minimize.cpp.preplumed

/*
 * 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) 2013,2014,2015,2016,2017,2018, 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.
 */
/*! \internal \file
 *
 * \brief This file defines integrators for energy minimization
 *
 * \author Berk Hess <hess@kth.se>
 * \author Erik Lindahl <erik@kth.se>
 * \ingroup module_mdlib
 */
#include "gmxpre.h"

#include "minimize.h"

#include "config.h"

#include <cmath>
#include <cstring>
#include <ctime>

#include <algorithm>
#include <vector>

#include "gromacs/commandline/filenm.h"
#include "gromacs/domdec/domdec.h"
#include "gromacs/domdec/domdec_struct.h"
#include "gromacs/ewald/pme.h"
#include "gromacs/fileio/confio.h"
#include "gromacs/fileio/mtxio.h"
#include "gromacs/gmxlib/network.h"
#include "gromacs/gmxlib/nrnb.h"
#include "gromacs/imd/imd.h"
#include "gromacs/linearalgebra/sparsematrix.h"
#include "gromacs/listed-forces/manage-threading.h"
#include "gromacs/math/functions.h"
#include "gromacs/math/vec.h"
#include "gromacs/mdlib/constr.h"
#include "gromacs/mdlib/force.h"
#include "gromacs/mdlib/forcerec.h"
#include "gromacs/mdlib/gmx_omp_nthreads.h"
#include "gromacs/mdlib/md_support.h"
#include "gromacs/mdlib/mdatoms.h"
#include "gromacs/mdlib/mdebin.h"
#include "gromacs/mdlib/mdrun.h"
#include "gromacs/mdlib/mdsetup.h"
#include "gromacs/mdlib/ns.h"
#include "gromacs/mdlib/shellfc.h"
#include "gromacs/mdlib/sim_util.h"
#include "gromacs/mdlib/tgroup.h"
#include "gromacs/mdlib/trajectory_writing.h"
#include "gromacs/mdlib/update.h"
#include "gromacs/mdlib/vsite.h"
#include "gromacs/mdtypes/commrec.h"
#include "gromacs/mdtypes/inputrec.h"
#include "gromacs/mdtypes/md_enums.h"
#include "gromacs/mdtypes/state.h"
#include "gromacs/pbcutil/mshift.h"
#include "gromacs/pbcutil/pbc.h"
#include "gromacs/timing/wallcycle.h"
#include "gromacs/timing/walltime_accounting.h"
#include "gromacs/topology/mtop_util.h"
#include "gromacs/topology/topology.h"
#include "gromacs/utility/cstringutil.h"
#include "gromacs/utility/exceptions.h"
#include "gromacs/utility/fatalerror.h"
#include "gromacs/utility/logger.h"
#include "gromacs/utility/smalloc.h"

//! Utility structure for manipulating states during EM
typedef struct {
    //! Copy of the global state
    t_state          s;
    //! Force array
    PaddedRVecVector f;
    //! Potential energy
    real             epot;
    //! Norm of the force
    real             fnorm;
    //! Maximum force
    real             fmax;
    //! Direction
    int              a_fmax;
} em_state_t;

//! Print the EM starting conditions
static void print_em_start(FILE                     *fplog,
                           t_commrec                *cr,
                           gmx_walltime_accounting_t walltime_accounting,
                           gmx_wallcycle_t           wcycle,
                           const char               *name)
{
    walltime_accounting_start(walltime_accounting);
    wallcycle_start(wcycle, ewcRUN);
    print_start(fplog, cr, walltime_accounting, name);
}

//! Stop counting time for EM
static void em_time_end(gmx_walltime_accounting_t walltime_accounting,
                        gmx_wallcycle_t           wcycle)
{
    wallcycle_stop(wcycle, ewcRUN);

    walltime_accounting_end(walltime_accounting);
}

//! Printing a log file and console header
static void sp_header(FILE *out, const char *minimizer, real ftol, int nsteps)
{
    fprintf(out, "\n");
    fprintf(out, "%s:\n", minimizer);
    fprintf(out, "   Tolerance (Fmax)   = %12.5e\n", ftol);
    fprintf(out, "   Number of steps    = %12d\n", nsteps);
}

//! Print warning message
static void warn_step(FILE *fp, real ftol, gmx_bool bLastStep, gmx_bool bConstrain)
{
    char buffer[2048];
    if (bLastStep)
    {
        sprintf(buffer,
                "\nEnergy minimization reached the maximum number "
                "of steps before the forces reached the requested "
                "precision Fmax < %g.\n", ftol);
    }
    else
    {
        sprintf(buffer,
                "\nEnergy minimization has stopped, but the forces have "
                "not converged to the requested precision Fmax < %g (which "
                "may not be possible for your system). It stopped "
                "because the algorithm tried to make a new step whose size "
                "was too small, or there was no change in the energy since "
                "last step. Either way, we regard the minimization as "
                "converged to within the available machine precision, "
                "given your starting configuration and EM parameters.\n%s%s",
                ftol,
                sizeof(real) < sizeof(double) ?
                "\nDouble precision normally gives you higher accuracy, but "
                "this is often not needed for preparing to run molecular "
                "dynamics.\n" :
                "",
                bConstrain ?
                "You might need to increase your constraint accuracy, or turn\n"
                "off constraints altogether (set constraints = none in mdp file)\n" :
                "");
    }
    fputs(wrap_lines(buffer, 78, 0, FALSE), fp);
}

//! Print message about convergence of the EM
static void print_converged(FILE *fp, const char *alg, real ftol,
                            gmx_int64_t count, gmx_bool bDone, gmx_int64_t nsteps,
                            const em_state_t *ems, double sqrtNumAtoms)
{
    char buf[STEPSTRSIZE];

    if (bDone)
    {
        fprintf(fp, "\n%s converged to Fmax < %g in %s steps\n",
                alg, ftol, gmx_step_str(count, buf));
    }
    else if (count < nsteps)
    {
        fprintf(fp, "\n%s converged to machine precision in %s steps,\n"
                "but did not reach the requested Fmax < %g.\n",
                alg, gmx_step_str(count, buf), ftol);
    }
    else
    {
        fprintf(fp, "\n%s did not converge to Fmax < %g in %s steps.\n",
                alg, ftol, gmx_step_str(count, buf));
    }

#if GMX_DOUBLE
    fprintf(fp, "Potential Energy  = %21.14e\n", ems->epot);
    fprintf(fp, "Maximum force     = %21.14e on atom %d\n", ems->fmax, ems->a_fmax + 1);
    fprintf(fp, "Norm of force     = %21.14e\n", ems->fnorm/sqrtNumAtoms);
#else
    fprintf(fp, "Potential Energy  = %14.7e\n", ems->epot);
    fprintf(fp, "Maximum force     = %14.7e on atom %d\n", ems->fmax, ems->a_fmax + 1);
    fprintf(fp, "Norm of force     = %14.7e\n", ems->fnorm/sqrtNumAtoms);
#endif
}

//! Compute the norm and max of the force array in parallel
static void get_f_norm_max(t_commrec *cr,
                           t_grpopts *opts, t_mdatoms *mdatoms, const rvec *f,
                           real *fnorm, real *fmax, int *a_fmax)
{
    double fnorm2, *sum;
    real   fmax2, fam;
    int    la_max, a_max, start, end, i, m, gf;

    /* This routine finds the largest force and returns it.
     * On parallel machines the global max is taken.
     */
    fnorm2 = 0;
    fmax2  = 0;
    la_max = -1;
    start  = 0;
    end    = mdatoms->homenr;
    if (mdatoms->cFREEZE)
    {
        for (i = start; i < end; i++)
        {
            gf  = mdatoms->cFREEZE[i];
            fam = 0;
            for (m = 0; m < DIM; m++)
            {
                if (!opts->nFreeze[gf][m])
                {
                    fam += gmx::square(f[i][m]);
                }
            }
            fnorm2 += fam;
            if (fam > fmax2)
            {
                fmax2  = fam;
                la_max = i;
            }
        }
    }
    else
    {
        for (i = start; i < end; i++)
        {
            fam     = norm2(f[i]);
            fnorm2 += fam;
            if (fam > fmax2)
            {
                fmax2  = fam;
                la_max = i;
            }
        }
    }

    if (la_max >= 0 && DOMAINDECOMP(cr))
    {
        a_max = cr->dd->gatindex[la_max];
    }
    else
    {
        a_max = la_max;
    }
    if (PAR(cr))
    {
        snew(sum, 2*cr->nnodes+1);
        sum[2*cr->nodeid]   = fmax2;
        sum[2*cr->nodeid+1] = a_max;
        sum[2*cr->nnodes]   = fnorm2;
        gmx_sumd(2*cr->nnodes+1, sum, cr);
        fnorm2 = sum[2*cr->nnodes];
        /* Determine the global maximum */
        for (i = 0; i < cr->nnodes; i++)
        {
            if (sum[2*i] > fmax2)
            {
                fmax2 = sum[2*i];
                a_max = (int)(sum[2*i+1] + 0.5);
            }
        }
        sfree(sum);
    }

    if (fnorm)
    {
        *fnorm = sqrt(fnorm2);
    }
    if (fmax)
    {
        *fmax  = sqrt(fmax2);
    }
    if (a_fmax)
    {
        *a_fmax = a_max;
    }
}

//! Compute the norm of the force
static void get_state_f_norm_max(t_commrec *cr,
                                 t_grpopts *opts, t_mdatoms *mdatoms,
                                 em_state_t *ems)
{
    get_f_norm_max(cr, opts, mdatoms, as_rvec_array(ems->f.data()),
                   &ems->fnorm, &ems->fmax, &ems->a_fmax);
}

//! Initialize the energy minimization
static void init_em(FILE *fplog, const char *title,
                    t_commrec *cr, gmx::IMDOutputProvider *outputProvider,
                    t_inputrec *ir,
                    const MdrunOptions &mdrunOptions,
                    t_state *state_global, gmx_mtop_t *top_global,
                    em_state_t *ems, gmx_localtop_t **top,
                    t_nrnb *nrnb, rvec mu_tot,
                    t_forcerec *fr, gmx_enerdata_t **enerd,
                    t_graph **graph, gmx::MDAtoms *mdAtoms, gmx_global_stat_t *gstat,
                    gmx_vsite_t *vsite, gmx_constr_t constr, gmx_shellfc_t **shellfc,
                    int nfile, const t_filenm fnm[],
                    gmx_mdoutf_t *outf, t_mdebin **mdebin,
                    gmx_wallcycle_t wcycle)
{
    real dvdl_constr;

    if (fplog)
    {
        fprintf(fplog, "Initiating %s\n", title);
    }

    if (MASTER(cr))
    {
        state_global->ngtc = 0;

        /* Initialize lambda variables */
        initialize_lambdas(fplog, ir, &(state_global->fep_state), state_global->lambda, nullptr);
    }

    init_nrnb(nrnb);

    /* Interactive molecular dynamics */
    init_IMD(ir, cr, top_global, fplog, 1,
             MASTER(cr) ? as_rvec_array(state_global->x.data()) : nullptr,
             nfile, fnm, nullptr, mdrunOptions);

    if (ir->eI == eiNM)
    {
        GMX_ASSERT(shellfc != nullptr, "With NM we always support shells");

        *shellfc = init_shell_flexcon(stdout,
                                      top_global,
                                      n_flexible_constraints(constr),
                                      ir->nstcalcenergy,
                                      DOMAINDECOMP(cr));
    }
    else
    {
        GMX_ASSERT(EI_ENERGY_MINIMIZATION(ir->eI), "This else currently only handles energy minimizers, consider if your algorithm needs shell/flexible-constraint support");

        /* With energy minimization, shells and flexible constraints are
         * automatically minimized when treated like normal DOFS.
         */
        if (shellfc != nullptr)
        {
            *shellfc = nullptr;
        }
    }

    auto mdatoms = mdAtoms->mdatoms();
    if (DOMAINDECOMP(cr))
    {
        *top = dd_init_local_top(top_global);

        dd_init_local_state(cr->dd, state_global, &ems->s);

        /* Distribute the charge groups over the nodes from the master node */
        dd_partition_system(fplog, ir->init_step, cr, TRUE, 1,
                            state_global, top_global, ir,
                            &ems->s, &ems->f, mdAtoms, *top,
                            fr, vsite, constr,
                            nrnb, nullptr, FALSE);
        dd_store_state(cr->dd, &ems->s);

        *graph = nullptr;
    }
    else
    {
        state_change_natoms(state_global, state_global->natoms);
        /* Just copy the state */
        ems->s = *state_global;
        state_change_natoms(&ems->s, ems->s.natoms);
        /* We need to allocate one element extra, since we might use
         * (unaligned) 4-wide SIMD loads to access rvec entries.
         */
        ems->f.resize(gmx::paddedRVecVectorSize(ems->s.natoms));

        snew(*top, 1);
        mdAlgorithmsSetupAtomData(cr, ir, top_global, *top, fr,
                                  graph, mdAtoms,
                                  vsite, shellfc ? *shellfc : nullptr);

        if (vsite)
        {
            set_vsite_top(vsite, *top, mdatoms);
        }
    }

    update_mdatoms(mdAtoms->mdatoms(), ems->s.lambda[efptMASS]);

    if (constr)
    {
        if (ir->eConstrAlg == econtSHAKE &&
            gmx_mtop_ftype_count(top_global, F_CONSTR) > 0)
        {
            gmx_fatal(FARGS, "Can not do energy minimization with %s, use %s\n",
                      econstr_names[econtSHAKE], econstr_names[econtLINCS]);
        }

        if (!DOMAINDECOMP(cr))
        {
            set_constraints(constr, *top, ir, mdatoms, cr);
        }

        if (!ir->bContinuation)
        {
            /* Constrain the starting coordinates */
            dvdl_constr = 0;
            constrain(PAR(cr) ? nullptr : fplog, TRUE, TRUE, constr, &(*top)->idef,
                      ir, cr, -1, 0, 1.0, mdatoms,
                      as_rvec_array(ems->s.x.data()),
                      as_rvec_array(ems->s.x.data()),
                      nullptr,
                      fr->bMolPBC, ems->s.box,
                      ems->s.lambda[efptFEP], &dvdl_constr,
                      nullptr, nullptr, nrnb, econqCoord);
        }
    }

    if (PAR(cr))
    {
        *gstat = global_stat_init(ir);
    }
    else
    {
        *gstat = nullptr;
    }

    *outf = init_mdoutf(fplog, nfile, fnm, mdrunOptions, cr, outputProvider, ir, top_global, nullptr, wcycle);

    snew(*enerd, 1);
    init_enerdata(top_global->groups.grps[egcENER].nr, ir->fepvals->n_lambda,
                  *enerd);

    if (mdebin != nullptr)
    {
        /* Init bin for energy stuff */
        *mdebin = init_mdebin(mdoutf_get_fp_ene(*outf), top_global, ir, nullptr);
    }

    clear_rvec(mu_tot);
    calc_shifts(ems->s.box, fr->shift_vec);
}

//! Finalize the minimization
static void finish_em(t_commrec *cr, gmx_mdoutf_t outf,
                      gmx_walltime_accounting_t walltime_accounting,
                      gmx_wallcycle_t wcycle)
{
    if (!thisRankHasDuty(cr, DUTY_PME))
    {
        /* Tell the PME only node to finish */
        gmx_pme_send_finish(cr);
    }

    done_mdoutf(outf);

    em_time_end(walltime_accounting, wcycle);
}

//! Swap two different EM states during minimization
static void swap_em_state(em_state_t **ems1, em_state_t **ems2)
{
    em_state_t *tmp;

    tmp   = *ems1;
    *ems1 = *ems2;
    *ems2 = tmp;
}

//! Save the EM trajectory
static void write_em_traj(FILE *fplog, t_commrec *cr,
                          gmx_mdoutf_t outf,
                          gmx_bool bX, gmx_bool bF, const char *confout,
                          gmx_mtop_t *top_global,
                          t_inputrec *ir, gmx_int64_t step,
                          em_state_t *state,
                          t_state *state_global,
                          ObservablesHistory *observablesHistory)
{
    int mdof_flags = 0;

    if (bX)
    {
        mdof_flags |= MDOF_X;
    }
    if (bF)
    {
        mdof_flags |= MDOF_F;
    }

    /* If we want IMD output, set appropriate MDOF flag */
    if (ir->bIMD)
    {
        mdof_flags |= MDOF_IMD;
    }

    mdoutf_write_to_trajectory_files(fplog, cr, outf, mdof_flags,
                                     top_global, step, (double)step,
                                     &state->s, state_global, observablesHistory,
                                     state->f);

    if (confout != nullptr)
    {
        if (DOMAINDECOMP(cr))
        {
            /* If bX=true, x was collected to state_global in the call above */
            if (!bX)
            {
                gmx::ArrayRef<gmx::RVec> globalXRef = MASTER(cr) ? gmx::makeArrayRef(state_global->x) : gmx::EmptyArrayRef();
                dd_collect_vec(cr->dd, &state->s, state->s.x, globalXRef);
            }
        }
        else
        {
            /* Copy the local state pointer */
            state_global = &state->s;
        }

        if (MASTER(cr))
        {
            if (ir->ePBC != epbcNONE && !ir->bPeriodicMols && DOMAINDECOMP(cr))
            {
                /* Make molecules whole only for confout writing */
                do_pbc_mtop(fplog, ir->ePBC, state->s.box, top_global,
                            as_rvec_array(state_global->x.data()));
            }

            write_sto_conf_mtop(confout,
                                *top_global->name, top_global,
                                as_rvec_array(state_global->x.data()), nullptr, ir->ePBC, state->s.box);
        }
    }
}

//! \brief Do one minimization step
//
// \returns true when the step succeeded, false when a constraint error occurred
static bool do_em_step(t_commrec *cr, t_inputrec *ir, t_mdatoms *md,
                       gmx_bool bMolPBC,
                       em_state_t *ems1, real a, const PaddedRVecVector *force,
                       em_state_t *ems2,
                       gmx_constr_t constr, gmx_localtop_t *top,
                       t_nrnb *nrnb, gmx_wallcycle_t wcycle,
                       gmx_int64_t count)

{
    t_state *s1, *s2;
    int      start, end;
    real     dvdl_constr;
    int      nthreads gmx_unused;

    bool     validStep = true;

    s1 = &ems1->s;
    s2 = &ems2->s;

    if (DOMAINDECOMP(cr) && s1->ddp_count != cr->dd->ddp_count)
    {
        gmx_incons("state mismatch in do_em_step");
    }

    s2->flags = s1->flags;

    if (s2->natoms != s1->natoms)
    {
        state_change_natoms(s2, s1->natoms);
        /* We need to allocate one element extra, since we might use
         * (unaligned) 4-wide SIMD loads to access rvec entries.
         */
        ems2->f.resize(gmx::paddedRVecVectorSize(s2->natoms));
    }
    if (DOMAINDECOMP(cr) && s2->cg_gl.size() != s1->cg_gl.size())
    {
        s2->cg_gl.resize(s1->cg_gl.size());
    }

    copy_mat(s1->box, s2->box);
    /* Copy free energy state */
    s2->lambda = s1->lambda;
    copy_mat(s1->box, s2->box);

    start = 0;
    end   = md->homenr;

    // cppcheck-suppress unreadVariable
    nthreads = gmx_omp_nthreads_get(emntUpdate);
#pragma omp parallel num_threads(nthreads)
    {
        const rvec *x1 = as_rvec_array(s1->x.data());
        rvec       *x2 = as_rvec_array(s2->x.data());
        const rvec *f  = as_rvec_array(force->data());

        int         gf = 0;
#pragma omp for schedule(static) nowait
        for (int i = start; i < end; i++)
        {
            try
            {
                if (md->cFREEZE)
                {
                    gf = md->cFREEZE[i];
                }
                for (int m = 0; m < DIM; m++)
                {
                    if (ir->opts.nFreeze[gf][m])
                    {
                        x2[i][m] = x1[i][m];
                    }
                    else
                    {
                        x2[i][m] = x1[i][m] + a*f[i][m];
                    }
                }
            }
            GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
        }

        if (s2->flags & (1<<estCGP))
        {
            /* Copy the CG p vector */
            const rvec *p1 = as_rvec_array(s1->cg_p.data());
            rvec       *p2 = as_rvec_array(s2->cg_p.data());
#pragma omp for schedule(static) nowait
            for (int i = start; i < end; i++)
            {
                // Trivial OpenMP block that does not throw
                copy_rvec(p1[i], p2[i]);
            }
        }

        if (DOMAINDECOMP(cr))
        {
            s2->ddp_count = s1->ddp_count;

            /* OpenMP does not supported unsigned loop variables */
#pragma omp for schedule(static) nowait
            for (int i = 0; i < static_cast<int>(s2->cg_gl.size()); i++)
            {
                s2->cg_gl[i] = s1->cg_gl[i];
            }
            s2->ddp_count_cg_gl = s1->ddp_count_cg_gl;
        }
    }

    if (constr)
    {
        wallcycle_start(wcycle, ewcCONSTR);
        dvdl_constr = 0;
        validStep   =
            constrain(nullptr, TRUE, TRUE, constr, &top->idef,
                      ir, cr, count, 0, 1.0, md,
                      as_rvec_array(s1->x.data()), as_rvec_array(s2->x.data()),
                      nullptr, bMolPBC, s2->box,
                      s2->lambda[efptBONDED], &dvdl_constr,
                      nullptr, nullptr, nrnb, econqCoord);
        wallcycle_stop(wcycle, ewcCONSTR);

        if (cr->nnodes > 1)
        {
            /* This global reduction will affect performance at high
             * parallelization, but we can not really avoid it.
             * But usually EM is not run at high parallelization.
             */
            int reductionBuffer = !validStep;
            gmx_sumi(1, &reductionBuffer, cr);
            validStep           = (reductionBuffer == 0);
        }

        // We should move this check to the different minimizers
        if (!validStep && ir->eI != eiSteep)
        {
            gmx_fatal(FARGS, "The coordinates could not be constrained. Minimizer '%s' can not handle constraint failures, use minimizer '%s' before using '%s'.",
                      EI(ir->eI), EI(eiSteep), EI(ir->eI));
        }
    }

    return validStep;
}

//! Prepare EM for using domain decomposition parallellization
static void em_dd_partition_system(FILE *fplog, int step, t_commrec *cr,
                                   gmx_mtop_t *top_global, t_inputrec *ir,
                                   em_state_t *ems, gmx_localtop_t *top,
                                   gmx::MDAtoms *mdAtoms, t_forcerec *fr,
                                   gmx_vsite_t *vsite, gmx_constr_t constr,
                                   t_nrnb *nrnb, gmx_wallcycle_t wcycle)
{
    /* Repartition the domain decomposition */
    dd_partition_system(fplog, step, cr, FALSE, 1,
                        nullptr, top_global, ir,
                        &ems->s, &ems->f,
                        mdAtoms, top, fr, vsite, constr,
                        nrnb, wcycle, FALSE);
    dd_store_state(cr->dd, &ems->s);
}

//! De one energy evaluation
static void evaluate_energy(FILE *fplog, t_commrec *cr,
                            gmx_mtop_t *top_global,
                            em_state_t *ems, gmx_localtop_t *top,
                            t_inputrec *inputrec,
                            t_nrnb *nrnb, gmx_wallcycle_t wcycle,
                            gmx_global_stat_t gstat,
                            gmx_vsite_t *vsite, gmx_constr_t constr,
                            t_fcdata *fcd,
                            t_graph *graph, gmx::MDAtoms *mdAtoms,
                            t_forcerec *fr, rvec mu_tot,
                            gmx_enerdata_t *enerd, tensor vir, tensor pres,
                            gmx_int64_t count, gmx_bool bFirst)
{
    real     t;
    gmx_bool bNS;
    tensor   force_vir, shake_vir, ekin;
    real     dvdl_constr, prescorr, enercorr, dvdlcorr;
    real     terminate = 0;

    /* Set the time to the initial time, the time does not change during EM */
    t = inputrec->init_t;

    if (bFirst ||
        (DOMAINDECOMP(cr) && ems->s.ddp_count < cr->dd->ddp_count))
    {
        /* This is the first state or an old state used before the last ns */
        bNS = TRUE;
    }
    else
    {
        bNS = FALSE;
        if (inputrec->nstlist > 0)
        {
            bNS = TRUE;
        }
    }

    if (vsite)
    {
        construct_vsites(vsite, as_rvec_array(ems->s.x.data()), 1, nullptr,
                         top->idef.iparams, top->idef.il,
                         fr->ePBC, fr->bMolPBC, cr, ems->s.box);
    }

    if (DOMAINDECOMP(cr) && bNS)
    {
        /* Repartition the domain decomposition */
        em_dd_partition_system(fplog, count, cr, top_global, inputrec,
                               ems, top, mdAtoms, fr, vsite, constr,
                               nrnb, wcycle);
    }

    /* Calc force & energy on new trial position  */
    /* do_force always puts the charge groups in the box and shifts again
     * We do not unshift, so molecules are always whole in congrad.c
     */
    do_force(fplog, cr, inputrec,
             count, nrnb, wcycle, top, &top_global->groups,
             ems->s.box, ems->s.x, &ems->s.hist,
             ems->f, force_vir, mdAtoms->mdatoms(), enerd, fcd,
             ems->s.lambda, graph, fr, vsite, mu_tot, t, nullptr, TRUE,
             GMX_FORCE_STATECHANGED | GMX_FORCE_ALLFORCES |
             GMX_FORCE_VIRIAL | GMX_FORCE_ENERGY |
             (bNS ? GMX_FORCE_NS : 0),
             DOMAINDECOMP(cr) ?
             DdOpenBalanceRegionBeforeForceComputation::yes :
             DdOpenBalanceRegionBeforeForceComputation::no,
             DOMAINDECOMP(cr) ?
             DdCloseBalanceRegionAfterForceComputation::yes :
             DdCloseBalanceRegionAfterForceComputation::no);

    /* Clear the unused shake virial and pressure */
    clear_mat(shake_vir);
    clear_mat(pres);

    /* Communicate stuff when parallel */
    if (PAR(cr) && inputrec->eI != eiNM)
    {
        wallcycle_start(wcycle, ewcMoveE);

        global_stat(gstat, cr, enerd, force_vir, shake_vir, mu_tot,
                    inputrec, nullptr, nullptr, nullptr, 1, &terminate,
                    nullptr, FALSE,
                    CGLO_ENERGY |
                    CGLO_PRESSURE |
                    CGLO_CONSTRAINT);

        wallcycle_stop(wcycle, ewcMoveE);
    }

    /* Calculate long range corrections to pressure and energy */
    calc_dispcorr(inputrec, fr, ems->s.box, ems->s.lambda[efptVDW],
                  pres, force_vir, &prescorr, &enercorr, &dvdlcorr);
    enerd->term[F_DISPCORR] = enercorr;
    enerd->term[F_EPOT]    += enercorr;
    enerd->term[F_PRES]    += prescorr;
    enerd->term[F_DVDL]    += dvdlcorr;

    ems->epot = enerd->term[F_EPOT];

    if (constr)
    {
        /* Project out the constraint components of the force */
        wallcycle_start(wcycle, ewcCONSTR);
        dvdl_constr = 0;
        rvec *f_rvec = as_rvec_array(ems->f.data());
        constrain(nullptr, FALSE, FALSE, constr, &top->idef,
                  inputrec, cr, count, 0, 1.0, mdAtoms->mdatoms(),
                  as_rvec_array(ems->s.x.data()), f_rvec, f_rvec,
                  fr->bMolPBC, ems->s.box,
                  ems->s.lambda[efptBONDED], &dvdl_constr,
                  nullptr, &shake_vir, nrnb, econqForceDispl);
        enerd->term[F_DVDL_CONSTR] += dvdl_constr;
        m_add(force_vir, shake_vir, vir);
        wallcycle_stop(wcycle, ewcCONSTR);
    }
    else
    {
        copy_mat(force_vir, vir);
    }

    clear_mat(ekin);
    enerd->term[F_PRES] =
        calc_pres(fr->ePBC, inputrec->nwall, ems->s.box, ekin, vir, pres);

    sum_dhdl(enerd, ems->s.lambda, inputrec->fepvals);

    if (EI_ENERGY_MINIMIZATION(inputrec->eI))
    {
        get_state_f_norm_max(cr, &(inputrec->opts), mdAtoms->mdatoms(), ems);
    }
}

//! Parallel utility summing energies and forces
static double reorder_partsum(t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
                              gmx_mtop_t *top_global,
                              em_state_t *s_min, em_state_t *s_b)
{
    t_block       *cgs_gl;
    int            ncg, *cg_gl, *index, c, cg, i, a0, a1, a, gf, m;
    double         partsum;
    unsigned char *grpnrFREEZE;

    if (debug)
    {
        fprintf(debug, "Doing reorder_partsum\n");
    }

    const rvec *fm = as_rvec_array(s_min->f.data());
    const rvec *fb = as_rvec_array(s_b->f.data());

    cgs_gl = dd_charge_groups_global(cr->dd);
    index  = cgs_gl->index;

    /* Collect fm in a global vector fmg.
     * This conflicts with the spirit of domain decomposition,
     * but to fully optimize this a much more complicated algorithm is required.
     */
    rvec *fmg;
    snew(fmg, top_global->natoms);

    ncg   = s_min->s.cg_gl.size();
    cg_gl = s_min->s.cg_gl.data();
    i     = 0;
    for (c = 0; c < ncg; c++)
    {
        cg = cg_gl[c];
        a0 = index[cg];
        a1 = index[cg+1];
        for (a = a0; a < a1; a++)
        {
            copy_rvec(fm[i], fmg[a]);
            i++;
        }
    }
    gmx_sum(top_global->natoms*3, fmg[0], cr);

    /* Now we will determine the part of the sum for the cgs in state s_b */
    ncg         = s_b->s.cg_gl.size();
    cg_gl       = s_b->s.cg_gl.data();
    partsum     = 0;
    i           = 0;
    gf          = 0;
    grpnrFREEZE = top_global->groups.grpnr[egcFREEZE];
    for (c = 0; c < ncg; c++)
    {
        cg = cg_gl[c];
        a0 = index[cg];
        a1 = index[cg+1];
        for (a = a0; a < a1; a++)
        {
            if (mdatoms->cFREEZE && grpnrFREEZE)
            {
                gf = grpnrFREEZE[i];
            }
            for (m = 0; m < DIM; m++)
            {
                if (!opts->nFreeze[gf][m])
                {
                    partsum += (fb[i][m] - fmg[a][m])*fb[i][m];
                }
            }
            i++;
        }
    }

    sfree(fmg);

    return partsum;
}

//! Print some stuff, like beta, whatever that means.
static real pr_beta(t_commrec *cr, t_grpopts *opts, t_mdatoms *mdatoms,
                    gmx_mtop_t *top_global,
                    em_state_t *s_min, em_state_t *s_b)
{
    double sum;

    /* This is just the classical Polak-Ribiere calculation of beta;
     * it looks a bit complicated since we take freeze groups into account,
     * and might have to sum it in parallel runs.
     */

    if (!DOMAINDECOMP(cr) ||
        (s_min->s.ddp_count == cr->dd->ddp_count &&
         s_b->s.ddp_count   == cr->dd->ddp_count))
    {
        const rvec *fm  = as_rvec_array(s_min->f.data());
        const rvec *fb  = as_rvec_array(s_b->f.data());
        sum             = 0;
        int         gf  = 0;
        /* This part of code can be incorrect with DD,
         * since the atom ordering in s_b and s_min might differ.
         */
        for (int i = 0; i < mdatoms->homenr; i++)
        {
            if (mdatoms->cFREEZE)
            {
                gf = mdatoms->cFREEZE[i];
            }
            for (int m = 0; m < DIM; m++)
            {
                if (!opts->nFreeze[gf][m])
                {
                    sum += (fb[i][m] - fm[i][m])*fb[i][m];
                }
            }
        }
    }
    else
    {
        /* We need to reorder cgs while summing */
        sum = reorder_partsum(cr, opts, mdatoms, top_global, s_min, s_b);
    }
    if (PAR(cr))
    {
        gmx_sumd(1, &sum, cr);
    }

    return sum/gmx::square(s_min->fnorm);
}

namespace gmx
{

/*! \brief Do conjugate gradients minimization
    \copydoc integrator_t(FILE *fplog, t_commrec *cr, const gmx::MDLogger &mdlog,
                           int nfile, const t_filenm fnm[],
                           const gmx_output_env_t *oenv,