fieldaligned_tmp.h

Go to the documentation of this file.

#pragma once
#include <cmath>
#include <array>
#include <cusp/csr_matrix.h>
 
#include "dg/algorithm.h"
#include "magnetic_field.h"
#include "fluxfunctions.h"
#include "curvilinear.h"
 
namespace dg{
namespace geo{
 
enum whichMatrix
{
    einsPlus = 0,   
    einsPlusT, 
    einsMinus, 
    einsMinusT,
    zeroPlus,  
    zeroMinus, 
    zeroPlusT, 
    zeroMinusT, 
    zeroForw  
};
 
typedef ONE FullLimiter;
 
typedef ZERO NoLimiter;
namespace detail{
 
 
struct DSFieldCylindrical3
{
    DSFieldCylindrical3( const dg::geo::CylindricalVectorLvl0& v): m_v(v){}
    void operator()( double t, const std::array<double,3>& y,
            std::array<double,3>& yp) const {
        double R = y[0], Z = y[1];
        double vz = m_v.z()(R, Z);
        yp[0] = m_v.x()(R, Z)/vz;
        yp[1] = m_v.y()(R, Z)/vz;
        yp[2] = 1./vz;
    }
    private:
    dg::geo::CylindricalVectorLvl0 m_v;
};
 
struct DSFieldCylindrical4
{
    DSFieldCylindrical4( const dg::geo::CylindricalVectorLvl1& v): m_v(v){}
    void operator()( double t, const std::array<double,3>& y,
            std::array<double,3>& yp) const {
        double R = y[0], Z = y[1];
        double vx = m_v.x()(R,Z);
        double vy = m_v.y()(R,Z);
        double vz = m_v.z()(R,Z);
        double divvvz = m_v.divvvz()(R,Z);
        yp[0] = vx/vz;
        yp[1] = vy/vz;
        yp[2] = divvvz*y[2];
    }
 
    private:
    dg::geo::CylindricalVectorLvl1 m_v;
};
 
struct DSField
{
    DSField() = default;
    //z component of v may not vanish
    DSField( const dg::geo::CylindricalVectorLvl1& v,
            const dg::aGeometry2d& g ):
        m_g(g)
    {
        dg::HVec v_zeta, v_eta;
        dg::pushForwardPerp( v.x(), v.y(), v_zeta, v_eta, g);
        dg::HVec vx = dg::pullback( v.x(), g);
        dg::HVec vy = dg::pullback( v.y(), g);
        dg::HVec vz = dg::pullback( v.z(), g);
        dg::HVec divvvz = dg::pullback( v.divvvz(), g);
        dg::blas1::pointwiseDivide(v_zeta,  vz, v_zeta);
        dg::blas1::pointwiseDivide(v_eta,   vz, v_eta);
        dzetadphi_  = dg::forward_transform( v_zeta, g );
        detadphi_   = dg::forward_transform( v_eta, g );
        dvdphi_     = dg::forward_transform( divvvz, g );
    }
    //interpolate the vectors given in the constructor on the given point
    void operator()(double t, const std::array<double,3>& y, std::array<double,3>& yp) const
    {
        // shift point into domain
        yp[0] = interpolate(dg::lspace, dzetadphi_, y[0], y[1], *m_g);
        yp[1] = interpolate(dg::lspace, detadphi_,  y[0], y[1], *m_g);
        yp[2] = interpolate(dg::lspace, dvdphi_,    y[0], y[1], *m_g)*y[2];
    }
    private:
    thrust::host_vector<double> dzetadphi_, detadphi_, dvdphi_;
    dg::ClonePtr<dg::aGeometry2d> m_g;
};
 
//used in constructor of Fieldaligned
template<class real_type>
void integrate_all_fieldlines2d( const dg::geo::CylindricalVectorLvl1& vec,
    const dg::aRealGeometry2d<real_type>& grid_field,
    const dg::aRealTopology2d<real_type>& grid_evaluate,
    std::array<thrust::host_vector<real_type>,3>& yp,
    const thrust::host_vector<double>& vol0,
    thrust::host_vector<real_type>& yp2b,
    thrust::host_vector<bool>& in_boxp,
    real_type deltaPhi, real_type eps)
{
    //grid_field contains the global geometry for the field and the boundaries
    //grid_evaluate contains the points to actually integrate
    std::array<thrust::host_vector<real_type>,3> y{
        dg::evaluate( dg::cooX2d, grid_evaluate),
        dg::evaluate( dg::cooY2d, grid_evaluate),
        vol0
    };
    yp.fill(dg::evaluate( dg::zero, grid_evaluate));
    //construct field on high polynomial grid, then integrate it
    dg::geo::detail::DSField field;
    if( !dynamic_cast<const dg::CartesianGrid2d*>( &grid_field))
        field = dg::geo::detail::DSField( vec, grid_field);
 
    //field in case of cartesian grid
    dg::geo::detail::DSFieldCylindrical4 cyl_field(vec);
    const unsigned size = grid_evaluate.size();
    dg::Adaptive<dg::ERKStep<std::array<real_type,3>>> adapt(
            "Dormand-Prince-7-4-5", std::array<real_type,3>{0,0,0});
    dg::AdaptiveTimeloop< std::array<real_type,3>> odeint;
    if( dynamic_cast<const dg::CartesianGrid2d*>( &grid_field))
        odeint = dg::AdaptiveTimeloop<std::array<real_type,3>>( adapt,
                cyl_field, dg::pid_control, dg::fast_l2norm, eps, 1e-10);
    else
        odeint = dg::AdaptiveTimeloop<std::array<real_type,3>>( adapt,
                field, dg::pid_control, dg::fast_l2norm, eps, 1e-10);
 
    for( unsigned i=0; i<size; i++)
    {
        std::array<real_type,3> coords{y[0][i],y[1][i],y[2][i]}, coordsP;
        //x,y,s
        real_type phi1 = deltaPhi;
        odeint.set_dt( deltaPhi/2.);
        odeint.integrate( 0, coords, phi1, coordsP);
        yp[0][i] = coordsP[0], yp[1][i] = coordsP[1], yp[2][i] = coordsP[2];
    }
    yp2b.assign( grid_evaluate.size(), deltaPhi); //allocate memory for output
    in_boxp.resize( yp2b.size());
    //Now integrate again but this time find the boundary distance
    for( unsigned i=0; i<size; i++)
    {
        std::array<real_type,3> coords{y[0][i],y[1][i],y[2][i]}, coordsP;
        in_boxp[i] = grid_field.contains( yp[0][i], yp[1][i]) ? true : false;
        if( false == in_boxp[i])
        {
            //x,y,s
            real_type phi1 = deltaPhi;
            odeint.integrate_in_domain( 0., coords, phi1, coordsP, 0., (const
                        dg::aRealTopology2d<real_type>&)grid_field, eps);
            yp2b[i] = phi1;
        }
    }
}
 
 
}//namespace detail
 
struct WallFieldlineDistance : public aCylindricalFunctor<WallFieldlineDistance>
{
    WallFieldlineDistance(
        const dg::geo::CylindricalVectorLvl0& vec,
        const dg::aRealTopology2d<double>& domain,
        double maxPhi, double eps, std::string type) :
        m_domain( domain), m_cyl_field(vec),
        m_deltaPhi( maxPhi), m_eps( eps), m_type(type)
    {
        if( m_type != "phi" && m_type != "s")
            throw std::runtime_error( "Distance type "+m_type+" not recognized!\n");
    }
    double do_compute( double R, double Z) const
    {
        std::array<double,3> coords{ R, Z, 0}, coordsP(coords);
        // determine sign
        m_cyl_field( 0., coords, coordsP);
        double sign = coordsP[2] > 0 ? +1. : -1.;
        double phi1 = sign*m_deltaPhi; // we integrate negative ...
        try{
            dg::Adaptive<dg::ERKStep<std::array<double,3>>> adapt(
                    "Dormand-Prince-7-4-5", coords);
            dg::AdaptiveTimeloop<std::array<double,3>> odeint( adapt,
                    m_cyl_field, dg::pid_control, dg::fast_l2norm, m_eps,
                    1e-10);
            odeint.integrate_in_domain( 0., coords, phi1, coordsP, 0.,
                    m_domain, m_eps);
            //integration
        }catch (std::exception& e)
        {
            // if not possible the distance is large
            //std::cerr << e.what();
            phi1 = sign*m_deltaPhi;
            coordsP[2] = 1e6*phi1;
        }
        if( m_type == "phi")
            return sign*phi1;
        return coordsP[2];
    }
 
    private:
    const dg::Grid2d m_domain;
    dg::geo::detail::DSFieldCylindrical3 m_cyl_field;
    double m_deltaPhi, m_eps;
    std::string m_type;
};
 
struct WallFieldlineCoordinate : public aCylindricalFunctor<WallFieldlineCoordinate>
{
    WallFieldlineCoordinate(
        const dg::geo::CylindricalVectorLvl0& vec,
        const dg::aRealTopology2d<double>& domain,
        double maxPhi, double eps, std::string type) :
        m_domain( domain), m_cyl_field(vec),
        m_deltaPhi( maxPhi), m_eps( eps), m_type(type)
    {
        if( m_type != "phi" && m_type != "s")
            throw std::runtime_error( "Distance type "+m_type+" not recognized!\n");
    }
    double do_compute( double R, double Z) const
    {
        double phiP = m_deltaPhi, phiM = -m_deltaPhi;
        std::array<double,3> coords{ R, Z, 0}, coordsP(coords), coordsM(coords);
        // determine sign
        m_cyl_field( 0., coords, coordsP);
        double sign = coordsP[2] > 0 ? +1. : -1.;
        try{
            dg::AdaptiveTimeloop<std::array<double,3>> odeint(
                    dg::Adaptive<dg::ERKStep<std::array<double,3>>>(
                        "Dormand-Prince-7-4-5", coords), m_cyl_field,
                    dg::pid_control, dg::fast_l2norm, m_eps, 1e-10);
            odeint.integrate_in_domain( 0., coords, phiP, coordsP, 0.,
                    m_domain, m_eps);
            odeint.integrate_in_domain( 0., coords, phiM, coordsM, 0.,
                    m_domain, m_eps);
        }catch (std::exception& e)
        {
            // if not possible the distance is large
            phiP = m_deltaPhi;
            coordsP[2] = 1e6*phiP;
            phiM = -m_deltaPhi;
            coordsM[2] = 1e6*phiM;
        }
        if( m_type == "phi")
            return sign*(-phiP-phiM)/(phiP-phiM);
        double sP = coordsP[2], sM = coordsM[2];
        double value = sign*(-sP-sM)/(sP-sM);
        if( (phiM <= -m_deltaPhi)  && (phiP >= m_deltaPhi))
            return 0.; //return exactly zero if sheath not reached
        if( (phiM <= -m_deltaPhi))
            return value*sign > 0 ? value : 0.; // avoid false negatives
        if( (phiP >= m_deltaPhi))
            return value*sign < 0 ? value : 0.; // avoid false positives
        return value;
    }
 
    private:
    const dg::Grid2d m_domain;
    dg::geo::detail::DSFieldCylindrical3 m_cyl_field;
    double m_deltaPhi, m_eps;
    std::string m_type;
};
 
 
 
template<class ProductGeometry, class IMatrix, class container >
struct Fieldaligned
{
 
    Fieldaligned(){}
    template <class Limiter>
    Fieldaligned(const dg::geo::TokamakMagneticField& vec,
        const ProductGeometry& grid,
        dg::bc bcx = dg::NEU,
        dg::bc bcy = dg::NEU,
        Limiter limit = FullLimiter(),
        double eps = 1e-5,
        unsigned mx=10, unsigned my=10,
        double deltaPhi = -1,
        std::string interpolation_method = "dg",
        bool benchmark=true
        ):
            Fieldaligned( dg::geo::createBHat(vec),
                grid, bcx, bcy, limit, eps, mx, my, deltaPhi, interpolation_method)
    {
    }
 
    template <class Limiter>
    Fieldaligned(const dg::geo::CylindricalVectorLvl1& vec,
        const ProductGeometry& grid,
        dg::bc bcx = dg::NEU,
        dg::bc bcy = dg::NEU,
        Limiter limit = FullLimiter(),
        double eps = 1e-5,
        unsigned mx=10, unsigned my=10,
        double deltaPhi = -1,
        std::string interpolation_method = "dg",
        bool benchmark=true
        );
    template<class ...Params>
    void construct( Params&& ...ps)
    {
        //construct and swap
        *this = Fieldaligned( std::forward<Params>( ps)...);
    }
 
    dg::bc bcx()const{
        return m_bcx;
    }
    dg::bc bcy()const{
        return m_bcy;
    }
 
 
    void set_boundaries( dg::bc bcz, double left, double right)
    {
        m_bcz = bcz;
        const dg::Grid1d g2d( 0, 1, 1, m_perp_size);
        m_left  = dg::evaluate( dg::CONSTANT(left), g2d);
        m_right = dg::evaluate( dg::CONSTANT(right),g2d);
    }
 
    void set_boundaries( dg::bc bcz, const container& left, const container& right)
    {
        m_bcz = bcz;
        m_left = left;
        m_right = right;
    }
 
    void set_boundaries( dg::bc bcz, const container& global, double scal_left, double scal_right)
    {
        dg::split( global, m_f, *m_g);
        dg::blas1::axpby( scal_left,  m_f[0],      0, m_left);
        dg::blas1::axpby( scal_right, m_f[m_Nz-1], 0, m_right);
        m_bcz = bcz;
    }
 
    void operator()(enum whichMatrix which, const container& in, container& out);
 
    double deltaPhi() const{return m_deltaPhi;}
    const container& hbm()const {
        return m_hbm;
    }
    const container& hbp()const {
        return m_hbp;
    }
    const container& sqrtG()const {
        return m_G;
    }
    const container& sqrtGm()const {
        return m_Gm;
    }
    const container& sqrtGp()const {
        return m_Gp;
    }
    const container& bphi()const {
        return m_bphi;
    }
    const container& bphiM()const {
        return m_bphiM;
    }
    const container& bphiP()const {
        return m_bphiP;
    }
    const container& bbm()const {
        return m_bbm;
    }
    const container& bbo()const {
        return m_bbo;
    }
    const container& bbp()const {
        return m_bbp;
    }
    const ProductGeometry& grid()const{return *m_g;}
 
    container interpolate_from_coarse_grid( const ProductGeometry& grid_coarse, const container& coarse);
    void integrate_between_coarse_grid( const ProductGeometry& grid_coarse, const container& coarse, container& out );
 
 
    template< class BinaryOp, class UnaryOp>
    container evaluate( BinaryOp binary, UnaryOp unary,
            unsigned p0, unsigned rounds) const;
 
    std::string method() const{return m_interpolation_method;}
 
    private:
    void ePlus( enum whichMatrix which, const container& in, container& out);
    void eMinus(enum whichMatrix which, const container& in, container& out);
    void zero( enum whichMatrix which, const container& in, container& out);
    IMatrix m_plus, m_minus, m_plusT, m_minusT; //2d interpolation matrices
    container m_tmp; // 3d size
    IMatrix m_back, m_forw;
    IMatrix m_stencil, m_stencilY;
    container m_hbm, m_hbp;         //3d size
    container m_G, m_Gm, m_Gp; // 3d size
    container m_bphi, m_bphiM, m_bphiP; // 3d size
    container m_bbm, m_bbp, m_bbo;  //3d size masks
 
    container m_left, m_right, m_temp2d;      //perp_size
    container m_limiter;            //perp_size
    container m_ghostM, m_ghostP;   //perp_size
    unsigned m_Nz, m_perp_size;
    dg::bc m_bcx, m_bcy, m_bcz;
    std::vector<dg::View<const container>> m_f;
    std::vector<dg::View< container>> m_temp, m_temp2;
    dg::ClonePtr<ProductGeometry> m_g;
    dg::InverseKroneckerTriDiagonal2d<container> m_inv_linear;
    double m_deltaPhi;
    std::string m_interpolation_method;
 
    bool m_have_adjoint = false;
    void updateAdjoint( )
    {
        m_plusT = dg::transpose( m_plus);
        m_minusT = dg::transpose( m_minus);
        m_have_adjoint = true;
    }
};
 
 
template<class Geometry, class IMatrix, class container>
template <class Limiter>
Fieldaligned<Geometry, IMatrix, container>::Fieldaligned(
    const dg::geo::CylindricalVectorLvl1& vec,
    const Geometry& grid,
    dg::bc bcx, dg::bc bcy, Limiter limit, double eps,
    unsigned mx, unsigned my, double deltaPhi, std::string interpolation_method, bool benchmark)
{
    m_stencil = dg::create::limiter_stencil( dg::coo3d::x, grid, grid.bcx());
    m_stencilY = dg::create::limiter_stencil( dg::coo3d::y, grid, grid.bcy());
    if( (grid.bcx() == PER && bcx != PER) || (grid.bcx() != PER && bcx == PER) )
        throw( dg::Error(dg::Message(_ping_)<<"Fieldaligned: Got conflicting periodicity in x. The grid says "<<bc2str(grid.bcx())<<" while the parameter says "<<bc2str(bcx)));
    if( (grid.bcy() == PER && bcy != PER) || (grid.bcy() != PER && bcy == PER) )
        throw( dg::Error(dg::Message(_ping_)<<"Fieldaligned: Got conflicting boundary conditions in y. The grid says "<<bc2str(grid.bcy())<<" while the parameter says "<<bc2str(bcy)));
    m_Nz=grid.Nz(), m_bcx = bcx, m_bcy = bcy, m_bcz=grid.bcz();
    m_g.reset(grid);
    if( deltaPhi <=0) deltaPhi = grid.hz();
    //if( interpolation_method != "dg")
    {
        m_back = dg::create::inv_backproject( grid); //from equidist to dg
        m_forw = dg::create::backproject( grid); // from dg to equidist
    }
    dg::Timer t;
    if( benchmark) t.tic();
    dg::ClonePtr<dg::aGeometry2d> grid_original( grid.perp_grid()) ;
    dg::ClonePtr<dg::aGeometry3d> grid_original3d( grid) ;
    dg::ClonePtr<dg::aGeometry2d> grid_transform( grid_original) ;
    dg::ClonePtr<dg::aGeometry3d> grid_transform3d( grid) ;
    //if( interpolation_method == "dg-const")
    //{
    //    grid_transform->set( 1, grid.gx().size(), grid.gy().size());
    //    grid_transform3d->set( 1, grid.gx().size(), grid.gy().size(), grid.gz().size());
    //}
    dg::ClonePtr<dg::aGeometry2d> grid_magnetic = grid_original;//INTEGRATE HIGH ORDER GRID
    grid_magnetic->set( grid_original->n() < 3 ? 4 : 7, grid_magnetic->Nx(), grid_magnetic->Ny());
    dg::ClonePtr<dg::aGeometry2d> grid_fine;
    if( interpolation_method == "dg-const" || interpolation_method == "dg")
    {
        grid_fine = grid_original;
        if( interpolation_method == "dg-const")
        {
            //mx *= grid.n(), my *= grid.n();
            //grid_original->set( 1, grid.gx().size(), grid.gy().size());
            //grid_original3d->set( 1, grid.gx().size(), grid.gy().size(), grid.gz().size());
            //m_inv_linear = dg::create::inv_fem_linear2const2d( *grid_original3d);
            m_forw = dg::create::smoothing( *grid_transform);
        }
        grid_fine->multiplyCellNumbers((double)mx, (double)my);
    }
    else
    {
        grid_original->set( 1, grid.gx().size(), grid.gy().size());
        grid_original3d->set( 1, grid.gx().size(), grid.gy().size(), grid.gz().size());
 
        if( interpolation_method == "linear")
        {
            m_inv_linear = dg::create::inv_fem_mass2d( *grid_original3d);
            dg::FemRefinement fem_refX( mx), fem_refY(my);
            dg::Grid2d tmp( *grid_original);
            dg::CartesianRefinedGrid2d ref( fem_refX, fem_refY, tmp.x0(), tmp.x1(),
                    tmp.y0(), tmp.y1(), tmp.n(), tmp.Nx(), tmp.Ny(),
                    tmp.bcx(), tmp.bcy());
            grid_fine = ref;
        }
        if ( interpolation_method == "linear-const")
        {
            grid_fine = grid_original;
            grid_fine->multiplyCellNumbers((double)mx, (double)my);
            m_inv_linear = dg::create::inv_fem_linear2const2d( *grid_original3d);
        }
    }
    if( benchmark)
    {
        t.toc();
        std::cout << "# DS: High order grid gen  took: "<<t.diff()<<"\n";
        t.tic();
    }
    std::array<thrust::host_vector<double>,3> yp_trafo, ym_trafo, yp, ym;
    thrust::host_vector<bool> in_boxp, in_boxm;
    thrust::host_vector<double> hbp, hbm;
    thrust::host_vector<double> vol = dg::tensor::volume(grid_transform3d->metric()), vol2d0;
    auto vol2d = dg::split( vol, *grid_transform3d);
    dg::assign( vol2d[0], vol2d0);
    detail::integrate_all_fieldlines2d( vec, *grid_magnetic, *grid_transform,
            yp_trafo, vol2d0, hbp, in_boxp, deltaPhi, eps);
    detail::integrate_all_fieldlines2d( vec, *grid_magnetic, *grid_transform,
            ym_trafo, vol2d0, hbm, in_boxm, -deltaPhi, eps);
    dg::HVec Xf = dg::pullback(  dg::cooX2d, *grid_fine);
    dg::HVec Yf = dg::pullback(  dg::cooY2d, *grid_fine);
    {
    dg::IHMatrix interpolate = dg::create::interpolation( Xf, Yf,
            *grid_transform, dg::NEU, dg::NEU, grid_transform->n() < 3 ? "cubic" : "dg");  //INTERPOLATE TO FINE GRID
    yp.fill(dg::evaluate( dg::zero, *grid_fine));
    ym = yp;
    for( int i=0; i<2; i++) //only R and Z get interpolated
    {
        dg::blas2::symv( interpolate, yp_trafo[i], yp[i]);
        dg::blas2::symv( interpolate, ym_trafo[i], ym[i]);
    }
    } // release memory for interpolate matrix
    if( benchmark)
    {
        t.toc();
        std::cout << "# DS: Computing all points took: "<<t.diff()<<"\n";
        t.tic();
    }
    dg::IHMatrix plusFine, minusFine, plus, minus;
    if( interpolation_method == "dg-const" || interpolation_method == "dg")
    {
        plusFine = dg::create::interpolation( yp[0], yp[1],
                *grid_transform, bcx, bcy, "dg");
        minusFine = dg::create::interpolation( ym[0], ym[1],
                *grid_transform, bcx, bcy, "dg");
    }
    else
    {
        dg::IHMatrix plusFineTmp = dg::create::interpolation( yp[0], yp[1],
                *grid_original, bcx, bcy, "linear");
        dg::IHMatrix minusFineTmp = dg::create::interpolation( ym[0], ym[1],
                *grid_original, bcx, bcy, "linear");
        dg::IHMatrix forw = dg::create::backproject( *grid_transform); // from dg to equidist
        cusp::multiply( plusFineTmp, forw, plusFine);
        cusp::multiply( minusFineTmp, forw, minusFine);
    }
    if( mx == 1 && my == 1)
    {
        plus = plusFine;
        minus = minusFine;
    }
    else
    {
        dg::IHMatrix projection;
        if( interpolation_method == "dg-const")
        {
            //dg::IHMatrix proj = dg::create::transformation( *grid_original, *grid_fine);
            //auto back = dg::create::inv_backproject( *grid_transform);
            //cusp::multiply( back, proj, projection);
            projection = dg::create::projection( *grid_original, *grid_fine);
        }
        else if( interpolation_method == "dg")
            projection = dg::create::projection( *grid_original, *grid_fine);
        else if( interpolation_method == "linear")
        {
            projection = dg::transpose(dg::create::interpolation( Xf,
                    Yf, *grid_original, dg::NEU, dg::NEU, "linear"));
            dg::IHMatrix Wf = dg::create::diagonal( dg::create::volume( *grid_fine));
            dg::IHMatrix Vf = dg::create::diagonal( dg::create::fem_inv_weights( *grid_original));
            dg::IHMatrix tmp;
            cusp::multiply( projection, Wf, tmp);
            cusp::multiply( Vf, tmp, Wf);
            auto back = dg::create::inv_backproject( *grid_transform);
            cusp::multiply( back, Wf, projection);
        }
        else
        {
            dg::IHMatrix proj = dg::create::projection( *grid_original, *grid_fine);
            auto back = dg::create::inv_backproject( *grid_transform);
            cusp::multiply( back, proj, projection);
        }
        cusp::multiply( projection, plusFine, plus);
        cusp::multiply( projection, minusFine, minus);
    }
    if( benchmark)
    {
        t.toc();
        std::cout << "# DS: Multiplication PI    took: "<<t.diff()<<"\n";
    }
    dg::blas2::transfer( plus, m_plus);
    dg::blas2::transfer( minus, m_minus);
    dg::HVec hbphi( yp_trafo[2]), hbphiP(hbphi), hbphiM(hbphi);
    hbphi = dg::pullback( vec.z(), *grid_transform);
    //this is a pullback bphi( R(zeta, eta), Z(zeta, eta)):
    if( dynamic_cast<const dg::CartesianGrid2d*>( grid_transform.get()))
    {
        for( unsigned i=0; i<hbphiP.size(); i++)
        {
            hbphiP[i] = vec.z()(yp_trafo[0][i], yp_trafo[1][i]);
            hbphiM[i] = vec.z()(ym_trafo[0][i], ym_trafo[1][i]);
        }
    }
    else
    {
        dg::HVec Ihbphi = dg::pullback( vec.z(), *grid_magnetic);
        dg::HVec Lhbphi = dg::forward_transform( Ihbphi, *grid_magnetic);
        for( unsigned i=0; i<yp_trafo[0].size(); i++)
        {
            hbphiP[i] = dg::interpolate( dg::lspace, Lhbphi, yp_trafo[0][i],
                    yp_trafo[1][i], *grid_magnetic);
            hbphiM[i] = dg::interpolate( dg::lspace, Lhbphi, ym_trafo[0][i],
                    ym_trafo[1][i], *grid_magnetic);
        }
    }
    dg::assign3dfrom2d( hbphi,  m_bphi,  *grid_transform3d);
    dg::assign3dfrom2d( hbphiM, m_bphiM, *grid_transform3d);
    dg::assign3dfrom2d( hbphiP, m_bphiP, *grid_transform3d);
 
    dg::assign3dfrom2d( yp_trafo[2], m_Gp, *grid_transform3d);
    dg::assign3dfrom2d( ym_trafo[2], m_Gm, *grid_transform3d);
    // The weights don't matter since they fall out in Div and Lap anyway
    // But they are good for testing
    m_G = vol;
    container weights = dg::create::weights( *grid_transform3d);
    dg::blas1::pointwiseDot( m_G, weights, m_G);
    dg::blas1::pointwiseDot( m_Gp, weights, m_Gp);
    dg::blas1::pointwiseDot( m_Gm, weights, m_Gm);
 
    dg::assign( dg::evaluate( dg::zero, *grid_transform3d), m_hbm);
    m_f     = dg::split( (const container&)m_hbm, *grid_transform3d);
    m_temp  = dg::split( m_hbm, *grid_transform3d);
    m_temp2  = dg::split( m_hbm, *grid_transform3d);
    dg::assign3dfrom2d( hbp, m_hbp, *grid_transform3d);
    dg::assign3dfrom2d( hbm, m_hbm, *grid_transform3d);
    dg::blas1::scal( m_hbm, -1.);
 
    thrust::host_vector<double> bbm( in_boxp.size(),0.), bbo(bbm), bbp(bbm);
    for( unsigned i=0; i<in_boxp.size(); i++)
    {
        if( !in_boxp[i] && !in_boxm[i])
            bbo[i] = 1.;
        else if( !in_boxp[i] && in_boxm[i])
            bbp[i] = 1.;
        else if( in_boxp[i] && !in_boxm[i])
            bbm[i] = 1.;
        // else all are 0
    }
    dg::assign3dfrom2d( bbm, m_bbm, *grid_transform3d);
    dg::assign3dfrom2d( bbo, m_bbo, *grid_transform3d);
    dg::assign3dfrom2d( bbp, m_bbp, *grid_transform3d);
 
    m_deltaPhi = deltaPhi; // store for evaluate
    m_tmp = m_bbm;
    m_interpolation_method = interpolation_method;
 
    m_perp_size = grid_transform->size();
    dg::assign( dg::pullback(limit, *grid_transform), m_limiter);
    dg::assign( dg::evaluate(dg::zero, *grid_transform), m_left);
    m_temp2d = m_ghostM = m_ghostP = m_right = m_left;
}
 
 
template<class G, class I, class container>
container Fieldaligned<G, I,container>::interpolate_from_coarse_grid(
        const G& grid, const container& in)
{
    //I think we need grid as input to split input vector and we need to interpret
    //the grid nodes as node centered not cell-centered!
    //idea: apply I+/I- cphi - 1 times in each direction and then apply interpolation formula
    assert( m_g->Nz() % grid.Nz() == 0);
    unsigned Nz_coarse = grid.Nz(), Nz = m_g->Nz();
    unsigned cphi = Nz / Nz_coarse;
 
    container out = dg::evaluate( dg::zero, *m_g);
    container helper = dg::evaluate( dg::zero, *m_g);
    dg::split( helper, m_temp, *m_g);
    std::vector<dg::View< container>> out_split = dg::split( out, *m_g);
    std::vector<dg::View< const container>> in_split = dg::split( in, grid);
    for ( int i=0; i<(int)Nz_coarse; i++)
    {
        //1. copy input vector to appropriate place in output
        dg::blas1::copy( in_split[i], out_split[i*cphi]);
        dg::blas1::copy( in_split[i], m_temp[i*cphi]);
    }
    //Step 1 needs to finish so that m_temp contains values everywhere
    //2. Now apply plus and minus T to fill in the rest
    for ( int i=0; i<(int)Nz_coarse; i++)
    {
        for( int j=1; j<(int)cphi; j++)
        {
            dg::blas2::symv( m_minus, out_split[i*cphi+j-1], out_split[i*cphi+j]);
            dg::blas2::symv( m_plus, m_temp[(i*cphi+cphi+1-j)%Nz], m_temp[i*cphi+cphi-j]);
        }
    }
    //3. Now add up with appropriate weights
    for( int i=0; i<(int)Nz_coarse; i++)
        for( int j=1; j<(int)cphi; j++)
        {
            double alpha = (double)(cphi-j)/(double)cphi;
            double beta = (double)j/(double)cphi;
            dg::blas1::axpby( alpha, out_split[i*cphi+j], beta, m_temp[i*cphi+j], out_split[i*cphi+j]);
        }
    return out;
}
template<class G, class I, class container>
void Fieldaligned<G, I,container>::integrate_between_coarse_grid( const G& grid, const container& in, container& out)
{
    assert( m_g->Nz() % grid.Nz() == 0);
    unsigned Nz_coarse = grid.Nz(), Nz = m_g->Nz();
    unsigned cphi = Nz / Nz_coarse;
 
    out = in;
    container helperP( in), helperM(in), tempP(in), tempM(in);
 
    //1. Apply plus and minus T and sum up
    for( int j=1; j<(int)cphi; j++)
    {
        dg::blas2::symv( m_minus, helperP, tempP);
        dg::blas1::axpby( (double)(cphi-j)/(double)cphi, tempP, 1., out  );
        helperP.swap(tempP);
        dg::blas2::symv( m_plus, helperM, tempM);
        dg::blas1::axpby( (double)(cphi-j)/(double)cphi, tempM, 1., out  );
        helperM.swap(tempM);
    }
    dg::blas1::scal( out, 1./(double)cphi);
}
 
template<class G, class I, class container>
void Fieldaligned<G, I, container >::operator()(enum whichMatrix which, const container& f, container& fe)
{
    if(     which == einsPlus  || which == einsMinusT ) ePlus(  which, f, fe);
    else if(which == einsMinus || which == einsPlusT  ) eMinus( which, f, fe);
    else if(which == zeroMinus || which == zeroPlus ||
            which == zeroMinusT|| which == zeroPlusT ||
            which == zeroForw  ) zero(   which, f, fe);
    if( m_interpolation_method == "dg-const")
    {
        //dg::split( m_tmp, m_temp2, *m_g);
        //dg::split( fe, m_temp, *m_g);
        //for( unsigned i0 = 0; i0 <m_Nz; i0++)
        //    dg::blas2::symv( m_forw, m_temp[i0], m_temp2[i0]);
        //m_tmp.swap(fe);
        dg::blas2::stencil( dg::CSRSlopeLimiter<double>(10), m_stencil, fe, m_tmp);
        dg::blas2::stencil( dg::CSRSlopeLimiter<double>(10), m_stencilY, m_tmp, fe);
    }
}
 
template< class G, class I, class container>
void Fieldaligned<G, I, container>::zero( enum whichMatrix which,
        const container& f, container& f0)
{
    dg::split( f, m_f, *m_g);
    dg::split( f0, m_temp, *m_g);
    //1. compute 2d interpolation in every plane and store in m_temp
    for( unsigned i0=0; i0<m_Nz; i0++)
    {
        if(which == zeroPlus)
            dg::blas2::symv( m_plus,   m_f[i0], m_temp[i0]);
        else if(which == zeroMinus)
            dg::blas2::symv( m_minus,  m_f[i0], m_temp[i0]);
        else if(which == zeroPlusT)
        {
            if( ! m_have_adjoint) updateAdjoint( );
            dg::blas2::symv( m_plusT,  m_f[i0], m_temp[i0]);
        }
        else if(which == zeroMinusT)
        {
            if( ! m_have_adjoint) updateAdjoint( );
            dg::blas2::symv( m_minusT, m_f[i0], m_temp[i0]);
        }
    }
    if( which == zeroForw)
    {
        if ( m_interpolation_method == "linear-const" )
        {
            dg::blas2::symv( m_forw, f, f0);
            dg::blas2::symv( m_inv_linear.tri(), f0, m_tmp);
            dg::blas2::symv( m_back, m_tmp, f0);
        }
        else if ( m_interpolation_method == "linear" )
        {
            dg::blas2::symv( m_forw, f, f0);
            dg::blas2::symv( m_inv_linear.tri(), f0, m_tmp);
            dg::blas2::symv( m_back, m_tmp, f0);
        }
        else
            dg::blas1::copy( f, f0);
    }
}
template< class G, class I, class container>
void Fieldaligned<G, I, container>::ePlus( enum whichMatrix which,
        const container& f, container& fpe)
{
    dg::split( f, m_f, *m_g);
    dg::split( fpe, m_temp, *m_g);
    //1. compute 2d interpolation in every plane and store in m_temp
    for( unsigned i0=0; i0<m_Nz; i0++)
    {
        unsigned ip = (i0==m_Nz-1) ? 0:i0+1;
        if(which == einsPlus)
            dg::blas2::symv( m_plus,   m_f[ip], m_temp[i0]);
        else if(which == einsMinusT)
        {
            if( ! m_have_adjoint) updateAdjoint( );
            dg::blas2::symv( m_minusT, m_f[ip], m_temp[i0]);
        }
    }
    //2. apply right boundary conditions in last plane
    unsigned i0=m_Nz-1;
    if( m_bcz != dg::PER)
    {
        if( m_bcz == dg::DIR || m_bcz == dg::NEU_DIR)
            dg::blas1::axpby( 2, m_right, -1., m_f[i0], m_ghostP);
        if( m_bcz == dg::NEU || m_bcz == dg::DIR_NEU)
            dg::blas1::axpby( m_deltaPhi, m_right, 1., m_f[i0], m_ghostP);
        //interlay ghostcells with periodic cells: L*g + (1-L)*fpe
        dg::blas1::axpby( 1., m_ghostP, -1., m_temp[i0], m_ghostP);
        dg::blas1::pointwiseDot( 1., m_limiter, m_ghostP, 1., m_temp[i0]);
    }
}
 
template< class G, class I, class container>
void Fieldaligned<G, I, container>::eMinus( enum whichMatrix which,
        const container& f, container& fme)
{
    dg::split( f, m_f, *m_g);
    dg::split( fme, m_temp, *m_g);
    //1. compute 2d interpolation in every plane and store in m_temp
    for( unsigned i0=0; i0<m_Nz; i0++)
    {
        unsigned im = (i0==0) ? m_Nz-1:i0-1;
        if(which == einsPlusT)
        {
            if( ! m_have_adjoint) updateAdjoint( );
            dg::blas2::symv( m_plusT, m_f[im], m_temp[i0]);
        }
        else if (which == einsMinus)
            dg::blas2::symv( m_minus, m_f[im], m_temp[i0]);
    }
    //2. apply left boundary conditions in first plane
    unsigned i0=0;
    if( m_bcz != dg::PER)
    {
        if( m_bcz == dg::DIR || m_bcz == dg::DIR_NEU)
            dg::blas1::axpby( 2., m_left,  -1., m_f[i0], m_ghostM);
        if( m_bcz == dg::NEU || m_bcz == dg::NEU_DIR)
            dg::blas1::axpby( -m_deltaPhi, m_left, 1., m_f[i0], m_ghostM);
        //interlay ghostcells with periodic cells: L*g + (1-L)*fme
        dg::blas1::axpby( 1., m_ghostM, -1., m_temp[i0], m_ghostM);
        dg::blas1::pointwiseDot( 1., m_limiter, m_ghostM, 1., m_temp[i0]);
    }
}
 
template<class G, class I, class container>
template< class BinaryOp, class UnaryOp>
container Fieldaligned<G, I,container>::evaluate( BinaryOp binary,
        UnaryOp unary, unsigned p0, unsigned rounds) const
{
    //idea: simply apply I+/I- enough times on the init2d vector to get the result in each plane
    //unary function is always such that the p0 plane is at x=0
    assert( p0 < m_g->Nz());
    const dg::ClonePtr<aGeometry2d> g2d = m_g->perp_grid();
    container init2d = dg::pullback( binary, *g2d);
    container zero2d = dg::evaluate( dg::zero, *g2d);
 
    container temp(init2d), tempP(init2d), tempM(init2d);
    container vec3d = dg::evaluate( dg::zero, *m_g);
    std::vector<container>  plus2d(m_Nz, zero2d), minus2d(plus2d), result(plus2d);
    unsigned turns = rounds;
    if( turns ==0) turns++;
    //first apply Interpolation many times, scale and store results
    for( unsigned r=0; r<turns; r++)
        for( unsigned i0=0; i0<m_Nz; i0++)
        {
            dg::blas1::copy( init2d, tempP);
            dg::blas1::copy( init2d, tempM);
            unsigned rep = r*m_Nz + i0;
            for(unsigned k=0; k<rep; k++)
            {
                dg::blas2::symv( m_minus, tempP, temp);
                temp.swap( tempP);
                dg::blas2::symv( m_plus, tempM, temp);
                temp.swap( tempM);
            }
            dg::blas1::scal( tempP, unary(  (double)rep*m_deltaPhi ) );
            dg::blas1::scal( tempM, unary( -(double)rep*m_deltaPhi ) );
            dg::blas1::axpby( 1., tempP, 1., plus2d[i0]);
            dg::blas1::axpby( 1., tempM, 1., minus2d[i0]);
        }
    //now we have the plus and the minus filaments
    if( rounds == 0) //there is a limiter
    {
        for( unsigned i0=0; i0<m_Nz; i0++)
        {
            int idx = (int)i0 - (int)p0;
            if(idx>=0)
                result[i0] = plus2d[idx];
            else
                result[i0] = minus2d[abs(idx)];
            thrust::copy( result[i0].begin(), result[i0].end(), vec3d.begin() + i0*m_perp_size);
        }
    }
    else //sum up plus2d and minus2d
    {
        for( unsigned i0=0; i0<m_Nz; i0++)
        {
            unsigned revi0 = (m_Nz - i0)%m_Nz; //reverted index
            dg::blas1::axpby( 1., plus2d[i0], 0., result[i0]);
            dg::blas1::axpby( 1., minus2d[revi0], 1., result[i0]);
        }
        dg::blas1::axpby( -1., init2d, 1., result[0]);
        for(unsigned i0=0; i0<m_Nz; i0++)
        {
            int idx = ((int)i0 -(int)p0 + m_Nz)%m_Nz; //shift index
            thrust::copy( result[idx].begin(), result[idx].end(), vec3d.begin() + i0*m_perp_size);
        }
    }
    return vec3d;
}
 
 
 
template<class BinaryOp, class UnaryOp>
thrust::host_vector<double> fieldaligned_evaluate(
        const aProductGeometry3d& grid,
        const CylindricalVectorLvl0& vec,
        const BinaryOp& binary,
        const UnaryOp& unary,
        unsigned p0,
        unsigned rounds,
        double eps = 1e-5)
{
    unsigned Nz = grid.Nz();
    const dg::ClonePtr<aGeometry2d> g2d = grid.perp_grid();
    // Construct for field-aligned output
    dg::HVec tempP = dg::evaluate( dg::zero, *g2d), tempM( tempP);
    std::vector<dg::HVec>  plus2d(Nz, tempP), minus2d(plus2d), result(plus2d);
    dg::HVec vec3d = dg::evaluate( dg::zero, grid);
    dg::HVec init2d = dg::pullback( binary, *g2d);
    std::array<dg::HVec,3> yy0{
        dg::pullback( dg::cooX2d, *g2d),
        dg::pullback( dg::cooY2d, *g2d),
        dg::evaluate( dg::zero, *g2d)}, yy1(yy0), xx0( yy0), xx1(yy0); //s
    dg::geo::detail::DSFieldCylindrical3 cyl_field(vec);
    double deltaPhi = grid.hz();
    double phiM0 = 0., phiP0 = 0.;
    unsigned turns = rounds;
    if( turns == 0) turns++;
    for( unsigned r=0; r<turns; r++)
        for( unsigned  i0=0; i0<Nz; i0++)
        {
            unsigned rep = r*Nz + i0;
            if( rep == 0)
                tempM = tempP = init2d;
            else
            {
                dg::Adaptive<dg::ERKStep<std::array<double,3>>> adapt(
                        "Dormand-Prince-7-4-5", std::array<double,3>{0,0,0});
                dg::AdaptiveTimeloop<std::array<double,3>> odeint( adapt,
                        cyl_field, dg::pid_control, dg::fast_l2norm, eps,
                        1e-10);
                for( unsigned i=0; i<g2d->size(); i++)
                {
                    // minus direction needs positive integration!
                    double phiM1 = phiM0 + deltaPhi;
                    std::array<double,3>
                        coords0{yy0[0][i],yy0[1][i],yy0[2][i]}, coords1;
                    odeint.integrate_in_domain( phiM0, coords0, phiM1, coords1,
                            deltaPhi, *g2d, eps);
                    yy1[0][i] = coords1[0], yy1[1][i] = coords1[1], yy1[2][i] =
                        coords1[2];
                    tempM[i] = binary( yy1[0][i], yy1[1][i]);
 
                    // plus direction needs negative integration!
                    double phiP1 = phiP0 - deltaPhi;
                    coords0 = std::array<double,3>{xx0[0][i],xx0[1][i],xx0[2][i]};
                    odeint.integrate_in_domain( phiP0, coords0, phiP1, coords1,
                            -deltaPhi, *g2d, eps);
                    xx1[0][i] = coords1[0], xx1[1][i] = coords1[1], xx1[2][i] =
                        coords1[2];
                    tempP[i] = binary( xx1[0][i], xx1[1][i]);
                }
                std::swap( yy0, yy1);
                std::swap( xx0, xx1);
                phiM0 += deltaPhi;
                phiP0 -= deltaPhi;
            }
            dg::blas1::scal( tempM, unary( -(double)rep*deltaPhi ) );
            dg::blas1::scal( tempP, unary(  (double)rep*deltaPhi ) );
            dg::blas1::axpby( 1., tempM, 1., minus2d[i0]);
            dg::blas1::axpby( 1., tempP, 1., plus2d[i0]);
        }
    //now we have the plus and the minus filaments
    if( rounds == 0) //there is a limiter
    {
        for( unsigned i0=0; i0<Nz; i0++)
        {
            int idx = (int)i0 - (int)p0;
            if(idx>=0)
                result[i0] = plus2d[idx];
            else
                result[i0] = minus2d[abs(idx)];
            thrust::copy( result[i0].begin(), result[i0].end(), vec3d.begin() +
                    i0*g2d->size());
        }
    }
    else //sum up plus2d and minus2d
    {
        for( unsigned i0=0; i0<Nz; i0++)
        {
            unsigned revi0 = (Nz - i0)%Nz; //reverted index
            dg::blas1::axpby( 1., plus2d[i0], 0., result[i0]);
            dg::blas1::axpby( 1., minus2d[revi0], 1., result[i0]);
        }
        dg::blas1::axpby( -1., init2d, 1., result[0]);
        for(unsigned i0=0; i0<Nz; i0++)
        {
            int idx = ((int)i0 -(int)p0 + Nz)%Nz; //shift index
            thrust::copy( result[idx].begin(), result[idx].end(), vec3d.begin()
                    + i0*g2d->size());
        }
    }
    return vec3d;
}
 
}//namespace geo
}//namespace dg