridj->bb);
 +        *ndistc += 2;
 +
 +        /* Check if the distance is within the distance where
 +         * we use only the bounding box distance rbb,
 +         * or within the cut-off and there is at least one atom pair
 +         * within the cut-off.
 +         */
 +        if (d2 < rbb2)
 +        {
 +            InRange = TRUE;
 +        }
 +        else if (d2 < rl2)
 +        {
 +            cjl_gl = gridj->cell0 + cjl;
 +            for (int i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE && !InRange; i++)
 +            {
 +                for (int j = 0; j < NBNXN_CPU_CLUSTER_I_SIZE; j++)
 +                {
 +                    InRange = InRange ||
 +                        (gmx::square(x_ci[i*STRIDE_XYZ+XX] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+XX]) +
 +                         gmx::square(x_ci[i*STRIDE_XYZ+YY] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+YY]) +
 +                         gmx::square(x_ci[i*STRIDE_XYZ+ZZ] - x_j[(cjl_gl*NBNXN_CPU_CLUSTER_I_SIZE+j)*STRIDE_XYZ+ZZ]) < rl2);
 +                }
 +            }
 +            *ndistc += NBNXN_CPU_CLUSTER_I_SIZE*NBNXN_CPU_CLUSTER_I_SIZE;
 +        }
 +        if (!InRange)
 +        {
 +            cjl--;
 +        }
 +    }
 +
 +    if (cjf <= cjl)
 +    {
 +        for (int cj = cjf; cj <= cjl; cj++)
 +        {
 +            /* Store cj and the interaction mask */
 +            nbl->cj[nbl->ncj].cj   = gridj->cell0 + cj;
 +            nbl->cj[nbl->ncj].excl = get_imask(remove_sub_diag, ci, cj);
 +            nbl->ncj++;
 +        }
 +        /* Increase the closing index in i super-cell list */
 +        nbl->ci[nbl->nci].cj_ind_end = nbl->ncj;
 +    }
 +}
 +
 +#ifdef GMX_NBNXN_SIMD_4XN
 +#include "gromacs/mdlib/nbnxn_search_simd_4xn.h"
 +#endif
 +#ifdef GMX_NBNXN_SIMD_2XNN
 +#include "gromacs/mdlib/nbnxn_search_simd_2xnn.h"
 +#endif
 +
 +/* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
 + * Checks bounding box distances and possibly atom pair distances.
 + */
 +static void make_cluster_list_supersub(const nbnxn_grid_t *gridi,
 +                                       const nbnxn_grid_t *gridj,
 +                                       nbnxn_pairlist_t *nbl,
 +                                       int sci, int scj,
 +                                       gmx_bool sci_equals_scj,
 +                                       int stride, const real *x,
 +                                       real rl2, float rbb2,
 +                                       int *ndistc)
 +{
 +    nbnxn_list_work_t *work   = nbl->work;
 +
 +#if NBNXN_BBXXXX
 +    const float       *pbb_ci = work->pbb_ci;
 +#else
 +    const nbnxn_bb_t  *bb_ci  = work->bb_ci;
 +#endif
 +
 +    assert(c_nbnxnGpuClusterSize == gridi->na_c);
 +    assert(c_nbnxnGpuClusterSize == gridj->na_c);
 +
 +    /* We generate the pairlist mainly based on bounding-box distances
 +     * and do atom pair distance based pruning on the GPU.
 +     * Only if a j-group contains a single cluster-pair, we try to prune
 +     * that pair based on atom distances on the CPU to avoid empty j-groups.
 +     */
 +#define PRUNE_LIST_CPU_ONE 1
 +#define PRUNE_LIST_CPU_ALL 0
 +
 +#if PRUNE_LIST_CPU_ONE
 +    int  ci_last = -1;
 +#endif
 +
 +    float *d2l = work->d2;
 +
 +    for (int subc = 0; subc < gridj->nsubc[scj]; subc++)
 +    {
 +        int          cj4_ind   = nbl->work->cj_ind/c_nbnxnGpuJgroupSize;
 +        int          cj_offset = nbl->work->cj_ind - cj4_ind*c_nbnxnGpuJgroupSize;
 +        nbnxn_cj4_t *cj4       = &nbl->cj4[cj4_ind];
 +
 +        int          cj        = scj*c_gpuNumClusterPerCell + subc;
 +
 +        int          cj_gl     = gridj->cell0*c_gpuNumClusterPerCell + cj;
 +
 +        /* Initialize this j-subcell i-subcell list */
 +        cj4->cj[cj_offset] = cj_gl;
 +
 +        int ci1;
 +        if (sci_equals_scj)
 +        {
 +            ci1 = subc + 1;
 +        }
 +        else
 +        {
 +            ci1 = gridi->nsubc[sci];
 +        }
 +
 +#if NBNXN_BBXXXX
 +        /* Determine all ci1 bb distances in one call with SIMD4 */
 +        subc_bb_dist2_simd4_xxxx(gridj->pbb+(cj>>STRIDE_PBB_2LOG)*NNBSBB_XXXX+(cj & (STRIDE_PBB-1)),
 +                                 ci1, pbb_ci, d2l);
 +        *ndistc += c_nbnxnGpuClusterSize*2;
 +#endif
 +
 +        int          npair = 0;
 +        unsigned int imask = 0;
 +        /* We use a fixed upper-bound instead of ci1 to help optimization */
 +        for (int ci = 0; ci < c_gpuNumClusterPerCell; ci++)
 +        {
 +            if (ci == ci1)
 +            {
 +                break;
 +            }
 +
 +#if !NBNXN_BBXXXX
 +            /* Determine the bb distance between ci and cj */
 +            d2l[ci]  = subc_bb_dist2(ci, bb_ci, cj, gridj->bb);
 +            *ndistc += 2;
 +#endif
 +            float d2 = d2l[ci];
 +
 +#if PRUNE_LIST_CPU_ALL
 +            /* Check if the distance is within the distance where
 +             * we use only the bounding box distance rbb,
 +             * or within the cut-off and there is at least one atom pair
 +             * within the cut-off. This check is very costly.
 +             */
 +            *ndistc += c_nbnxnGpuClusterSize*c_nbnxnGpuClusterSize;
 +            if (d2 < rbb2 ||
 +                (d2 < rl2 &&
 +                 clusterpair_in_range(work, ci, cj_gl, stride, x, rl2)))
 +#else
 +            /* Check if the distance between the two bounding boxes
 +             * in within the pair-list cut-off.
 +             */
 +            if (d2 < rl2)
 +#endif
 +            {
 +                /* Flag this i-subcell to be taken into account */
 +                imask |= (1U << (cj_offset*c_gpuNumClusterPerCell + ci));
 +
 +#if PRUNE_LIST_CPU_ONE
 +                ci_last = ci;
 +#endif
 +
 +                npair++;
 +            }
 +        }
 +
 +#if PRUNE_LIST_CPU_ONE
 +        /* If we only found 1 pair, check if any atoms are actually
 +         * within the cut-off, so we could get rid of it.
 +         */
 +        if (npair == 1 && d2l[ci_last] >= rbb2 &&
 +            !clusterpair_in_range(work, ci_last, cj_gl, stride, x, rl2))
 +        {
 +            imask &= ~(1U << (cj_offset*c_gpuNumClusterPerCell + ci_last));
 +            npair--;
 +        }
 +#endif
 +
 +        if (npair > 0)
 +        {
 +            /* We have a useful sj entry, close it now */
 +
 +            /* Set the exclucions for the ci== sj entry.
 +             * Here we don't bother to check if this entry is actually flagged,
 +             * as it will nearly always be in the list.
 +             */
 +            if (sci_equals_scj)
 +            {
 +                set_self_and_newton_excls_supersub(nbl, cj4_ind, cj_offset, subc);
 +            }
 +
 +            /* Copy the cluster interaction mask to the list */
 +            for (int w = 0; w < c_nbnxnGpuClusterpairSplit; w++)
 +            {
 +                cj4->imei[w].imask |= imask;
 +            }
 +
 +            nbl->work->cj_ind++;
 +
 +            /* Keep the count */
 +            nbl->nci_tot += npair;
 +
 +            /* Increase the closing index in i super-cell list */
 +            nbl->sci[nbl->nsci].cj4_ind_end =
 +                (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize;
 +        }
 +    }
 +}
 +
 +/* Set all atom-pair exclusions from the topology stored in excl
 + * as masks in the pair-list for simple list i-entry nbl_ci
 + */
 +static void set_ci_top_excls(const nbnxn_search_t nbs,
 +                             nbnxn_pairlist_t    *nbl,
 +                             gmx_bool             diagRemoved,
 +                             int                  na_ci_2log,
 +                             int                  na_cj_2log,
 +                             const nbnxn_ci_t    *nbl_ci,
 +                             const t_blocka      *excl)
 +{
 +    const int    *cell;
 +    int           ci;
 +    int           cj_ind_first, cj_ind_last;
 +    int           cj_first, cj_last;
 +    int           ndirect;
 +    int           ai, aj, si, ge, se;
 +    int           found, cj_ind_0, cj_ind_1, cj_ind_m;
 +    int           cj_m;
 +    int           inner_i, inner_e;
 +
 +    cell = nbs->cell;
 +
 +    if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
 +    {
 +        /* Empty list */
 +        return;
 +    }
 +
 +    ci = nbl_ci->ci;
 +
 +    cj_ind_first = nbl_ci->cj_ind_start;
 +    cj_ind_last  = nbl->ncj - 1;
 +
 +    cj_first = nbl->cj[cj_ind_first].cj;
 +    cj_last  = nbl->cj[cj_ind_last].cj;
 +
 +    /* Determine how many contiguous j-cells we have starting
 +     * from the first i-cell. This number can be used to directly
 +     * calculate j-cell indices for excluded atoms.
 +     */
 +    ndirect = 0;
 +    if (na_ci_2log == na_cj_2log)
 +    {
 +        while (cj_ind_first + ndirect <= cj_ind_last &&
 +               nbl->cj[cj_ind_first+ndirect].cj == ci + ndirect)
 +        {
 +            ndirect++;
 +        }
 +    }
 +#if NBNXN_SEARCH_BB_SIMD4
 +    else
 +    {
 +        while (cj_ind_first + ndirect <= cj_ind_last &&
 +               nbl->cj[cj_ind_first+ndirect].cj == ci_to_cj(ci, na_cj_2log) + ndirect)
 +        {
 +            ndirect++;
 +        }
 +    }
 +#endif
 +
 +    /* Loop over the atoms in the i super-cell */
 +    for (int i = 0; i < nbl->na_sc; i++)
 +    {
 +        ai = nbs->a[ci*nbl->na_sc+i];
 +        if (ai >= 0)
 +        {
 +            si  = (i>>na_ci_2log);
 +
 +            /* Loop over the topology-based exclusions for this i-atom */
 +            for (int eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
 +            {
 +                aj = excl->a[eind];
 +
 +                if (aj == ai)
 +                {
 +                    /* The self exclusion are already set, save some time */
 +                    continue;
 +                }
 +
 +                ge = cell[aj];
 +
 +                /* Without shifts we only calculate interactions j>i
 +                 * for one-way pair-lists.
 +                 */
 +                if (diagRemoved && ge <= ci*nbl->na_sc + i)
 +                {
 +                    continue;
 +                }
 +
 +                se = (ge >> na_cj_2log);
 +
 +                /* Could the cluster se be in our list? */
 +                if (se >= cj_first && se <= cj_last)
 +                {
 +                    if (se < cj_first + ndirect)
 +                    {
 +                        /* We can calculate cj_ind directly from se */
 +                        found = cj_ind_first + se - cj_first;
 +                    }
 +                    else
 +                    {
 +                        /* Search for se using bisection */
 +                        found    = -1;
 +                        cj_ind_0 = cj_ind_first + ndirect;
 +                        cj_ind_1 = cj_ind_last + 1;
 +                        while (found == -1 && cj_ind_0 < cj_ind_1)
 +                        {
 +                            cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
 +
 +                            cj_m = nbl->cj[cj_ind_m].cj;
 +
 +                            if (se == cj_m)
 +                            {
 +                                found = cj_ind_m;
 +                            }
 +                            else if (se < cj_m)
 +                            {
 +                                cj_ind_1 = cj_ind_m;
 +                            }
 +                            else
 +                            {
 +                                cj_ind_0 = cj_ind_m + 1;
 +                            }
 +                        }
 +                    }
 +
 +                    if (found >= 0)
 +                    {
 +                        inner_i = i  - (si << na_ci_2log);
 +                        inner_e = ge - (se << na_cj_2log);
 +
 +                        nbl->cj[found].excl &= ~(1U<<((inner_i<<na_cj_2log) + inner_e));
 +                    }
 +                }
 +            }
 +        }
 +    }
 +}
 +
 +/* Add a new i-entry to the FEP list and copy the i-properties */
 +static gmx_inline void fep_list_new_nri_copy(t_nblist *nlist)
 +{
 +    /* Add a new i-entry */
 +    nlist->nri++;
 +
 +    assert(nlist->nri < nlist->maxnri);
 +
 +    /* Duplicate the last i-entry, except for jindex, which continues */
 +    nlist->iinr[nlist->nri]   = nlist->iinr[nlist->nri-1];
 +    nlist->shift[nlist->nri]  = nlist->shift[nlist->nri-1];
 +    nlist->gid[nlist->nri]    = nlist->gid[nlist->nri-1];
 +    nlist->jindex[nlist->nri] = nlist->nrj;
 +}
 +
 +/* For load balancing of the free-energy lists over threads, we set
 + * the maximum nrj size of an i-entry to 40. This leads to good
 + * load balancing in the worst case scenario of a single perturbed
 + * particle on 16 threads, while not introducing significant overhead.
 + * Note that half of the perturbed pairs will anyhow end up in very small lists,
 + * since non perturbed i-particles will see few perturbed j-particles).
 + */
 +const int max_nrj_fep = 40;
 +
 +/* Exclude the perturbed pairs from the Verlet list. This is only done to avoid
 + * singularities for overlapping particles (0/0), since the charges and
 + * LJ parameters have been zeroed in the nbnxn data structure.
 + * Simultaneously make a group pair list for the perturbed pairs.
 + */
 +static void make_fep_list(const nbnxn_search_t    nbs,
 +                          const nbnxn_atomdata_t *nbat,
 +                          nbnxn_pairlist_t       *nbl,
 +                          gmx_bool                bDiagRemoved,
 +                          nbnxn_ci_t             *nbl_ci,
 +                          const nbnxn_grid_t     *gridi,
 +                          const nbnxn_grid_t     *gridj,
 +                          t_nblist               *nlist)
 +{
 +    int      ci, cj_ind_start, cj_ind_end, cja, cjr;
 +    int      nri_max;
 +    int      ngid, gid_i = 0, gid_j, gid;
 +    int      egp_shift, egp_mask;
 +    int      gid_cj = 0;
 +    int      ind_i, ind_j, ai, aj;
 +    int      nri;
 +    gmx_bool bFEP_i, bFEP_i_all;
 +
 +    if (nbl_ci->cj_ind_end == nbl_ci->cj_ind_start)
 +    {
 +        /* Empty list */
 +        return;
 +    }
 +
 +    ci = nbl_ci->ci;
 +
 +    cj_ind_start = nbl_ci->cj_ind_start;
 +    cj_ind_end   = nbl_ci->cj_ind_end;
 +
 +    /* In worst case we have alternating energy groups
 +     * and create #atom-pair lists, which means we need the size
 +     * of a cluster pair (na_ci*na_cj) times the number of cj's.
 +     */
 +    nri_max = nbl->na_ci*nbl->na_cj*(cj_ind_end - cj_ind_start);
 +    if (nlist->nri + nri_max > nlist->maxnri)
 +    {
 +        nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
 +        reallocate_nblist(nlist);
 +    }
 +
 +    ngid = nbat->nenergrp;
 +
 +    if (static_cast<std::size_t>(ngid*gridj->na_cj) > sizeof(gid_cj)*8)
 +    {
 +        gmx_fatal(FARGS, "The Verlet scheme with %dx%d kernels and free-energy only supports up to %d energy groups",
 +                  gridi->na_c, gridj->na_cj, (sizeof(gid_cj)*8)/gridj->na_cj);
 +    }
 +
 +    egp_shift = nbat->neg_2log;
 +    egp_mask  = (1<<nbat->neg_2log) - 1;
 +
 +    /* Loop over the atoms in the i sub-cell */
 +    bFEP_i_all = TRUE;
 +    for (int i = 0; i < nbl->na_ci; i++)
 +    {
 +        ind_i = ci*nbl->na_ci + i;
 +        ai    = nbs->a[ind_i];
 +        if (ai >= 0)
 +        {
 +            nri                  = nlist->nri;
 +            nlist->jindex[nri+1] = nlist->jindex[nri];
 +            nlist->iinr[nri]     = ai;
 +            /* The actual energy group pair index is set later */
 +            nlist->gid[nri]      = 0;
 +            nlist->shift[nri]    = nbl_ci->shift & NBNXN_CI_SHIFT;
 +
 +            bFEP_i = gridi->fep[ci - gridi->cell0] & (1 << i);
 +
 +            bFEP_i_all = bFEP_i_all && bFEP_i;
 +
 +            if ((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj > nlist->maxnrj)
 +            {
 +                nlist->maxnrj = over_alloc_small((nlist->nrj + cj_ind_end - cj_ind_start)*nbl->na_cj);
 +                srenew(nlist->jjnr,     nlist->maxnrj);
 +                srenew(nlist->excl_fep, nlist->maxnrj);
 +            }
 +
 +            if (ngid > 1)
 +            {
 +                gid_i = (nbat->energrp[ci] >> (egp_shift*i)) & egp_mask;
 +            }
 +
 +            for (int cj_ind = cj_ind_start; cj_ind < cj_ind_end; cj_ind++)
 +            {
 +                unsigned int fep_cj;
 +
 +                cja = nbl->cj[cj_ind].cj;
 +
 +                if (gridj->na_cj == gridj->na_c)
 +                {
 +                    cjr    = cja - gridj->cell0;
 +                    fep_cj = gridj->fep[cjr];
 +                    if (ngid > 1)
 +                    {
 +                        gid_cj = nbat->energrp[cja];
 +                    }
 +                }
 +                else if (2*gridj->na_cj == gridj->na_c)
 +                {
 +                    cjr    = cja - gridj->cell0*2;
 +                    /* Extract half of the ci fep/energrp mask */
 +                    fep_cj = (gridj->fep[cjr>>1] >> ((cjr&1)*gridj->na_cj)) & ((1<<gridj->na_cj) - 1);
 +                    if (ngid > 1)
 +                    {
 +                        gid_cj = nbat->energrp[cja>>1] >> ((cja&1)*gridj->na_cj*egp_shift) & ((1<<(gridj->na_cj*egp_shift)) - 1);
 +                    }
 +                }
 +                else
 +                {
 +                    cjr    = cja - (gridj->cell0>>1);
 +                    /* Combine two ci fep masks/energrp */
 +                    fep_cj = gridj->fep[cjr*2] + (gridj->fep[cjr*2+1] << gridj->na_c);
 +                    if (ngid > 1)
 +                    {
 +                        gid_cj = nbat->energrp[cja*2] + (nbat->energrp[cja*2+1] << (gridj->na_c*egp_shift));
 +                    }
 +                }
 +
 +                if (bFEP_i || fep_cj != 0)
 +                {
 +                    for (int j = 0; j < nbl->na_cj; j++)
 +                    {
 +                        /* Is this interaction perturbed and not excluded? */
 +                        ind_j = cja*nbl->na_cj + j;
 +                        aj    = nbs->a[ind_j];
 +                        if (aj >= 0 &&
 +                            (bFEP_i || (fep_cj & (1 << j))) &&
 +                            (!bDiagRemoved || ind_j >= ind_i))
 +                        {
 +                            if (ngid > 1)
 +                            {
 +                                gid_j = (gid_cj >> (j*egp_shift)) & egp_mask;
 +                                gid   = GID(gid_i, gid_j, ngid);
 +
 +                                if (nlist->nrj > nlist->jindex[nri] &&
 +                                    nlist->gid[nri] != gid)
 +                                {
 +                                    /* Energy group pair changed: new list */
 +                                    fep_list_new_nri_copy(nlist);
 +                                    nri = nlist->nri;
 +                                }
 +                                nlist->gid[nri] = gid;
 +                            }
 +
 +                            if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
 +                            {
 +                                fep_list_new_nri_copy(nlist);
 +                                nri = nlist->nri;
 +                            }
 +
 +                            /* Add it to the FEP list */
 +                            nlist->jjnr[nlist->nrj]     = aj;
 +                            nlist->excl_fep[nlist->nrj] = (nbl->cj[cj_ind].excl >> (i*nbl->na_cj + j)) & 1;
 +                            nlist->nrj++;
 +
 +                            /* Exclude it from the normal list.
 +                             * Note that the charge has been set to zero,
 +                             * but we need to avoid 0/0, as perturbed atoms
 +                             * can be on top of each other.
 +                             */
 +                            nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
 +                        }
 +                    }
 +                }
 +            }
 +
 +            if (nlist->nrj > nlist->jindex[nri])
 +            {
 +                /* Actually add this new, non-empty, list */
 +                nlist->nri++;
 +                nlist->jindex[nlist->nri] = nlist->nrj;
 +            }
 +        }
 +    }
 +
 +    if (bFEP_i_all)
 +    {
 +        /* All interactions are perturbed, we can skip this entry */
 +        nbl_ci->cj_ind_end = cj_ind_start;
 +    }
 +}
 +
 +/* Return the index of atom a within a cluster */
 +static gmx_inline int cj_mod_cj4(int cj)
 +{
 +    return cj & (c_nbnxnGpuJgroupSize - 1);
 +}
 +
 +/* Convert a j-cluster to a cj4 group */
 +static gmx_inline int cj_to_cj4(int cj)
 +{
 +    return cj/c_nbnxnGpuJgroupSize;
 +}
 +
 +/* Return the index of an j-atom within a warp */
 +static gmx_inline int a_mod_wj(int a)
 +{
 +    return a & (c_nbnxnGpuClusterSize/c_nbnxnGpuClusterpairSplit - 1);
 +}
 +
 +/* As make_fep_list above, but for super/sub lists. */
 +static void make_fep_list_supersub(const nbnxn_search_t    nbs,
 +                                   const nbnxn_atomdata_t *nbat,
 +                                   nbnxn_pairlist_t       *nbl,
 +                                   gmx_bool                bDiagRemoved,
 +                                   const nbnxn_sci_t      *nbl_sci,
 +                                   real                    shx,
 +                                   real                    shy,
 +                                   real                    shz,
 +                                   real                    rlist_fep2,
 +                                   const nbnxn_grid_t     *gridi,
 +                                   const nbnxn_grid_t     *gridj,
 +                                   t_nblist               *nlist)
 +{
 +    int                sci, cj4_ind_start, cj4_ind_end, cjr;
 +    int                nri_max;
 +    int                c_abs;
 +    int                ind_i, ind_j, ai, aj;
 +    int                nri;
 +    gmx_bool           bFEP_i;
 +    real               xi, yi, zi;
 +    const nbnxn_cj4_t *cj4;
 +
 +    if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
 +    {
 +        /* Empty list */
 +        return;
 +    }
 +
 +    sci = nbl_sci->sci;
 +
 +    cj4_ind_start = nbl_sci->cj4_ind_start;
 +    cj4_ind_end   = nbl_sci->cj4_ind_end;
 +
 +    /* Here we process one super-cell, max #atoms na_sc, versus a list
 +     * cj4 entries, each with max c_nbnxnGpuJgroupSize cj's, each
 +     * of size na_cj atoms.
 +     * On the GPU we don't support energy groups (yet).
 +     * So for each of the na_sc i-atoms, we need max one FEP list
 +     * for each max_nrj_fep j-atoms.
 +     */
 +    nri_max = nbl->na_sc*nbl->na_cj*(1 + ((cj4_ind_end - cj4_ind_start)*c_nbnxnGpuJgroupSize)/max_nrj_fep);
 +    if (nlist->nri + nri_max > nlist->maxnri)
 +    {
 +        nlist->maxnri = over_alloc_large(nlist->nri + nri_max);
 +        reallocate_nblist(nlist);
 +    }
 +
 +    /* Loop over the atoms in the i super-cluster */
 +    for (int c = 0; c < c_gpuNumClusterPerCell; c++)
 +    {
 +        c_abs = sci*c_gpuNumClusterPerCell + c;
 +
 +        for (int i = 0; i < nbl->na_ci; i++)
 +        {
 +            ind_i = c_abs*nbl->na_ci + i;
 +            ai    = nbs->a[ind_i];
 +            if (ai >= 0)
 +            {
 +                nri                  = nlist->nri;
 +                nlist->jindex[nri+1] = nlist->jindex[nri];
 +                nlist->iinr[nri]     = ai;
 +                /* With GPUs, energy groups are not supported */
 +                nlist->gid[nri]      = 0;
 +                nlist->shift[nri]    = nbl_sci->shift & NBNXN_CI_SHIFT;
 +
 +                bFEP_i = (gridi->fep[c_abs - gridi->cell0*c_gpuNumClusterPerCell] & (1 << i));
 +
 +                xi = nbat->x[ind_i*nbat->xstride+XX] + shx;
 +                yi = nbat->x[ind_i*nbat->xstride+YY] + shy;
 +                zi = nbat->x[ind_i*nbat->xstride+ZZ] + shz;
 +
 +                if ((nlist->nrj + cj4_ind_end - cj4_ind_start)*c_nbnxnGpuJgroupSize*nbl->na_cj > nlist->maxnrj)
 +                {
 +                    nlist->maxnrj = over_alloc_small((nlist->nrj + cj4_ind_end - cj4_ind_start)*c_nbnxnGpuJgroupSize*nbl->na_cj);
 +                    srenew(nlist->jjnr,     nlist->maxnrj);
 +                    srenew(nlist->excl_fep, nlist->maxnrj);
 +                }
 +
 +                for (int cj4_ind = cj4_ind_start; cj4_ind < cj4_ind_end; cj4_ind++)
 +                {
 +                    cj4 = &nbl->cj4[cj4_ind];
 +
 +                    for (int gcj = 0; gcj < c_nbnxnGpuJgroupSize; gcj++)
 +                    {
 +                        unsigned int fep_cj;
 +
 +                        if ((cj4->imei[0].imask & (1U << (gcj*c_gpuNumClusterPerCell + c))) == 0)
 +                        {
 +                            /* Skip this ci for this cj */
 +                            continue;
 +                        }
 +
 +                        cjr = cj4->cj[gcj] - gridj->cell0*c_gpuNumClusterPerCell;
 +
 +                        fep_cj = gridj->fep[cjr];
 +
 +                        if (bFEP_i || fep_cj != 0)
 +                        {
 +                            for (int j = 0; j < nbl->na_cj; j++)
 +                            {
 +                                /* Is this interaction perturbed and not excluded? */
 +                                ind_j = (gridj->cell0*c_gpuNumClusterPerCell + cjr)*nbl->na_cj + j;
 +                                aj    = nbs->a[ind_j];
 +                                if (aj >= 0 &&
 +                                    (bFEP_i || (fep_cj & (1 << j))) &&
 +                                    (!bDiagRemoved || ind_j >= ind_i))
 +                                {
 +                                    nbnxn_excl_t *excl;
 +                                    int           excl_pair;
 +                                    unsigned int  excl_bit;
 +                                    real          dx, dy, dz;
 +
 +                                    get_nbl_exclusions_1(nbl, cj4_ind, j>>2, &excl);
 +
 +                                    excl_pair = a_mod_wj(j)*nbl->na_ci + i;
 +                                    excl_bit  = (1U << (gcj*c_gpuNumClusterPerCell + c));
 +
 +                                    dx = nbat->x[ind_j*nbat->xstride+XX] - xi;
 +                                    dy = nbat->x[ind_j*nbat->xstride+YY] - yi;
 +                                    dz = nbat->x[ind_j*nbat->xstride+ZZ] - zi;
 +
 +                                    /* The unpruned GPU list has more than 2/3
 +                                     * of the atom pairs beyond rlist. Using
 +                                     * this list will cause a lot of overhead
 +                                     * in the CPU FEP kernels, especially
 +                                     * relative to the fast GPU kernels.
 +                                     * So we prune the FEP list here.
 +                                     */
 +                                    if (dx*dx + dy*dy + dz*dz < rlist_fep2)
 +                                    {
 +                                        if (nlist->nrj - nlist->jindex[nri] >= max_nrj_fep)
 +                                        {
 +                                            fep_list_new_nri_copy(nlist);
 +                                            nri = nlist->nri;
 +                                        }
 +
 +                                        /* Add it to the FEP list */
 +                                        nlist->jjnr[nlist->nrj]     = aj;
 +                                        nlist->excl_fep[nlist->nrj] = (excl->pair[excl_pair] & excl_bit) ? 1 : 0;
 +                                        nlist->nrj++;
 +                                    }
 +
 +                                    /* Exclude it from the normal list.
 +                                     * Note that the charge and LJ parameters have
 +                                     * been set to zero, but we need to avoid 0/0,
 +                                     * as perturbed atoms can be on top of each other.
 +                                     */
 +                                    excl->pair[excl_pair] &= ~excl_bit;
 +                                }
 +                            }
 +
 +                            /* Note that we could mask out this pair in imask
 +                             * if all i- and/or all j-particles are perturbed.
 +                             * But since the perturbed pairs on the CPU will
 +                             * take an order of magnitude more time, the GPU
 +                             * will finish before the CPU and there is no gain.
 +                             */
 +                        }
 +                    }
 +                }
 +
 +                if (nlist->nrj > nlist->jindex[nri])
 +                {
 +                    /* Actually add this new, non-empty, list */
 +                    nlist->nri++;
 +                    nlist->jindex[nlist->nri] = nlist->nrj;
 +                }
 +            }
 +        }
 +    }
 +}
 +
 +/* Set all atom-pair exclusions from the topology stored in excl
 + * as masks in the pair-list for i-super-cell entry nbl_sci
 + */
 +static void set_sci_top_excls(const nbnxn_search_t nbs,
 +                              nbnxn_pairlist_t    *nbl,
 +                              gmx_bool             diagRemoved,
 +                              int                  na_c_2log,
 +                              const nbnxn_sci_t   *nbl_sci,
 +                              const t_blocka      *excl)
 +{
 +    const int    *cell;
 +    int           na_c;
 +    int           sci;
 +    int           cj_ind_first, cj_ind_last;
 +    int           cj_first, cj_last;
 +    int           ndirect;
 +    int           ai, aj, si, ge, se;
 +    int           found, cj_ind_0, cj_ind_1, cj_ind_m;
 +    int           cj_m;
 +    nbnxn_excl_t *nbl_excl;
 +    int           inner_i, inner_e, w;
 +
 +    cell = nbs->cell;
 +
 +    na_c = nbl->na_ci;
 +
 +    if (nbl_sci->cj4_ind_end == nbl_sci->cj4_ind_start)
 +    {
 +        /* Empty list */
 +        return;
 +    }
 +
 +    sci = nbl_sci->sci;
 +
 +    cj_ind_first = nbl_sci->cj4_ind_start*c_nbnxnGpuJgroupSize;
 +    cj_ind_last  = nbl->work->cj_ind - 1;
 +
 +    cj_first = nbl->cj4[nbl_sci->cj4_ind_start].cj[0];
 +    cj_last  = nbl_cj(nbl, cj_ind_last);
 +
 +    /* Determine how many contiguous j-clusters we have starting
 +     * from the first i-cluster. This number can be used to directly
 +     * calculate j-cluster indices for excluded atoms.
 +     */
 +    ndirect = 0;
 +    while (cj_ind_first + ndirect <= cj_ind_last &&
 +           nbl_cj(nbl, cj_ind_first+ndirect) == sci*c_gpuNumClusterPerCell + ndirect)
 +    {
 +        ndirect++;
 +    }
 +
 +    /* Loop over the atoms in the i super-cell */
 +    for (int i = 0; i < nbl->na_sc; i++)
 +    {
 +        ai = nbs->a[sci*nbl->na_sc+i];
 +        if (ai >= 0)
 +        {
 +            si  = (i>>na_c_2log);
 +
 +            /* Loop over the topology-based exclusions for this i-atom */
 +            for (int eind = excl->index[ai]; eind < excl->index[ai+1]; eind++)
 +            {
 +                aj = excl->a[eind];
 +
 +                if (aj == ai)
 +                {
 +                    /* The self exclusion are already set, save some time */
 +                    continue;
 +                }
 +
 +                ge = cell[aj];
 +
 +                /* Without shifts we only calculate interactions j>i
 +                 * for one-way pair-lists.
 +                 */
 +                if (diagRemoved && ge <= sci*nbl->na_sc + i)
 +                {
 +                    continue;
 +                }
 +
 +                se = ge>>na_c_2log;
 +                /* Could the cluster se be in our list? */
 +                if (se >= cj_first && se <= cj_last)
 +                {
 +                    if (se < cj_first + ndirect)
 +                    {
 +                        /* We can calculate cj_ind directly from se */
 +                        found = cj_ind_first + se - cj_first;
 +                    }
 +                    else
 +                    {
 +                        /* Search for se using bisection */
 +                        found    = -1;
 +                        cj_ind_0 = cj_ind_first + ndirect;
 +                        cj_ind_1 = cj_ind_last + 1;
 +                        while (found == -1 && cj_ind_0 < cj_ind_1)
 +                        {
 +                            cj_ind_m = (cj_ind_0 + cj_ind_1)>>1;
 +
 +                            cj_m = nbl_cj(nbl, cj_ind_m);
 +
 +                            if (se == cj_m)
 +                            {
 +                                found = cj_ind_m;
 +                            }
 +                            else if (se < cj_m)
 +                            {
 +                                cj_ind_1 = cj_ind_m;
 +                            }
 +                            else
 +                            {
 +                                cj_ind_0 = cj_ind_m + 1;
 +                            }
 +                        }
 +                    }
 +
 +                    if (found >= 0)
 +                    {
 +                        inner_i = i  - si*na_c;
 +                        inner_e = ge - se*na_c;
 +
 +                        if (nbl_imask0(nbl, found) & (1U << (cj_mod_cj4(found)*c_gpuNumClusterPerCell + si)))
 +                        {
 +                            w       = (inner_e >> 2);
 +
 +                            get_nbl_exclusions_1(nbl, cj_to_cj4(found), w, &nbl_excl);
 +
 +                            nbl_excl->pair[a_mod_wj(inner_e)*nbl->na_ci+inner_i] &=
 +                                ~(1U << (cj_mod_cj4(found)*c_gpuNumClusterPerCell + si));
 +                        }
 +                    }
 +                }
 +            }
 +        }
 +    }
 +}
 +
 +/* Reallocate the simple ci list for at least n entries */
 +static void nb_realloc_ci(nbnxn_pairlist_t *nbl, int n)
 +{
 +    nbl->ci_nalloc = over_alloc_small(n);
 +    nbnxn_realloc_void((void **)&nbl->ci,
 +                       nbl->nci*sizeof(*nbl->ci),
 +                       nbl->ci_nalloc*sizeof(*nbl->ci),
 +                       nbl->alloc, nbl->free);
 +}
 +
 +/* Reallocate the super-cell sci list for at least n entries */
 +static void nb_realloc_sci(nbnxn_pairlist_t *nbl, int n)
 +{
 +    nbl->sci_nalloc = over_alloc_small(n);
 +    nbnxn_realloc_void((void **)&nbl->sci,
 +                       nbl->nsci*sizeof(*nbl->sci),
 +                       nbl->sci_nalloc*sizeof(*nbl->sci),
 +                       nbl->alloc, nbl->free);
 +}
 +
 +/* Make a new ci entry at index nbl->nci */
 +static void new_ci_entry(nbnxn_pairlist_t *nbl, int ci, int shift, int flags)
 +{
 +    if (nbl->nci + 1 > nbl->ci_nalloc)
 +    {
 +        nb_realloc_ci(nbl, nbl->nci+1);
 +    }
 +    nbl->ci[nbl->nci].ci            = ci;
 +    nbl->ci[nbl->nci].shift         = shift;
 +    /* Store the interaction flags along with the shift */
 +    nbl->ci[nbl->nci].shift        |= flags;
 +    nbl->ci[nbl->nci].cj_ind_start  = nbl->ncj;
 +    nbl->ci[nbl->nci].cj_ind_end    = nbl->ncj;
 +}
 +
 +/* Make a new sci entry at index nbl->nsci */
 +static void new_sci_entry(nbnxn_pairlist_t *nbl, int sci, int shift)
 +{
 +    if (nbl->nsci + 1 > nbl->sci_nalloc)
 +    {
 +        nb_realloc_sci(nbl, nbl->nsci+1);
 +    }
 +    nbl->sci[nbl->nsci].sci           = sci;
 +    nbl->sci[nbl->nsci].shift         = shift;
 +    nbl->sci[nbl->nsci].cj4_ind_start = nbl->ncj4;
 +    nbl->sci[nbl->nsci].cj4_ind_end   = nbl->ncj4;
 +}
 +
 +/* Sort the simple j-list cj on exclusions.
 + * Entries with exclusions will all be sorted to the beginning of the list.
 + */
 +static void sort_cj_excl(nbnxn_cj_t *cj, int ncj,
 +                         nbnxn_list_work_t *work)
 +{
 +    int jnew;
 +
 +    if (ncj > work->cj_nalloc)
 +    {
 +        work->cj_nalloc = over_alloc_large(ncj);
 +        srenew(work->cj, work->cj_nalloc);
 +    }
 +
 +    /* Make a list of the j-cells involving exclusions */
 +    jnew = 0;
 +    for (int j = 0; j < ncj; j++)
 +    {
 +        if (cj[j].excl != NBNXN_INTERACTION_MASK_ALL)
 +        {
 +            work->cj[jnew++] = cj[j];
 +        }
 +    }
 +    /* Check if there are exclusions at all or not just the first entry */
 +    if (!((jnew == 0) ||
 +          (jnew == 1 && cj[0].excl != NBNXN_INTERACTION_MASK_ALL)))
 +    {
 +        for (int j = 0; j < ncj; j++)
 +        {
 +            if (cj[j].excl == NBNXN_INTERACTION_MASK_ALL)
 +            {
 +                work->cj[jnew++] = cj[j];
 +            }
 +        }
 +        for (int j = 0; j < ncj; j++)
 +        {
 +            cj[j] = work->cj[j];
 +        }
 +    }
 +}
 +
 +/* Close this simple list i entry */
 +static void close_ci_entry_simple(nbnxn_pairlist_t *nbl)
 +{
 +    int jlen;
 +
 +    /* All content of the new ci entry have already been filled correctly,
 +     * we only need to increase the count here (for non empty lists).
 +     */
 +    jlen = nbl->ci[nbl->nci].cj_ind_end - nbl->ci[nbl->nci].cj_ind_start;
 +    if (jlen > 0)
 +    {
 +        sort_cj_excl(nbl->cj+nbl->ci[nbl->nci].cj_ind_start, jlen, nbl->work);
 +
 +        /* The counts below are used for non-bonded pair/flop counts
 +         * and should therefore match the available kernel setups.
 +         */
 +        if (!(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_COUL(0)))
 +        {
 +            nbl->work->ncj_noq += jlen;
 +        }
 +        else if ((nbl->ci[nbl->nci].shift & NBNXN_CI_HALF_LJ(0)) ||
 +                 !(nbl->ci[nbl->nci].shift & NBNXN_CI_DO_LJ(0)))
 +        {
 +            nbl->work->ncj_hlj += jlen;
 +        }
 +
 +        nbl->nci++;
 +    }
 +}
 +
 +/* Split sci entry for load balancing on the GPU.
 + * Splitting ensures we have enough lists to fully utilize the whole GPU.
 + * With progBal we generate progressively smaller lists, which improves
 + * load balancing. As we only know the current count on our own thread,
 + * we will need to estimate the current total amount of i-entries.
 + * As the lists get concatenated later, this estimate depends
 + * both on nthread and our own thread index.
 + */
 +static void split_sci_entry(nbnxn_pairlist_t *nbl,
 +                            int nsp_target_av,
-                             gmx_bool progBal, int nsp_tot_est,
++                            gmx_bool progBal, float nsp_tot_est,
 +                            int thread, int nthread)
 +{
-     int nsp_est;
 +    int nsp_max;
 +    int cj4_start, cj4_end, j4len;
 +    int sci;
 +    int nsp, nsp_sci, nsp_cj4, nsp_cj4_e, nsp_cj4_p;
 +
 +    if (progBal)
 +    {
++        float nsp_est;
++
 +        /* Estimate the total numbers of ci's of the nblist combined
 +         * over all threads using the target number of ci's.
 +         */
 +        nsp_est = (nsp_tot_est*thread)/nthread + nbl->nci_tot;
 +
 +        /* The first ci blocks should be larger, to avoid overhead.
 +         * The last ci blocks should be smaller, to improve load balancing.
 +         * The factor 3/2 makes the first block 3/2 times the target average
 +         * and ensures that the total number of blocks end up equal to
-          * that with of equally sized blocks of size nsp_target_av.
++         * that of equally sized blocks of size nsp_target_av.
 +         */
-         nsp_max = nsp_target_av*nsp_tot_est*3/(2*(nsp_est + nsp_tot_est));
++        nsp_max = static_cast<int>(nsp_target_av*(nsp_tot_est*1.5/(nsp_est + nsp_tot_est)));
 +    }
 +    else
 +    {
 +        nsp_max = nsp_target_av;
 +    }
 +
 +    /* Since nsp_max is a maximum/cut-off (this avoids high outliers,
 +     * which lead to load imbalance), not an average, we add half the
 +     * number of pairs in a cj4 block to get the average about right.
 +     */
 +    nsp_max += c_gpuNumClusterPerCell*c_nbnxnGpuJgroupSize/2;
 +
 +    cj4_start = nbl->sci[nbl->nsci-1].cj4_ind_start;
 +    cj4_end   = nbl->sci[nbl->nsci-1].cj4_ind_end;
 +    j4len     = cj4_end - cj4_start;
 +
 +    if (j4len > 1 && j4len*c_gpuNumClusterPerCell*c_nbnxnGpuJgroupSize > nsp_max)
 +    {
 +        /* Remove the last ci entry and process the cj4's again */
 +        nbl->nsci -= 1;
 +
 +        sci        = nbl->nsci;
 +        nsp        = 0;
 +        nsp_sci    = 0;
 +        nsp_cj4_e  = 0;
 +        nsp_cj4    = 0;
 +        for (int cj4 = cj4_start; cj4 < cj4_end; cj4++)
 +        {
 +            nsp_cj4_p = nsp_cj4;
 +            /* Count the number of cluster pairs in this cj4 group */
 +            nsp_cj4   = 0;
 +            for (int p = 0; p < c_gpuNumClusterPerCell*c_nbnxnGpuJgroupSize; p++)
 +            {
 +                nsp_cj4 += (nbl->cj4[cj4].imei[0].imask >> p) & 1;
 +            }
 +
 +            /* Check if we should split at this cj4 to get a list of size nsp */
 +            if (nsp > 0 && nsp + nsp_cj4 > nsp_max)
 +            {
 +                /* Split the list at cj4 */
 +                nbl->sci[sci].cj4_ind_end = cj4;
 +                /* Create a new sci entry */
 +                sci++;
 +                nbl->nsci++;
 +                if (nbl->nsci+1 > nbl->sci_nalloc)
 +                {
 +                    nb_realloc_sci(nbl, nbl->nsci+1);
 +                }
 +                nbl->sci[sci].sci           = nbl->sci[nbl->nsci-1].sci;
 +                nbl->sci[sci].shift         = nbl->sci[nbl->nsci-1].shift;
 +                nbl->sci[sci].cj4_ind_start = cj4;
 +                nsp_sci                     = nsp;
 +                nsp_cj4_e                   = nsp_cj4_p;
 +                nsp                         = 0;
 +            }
 +            nsp += nsp_cj4;
 +        }
 +
 +        /* Put the remaining cj4's in the last sci entry */
 +        nbl->sci[sci].cj4_ind_end = cj4_end;
 +
 +        /* Possibly balance out the last two sci's
 +         * by moving the last cj4 of the second last sci.
 +         */
 +        if (nsp_sci - nsp_cj4_e >= nsp + nsp_cj4_e)
 +        {
 +            nbl->sci[sci-1].cj4_ind_end--;
 +            nbl->sci[sci].cj4_ind_start--;
 +        }
 +
 +        nbl->nsci++;
 +    }
 +}
 +
 +/* Clost this super/sub list i entry */
 +static void close_ci_entry_supersub(nbnxn_pairlist_t *nbl,
 +                                    int nsp_max_av,
-                                     gmx_bool progBal, int nsp_tot_est,
++                                    gmx_bool progBal, float nsp_tot_est,
 +                                    int thread, int nthread)
 +{
 +    /* All content of the new ci entry have already been filled correctly,
 +     * we only need to increase the count here (for non empty lists).
 +     */
 +    int j4len = nbl->sci[nbl->nsci].cj4_ind_end - nbl->sci[nbl->nsci].cj4_ind_start;
 +    if (j4len > 0)
 +    {
 +        /* We can only have complete blocks of 4 j-entries in a list,
 +         * so round the count up before closing.
 +         */
 +        nbl->ncj4         = (nbl->work->cj_ind + c_nbnxnGpuJgroupSize - 1)/c_nbnxnGpuJgroupSize;
 +        nbl->work->cj_ind = nbl->ncj4*c_nbnxnGpuJgroupSize;
 +
 +        nbl->nsci++;
 +
 +        if (nsp_max_av > 0)
 +        {
 +            /* Measure the size of the new entry and potentially split it */
 +            split_sci_entry(nbl, nsp_max_av, progBal, nsp_tot_est,
 +                            thread, nthread);
 +        }
 +    }
 +}
 +
 +/* Syncs the working array before adding another grid pair to the list */
 +static void sync_work(nbnxn_pairlist_t *nbl)
 +{
 +    if (!nbl->bSimple)
 +    {
 +        nbl->work->cj_ind   = nbl->ncj4*c_nbnxnGpuJgroupSize;
 +        nbl->work->cj4_init = nbl->ncj4;
 +    }
 +}
 +
 +/* Clears an nbnxn_pairlist_t data structure */
 +static void clear_pairlist(nbnxn_pairlist_t *nbl)
 +{
 +    nbl->nci           = 0;
 +    nbl->nsci          = 0;
 +    nbl->ncj           = 0;
 +    nbl->ncj4          = 0;
 +    nbl->nci_tot       = 0;
 +    nbl->nexcl         = 1;
 +
 +    nbl->work->ncj_noq = 0;
 +    nbl->work->ncj_hlj = 0;
 +}
 +
 +/* Clears a group scheme pair list */
 +static void clear_pairlist_fep(t_nblist *nl)
 +{
 +    nl->nri = 0;
 +    nl->nrj = 0;
 +    if (nl->jindex == NULL)
 +    {
 +        snew(nl->jindex, 1);
 +    }
 +    nl->jindex[0] = 0;
 +}
 +
 +/* Sets a simple list i-cell bounding box, including PBC shift */
 +static gmx_inline void set_icell_bb_simple(const nbnxn_bb_t *bb, int ci,
 +                                           real shx, real shy, real shz,
 +                                           nbnxn_bb_t *bb_ci)
 +{
 +    bb_ci->lower[BB_X] = bb[ci].lower[BB_X] + shx;
 +    bb_ci->lower[BB_Y] = bb[ci].lower[BB_Y] + shy;
 +    bb_ci->lower[BB_Z] = bb[ci].lower[BB_Z] + shz;
 +    bb_ci->upper[BB_X] = bb[ci].upper[BB_X] + shx;
 +    bb_ci->upper[BB_Y] = bb[ci].upper[BB_Y] + shy;
 +    bb_ci->upper[BB_Z] = bb[ci].upper[BB_Z] + shz;
 +}
 +
 +#if NBNXN_BBXXXX
 +/* Sets a super-cell and sub cell bounding boxes, including PBC shift */
 +static void set_icell_bbxxxx_supersub(const float *bb, int ci,
 +                                      real shx, real shy, real shz,
 +                                      float *bb_ci)
 +{
 +    int ia = ci*(c_gpuNumClusterPerCell >> STRIDE_PBB_2LOG)*NNBSBB_XXXX;
 +    for (int m = 0; m < (c_gpuNumClusterPerCell >> STRIDE_PBB_2LOG)*NNBSBB_XXXX; m += NNBSBB_XXXX)
 +    {
 +        for (int i = 0; i < STRIDE_PBB; i++)
 +        {
 +            bb_ci[m+0*STRIDE_PBB+i] = bb[ia+m+0*STRIDE_PBB+i] + shx;
 +            bb_ci[m+1*STRIDE_PBB+i] = bb[ia+m+1*STRIDE_PBB+i] + shy;
 +            bb_ci[m+2*STRIDE_PBB+i] = bb[ia+m+2*STRIDE_PBB+i] + shz;
 +            bb_ci[m+3*STRIDE_PBB+i] = bb[ia+m+3*STRIDE_PBB+i] + shx;
 +            bb_ci[m+4*STRIDE_PBB+i] = bb[ia+m+4*STRIDE_PBB+i] + shy;
 +            bb_ci[m+5*STRIDE_PBB+i] = bb[ia+m+5*STRIDE_PBB+i] + shz;
 +        }
 +    }
 +}
 +#endif
 +
 +/* Sets a super-cell and sub cell bounding boxes, including PBC shift */
 +static void set_icell_bb_supersub(const nbnxn_bb_t *bb, int ci,
 +                                  real shx, real shy, real shz,
 +                                  nbnxn_bb_t *bb_ci)
 +{
 +    for (int i = 0; i < c_gpuNumClusterPerCell; i++)
 +    {
 +        set_icell_bb_simple(bb, ci*c_gpuNumClusterPerCell+i,
 +                            shx, shy, shz,
 +                            &bb_ci[i]);
 +    }
 +}
 +
 +/* Copies PBC shifted i-cell atom coordinates x,y,z to working array */
 +static void icell_set_x_simple(int ci,
 +                               real shx, real shy, real shz,
 +                               int stride, const real *x,
 +                               nbnxn_list_work_t *work)
 +{
 +    int ia = ci*NBNXN_CPU_CLUSTER_I_SIZE;
 +
 +    for (int i = 0; i < NBNXN_CPU_CLUSTER_I_SIZE; i++)
 +    {
 +        work->x_ci[i*STRIDE_XYZ+XX] = x[(ia+i)*stride+XX] + shx;
 +        work->x_ci[i*STRIDE_XYZ+YY] = x[(ia+i)*stride+YY] + shy;
 +        work->x_ci[i*STRIDE_XYZ+ZZ] = x[(ia+i)*stride+ZZ] + shz;
 +    }
 +}
 +
 +/* Copies PBC shifted super-cell atom coordinates x,y,z to working array */
 +static void icell_set_x_supersub(int ci,
 +                                 real shx, real shy, real shz,
 +                                 int stride, const real *x,
 +                                 nbnxn_list_work_t *work)
 +{
 +#if !NBNXN_SEARCH_BB_SIMD4
 +
 +    real * x_ci = work->x_ci;
 +
 +    int    ia = ci*c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize;
 +    for (int i = 0; i < c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize; i++)
 +    {
 +        x_ci[i*DIM + XX] = x[(ia+i)*stride + XX] + shx;
 +        x_ci[i*DIM + YY] = x[(ia+i)*stride + YY] + shy;
 +        x_ci[i*DIM + ZZ] = x[(ia+i)*stride + ZZ] + shz;
 +    }
 +
 +#else /* !NBNXN_SEARCH_BB_SIMD4 */
 +
 +    real * x_ci = work->x_ci_simd;
 +
 +    for (int si = 0; si < c_gpuNumClusterPerCell; si++)
 +    {
 +        for (int i = 0; i < c_nbnxnGpuClusterSize; i += GMX_SIMD4_WIDTH)
 +        {
 +            int io = si*c_nbnxnGpuClusterSize + i;
 +            int ia = ci*c_gpuNumClusterPerCell*c_nbnxnGpuClusterSize + io;
 +            for (int j = 0; j < GMX_SIMD4_WIDTH; j++)
 +            {
 +                x_ci[io*DIM + j + XX*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + XX] + shx;
 +                x_ci[io*DIM + j + YY*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + YY] + shy;
 +                x_ci[io*DIM + j + ZZ*GMX_SIMD4_WIDTH] = x[(ia + j)*stride + ZZ] + shz;
 +            }
 +        }
 +    }
 +
 +#endif /* !NBNXN_SEARCH_BB_SIMD4 */
 +}
 +
 +static real minimum_subgrid_size_xy(const nbnxn_grid_t *grid)
 +{
 +    if (grid->bSimple)
 +    {
 +        return std::min(grid->sx, grid->sy);
 +    }
 +    else
 +    {
 +        return std::min(grid->sx/c_gpuNumClusterPerCellX,
 +                        grid->sy/c_gpuNumClusterPerCellY);
 +    }
 +}
 +
 +static real effective_buffer_1x1_vs_MxN(const nbnxn_grid_t *gridi,
 +                                        const nbnxn_grid_t *gridj)
 +{
 +    const real eff_1x1_buffer_fac_overest = 0.1;
 +
 +    /* Determine an atom-pair list cut-off buffer size for atom pairs,
 +     * to be added to rlist (including buffer) used for MxN.
 +     * This is for converting an MxN list to a 1x1 list. This means we can't
 +     * use the normal buffer estimate, as we have an MxN list in which
 +     * some atom pairs beyond rlist are missing. We want to capture
 +     * the beneficial effect of buffering by extra pairs just outside rlist,
 +     * while removing the useless pairs that are further away from rlist.
 +     * (Also the buffer could have been set manually not using the estimate.)
 +     * This buffer size is an overestimate.
 +     * We add 10% of the smallest grid sub-cell dimensions.
 +     * Note that the z-size differs per cell and we don't use this,
 +     * so we overestimate.
 +     * With PME, the 10% value gives a buffer that is somewhat larger
 +     * than the effective buffer with a tolerance of 0.005 kJ/mol/ps.
 +     * Smaller tolerances or using RF lead to a smaller effective buffer,
 +     * so 10% gives a safe overestimate.
 +     */
 +    return eff_1x1_buffer_fac_overest*(minimum_subgrid_size_xy(gridi) +
 +                                       minimum_subgrid_size_xy(gridj));
 +}
 +
 +/* Clusters at the cut-off only increase rlist by 60% of their size */
 +static real nbnxn_rlist_inc_outside_fac = 0.6;
 +
 +/* Due to the cluster size the effective pair-list is longer than
 + * that of a simple atom pair-list. This function gives the extra distance.
 + */
 +real nbnxn_get_rlist_effective_inc(int cluster_size_j, real atom_density)
 +{
 +    int  cluster_size_i;
 +    real vol_inc_i, vol_inc_j;
 +
 +    /* We should get this from the setup, but currently it's the same for
 +     * all setups, including GPUs.
 +     */
 +    cluster_size_i = NBNXN_CPU_CLUSTER_I_SIZE;
 +
 +    vol_inc_i = (cluster_size_i - 1)/atom_density;
 +    vol_inc_j = (cluster_size_j - 1)/atom_density;
 +
 +    return nbnxn_rlist_inc_outside_fac*std::cbrt(vol_inc_i + vol_inc_j);
 +}
 +
 +/* Estimates the interaction volume^2 for non-local interactions */
 +static real nonlocal_vol2(const struct gmx_domdec_zones_t *zones, rvec ls, real r)
 +{
 +    real cl, ca, za;
 +    real vold_est;
 +    real vol2_est_tot;
 +
 +    vol2_est_tot = 0;
 +
 +    /* Here we simply add up the volumes of 1, 2 or 3 1D decomposition
 +     * not home interaction volume^2. As these volumes are not additive,
 +     * this is an overestimate, but it would only be significant in the limit
 +     * of small cells, where we anyhow need to split the lists into
 +     * as small parts as possible.
 +     */
 +
 +    for (int z = 0; z < zones->n; z++)
 +    {
 +        if (zones->shift[z][XX] + zones->shift[z][YY] + zones->shift[z][ZZ] == 1)
 +        {
 +            cl = 0;
 +            ca = 1;
 +            za = 1;
 +            for (int d = 0; d < DIM; d++)
 +            {
 +                if (zones->shift[z][d] == 0)
 +                {
 +                    cl += 0.5*ls[d];
 +                    ca *= ls[d];
 +                    za *= zones->size[z].x1[d] - zones->size[z].x0[d];
 +                }
 +            }
 +
 +            /* 4 octants of a sphere */
 +            vold_est  = 0.25*M_PI*r*r*r*r;
 +            /* 4 quarter pie slices on the edges */
 +            vold_est += 4*cl*M_PI/6.0*r*r*r;
 +            /* One rectangular volume on a face */
 +            vold_est += ca*0.5*r*r;
 +
 +            vol2_est_tot += vold_est*za;
 +        }
 +    }
 +
 +    return vol2_est_tot;
 +}
 +
 +/* Estimates the average size of a full j-list for super/sub setup */
 +static void get_nsubpair_target(const nbnxn_search_t  nbs,
 +                                int                   iloc,
 +                                real                  rlist,
 +                                int                   min_ci_balanced,
 +                                int                  *nsubpair_target,
-                                 int                  *nsubpair_tot_est)
++                                float                *nsubpair_tot_est)
 +{
 +    /* The target value of 36 seems to be the optimum for Kepler.
 +     * Maxwell is less sensitive to the exact value.
 +     */
 +    const int           nsubpair_target_min = 36;
 +    const nbnxn_grid_t *grid;
 +    rvec                ls;
 +    real                xy_diag2, r_eff_sup, vol_est, nsp_est, nsp_est_nl;
 +
 +    grid = &nbs->grid[0];
 +
++    /* We don't need to balance list sizes if:
++     * - We didn't request balancing.
++     * - The number of grid cells >= the number of lists requested,
++     *   since we will always generate at least #cells lists.
++     * - We don't have any cells, since then there won't be any lists.
++     */
 +    if (min_ci_balanced <= 0 || grid->nc >= min_ci_balanced || grid->nc == 0)
 +    {
-         /* We don't need to balance the list sizes */
++        /* nsubpair_target==0 signals no balancing */
 +        *nsubpair_target  = 0;
 +        *nsubpair_tot_est = 0;
 +
 +        return;
 +    }
 +
 +    ls[XX] = (grid->c1[XX] - grid->c0[XX])/(grid->ncx*c_gpuNumClusterPerCellX);
 +    ls[YY] = (grid->c1[YY] - grid->c0[YY])/(grid->ncy*c_gpuNumClusterPerCellY);
 +    ls[ZZ] = grid->na_c/(grid->atom_density*ls[XX]*ls[YY]);
 +
 +    /* The average squared length of the diagonal of a sub cell */
 +    xy_diag2 = ls[XX]*ls[XX] + ls[YY]*ls[YY] + ls[ZZ]*ls[ZZ];
 +
 +    /* The formulas below are a heuristic estimate of the average nsj per si*/
 +    r_eff_sup = rlist + nbnxn_rlist_inc_outside_fac*gmx::square((grid->na_c - 1.0)/grid->na_c)*std::sqrt(xy_diag2/3);
 +
 +    if (!nbs->DomDec || nbs->zones->n == 1)
 +    {
 +        nsp_est_nl = 0;
 +    }
 +    else
 +    {
 +        nsp_est_nl =
 +            gmx::square(grid->atom_density/grid->na_c)*
 +            nonlocal_vol2(nbs->zones, ls, r_eff_sup);
 +    }
 +
 +    if (LOCAL_I(iloc))
 +    {
 +        /* Sub-cell interacts with itself */
 +        vol_est  = ls[XX]*ls[YY]*ls[ZZ];
 +        /* 6/2 rectangular volume on the faces */
 +        vol_est += (ls[XX]*ls[YY] + ls[XX]*ls[ZZ] + ls[YY]*ls[ZZ])*r_eff_sup;
 +        /* 12/2 quarter pie slices on the edges */
 +        vol_est += 2*(ls[XX] + ls[YY] + ls[ZZ])*0.25*M_PI*gmx::square(r_eff_sup);
 +        /* 4 octants of a sphere */
 +        vol_est += 0.5*4.0/3.0*M_PI*gmx::power3(r_eff_sup);
 +
 +        /* Estimate the number of cluster pairs as the local number of
 +         * clusters times the volume they interact with times the density.
 +         */
 +        nsp_est = grid->nsubc_tot*vol_est*grid->atom_density/grid->na_c;
 +
 +        /* Subtract the non-local pair count */
 +        nsp_est -= nsp_est_nl;
 +
 +        /* For small cut-offs nsp_est will be an underesimate.
 +         * With DD nsp_est_nl is an overestimate so nsp_est can get negative.
 +         * So to avoid too small or negative nsp_est we set a minimum of
 +         * all cells interacting with all 3^3 direct neighbors (3^3-1)/2+1=14.
 +         * This might be a slight overestimate for small non-periodic groups of
 +         * atoms as will occur for a local domain with DD, but for small
 +         * groups of atoms we'll anyhow be limited by nsubpair_target_min,
 +         * so this overestimation will not matter.
 +         */
 +        nsp_est = std::max(nsp_est, grid->nsubc_tot*static_cast<real>(14));
 +
 +        if (debug)
 +        {
 +            fprintf(debug, "nsp_est local %5.1f non-local %5.1f\n",
 +                    nsp_est, nsp_est_nl);
 +        }
 +    }
 +    else
 +    {
 +        nsp_est = nsp_est_nl;
 +    }
 +
 +    /* Thus the (average) maximum j-list size should be as follows.
 +     * Since there is overhead, we shouldn't make the lists too small
 +     * (and we can't chop up j-groups) so we use a minimum target size of 36.
 +     */
 +    *nsubpair_target  = std::max(nsubpair_target_min,
 +                                 static_cast<int>(nsp_est/min_ci_balanced + 0.5));
 +    *nsubpair_tot_est = static_cast<int>(nsp_est);
 +
 +    if (debug)
 +    {
 +        fprintf(debug, "nbl nsp estimate %.1f, nsubpair_target %d\n",
 +                nsp_est, *nsubpair_target);
 +    }
 +}
 +
 +/* Debug list print function */
 +static void print_nblist_ci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
 +{
 +    for (int i = 0; i < nbl->nci; i++)
 +    {
 +        fprintf(fp, "ci %4d  shift %2d  ncj %3d\n",
 +                nbl->ci[i].ci, nbl->ci[i].shift,
 +                nbl->ci[i].cj_ind_end - nbl->ci[i].cj_ind_start);
 +
 +        for (int j = nbl->ci[i].cj_ind_start; j < nbl->ci[i].cj_ind_end; j++)
 +        {
 +            fprintf(fp, "  cj %5d  imask %x\n",
 +                    nbl->cj[j].cj,
 +                    nbl->cj[j].excl);
 +        }
 +    }
 +}
 +
 +/* Debug list print function */
 +static void print_nblist_sci_cj(FILE *fp, const nbnxn_pairlist_t *nbl)
 +{
 +    for (int i = 0; i < nbl->nsci; i++)
 +    {
 +        fprintf(fp, "ci %4d  shift %2d  ncj4 %2d\n",
 +                nbl->sci[i].sci, nbl->sci[i].shift,
 +                nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start);
 +
 +        int ncp = 0;
 +        for (int j4 = nbl->sci[i].cj4_ind_start; j4 < nbl->sci[i].cj4_ind_end; j4++)
 +        {
 +            for (int j = 0; j < c_nbnxnGpuJgroupSize; j++)
 +            {
 +                fprintf(fp, "  sj %5d  imask %x\n",
 +                        nbl->cj4[j4].cj[j],
 +                        nbl->cj4[j4].imei[0].imask);
 +                for (int si = 0; si < c_gpuNumClusterPerCell; si++)
 +                {
 +                    if (nbl->cj4[j4].imei[0].imask & (1U << (j*c_gpuNumClusterPerCell + si)))
 +                    {
 +                        ncp++;
 +                    }
 +                }
 +            }
 +        }
 +        fprintf(fp, "ci %4d  shift %2d  ncj4 %2d ncp %3d\n",
 +                nbl->sci[i].sci, nbl->sci[i].shift,
 +                nbl->sci[i].cj4_ind_end - nbl->sci[i].cj4_ind_start,
 +                ncp);
 +    }
 +}
 +
 +/* Combine pair lists *nbl generated on multiple threads nblc */
 +static void combine_nblists(int nnbl, nbnxn_pairlist_t **nbl,
 +                            nbnxn_pairlist_t *nblc)
 +{
 +    int nsci, ncj4, nexcl;
 +
 +    if (nblc->bSimple)
 +    {
 +        gmx_incons("combine_nblists does not support simple lists");
 +    }
 +
 +    nsci  = nblc->nsci;
 +    ncj4  = nblc->ncj4;
 +    nexcl = nblc->nexcl;
 +    for (int i = 0; i < nnbl; i++)
 +    {
 +        nsci  += nbl[i]->nsci;
 +        ncj4  += nbl[i]->ncj4;
 +        nexcl += nbl[i]->nexcl;
 +    }
 +
 +    if (nsci > nblc->sci_nalloc)
 +    {
 +        nb_realloc_sci(nblc, nsci);
 +    }
 +    if (ncj4 > nblc->cj4_nalloc)
 +    {
 +        nblc->cj4_nalloc = over_alloc_small(ncj4);
 +        nbnxn_realloc_void((void **)&nblc->cj4,
 +                           nblc->ncj4*sizeof(*nblc->cj4),
 +                           nblc->cj4_nalloc*sizeof(*nblc->cj4),
 +                           nblc->alloc, nblc->free);
 +    }
 +    if (nexcl > nblc->excl_nalloc)
 +    {
 +        nblc->excl_nalloc = over_alloc_small(nexcl);
 +        nbnxn_realloc_void((void **)&nblc->excl,
 +                           nblc->nexcl*sizeof(*nblc->excl),
 +                           nblc->excl_nalloc*sizeof(*nblc->excl),
 +                           nblc->alloc, nblc->free);
 +    }
 +
 +    /* Each thread should copy its own data to the combined arrays,
 +     * as otherwise data will go back and forth between different caches.
 +     */
 +#if GMX_OPENMP && !(defined __clang_analyzer__)
 +    // cppcheck-suppress unreadVariable
 +    int nthreads = gmx_omp_nthreads_get(emntPairsearch);
 +#endif
 +
 +#pragma omp parallel for num_threads(nthreads) schedule(static)
 +    for (int n = 0; n < nnbl; n++)
 +    {
 +        try
 +        {
 +            int                     sci_offset;
 +            int                     cj4_offset;
 +            int                     excl_offset;
 +            const nbnxn_pairlist_t *nbli;
 +
 +            /* Determine the offset in the combined data for our thread */
 +            sci_offset  = nblc->nsci;
 +            cj4_offset  = nblc->ncj4;
 +            excl_offset = nblc->nexcl;
 +
 +            for (int i = 0; i < n; i++)
 +            {
 +                sci_offset  += nbl[i]->nsci;
 +                cj4_offset  += nbl[i]->ncj4;
 +                excl_offset += nbl[i]->nexcl;
 +            }
 +
 +            nbli = nbl[n];
 +
 +            for (int i = 0; i < nbli->nsci; i++)
 +            {
 +                nblc->sci[sci_offset+i]                = nbli->sci[i];
 +                nblc->sci[sci_offset+i].cj4_ind_start += cj4_offset;
 +                nblc->sci[sci_offset+i].cj4_ind_end   += cj4_offset;
 +            }
 +
 +            for (int j4 = 0; j4 < nbli->ncj4; j4++)
 +            {
 +                nblc->cj4[cj4_offset+j4]                   = nbli->cj4[j4];
 +                nblc->cj4[cj4_offset+j4].imei[0].excl_ind += excl_offset;
 +                nblc->cj4[cj4_offset+j4].imei[1].excl_ind += excl_offset;
 +            }
 +
 +            for (int j4 = 0; j4 < nbli->nexcl; j4++)
 +            {
 +                nblc->excl[excl_offset+j4] = nbli->excl[j4];
 +            }
 +        }
 +        GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 +    }
 +
 +    for (int n = 0; n < nnbl; n++)
 +    {
 +        nblc->nsci    += nbl[n]->nsci;
 +        nblc->ncj4    += nbl[n]->ncj4;
 +        nblc->nci_tot += nbl[n]->nci_tot;
 +        nblc->nexcl   += nbl[n]->nexcl;
 +    }
 +}
 +
 +static void balance_fep_lists(const nbnxn_search_t  nbs,
 +                              nbnxn_pairlist_set_t *nbl_lists)
 +{
 +    int       nnbl;
 +    int       nri_tot, nrj_tot, nrj_target;
 +    int       th_dest;
 +    t_nblist *nbld;
 +
 +    nnbl = nbl_lists->nnbl;
 +
 +    if (nnbl == 1)
 +    {
 +        /* Nothing to balance */
 +        return;
 +    }
 +
 +    /* Count the total i-lists and pairs */
 +    nri_tot = 0;
 +    nrj_tot = 0;
 +    for (int th = 0; th < nnbl; th++)
 +    {
 +        nri_tot += nbl_lists->nbl_fep[th]->nri;
 +        nrj_tot += nbl_lists->nbl_fep[th]->nrj;
 +    }
 +
 +    nrj_target = (nrj_tot + nnbl - 1)/nnbl;
 +
 +    assert(gmx_omp_nthreads_get(emntNonbonded) == nnbl);
 +
 +#pragma omp parallel for schedule(static) num_threads(nnbl)
 +    for (int th = 0; th < nnbl; th++)
 +    {
 +        try
 +        {
 +            t_nblist *nbl;
 +
 +            nbl = nbs->work[th].nbl_fep;
 +
 +            /* Note that here we allocate for the total size, instead of
 +             * a per-thread esimate (which is hard to obtain).
 +             */
 +            if (nri_tot > nbl->maxnri)
 +            {
 +                nbl->maxnri = over_alloc_large(nri_tot);
 +                reallocate_nblist(nbl);
 +            }
 +            if (nri_tot > nbl->maxnri || nrj_tot > nbl->maxnrj)
 +            {
 +                nbl->maxnrj = over_alloc_small(nrj_tot);
 +                srenew(nbl->jjnr, nbl->maxnrj);
 +                srenew(nbl->excl_fep, nbl->maxnrj);
 +            }
 +
 +            clear_pairlist_fep(nbl);
 +        }
 +        GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 +    }
 +
 +    /* Loop over the source lists and assign and copy i-entries */
 +    th_dest = 0;
 +    nbld    = nbs->work[th_dest].nbl_fep;
 +    for (int th = 0; th < nnbl; th++)
 +    {
 +        t_nblist *nbls;
 +
 +        nbls = nbl_lists->nbl_fep[th];
 +
 +        for (int i = 0; i < nbls->nri; i++)
 +        {
 +            int nrj;
 +
 +            /* The number of pairs in this i-entry */
 +            nrj = nbls->jindex[i+1] - nbls->jindex[i];
 +
 +            /* Decide if list th_dest is too large and we should procede
 +             * to the next destination list.
 +             */
 +            if (th_dest+1 < nnbl && nbld->nrj > 0 &&
 +                nbld->nrj + nrj - nrj_target > nrj_target - nbld->nrj)
 +            {
 +                th_dest++;
 +                nbld = nbs->work[th_dest].nbl_fep;
 +            }
 +
 +            nbld->iinr[nbld->nri]  = nbls->iinr[i];
 +            nbld->gid[nbld->nri]   = nbls->gid[i];
 +            nbld->shift[nbld->nri] = nbls->shift[i];
 +
 +            for (int j = nbls->jindex[i]; j < nbls->jindex[i+1]; j++)
 +            {
 +                nbld->jjnr[nbld->nrj]     = nbls->jjnr[j];
 +                nbld->excl_fep[nbld->nrj] = nbls->excl_fep[j];
 +                nbld->nrj++;
 +            }
 +            nbld->nri++;
 +            nbld->jindex[nbld->nri] = nbld->nrj;
 +        }
 +    }
 +
 +    /* Swap the list pointers */
 +    for (int th = 0; th < nnbl; th++)
 +    {
 +        t_nblist *nbl_tmp;
 +
 +        nbl_tmp                = nbl_lists->nbl_fep[th];
 +        nbl_lists->nbl_fep[th] = nbs->work[th].nbl_fep;
 +        nbs->work[th].nbl_fep  = nbl_tmp;
 +
 +        if (debug)
 +        {
 +            fprintf(debug, "nbl_fep[%d] nri %4d nrj %4d\n",
 +                    th,
 +                    nbl_lists->nbl_fep[th]->nri,
 +                    nbl_lists->nbl_fep[th]->nrj);
 +        }
 +    }
 +}
 +
 +/* Returns the next ci to be processes by our thread */
 +static gmx_bool next_ci(const nbnxn_grid_t *grid,
 +                        int conv,
 +                        int nth, int ci_block,
 +                        int *ci_x, int *ci_y,
 +                        int *ci_b, int *ci)
 +{
 +    (*ci_b)++;
 +    (*ci)++;
 +
 +    if (*ci_b == ci_block)
 +    {
 +        /* Jump to the next block assigned to this task */
 +        *ci   += (nth - 1)*ci_block;
 +        *ci_b  = 0;
 +    }
 +
 +    if (*ci >= grid->nc*conv)
 +    {
 +        return FALSE;
 +    }
 +
 +    while (*ci >= grid->cxy_ind[*ci_x*grid->ncy + *ci_y + 1]*conv)
 +    {
 +        *ci_y += 1;
 +        if (*ci_y == grid->ncy)
 +        {
 +            *ci_x += 1;
 +            *ci_y  = 0;
 +        }
 +    }
 +
 +    return TRUE;
 +}
 +
 +/* Returns the distance^2 for which we put cell pairs in the list
 + * without checking atom pair distances. This is usually < rlist^2.
 + */
 +static float boundingbox_only_distance2(const nbnxn_grid_t *gridi,
 +                                        const nbnxn_grid_t *gridj,
 +                                        real                rlist,
 +                                        gmx_bool            simple)
 +{
 +    /* If the distance between two sub-cell bounding boxes is less
 +     * than this distance, do not check the distance between
 +     * all particle pairs in the sub-cell, since then it is likely
 +     * that the box pair has atom pairs within the cut-off.
 +     * We use the nblist cut-off minus 0.5 times the average x/y diagonal
 +     * spacing of the sub-cells. Around 40% of the checked pairs are pruned.
 +     * Using more than 0.5 gains at most 0.5%.
 +     * If forces are calculated more than twice, the performance gain
 +     * in the force calculation outweighs the cost of checking.
 +     * Note that with subcell lists, the atom-pair distance check
 +     * is only performed when only 1 out of 8 sub-cells in within range,
 +     * this is because the GPU is much faster than the cpu.
 +     */
 +    real bbx, bby;
 +    real rbb2;
 +
 +    bbx = 0.5*(gridi->sx + gridj->sx);
 +    bby = 0.5*(gridi->sy + gridj->sy);
 +    if (!simple)
 +    {
 +        bbx /= c_gpuNumClusterPerCellX;
 +        bby /= c_gpuNumClusterPerCellY;
 +    }
 +
 +    rbb2 = std::max(0.0, rlist - 0.5*std::sqrt(bbx*bbx + bby*bby));
 +    rbb2 = rbb2 * rbb2;
 +
 +#if !GMX_DOUBLE
 +    return rbb2;
 +#else
 +    return (float)((1+GMX_FLOAT_EPS)*rbb2);
 +#endif
 +}
 +
 +static int get_ci_block_size(const nbnxn_grid_t *gridi,
 +                             gmx_bool bDomDec, int nth)
 +{
 +    const int ci_block_enum      = 5;
 +    const int ci_block_denom     = 11;
 +    const int ci_block_min_atoms = 16;
 +    int       ci_block;
 +
 +    /* Here we decide how to distribute the blocks over the threads.
 +     * We use prime numbers to try to avoid that the grid size becomes
 +     * a multiple of the number of threads, which would lead to some
 +     * threads getting "inner" pairs and others getting boundary pairs,
 +     * which in turns will lead to load imbalance between threads.
 +     * Set the block size as 5/11/ntask times the average number of cells
 +     * in a y,z slab. This should ensure a quite uniform distribution
 +     * of the grid parts of the different thread along all three grid
 +     * zone boundaries with 3D domain decomposition. At the same time
 +     * the blocks will not become too small.
 +     */
 +    ci_block = (gridi->nc*ci_block_enum)/(ci_block_denom*gridi->ncx*nth);
 +
 +    /* Ensure the blocks are not too small: avoids cache invalidation */
 +    if (ci_block*gridi->na_sc < ci_block_min_atoms)
 +    {
 +        ci_block = (ci_block_min_atoms + gridi->na_sc - 1)/gridi->na_sc;
 +    }
 +
 +    /* Without domain decomposition
 +     * or with less than 3 blocks per task, divide in nth blocks.
 +     */
 +    if (!bDomDec || nth*3*ci_block > gridi->nc)
 +    {
 +        ci_block = (gridi->nc + nth - 1)/nth;
 +    }
 +
 +    if (ci_block > 1 && (nth - 1)*ci_block >= gridi->nc)
 +    {
 +        /* Some threads have no work. Although reducing the block size
 +         * does not decrease the block count on the first few threads,
 +         * with GPUs better mixing of "upper" cells that have more empty
 +         * clusters results in a somewhat lower max load over all threads.
 +         * Without GPUs the regime of so few atoms per thread is less
 +         * performance relevant, but with 8-wide SIMD the same reasoning
 +         * applies, since the pair list uses 4 i-atom "sub-clusters".
 +         */
 +        ci_block--;
 +    }
 +
 +    return ci_block;
 +}
 +
 +/* Generates the part of pair-list nbl assigned to our thread */
 +static void nbnxn_make_pairlist_part(const nbnxn_search_t nbs,
 +                                     const nbnxn_grid_t *gridi,
 +                                     const nbnxn_grid_t *gridj,
 +                                     nbnxn_search_work_t *work,
 +                                     const nbnxn_atomdata_t *nbat,
 +                                     const t_blocka *excl,
 +                                     real rlist,
 +                                     int nb_kernel_type,
 +                                     int ci_block,
 +                                     gmx_bool bFBufferFlag,
 +                                     int nsubpair_max,
 +                                     gmx_bool progBal,
-                                      int nsubpair_tot_est,
++                                     float nsubpair_tot_est,
 +                                     int th, int nth,
 +                                     nbnxn_pairlist_t *nbl,
 +                                     t_nblist *nbl_fep)
 +{
 +    int               na_cj_2log;
 +    matrix            box;
 +    real              rl2, rl_fep2 = 0;
 +    float             rbb2;
 +    int               ci_b, ci, ci_x, ci_y, ci_xy, cj;
 +    ivec              shp;
 +    int               shift;
 +    real              shx, shy, shz;
 +    int               conv_i, cell0_i;
 +    const nbnxn_bb_t *bb_i = NULL;
 +#if NBNXN_BBXXXX
 +    const float      *pbb_i = NULL;
 +#endif
 +    const float      *bbcz_i, *bbcz_j;
 +    const int        *flags_i;
 +    real              bx0, bx1, by0, by1, bz0, bz1;
 +    real              bz1_frac;
 +    real              d2cx, d2z, d2z_cx, d2z_cy, d2zx, d2zxy, d2xy;
 +    int               cxf, cxl, cyf, cyf_x, cyl;
 +    int               c0, c1, cs, cf, cl;
 +    int               ndistc;
 +    int               ncpcheck;
 +    int               gridi_flag_shift = 0, gridj_flag_shift = 0;
 +    gmx_bitmask_t    *gridj_flag       = NULL;
 +    int               ncj_old_i, ncj_old_j;
 +
 +    nbs_cycle_start(&work->cc[enbsCCsearch]);
 +
 +    if (gridj->bSimple != nbl->bSimple)
 +    {
 +        gmx_incons("Grid incompatible with pair-list");
 +    }
 +
 +    sync_work(nbl);
 +    nbl->na_sc = gridj->na_sc;
 +    nbl->na_ci = gridj->na_c;
 +    nbl->na_cj = nbnxn_kernel_to_cluster_j_size(nb_kernel_type);
 +    na_cj_2log = get_2log(nbl->na_cj);
 +
 +    nbl->rlist  = rlist;
 +
 +    if (bFBufferFlag)
 +    {
 +        /* Determine conversion of clusters to flag blocks */
 +        gridi_flag_shift = 0;
 +        while ((nbl->na_ci<<gridi_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
 +        {
 +            gridi_flag_shift++;
 +        }
 +        gridj_flag_shift = 0;
 +        while ((nbl->na_cj<<gridj_flag_shift) < NBNXN_BUFFERFLAG_SIZE)
 +        {
 +            gridj_flag_shift++;
 +        }
 +
 +        gridj_flag = work->buffer_flags.flag;
 +    }
 +
 +    copy_mat(nbs->box, box);
 +
 +    rl2 = nbl->rlist*nbl->rlist;
 +
 +    if (nbs->bFEP && !nbl->bSimple)
 +    {
 +        /* Determine an atom-pair list cut-off distance for FEP atom pairs.
 +         * We should not simply use rlist, since then we would not have
 +         * the small, effective buffering of the NxN lists.
 +         * The buffer is on overestimate, but the resulting cost for pairs
 +         * beyond rlist is neglible compared to the FEP pairs within rlist.
 +         */
 +        rl_fep2 = nbl->rlist + effective_buffer_1x1_vs_MxN(gridi, gridj);
 +
 +        if (debug)
 +        {
 +            fprintf(debug, "nbl_fep atom-pair rlist %f\n", rl_fep2);
 +        }
 +        rl_fep2 = rl_fep2*rl_fep2;
 +    }
 +
 +    rbb2 = boundingbox_only_distance2(gridi, gridj, nbl->rlist, nbl->bSimple);
 +
 +    if (debug)
 +    {
 +        fprintf(debug, "nbl bounding box only distance %f\n", std::sqrt(rbb2));
 +    }
 +
 +    /* Set the shift range */
 +    for (int d = 0; d < DIM; d++)
 +    {
 +        /* Check if we need periodicity shifts.
 +         * Without PBC or with domain decomposition we don't need them.
 +         */
 +        if (d >= ePBC2npbcdim(nbs->ePBC) || nbs->dd_dim[d])
 +        {
 +            shp[d] = 0;
 +        }
 +        else
 +        {
 +            if (d == XX &&
 +                box[XX][XX] - fabs(box[YY][XX]) - fabs(box[ZZ][XX]) < std::sqrt(rl2))
 +            {
 +                shp[d] = 2;
 +            }
 +            else
 +            {
 +                shp[d] = 1;
 +            }
 +        }
 +    }
 +
 +    if (nbl->bSimple && !gridi->bSimple)
 +    {
 +        conv_i  = gridi->na_sc/gridj->na_sc;
 +        bb_i    = gridi->bb_simple;
 +        bbcz_i  = gridi->bbcz_simple;
 +        flags_i = gridi->flags_simple;
 +    }
 +    else
 +    {
 +        conv_i  = 1;
 +#if NBNXN_BBXXXX
 +        if (gridi->bSimple)
 +        {
 +            bb_i  = gridi->bb;
 +        }
 +        else
 +        {
 +            pbb_i = gridi->pbb;
 +        }
 +#else
 +        /* We use the normal bounding box format for both grid types */
 +        bb_i  = gridi->bb;
 +#endif
 +        bbcz_i  = gridi->bbcz;
 +        flags_i = gridi->flags;
 +    }
 +    cell0_i = gridi->cell0*conv_i;
 +
 +    bbcz_j = gridj->bbcz;
 +
 +    if (conv_i != 1)
 +    {
 +        /* Blocks of the conversion factor - 1 give a large repeat count
 +         * combined with a small block size. This should result in good
 +         * load balancing for both small and large domains.
 +         */
 +        ci_block = conv_i - 1;
 +    }
 +    if (debug)
 +    {
 +        fprintf(debug, "nbl nc_i %d col.av. %.1f ci_block %d\n",
 +                gridi->nc, gridi->nc/(double)(gridi->ncx*gridi->ncy), ci_block);
 +    }
 +
 +    ndistc   = 0;
 +    ncpcheck = 0;
 +
 +    /* Initially ci_b and ci to 1 before where we want them to start,
 +     * as they will both be incremented in next_ci.
 +     */
 +    ci_b = -1;
 +    ci   = th*ci_block - 1;
 +    ci_x = 0;
 +    ci_y = 0;
 +    while (next_ci(gridi, conv_i, nth, ci_block, &ci_x, &ci_y, &ci_b, &ci))
 +    {
 +        if (nbl->bSimple && flags_i[ci] == 0)
 +        {
 +            continue;
 +        }
 +
 +        ncj_old_i = nbl->ncj;
 +
 +        d2cx = 0;
 +        if (gridj != gridi && shp[XX] == 0)
 +        {
 +            if (nbl->bSimple)
 +            {
 +                bx1 = bb_i[ci].upper[BB_X];
 +            }
 +            else
 +            {
 +                bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx;
 +            }
 +            if (bx1 < gridj->c0[XX])
 +            {
 +                d2cx = gmx::square(gridj->c0[XX] - bx1);
 +
 +                if (d2cx >= rl2)
 +                {
 +                    continue;
 +                }
 +            }
 +        }
 +
 +        ci_xy = ci_x*gridi->ncy + ci_y;
 +
 +        /* Loop over shift vectors in three dimensions */
 +        for (int tz = -shp[ZZ]; tz <= shp[ZZ]; tz++)
 +        {
 +            shz = tz*box[ZZ][ZZ];
 +
 +            bz0 = bbcz_i[ci*NNBSBB_D  ] + shz;
 +            bz1 = bbcz_i[ci*NNBSBB_D+1] + shz;
 +
 +            if (tz == 0)
 +            {
 +                d2z = 0;
 +            }
 +            else if (tz < 0)
 +            {
 +                d2z = gmx::square(bz1);
 +            }
 +            else
 +            {
 +                d2z = gmx::square(bz0 - box[ZZ][ZZ]);
 +            }
 +
 +            d2z_cx = d2z + d2cx;
 +
 +            if (d2z_cx >= rl2)
 +            {
 +                continue;
 +            }
 +
 +            bz1_frac = bz1/(gridi->cxy_ind[ci_xy+1] - gridi->cxy_ind[ci_xy]);
 +            if (bz1_frac < 0)
 +            {
 +                bz1_frac = 0;
 +            }
 +            /* The check with bz1_frac close to or larger than 1 comes later */
 +
 +            for (int ty = -shp[YY]; ty <= shp[YY]; ty++)
 +            {
 +                shy = ty*box[YY][YY] + tz*box[ZZ][YY];
 +
 +                if (nbl->bSimple)
 +                {
 +                    by0 = bb_i[ci].lower[BB_Y] + shy;
 +                    by1 = bb_i[ci].upper[BB_Y] + shy;
 +                }
 +                else
 +                {
 +                    by0 = gridi->c0[YY] + (ci_y  )*gridi->sy + shy;
 +                    by1 = gridi->c0[YY] + (ci_y+1)*gridi->sy + shy;
 +                }
 +
 +                get_cell_range(by0, by1,
 +                               gridj->ncy, gridj->c0[YY], gridj->sy, gridj->inv_sy,
 +                               d2z_cx, rl2,
 +                               &cyf, &cyl);
 +
 +                if (cyf > cyl)
 +                {
 +                    continue;
 +                }
 +
 +                d2z_cy = d2z;
 +                if (by1 < gridj->c0[YY])
 +                {
 +                    d2z_cy += gmx::square(gridj->c0[YY] - by1);
 +                }
 +                else if (by0 > gridj->c1[YY])
 +                {
 +                    d2z_cy += gmx::square(by0 - gridj->c1[YY]);
 +                }
 +
 +                for (int tx = -shp[XX]; tx <= shp[XX]; tx++)
 +                {
 +                    shift = XYZ2IS(tx, ty, tz);
 +
 +                    if (pbc_shift_backward && gridi == gridj && shift > CENTRAL)
 +                    {
 +                        continue;
 +                    }
 +
 +                    shx = tx*box[XX][XX] + ty*box[YY][XX] + tz*box[ZZ][XX];
 +
 +                    if (nbl->bSimple)
 +                    {
 +                        bx0 = bb_i[ci].lower[BB_X] + shx;
 +                        bx1 = bb_i[ci].upper[BB_X] + shx;
 +                    }
 +                    else
 +                    {
 +                        bx0 = gridi->c0[XX] + (ci_x  )*gridi->sx + shx;
 +                        bx1 = gridi->c0[XX] + (ci_x+1)*gridi->sx + shx;
 +                    }
 +
 +                    get_cell_range(bx0, bx1,
 +                                   gridj->ncx, gridj->c0[XX], gridj->sx, gridj->inv_sx,
 +                                   d2z_cy, rl2,
 +                                   &cxf, &cxl);
 +
 +                    if (cxf > cxl)
 +                    {
 +                        continue;
 +                    }
 +
 +                    if (nbl->bSimple)
 +                    {
 +                        new_ci_entry(nbl, cell0_i+ci, shift, flags_i[ci]);
 +                    }
 +                    else
 +                    {
 +                        new_sci_entry(nbl, cell0_i+ci, shift);
 +                    }
 +
 +                    if ((!pbc_shift_backward || (shift == CENTRAL &&
 +                                                 gridi == gridj)) &&
 +                        cxf < ci_x)
 +                    {
 +                        /* Leave the pairs with i > j.
 +                         * x is the major index, so skip half of it.
 +                         */
 +                        cxf = ci_x;
 +                    }
 +
 +                    if (nbl->bSimple)
 +                    {
 +                        set_icell_bb_simple(bb_i, ci, shx, shy, shz,
 +                                            nbl->work->bb_ci);
 +                    }
 +                    else
 +                    {
 +#if NBNXN_BBXXXX
 +                        set_icell_bbxxxx_supersub(pbb_i, ci, shx, shy, shz,
 +                                                  nbl->work->pbb_ci);
 +#else
 +                        set_icell_bb_supersub(bb_i, ci, shx, shy, shz,
 +                                              nbl->work->bb_ci);
 +#endif
 +                    }
 +
 +                    nbs->icell_set_x(cell0_i+ci, shx, shy, shz,
 +                                     nbat->xstride, nbat->x,
 +                                     nbl->work);
 +
 +                    for (int cx = cxf; cx <= cxl; cx++)
 +                    {
 +                        d2zx = d2z;
 +                        if (gridj->c0[XX] + cx*gridj->sx > bx1)
 +                        {
 +                            d2zx += gmx::square(gridj->c0[XX] + cx*gridj->sx - bx1);
 +                        }
 +                        else if (gridj->c0[XX] + (cx+1)*gridj->sx < bx0)
 +                        {
 +                            d2zx += gmx::square(gridj->c0[XX] + (cx+1)*gridj->sx - bx0);
 +                        }
 +
 +                        if (gridi == gridj &&
 +                            cx == 0 &&
 +                            (!pbc_shift_backward || shift == CENTRAL) &&
 +                            cyf < ci_y)
 +                        {
 +                            /* Leave the pairs with i > j.
 +                             * Skip half of y when i and j have the same x.
 +                             */
 +                            cyf_x = ci_y;
 +                        }
 +                        else
 +                        {
 +                            cyf_x = cyf;
 +                        }
 +
 +                        for (int cy = cyf_x; cy <= cyl; cy++)
 +                        {
 +                            c0 = gridj->cxy_ind[cx*gridj->ncy+cy];
 +                            c1 = gridj->cxy_ind[cx*gridj->ncy+cy+1];
 +
 +                            if (pbc_shift_backward &&
 +                                gridi == gridj &&
 +                                shift == CENTRAL && c0 < ci)
 +                            {
 +                                c0 = ci;
 +                            }
 +
 +                            d2zxy = d2zx;
 +                            if (gridj->c0[YY] + cy*gridj->sy > by1)
 +                            {
 +                                d2zxy += gmx::square(gridj->c0[YY] + cy*gridj->sy - by1);
 +                            }
 +                            else if (gridj->c0[YY] + (cy+1)*gridj->sy < by0)
 +                            {
 +                                d2zxy += gmx::square(gridj->c0[YY] + (cy+1)*gridj->sy - by0);
 +                            }
 +                            if (c1 > c0 && d2zxy < rl2)
 +                            {
 +                                cs = c0 + static_cast<int>(bz1_frac*(c1 - c0));
 +                                if (cs >= c1)
 +                                {
 +                                    cs = c1 - 1;
 +                                }
 +
 +                                d2xy = d2zxy - d2z;
 +
 +                                /* Find the lowest cell that can possibly
 +                                 * be within range.
 +                                 */
 +                                cf = cs;
 +                                while (cf > c0 &&
 +                                       (bbcz_j[cf*NNBSBB_D+1] >= bz0 ||
 +                                        d2xy + gmx::square(bbcz_j[cf*NNBSBB_D+1] - bz0) < rl2))
 +                                {
 +                                    cf--;
 +                                }
 +
 +                                /* Find the highest cell that can possibly
 +                                 * be within range.
 +                                 */
 +                                cl = cs;
 +                                while (cl < c1-1 &&
 +                                       (bbcz_j[cl*NNBSBB_D] <= bz1 ||
 +                                        d2xy + gmx::square(bbcz_j[cl*NNBSBB_D] - bz1) < rl2))
 +                                {
 +                                    cl++;
 +                                }
 +
 +#ifdef NBNXN_REFCODE
 +                                {
 +                                    /* Simple reference code, for debugging,
 +                                     * overrides the more complex code above.
 +                                     */
 +                                    cf = c1;
 +                                    cl = -1;
 +                                    for (int k = c0; k < c1; k++)
 +                                    {
 +                                        if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
 +                                            k < cf)
 +                                        {
 +                                            cf = k;
 +                                        }
 +                                        if (box_dist2(bx0, bx1, by0, by1, bz0, bz1, bb+k) < rl2 &&
 +                                            k > cl)
 +                                        {
 +                                            cl = k;
 +                                        }
 +                                    }
 +                                }
 +#endif
 +
 +                                if (gridi == gridj)
 +                                {
 +                                    /* We want each atom/cell pair only once,
 +                                     * only use cj >= ci.
 +                                     */
 +                                    if (!pbc_shift_backward || shift == CENTRAL)
 +                                    {
 +                                        cf = std::max(cf, ci);
 +                                    }
 +                                }
 +
 +                                if (cf <= cl)
 +                                {
 +                                    /* For f buffer flags with simple lists */
 +                                    ncj_old_j = nbl->ncj;
 +
 +                                    switch (nb_kernel_type)
 +                                    {
 +                                        case nbnxnk4x4_PlainC:
 +                                            check_cell_list_space_simple(nbl, cl-cf+1);
 +
 +                                            make_cluster_list_simple(gridj,
 +                                                                     nbl, ci, cf, cl,
 +                                                                     (gridi == gridj && shift == CENTRAL),
 +                                                                     nbat->x,
 +                                                                     rl2, rbb2,
 +                                                                     &ndistc);
 +                                            break;
 +#ifdef GMX_NBNXN_SIMD_4XN
 +                                        case nbnxnk4xN_SIMD_4xN:
 +                                            check_cell_list_space_simple(nbl, ci_to_cj_simd_4xn(cl - cf) + 2);
 +                                            make_cluster_list_simd_4xn(gridj,
 +                                                                       nbl, ci, cf, cl,
 +                                                                       (gridi == gridj && shift == CENTRAL),
 +                                                                       nbat->x,
 +                                                                       rl2, rbb2,
 +                                                                       &ndistc);
 +                                            break;
 +#endif
 +#ifdef GMX_NBNXN_SIMD_2XNN
 +                                        case nbnxnk4xN_SIMD_2xNN:
 +                                            check_cell_list_space_simple(nbl, ci_to_cj_simd_2xnn(cl - cf) + 2);
 +                                            make_cluster_list_simd_2xnn(gridj,
 +                                                                        nbl, ci, cf, cl,
 +                                                                        (gridi == gridj && shift == CENTRAL),
 +                                                                        nbat->x,
 +                                                                        rl2, rbb2,
 +                                                                        &ndistc);
 +                                            break;
 +#endif
 +                                        case nbnxnk8x8x8_PlainC:
 +                                        case nbnxnk8x8x8_GPU:
 +                                            check_cell_list_space_supersub(nbl, cl-cf+1);
 +                                            for (cj = cf; cj <= cl; cj++)
 +                                            {
 +                                                make_cluster_list_supersub(gridi, gridj,
 +                                                                           nbl, ci, cj,
 +                                                                           (gridi == gridj && shift == CENTRAL && ci == cj),
 +                                                                           nbat->xstride, nbat->x,
 +                                                                           rl2, rbb2,
 +                                                                           &ndistc);
 +                                            }
 +                                            break;
 +                                    }
 +                                    ncpcheck += cl - cf + 1;
 +
 +                                    if (bFBufferFlag && nbl->ncj > ncj_old_j)
 +                                    {
 +                                        int cbf = nbl->cj[ncj_old_j].cj >> gridj_flag_shift;
 +                                        int cbl = nbl->cj[nbl->ncj-1].cj >> gridj_flag_shift;
 +                                        for (int cb = cbf; cb <= cbl; cb++)
 +                                        {
 +                                            bitmask_init_bit(&gridj_flag[cb], th);
 +                                        }
 +                                    }
 +                                }
 +                            }
 +                        }
 +                    }
 +
 +                    /* Set the exclusions for this ci list */
 +                    if (nbl->bSimple)
 +                    {
 +                        set_ci_top_excls(nbs,
 +                                         nbl,
 +                                         shift == CENTRAL && gridi == gridj,
 +                                         gridj->na_c_2log,
 +                                         na_cj_2log,
 +                                         &(nbl->ci[nbl->nci]),
 +                                         excl);
 +
 +                        if (nbs->bFEP)
 +                        {
 +                            make_fep_list(nbs, nbat, nbl,
 +                                          shift == CENTRAL && gridi == gridj,
 +                                          &(nbl->ci[nbl->nci]),
 +                                          gridi, gridj, nbl_fep);
 +                        }
 +                    }
 +                    else
 +                    {
 +                        set_sci_top_excls(nbs,
 +                                          nbl,
 +                                          shift == CENTRAL && gridi == gridj,
 +                                          gridj->na_c_2log,
 +                                          &(nbl->sci[nbl->nsci]),
 +                                          excl);
 +
 +                        if (nbs->bFEP)
 +                        {
 +                            make_fep_list_supersub(nbs, nbat, nbl,
 +                                                   shift == CENTRAL && gridi == gridj,
 +                                                   &(nbl->sci[nbl->nsci]),
 +                                                   shx, shy, shz,
 +                                                   rl_fep2,
 +                                                   gridi, gridj, nbl_fep);
 +                        }
 +                    }
 +
 +                    /* Close this ci list */
 +                    if (nbl->bSimple)
 +                    {
 +                        close_ci_entry_simple(nbl);
 +                    }
 +                    else
 +                    {
 +                        close_ci_entry_supersub(nbl,
 +                                                nsubpair_max,
 +                                                progBal, nsubpair_tot_est,
 +                                                th, nth);
 +                    }
 +                }
 +            }
 +        }
 +
 +        if (bFBufferFlag && nbl->ncj > ncj_old_i)
 +        {
 +            bitmask_init_bit(&(work->buffer_flags.flag[(gridi->cell0+ci)>>gridi_flag_shift]), th);
 +        }
 +    }
 +
 +    work->ndistc = ndistc;
 +
 +    nbs_cycle_stop(&work->cc[enbsCCsearch]);
 +
 +    if (debug)
 +    {
 +        fprintf(debug, "number of distance checks %d\n", ndistc);
 +        fprintf(debug, "ncpcheck %s %d\n", gridi == gridj ? "local" : "non-local",
 +                ncpcheck);
 +
 +        if (nbl->bSimple)
 +        {
 +            print_nblist_statistics_simple(debug, nbl, nbs, rlist);
 +        }
 +        else
 +        {
 +            print_nblist_statistics_supersub(debug, nbl, nbs, rlist);
 +        }
 +
 +        if (nbs->bFEP)
 +        {
 +            fprintf(debug, "nbl FEP list pairs: %d\n", nbl_fep->nrj);
 +        }
 +    }
 +}
 +
 +static void reduce_buffer_flags(const nbnxn_search_t        nbs,
 +                                int                         nsrc,
 +                                const nbnxn_buffer_flags_t *dest)
 +{
 +    for (int s = 0; s < nsrc; s++)
 +    {
 +        gmx_bitmask_t * flag = nbs->work[s].buffer_flags.flag;
 +
 +        for (int b = 0; b < dest->nflag; b++)
 +        {
 +            bitmask_union(&(dest->flag[b]), flag[b]);
 +        }
 +    }
 +}
 +
 +static void print_reduction_cost(const nbnxn_buffer_flags_t *flags, int nout)
 +{
 +    int           nelem, nkeep, ncopy, nred, out;
 +    gmx_bitmask_t mask_0;
 +
 +    nelem = 0;
 +    nkeep = 0;
 +    ncopy = 0;
 +    nred  = 0;
 +    bitmask_init_bit(&mask_0, 0);
 +    for (int b = 0; b < flags->nflag; b++)
 +    {
 +        if (bitmask_is_equal(flags->flag[b], mask_0))
 +        {
 +            /* Only flag 0 is set, no copy of reduction required */
 +            nelem++;
 +            nkeep++;
 +        }
 +        else if (!bitmask_is_zero(flags->flag[b]))
 +        {
 +            int c = 0;
 +            for (out = 0; out < nout; out++)
 +            {
 +                if (bitmask_is_set(flags->flag[b], out))
 +                {
 +                    c++;
 +                }
 +            }
 +            nelem += c;
 +            if (c == 1)
 +            {
 +                ncopy++;
 +            }
 +            else
 +            {
 +                nred += c;
 +            }
 +        }
 +    }
 +
 +    fprintf(debug, "nbnxn reduction: #flag %d #list %d elem %4.2f, keep %4.2f copy %4.2f red %4.2f\n",
 +            flags->nflag, nout,
 +            nelem/(double)(flags->nflag),
 +            nkeep/(double)(flags->nflag),
 +            ncopy/(double)(flags->nflag),
 +            nred/(double)(flags->nflag));
 +}
 +
 +/* Perform a count (linear) sort to sort the smaller lists to the end.
 + * This avoids load imbalance on the GPU, as large lists will be
 + * scheduled and executed first and the smaller lists later.
 + * Load balancing between multi-processors only happens at the end
 + * and there smaller lists lead to more effective load balancing.
 + * The sorting is done on the cj4 count, not on the actual pair counts.
 + * Not only does this make the sort faster, but it also results in
 + * better load balancing than using a list sorted on exact load.
 + * This function swaps the pointer in the pair list to avoid a copy operation.
 + */
 +static void sort_sci(nbnxn_pairlist_t *nbl)
 +{
 +    nbnxn_list_work_t *work;
 +    int                m, s0, s1;
 +    nbnxn_sci_t       *sci_sort;
 +
 +    if (nbl->ncj4 <= nbl->nsci)
 +    {
 +        /* nsci = 0 or all sci have size 1, sorting won't change the order */
 +        return;
 +    }
 +
 +    work = nbl->work;
 +
 +    /* We will distinguish differences up to double the average */
 +    m = (2*nbl->ncj4)/nbl->nsci;
 +
 +    if (m + 1 > work->sort_nalloc)
 +    {
 +        work->sort_nalloc = over_alloc_large(m + 1);
 +        srenew(work->sort, work->sort_nalloc);
 +    }
 +
 +    if (work->sci_sort_nalloc != nbl->sci_nalloc)
 +    {
 +        work->sci_sort_nalloc = nbl->sci_nalloc;
 +        nbnxn_realloc_void((void **)&work->sci_sort,
 +                           0,
 +                           work->sci_sort_nalloc*sizeof(*work->sci_sort),
 +                           nbl->alloc, nbl->free);
 +    }
 +
 +    /* Count the entries of each size */
 +    for (int i = 0; i <= m; i++)
 +    {
 +        work->sort[i] = 0;
 +    }
 +    for (int s = 0; s < nbl->nsci; s++)
 +    {
 +        int i = std::min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
 +        work->sort[i]++;
 +    }
 +    /* Calculate the offset for each count */
 +    s0            = work->sort[m];
 +    work->sort[m] = 0;
 +    for (int i = m - 1; i >= 0; i--)
 +    {
 +        s1            = work->sort[i];
 +        work->sort[i] = work->sort[i + 1] + s0;
 +        s0            = s1;
 +    }
 +
 +    /* Sort entries directly into place */
 +    sci_sort = work->sci_sort;
 +    for (int s = 0; s < nbl->nsci; s++)
 +    {
 +        int i = std::min(m, nbl->sci[s].cj4_ind_end - nbl->sci[s].cj4_ind_start);
 +        sci_sort[work->sort[i]++] = nbl->sci[s];
 +    }
 +
 +    /* Swap the sci pointers so we use the new, sorted list */
 +    work->sci_sort = nbl->sci;
 +    nbl->sci       = sci_sort;
 +}
 +
 +/* Make a local or non-local pair-list, depending on iloc */
 +void nbnxn_make_pairlist(const nbnxn_search_t  nbs,
 +                         nbnxn_atomdata_t     *nbat,
 +                         const t_blocka       *excl,
 +                         real                  rlist,
 +                         int                   min_ci_balanced,
 +                         nbnxn_pairlist_set_t *nbl_list,
 +                         int                   iloc,
 +                         int                   nb_kernel_type,
 +                         t_nrnb               *nrnb)
 +{
 +    nbnxn_grid_t      *gridi, *gridj;
 +    gmx_bool           bGPUCPU;
 +    int                nzi, zj0, zj1;
-     int                nsubpair_target, nsubpair_tot_est;
++    int                nsubpair_target;
++    float              nsubpair_tot_est;
 +    int                nnbl;
 +    nbnxn_pairlist_t **nbl;
 +    int                ci_block;
 +    gmx_bool           CombineNBLists;
 +    gmx_bool           progBal;
 +    int                np_tot, np_noq, np_hlj, nap;
 +
 +    /* Check if we are running hybrid GPU + CPU nbnxn mode */
 +    bGPUCPU = (!nbs->grid[0].bSimple && nbl_list->bSimple);
 +
 +    nnbl            = nbl_list->nnbl;
 +    nbl             = nbl_list->nbl;
 +    CombineNBLists  = nbl_list->bCombined;
 +
 +    if (debug)
 +    {
 +        fprintf(debug, "ns making %d nblists\n", nnbl);
 +    }
 +
 +    nbat->bUseBufferFlags = (nbat->nout > 1);
 +    /* We should re-init the flags before making the first list */
 +    if (nbat->bUseBufferFlags && (LOCAL_I(iloc) || bGPUCPU))
 +    {
 +        init_buffer_flags(&nbat->buffer_flags, nbat->natoms);
 +    }
 +
 +    if (nbl_list->bSimple)
 +    {
 +#if GMX_SIMD
 +        switch (nb_kernel_type)
 +        {
 +#ifdef GMX_NBNXN_SIMD_4XN
 +            case nbnxnk4xN_SIMD_4xN:
 +                nbs->icell_set_x = icell_set_x_simd_4xn;
 +                break;
 +#endif
 +#ifdef GMX_NBNXN_SIMD_2XNN
 +            case nbnxnk4xN_SIMD_2xNN:
 +                nbs->icell_set_x = icell_set_x_simd_2xnn;
 +                break;
 +#endif
 +            default:
 +                nbs->icell_set_x = icell_set_x_simple;
 +                break;
 +        }
 +#else   // GMX_SIMD
 +        /* MSVC 2013 complains about switch statements without case */
 +        nbs->icell_set_x = icell_set_x_simple;
 +#endif  // GMX_SIMD
 +    }
 +    else
 +    {
 +        nbs->icell_set_x = icell_set_x_supersub;
 +    }
 +
 +    if (LOCAL_I(iloc))
 +    {
 +        /* Only zone (grid) 0 vs 0 */
 +        nzi = 1;
 +        zj0 = 0;
 +        zj1 = 1;
 +    }
 +    else
 +    {
 +        nzi = nbs->zones->nizone;
 +    }
 +
 +    if (!nbl_list->bSimple && min_ci_balanced > 0)
 +    {
 +        get_nsubpair_target(nbs, iloc, rlist, min_ci_balanced,
 +                            &nsubpair_target, &nsubpair_tot_est);
 +    }
 +    else
 +    {
 +        nsubpair_target  = 0;
 +        nsubpair_tot_est = 0;
 +    }
 +
 +    /* Clear all pair-lists */
 +    for (int th = 0; th < nnbl; th++)
 +    {
 +        clear_pairlist(nbl[th]);
 +
 +        if (nbs->bFEP)
 +        {
 +            clear_pairlist_fep(nbl_list->nbl_fep[th]);
 +        }
 +    }
 +
 +    for (int zi = 0; zi < nzi; zi++)
 +    {
 +        gridi = &nbs->grid[zi];
 +
 +        if (NONLOCAL_I(iloc))
 +        {
 +            zj0 = nbs->zones->izone[zi].j0;
 +            zj1 = nbs->zones->izone[zi].j1;
 +            if (zi == 0)
 +            {
 +                zj0++;
 +            }
 +        }
 +        for (int zj = zj0; zj < zj1; zj++)
 +        {
 +            gridj = &nbs->grid[zj];
 +
 +            if (debug)
 +            {
 +                fprintf(debug, "ns search grid %d vs %d\n", zi, zj);
 +            }
 +
 +            nbs_cycle_start(&nbs->cc[enbsCCsearch]);
 +
 +            if (nbl[0]->bSimple && !gridi->bSimple)
 +            {
 +                /* Hybrid list, determine blocking later */
 +                ci_block = 0;
 +            }
 +            else
 +            {
 +                ci_block = get_ci_block_size(gridi, nbs->DomDec, nnbl);
 +            }
 +
 +            /* With GPU: generate progressively smaller lists for
 +             * load balancing for local only or non-local with 2 zones.
 +             */
 +            progBal = (LOCAL_I(iloc) || nbs->zones->n <= 2);
 +
 +#pragma omp parallel for num_threads(nnbl) schedule(static)
 +            for (int th = 0; th < nnbl; th++)
 +            {
 +                try
 +                {
 +                    /* Re-init the thread-local work flag data before making
 +                     * the first list (not an elegant conditional).
 +                     */
 +                    if (nbat->bUseBufferFlags && ((zi == 0 && zj == 0) ||
 +                                                  (bGPUCPU && zi == 0 && zj == 1)))
 +                    {
 +                        init_buffer_flags(&nbs->work[th].buffer_flags, nbat->natoms);
 +                    }
 +
 +                    if (CombineNBLists && th > 0)
 +                    {
 +                        clear_pairlist(nbl[th]);
 +                    }
 +
 +                    /* Divide the i super cell equally over the nblists */
 +                    nbnxn_make_pairlist_part(nbs, gridi, gridj,
 +                                             &nbs->work[th], nbat, excl,
 +                                             rlist,
 +                                             nb_kernel_type,
 +                                             ci_block,
 +                                             nbat->bUseBufferFlags,
 +                                             nsubpair_target,
 +                                             progBal, nsubpair_tot_est,
 +                                             th, nnbl,
 +                                             nbl[th],
 +                                             nbl_list->nbl_fep[th]);
 +                }
 +                GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 +            }
 +            nbs_cycle_stop(&nbs->cc[enbsCCsearch]);
 +
 +            np_tot = 0;
 +            np_noq = 0;
 +            np_hlj = 0;
 +            for (int th = 0; th < nnbl; th++)
 +            {
 +                inc_nrnb(nrnb, eNR_NBNXN_DIST2, nbs->work[th].ndistc);
 +
 +                if (nbl_list->bSimple)
 +                {
 +                    np_tot += nbl[th]->ncj;
 +                    np_noq += nbl[th]->work->ncj_noq;
 +                    np_hlj += nbl[th]->work->ncj_hlj;
 +                }
 +                else
 +                {
 +                    /* This count ignores potential subsequent pair pruning */
 +                    np_tot += nbl[th]->nci_tot;
 +                }
 +            }
 +            nap                   = nbl[0]->na_ci*nbl[0]->na_cj;
 +            nbl_list->natpair_ljq = (np_tot - np_noq)*nap - np_hlj*nap/2;
 +            nbl_list->natpair_lj  = np_noq*nap;
 +            nbl_list->natpair_q   = np_hlj*nap/2;
 +
 +            if (CombineNBLists && nnbl > 1)
 +            {
 +                nbs_cycle_start(&nbs->cc[enbsCCcombine]);
 +
 +                combine_nblists(nnbl-1, nbl+1, nbl[0]);
 +
 +                nbs_cycle_stop(&nbs->cc[enbsCCcombine]);
 +            }
 +        }
 +    }
 +
 +    if (!nbl_list->bSimple)
 +    {
 +        /* Sort the entries on size, large ones first */
 +        if (CombineNBLists || nnbl == 1)
 +        {
 +            sort_sci(nbl[0]);
 +        }
 +        else
 +        {
 +#pragma omp parallel for num_threads(nnbl) schedule(static)
 +            for (int th = 0; th < nnbl; th++)
 +            {
 +                try
 +                {
 +                    sort_sci(nbl[th]);
 +                }
 +                GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 +            }
 +        }
 +    }
 +
 +    if (nbat->bUseBufferFlags)
 +    {
 +        reduce_buffer_flags(nbs, nnbl, &nbat->buffer_flags);
 +    }
 +
 +    if (nbs->bFEP)
 +    {
 +        /* Balance the free-energy lists over all the threads */
 +        balance_fep_lists(nbs, nbl_list);
 +    }
 +
 +    /* Special performance logging stuff (env.var. GMX_NBNXN_CYCLE) */
 +    if (LOCAL_I(iloc))
 +    {
 +        nbs->search_count++;
 +    }
 +    if (nbs->print_cycles &&
 +        (!nbs->DomDec || !LOCAL_I(iloc)) &&
 +        nbs->search_count % 100 == 0)
 +    {
 +        nbs_cycle_print(stderr, nbs);
 +    }
 +
 +    if (debug && (CombineNBLists && nnbl > 1))
 +    {
 +        if (nbl[0]->bSimple)
 +        {
 +            print_nblist_statistics_simple(debug, nbl[0], nbs, rlist);
 +        }
 +        else
 +        {
 +            print_nblist_statistics_supersub(debug, nbl[0], nbs, rlist);
 +        }
 +    }
 +
 +    if (debug)
 +    {
 +        if (gmx_debug_at)
 +        {
 +            if (nbl[0]->bSimple)
 +            {
 +                print_nblist_ci_cj(debug, nbl[0]);
 +            }
 +            else
 +            {
 +                print_nblist_sci_cj(debug, nbl[0]);
 +            }
 +        }
 +
 +        if (nbat->bUseBufferFlags)
 +        {
 +            print_reduction_cost(&nbat->buffer_flags, nnbl);
 +        }
 +    }
 +}
diff --cc src/gromacs/mdlib/tgroup.cpp
index e8e117b428,0000000000..78ece843ca
mode 100644,000000..100644
--- a/src/gromacs/mdlib/tgroup.cpp
+++ b/src/gromacs/mdlib/tgroup.cpp
@@@ -1,333 -1,0 +1,333 @@@
 +/*
 + * This file is part of the GROMACS molecular simulation package.
 + *
 + * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
 + * Copyright (c) 2001-2004, The GROMACS development team.
-  * Copyright (c) 2013,2014,2015, by the GROMACS development team, led by
++ * Copyright (c) 2013,2014,2015,2016, by the GROMACS development team, led by
 + * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
 + * and including many others, as listed in the AUTHORS file in the
 + * top-level source directory and at http://www.gromacs.org.
 + *
 + * GROMACS is free software; you can redistribute it and/or
 + * modify it under the terms of the GNU Lesser General Public License
 + * as published by the Free Software Foundation; either version 2.1
 + * of the License, or (at your option) any later version.
 + *
 + * GROMACS is distributed in the hope that it will be useful,
 + * but WITHOUT ANY WARRANTY; without even the implied warranty of
 + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 + * Lesser General Public License for more details.
 + *
 + * You should have received a copy of the GNU Lesser General Public
 + * License along with GROMACS; if not, see
 + * http://www.gnu.org/licenses, or write to the Free Software Foundation,
 + * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
 + *
 + * If you want to redistribute modifications to GROMACS, please
 + * consider that scientific software is very special. Version
 + * control is crucial - bugs must be traceable. We will be happy to
 + * consider code for inclusion in the official distribution, but
 + * derived work must not be called official GROMACS. Details are found
 + * in the README & COPYING files - if they are missing, get the
 + * official version at http://www.gromacs.org.
 + *
 + * To help us fund GROMACS development, we humbly ask that you cite
 + * the research papers on the package. Check out http://www.gromacs.org.
 + */
 +/* This file is completely threadsafe - keep it that way! */
 +#include "gmxpre.h"
 +
 +#include "tgroup.h"
 +
 +#include <math.h>
 +
 +#include "gromacs/gmxlib/network.h"
 +#include "gromacs/math/vec.h"
 +#include "gromacs/mdlib/gmx_omp_nthreads.h"
 +#include "gromacs/mdlib/rbin.h"
 +#include "gromacs/mdlib/update.h"
 +#include "gromacs/mdtypes/group.h"
 +#include "gromacs/mdtypes/inputrec.h"
 +#include "gromacs/mdtypes/mdatom.h"
 +#include "gromacs/topology/mtop_util.h"
 +#include "gromacs/topology/topology.h"
 +#include "gromacs/utility/exceptions.h"
 +#include "gromacs/utility/fatalerror.h"
 +#include "gromacs/utility/futil.h"
 +#include "gromacs/utility/smalloc.h"
 +
 +static void init_grptcstat(int ngtc, t_grp_tcstat tcstat[])
 +{
 +    int i;
 +
 +    for (i = 0; (i < ngtc); i++)
 +    {
 +        tcstat[i].T = 0;
 +        clear_mat(tcstat[i].ekinh);
 +        clear_mat(tcstat[i].ekinh_old);
 +        clear_mat(tcstat[i].ekinf);
 +    }
 +}
 +
 +static void init_grpstat(gmx_mtop_t *mtop, int ngacc, t_grp_acc gstat[])
 +{
 +    gmx_groups_t           *groups;
 +    gmx_mtop_atomloop_all_t aloop;
 +    int                     i, grp;
 +    t_atom                 *atom;
 +
 +    if (ngacc > 0)
 +    {
 +        groups = &mtop->groups;
 +        aloop  = gmx_mtop_atomloop_all_init(mtop);
 +        while (gmx_mtop_atomloop_all_next(aloop, &i, &atom))
 +        {
 +            grp = ggrpnr(groups, egcACC, i);
 +            if ((grp < 0) && (grp >= ngacc))
 +            {
 +                gmx_incons("Input for acceleration groups wrong");
 +            }
 +            gstat[grp].nat++;
 +            /* This will not work for integrator BD */
 +            gstat[grp].mA += atom->m;
 +            gstat[grp].mB += atom->mB;
 +        }
 +    }
 +}
 +
 +void init_ekindata(FILE gmx_unused *log, gmx_mtop_t *mtop, t_grpopts *opts,
 +                   gmx_ekindata_t *ekind)
 +{
 +    int i;
 +    int nthread, thread;
 +#ifdef DEBUG
 +    fprintf(log, "ngtc: %d, ngacc: %d, ngener: %d\n", opts->ngtc, opts->ngacc,
 +            opts->ngener);
 +#endif
 +
 +    /* bNEMD tells if we should remove remove the COM velocity
 +     * from the velocities during velocity scaling in T-coupling.
 +     * Turn this on when we have multiple acceleration groups
 +     * or one accelerated group.
 +     */
-     ekind->bNEMD = (opts->ngacc > 1 || norm(opts->acc[0]) > 0);
++    ekind->bNEMD = (opts->ngacc > 1 || norm2(opts->acc[0]) > 0);
 +
 +    ekind->ngtc = opts->ngtc;
 +    snew(ekind->tcstat, opts->ngtc);
 +    init_grptcstat(opts->ngtc, ekind->tcstat);
 +    /* Set Berendsen tcoupl lambda's to 1,
 +     * so runs without Berendsen coupling are not affected.
 +     */
 +    for (i = 0; i < opts->ngtc; i++)
 +    {
 +        ekind->tcstat[i].lambda         = 1.0;
 +        ekind->tcstat[i].vscale_nhc     = 1.0;
 +        ekind->tcstat[i].ekinscaleh_nhc = 1.0;
 +        ekind->tcstat[i].ekinscalef_nhc = 1.0;
 +    }
 +
 +    nthread = gmx_omp_nthreads_get(emntUpdate);
 +
 +    snew(ekind->ekin_work_alloc, nthread);
 +    snew(ekind->ekin_work, nthread);
 +    snew(ekind->dekindl_work, nthread);
 +#pragma omp parallel for num_threads(nthread) schedule(static)
 +    for (thread = 0; thread < nthread; thread++)
 +    {
 +        try
 +        {
 +#define EKIN_WORK_BUFFER_SIZE 2
 +            /* Allocate 2 extra elements on both sides, so in single
 +             * precision we have
 +             * EKIN_WORK_BUFFER_SIZE*DIM*DIM*sizeof(real) = 72/144 bytes
 +             * buffer on both sides to avoid cache pollution.
 +             */
 +            snew(ekind->ekin_work_alloc[thread], ekind->ngtc+2*EKIN_WORK_BUFFER_SIZE);
 +            ekind->ekin_work[thread] = ekind->ekin_work_alloc[thread] + EKIN_WORK_BUFFER_SIZE;
 +            /* Nasty hack so we can have the per-thread accumulation
 +             * variable for dekindl in the same thread-local cache lines
 +             * as the per-thread accumulation tensors for ekin[fh],
 +             * because they are accumulated in the same loop. */
 +            ekind->dekindl_work[thread] = &(ekind->ekin_work[thread][ekind->ngtc][0][0]);
 +#undef EKIN_WORK_BUFFER_SIZE
 +        }
 +        GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
 +    }
 +
 +    ekind->ngacc = opts->ngacc;
 +    snew(ekind->grpstat, opts->ngacc);
 +    init_grpstat(mtop, opts->ngacc, ekind->grpstat);
 +}
 +
 +void accumulate_u(t_commrec *cr, t_grpopts *opts, gmx_ekindata_t *ekind)
 +{
 +    /* This routine will only be called when it's necessary */
 +    t_bin *rb;
 +    int    g;
 +
 +    rb = mk_bin();
 +
 +    for (g = 0; (g < opts->ngacc); g++)
 +    {
 +        add_binr(rb, DIM, ekind->grpstat[g].u);
 +    }
 +    sum_bin(rb, cr);
 +
 +    for (g = 0; (g < opts->ngacc); g++)
 +    {
 +        extract_binr(rb, DIM*g, DIM, ekind->grpstat[g].u);
 +    }
 +    destroy_bin(rb);
 +}
 +
 +/* I don't think accumulate_ekin is used anymore? */
 +
 +#if 0
 +static void accumulate_ekin(t_commrec *cr, t_grpopts *opts,
 +                            gmx_ekindata_t *ekind)
 +{
 +    int g;
 +
 +    if (PAR(cr))
 +    {
 +        for (g = 0; (g < opts->ngtc); g++)
 +        {
 +            gmx_sum(DIM*DIM, ekind->tcstat[g].ekinf[0], cr);
 +        }
 +    }
 +}
 +#endif
 +
 +void update_ekindata(int start, int homenr, gmx_ekindata_t *ekind,
 +                     t_grpopts *opts, rvec v[], t_mdatoms *md, real lambda)
 +{
 +    int  d, g, n;
 +    real mv;
 +
 +    /* calculate mean velocities at whole timestep */
 +    for (g = 0; (g < opts->ngtc); g++)
 +    {
 +        ekind->tcstat[g].T = 0;
 +    }
 +
 +    if (ekind->bNEMD)
 +    {
 +        for (g = 0; (g < opts->ngacc); g++)
 +        {
 +            clear_rvec(ekind->grpstat[g].u);
 +        }
 +
 +        g = 0;
 +        for (n = start; (n < start+homenr); n++)
 +        {
 +            if (md->cACC)
 +            {
 +                g = md->cACC[n];
 +            }
 +            for (d = 0; (d < DIM); d++)
 +            {
 +                mv                      = md->massT[n]*v[n][d];
 +                ekind->grpstat[g].u[d] += mv;
 +            }
 +        }
 +
 +        for (g = 0; (g < opts->ngacc); g++)
 +        {
 +            for (d = 0; (d < DIM); d++)
 +            {
 +                ekind->grpstat[g].u[d] /=
 +                    (1-lambda)*ekind->grpstat[g].mA + lambda*ekind->grpstat[g].mB;
 +            }
 +        }
 +    }
 +}
 +
 +real sum_ekin(t_grpopts *opts, gmx_ekindata_t *ekind, real *dekindlambda,
 +              gmx_bool bEkinAveVel, gmx_bool bScaleEkin)
 +{
 +    int           i, j, m, ngtc;
 +    real          T;
 +    t_grp_tcstat *tcstat;
 +    real          nrdf, nd, *ndf;
 +
 +    ngtc = opts->ngtc;
 +    ndf  = opts->nrdf;
 +
 +    T    = 0;
 +    nrdf = 0;
 +
 +    clear_mat(ekind->ekin);
 +
 +    for (i = 0; (i < ngtc); i++)
 +    {
 +
 +        nd     = ndf[i];
 +        tcstat = &ekind->tcstat[i];
 +        /* Sometimes a group does not have degrees of freedom, e.g.
 +         * when it consists of shells and virtual sites, then we just
 +         * set the temperatue to 0 and also neglect the kinetic
 +         * energy, which should be  zero anyway.
 +         */
 +
 +        if (nd > 0)
 +        {
 +            if (bEkinAveVel)
 +            {
 +                if (!bScaleEkin)
 +                {
 +                    /* in this case, kinetic energy is from the current velocities already */
 +                    msmul(tcstat->ekinf, tcstat->ekinscalef_nhc, tcstat->ekinf);
 +                }
 +            }
 +            else
 +            {
 +                /* Calculate the full step Ekin as the average of the half steps */
 +                for (j = 0; (j < DIM); j++)
 +                {
 +                    for (m = 0; (m < DIM); m++)
 +                    {
 +                        tcstat->ekinf[j][m] =
 +                            0.5*(tcstat->ekinh[j][m]*tcstat->ekinscaleh_nhc + tcstat->ekinh_old[j][m]);
 +                    }
 +                }
 +            }
 +            m_add(tcstat->ekinf, ekind->ekin, ekind->ekin);
 +
 +            tcstat->Th = calc_temp(trace(tcstat->ekinh), nd);
 +            tcstat->T  = calc_temp(trace(tcstat->ekinf), nd);
 +
 +            /* after the scaling factors have been multiplied in, we can remove them */
 +            if (bEkinAveVel)
 +            {
 +                tcstat->ekinscalef_nhc = 1.0;
 +            }
 +            else
 +            {
 +                tcstat->ekinscaleh_nhc = 1.0;
 +            }
 +        }
 +        else
 +        {
 +            tcstat->T  = 0;
 +            tcstat->Th = 0;
 +        }
 +        T    += nd*tcstat->T;
 +        nrdf += nd;
 +    }
 +    if (nrdf > 0)
 +    {
 +        T /= nrdf;
 +    }
 +    if (dekindlambda)
 +    {
 +        if (bEkinAveVel)
 +        {
 +            *dekindlambda = ekind->dekindl;
 +        }
 +        else
 +        {
 +            *dekindlambda = 0.5*(ekind->dekindl + ekind->dekindl_old);
 +        }
 +    }
 +    return T;
 +}
diff --cc src/gromacs/tools/compare.cpp
index 95e13233f9,0000000000..c5afe10f58
mode 100644,000000..100644
--- a/src/gromacs/tools/compare.cpp
+++ b/src/gromacs/tools/compare.cpp
@@@ -1,1445 -1,0 +1,1469 @@@
 +/*
 + * This file is part of the GROMACS molecular simulation package.
 + *
 + * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
 + * Copyright (c) 2001-2004, The GROMACS development team.
 + * Copyright (c) 2013,2014,2015,2016, by the GROMACS development team, led by
 + * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
 + * and including many others, as listed in the AUTHORS file in the
 + * top-level source directory and at http://www.gromacs.org.
 + *
 + * GROMACS is free software; you can redistribute it and/or
 + * modify it under the terms of the GNU Lesser General Public License
 + * as published by the Free Software Foundation; either version 2.1
 + * of the License, or (at your option) any later version.
 + *
 + * GROMACS is distributed in the hope that it will be useful,
 + * but WITHOUT ANY WARRANTY; without even the implied warranty of
 + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 + * Lesser General Public License for more details.
 + *
 + * You should have received a copy of the GNU Lesser General Public
 + * License along with GROMACS; if not, see
 + * http://www.gnu.org/licenses, or write to the Free Software Foundation,
 + * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
 + *
 + * If you want to redistribute modifications to GROMACS, please
 + * consider that scientific software is very special. Version
 + * control is crucial - bugs must be traceable. We will be happy to
 + * consider code for inclusion in the official distribution, but
 + * derived work must not be called official GROMACS. Details are found
 + * in the README & COPYING files - if they are missing, get the
 + * official version at http://www.gromacs.org.
 + *
 + * To help us fund GROMACS development, we humbly ask that you cite
 + * the research papers on the package. Check out http://www.gromacs.org.
 + */
 +/* This file is completely threadsafe - keep it that way! */
 +
 +#include "gmxpre.h"
 +
 +#include <cmath>
 +#include <cstdio>
 +#include <cstring>
 +
 +#include <algorithm>
 +
 +#include "gromacs/fileio/enxio.h"
 +#include "gromacs/fileio/tpxio.h"
 +#include "gromacs/fileio/trxio.h"
 +#include "gromacs/mdtypes/inputrec.h"
 +#include "gromacs/mdtypes/md_enums.h"
 +#include "gromacs/mdtypes/pull-params.h"
 +#include "gromacs/topology/ifunc.h"
 +#include "gromacs/topology/mtop_util.h"
 +#include "gromacs/topology/topology.h"
 +#include "gromacs/trajectory/trajectoryframe.h"
 +#include "gromacs/utility/cstringutil.h"
 +#include "gromacs/utility/fatalerror.h"
 +#include "gromacs/utility/futil.h"
 +#include "gromacs/utility/smalloc.h"
 +#include "gromacs/utility/stringutil.h"
 +
 +static void cmp_int(FILE *fp, const char *s, int index, int i1, int i2)
 +{
 +    if (i1 != i2)
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%d] (%d - %d)\n", s, index, i1, i2);
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%d - %d)\n", s, i1, i2);
 +        }
 +    }
 +}
 +
 +static void cmp_int64(FILE *fp, const char *s, gmx_int64_t i1, gmx_int64_t i2)
 +{
 +    if (i1 != i2)
 +    {
 +        fprintf(fp, "%s (", s);
 +        fprintf(fp, "%" GMX_PRId64, i1);
 +        fprintf(fp, " - ");
 +        fprintf(fp, "%" GMX_PRId64, i2);
 +        fprintf(fp, ")\n");
 +    }
 +}
 +
 +static void cmp_us(FILE *fp, const char *s, int index, unsigned short i1, unsigned short i2)
 +{
 +    if (i1 != i2)
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%d] (%hu - %hu)\n", s, index, i1, i2);
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%hu - %hu)\n", s, i1, i2);
 +        }
 +    }
 +}
 +
 +static void cmp_uc(FILE *fp, const char *s, int index, unsigned char i1, unsigned char i2)
 +{
 +    if (i1 != i2)
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%d] (%d - %d)\n", s, index, i1, i2);
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%d - %d)\n", s, i1, i2);
 +        }
 +    }
 +}
 +
 +static gmx_bool cmp_bool(FILE *fp, const char *s, int index, gmx_bool b1, gmx_bool b2)
 +{
 +    if (b1)
 +    {
 +        b1 = 1;
 +    }
 +    else
 +    {
 +        b1 = 0;
 +    }
 +    if (b2)
 +    {
 +        b2 = 1;
 +    }
 +    else
 +    {
 +        b2 = 0;
 +    }
 +    if (b1 != b2)
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%d] (%s - %s)\n", s, index,
 +                    gmx::boolToString(b1), gmx::boolToString(b2));
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%s - %s)\n", s,
 +                    gmx::boolToString(b1), gmx::boolToString(b2));
 +        }
 +    }
 +    return b1 && b2;
 +}
 +
 +static void cmp_str(FILE *fp, const char *s, int index,
 +                    const char *s1, const char *s2)
 +{
 +    if (std::strcmp(s1, s2) != 0)
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%d] (%s - %s)\n", s, index, s1, s2);
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%s - %s)\n", s, s1, s2);
 +        }
 +    }
 +}
 +
 +static gmx_bool equal_real(real i1, real i2, real ftol, real abstol)
 +{
 +    return ( ( 2*fabs(i1 - i2) <= (fabs(i1) + fabs(i2))*ftol ) || fabs(i1-i2) <= abstol );
 +}
 +
 +static gmx_bool equal_float(float i1, float i2, float ftol, float abstol)
 +{
 +    return ( ( 2*fabs(i1 - i2) <= (fabs(i1) + fabs(i2))*ftol ) || fabs(i1-i2) <= abstol );
 +}
 +
 +static gmx_bool equal_double(double i1, double i2, real ftol, real abstol)
 +{
 +    return ( ( 2*fabs(i1 - i2) <= (fabs(i1) + fabs(i2))*ftol ) || fabs(i1-i2) <= abstol );
 +}
 +
 +static void
 +cmp_real(FILE *fp, const char *s, int index, real i1, real i2, real ftol, real abstol)
 +{
 +    if (!equal_real(i1, i2, ftol, abstol))
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%2d] (%e - %e)\n", s, index, i1, i2);
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%e - %e)\n", s, i1, i2);
 +        }
 +    }
 +}
 +
 +static void
 +cmp_float(FILE *fp, const char *s, int index, float i1, float i2, float ftol, float abstol)
 +{
 +    if (!equal_float(i1, i2, ftol, abstol))
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%2d] (%e - %e)\n", s, index, i1, i2);
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%e - %e)\n", s, i1, i2);
 +        }
 +    }
 +}
 +
 +
 +
 +static void
 +cmp_double(FILE *fp, const char *s, int index, double i1, double i2, double ftol, double abstol)
 +{
 +    if (!equal_double(i1, i2, ftol, abstol))
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%2d] (%16.9e - %16.9e)\n", s, index, i1, i2);
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%16.9e - %16.9e)\n", s, i1, i2);
 +        }
 +    }
 +}
 +
 +static void cmp_rvec(FILE *fp, const char *s, int index, rvec i1, rvec i2, real ftol, real abstol)
 +{
 +    if (!equal_real(i1[XX], i2[XX], ftol, abstol) ||
 +        !equal_real(i1[YY], i2[YY], ftol, abstol) ||
 +        !equal_real(i1[ZZ], i2[ZZ], ftol, abstol))
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%5d] (%12.5e %12.5e %12.5e) - (%12.5e %12.5e %12.5e)\n",
 +                    s, index, i1[XX], i1[YY], i1[ZZ], i2[XX], i2[YY], i2[ZZ]);
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%12.5e %12.5e %12.5e) - (%12.5e %12.5e %12.5e)\n",
 +                    s, i1[XX], i1[YY], i1[ZZ], i2[XX], i2[YY], i2[ZZ]);
 +        }
 +    }
 +}
 +
 +static void cmp_ivec(FILE *fp, const char *s, int index, ivec i1, ivec i2)
 +{
 +    if ((i1[XX] != i2[XX]) || (i1[YY] != i2[YY]) || (i1[ZZ] != i2[ZZ]))
 +    {
 +        if (index != -1)
 +        {
 +            fprintf(fp, "%s[%5d] (%8d,%8d,%8d - %8d,%8d,%8d)\n", s, index,
 +                    i1[XX], i1[YY], i1[ZZ], i2[XX], i2[YY], i2[ZZ]);
 +        }
 +        else
 +        {
 +            fprintf(fp, "%s (%8d,%8d,%8d - %8d,%8d,%8d)\n", s,
 +                    i1[XX], i1[YY], i1[ZZ], i2[XX], i2[YY], i2[ZZ]);
 +        }
 +    }
 +}
 +
 +static void cmp_ilist(FILE *fp, int ftype, t_ilist *il1, t_ilist *il2)
 +{
 +    int  i;
 +    char buf[256];
 +
 +    fprintf(fp, "comparing ilist %s\n", interaction_function[ftype].name);
 +    sprintf(buf, "%s->nr", interaction_function[ftype].name);
 +    cmp_int(fp, buf, -1, il1->nr, il2->nr);
 +    sprintf(buf, "%s->iatoms", interaction_function[ftype].name);
 +    if (((il1->nr > 0) && (!il1->iatoms)) ||
 +        ((il2->nr > 0) && (!il2->iatoms)) ||
 +        ((il1->nr != il2->nr)))
 +    {
 +        fprintf(fp, "Comparing radically different topologies - %s is different\n",
 +                buf);
 +    }
 +    else
 +    {
 +        for (i = 0; (i < il1->nr); i++)
 +        {
 +            cmp_int(fp, buf, i, il1->iatoms[i], il2->iatoms[i]);
 +        }
 +    }
 +}
 +
 +void cmp_iparm(FILE *fp, const char *s, t_functype ft,
 +               t_iparams ip1, t_iparams ip2, real ftol, real abstol)
 +{
 +    int      i;
 +    gmx_bool bDiff;
 +
 +    bDiff = FALSE;
 +    for (i = 0; i < MAXFORCEPARAM && !bDiff; i++)
 +    {
 +        bDiff = !equal_real(ip1.generic.buf[i], ip2.generic.buf[i], ftol, abstol);
 +    }
 +    if (bDiff)
 +    {
 +        fprintf(fp, "%s1: ", s);
 +        pr_iparams(fp, ft, &ip1);
 +        fprintf(fp, "%s2: ", s);
 +        pr_iparams(fp, ft, &ip2);
 +    }
 +}
 +
 +void cmp_iparm_AB(FILE *fp, const char *s, t_functype ft, t_iparams ip1, real ftol, real abstol)
 +{
 +    int      nrfpA, nrfpB, p0, i;
 +    gmx_bool bDiff;
 +
 +    /* Normally the first parameter is perturbable */
 +    p0    = 0;
 +    nrfpA = interaction_function[ft].nrfpA;
 +    nrfpB = interaction_function[ft].nrfpB;
 +    if (ft == F_PDIHS)
 +    {
 +        nrfpB = 2;
 +    }
 +    else if (interaction_function[ft].flags & IF_TABULATED)
 +    {
 +        /* For tabulated interactions only the second parameter is perturbable */
 +        p0    = 1;
 +        nrfpB = 1;
 +    }
 +    bDiff = FALSE;
 +    for (i = 0; i < nrfpB && !bDiff; i++)
 +    {
 +        bDiff = !equal_real(ip1.generic.buf[p0+i], ip1.generic.buf[nrfpA+i], ftol, abstol);
 +    }
 +    if (bDiff)
 +    {
 +        fprintf(fp, "%s: ", s);
 +        pr_iparams(fp, ft, &ip1);
 +    }
 +}
 +
++static void cmp_cmap(FILE *fp, const gmx_cmap_t *cmap1, const gmx_cmap_t *cmap2, real ftol, real abstol)
++{
++    cmp_int(fp, "cmap ngrid", -1, cmap1->ngrid, cmap2->ngrid);
++    cmp_int(fp, "cmap grid_spacing", -1, cmap1->grid_spacing, cmap2->grid_spacing);
++    if (cmap1->ngrid == cmap2->ngrid &&
++        cmap1->grid_spacing == cmap2->grid_spacing)
++    {
++        int g;
++
++        for (g = 0; g < cmap1->ngrid; g++)
++        {
++            int i;
++
++            fprintf(fp, "comparing cmap %d\n", g);
++
++            for (i = 0; i < 4*cmap1->grid_spacing*cmap1->grid_spacing; i++)
++            {
++                cmp_real(fp, "", i, cmap1->cmapdata[g].cmap[i], cmap2->cmapdata[g].cmap[i], ftol, abstol);
++            }
++        }
++    }
++}
++
 +static void cmp_idef(FILE *fp, t_idef *id1, t_idef *id2, real ftol, real abstol)
 +{
 +    int  i;
 +    char buf1[64], buf2[64];
 +
 +    fprintf(fp, "comparing idef\n");
 +    if (id2)
 +    {
 +        cmp_int(fp, "idef->ntypes", -1, id1->ntypes, id2->ntypes);
 +        cmp_int(fp, "idef->atnr",  -1, id1->atnr, id2->atnr);
 +        for (i = 0; (i < std::min(id1->ntypes, id2->ntypes)); i++)
 +        {
 +            sprintf(buf1, "idef->functype[%d]", i);
 +            sprintf(buf2, "idef->iparam[%d]", i);
 +            cmp_int(fp, buf1, i, (int)id1->functype[i], (int)id2->functype[i]);
 +            cmp_iparm(fp, buf2, id1->functype[i],
 +                      id1->iparams[i], id2->iparams[i], ftol, abstol);
 +        }
 +        cmp_real(fp, "fudgeQQ", -1, id1->fudgeQQ, id2->fudgeQQ, ftol, abstol);
++        cmp_cmap(fp, &id1->cmap_grid, &id2->cmap_grid, ftol, abstol);
 +        for (i = 0; (i < F_NRE); i++)
 +        {
 +            cmp_ilist(fp, i, &(id1->il[i]), &(id2->il[i]));
 +        }
 +    }
 +    else
 +    {
 +        for (i = 0; (i < id1->ntypes); i++)
 +        {
 +            cmp_iparm_AB(fp, "idef->iparam", id1->functype[i], id1->iparams[i], ftol, abstol);
 +        }
 +    }
 +}
 +
 +static void cmp_block(FILE *fp, t_block *b1, t_block *b2, const char *s)
 +{
 +    char buf[32];
 +
 +    fprintf(fp, "comparing block %s\n", s);
 +    sprintf(buf, "%s.nr", s);
 +    cmp_int(fp, buf, -1, b1->nr, b2->nr);
 +}
 +
 +static void cmp_blocka(FILE *fp, t_blocka *b1, t_blocka *b2, const char *s)
 +{
 +    char buf[32];
 +
 +    fprintf(fp, "comparing blocka %s\n", s);
 +    sprintf(buf, "%s.nr", s);
 +    cmp_int(fp, buf, -1, b1->nr, b2->nr);
 +    sprintf(buf, "%s.nra", s);
 +    cmp_int(fp, buf, -1, b1->nra, b2->nra);
 +}
 +
 +static void cmp_atom(FILE *fp, int index, t_atom *a1, t_atom *a2, real ftol, real abstol)
 +{
 +    if (a2)
 +    {
 +        cmp_us(fp, "atom.type", index, a1->type, a2->type);
 +        cmp_us(fp, "atom.ptype", index, a1->ptype, a2->ptype);
 +        cmp_int(fp, "atom.resind", index, a1->resind, a2->resind);
 +        cmp_int(fp, "atom.atomnumber", index, a1->atomnumber, a2->atomnumber);
 +        cmp_real(fp, "atom.m", index, a1->m, a2->m, ftol, abstol);
 +        cmp_real(fp, "atom.q", index, a1->q, a2->q, ftol, abstol);
 +        cmp_us(fp, "atom.typeB", index, a1->typeB, a2->typeB);
 +        cmp_real(fp, "atom.mB", index, a1->mB, a2->mB, ftol, abstol);
 +        cmp_real(fp, "atom.qB", index, a1->qB, a2->qB, ftol, abstol);
 +    }
 +    else
 +    {
 +        cmp_us(fp, "atom.type", index, a1->type, a1->typeB);
 +        cmp_real(fp, "atom.m", index, a1->m, a1->mB, ftol, abstol);
 +        cmp_real(fp, "atom.q", index, a1->q, a1->qB, ftol, abstol);
 +    }
 +}
 +
 +static void cmp_atoms(FILE *fp, t_atoms *a1, t_atoms *a2, real ftol, real abstol)
 +{
 +    int i;
 +
 +    fprintf(fp, "comparing atoms\n");
 +
 +    if (a2)
 +    {
 +        cmp_int(fp, "atoms->nr", -1, a1->nr, a2->nr);
 +        for (i = 0; (i < a1->nr); i++)
 +        {
 +            cmp_atom(fp, i, &(a1->atom[i]), &(a2->atom[i]), ftol, abstol);
 +        }
 +    }
 +    else
 +    {
 +        for (i = 0; (i < a1->nr); i++)
 +        {
 +            cmp_atom(fp, i, &(a1->atom[i]), NULL, ftol, abstol);
 +        }
 +    }
 +}
 +
 +static void cmp_top(FILE *fp, t_topology *t1, t_topology *t2, real ftol, real abstol)
 +{
 +    fprintf(fp, "comparing top\n");
 +    if (t2)
 +    {
 +        cmp_idef(fp, &(t1->idef), &(t2->idef), ftol, abstol);
 +        cmp_atoms(fp, &(t1->atoms), &(t2->atoms), ftol, abstol);
 +        cmp_block(fp, &t1->cgs, &t2->cgs, "cgs");
 +        cmp_block(fp, &t1->mols, &t2->mols, "mols");
 +        cmp_bool(fp, "bIntermolecularInteractions", -1, t1->bIntermolecularInteractions, t2->bIntermolecularInteractions);
 +        cmp_blocka(fp, &t1->excls, &t2->excls, "excls");
 +    }
 +    else
 +    {
 +        cmp_idef(fp, &(t1->idef), NULL, ftol, abstol);
 +        cmp_atoms(fp, &(t1->atoms), NULL, ftol, abstol);
 +    }
 +}
 +
 +static void cmp_groups(FILE *fp, gmx_groups_t *g0, gmx_groups_t *g1,
 +                       int natoms0, int natoms1)
 +{
 +    int  i, j;
 +    char buf[32];
 +
 +    fprintf(fp, "comparing groups\n");
 +
 +    for (i = 0; i < egcNR; i++)
 +    {
 +        sprintf(buf, "grps[%d].nr", i);
 +        cmp_int(fp, buf, -1, g0->grps[i].nr, g1->grps[i].nr);
 +        if (g0->grps[i].nr == g1->grps[i].nr)
 +        {
 +            for (j = 0; j < g0->grps[i].nr; j++)
 +            {
 +                sprintf(buf, "grps[%d].name[%d]", i, j);
 +                cmp_str(fp, buf, -1,
 +                        *g0->grpname[g0->grps[i].nm_ind[j]],
 +                        *g1->grpname[g1->grps[i].nm_ind[j]]);
 +            }
 +        }
 +        cmp_int(fp, "ngrpnr", i, g0->ngrpnr[i], g1->ngrpnr[i]);
 +        if (g0->ngrpnr[i] == g1->ngrpnr[i] && natoms0 == natoms1 &&
 +            (g0->grpnr[i] != NULL || g1->grpnr[i] != NULL))
 +        {
 +            for (j = 0; j < natoms0; j++)
 +            {
 +                cmp_int(fp, gtypes[i], j, ggrpnr(g0, i, j), ggrpnr(g1, i, j));
 +            }
 +        }
 +    }
 +    /* We have compared the names in the groups lists,
 +     * so we can skip the grpname list comparison.
 +     */
 +}
 +
 +static void cmp_rvecs_rmstol(FILE *fp, const char *title, int n, rvec x1[], rvec x2[],
 +                             real ftol, real abstol)
 +{
 +    int    i, m;
 +    double rms;
 +
 +    /* For a vector you are usally not interested in a relative difference
 +     * on a component that is very small compared to the other components.
 +     * Therefore we do the relative comparision relative to the RMS component.
 +     */
 +    rms = 0.0;
 +    for (i = 0; (i < n); i++)
 +    {
 +        for (m = 0; m < DIM; m++)
 +        {
 +            rms += x1[i][m]*x1[i][m] + x2[i][m]*x2[i][m];
 +        }
 +    }
 +    rms = sqrt(rms/(2*n*DIM));
 +
 +    /* Convert the relative tolerance into an absolute tolerance */
 +    if (ftol*rms < abstol)
 +    {
 +        abstol = ftol*rms;
 +    }
 +
 +    /* And now do the actual comparision */
 +    for (i = 0; (i < n); i++)
 +    {
 +        cmp_rvec(fp, title, i, x1[i], x2[i], 0.0, abstol);
 +    }
 +}
 +
 +static void cmp_rvecs(FILE *fp, const char *title, int n, rvec x1[], rvec x2[],
 +                      gmx_bool bRMSD, real ftol, real abstol)
 +{
 +    int    i, m;
 +    double d, ssd;
 +
 +    if (bRMSD)
 +    {
 +        ssd = 0;
 +        for (i = 0; (i < n); i++)
 +        {
 +            for (m = 0; m < DIM; m++)
 +            {
 +                d    = x1[i][m] - x2[i][m];
 +                ssd += d*d;
 +            }
 +        }
 +        fprintf(fp, "%s RMSD %g\n", title, std::sqrt(ssd/n));
 +    }
 +    else
 +    {
 +        cmp_rvecs_rmstol(fp, title, n, x1, x2, ftol, abstol);
 +    }
 +}
 +
 +static void cmp_grpopts(FILE *fp, t_grpopts *opt1, t_grpopts *opt2, real ftol, real abstol)
 +{
 +    int  i, j;
 +    char buf1[256], buf2[256];
 +
 +    cmp_int(fp, "inputrec->grpopts.ngtc", -1,  opt1->ngtc, opt2->ngtc);
 +    cmp_int(fp, "inputrec->grpopts.ngacc", -1, opt1->ngacc, opt2->ngacc);
 +    cmp_int(fp, "inputrec->grpopts.ngfrz", -1, opt1->ngfrz, opt2->ngfrz);
 +    cmp_int(fp, "inputrec->grpopts.ngener", -1, opt1->ngener, opt2->ngener);
 +    for (i = 0; (i < std::min(opt1->ngtc, opt2->ngtc)); i++)
 +    {
 +        cmp_real(fp, "inputrec->grpopts.nrdf", i, opt1->nrdf[i], opt2->nrdf[i], ftol, abstol);
 +        cmp_real(fp, "inputrec->grpopts.ref_t", i, opt1->ref_t[i], opt2->ref_t[i], ftol, abstol);
 +        cmp_real(fp, "inputrec->grpopts.tau_t", i, opt1->tau_t[i], opt2->tau_t[i], ftol, abstol);
 +        cmp_int(fp, "inputrec->grpopts.annealing", i, opt1->annealing[i], opt2->annealing[i]);
 +        cmp_int(fp, "inputrec->grpopts.anneal_npoints", i,
 +                opt1->anneal_npoints[i], opt2->anneal_npoints[i]);
 +        if (opt1->anneal_npoints[i] == opt2->anneal_npoints[i])
 +        {
 +            sprintf(buf1, "inputrec->grpopts.anneal_time[%d]", i);
 +            sprintf(buf2, "inputrec->grpopts.anneal_temp[%d]", i);
 +            for (j = 0; j < opt1->anneal_npoints[i]; j++)
 +            {
 +                cmp_real(fp, buf1, j, opt1->anneal_time[i][j], opt2->anneal_time[i][j], ftol, abstol);
 +                cmp_real(fp, buf2, j, opt1->anneal_temp[i][j], opt2->anneal_temp[i][j], ftol, abstol);
 +            }
 +        }
 +    }
 +    if (opt1->ngener == opt2->ngener)
 +    {
 +        for (i = 0; i < opt1->ngener; i++)
 +        {
 +            for (j = i; j < opt1->ngener; j++)
 +            {
 +                sprintf(buf1, "inputrec->grpopts.egp_flags[%d]", i);
 +                cmp_int(fp, buf1, j,
 +                        opt1->egp_flags[opt1->ngener*i+j],
 +                        opt2->egp_flags[opt1->ngener*i+j]);
 +            }
 +        }
 +    }
 +    for (i = 0; (i < std::min(opt1->ngacc, opt2->ngacc)); i++)
 +    {
 +        cmp_rvec(fp, "inputrec->grpopts.acc", i, opt1->acc[i], opt2->acc[i], ftol, abstol);
 +    }
 +    for (i = 0; (i < std::min(opt1->ngfrz, opt2->ngfrz)); i++)
 +    {
 +        cmp_ivec(fp, "inputrec->grpopts.nFreeze", i, opt1->nFreeze[i], opt2->nFreeze[i]);
 +    }
 +}
 +
 +static void cmp_cosines(FILE *fp, const char *s, t_cosines c1[DIM], t_cosines c2[DIM], real ftol, real abstol)
 +{
 +    int  i, m;
 +    char buf[256];
 +
 +    for (m = 0; (m < DIM); m++)
 +    {
 +        sprintf(buf, "inputrec->%s[%d]", s, m);
 +        cmp_int(fp, buf, 0, c1->n, c2->n);
 +        for (i = 0; (i < std::min(c1->n, c2->n)); i++)
 +        {
 +            cmp_real(fp, buf, i, c1->a[i], c2->a[i], ftol, abstol);
 +            cmp_real(fp, buf, i, c1->phi[i], c2->phi[i], ftol, abstol);
 +        }
 +    }
 +}
 +static void cmp_pull(FILE *fp)
 +{
 +    fprintf(fp, "WARNING: Both files use COM pulling, but comparing of the pull struct is not implemented (yet). The pull parameters could be the same or different.\n");
 +}
 +
 +static void cmp_simtempvals(FILE *fp, t_simtemp *simtemp1, t_simtemp *simtemp2, int n_lambda, real ftol, real abstol)
 +{
 +    int i;
 +    cmp_int(fp, "inputrec->simtempvals->eSimTempScale", -1, simtemp1->eSimTempScale, simtemp2->eSimTempScale);
 +    cmp_real(fp, "inputrec->simtempvals->simtemp_high", -1, simtemp1->simtemp_high, simtemp2->simtemp_high, ftol, abstol);
 +    cmp_real(fp, "inputrec->simtempvals->simtemp_low", -1, simtemp1->simtemp_low, simtemp2->simtemp_low, ftol, abstol);
 +    for (i = 0; i < n_lambda; i++)
 +    {
 +        cmp_real(fp, "inputrec->simtempvals->temperatures", -1, simtemp1->temperatures[i], simtemp2->temperatures[i], ftol, abstol);
 +    }
 +}
 +
 +static void cmp_expandedvals(FILE *fp, t_expanded *expand1, t_expanded *expand2, int n_lambda, real ftol, real abstol)
 +{
 +    int i;
 +
 +    cmp_bool(fp, "inputrec->fepvals->bInit_weights", -1, expand1->bInit_weights, expand2->bInit_weights);
 +    cmp_bool(fp, "inputrec->fepvals->bWLoneovert", -1, expand1->bWLoneovert, expand2->bWLoneovert);
 +
 +    for (i = 0; i < n_lambda; i++)
 +    {
 +        cmp_real(fp, "inputrec->expandedvals->init_lambda_weights", -1,
 +                 expand1->init_lambda_weights[i], expand2->init_lambda_weights[i], ftol, abstol);
 +    }
 +
 +    cmp_int(fp, "inputrec->expandedvals->lambda-stats", -1, expand1->elamstats, expand2->elamstats);
 +    cmp_int(fp, "inputrec->expandedvals->lambda-mc-move", -1, expand1->elmcmove, expand2->elmcmove);
 +    cmp_int(fp, "inputrec->expandedvals->lmc-repeats", -1, expand1->lmc_repeats, expand2->lmc_repeats);
 +    cmp_int(fp, "inputrec->expandedvals->lmc-gibbsdelta", -1, expand1->gibbsdeltalam, expand2->gibbsdeltalam);
 +    cmp_int(fp, "inputrec->expandedvals->lmc-forced-nstart", -1, expand1->lmc_forced_nstart, expand2->lmc_forced_nstart);
 +    cmp_int(fp, "inputrec->expandedvals->lambda-weights-equil", -1, expand1->elmceq, expand2->elmceq);
 +    cmp_int(fp, "inputrec->expandedvals->,weight-equil-number-all-lambda", -1, expand1->equil_n_at_lam, expand2->equil_n_at_lam);
 +    cmp_int(fp, "inputrec->expandedvals->weight-equil-number-samples", -1, expand1->equil_samples, expand2->equil_samples);
 +    cmp_int(fp, "inputrec->expandedvals->weight-equil-number-steps", -1, expand1->equil_steps, expand2->equil_steps);
 +    cmp_real(fp, "inputrec->expandedvals->weight-equil-wl-delta", -1, expand1->equil_wl_delta, expand2->equil_wl_delta, ftol, abstol);
 +    cmp_real(fp, "inputrec->expandedvals->weight-equil-count-ratio", -1, expand1->equil_ratio, expand2->equil_ratio, ftol, abstol);
 +    cmp_bool(fp, "inputrec->expandedvals->symmetrized-transition-matrix", -1, expand1->bSymmetrizedTMatrix, expand2->bSymmetrizedTMatrix);
 +    cmp_int(fp, "inputrec->expandedvals->nstTij", -1, expand1->nstTij, expand2->nstTij);
 +    cmp_int(fp, "inputrec->expandedvals->mininum-var-min", -1, expand1->minvarmin, expand2->minvarmin); /*default is reasonable */
 +    cmp_int(fp, "inputrec->expandedvals->weight-c-range", -1, expand1->c_range, expand2->c_range);      /* default is just C=0 */
 +    cmp_real(fp, "inputrec->expandedvals->wl-scale", -1, expand1->wl_scale, expand2->wl_scale, ftol, abstol);
 +    cmp_real(fp, "inputrec->expandedvals->init-wl-delta", -1, expand1->init_wl_delta, expand2->init_wl_delta, ftol, abstol);
 +    cmp_real(fp, "inputrec->expandedvals->wl-ratio", -1, expand1->wl_ratio, expand2->wl_ratio, ftol, abstol);
 +    cmp_int(fp, "inputrec->expandedvals->nstexpanded", -1, expand1->nstexpanded, expand2->nstexpanded);
 +    cmp_int(fp, "inputrec->expandedvals->lmc-seed", -1, expand1->lmc_seed, expand2->lmc_seed);
 +    cmp_real(fp, "inputrec->expandedvals->mc-temperature", -1, expand1->mc_temp, expand2->mc_temp, ftol, abstol);
 +}
 +
 +static void cmp_fepvals(FILE *fp, t_lambda *fep1, t_lambda *fep2, real ftol, real abstol)
 +{
 +    int i, j;
 +    cmp_int(fp, "inputrec->nstdhdl", -1, fep1->nstdhdl, fep2->nstdhdl);
 +    cmp_double(fp, "inputrec->fepvals->init_fep_state", -1, fep1->init_fep_state, fep2->init_fep_state, ftol, abstol);
 +    cmp_double(fp, "inputrec->fepvals->delta_lambda", -1, fep1->delta_lambda, fep2->delta_lambda, ftol, abstol);
 +    cmp_int(fp, "inputrec->fepvals->n_lambda", -1, fep1->n_lambda, fep2->n_lambda);
 +    for (i = 0; i < efptNR; i++)
 +    {
 +        for (j = 0; j < std::min(fep1->n_lambda, fep2->n_lambda); j++)
 +        {
 +            cmp_double(fp, "inputrec->fepvals->all_lambda", -1, fep1->all_lambda[i][j], fep2->all_lambda[i][j], ftol, abstol);
 +        }
 +    }
 +    cmp_int(fp, "inputrec->fepvals->lambda_neighbors", 1, fep1->lambda_neighbors,
 +            fep2->lambda_neighbors);
 +    cmp_real(fp, "inputrec->fepvals->sc_alpha", -1, fep1->sc_alpha, fep2->sc_alpha, ftol, abstol);
 +    cmp_int(fp, "inputrec->fepvals->sc_power", -1, fep1->sc_power, fep2->sc_power);
 +    cmp_real(fp, "inputrec->fepvals->sc_r_power", -1, fep1->sc_r_power, fep2->sc_r_power, ftol, abstol);
 +    cmp_real(fp, "inputrec->fepvals->sc_sigma", -1, fep1->sc_sigma, fep2->sc_sigma, ftol, abstol);
 +    cmp_int(fp, "inputrec->fepvals->edHdLPrintEnergy", -1, fep1->edHdLPrintEnergy, fep1->edHdLPrintEnergy);
 +    cmp_bool(fp, "inputrec->fepvals->bScCoul", -1, fep1->bScCoul, fep1->bScCoul);
 +    cmp_int(fp, "inputrec->separate_dhdl_file", -1, fep1->separate_dhdl_file, fep2->separate_dhdl_file);
 +    cmp_int(fp, "inputrec->dhdl_derivatives", -1, fep1->dhdl_derivatives, fep2->dhdl_derivatives);
 +    cmp_int(fp, "inputrec->dh_hist_size", -1, fep1->dh_hist_size, fep2->dh_hist_size);
 +    cmp_double(fp, "inputrec->dh_hist_spacing", -1, fep1->dh_hist_spacing, fep2->dh_hist_spacing, ftol, abstol);
 +}
 +
 +static void cmp_inputrec(FILE *fp, t_inputrec *ir1, t_inputrec *ir2, real ftol, real abstol)
 +{
 +    fprintf(fp, "comparing inputrec\n");
 +
 +    /* gcc 2.96 doesnt like these defines at all, but issues a huge list
 +     * of warnings. Maybe it will change in future versions, but for the
 +     * moment I've spelled them out instead. /EL 000820
 +     * #define CIB(s) cmp_int(fp,"inputrec->"#s,0,ir1->##s,ir2->##s)
 +     * #define CII(s) cmp_int(fp,"inputrec->"#s,0,ir1->##s,ir2->##s)
 +     * #define CIR(s) cmp_real(fp,"inputrec->"#s,0,ir1->##s,ir2->##s,ftol)
 +     */
 +    cmp_int(fp, "inputrec->eI", -1, ir1->eI, ir2->eI);
 +    cmp_int64(fp, "inputrec->nsteps", ir1->nsteps, ir2->nsteps);
 +    cmp_int64(fp, "inputrec->init_step", ir1->init_step, ir2->init_step);
 +    cmp_int(fp, "inputrec->simulation_part", -1, ir1->simulation_part, ir2->simulation_part);
 +    cmp_int(fp, "inputrec->ePBC", -1, ir1->ePBC, ir2->ePBC);
 +    cmp_int(fp, "inputrec->bPeriodicMols", -1, ir1->bPeriodicMols, ir2->bPeriodicMols);
 +    cmp_int(fp, "inputrec->cutoff_scheme", -1, ir1->cutoff_scheme, ir2->cutoff_scheme);
 +    cmp_int(fp, "inputrec->ns_type", -1, ir1->ns_type, ir2->ns_type);
 +    cmp_int(fp, "inputrec->nstlist", -1, ir1->nstlist, ir2->nstlist);
 +    cmp_int(fp, "inputrec->nstcomm", -1, ir1->nstcomm, ir2->nstcomm);
 +    cmp_int(fp, "inputrec->comm_mode", -1, ir1->comm_mode, ir2->comm_mode);
 +    cmp_int(fp, "inputrec->nstlog", -1, ir1->nstlog, ir2->nstlog);
 +    cmp_int(fp, "inputrec->nstxout", -1, ir1->nstxout, ir2->nstxout);
 +    cmp_int(fp, "inputrec->nstvout", -1, ir1->nstvout, ir2->nstvout);
 +    cmp_int(fp, "inputrec->nstfout", -1, ir1->nstfout, ir2->nstfout);
 +    cmp_int(fp, "inputrec->nstcalcenergy", -1, ir1->nstcalcenergy, ir2->nstcalcenergy);
 +    cmp_int(fp, "inputrec->nstenergy", -1, ir1->nstenergy, ir2->nstenergy);
 +    cmp_int(fp, "inputrec->nstxout_compressed", -1, ir1->nstxout_compressed, ir2->nstxout_compressed);
 +    cmp_double(fp, "inputrec->init_t", -1, ir1->init_t, ir2->init_t, ftol, abstol);
 +    cmp_double(fp, "inputrec->delta_t", -1, ir1->delta_t, ir2->delta_t, ftol, abstol);
 +    cmp_real(fp, "inputrec->x_compression_precision", -1, ir1->x_compression_precision, ir2->x_compression_precision, ftol, abstol);
 +    cmp_real(fp, "inputrec->fourierspacing", -1, ir1->fourier_spacing, ir2->fourier_spacing, ftol, abstol);
 +    cmp_int(fp, "inputrec->nkx", -1, ir1->nkx, ir2->nkx);
 +    cmp_int(fp, "inputrec->nky", -1, ir1->nky, ir2->nky);
 +    cmp_int(fp, "inputrec->nkz", -1, ir1->nkz, ir2->nkz);
 +    cmp_int(fp, "inputrec->pme_order", -1, ir1->pme_order, ir2->pme_order);
 +    cmp_real(fp, "inputrec->ewald_rtol", -1, ir1->ewald_rtol, ir2->ewald_rtol, ftol, abstol);
 +    cmp_int(fp, "inputrec->ewald_geometry", -1, ir1->ewald_geometry, ir2->ewald_geometry);
 +    cmp_real(fp, "inputrec->epsilon_surface", -1, ir1->epsilon_surface, ir2->epsilon_surface, ftol, abstol);
 +    cmp_int(fp, "inputrec->bContinuation", -1, ir1->bContinuation, ir2->bContinuation);
 +    cmp_int(fp, "inputrec->bShakeSOR", -1, ir1->bShakeSOR, ir2->bShakeSOR);
 +    cmp_int(fp, "inputrec->etc", -1, ir1->etc, ir2->etc);
 +    cmp_int(fp, "inputrec->bPrintNHChains", -1, ir1->bPrintNHChains, ir2->bPrintNHChains);
 +    cmp_int(fp, "inputrec->epc", -1, ir1->epc, ir2->epc);
 +    cmp_int(fp, "inputrec->epct", -1, ir1->epct, ir2->epct);
 +    cmp_real(fp, "inputrec->tau_p", -1, ir1->tau_p, ir2->tau_p, ftol, abstol);
 +    cmp_rvec(fp, "inputrec->ref_p(x)", -1, ir1->ref_p[XX], ir2->ref_p[XX], ftol, abstol);
 +    cmp_rvec(fp, "inputrec->ref_p(y)", -1, ir1->ref_p[YY], ir2->ref_p[YY], ftol, abstol);
 +    cmp_rvec(fp, "inputrec->ref_p(z)", -1, ir1->ref_p[ZZ], ir2->ref_p[ZZ], ftol, abstol);
 +    cmp_rvec(fp, "inputrec->compress(x)", -1, ir1->compress[XX], ir2->compress[XX], ftol, abstol);
 +    cmp_rvec(fp, "inputrec->compress(y)", -1, ir1->compress[YY], ir2->compress[YY], ftol, abstol);
 +    cmp_rvec(fp, "inputrec->compress(z)", -1, ir1->compress[ZZ], ir2->compress[ZZ], ftol, abstol);
 +    cmp_int(fp, "refcoord_scaling", -1, ir1->refcoord_scaling, ir2->refcoord_scaling);
 +    cmp_rvec(fp, "inputrec->posres_com", -1, ir1->posres_com, ir2->posres_com, ftol, abstol);
 +    cmp_rvec(fp, "inputrec->posres_comB", -1, ir1->posres_comB, ir2->posres_comB, ftol, abstol);
 +    cmp_real(fp, "inputrec->verletbuf_tol", -1, ir1->verletbuf_tol, ir2->verletbuf_tol, ftol, abstol);
 +    cmp_real(fp, "inputrec->rlist", -1, ir1->rlist, ir2->rlist, ftol, abstol);
 +    cmp_real(fp, "inputrec->rtpi", -1, ir1->rtpi, ir2->rtpi, ftol, abstol);
 +    cmp_int(fp, "inputrec->coulombtype", -1, ir1->coulombtype, ir2->coulombtype);
 +    cmp_int(fp, "inputrec->coulomb_modifier", -1, ir1->coulomb_modifier, ir2->coulomb_modifier);
 +    cmp_real(fp, "inputrec->rcoulomb_switch", -1, ir1->rcoulomb_switch, ir2->rcoulomb_switch, ftol, abstol);
 +    cmp_real(fp, "inputrec->rcoulomb", -1, ir1->rcoulomb, ir2->rcoulomb, ftol, abstol);
 +    cmp_int(fp, "inputrec->vdwtype", -1, ir1->vdwtype, ir2->vdwtype);
 +    cmp_int(fp, "inputrec->vdw_modifier", -1, ir1->vdw_modifier, ir2->vdw_modifier);  cmp_real(fp, "inputrec->rvdw_switch", -1, ir1->rvdw_switch, ir2->rvdw_switch, ftol, abstol);
 +    cmp_real(fp, "inputrec->rvdw", -1, ir1->rvdw, ir2->rvdw, ftol, abstol);
 +    cmp_real(fp, "inputrec->epsilon_r", -1, ir1->epsilon_r, ir2->epsilon_r, ftol, abstol);
 +    cmp_real(fp, "inputrec->epsilon_rf", -1, ir1->epsilon_rf, ir2->epsilon_rf, ftol, abstol);
 +    cmp_real(fp, "inputrec->tabext", -1, ir1->tabext, ir2->tabext, ftol, abstol);
 +    cmp_int(fp, "inputrec->implicit_solvent", -1, ir1->implicit_solvent, ir2->implicit_solvent);
 +    cmp_int(fp, "inputrec->gb_algorithm", -1, ir1->gb_algorithm, ir2->gb_algorithm);
 +    cmp_int(fp, "inputrec->nstgbradii", -1, ir1->nstgbradii, ir2->nstgbradii);
 +    cmp_real(fp, "inputrec->rgbradii", -1, ir1->rgbradii, ir2->rgbradii, ftol, abstol);
 +    cmp_real(fp, "inputrec->gb_saltconc", -1, ir1->gb_saltconc, ir2->gb_saltconc, ftol, abstol);
 +    cmp_real(fp, "inputrec->gb_epsilon_solvent", -1, ir1->gb_epsilon_solvent, ir2->gb_epsilon_solvent, ftol, abstol);
 +    cmp_real(fp, "inputrec->gb_obc_alpha", -1, ir1->gb_obc_alpha, ir2->gb_obc_alpha, ftol, abstol);
 +    cmp_real(fp, "inputrec->gb_obc_beta", -1, ir1->gb_obc_beta, ir2->gb_obc_beta, ftol, abstol);
 +    cmp_real(fp, "inputrec->gb_obc_gamma", -1, ir1->gb_obc_gamma, ir2->gb_obc_gamma, ftol, abstol);
 +    cmp_real(fp, "inputrec->gb_dielectric_offset", -1, ir1->gb_dielectric_offset, ir2->gb_dielectric_offset, ftol, abstol);
 +    cmp_int(fp, "inputrec->sa_algorithm", -1, ir1->sa_algorithm, ir2->sa_algorithm);
 +    cmp_real(fp, "inputrec->sa_surface_tension", -1, ir1->sa_surface_tension, ir2->sa_surface_tension, ftol, abstol);
 +
 +    cmp_int(fp, "inputrec->eDispCorr", -1, ir1->eDispCorr, ir2->eDispCorr);
 +    cmp_real(fp, "inputrec->shake_tol", -1, ir1->shake_tol, ir2->shake_tol, ftol, abstol);
 +    cmp_int(fp, "inputrec->efep", -1, ir1->efep, ir2->efep);
 +    cmp_fepvals(fp, ir1->fepvals, ir2->fepvals, ftol, abstol);
 +    cmp_int(fp, "inputrec->bSimTemp", -1, ir1->bSimTemp, ir2->bSimTemp);
 +    if ((ir1->bSimTemp == ir2->bSimTemp) && (ir1->bSimTemp))
 +    {
 +        cmp_simtempvals(fp, ir1->simtempvals, ir2->simtempvals, std::min(ir1->fepvals->n_lambda, ir2->fepvals->n_lambda), ftol, abstol);
 +    }
 +    cmp_int(fp, "inputrec->bExpanded", -1, ir1->bExpanded, ir2->bExpanded);
 +    if ((ir1->bExpanded == ir2->bExpanded) && (ir1->bExpanded))
 +    {
 +        cmp_expandedvals(fp, ir1->expandedvals, ir2->expandedvals, std::min(ir1->fepvals->n_lambda, ir2->fepvals->n_lambda), ftol, abstol);
 +    }
 +    cmp_int(fp, "inputrec->nwall", -1, ir1->nwall, ir2->nwall);
 +    cmp_int(fp, "inputrec->wall_type", -1, ir1->wall_type, ir2->wall_type);
 +    cmp_int(fp, "inputrec->wall_atomtype[0]", -1, ir1->wall_atomtype[0], ir2->wall_atomtype[0]);
 +    cmp_int(fp, "inputrec->wall_atomtype[1]", -1, ir1->wall_atomtype[1], ir2->wall_atomtype[1]);
 +    cmp_real(fp, "inputrec->wall_density[0]", -1, ir1->wall_density[0], ir2->wall_density[0], ftol, abstol);
 +    cmp_real(fp, "inputrec->wall_density[1]", -1, ir1->wall_density[1], ir2->wall_density[1], ftol, abstol);
 +    cmp_real(fp, "inputrec->wall_ewald_zfac", -1, ir1->wall_ewald_zfac, ir2->wall_ewald_zfac, ftol, abstol);
 +
 +    cmp_bool(fp, "inputrec->bPull", -1, ir1->bPull, ir2->bPull);
 +    if (ir1->bPull && ir2->bPull)
 +    {
 +        cmp_pull(fp);
 +    }
 +
 +    cmp_int(fp, "inputrec->eDisre", -1, ir1->eDisre, ir2->eDisre);
 +    cmp_real(fp, "inputrec->dr_fc", -1, ir1->dr_fc, ir2->dr_fc, ftol, abstol);
 +    cmp_int(fp, "inputrec->eDisreWeighting", -1, ir1->eDisreWeighting, ir2->eDisreWeighting);
 +    cmp_int(fp, "inputrec->bDisreMixed", -1, ir1->bDisreMixed, ir2->bDisreMixed);
 +    cmp_int(fp, "inputrec->nstdisreout", -1, ir1->nstdisreout, ir2->nstdisreout);
 +    cmp_real(fp, "inputrec->dr_tau", -1, ir1->dr_tau, ir2->dr_tau, ftol, abstol);
 +    cmp_real(fp, "inputrec->orires_fc", -1, ir1->orires_fc, ir2->orires_fc, ftol, abstol);
 +    cmp_real(fp, "inputrec->orires_tau", -1, ir1->orires_tau, ir2->orires_tau, ftol, abstol);
 +    cmp_int(fp, "inputrec->nstorireout", -1, ir1->nstorireout, ir2->nstorireout);
 +    cmp_real(fp, "inputrec->em_stepsize", -1, ir1->em_stepsize, ir2->em_stepsize, ftol, abstol);
 +    cmp_real(fp, "inputrec->em_tol", -1, ir1->em_tol, ir2->em_tol, ftol, abstol);
 +    cmp_int(fp, "inputrec->niter", -1, ir1->niter, ir2->niter);
 +    cmp_real(fp, "inputrec->fc_stepsize", -1, ir1->fc_stepsize, ir2->fc_stepsize, ftol, abstol);
 +    cmp_int(fp, "inputrec->nstcgsteep", -1, ir1->nstcgsteep, ir2->nstcgsteep);
 +    cmp_int(fp, "inputrec->nbfgscorr", 0, ir1->nbfgscorr, ir2->nbfgscorr);
 +    cmp_int(fp, "inputrec->eConstrAlg", -1, ir1->eConstrAlg, ir2->eConstrAlg);
 +    cmp_int(fp, "inputrec->nProjOrder", -1, ir1->nProjOrder, ir2->nProjOrder);
 +    cmp_real(fp, "inputrec->LincsWarnAngle", -1, ir1->LincsWarnAngle, ir2->LincsWarnAngle, ftol, abstol);
 +    cmp_int(fp, "inputrec->nLincsIter", -1, ir1->nLincsIter, ir2->nLincsIter);
 +    cmp_real(fp, "inputrec->bd_fric", -1, ir1->bd_fric, ir2->bd_fric, ftol, abstol);
 +    cmp_int64(fp, "inputrec->ld_seed", ir1->ld_seed, ir2->ld_seed);
 +    cmp_real(fp, "inputrec->cos_accel", -1, ir1->cos_accel, ir2->cos_accel, ftol, abstol);
 +    cmp_rvec(fp, "inputrec->deform(a)", -1, ir1->deform[XX], ir2->deform[XX], ftol, abstol);
 +    cmp_rvec(fp, "inputrec->deform(b)", -1, ir1->deform[YY], ir2->deform[YY], ftol, abstol);
 +    cmp_rvec(fp, "inputrec->deform(c)", -1, ir1->deform[ZZ], ir2->deform[ZZ], ftol, abstol);
 +
 +
 +    cmp_int(fp, "inputrec->userint1", -1, ir1->userint1, ir2->userint1);
 +    cmp_int(fp, "inputrec->userint2", -1, ir1->userint2, ir2->userint2);
 +    cmp_int(fp, "inputrec->userint3", -1, ir1->userint3, ir2->userint3);
 +    cmp_int(fp, "inputrec->userint4", -1, ir1->userint4, ir2->userint4);
 +    cmp_real(fp, "inputrec->userreal1", -1, ir1->userreal1, ir2->userreal1, ftol, abstol);
 +    cmp_real(fp, "inputrec->userreal2", -1, ir1->userreal2, ir2->userreal2, ftol, abstol);
 +    cmp_real(fp, "inputrec->userreal3", -1, ir1->userreal3, ir2->userreal3, ftol, abstol);
 +    cmp_real(fp, "inputrec->userreal4", -1, ir1->userreal4, ir2->userreal4, ftol, abstol);
 +    cmp_grpopts(fp, &(ir1->opts), &(ir2->opts), ftol, abstol);
 +    cmp_cosines(fp, "ex", ir1->ex, ir2->ex, ftol, abstol);
 +    cmp_cosines(fp, "et", ir1->et, ir2->et, ftol, abstol);
 +}
 +
 +static void comp_pull_AB(FILE *fp, pull_params_t *pull, real ftol, real abstol)
 +{
 +    int i;
 +
 +    for (i = 0; i < pull->ncoord; i++)
 +    {
 +        fprintf(fp, "comparing pull coord %d\n", i);
 +        cmp_real(fp, "pull-coord->k", -1, pull->coord[i].k, pull->coord[i].kB, ftol, abstol);
 +    }
 +}
 +
 +static void comp_state(t_state *st1, t_state *st2,
 +                       gmx_bool bRMSD, real ftol, real abstol)
 +{
 +    int i, j, nc;
 +
 +    fprintf(stdout, "comparing flags\n");
 +    cmp_int(stdout, "flags", -1, st1->flags, st2->flags);
 +    fprintf(stdout, "comparing box\n");
 +    cmp_rvecs(stdout, "box", DIM, st1->box, st2->box, FALSE, ftol, abstol);
 +    fprintf(stdout, "comparing box_rel\n");
 +    cmp_rvecs(stdout, "box_rel", DIM, st1->box_rel, st2->box_rel, FALSE, ftol, abstol);
 +    fprintf(stdout, "comparing boxv\n");
 +    cmp_rvecs(stdout, "boxv", DIM, st1->boxv, st2->boxv, FALSE, ftol, abstol);
 +    if (st1->flags & (1<<estSVIR_PREV))
 +    {
 +        fprintf(stdout, "comparing shake vir_prev\n");
 +        cmp_rvecs(stdout, "svir_prev", DIM, st1->svir_prev, st2->svir_prev, FALSE, ftol, abstol);
 +    }
 +    if (st1->flags & (1<<estFVIR_PREV))
 +    {
 +        fprintf(stdout, "comparing force vir_prev\n");
 +        cmp_rvecs(stdout, "fvir_prev", DIM, st1->fvir_prev, st2->fvir_prev, FALSE, ftol, abstol);
 +    }
 +    if (st1->flags & (1<<estPRES_PREV))
 +    {
 +        fprintf(stdout, "comparing prev_pres\n");
 +        cmp_rvecs(stdout, "pres_prev", DIM, st1->pres_prev, st2->pres_prev, FALSE, ftol, abstol);
 +    }
 +    cmp_int(stdout, "ngtc", -1, st1->ngtc, st2->ngtc);
 +    cmp_int(stdout, "nhchainlength", -1, st1->nhchainlength, st2->nhchainlength);
 +    if (st1->ngtc == st2->ngtc && st1->nhchainlength == st2->nhchainlength)
 +    {
 +        for (i = 0; i < st1->ngtc; i++)
 +        {
 +            nc = i*st1->nhchainlength;
 +            for (j = 0; j < nc; j++)
 +            {
 +                cmp_real(stdout, "nosehoover_xi",
 +                         i, st1->nosehoover_xi[nc+j], st2->nosehoover_xi[nc+j], ftol, abstol);
 +            }
 +        }
 +    }
 +    cmp_int(stdout, "nnhpres", -1, st1->nnhpres, st2->nnhpres);
 +    if (st1->nnhpres == st2->nnhpres && st1->nhchainlength == st2->nhchainlength)
 +    {
 +        for (i = 0; i < st1->nnhpres; i++)
 +        {
 +            nc = i*st1->nhchainlength;
 +            for (j = 0; j < nc; j++)
 +            {
 +                cmp_real(stdout, "nosehoover_xi",
 +                         i, st1->nhpres_xi[nc+j], st2->nhpres_xi[nc+j], ftol, abstol);
 +            }
 +        }
 +    }
 +
 +    cmp_int(stdout, "natoms", -1, st1->natoms, st2->natoms);
 +    if (st1->natoms == st2->natoms)
 +    {
 +        if ((st1->flags & (1<<estX)) && (st2->flags & (1<<estX)))
 +        {
 +            fprintf(stdout, "comparing x\n");
 +            cmp_rvecs(stdout, "x", st1->natoms, st1->x, st2->x, bRMSD, ftol, abstol);
 +        }
 +        if ((st1->flags & (1<<estV)) && (st2->flags & (1<<estV)))
 +        {
 +            fprintf(stdout, "comparing v\n");
 +            cmp_rvecs(stdout, "v", st1->natoms, st1->v, st2->v, bRMSD, ftol, abstol);
 +        }
 +    }
 +}
 +
 +void comp_tpx(const char *fn1, const char *fn2,
 +              gmx_bool bRMSD, real ftol, real abstol)
 +{
 +    const char  *ff[2];
 +    t_inputrec   ir[2];
 +    t_state      state[2];
 +    gmx_mtop_t   mtop[2];
 +    t_topology   top[2];
 +    int          i;
 +
 +    ff[0] = fn1;
 +    ff[1] = fn2;
 +    for (i = 0; i < (fn2 ? 2 : 1); i++)
 +    {
 +        read_tpx_state(ff[i], &(ir[i]), &state[i], &(mtop[i]));
 +    }
 +    if (fn2)
 +    {
 +        cmp_inputrec(stdout, &ir[0], &ir[1], ftol, abstol);
 +        /* Convert gmx_mtop_t to t_topology.
 +         * We should implement direct mtop comparison,
 +         * but it might be useful to keep t_topology comparison as an option.
 +         */
 +        top[0] = gmx_mtop_t_to_t_topology(&mtop[0]);
 +        top[1] = gmx_mtop_t_to_t_topology(&mtop[1]);
 +        cmp_top(stdout, &top[0], &top[1], ftol, abstol);
 +        cmp_groups(stdout, &mtop[0].groups, &mtop[1].groups,
 +                   mtop[0].natoms, mtop[1].natoms);
 +        comp_state(&state[0], &state[1], bRMSD, ftol, abstol);
 +    }
 +    else
 +    {
 +        if (ir[0].efep == efepNO)
 +        {
 +            fprintf(stdout, "inputrec->efep = %s\n", efep_names[ir[0].efep]);
 +        }
 +        else
 +        {
 +            if (ir[0].bPull)
 +            {
 +                comp_pull_AB(stdout, ir->pull, ftol, abstol);
 +            }
 +            /* Convert gmx_mtop_t to t_topology.
 +             * We should implement direct mtop comparison,
 +             * but it might be useful to keep t_topology comparison as an option.
 +             */
 +            top[0] = gmx_mtop_t_to_t_topology(&mtop[0]);
 +            cmp_top(stdout, &top[0], NULL, ftol, abstol);
 +        }
 +    }
 +}
 +
 +void comp_frame(FILE *fp, t_trxframe *fr1, t_trxframe *fr2,
 +                gmx_bool bRMSD, real ftol, real abstol)
 +{
 +    fprintf(fp, "\n");
 +    cmp_int(fp, "not_ok", -1, fr1->not_ok, fr2->not_ok);
 +    cmp_int(fp, "natoms", -1, fr1->natoms, fr2->natoms);
 +    if (cmp_bool(fp, "bTitle", -1, fr1->bTitle, fr2->bTitle))
 +    {
 +        cmp_str(fp, "title", -1, fr1->title, fr2->title);
 +    }
 +    if (cmp_bool(fp, "bStep", -1, fr1->bStep, fr2->bStep))
 +    {
 +        cmp_int(fp, "step", -1, fr1->step, fr2->step);
 +    }
 +    cmp_int(fp, "step", -1, fr1->step, fr2->step);
 +    if (cmp_bool(fp, "bTime", -1, fr1->bTime, fr2->bTime))
 +    {
 +        cmp_real(fp, "time", -1, fr1->time, fr2->time, ftol, abstol);
 +    }
 +    if (cmp_bool(fp, "bLambda", -1, fr1->bLambda, fr2->bLambda))
 +    {
 +        cmp_real(fp, "lambda", -1, fr1->lambda, fr2->lambda, ftol, abstol);
 +    }
 +    if (cmp_bool(fp, "bAtoms", -1, fr1->bAtoms, fr2->bAtoms))
 +    {
 +        cmp_atoms(fp, fr1->atoms, fr2->atoms, ftol, abstol);
 +    }
 +    if (cmp_bool(fp, "bPrec", -1, fr1->bPrec, fr2->bPrec))
 +    {
 +        cmp_real(fp, "prec", -1, fr1->prec, fr2->prec, ftol, abstol);
 +    }
 +    if (cmp_bool(fp, "bX", -1, fr1->bX, fr2->bX))
 +    {
 +        cmp_rvecs(fp, "x", std::min(fr1->natoms, fr2->natoms), fr1->x, fr2->x, bRMSD, ftol, abstol);
 +    }
 +    if (cmp_bool(fp, "bV", -1, fr1->bV, fr2->bV))
 +    {
 +        cmp_rvecs(fp, "v", std::min(fr1->natoms, fr2->natoms), fr1->v, fr2->v, bRMSD, ftol, abstol);
 +    }
 +    if (cmp_bool(fp, "bF", -1, fr1->bF, fr2->bF))
 +    {
 +        cmp_rvecs(fp, "f", std::min(fr1->natoms, fr2->natoms), fr1->f, fr2->f, bRMSD, ftol, abstol);
 +    }
 +    if (cmp_bool(fp, "bBox", -1, fr1->bBox, fr2->bBox))
 +    {
 +        cmp_rvecs(fp, "box", 3, fr1->box, fr2->box, FALSE, ftol, abstol);
 +    }
 +}
 +
 +void comp_trx(const gmx_output_env_t *oenv, const char *fn1, const char *fn2,
 +              gmx_bool bRMSD, real ftol, real abstol)
 +{
 +    int          i;
 +    const char  *fn[2];
 +    t_trxframe   fr[2];
 +    t_trxstatus *status[2];
 +    gmx_bool     b[2];
 +
 +    fn[0] = fn1;
 +    fn[1] = fn2;
 +    fprintf(stderr, "Comparing trajectory files %s and %s\n", fn1, fn2);
 +    for (i = 0; i < 2; i++)
 +    {
 +        b[i] = read_first_frame(oenv, &status[i], fn[i], &fr[i], TRX_READ_X|TRX_READ_V|TRX_READ_F);
 +    }
 +
 +    if (b[0] && b[1])
 +    {
 +        do
 +        {
 +            comp_frame(stdout, &(fr[0]), &(fr[1]), bRMSD, ftol, abstol);
 +
 +            for (i = 0; i < 2; i++)
 +            {
 +                b[i] = read_next_frame(oenv, status[i], &fr[i]);
 +            }
 +        }
 +        while (b[0] && b[1]);
 +
 +        for (i = 0; i < 2; i++)
 +        {
 +            if (b[i] && !b[1-i])
 +            {
 +                fprintf(stdout, "\nEnd of file on %s but not on %s\n", fn[1-i], fn[i]);
 +            }
 +            close_trj(status[i]);
 +        }
 +    }
 +    if (!b[0] && !b[1])
 +    {
 +        fprintf(stdout, "\nBoth files read correctly\n");
 +    }
 +}
 +
 +static real ener_tensor_diag(int n, int *ind1, int *ind2,
 +                             gmx_enxnm_t *enm1,
 +                             int *tensi, int i,
 +                             t_energy e1[], t_energy e2[])
 +{
 +    int    d1, d2;
 +    int    j;
 +    real   prod1, prod2;
 +    int    nfound;
 +    size_t len;
 +
 +    d1 = tensi[i]/DIM;
 +    d2 = tensi[i] - d1*DIM;
 +
 +    /* Find the diagonal elements d1 and d2 */
 +    len    = std::strlen(enm1[ind1[i]].name);
 +    prod1  = 1;
 +    prod2  = 1;
 +    nfound = 0;
 +    for (j = 0; j < n; j++)
 +    {
 +        if (tensi[j] >= 0 &&
 +            std::strlen(enm1[ind1[j]].name) == len &&
 +            std::strncmp(enm1[ind1[i]].name, enm1[ind1[j]].name, len-2) == 0 &&
 +            (tensi[j] == d1*DIM+d1 || tensi[j] == d2*DIM+d2))
 +        {
 +            prod1 *= fabs(e1[ind1[j]].e);
 +            prod2 *= fabs(e2[ind2[j]].e);
 +            nfound++;
 +        }
 +    }
 +
 +    if (nfound == 2)
 +    {
 +        return 0.5*(std::sqrt(prod1) + std::sqrt(prod2));
 +    }
 +    else
 +    {
 +        return 0;
 +    }
 +}
 +
 +static gmx_bool enernm_equal(const char *nm1, const char *nm2)
 +{
 +    int len1, len2;
 +
 +    len1 = std::strlen(nm1);
 +    len2 = std::strlen(nm2);
 +
 +    /* Remove " (bar)" at the end of a name */
 +    if (len1 > 6 && std::strcmp(nm1+len1-6, " (bar)") == 0)
 +    {
 +        len1 -= 6;
 +    }
 +    if (len2 > 6 && std::strcmp(nm2+len2-6, " (bar)") == 0)
 +    {
 +        len2 -= 6;
 +    }
 +
 +    return (len1 == len2 && gmx_strncasecmp(nm1, nm2, len1) == 0);
 +}
 +
 +static void cmp_energies(FILE *fp, int step1, int step2,
 +                         t_energy e1[], t_energy e2[],
 +                         gmx_enxnm_t *enm1,
 +                         real ftol, real abstol,
 +                         int nre, int *ind1, int *ind2, int maxener)
 +{
 +    int   i, ii;
 +    int  *tensi, len, d1, d2;
 +    real  ftol_i, abstol_i;
 +
 +    snew(tensi, maxener);
 +    /* Check for tensor elements ending on "-XX", "-XY", ... , "-ZZ" */
 +    for (i = 0; (i < maxener); i++)
 +    {
 +        ii       = ind1[i];
 +        tensi[i] = -1;
 +        len      = std::strlen(enm1[ii].name);
 +        if (len > 3 && enm1[ii].name[len-3] == '-')
 +        {
 +            d1 = enm1[ii].name[len-2] - 'X';
 +            d2 = enm1[ii].name[len-1] - 'X';
 +            if (d1 >= 0 && d1 < DIM &&
 +                d2 >= 0 && d2 < DIM)
 +            {
 +                tensi[i] = d1*DIM + d2;
 +            }
 +        }
 +    }
 +
 +    for (i = 0; (i < maxener); i++)
 +    {
 +        /* Check if this is an off-diagonal tensor element */
 +        if (tensi[i] >= 0 && tensi[i] != 0 && tensi[i] != 4 && tensi[i] != 8)
 +        {
 +            /* Turn on the relative tolerance check (4 is maximum relative diff.) */
 +            ftol_i = 5;
 +            /* Do the relative tolerance through an absolute tolerance times
 +             * the size of diagonal components of the tensor.
 +             */
 +            abstol_i = ftol*ener_tensor_diag(nre, ind1, ind2, enm1, tensi, i, e1, e2);
 +            if (debug)
 +            {
 +                fprintf(debug, "tensor '%s' val %f diag %f\n",
 +                        enm1[i].name, e1[i].e, abstol_i/ftol);
 +            }
 +            if (abstol_i > 0)
 +            {
 +                /* We found a diagonal, we need to check with the minimum tolerance */
 +                abstol_i = std::min(abstol_i, abstol);
 +            }
 +            else
 +            {
 +                /* We did not find a diagonal, ignore the relative tolerance check */
 +                abstol_i = abstol;
 +            }
 +        }
 +        else
 +        {
 +            ftol_i   = ftol;
 +            abstol_i = abstol;
 +        }
 +        if (!equal_real(e1[ind1[i]].e, e2[ind2[i]].e, ftol_i, abstol_i))
 +        {
 +            fprintf(fp, "%-15s  step %3d:  %12g,  step %3d: %12g\n",
 +                    enm1[ind1[i]].name,
 +                    step1, e1[ind1[i]].e,
 +                    step2, e2[ind2[i]].e);
 +        }
 +    }
 +
 +    sfree(tensi);
 +}
 +
 +#if 0
 +static void cmp_disres(t_enxframe *fr1, t_enxframe *fr2, real ftol, real abstol)
 +{
 +    int  i;
 +    char bav[64], bt[64], bs[22];
 +
 +    cmp_int(stdout, "ndisre", -1, fr1->ndisre, fr2->ndisre);
 +    if ((fr1->ndisre == fr2->ndisre) && (fr1->ndisre > 0))
 +    {
 +        sprintf(bav, "step %s: disre rav", gmx_step_str(fr1->step, bs));
 +        sprintf(bt, "step %s: disre  rt", gmx_step_str(fr1->step, bs));
 +        for (i = 0; (i < fr1->ndisre); i++)
 +        {
 +            cmp_real(stdout, bav, i, fr1->disre_rm3tav[i], fr2->disre_rm3tav[i], ftol, abstol);
 +            cmp_real(stdout, bt, i, fr1->disre_rt[i], fr2->disre_rt[i], ftol, abstol);
 +        }
 +    }
 +}
 +#endif
 +
 +static void cmp_eblocks(t_enxframe *fr1, t_enxframe *fr2, real ftol, real abstol)
 +{
 +    int  i, j, k;
 +    char buf[64], bs[22];
 +
 +    cmp_int(stdout, "nblock", -1, fr1->nblock, fr2->nblock);
 +    if ((fr1->nblock == fr2->nblock) && (fr1->nblock > 0))
 +    {
 +        for (j = 0; (j < fr1->nblock); j++)
 +        {
 +            t_enxblock *b1, *b2; /* convenience vars */
 +
 +            b1 = &(fr1->block[j]);
 +            b2 = &(fr2->block[j]);
 +
 +            sprintf(buf, "step %s: block[%d]", gmx_step_str(fr1->step, bs), j);
 +            cmp_int(stdout, buf, -1, b1->nsub, b2->nsub);
 +            cmp_int(stdout, buf, -1, b1->id, b2->id);
 +
 +            if ( (b1->nsub == b2->nsub) && (b1->id == b2->id) )
 +            {
 +                for (i = 0; i < b1->nsub; i++)
 +                {
 +                    t_enxsubblock *s1, *s2;
 +
 +                    s1 = &(b1->sub[i]);
 +                    s2 = &(b2->sub[i]);
 +
 +                    cmp_int(stdout, buf, -1, (int)s1->type, (int)s2->type);
 +                    cmp_int64(stdout, buf, s1->nr, s2->nr);
 +
 +                    if ((s1->type == s2->type) && (s1->nr == s2->nr))
 +                    {
 +                        switch (s1->type)
 +                        {
 +                            case xdr_datatype_float:
 +                                for (k = 0; k < s1->nr; k++)
 +                                {
 +                                    cmp_float(stdout, buf, i,
 +                                              s1->fval[k], s2->fval[k],
 +                                              ftol, abstol);
 +                                }
 +                                break;
 +                            case xdr_datatype_double:
 +                                for (k = 0; k < s1->nr; k++)
 +                                {
 +                                    cmp_double(stdout, buf, i,
 +                                               s1->dval[k], s2->dval[k],
 +                                               ftol, abstol);
 +                                }
 +                                break;
 +                            case xdr_datatype_int:
 +                                for (k = 0; k < s1->nr; k++)
 +                                {
 +                                    cmp_int(stdout, buf, i,
 +                                            s1->ival[k], s2->ival[k]);
 +                                }
 +                                break;
 +                            case xdr_datatype_int64:
 +                                for (k = 0; k < s1->nr; k++)
 +                                {
 +                                    cmp_int64(stdout, buf,
 +                                              s1->lval[k], s2->lval[k]);
 +                                }
 +                                break;
 +                            case xdr_datatype_char:
 +                                for (k = 0; k < s1->nr; k++)
 +                                {
 +                                    cmp_uc(stdout, buf, i,
 +                                           s1->cval[k], s2->cval[k]);
 +                                }
 +                                break;
 +                            case xdr_datatype_string:
 +                                for (k = 0; k < s1->nr; k++)
 +                                {
 +                                    cmp_str(stdout, buf, i,
 +                                            s1->sval[k], s2->sval[k]);
 +                                }
 +                                break;
 +                            default:
 +                                gmx_incons("Unknown data type!!");
 +                        }
 +                    }
 +                }
 +            }
 +        }
 +    }
 +}
 +
 +void comp_enx(const char *fn1, const char *fn2, real ftol, real abstol, const char *lastener)
 +{
 +    int            nre, nre1, nre2;
 +    ener_file_t    in1, in2;
 +    int            i, j, maxener, *ind1, *ind2, *have;
 +    gmx_enxnm_t   *enm1 = NULL, *enm2 = NULL;
 +    t_enxframe    *fr1, *fr2;
 +    gmx_bool       b1, b2;
 +
 +    fprintf(stdout, "comparing energy file %s and %s\n\n", fn1, fn2);
 +
 +    in1 = open_enx(fn1, "r");
 +    in2 = open_enx(fn2, "r");
 +    do_enxnms(in1, &nre1, &enm1);
 +    do_enxnms(in2, &nre2, &enm2);
 +    if (nre1 != nre2)
 +    {
 +        fprintf(stdout, "There are %d and %d terms in the energy files\n\n",
 +                nre1, nre2);
 +    }
 +    else
 +    {
 +        fprintf(stdout, "There are %d terms in the energy files\n\n", nre1);
 +    }
 +
 +    snew(ind1, nre1);
 +    snew(ind2, nre2);
 +    snew(have, nre2);
 +    nre = 0;
 +    for (i = 0; i < nre1; i++)
 +    {
 +        for (j = 0; j < nre2; j++)
 +        {
 +            if (enernm_equal(enm1[i].name, enm2[j].name))
 +            {
 +                ind1[nre] = i;
 +                ind2[nre] = j;
 +                have[j]   = 1;
 +                nre++;
 +                break;
 +            }
 +        }
 +        if (nre == 0 || ind1[nre-1] != i)
 +        {
 +            cmp_str(stdout, "enm", i, enm1[i].name, "-");
 +        }
 +    }
 +    for (i = 0; i < nre2; i++)
 +    {
 +        if (have[i] == 0)
 +        {
 +            cmp_str(stdout, "enm", i, "-", enm2[i].name);
 +        }
 +    }
 +
 +    maxener = nre;
 +    for (i = 0; i < nre; i++)
 +    {
 +        if ((lastener != NULL) && (std::strstr(enm1[i].name, lastener) != NULL))
 +        {
 +            maxener = i+1;
 +            break;
 +        }
 +    }
 +
 +    fprintf(stdout, "There are %d terms to compare in the energy files\n\n",
 +            maxener);
 +
 +    for (i = 0; i < maxener; i++)
 +    {
 +        cmp_str(stdout, "unit", i, enm1[ind1[i]].unit, enm2[ind2[i]].unit);
 +    }
 +
 +    snew(fr1, 1);
 +    snew(fr2, 1);
 +    do
 +    {
 +        b1 = do_enx(in1, fr1);
 +        b2 = do_enx(in2, fr2);
 +        if (b1 && !b2)
 +        {
 +            fprintf(stdout, "\nEnd of file on %s but not on %s\n", fn2, fn1);
 +        }
 +        else if (!b1 && b2)
 +        {
 +            fprintf(stdout, "\nEnd of file on %s but not on %s\n", fn1, fn2);
 +        }
 +        else if (!b1 && !b2)
 +        {
 +            fprintf(stdout, "\nFiles read successfully\n");
 +        }
 +        else
 +        {
 +            cmp_real(stdout, "t", -1, fr1->t, fr2->t, ftol, abstol);
 +            cmp_int(stdout, "step", -1, fr1->step, fr2->step);
 +            /* We don't want to print the nre mismatch for every frame */
 +            /* cmp_int(stdout,"nre",-1,fr1->nre,fr2->nre); */
 +            if ((fr1->nre >= nre) && (fr2->nre >= nre))
 +            {
 +                cmp_energies(stdout, fr1->step, fr1->step, fr1->ener, fr2->ener,
 +                             enm1, ftol, abstol, nre, ind1, ind2, maxener);
 +            }
 +            /*cmp_disres(fr1,fr2,ftol,abstol);*/
 +            cmp_eblocks(fr1, fr2, ftol, abstol);
 +        }
 +    }
 +    while (b1 && b2);
 +
 +    close_enx(in1);
 +    close_enx(in2);
 +
 +    free_enxframe(fr2);
 +    sfree(fr2);
 +    free_enxframe(fr1);
 +    sfree(fr1);
 +}