Further improve getDDGridSetup

[gromacs.git] / src / gromacs / mdlib / lincs.cpp

/*
 * 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) 2013,2014,2015,2016,2017,2018,2019, 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 Defines LINCS code.
 *
 * \author Berk Hess <hess@kth.se>
 * \author Mark Abraham <mark.j.abraham@gmail.com>
 * \ingroup module_mdlib
 */
#include "gmxpre.h"

#include "lincs.h"

#include "config.h"

#include <cassert>
#include <cmath>
#include <cstdlib>

#include <algorithm>
#include <vector>

#include "gromacs/domdec/domdec.h"
#include "gromacs/domdec/domdec_struct.h"
#include "gromacs/gmxlib/nrnb.h"
#include "gromacs/math/functions.h"
#include "gromacs/math/paddedvector.h"
#include "gromacs/math/units.h"
#include "gromacs/math/vec.h"
#include "gromacs/mdlib/constr.h"
#include "gromacs/mdlib/gmx_omp_nthreads.h"
#include "gromacs/mdrunutility/multisim.h"
#include "gromacs/mdtypes/commrec.h"
#include "gromacs/mdtypes/inputrec.h"
#include "gromacs/mdtypes/md_enums.h"
#include "gromacs/mdtypes/mdatom.h"
#include "gromacs/pbcutil/pbc.h"
#include "gromacs/pbcutil/pbc_simd.h"
#include "gromacs/simd/simd.h"
#include "gromacs/simd/simd_math.h"
#include "gromacs/simd/vector_operations.h"
#include "gromacs/topology/block.h"
#include "gromacs/topology/mtop_util.h"
#include "gromacs/utility/alignedallocator.h"
#include "gromacs/utility/arrayref.h"
#include "gromacs/utility/basedefinitions.h"
#include "gromacs/utility/bitmask.h"
#include "gromacs/utility/cstringutil.h"
#include "gromacs/utility/exceptions.h"
#include "gromacs/utility/fatalerror.h"
#include "gromacs/utility/gmxomp.h"
#include "gromacs/utility/pleasecite.h"

using namespace gmx; // TODO: Remove when this file is moved into gmx namespace

namespace gmx
{

//! Indices of the two atoms involved in a single constraint
struct AtomPair
{
    //! \brief Constructor, does not initialize to catch bugs and faster construction
    AtomPair()
    {
    }

    //! Index of atom 1
    int index1;
    //! Index of atom 2
    int index2;
};

//! Unit of work within LINCS.
struct Task
{
    //! First constraint for this task.
    int              b0 = 0;
    //! b1-1 is the last constraint for this task.
    int              b1 = 0;
    //! The number of constraints in triangles.
    int              ntriangle = 0;
    //! The list of triangle constraints.
    std::vector<int> triangle;
    //! The bits tell if the matrix element should be used.
    std::vector<int> tri_bits;
    //! Constraint index for updating atom data.
    std::vector<int> ind;
    //! Constraint index for updating atom data.
    std::vector<int> ind_r;
    //! Temporary variable for virial calculation.
    tensor           vir_r_m_dr = {{0}};
    //! Temporary variable for lambda derivative.
    real             dhdlambda;
};

/*! \brief Data for LINCS algorithm.
 */
class Lincs
{
    public:
        //! The global number of constraints.
        int             ncg = 0;
        //! The global number of flexible constraints.
        int             ncg_flex = 0;
        //! The global number of constraints in triangles.
        int             ncg_triangle = 0;
        //! The number of iterations.
        int             nIter = 0;
        //! The order of the matrix expansion.
        int             nOrder = 0;
        //! The maximum number of constraints connected to a single atom.
        int             max_connect = 0;

        //! The number of real constraints.
        int              nc_real = 0;
        //! The number of constraints including padding for SIMD.
        int              nc = 0;
        //! The number of constraint connections.
        int              ncc = 0;
        //! The FE lambda value used for filling blc and blmf.
        real             matlam = 0;
        //! mapping from topology to LINCS constraints.
        std::vector<int> con_index;
        //! The reference distance in topology A.
        std::vector < real, AlignedAllocator < real>> bllen0;
        //! The reference distance in top B - the r.d. in top A.
        std::vector < real, AlignedAllocator < real>> ddist;
        //! The atom pairs involved in the constraints.
        std::vector<AtomPair> atoms;
        //! 1/sqrt(invmass1  invmass2).
        std::vector < real, AlignedAllocator < real>> blc;
        //! As blc, but with all masses 1.
        std::vector < real, AlignedAllocator < real>> blc1;
        //! Index into blbnb and blmf.
        std::vector<int>  blnr;
        //! List of constraint connections.
        std::vector<int>  blbnb;
        //! The local number of constraints in triangles.
        int               ntriangle = 0;
        //! The number of constraint connections in triangles.
        int               ncc_triangle = 0;
        //! Communicate before each LINCS interation.
        bool              bCommIter = false;
        //! Matrix of mass factors for constraint connections.
        std::vector<real> blmf;
        //! As blmf, but with all masses 1.
        std::vector<real> blmf1;
        //! The reference bond length.
        std::vector < real, AlignedAllocator < real>> bllen;
        //! The local atom count per constraint, can be NULL.
        std::vector<int>  nlocat;

        /*! \brief The number of tasks used for LINCS work.
         *
         * \todo This is mostly used to loop over \c task, which would
         * be nicer to do with range-based for loops, but the thread
         * index is used for constructing bit masks and organizing the
         * virial output buffer, so other things need to change,
         * first. */
        int                        ntask = 0;
        /*! \brief LINCS thread division */
        std::vector<Task>          task;
        //! Atom flags for thread parallelization.
        std::vector<gmx_bitmask_t> atf;
        //! Are the LINCS tasks interdependent?
        bool                       bTaskDep = false;
        //! Are there triangle constraints that cross task borders?
        bool                       bTaskDepTri = false;
        //! Arrays for temporary storage in the LINCS algorithm.
        /*! @{ */
        PaddedVector<gmx::RVec>                   tmpv;
        std::vector<real>                         tmpncc;
        std::vector < real, AlignedAllocator < real>> tmp1;
        std::vector < real, AlignedAllocator < real>> tmp2;
        std::vector < real, AlignedAllocator < real>> tmp3;
        std::vector < real, AlignedAllocator < real>> tmp4;
        /*! @} */
        //! The Lagrange multipliers times -1.
        std::vector < real, AlignedAllocator < real>> mlambda;
        //! Storage for the constraint RMS relative deviation output.
        std::array<real, 2> rmsdData = {{0}};
};

/*! \brief Define simd_width for memory allocation used for SIMD code */
#if GMX_SIMD_HAVE_REAL
static const int simd_width = GMX_SIMD_REAL_WIDTH;
#else
static const int simd_width = 1;
#endif

ArrayRef<real> lincs_rmsdData(Lincs *lincsd)
{
    return lincsd->rmsdData;
}

real lincs_rmsd(const Lincs *lincsd)
{
    if (lincsd->rmsdData[0] > 0)
    {
        return std::sqrt(lincsd->rmsdData[1]/lincsd->rmsdData[0]);
    }
    else
    {
        return 0;
    }
}

/*! \brief Do a set of nrec LINCS matrix multiplications.
 *
 * This function will return with up to date thread-local
 * constraint data, without an OpenMP barrier.
 */
static void lincs_matrix_expand(const Lincs               &lincsd,
                                const Task                &li_task,
                                gmx::ArrayRef<const real>  blcc,
                                gmx::ArrayRef<real>        rhs1,
                                gmx::ArrayRef<real>        rhs2,
                                gmx::ArrayRef<real>        sol)
{
    gmx::ArrayRef<const int> blnr  = lincsd.blnr;
    gmx::ArrayRef<const int> blbnb = lincsd.blbnb;

    const int                b0   = li_task.b0;
    const int                b1   = li_task.b1;
    const int                nrec = lincsd.nOrder;

    for (int rec = 0; rec < nrec; rec++)
    {
        if (lincsd.bTaskDep)
        {
#pragma omp barrier
        }
        for (int b = b0; b < b1; b++)
        {
            real mvb;
            int  n;

            mvb = 0;
            for (n = blnr[b]; n < blnr[b+1]; n++)
            {
                mvb = mvb + blcc[n]*rhs1[blbnb[n]];
            }
            rhs2[b] = mvb;
            sol[b]  = sol[b] + mvb;
        }

        std::swap(rhs1, rhs2);
    }   /* nrec*(ncons+2*nrtot) flops */

    if (lincsd.ntriangle > 0)
    {
        /* Perform an extra nrec recursions for only the constraints
         * involved in rigid triangles.
         * In this way their accuracy should come close to those of the other
         * constraints, since traingles of constraints can produce eigenvalues
         * around 0.7, while the effective eigenvalue for bond constraints
         * is around 0.4 (and 0.7*0.7=0.5).
         */

        if (lincsd.bTaskDep)
        {
            /* We need a barrier here, since other threads might still be
             * reading the contents of rhs1 and/o rhs2.
             * We could avoid this barrier by introducing two extra rhs
             * arrays for the triangle constraints only.
             */
#pragma omp barrier
        }

        /* Constraints involved in a triangle are ensured to be in the same
         * LINCS task. This means no barriers are required during the extra
         * iterations for the triangle constraints.
         */
        gmx::ArrayRef<const int> triangle = li_task.triangle;
        gmx::ArrayRef<const int> tri_bits = li_task.tri_bits;

        for (int rec = 0; rec < nrec; rec++)
        {
            for (int tb = 0; tb < li_task.ntriangle; tb++)
            {
                int  b, bits, nr0, nr1, n;
                real mvb;

                b    = triangle[tb];
                bits = tri_bits[tb];
                mvb  = 0;
                nr0  = blnr[b];
                nr1  = blnr[b+1];
                for (n = nr0; n < nr1; n++)
                {
                    if (bits & (1 << (n - nr0)))
                    {
                        mvb = mvb + blcc[n]*rhs1[blbnb[n]];
                    }
                }
                rhs2[b] = mvb;
                sol[b]  = sol[b] + mvb;
            }

            std::swap(rhs1, rhs2);
        }   /* nrec*(ntriangle + ncc_triangle*2) flops */

        if (lincsd.bTaskDepTri)
        {
            /* The constraints triangles are decoupled from each other,
             * but constraints in one triangle cross thread task borders.
             * We could probably avoid this with more advanced setup code.
             */
#pragma omp barrier
        }
    }
}

//! Update atomic coordinates when an index is not required.
static void
lincs_update_atoms_noind(int                             ncons,
                         gmx::ArrayRef<const AtomPair>   atoms,
                         real                            preFactor,
                         gmx::ArrayRef<const real>       fac,
                         gmx::ArrayRef<const gmx::RVec>  r,
                         const real                     *invmass,
                         rvec                           *x)
{
    if (invmass != nullptr)
    {
        for (int b = 0; b < ncons; b++)
        {
            int  i     = atoms[b].index1;
            int  j     = atoms[b].index2;
            real mvb   = preFactor*fac[b];
            real im1   = invmass[i];
            real im2   = invmass[j];
            real tmp0  = r[b][0]*mvb;
            real tmp1  = r[b][1]*mvb;
            real tmp2  = r[b][2]*mvb;
            x[i][0]   -= tmp0*im1;
            x[i][1]   -= tmp1*im1;
            x[i][2]   -= tmp2*im1;
            x[j][0]   += tmp0*im2;
            x[j][1]   += tmp1*im2;
            x[j][2]   += tmp2*im2;
        }   /* 16 ncons flops */
    }
    else
    {
        for (int b = 0; b < ncons; b++)
        {
            int  i     = atoms[b].index1;
            int  j     = atoms[b].index2;
            real mvb   = preFactor*fac[b];
            real tmp0  = r[b][0]*mvb;
            real tmp1  = r[b][1]*mvb;
            real tmp2  = r[b][2]*mvb;
            x[i][0]   -= tmp0;
            x[i][1]   -= tmp1;
            x[i][2]   -= tmp2;
            x[j][0]   += tmp0;
            x[j][1]   += tmp1;
            x[j][2]   += tmp2;
        }
    }
}

//! Update atomic coordinates when an index is required.
static void
lincs_update_atoms_ind(gmx::ArrayRef<const int>        ind,
                       gmx::ArrayRef<const AtomPair>   atoms,
                       real                            preFactor,
                       gmx::ArrayRef<const real>       fac,
                       gmx::ArrayRef<const gmx::RVec>  r,
                       const real                     *invmass,
                       rvec                           *x)
{
    if (invmass != nullptr)
    {
        for (int b : ind)
        {
            int  i     = atoms[b].index1;
            int  j     = atoms[b].index2;
            real mvb   = preFactor*fac[b];
            real im1   = invmass[i];
            real im2   = invmass[j];
            real tmp0  = r[b][0]*mvb;
            real tmp1  = r[b][1]*mvb;
            real tmp2  = r[b][2]*mvb;
            x[i][0]   -= tmp0*im1;
            x[i][1]   -= tmp1*im1;
            x[i][2]   -= tmp2*im1;
            x[j][0]   += tmp0*im2;
            x[j][1]   += tmp1*im2;
            x[j][2]   += tmp2*im2;
        }   /* 16 ncons flops */
    }
    else
    {
        for (int b : ind)
        {
            int  i     = atoms[b].index1;
            int  j     = atoms[b].index2;
            real mvb   = preFactor*fac[b];
            real tmp0  = r[b][0]*mvb;
            real tmp1  = r[b][1]*mvb;
            real tmp2  = r[b][2]*mvb;
            x[i][0]   -= tmp0;
            x[i][1]   -= tmp1;
            x[i][2]   -= tmp2;
            x[j][0]   += tmp0;
            x[j][1]   += tmp1;
            x[j][2]   += tmp2;
        }   /* 16 ncons flops */
    }
}

//! Update coordinates for atoms.
static void
lincs_update_atoms(Lincs                          *li,
                   int                             th,
                   real                            preFactor,
                   gmx::ArrayRef<const real>       fac,
                   gmx::ArrayRef<const gmx::RVec>  r,
                   const real                     *invmass,
                   rvec                           *x)
{
    if (li->ntask == 1)
    {
        /* Single thread, we simply update for all constraints */
        lincs_update_atoms_noind(li->nc_real,
                                 li->atoms, preFactor, fac, r, invmass, x);
    }
    else
    {
        /* Update the atom vector components for our thread local
         * constraints that only access our local atom range.
         * This can be done without a barrier.
         */
        lincs_update_atoms_ind(li->task[th].ind,
                               li->atoms, preFactor, fac, r, invmass, x);

        if (!li->task[li->ntask].ind.empty())
        {
            /* Update the constraints that operate on atoms
             * in multiple thread atom blocks on the master thread.
             */
#pragma omp barrier
#pragma omp master
            {
                lincs_update_atoms_ind(li->task[li->ntask].ind,
                                       li->atoms, preFactor, fac, r, invmass, x);
            }
        }
    }
}

#if GMX_SIMD_HAVE_REAL
//! Helper function so that we can run TSAN with SIMD support (where implemented).
template <int align>
static inline void gmx_simdcall
gatherLoadUTransposeTSANSafe(const real         *base,
                             const std::int32_t *offset,
                             SimdReal           *v0,
                             SimdReal           *v1,
                             SimdReal           *v2)
{
#if (CMAKE_BUILD_TYPE == CMAKE_BUILD_TYPE_TSAN) && GMX_SIMD_X86_AVX2_256
    // This function is only implemented in this case
    gatherLoadUTransposeSafe<align>(base, offset, v0, v1, v2);
#else
    gatherLoadUTranspose<align>(base, offset, v0, v1, v2);
#endif
}

/*! \brief Calculate the constraint distance vectors r to project on from x.
 *
 * Determine the right-hand side of the matrix equation using quantity f.
 * This function only differs from calc_dr_x_xp_simd below in that
 * no constraint length is subtracted and no PBC is used for f. */
static void gmx_simdcall
calc_dr_x_f_simd(int                            b0,
                 int                            b1,
                 gmx::ArrayRef<const AtomPair>  atoms,
                 const rvec * gmx_restrict      x,
                 const rvec * gmx_restrict      f,
                 const real * gmx_restrict      blc,
                 const real *                   pbc_simd,
                 rvec * gmx_restrict            r,
                 real * gmx_restrict            rhs,
                 real * gmx_restrict            sol)
{
    assert(b0 % GMX_SIMD_REAL_WIDTH == 0);

    alignas(GMX_SIMD_ALIGNMENT) std::int32_t offset2[GMX_SIMD_REAL_WIDTH];

    for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
    {
        offset2[i] = i;
    }

    for (int bs = b0; bs < b1; bs += GMX_SIMD_REAL_WIDTH)
    {
        SimdReal x0_S, y0_S, z0_S;
        SimdReal x1_S, y1_S, z1_S;
        SimdReal rx_S, ry_S, rz_S, n2_S, il_S;
        SimdReal fx_S, fy_S, fz_S, ip_S, rhs_S;
        alignas(GMX_SIMD_ALIGNMENT) std::int32_t      offset0[GMX_SIMD_REAL_WIDTH];
        alignas(GMX_SIMD_ALIGNMENT) std::int32_t      offset1[GMX_SIMD_REAL_WIDTH];

        for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
        {
            offset0[i] = atoms[bs + i].index1;
            offset1[i] = atoms[bs + i].index2;
        }

        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(x), offset0, &x0_S, &y0_S, &z0_S);
        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(x), offset1, &x1_S, &y1_S, &z1_S);
        rx_S = x0_S - x1_S;
        ry_S = y0_S - y1_S;
        rz_S = z0_S - z1_S;

        pbc_correct_dx_simd(&rx_S, &ry_S, &rz_S, pbc_simd);

        n2_S  = norm2(rx_S, ry_S, rz_S);
        il_S  = invsqrt(n2_S);

        rx_S  = rx_S * il_S;
        ry_S  = ry_S * il_S;
        rz_S  = rz_S * il_S;

        transposeScatterStoreU<3>(reinterpret_cast<real *>(r + bs), offset2, rx_S, ry_S, rz_S);

        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(f), offset0, &x0_S, &y0_S, &z0_S);
        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(f), offset1, &x1_S, &y1_S, &z1_S);
        fx_S = x0_S - x1_S;
        fy_S = y0_S - y1_S;
        fz_S = z0_S - z1_S;

        ip_S  = iprod(rx_S, ry_S, rz_S, fx_S, fy_S, fz_S);

        rhs_S = load<SimdReal>(blc + bs) * ip_S;

        store(rhs + bs, rhs_S);
        store(sol + bs, rhs_S);
    }
}
#endif // GMX_SIMD_HAVE_REAL

/*! \brief LINCS projection, works on derivatives of the coordinates. */
static void do_lincsp(const rvec *x, rvec *f, rvec *fp, t_pbc *pbc,
                      Lincs *lincsd, int th,
                      real *invmass,
                      ConstraintVariable econq, bool bCalcDHDL,
                      bool bCalcVir, tensor rmdf)
{
    const int                     b0 = lincsd->task[th].b0;
    const int                     b1 = lincsd->task[th].b1;

    gmx::ArrayRef<const AtomPair> atoms = lincsd->atoms;
    gmx::ArrayRef<gmx::RVec>      r     = lincsd->tmpv;
    gmx::ArrayRef<const int>      blnr  = lincsd->blnr;
    gmx::ArrayRef<const int>      blbnb = lincsd->blbnb;

    gmx::ArrayRef<const real>     blc;
    gmx::ArrayRef<const real>     blmf;
    if (econq != ConstraintVariable::Force)
    {
        /* Use mass-weighted parameters */
        blc  = lincsd->blc;
        blmf = lincsd->blmf;
    }
    else
    {
        /* Use non mass-weighted parameters */
        blc  = lincsd->blc1;
        blmf = lincsd->blmf1;
    }
    gmx::ArrayRef<real> blcc = lincsd->tmpncc;
    gmx::ArrayRef<real> rhs1 = lincsd->tmp1;
    gmx::ArrayRef<real> rhs2 = lincsd->tmp2;
    gmx::ArrayRef<real> sol  = lincsd->tmp3;

#if GMX_SIMD_HAVE_REAL
    /* This SIMD code does the same as the plain-C code after the #else.
     * The only difference is that we always call pbc code, as with SIMD
     * the overhead of pbc computation (when not needed) is small.
     */
    alignas(GMX_SIMD_ALIGNMENT) real    pbc_simd[9*GMX_SIMD_REAL_WIDTH];

    /* Convert the pbc struct for SIMD */
    set_pbc_simd(pbc, pbc_simd);

    /* Compute normalized x i-j vectors, store in r.
     * Compute the inner product of r and xp i-j and store in rhs1.
     */
    calc_dr_x_f_simd(b0, b1, atoms, x, f, blc.data(),
                     pbc_simd,
                     as_rvec_array(r.data()), rhs1.data(), sol.data());

#else   // GMX_SIMD_HAVE_REAL

    /* Compute normalized i-j vectors */
    if (pbc)
    {
        for (int b = b0; b < b1; b++)
        {
            rvec dx;

            pbc_dx_aiuc(pbc, x[atoms[b].index1], x[atoms[b].index2], dx);
            unitv(dx, r[b]);
        }
    }
    else
    {
        for (int b = b0; b < b1; b++)
        {
            rvec dx;

            rvec_sub(x[atoms[b].index1], x[atoms[b].index2], dx);
            unitv(dx, r[b]);
        }   /* 16 ncons flops */
    }

    for (int b = b0; b < b1; b++)
    {
        int  i   = atoms[b].index1;
        int  j   = atoms[b].index2;
        real mvb = blc[b]*(r[b][0]*(f[i][0] - f[j][0]) +
                           r[b][1]*(f[i][1] - f[j][1]) +
                           r[b][2]*(f[i][2] - f[j][2]));
        rhs1[b] = mvb;
        sol[b]  = mvb;
        /* 7 flops */
    }

#endif  // GMX_SIMD_HAVE_REAL

    if (lincsd->bTaskDep)
    {
        /* We need a barrier, since the matrix construction below
         * can access entries in r of other threads.
         */
#pragma omp barrier
    }

    /* Construct the (sparse) LINCS matrix */
    for (int b = b0; b < b1; b++)
    {
        for (int n = blnr[b]; n < blnr[b+1]; n++)
        {
            blcc[n] = blmf[n]*::iprod(r[b], r[blbnb[n]]);
        }   /* 6 nr flops */
    }
    /* Together: 23*ncons + 6*nrtot flops */

    lincs_matrix_expand(*lincsd, lincsd->task[th], blcc, rhs1, rhs2, sol);
    /* nrec*(ncons+2*nrtot) flops */

    if (econq == ConstraintVariable::Deriv_FlexCon)
    {
        /* We only want to constraint the flexible constraints,
         * so we mask out the normal ones by setting sol to 0.
         */
        for (int b = b0; b < b1; b++)
        {
            if (!(lincsd->bllen0[b] == 0 && lincsd->ddist[b] == 0))
            {
                sol[b] = 0;
            }
        }
    }

    /* We multiply sol by blc, so we can use lincs_update_atoms for OpenMP */
    for (int b = b0; b < b1; b++)
    {
        sol[b] *= blc[b];
    }

    /* When constraining forces, we should not use mass weighting,
     * so we pass invmass=NULL, which results in the use of 1 for all atoms.
     */
    lincs_update_atoms(lincsd, th, 1.0, sol, r,
                       (econq != ConstraintVariable::Force) ? invmass : nullptr, fp);

    if (bCalcDHDL)
    {
        real dhdlambda = 0;
        for (int b = b0; b < b1; b++)
        {
            dhdlambda -= sol[b]*lincsd->ddist[b];
        }

        lincsd->task[th].dhdlambda = dhdlambda;
    }

    if (bCalcVir)
    {
        /* Constraint virial,
         * determines sum r_bond x delta f,
         * where delta f is the constraint correction
         * of the quantity that is being constrained.
         */
        for (int b = b0; b < b1; b++)
        {
            const real mvb = lincsd->bllen[b]*sol[b];
            for (int i = 0; i < DIM; i++)
            {
                const real tmp1 = mvb*r[b][i];
                for (int j = 0; j < DIM; j++)
                {
                    rmdf[i][j] += tmp1*r[b][j];
                }
            }
        }   /* 23 ncons flops */
    }
}

#if GMX_SIMD_HAVE_REAL

/*! \brief Calculate the constraint distance vectors r to project on from x.
 *
 * Determine the right-hand side of the matrix equation using coordinates xp. */
static void gmx_simdcall
calc_dr_x_xp_simd(int                            b0,
                  int                            b1,
                  gmx::ArrayRef<const AtomPair>  atoms,
                  const rvec * gmx_restrict      x,
                  const rvec * gmx_restrict      xp,
                  const real * gmx_restrict      bllen,
                  const real * gmx_restrict      blc,
                  const real *                   pbc_simd,
                  rvec * gmx_restrict            r,
                  real * gmx_restrict            rhs,
                  real * gmx_restrict            sol)
{
    assert(b0 % GMX_SIMD_REAL_WIDTH == 0);
    alignas(GMX_SIMD_ALIGNMENT) std::int32_t offset2[GMX_SIMD_REAL_WIDTH];

    for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
    {
        offset2[i] = i;
    }

    for (int bs = b0; bs < b1; bs += GMX_SIMD_REAL_WIDTH)
    {
        SimdReal x0_S, y0_S, z0_S;
        SimdReal x1_S, y1_S, z1_S;
        SimdReal rx_S, ry_S, rz_S, n2_S, il_S;
        SimdReal rxp_S, ryp_S, rzp_S, ip_S, rhs_S;
        alignas(GMX_SIMD_ALIGNMENT) std::int32_t      offset0[GMX_SIMD_REAL_WIDTH];
        alignas(GMX_SIMD_ALIGNMENT) std::int32_t      offset1[GMX_SIMD_REAL_WIDTH];

        for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
        {
            offset0[i] = atoms[bs + i].index1;
            offset1[i] = atoms[bs + i].index2;
        }

        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(x), offset0, &x0_S, &y0_S, &z0_S);
        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(x), offset1, &x1_S, &y1_S, &z1_S);
        rx_S = x0_S - x1_S;
        ry_S = y0_S - y1_S;
        rz_S = z0_S - z1_S;

        pbc_correct_dx_simd(&rx_S, &ry_S, &rz_S, pbc_simd);

        n2_S  = norm2(rx_S, ry_S, rz_S);
        il_S  = invsqrt(n2_S);

        rx_S  = rx_S * il_S;
        ry_S  = ry_S * il_S;
        rz_S  = rz_S * il_S;

        transposeScatterStoreU<3>(reinterpret_cast<real *>(r + bs), offset2, rx_S, ry_S, rz_S);

        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(xp), offset0, &x0_S, &y0_S, &z0_S);
        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(xp), offset1, &x1_S, &y1_S, &z1_S);

        rxp_S = x0_S - x1_S;
        ryp_S = y0_S - y1_S;
        rzp_S = z0_S - z1_S;

        pbc_correct_dx_simd(&rxp_S, &ryp_S, &rzp_S, pbc_simd);

        ip_S  = iprod(rx_S, ry_S, rz_S, rxp_S, ryp_S, rzp_S);

        rhs_S = load<SimdReal>(blc + bs) * (ip_S - load<SimdReal>(bllen + bs));

        store(rhs + bs, rhs_S);
        store(sol + bs, rhs_S);
    }
}
#endif // GMX_SIMD_HAVE_REAL

/*! \brief Determine the distances and right-hand side for the next iteration. */
gmx_unused static void calc_dist_iter(
        int                           b0,
        int                           b1,
        gmx::ArrayRef<const AtomPair> atoms,
        const rvec * gmx_restrict     xp,
        const real * gmx_restrict     bllen,
        const real * gmx_restrict     blc,
        const t_pbc *                 pbc,
        real                          wfac,
        real * gmx_restrict           rhs,
        real * gmx_restrict           sol,
        bool *                        bWarn)
{
    for (int b = b0; b < b1; b++)
    {
        real len = bllen[b];
        rvec dx;
        if (pbc)
        {
            pbc_dx_aiuc(pbc, xp[atoms[b].index1], xp[atoms[b].index2], dx);
        }
        else
        {
            rvec_sub(xp[atoms[b].index1], xp[atoms[b].index2], dx);
        }
        real len2  = len*len;
        real dlen2 = 2*len2 - ::norm2(dx);
        if (dlen2 < wfac*len2)
        {
            /* not race free - see detailed comment in caller */
            *bWarn = TRUE;
        }
        real mvb;
        if (dlen2 > 0)
        {
            mvb = blc[b]*(len - dlen2*gmx::invsqrt(dlen2));
        }
        else
        {
            mvb = blc[b]*len;
        }
        rhs[b]  = mvb;
        sol[b]  = mvb;
    }   /* 20*ncons flops */
}

#if GMX_SIMD_HAVE_REAL
/*! \brief As calc_dist_iter(), but using SIMD intrinsics. */
static void gmx_simdcall
calc_dist_iter_simd(int                           b0,
                    int                           b1,
                    gmx::ArrayRef<const AtomPair> atoms,
                    const rvec * gmx_restrict     x,
                    const real * gmx_restrict     bllen,
                    const real * gmx_restrict     blc,
                    const real *                  pbc_simd,
                    real                          wfac,
                    real * gmx_restrict           rhs,
                    real * gmx_restrict           sol,
                    bool *                        bWarn)
{
    SimdReal        min_S(GMX_REAL_MIN);
    SimdReal        two_S(2.0);
    SimdReal        wfac_S(wfac);
    SimdBool        warn_S;

    assert(b0 % GMX_SIMD_REAL_WIDTH == 0);

    /* Initialize all to FALSE */
    warn_S = (two_S < setZero());

    for (int bs = b0; bs < b1; bs += GMX_SIMD_REAL_WIDTH)
    {
        SimdReal x0_S, y0_S, z0_S;
        SimdReal x1_S, y1_S, z1_S;
        SimdReal rx_S, ry_S, rz_S, n2_S;
        SimdReal len_S, len2_S, dlen2_S, lc_S, blc_S;
        alignas(GMX_SIMD_ALIGNMENT) std::int32_t      offset0[GMX_SIMD_REAL_WIDTH];
        alignas(GMX_SIMD_ALIGNMENT) std::int32_t      offset1[GMX_SIMD_REAL_WIDTH];

        for (int i = 0; i < GMX_SIMD_REAL_WIDTH; i++)
        {
            offset0[i] = atoms[bs + i].index1;
            offset1[i] = atoms[bs + i].index2;
        }

        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(x), offset0, &x0_S, &y0_S, &z0_S);
        gatherLoadUTransposeTSANSafe<3>(reinterpret_cast<const real *>(x), offset1, &x1_S, &y1_S, &z1_S);

        rx_S = x0_S - x1_S;
        ry_S = y0_S - y1_S;
        rz_S = z0_S - z1_S;

        pbc_correct_dx_simd(&rx_S, &ry_S, &rz_S, pbc_simd);

        n2_S    = norm2(rx_S, ry_S, rz_S);

        len_S   = load<SimdReal>(bllen + bs);
        len2_S  = len_S * len_S;

        dlen2_S = fms(two_S, len2_S, n2_S);

        warn_S  = warn_S || (dlen2_S < (wfac_S * len2_S));

        /* Avoid 1/0 by taking the max with REAL_MIN.
         * Note: when dlen2 is close to zero (90 degree constraint rotation),
         * the accuracy of the algorithm is no longer relevant.
         */
        dlen2_S = max(dlen2_S, min_S);

        lc_S    = fnma(dlen2_S, invsqrt(dlen2_S), len_S);

        blc_S   = load<SimdReal>(blc + bs);

        lc_S    = blc_S * lc_S;

        store(rhs + bs, lc_S);
        store(sol + bs, lc_S);
    }

    if (anyTrue(warn_S))
    {
        *bWarn = TRUE;
    }
}
#endif // GMX_SIMD_HAVE_REAL

//! Implements LINCS constraining.
static void do_lincs(const rvec *x, rvec *xp, const matrix box, t_pbc *pbc,
                     Lincs *lincsd, int th,
                     const real *invmass,
                     const t_commrec *cr,
                     bool bCalcDHDL,
                     real wangle, bool *bWarn,
                     real invdt, rvec * gmx_restrict v,
                     bool bCalcVir, tensor vir_r_m_dr)
{
    const int                     b0 = lincsd->task[th].b0;
    const int                     b1 = lincsd->task[th].b1;

    gmx::ArrayRef<const AtomPair> atoms   = lincsd->atoms;
    gmx::ArrayRef<gmx::RVec>      r       = lincsd->tmpv;
    gmx::ArrayRef<const int>      blnr    = lincsd->blnr;
    gmx::ArrayRef<const int>      blbnb   = lincsd->blbnb;
    gmx::ArrayRef<const real>     blc     = lincsd->blc;
    gmx::ArrayRef<const real>     blmf    = lincsd->blmf;
    gmx::ArrayRef<const real>     bllen   = lincsd->bllen;
    gmx::ArrayRef<real>           blcc    = lincsd->tmpncc;
    gmx::ArrayRef<real>           rhs1    = lincsd->tmp1;
    gmx::ArrayRef<real>           rhs2    = lincsd->tmp2;
    gmx::ArrayRef<real>           sol     = lincsd->tmp3;
    gmx::ArrayRef<real>           blc_sol = lincsd->tmp4;
    gmx::ArrayRef<real>           mlambda = lincsd->mlambda;
    gmx::ArrayRef<const int>      nlocat  = lincsd->nlocat;

#if GMX_SIMD_HAVE_REAL

    /* This SIMD code does the same as the plain-C code after the #else.
     * The only difference is that we always call pbc code, as with SIMD
     * the overhead of pbc computation (when not needed) is small.
     */
    alignas(GMX_SIMD_ALIGNMENT) real    pbc_simd[9*GMX_SIMD_REAL_WIDTH];

    /* Convert the pbc struct for SIMD */
    set_pbc_simd(pbc, pbc_simd);

    /* Compute normalized x i-j vectors, store in r.
     * Compute the inner product of r and xp i-j and store in rhs1.
     */
    calc_dr_x_xp_simd(b0, b1, atoms, x, xp, bllen.data(), blc.data(),
                      pbc_simd,
                      as_rvec_array(r.data()), rhs1.data(), sol.data());

#else   // GMX_SIMD_HAVE_REAL

    if (pbc)
    {
        /* Compute normalized i-j vectors */
        for (int b = b0; b < b1; b++)
        {
            rvec dx;
            pbc_dx_aiuc(pbc, x[atoms[b].index1], x[atoms[b].index2], dx);
            unitv(dx, r[b]);

            pbc_dx_aiuc(pbc, xp[atoms[b].index1], xp[atoms[b].index2], dx);
            real mvb  = blc[b]*(::iprod(r[b], dx) - bllen[b]);
            rhs1[b]   = mvb;
            sol[b]    = mvb;
        }
    }
    else
    {
        /* Compute normalized i-j vectors */
        for (int b = b0; b < b1; b++)
        {
            int  i    = atoms[b].index1;
            int  j    = atoms[b].index2;
            real tmp0 = x[i][0] - x[j][0];
            real tmp1 = x[i][1] - x[j][1];
            real tmp2 = x[i][2] - x[j][2];
            real rlen = gmx::invsqrt(tmp0*tmp0 + tmp1*tmp1 + tmp2*tmp2);
            r[b][0]   = rlen*tmp0;
            r[b][1]   = rlen*tmp1;
            r[b][2]   = rlen*tmp2;
            /* 16 ncons flops */

            real mvb  = blc[b]*(r[b][0]*(xp[i][0] - xp[j][0]) +
                                r[b][1]*(xp[i][1] - xp[j][1]) +
                                r[b][2]*(xp[i][2] - xp[j][2]) - bllen[b]);
            rhs1[b]   = mvb;
            sol[b]    = mvb;
            /* 10 flops */
        }
        /* Together: 26*ncons + 6*nrtot flops */
    }

#endif  // GMX_SIMD_HAVE_REAL

    if (lincsd->bTaskDep)
    {
        /* We need a barrier, since the matrix construction below
         * can access entries in r of other threads.
         */
#pragma omp barrier
    }

    /* Construct the (sparse) LINCS matrix */
    for (int b = b0; b < b1; b++)
    {
        for (int n = blnr[b]; n < blnr[b+1]; n++)
        {
            blcc[n] = blmf[n]*gmx::dot(r[b], r[blbnb[n]]);
        }
    }
    /* Together: 26*ncons + 6*nrtot flops */

    lincs_matrix_expand(*lincsd, lincsd->task[th], blcc, rhs1, rhs2, sol);
    /* nrec*(ncons+2*nrtot) flops */

#if GMX_SIMD_HAVE_REAL
    for (int b = b0; b < b1; b += GMX_SIMD_REAL_WIDTH)
    {
        SimdReal t1 = load<SimdReal>(blc.data() + b);
        SimdReal t2 = load<SimdReal>(sol.data() + b);
        store(mlambda.data() + b, t1 * t2);
    }
#else
    for (int b = b0; b < b1; b++)
    {
        mlambda[b] = blc[b]*sol[b];
    }
#endif  // GMX_SIMD_HAVE_REAL

    /* Update the coordinates */
    lincs_update_atoms(lincsd, th, 1.0, mlambda, r, invmass, xp);

    /*
     ********  Correction for centripetal effects  ********
     */

    real wfac;

    wfac = std::cos(DEG2RAD*wangle);
    wfac = wfac*wfac;

    for (int iter = 0; iter < lincsd->nIter; iter++)
    {
        if ((lincsd->bCommIter && DOMAINDECOMP(cr) && cr->dd->constraints))
        {
#pragma omp barrier
#pragma omp master
            {
                /* Communicate the corrected non-local coordinates */
                if (DOMAINDECOMP(cr))
                {
                    dd_move_x_constraints(cr->dd, box, xp, nullptr, FALSE);
                }
            }
#pragma omp barrier
        }
        else if (lincsd->bTaskDep)
        {
#pragma omp barrier
        }

#if GMX_SIMD_HAVE_REAL
        calc_dist_iter_simd(b0, b1, atoms,
                            xp, bllen.data(), blc.data(), pbc_simd, wfac,
                            rhs1.data(), sol.data(), bWarn);
#else
        calc_dist_iter(b0, b1, atoms, xp,
                       bllen.data(), blc.data(), pbc, wfac,
                       rhs1.data(), sol.data(), bWarn);
        /* 20*ncons flops */
#endif      // GMX_SIMD_HAVE_REAL

        lincs_matrix_expand(*lincsd, lincsd->task[th], blcc, rhs1, rhs2, sol);
        /* nrec*(ncons+2*nrtot) flops */

#if GMX_SIMD_HAVE_REAL
        for (int b = b0; b < b1; b += GMX_SIMD_REAL_WIDTH)
        {
            SimdReal t1  = load<SimdReal>(blc.data() + b);
            SimdReal t2  = load<SimdReal>(sol.data() + b);
            SimdReal mvb = t1 * t2;
            store(blc_sol.data() + b, mvb);
            store(mlambda.data() + b, load<SimdReal>(mlambda.data() + b) + mvb);
        }
#else
        for (int b = b0; b < b1; b++)
        {
            real mvb    = blc[b]*sol[b];
            blc_sol[b]  = mvb;
            mlambda[b] += mvb;
        }
#endif      // GMX_SIMD_HAVE_REAL

        /* Update the coordinates */
        lincs_update_atoms(lincsd, th, 1.0, blc_sol, r, invmass, xp);
    }
    /* nit*ncons*(37+9*nrec) flops */

    if (v != nullptr)
    {
        /* Update the velocities */
        lincs_update_atoms(lincsd, th, invdt, mlambda, r, invmass, v);
        /* 16 ncons flops */
    }

    if (!nlocat.empty() && (bCalcDHDL || bCalcVir))
    {
        if (lincsd->bTaskDep)
        {
            /* In lincs_update_atoms threads might cross-read mlambda */
#pragma omp barrier
        }

        /* Only account for local atoms */
        for (int b = b0; b < b1; b++)
        {
            mlambda[b] *= 0.5*nlocat[b];
        }
    }

    if (bCalcDHDL)
    {
        real dhdl = 0;
        for (int b = b0; b < b1; b++)
        {
            /* Note that this this is dhdl*dt^2, the dt^2 factor is corrected
             * later after the contributions are reduced over the threads.
             */
            dhdl -= lincsd->mlambda[b]*lincsd->ddist[b];
        }
        lincsd->task[th].dhdlambda = dhdl;
    }

    if (bCalcVir)
    {
        /* Constraint virial */
        for (int b = b0; b < b1; b++)
        {
            real tmp0 = -bllen[b]*mlambda[b];
            for (int i = 0; i < DIM; i++)
            {
                real tmp1 = tmp0*r[b][i];
                for (int j = 0; j < DIM; j++)
                {
                    vir_r_m_dr[i][j] -= tmp1*r[b][j];
                }
            }
        }   /* 22 ncons flops */
    }

    /* Total:
     * 26*ncons + 6*nrtot + nrec*(ncons+2*nrtot)
     * + nit * (20*ncons + nrec*(ncons+2*nrtot) + 17 ncons)
     *
     * (26+nrec)*ncons + (6+2*nrec)*nrtot
     * + nit * ((37+nrec)*ncons + 2*nrec*nrtot)
     * if nit=1
     * (63+nrec)*ncons + (6+4*nrec)*nrtot
     */
}

/*! \brief Sets the elements in the LINCS matrix for task task. */
static void set_lincs_matrix_task(Lincs                *li,
                                  Task                 *li_task,
                                  const real           *invmass,
                                  int                  *ncc_triangle,
                                  int                  *nCrossTaskTriangles)
{
    /* Construct the coupling coefficient matrix blmf */
    li_task->ntriangle   = 0;
    *ncc_triangle        = 0;
    *nCrossTaskTriangles = 0;
    for (int i = li_task->b0; i < li_task->b1; i++)
    {
        const int a1 = li->atoms[i].index1;
        const int a2 = li->atoms[i].index2;
        for (int n = li->blnr[i]; n < li->blnr[i + 1]; n++)
        {
            const int k = li->blbnb[n];

            /* If we are using multiple, independent tasks for LINCS,
             * the calls to check_assign_connected should have
             * put all connected constraints in our task.
             */
            assert(li->bTaskDep || (k >= li_task->b0 && k < li_task->b1));

            int sign;
            if (a1 == li->atoms[k].index1 || a2 == li->atoms[k].index2)
            {
                sign = -1;
            }
            else
            {
                sign = 1;
            }
            int center;
            int end;
            if (a1 == li->atoms[k].index1 || a1 == li->atoms[k].index2)
            {
                center = a1;
                end    = a2;
            }
            else
            {
                center = a2;
                end    = a1;
            }
            li->blmf[n]  = sign*invmass[center]*li->blc[i]*li->blc[k];
            li->blmf1[n] = sign*0.5;
            if (li->ncg_triangle > 0)
            {
                /* Look for constraint triangles */
                for (int nk = li->blnr[k]; nk < li->blnr[k + 1]; nk++)
                {
                    const int kk = li->blbnb[nk];
                    if (kk != i && kk != k &&
                        (li->atoms[kk].index1 == end ||
                         li->atoms[kk].index2 == end))
                    {
                        /* Check if the constraints in this triangle actually
                         * belong to a different task. We still assign them
                         * here, since it's convenient for the triangle
                         * iterations, but we then need an extra barrier.
                         */
                        if (k  < li_task->b0 || k  >= li_task->b1 ||
                            kk < li_task->b0 || kk >= li_task->b1)
                        {
                            (*nCrossTaskTriangles)++;
                        }

                        if (li_task->ntriangle == 0 ||
                            li_task->triangle[li_task->ntriangle - 1] < i)
                        {
                            /* Add this constraint to the triangle list */
                            li_task->triangle[li_task->ntriangle] = i;
                            li_task->tri_bits[li_task->ntriangle] = 0;
                            li_task->ntriangle++;
                            if (li->blnr[i+1] - li->blnr[i] > static_cast<int>(sizeof(li_task->tri_bits[0])*8 - 1))
                            {
                                gmx_fatal(FARGS, "A constraint is connected to %d constraints, this is more than the %zu allowed for constraints participating in triangles",
                                          li->blnr[i+1] - li->blnr[i],
                                          sizeof(li_task->tri_bits[0])*8-1);
                            }
                        }
                        li_task->tri_bits[li_task->ntriangle-1] |= (1 << (n - li->blnr[i]));
                        (*ncc_triangle)++;
                    }
                }
            }
        }
    }
}

/*! \brief Sets the elements in the LINCS matrix. */
static void set_lincs_matrix(Lincs *li, real *invmass, real lambda)
{
    const real invsqrt2 = 0.7071067811865475244;

    for (int i = 0; (i < li->nc); i++)
    {
        const int a1 = li->atoms[i].index1;
        const int a2 = li->atoms[i].index2;
        li->blc[i]   = gmx::invsqrt(invmass[a1] + invmass[a2]);
        li->blc1[i]  = invsqrt2;
    }

    /* Construct the coupling coefficient matrix blmf */
    int ntriangle = 0, ncc_triangle = 0, nCrossTaskTriangles = 0;
#pragma omp parallel for reduction(+: ntriangle, ncc_triangle, nCrossTaskTriangles) num_threads(li->ntask) schedule(static)
    for (int th = 0; th < li->ntask; th++)
    {
        try
        {
            set_lincs_matrix_task(li, &li->task[th], invmass,
                                  &ncc_triangle, &nCrossTaskTriangles);
            ntriangle += li->task[th].ntriangle;
        }
        GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
    }
    li->ntriangle    = ntriangle;
    li->ncc_triangle = ncc_triangle;
    li->bTaskDepTri  = (nCrossTaskTriangles > 0);

    if (debug)
    {
        fprintf(debug, "The %d constraints participate in %d triangles\n",
                li->nc, li->ntriangle);
        fprintf(debug, "There are %d constraint couplings, of which %d in triangles\n",
                li->ncc, li->ncc_triangle);
        if (li->ntriangle > 0 && li->ntask > 1)
        {
            fprintf(debug, "%d constraint triangles contain constraints assigned to different tasks\n",
                    nCrossTaskTriangles);
        }
    }

    /* Set matlam,
     * so we know with which lambda value the masses have been set.
     */
    li->matlam = lambda;
}

//! Finds all triangles of atoms that share constraints to a central atom.
static int count_triangle_constraints(const InteractionLists &ilist,
                                      const t_blocka         &at2con)
{
    int      ncon1, ncon_tot;
    int      c0, n1, c1, ac1, n2, c2;
    int      ncon_triangle;

    ncon1    = ilist[F_CONSTR].size()/3;
    ncon_tot = ncon1 + ilist[F_CONSTRNC].size()/3;

    gmx::ArrayRef<const int> ia1 = ilist[F_CONSTR].iatoms;
    gmx::ArrayRef<const int> ia2 = ilist[F_CONSTRNC].iatoms;

    ncon_triangle = 0;
    for (c0 = 0; c0 < ncon_tot; c0++)
    {
        bool       bTriangle = FALSE;
        const int *iap       = constr_iatomptr(ia1, ia2, c0);
        const int  a00       = iap[1];
        const int  a01       = iap[2];
        for (n1 = at2con.index[a01]; n1 < at2con.index[a01+1]; n1++)
        {
            c1 = at2con.a[n1];
            if (c1 != c0)
            {
                const int *iap = constr_iatomptr(ia1, ia2, c1);
                const int  a10 = iap[1];
                const int  a11 = iap[2];
                if (a10 == a01)
                {
                    ac1 = a11;
                }
                else
                {
                    ac1 = a10;
                }
                for (n2 = at2con.index[ac1]; n2 < at2con.index[ac1+1]; n2++)
                {
                    c2 = at2con.a[n2];
                    if (c2 != c0 && c2 != c1)
                    {
                        const int *iap = constr_iatomptr(ia1, ia2, c2);
                        const int  a20 = iap[1];
                        const int  a21 = iap[2];
                        if (a20 == a00 || a21 == a00)
                        {
                            bTriangle = TRUE;
                        }
                    }
                }
            }
        }
        if (bTriangle)
        {
            ncon_triangle++;
        }
    }

    return ncon_triangle;
}

//! Finds sequences of sequential constraints.
static bool more_than_two_sequential_constraints(const InteractionLists &ilist,
                                                 const t_blocka         &at2con)
{
    int       ncon1, ncon_tot, c;
    bool      bMoreThanTwoSequentialConstraints;

    ncon1    = ilist[F_CONSTR].size()/3;
    ncon_tot = ncon1 + ilist[F_CONSTRNC].size()/3;

    gmx::ArrayRef<const int> ia1 = ilist[F_CONSTR].iatoms;
    gmx::ArrayRef<const int> ia2 = ilist[F_CONSTRNC].iatoms;

    bMoreThanTwoSequentialConstraints = FALSE;
    for (c = 0; c < ncon_tot && !bMoreThanTwoSequentialConstraints; c++)
    {
        const int *iap = constr_iatomptr(ia1, ia2, c);
        const int  a1  = iap[1];
        const int  a2  = iap[2];
        /* Check if this constraint has constraints connected at both atoms */
        if (at2con.index[a1+1] - at2con.index[a1] > 1 &&
            at2con.index[a2+1] - at2con.index[a2] > 1)
        {
            bMoreThanTwoSequentialConstraints = TRUE;
        }
    }

    return bMoreThanTwoSequentialConstraints;
}

Lincs *init_lincs(FILE *fplog, const gmx_mtop_t &mtop,
                  int nflexcon_global, ArrayRef<const t_blocka> at2con,
                  bool bPLINCS, int nIter, int nProjOrder)
{
    // TODO this should become a unique_ptr
    Lincs                *li;
    bool                  bMoreThanTwoSeq;

    if (fplog)
    {
        fprintf(fplog, "\nInitializing%s LINear Constraint Solver\n",
                bPLINCS ? " Parallel" : "");
    }

    li = new Lincs;

    li->ncg      =
        gmx_mtop_ftype_count(mtop, F_CONSTR) +
        gmx_mtop_ftype_count(mtop, F_CONSTRNC);
    li->ncg_flex = nflexcon_global;

    li->nIter  = nIter;
    li->nOrder = nProjOrder;

    li->max_connect = 0;
    for (size_t mt = 0; mt < mtop.moltype.size(); mt++)
    {
        for (int a = 0; a < mtop.moltype[mt].atoms.nr; a++)
        {
            li->max_connect = std::max(li->max_connect,
                                       at2con[mt].index[a + 1] - at2con[mt].index[a]);
        }
    }

    li->ncg_triangle = 0;
    bMoreThanTwoSeq  = FALSE;
    for (const gmx_molblock_t &molb : mtop.molblock)
    {
        const gmx_moltype_t &molt = mtop.moltype[molb.type];

        li->ncg_triangle +=
            molb.nmol*
            count_triangle_constraints(molt.ilist, at2con[molb.type]);

        if (!bMoreThanTwoSeq &&
            more_than_two_sequential_constraints(molt.ilist, at2con[molb.type]))
        {
            bMoreThanTwoSeq = TRUE;
        }
    }

    /* Check if we need to communicate not only before LINCS,
     * but also before each iteration.
     * The check for only two sequential constraints is only
     * useful for the common case of H-bond only constraints.
     * With more effort we could also make it useful for small
     * molecules with nr. sequential constraints <= nOrder-1.
     */
    li->bCommIter = (bPLINCS && (li->nOrder < 1 || bMoreThanTwoSeq));

    if (debug && bPLINCS)
    {
        fprintf(debug, "PLINCS communication before each iteration: %d\n",
                static_cast<int>(li->bCommIter));
    }

    /* LINCS can run on any number of threads.
     * Currently the number is fixed for the whole simulation,
     * but it could be set in set_lincs().
     * The current constraint to task assignment code can create independent
     * tasks only when not more than two constraints are connected sequentially.
     */
    li->ntask    = gmx_omp_nthreads_get(emntLINCS);
    li->bTaskDep = (li->ntask > 1 && bMoreThanTwoSeq);
    if (debug)
    {
        fprintf(debug, "LINCS: using %d threads, tasks are %sdependent\n",
                li->ntask, li->bTaskDep ? "" : "in");
    }
    if (li->ntask == 1)
    {
        li->task.resize(1);
    }
    else
    {
        /* Allocate an extra elements for "task-overlap" constraints */
        li->task.resize(li->ntask + 1);
    }

    if (bPLINCS || li->ncg_triangle > 0)
    {
        please_cite(fplog, "Hess2008a");
    }
    else
    {
        please_cite(fplog, "Hess97a");
    }

    if (fplog)
    {
        fprintf(fplog, "The number of constraints is %d\n", li->ncg);
        if (bPLINCS)
        {
            fprintf(fplog, "There are constraints between atoms in different decomposition domains,\n"
                    "will communicate selected coordinates each lincs iteration\n");
        }
        if (li->ncg_triangle > 0)
        {
            fprintf(fplog,
                    "%d constraints are involved in constraint triangles,\n"
                    "will apply an additional matrix expansion of order %d for couplings\n"
                    "between constraints inside triangles\n",
                    li->ncg_triangle, li->nOrder);
        }
    }

    return li;
}

void done_lincs(Lincs *li)
{
    delete li;
}

/*! \brief Sets up the work division over the threads. */
static void lincs_thread_setup(Lincs *li, int natoms)
{
    li->atf.resize(natoms);

    gmx::ArrayRef<gmx_bitmask_t> atf = li->atf;

    /* Clear the atom flags */
    for (gmx_bitmask_t &mask : atf)
    {
        bitmask_clear(&mask);
    }

    if (li->ntask > BITMASK_SIZE)
    {
        gmx_fatal(FARGS, "More than %d threads is not supported for LINCS.", BITMASK_SIZE);
    }

    for (int th = 0; th < li->ntask; th++)
    {
        const Task *li_task = &li->task[th];

        /* For each atom set a flag for constraints from each */
        for (int b = li_task->b0; b < li_task->b1; b++)
        {
            bitmask_set_bit(&atf[li->atoms[b].index1], th);
            bitmask_set_bit(&atf[li->atoms[b].index2], th);
        }
    }

#pragma omp parallel for num_threads(li->ntask) schedule(static)
    for (int th = 0; th < li->ntask; th++)
    {
        try
        {
            Task          *li_task;
            gmx_bitmask_t  mask;
            int            b;

            li_task = &li->task[th];

            bitmask_init_low_bits(&mask, th);

            li_task->ind.clear();
            li_task->ind_r.clear();
            for (b = li_task->b0; b < li_task->b1; b++)
            {
                /* We let the constraint with the lowest thread index
                 * operate on atoms with constraints from multiple threads.
                 */
                if (bitmask_is_disjoint(atf[li->atoms[b].index1], mask) &&
                    bitmask_is_disjoint(atf[li->atoms[b].index2], mask))
                {
                    /* Add the constraint to the local atom update index */
                    li_task->ind.push_back(b);
                }
                else
                {
                    /* Add the constraint to the rest block */
                    li_task->ind_r.push_back(b);
                }
            }
        }
        GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
    }

    /* We need to copy all constraints which have not be assigned
     * to a thread to a separate list which will be handled by one thread.
     */
    Task *li_m = &li->task[li->ntask];

    li_m->ind.clear();
    for (int th = 0; th < li->ntask; th++)
    {
        const Task &li_task = li->task[th];

        for (int ind_r : li_task.ind_r)
        {
            li_m->ind.push_back(ind_r);
        }

        if (debug)
        {
            fprintf(debug, "LINCS thread %d: %zu constraints\n",
                    th, li_task.ind.size());
        }
    }

    if (debug)
    {
        fprintf(debug, "LINCS thread r: %zu constraints\n",
                li_m->ind.size());
    }
}

//! Assign a constraint.
static void assign_constraint(Lincs *li,
                              int constraint_index,
                              int a1, int a2,
                              real lenA, real lenB,
                              const t_blocka *at2con)
{
    int con;

    con = li->nc;

    /* Make an mapping of local topology constraint index to LINCS index */
    li->con_index[constraint_index] = con;

    li->bllen0[con]       = lenA;
    li->ddist[con]        = lenB - lenA;
    /* Set the length to the topology A length */
    li->bllen[con]        = lenA;
    li->atoms[con].index1 = a1;
    li->atoms[con].index2 = a2;

    /* Make space in the constraint connection matrix for constraints
     * connected to both end of the current constraint.
     */
    li->ncc +=
        at2con->index[a1 + 1] - at2con->index[a1] - 1 +
        at2con->index[a2 + 1] - at2con->index[a2] - 1;

    li->blnr[con + 1] = li->ncc;

    /* Increase the constraint count */
    li->nc++;
}

/*! \brief Check if constraint with topology index constraint_index is connected
 * to other constraints, and if so add those connected constraints to our task. */
static void check_assign_connected(Lincs *li,
                                   const t_iatom *iatom,
                                   const t_idef &idef,
                                   bool bDynamics,
                                   int a1, int a2,
                                   const t_blocka *at2con)
{
    /* Currently this function only supports constraint groups
     * in which all constraints share at least one atom
     * (e.g. H-bond constraints).
     * Check both ends of the current constraint for
     * connected constraints. We need to assign those
     * to the same task.
     */
    int end;

    for (end = 0; end < 2; end++)
    {
        int a, k;

        a = (end == 0 ? a1 : a2);

        for (k = at2con->index[a]; k < at2con->index[a + 1]; k++)
        {
            int cc;

            cc = at2con->a[k];
            /* Check if constraint cc has not yet been assigned */
            if (li->con_index[cc] == -1)
            {
                int  type;
                real lenA, lenB;

                type = iatom[cc*3];
                lenA = idef.iparams[type].constr.dA;
                lenB = idef.iparams[type].constr.dB;

                if (bDynamics || lenA != 0 || lenB != 0)
                {
                    assign_constraint(li, cc, iatom[3*cc + 1], iatom[3*cc + 2], lenA, lenB, at2con);
                }
            }
        }
    }
}

/*! \brief Check if constraint with topology index constraint_index is involved
 * in a constraint triangle, and if so add the other two constraints
 * in the triangle to our task. */
static void check_assign_triangle(Lincs *li,
                                  const t_iatom *iatom,
                                  const t_idef &idef,
                                  bool bDynamics,
                                  int constraint_index,
                                  int a1, int a2,
                                  const t_blocka *at2con)
{
    int nca, cc[32], ca[32], k;
    int c_triangle[2] = { -1, -1 };

    nca = 0;
    for (k = at2con->index[a1]; k < at2con->index[a1 + 1]; k++)
    {
        int c;

        c = at2con->a[k];
        if (c != constraint_index)
        {
            int aa1, aa2;

            aa1 = iatom[c*3 + 1];
            aa2 = iatom[c*3 + 2];
            if (aa1 != a1)
            {
                cc[nca] = c;
                ca[nca] = aa1;
                nca++;
            }
            if (aa2 != a1)
            {
                cc[nca] = c;
                ca[nca] = aa2;
                nca++;
            }
        }
    }

    for (k = at2con->index[a2]; k < at2con->index[a2 + 1]; k++)
    {
        int c;

        c = at2con->a[k];
        if (c != constraint_index)
        {
            int aa1, aa2, i;

            aa1 = iatom[c*3 + 1];
            aa2 = iatom[c*3 + 2];
            if (aa1 != a2)
            {
                for (i = 0; i < nca; i++)
                {
                    if (aa1 == ca[i])
                    {
                        c_triangle[0] = cc[i];
                        c_triangle[1] = c;
                    }
                }
            }
            if (aa2 != a2)
            {
                for (i = 0; i < nca; i++)
                {
                    if (aa2 == ca[i])
                    {
                        c_triangle[0] = cc[i];
                        c_triangle[1] = c;
                    }
                }
            }
        }
    }

    if (c_triangle[0] >= 0)
    {
        int end;

        for (end = 0; end < 2; end++)
        {
            /* Check if constraint c_triangle[end] has not yet been assigned */
            if (li->con_index[c_triangle[end]] == -1)
            {
                int  i, type;
                real lenA, lenB;

                i    = c_triangle[end]*3;
                type = iatom[i];
                lenA = idef.iparams[type].constr.dA;
                lenB = idef.iparams[type].constr.dB;

                if (bDynamics || lenA != 0 || lenB != 0)
                {
                    assign_constraint(li, c_triangle[end], iatom[i + 1], iatom[i + 2], lenA, lenB, at2con);
                }
            }
        }
    }
}

//! Sets matrix indices.
static void set_matrix_indices(Lincs                *li,
                               const Task           &li_task,
                               const t_blocka       *at2con,
                               bool                  bSortMatrix)
{
    for (int b = li_task.b0; b < li_task.b1; b++)
    {
        const int a1 = li->atoms[b].index1;
        const int a2 = li->atoms[b].index2;

        int       i = li->blnr[b];
        for (int k = at2con->index[a1]; k < at2con->index[a1 + 1]; k++)
        {
            int concon = li->con_index[at2con->a[k]];
            if (concon != b)
            {
                li->blbnb[i++] = concon;
            }
        }
        for (int k = at2con->index[a2]; k < at2con->index[a2 + 1]; k++)
        {
            int concon = li->con_index[at2con->a[k]];
            if (concon != b)
            {
                li->blbnb[i++] = concon;
            }
        }

        if (bSortMatrix)
        {
            /* Order the blbnb matrix to optimize memory access */
            std::sort(li->blbnb.begin()+li->blnr[b], li->blbnb.begin()+li->blnr[b+1]);
        }
    }
}

void set_lincs(const t_idef         &idef,
               const t_mdatoms      &md,
               bool                  bDynamics,
               const t_commrec      *cr,
               Lincs                *li)
{
    int          natoms;
    t_blocka     at2con;
    t_iatom     *iatom;

    li->nc_real = 0;
    li->nc      = 0;
    li->ncc     = 0;
    /* Zero the thread index ranges.
     * Otherwise without local constraints we could return with old ranges.
     */
    for (int i = 0; i < li->ntask; i++)
    {
        li->task[i].b0   = 0;
        li->task[i].b1   = 0;
        li->task[i].ind.clear();
    }
    if (li->ntask > 1)
    {
        li->task[li->ntask].ind.clear();
    }

    /* This is the local topology, so there are only F_CONSTR constraints */
    if (idef.il[F_CONSTR].nr == 0)
    {
        /* There are no constraints,
         * we do not need to fill any data structures.
         */
        return;
    }

    if (debug)
    {
        fprintf(debug, "Building the LINCS connectivity\n");
    }

    if (DOMAINDECOMP(cr))
    {
        if (cr->dd->constraints)
        {
            int start;

            dd_get_constraint_range(cr->dd, &start, &natoms);
        }
        else
        {
            natoms = dd_numHomeAtoms(*cr->dd);
        }
    }
    else
    {
        natoms = md.homenr;
    }

    at2con = make_at2con(natoms, idef.il, idef.iparams,
                         flexibleConstraintTreatment(bDynamics));

    const int ncon_tot = idef.il[F_CONSTR].nr/3;

    /* Ensure we have enough padding for aligned loads for each thread */
    const int numEntries = ncon_tot + li->ntask*simd_width;
    li->con_index.resize(numEntries);
    li->bllen0.resize(numEntries);
    li->ddist.resize(numEntries);
    li->atoms.resize(numEntries);
    li->blc.resize(numEntries);
    li->blc1.resize(numEntries);
    li->blnr.resize(numEntries + 1);
    li->bllen.resize(numEntries);
    li->tmpv.resizeWithPadding(numEntries);
    if (DOMAINDECOMP(cr))
    {
        li->nlocat.resize(numEntries);
    }
    li->tmp1.resize(numEntries);
    li->tmp2.resize(numEntries);
    li->tmp3.resize(numEntries);
    li->tmp4.resize(numEntries);
    li->mlambda.resize(numEntries);

    iatom = idef.il[F_CONSTR].iatoms;

    li->blnr[0]   = li->ncc;

    /* Assign the constraints for li->ntask LINCS tasks.
     * We target a uniform distribution of constraints over the tasks.
     * Note that when flexible constraints are present, but are removed here
     * (e.g. because we are doing EM) we get imbalance, but since that doesn't
     * happen during normal MD, that's ok.
     */

    /* Determine the number of constraints we need to assign here */
    int ncon_assign = ncon_tot;
    if (!bDynamics)
    {
        /* With energy minimization, flexible constraints are ignored
         * (and thus minimized, as they should be).
         */
        ncon_assign -= countFlexibleConstraints(idef.il, idef.iparams);
    }

    /* Set the target constraint count per task to exactly uniform,
     * this might be overridden below.
     */
    int ncon_target = (ncon_assign + li->ntask - 1)/li->ntask;

    /* Mark all constraints as unassigned by setting their index to -1 */
    for (int con = 0; con < ncon_tot; con++)
    {
        li->con_index[con] = -1;
    }

    int con = 0;
    for (int th = 0; th < li->ntask; th++)
    {
        Task *li_task;

        li_task = &li->task[th];

#if GMX_SIMD_HAVE_REAL
        /* With indepedent tasks we likely have H-bond constraints or constraint
         * pairs. The connected constraints will be pulled into the task, so the
         * constraints per task will often exceed ncon_target.
         * Triangle constraints can also increase the count, but there are
         * relatively few of those, so we usually expect to get ncon_target.
         */
        if (li->bTaskDep)
        {
            /* We round ncon_target to a multiple of GMX_SIMD_WIDTH,
             * since otherwise a lot of operations can be wasted.
             * There are several ways to round here, we choose the one
             * that alternates block sizes, which helps with Intel HT.
             */
            ncon_target = ((ncon_assign*(th + 1))/li->ntask - li->nc_real + GMX_SIMD_REAL_WIDTH - 1) & ~(GMX_SIMD_REAL_WIDTH - 1);
        }
#endif      // GMX_SIMD==2 && GMX_SIMD_HAVE_REAL

        /* Continue filling the arrays where we left off with the previous task,
         * including padding for SIMD.
         */
        li_task->b0 = li->nc;

        while (con < ncon_tot && li->nc - li_task->b0 < ncon_target)
        {
            if (li->con_index[con] == -1)
            {
                int  type, a1, a2;
                real lenA, lenB;

                type   = iatom[3*con];
                a1     = iatom[3*con + 1];
                a2     = iatom[3*con + 2];
                lenA   = idef.iparams[type].constr.dA;
                lenB   = idef.iparams[type].constr.dB;
                /* Skip the flexible constraints when not doing dynamics */
                if (bDynamics || lenA != 0 || lenB != 0)
                {
                    assign_constraint(li, con, a1, a2, lenA, lenB, &at2con);

                    if (li->ntask > 1 && !li->bTaskDep)
                    {
                        /* We can generate independent tasks. Check if we
                         * need to assign connected constraints to our task.
                         */
                        check_assign_connected(li, iatom, idef, bDynamics,
                                               a1, a2, &at2con);
                    }
                    if (li->ntask > 1 && li->ncg_triangle > 0)
                    {
                        /* Ensure constraints in one triangle are assigned
                         * to the same task.
                         */
                        check_assign_triangle(li, iatom, idef, bDynamics,
                                              con, a1, a2, &at2con);
                    }
                }
            }

            con++;
        }

        li_task->b1 = li->nc;

        if (simd_width > 1)
        {
            /* Copy the last atom pair indices and lengths for constraints
             * up to a multiple of simd_width, such that we can do all
             * SIMD operations without having to worry about end effects.
             */
            int i, last;

            li->nc = ((li_task->b1 + simd_width - 1)/simd_width)*simd_width;
            last   = li_task->b1 - 1;
            for (i = li_task->b1; i < li->nc; i++)
            {
                li->atoms[i]     = li->atoms[last];
                li->bllen0[i]    = li->bllen0[last];
                li->ddist[i]     = li->ddist[last];
                li->bllen[i]     = li->bllen[last];
                li->blnr[i + 1]  = li->blnr[last + 1];
            }
        }

        /* Keep track of how many constraints we assigned */
        li->nc_real += li_task->b1 - li_task->b0;

        if (debug)
        {
            fprintf(debug, "LINCS task %d constraints %d - %d\n",
                    th, li_task->b0, li_task->b1);
        }
    }

    assert(li->nc_real == ncon_assign);

    bool bSortMatrix;

    /* Without DD we order the blbnb matrix to optimize memory access.
     * With DD the overhead of sorting is more than the gain during access.
     */
    bSortMatrix = !DOMAINDECOMP(cr);

    li->blbnb.resize(li->ncc);

#pragma omp parallel for num_threads(li->ntask) schedule(static)
    for (int th = 0; th < li->ntask; th++)
    {
        try
        {
            Task &li_task = li->task[th];

            if (li->ncg_triangle > 0)
            {
                /* This is allocating too much, but it is difficult to improve */
                li_task.triangle.resize(li_task.b1 - li_task.b0);
                li_task.tri_bits.resize(li_task.b1 - li_task.b0);
            }

            set_matrix_indices(li, li_task, &at2con, bSortMatrix);
        }
        GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
    }

    done_blocka(&at2con);

    if (cr->dd == nullptr)
    {
        /* Since the matrix is static, we should free some memory */
        li->blbnb.resize(li->ncc);
    }

    li->blmf.resize(li->ncc);
    li->blmf1.resize(li->ncc);
    li->tmpncc.resize(li->ncc);

    gmx::ArrayRef<const int> nlocat_dd = dd_constraints_nlocalatoms(cr->dd);
    if (!nlocat_dd.empty())
    {
        /* Convert nlocat from local topology to LINCS constraint indexing */
        for (con = 0; con < ncon_tot; con++)
        {
            li->nlocat[li->con_index[con]] = nlocat_dd[con];
        }
    }
    else
    {
        li->nlocat.clear();
    }

    if (debug)
    {
        fprintf(debug, "Number of constraints is %d, padded %d, couplings %d\n",
                li->nc_real, li->nc, li->ncc);
    }

    if (li->ntask > 1)
    {
        lincs_thread_setup(li, md.nr);
    }

    set_lincs_matrix(li, md.invmass, md.lambda);
}

//! Issues a warning when LINCS constraints cannot be satisfied.
static void lincs_warning(gmx_domdec_t *dd, const rvec *x, rvec *xprime, t_pbc *pbc,
                          int ncons, gmx::ArrayRef<const AtomPair> atoms,
                          gmx::ArrayRef<const real> bllen, real wangle,
                          int maxwarn, int *warncount)
{
    real wfac = std::cos(DEG2RAD*wangle);

    fprintf(stderr,
            "bonds that rotated more than %g degrees:\n"
            " atom 1 atom 2  angle  previous, current, constraint length\n",
            wangle);

    for (int b = 0; b < ncons; b++)
    {
        const int i = atoms[b].index1;
        const int j = atoms[b].index2;
        rvec      v0;
        rvec      v1;
        if (pbc)
        {
            pbc_dx_aiuc(pbc, x[i], x[j], v0);
            pbc_dx_aiuc(pbc, xprime[i], xprime[j], v1);
        }
        else
        {
            rvec_sub(x[i], x[j], v0);
            rvec_sub(xprime[i], xprime[j], v1);
        }
        real d0     = norm(v0);
        real d1     = norm(v1);
        real cosine = ::iprod(v0, v1)/(d0*d1);
        if (cosine < wfac)
        {
            fprintf(stderr,
                    " %6d %6d  %5.1f  %8.4f %8.4f    %8.4f\n",
                    ddglatnr(dd, i), ddglatnr(dd, j),
                    RAD2DEG*std::acos(cosine), d0, d1, bllen[b]);
            if (!std::isfinite(d1))
            {
                gmx_fatal(FARGS, "Bond length not finite.");
            }

            (*warncount)++;
        }
    }
    if (*warncount > maxwarn)
    {
        too_many_constraint_warnings(econtLINCS, *warncount);
    }
}

//! Status information about how well LINCS satisified the constraints in this domain
struct LincsDeviations
{
    //! The maximum over all bonds in this domain of the relative deviation in bond lengths
    real maxDeviation = 0;
    //! Sum over all bonds in this domain of the squared relative deviation
    real sumSquaredDeviation = 0;
    //! Index of bond with max deviation
    int  indexOfMaxDeviation = -1;
    //! Number of bonds constrained in this domain
    int  numConstraints = 0;
};

//! Determine how well the constraints have been satisfied.
static LincsDeviations
makeLincsDeviations(const Lincs &lincsd,
                    const rvec  *x,
                    const t_pbc *pbc)
{
    LincsDeviations                    result;
    const ArrayRef<const AtomPair>     atoms  = lincsd.atoms;
    const ArrayRef<const real>         bllen  = lincsd.bllen;
    const ArrayRef<const int>          nlocat = lincsd.nlocat;

    for (int task = 0; task < lincsd.ntask; task++)
    {
        for (int b = lincsd.task[task].b0; b < lincsd.task[task].b1; b++)
        {
            rvec dx;
            if (pbc)
            {
                pbc_dx_aiuc(pbc, x[atoms[b].index1], x[atoms[b].index2], dx);
            }
            else
            {
                rvec_sub(x[atoms[b].index1], x[atoms[b].index2], dx);
            }
            real r2  = ::norm2(dx);
            real len = r2*gmx::invsqrt(r2);
            real d   = std::abs(len/bllen[b] - 1.0_real);
            if (d > result.maxDeviation && (nlocat.empty() || nlocat[b]))
            {
                result.maxDeviation        = d;
                result.indexOfMaxDeviation = b;
            }
            if (nlocat.empty())
            {
                result.sumSquaredDeviation += d*d;
                result.numConstraints++;
            }
            else
            {
                result.sumSquaredDeviation += nlocat[b]*d*d;
                result.numConstraints      += nlocat[b];
            }
        }
    }

    if (!nlocat.empty())
    {
        result.numConstraints      /= 2;
        result.sumSquaredDeviation *= 0.5;
    }
    return result;
}

bool constrain_lincs(bool computeRmsd,
                     const t_inputrec &ir,
                     int64_t step,
                     Lincs *lincsd, const t_mdatoms &md,
                     const t_commrec *cr,
                     const gmx_multisim_t *ms,
                     const rvec *x, rvec *xprime, rvec *min_proj,
                     const matrix box, t_pbc *pbc,
                     real lambda, real *dvdlambda,
                     real invdt, rvec *v,
                     bool bCalcVir, tensor vir_r_m_dr,
                     ConstraintVariable econq,
                     t_nrnb *nrnb,
                     int maxwarn, int *warncount)
{
    bool bOK = TRUE;

    /* This boolean should be set by a flag passed to this routine.
     * We can also easily check if any constraint length is changed,
     * if not dH/dlambda=0 and we can also set the boolean to FALSE.
     */
    bool bCalcDHDL = (ir.efep != efepNO && dvdlambda != nullptr);

    if (lincsd->nc == 0 && cr->dd == nullptr)
    {
        if (computeRmsd)
        {
            lincsd->rmsdData = {{0}};
        }

        return bOK;
    }

    if (econq == ConstraintVariable::Positions)
    {
        /* We can't use bCalcDHDL here, since NULL can be passed for dvdlambda
         * also with efep!=fepNO.
         */
        if (ir.efep != efepNO)
        {
            if (md.nMassPerturbed && lincsd->matlam != md.lambda)
            {
                set_lincs_matrix(lincsd, md.invmass, md.lambda);
            }

            for (int i = 0; i < lincsd->nc; i++)
            {
                lincsd->bllen[i] = lincsd->bllen0[i] + lambda*lincsd->ddist[i];
            }
        }

        if (lincsd->ncg_flex)
        {
            /* Set the flexible constraint lengths to the old lengths */
            if (pbc != nullptr)
            {
                for (int i = 0; i < lincsd->nc; i++)
                {
                    if (lincsd->bllen[i] == 0)
                    {
                        rvec dx;
                        pbc_dx_aiuc(pbc, x[lincsd->atoms[i].index1], x[lincsd->atoms[i].index2], dx);
                        lincsd->bllen[i] = norm(dx);
                    }
                }
            }
            else
            {
                for (int i = 0; i < lincsd->nc; i++)
                {
                    if (lincsd->bllen[i] == 0)
                    {
                        lincsd->bllen[i] =
                            std::sqrt(distance2(x[lincsd->atoms[i].index1],
                                                x[lincsd->atoms[i].index2]));
                    }
                }
            }
        }

        const bool printDebugOutput = ((debug != nullptr) && lincsd->nc > 0);
        if (printDebugOutput)
        {
            LincsDeviations deviations = makeLincsDeviations(*lincsd, xprime, pbc);
            fprintf(debug, "   Rel. Constraint Deviation:  RMS         MAX     between atoms\n");
            fprintf(debug, "       Before LINCS          %.6f    %.6f %6d %6d\n",
                    std::sqrt(deviations.sumSquaredDeviation/deviations.numConstraints),
                    deviations.maxDeviation,
                    ddglatnr(cr->dd, lincsd->atoms[deviations.indexOfMaxDeviation].index1),
                    ddglatnr(cr->dd, lincsd->atoms[deviations.indexOfMaxDeviation].index2));
        }

        /* This bWarn var can be updated by multiple threads
         * at the same time. But as we only need to detect
         * if a warning occurred or not, this is not an issue.
         */
        bool bWarn = FALSE;

        /* The OpenMP parallel region of constrain_lincs for coords */
#pragma omp parallel num_threads(lincsd->ntask)
        {
            try
            {
                int th = gmx_omp_get_thread_num();

                clear_mat(lincsd->task[th].vir_r_m_dr);

                do_lincs(x, xprime, box, pbc, lincsd, th,
                         md.invmass, cr,
                         bCalcDHDL,
                         ir.LincsWarnAngle, &bWarn,
                         invdt, v, bCalcVir,
                         th == 0 ? vir_r_m_dr : lincsd->task[th].vir_r_m_dr);
            }
            GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
        }

        if (computeRmsd || printDebugOutput || bWarn)
        {
            LincsDeviations deviations = makeLincsDeviations(*lincsd, xprime, pbc);

            if (computeRmsd)
            {
                // This is reduced across domains in compute_globals and
                // reported to the log file.
                lincsd->rmsdData[0] = deviations.numConstraints;
                lincsd->rmsdData[1] = deviations.sumSquaredDeviation;
            }
            else
            {
                // This is never read
                lincsd->rmsdData = {{0}};
            }
            if (printDebugOutput)
            {
                fprintf(debug,
                        "        After LINCS          %.6f    %.6f %6d %6d\n\n",
                        std::sqrt(deviations.sumSquaredDeviation/deviations.numConstraints),
                        deviations.maxDeviation,
                        ddglatnr(cr->dd, lincsd->atoms[deviations.indexOfMaxDeviation].index1),
                        ddglatnr(cr->dd, lincsd->atoms[deviations.indexOfMaxDeviation].index2));
            }

            if (bWarn)
            {
                if (maxwarn < INT_MAX)
                {
                    std::string simMesg;
                    if (isMultiSim(ms))
                    {
                        simMesg += gmx::formatString(" in simulation %d", ms->sim);
                    }
                    fprintf(stderr,
                            "\nStep %" PRId64 ", time %g (ps)  LINCS WARNING%s\n"
                            "relative constraint deviation after LINCS:\n"
                            "rms %.6f, max %.6f (between atoms %d and %d)\n",
                            step, ir.init_t+step*ir.delta_t,
                            simMesg.c_str(),
                            std::sqrt(deviations.sumSquaredDeviation/deviations.numConstraints),
                            deviations.maxDeviation,
                            ddglatnr(cr->dd, lincsd->atoms[deviations.indexOfMaxDeviation].index1),
                            ddglatnr(cr->dd, lincsd->atoms[deviations.indexOfMaxDeviation].index2));

                    lincs_warning(cr->dd, x, xprime, pbc,
                                  lincsd->nc, lincsd->atoms, lincsd->bllen,
                                  ir.LincsWarnAngle, maxwarn, warncount);
                }
                bOK = (deviations.maxDeviation < 0.5);
            }
        }

        if (lincsd->ncg_flex)
        {
            for (int i = 0; (i < lincsd->nc); i++)
            {
                if (lincsd->bllen0[i] == 0 && lincsd->ddist[i] == 0)
                {
                    lincsd->bllen[i] = 0;
                }
            }
        }
    }
    else
    {
        /* The OpenMP parallel region of constrain_lincs for derivatives */
#pragma omp parallel num_threads(lincsd->ntask)
        {
            try
            {
                int th = gmx_omp_get_thread_num();

                do_lincsp(x, xprime, min_proj, pbc, lincsd, th,
                          md.invmass, econq, bCalcDHDL,
                          bCalcVir, th == 0 ? vir_r_m_dr : lincsd->task[th].vir_r_m_dr);
            }
            GMX_CATCH_ALL_AND_EXIT_WITH_FATAL_ERROR;
        }
    }

    if (bCalcDHDL)
    {
        /* Reduce the dH/dlambda contributions over the threads */
        real dhdlambda;
        int  th;

        dhdlambda = 0;
        for (th = 0; th < lincsd->ntask; th++)
        {
            dhdlambda += lincsd->task[th].dhdlambda;
        }
        if (econq == ConstraintVariable::Positions)
        {
            /* dhdlambda contains dH/dlambda*dt^2, correct for this */
            /* TODO This should probably use invdt, so that sd integrator scaling works properly */
            dhdlambda /= ir.delta_t*ir.delta_t;
        }
        *dvdlambda += dhdlambda;
    }

    if (bCalcVir && lincsd->ntask > 1)
    {
        for (int i = 1; i < lincsd->ntask; i++)
        {
            m_add(vir_r_m_dr, lincsd->task[i].vir_r_m_dr, vir_r_m_dr);
        }
    }

    /* count assuming nit=1 */
    inc_nrnb(nrnb, eNR_LINCS, lincsd->nc_real);
    inc_nrnb(nrnb, eNR_LINCSMAT, (2+lincsd->nOrder)*lincsd->ncc);
    if (lincsd->ntriangle > 0)
    {
        inc_nrnb(nrnb, eNR_LINCSMAT, lincsd->nOrder*lincsd->ncc_triangle);
    }
    if (v)
    {
        inc_nrnb(nrnb, eNR_CONSTR_V, lincsd->nc_real*2);
    }
    if (bCalcVir)
    {
        inc_nrnb(nrnb, eNR_CONSTR_VIR, lincsd->nc_real);
    }

    return bOK;
}

}  // namespace gmx