elliptic.h

Go to the documentation of this file.

#pragma once
 
#include "blas.h"
#include "enums.h"
#include "backend/memory.h"
#include "topology/evaluation.h"
#include "topology/derivativesA.h"
#ifdef MPI_VERSION
#include "topology/mpi_evaluation.h"
#endif
#include "topology/geometry.h"
 
namespace dg
{
    //TODO Elliptic can be made complex aware with a 2nd complex ContainerType
// Note that there are many tests for this file : elliptic2d_b,
// elliptic2d_mpib, elliptic_b, elliptic_mpib, ellipticX2d_b
// And don't forget inc/geometries/elliptic3d_t (testing alignment and
// projection tensors as Chi) geometry_elliptic_b, geometry_elliptic_mpib,
// and geometryX_elliptic_b and geometryX_refined_elliptic_b
 
 
template <class Geometry, class Matrix, class Container>
class Elliptic1d
{
    public:
    using geometry_type = Geometry;
    using matrix_type = Matrix;
    using container_type = Container;
    using value_type = get_value_type<Container>;
    Elliptic1d() = default;
    Elliptic1d( const Geometry& g,
        direction dir = forward, value_type jfactor=1.):
        Elliptic1d( g, g.bcx(), dir, jfactor)
    {
    }
 
    Elliptic1d( const Geometry& g, bc bcx,
        direction dir = forward,
        value_type jfactor=1.)
    {
        m_jfactor=jfactor;
        dg::blas2::transfer( dg::create::dx( g, inverse( bcx), inverse(dir)), m_leftx);
        dg::blas2::transfer( dg::create::dx( g, bcx, dir), m_rightx);
        dg::blas2::transfer( dg::create::jumpX( g, bcx),   m_jumpX);
 
        dg::assign( dg::create::weights(g),       m_weights);
        dg::assign( dg::evaluate( dg::one, g),    m_precond);
        m_tempx = m_sigma = m_precond;
    }
 
    template<class ...Params>
    void construct( Params&& ...ps)
    {
        //construct and swap
        *this = Elliptic1d( std::forward<Params>( ps)...);
    }
 
    template<class ContainerType0>
    void set_chi( const ContainerType0& sigma)
    {
        dg::blas1::copy( sigma, m_sigma);
        //update preconditioner
        dg::blas1::pointwiseDivide( 1., sigma, m_precond);
        // sigma is possibly zero, which will invalidate the preconditioner
        // it is important to call this blas1 function because it can
        // overwrite NaN in m_precond in the next update
    }
 
    const Container& weights()const {
        return m_weights;
    }
    const Container& precond()const {
        return m_precond;
    }
    void set_jfactor( value_type new_jfactor) {m_jfactor = new_jfactor;}
    value_type get_jfactor() const {return m_jfactor;}
    template<class ContainerType0, class ContainerType1>
    void operator()( const ContainerType0& x, ContainerType1& y){
        symv( 1, x, 0, y);
    }
 
    template<class ContainerType0, class ContainerType1>
    void symv( const ContainerType0& x, ContainerType1& y){
        symv( 1, x, 0, y);
    }
    template<class ContainerType0, class ContainerType1>
    void symv( value_type alpha, const ContainerType0& x, value_type beta, ContainerType1& y)
    {
        dg::blas2::gemv( m_rightx, x, m_tempx);
        dg::blas1::pointwiseDot( m_tempx, m_sigma, m_tempx);
        dg::blas2::symv( -alpha, m_leftx, m_tempx, beta, y);
        //add jump terms
        if( 0.0 != m_jfactor )
        {
            dg::blas2::symv( m_jfactor*alpha, m_jumpX, x, 1., y);
        }
    }
 
    private:
    Matrix m_leftx, m_rightx, m_jumpX;
    Container m_weights, m_precond;
    Container m_tempx;
    Container m_sigma;
    value_type m_jfactor;
};
 
template <class Geometry, class Matrix, class Container>
class Elliptic2d
{
    public:
    using geometry_type = Geometry;
    using matrix_type = Matrix;
    using container_type = Container;
    using value_type = get_value_type<Container>;
    Elliptic2d() = default;
    Elliptic2d( const Geometry& g,
        direction dir = forward, value_type jfactor=1., bool chi_weight_jump = false):
        Elliptic2d( g, g.bcx(), g.bcy(), dir, jfactor, chi_weight_jump)
    {
    }
 
    Elliptic2d( const Geometry& g, bc bcx, bc bcy,
        direction dir = forward,
        value_type jfactor=1., bool chi_weight_jump = false)
    {
        m_jfactor=jfactor;
        m_chi_weight_jump = chi_weight_jump;
        dg::blas2::transfer( dg::create::dx( g, inverse( bcx), inverse(dir)), m_leftx);
        dg::blas2::transfer( dg::create::dy( g, inverse( bcy), inverse(dir)), m_lefty);
        dg::blas2::transfer( dg::create::dx( g, bcx, dir), m_rightx);
        dg::blas2::transfer( dg::create::dy( g, bcy, dir), m_righty);
        dg::blas2::transfer( dg::create::jumpX( g, bcx),   m_jumpX);
        dg::blas2::transfer( dg::create::jumpY( g, bcy),   m_jumpY);
 
        dg::assign( dg::create::volume(g),        m_weights);
        dg::assign( dg::evaluate( dg::one, g),    m_precond);
        m_temp = m_tempx = m_tempy = m_weights;
        m_chi=g.metric();
        m_sigma = m_vol = dg::tensor::volume(m_chi);
    }
 
    template<class ...Params>
    void construct( Params&& ...ps)
    {
        //construct and swap
        *this = Elliptic2d( std::forward<Params>( ps)...);
    }
 
    template<class ContainerType0>
    void set_chi( const ContainerType0& sigma)
    {
        dg::blas1::pointwiseDot( sigma, m_vol, m_sigma);
        //update preconditioner
        dg::blas1::pointwiseDivide( 1., sigma, m_precond);
        // sigma is possibly zero, which will invalidate the preconditioner
        // it is important to call this blas1 function because it can
        // overwrite NaN in m_precond in the next update
    }
    template<class ContainerType0>
    void set_chi( const SparseTensor<ContainerType0>& tau)
    {
        m_chi = SparseTensor<Container>(tau);
    }
 
    const Container& weights()const {
        return m_weights;
    }
    const Container& precond()const {
        return m_precond;
    }
    void set_jfactor( value_type new_jfactor) {m_jfactor = new_jfactor;}
    value_type get_jfactor() const {return m_jfactor;}
    void set_jump_weighting( bool jump_weighting) {m_chi_weight_jump = jump_weighting;}
    bool get_jump_weighting() const {return m_chi_weight_jump;}
    template<class ContainerType0, class ContainerType1>
    void operator()( const ContainerType0& x, ContainerType1& y){
        symv( 1, x, 0, y);
    }
 
    template<class ContainerType0, class ContainerType1>
    void symv( const ContainerType0& x, ContainerType1& y){
        symv( 1, x, 0, y);
    }
    template<class ContainerType0, class ContainerType1>
    void symv( value_type alpha, const ContainerType0& x, value_type beta, ContainerType1& y)
    {
        //compute gradient
        dg::blas2::gemv( m_rightx, x, m_tempx); //R_x*f
        dg::blas2::gemv( m_righty, x, m_tempy); //R_y*f
 
        //multiply with tensor (note the alias)
        dg::tensor::multiply2d(m_sigma, m_chi, m_tempx, m_tempy, 0., m_tempx, m_tempy);
 
        //now take divergence
        dg::blas2::symv( m_lefty, m_tempy, m_temp);
        dg::blas2::symv( -1., m_leftx, m_tempx, -1., m_temp);
 
        //add jump terms
        if( 0.0 != m_jfactor )
        {
            if(m_chi_weight_jump)
            {
                dg::blas2::symv( m_jfactor, m_jumpX, x, 0., m_tempx);
                dg::blas2::symv( m_jfactor, m_jumpY, x, 0., m_tempy);
                dg::tensor::multiply2d(m_sigma, m_chi, m_tempx, m_tempy, 0., m_tempx, m_tempy);
                dg::blas1::axpbypgz(1.0,m_tempx,1.0,m_tempy,1.0,m_temp);
            }
            else
            {
                dg::blas2::symv( m_jfactor, m_jumpX, x, 1., m_temp);
                dg::blas2::symv( m_jfactor, m_jumpY, x, 1., m_temp);
            }
        }
        dg::blas1::pointwiseDivide( alpha, m_temp, m_vol, beta, y);
    }
 
    template<class ContainerType0, class ContainerType1>
    void variation(const ContainerType0& phi, ContainerType1& sigma){
        variation(1., 1., phi, 0., sigma);
    }
    template<class ContainerTypeL, class ContainerType0, class ContainerType1>
    void variation(const ContainerTypeL& lambda, const ContainerType0& phi, ContainerType1& sigma){
        variation(1.,lambda, phi, 0., sigma);
    }
    template<class ContainerTypeL, class ContainerType0, class ContainerType1>
    void variation(value_type alpha, const ContainerTypeL& lambda, const ContainerType0& phi, value_type beta, ContainerType1& sigma)
    {
        dg::blas2::gemv( m_rightx, phi, m_tempx); //R_x*f
        dg::blas2::gemv( m_righty, phi, m_tempy); //R_y*f
        dg::tensor::scalar_product2d(alpha, lambda, m_tempx, m_tempy, m_chi, lambda, m_tempx, m_tempy, beta, sigma);
    }
 
 
    private:
    Matrix m_leftx, m_lefty, m_rightx, m_righty, m_jumpX, m_jumpY;
    Container m_weights, m_precond;
    Container m_tempx, m_tempy, m_temp;
    SparseTensor<Container> m_chi;
    Container m_sigma, m_vol;
    value_type m_jfactor;
    bool m_chi_weight_jump;
};
 
template <class Geometry, class Matrix, class Container>
using Elliptic = Elliptic2d<Geometry, Matrix, Container>;
 
//Elliptic3d is tested in inc/geometries/elliptic3d_t.cu
template <class Geometry, class Matrix, class Container>
class Elliptic3d
{
    public:
    using geometry_type = Geometry;
    using matrix_type = Matrix;
    using container_type = Container;
    using value_type = get_value_type<Container>;
    Elliptic3d() = default;
    Elliptic3d( const Geometry& g, direction dir = forward, value_type jfactor=1., bool chi_weight_jump = false):
        Elliptic3d( g, g.bcx(), g.bcy(), g.bcz(), dir, jfactor, chi_weight_jump)
    {
    }
 
    Elliptic3d( const Geometry& g, bc bcx, bc bcy, bc bcz, direction dir = forward, value_type jfactor = 1., bool chi_weight_jump = false)
    {
        // MW we should create an if guard for nx, ny, or nz = 1 and periodic boundaries
        m_jfactor=jfactor;
        m_chi_weight_jump = chi_weight_jump;
        dg::blas2::transfer( dg::create::dx( g, inverse( bcx), inverse(dir)), m_leftx);
        dg::blas2::transfer( dg::create::dy( g, inverse( bcy), inverse(dir)), m_lefty);
        dg::blas2::transfer( dg::create::dx( g, bcx, dir), m_rightx);
        dg::blas2::transfer( dg::create::dy( g, bcy, dir), m_righty);
        dg::blas2::transfer( dg::create::jumpX( g, bcx),   m_jumpX);
        dg::blas2::transfer( dg::create::jumpY( g, bcy),   m_jumpY);
        if( g.nz() == 1)
        {
            dg::blas2::transfer( dg::create::dz( g, bcz, dg::centered), m_rightz);
            dg::blas2::transfer( dg::create::dz( g, inverse( bcz), inverse(dg::centered)), m_leftz);
            m_addJumpZ = false;
        }
        else
        {
            dg::blas2::transfer( dg::create::dz( g, bcz, dir), m_rightz);
            dg::blas2::transfer( dg::create::dz( g, inverse( bcz), inverse(dir)), m_leftz);
            dg::blas2::transfer( dg::create::jumpZ( g, bcz),   m_jumpZ);
            m_addJumpZ = true;
        }
 
        dg::assign( dg::create::volume(g),        m_weights);
        dg::assign( dg::evaluate( dg::one, g),    m_precond);
        m_temp = m_tempx = m_tempy = m_tempz = m_weights;
        m_chi=g.metric();
        m_sigma = m_vol = dg::tensor::volume(m_chi);
    }
    template<class ...Params>
    void construct( Params&& ...ps)
    {
        //construct and swap
        *this = Elliptic3d( std::forward<Params>( ps)...);
    }
 
    template<class ContainerType0>
    void set_chi( const ContainerType0& sigma)
    {
        dg::blas1::pointwiseDot( sigma, m_vol, m_sigma);
        //update preconditioner
        dg::blas1::pointwiseDivide( 1., sigma, m_precond);
        // sigma is possibly zero, which will invalidate the preconditioner
        // it is important to call this blas1 function because it can
        // overwrite NaN in m_precond in the next update
    }
 
    template<class ContainerType0>
    void set_chi( const SparseTensor<ContainerType0>& tau)
    {
        m_chi = SparseTensor<Container>(tau);
    }
 
    const Container& weights()const {
        return m_weights;
    }
    const Container& precond()const {
        return m_precond;
    }
    void set_jfactor( value_type new_jfactor) {m_jfactor = new_jfactor;}
    value_type get_jfactor() const {return m_jfactor;}
    void set_jump_weighting( bool jump_weighting) {m_chi_weight_jump = jump_weighting;}
    bool get_jump_weighting() const {return m_chi_weight_jump;}
 
    void set_compute_in_2d( bool compute_in_2d ) {
        m_multiplyZ = !compute_in_2d;
    }
 
    template<class ContainerType0, class ContainerType1>
    void symv( const ContainerType0& x, ContainerType1& y){
        symv( 1, x, 0, y);
    }
    template<class ContainerType0, class ContainerType1>
    void symv( value_type alpha, const ContainerType0& x, value_type beta, ContainerType1& y)
    {
        //compute gradient
        dg::blas2::gemv( m_rightx, x, m_tempx); //R_x*f
        dg::blas2::gemv( m_righty, x, m_tempy); //R_y*f
        if( m_multiplyZ )
        {
            dg::blas2::gemv( m_rightz, x, m_tempz); //R_z*f
 
            //multiply with tensor (note the alias)
            dg::tensor::multiply3d(m_sigma, m_chi, m_tempx, m_tempy, m_tempz, 0., m_tempx, m_tempy, m_tempz);
            //now take divergence
            dg::blas2::symv( -1., m_leftz, m_tempz, 0., m_temp);
            dg::blas2::symv( -1., m_lefty, m_tempy, 1., m_temp);
        }
        else
        {
            dg::tensor::multiply2d(m_sigma, m_chi, m_tempx, m_tempy, 0., m_tempx, m_tempy);
            dg::blas2::symv( -1.,m_lefty, m_tempy, 0., m_temp);
        }
        dg::blas2::symv( -1., m_leftx, m_tempx, 1., m_temp);
 
        //add jump terms
        if( 0 != m_jfactor )
        {
            if(m_chi_weight_jump)
            {
                dg::blas2::symv( m_jfactor, m_jumpX, x, 0., m_tempx);
                dg::blas2::symv( m_jfactor, m_jumpY, x, 0., m_tempy);
                if( m_addJumpZ)
                {
                    dg::blas2::symv( m_jfactor, m_jumpZ, x, 0., m_tempz);
                    dg::tensor::multiply3d(m_sigma, m_chi, m_tempx, m_tempy,
                            m_tempz, 0., m_tempx, m_tempy, m_tempz);
                }
                else
                    dg::tensor::multiply2d(m_sigma, m_chi, m_tempx, m_tempy,
                            0., m_tempx, m_tempy);
 
                dg::blas1::axpbypgz(1., m_tempx, 1., m_tempy, 1., m_temp);
                if( m_addJumpZ)
                    dg::blas1::axpby( 1., m_tempz, 1., m_temp);
            }
            else
            {
                dg::blas2::symv( m_jfactor, m_jumpX, x, 1., m_temp);
                dg::blas2::symv( m_jfactor, m_jumpY, x, 1., m_temp);
                if( m_addJumpZ)
                    dg::blas2::symv( m_jfactor, m_jumpZ, x, 1., m_temp);
            }
        }
        dg::blas1::pointwiseDivide( alpha, m_temp, m_vol, beta, y);
    }
 
    template<class ContainerType0, class ContainerType1>
    void variation(const ContainerType0& phi, ContainerType1& sigma){
        variation(1.,1., phi, 0., sigma);
    }
    template<class ContainerTypeL, class ContainerType0, class ContainerType1>
    void variation(const ContainerTypeL& lambda, const ContainerType0& phi, ContainerType1& sigma){
        variation(1.,lambda, phi, 0., sigma);
    }
    template<class ContainerTypeL, class ContainerType0, class ContainerType1>
    void variation(value_type alpha, const ContainerTypeL& lambda, const ContainerType0& phi, value_type beta, ContainerType1& sigma)
    {
        dg::blas2::gemv( m_rightx, phi, m_tempx); //R_x*f
        dg::blas2::gemv( m_righty, phi, m_tempy); //R_y*f
        if( m_multiplyZ)
            dg::blas2::gemv( m_rightz, phi, m_tempz); //R_y*f
        else
            dg::blas1::scal( m_tempz, 0.);
        dg::tensor::scalar_product3d(alpha, lambda,  m_tempx, m_tempy, m_tempz, m_chi, lambda, m_tempx, m_tempy, m_tempz, beta, sigma);
    }
 
    private:
    Matrix m_leftx, m_lefty, m_leftz, m_rightx, m_righty, m_rightz, m_jumpX, m_jumpY, m_jumpZ;
    Container m_weights, m_precond;
    Container m_tempx, m_tempy, m_tempz, m_temp;
    SparseTensor<Container> m_chi;
    Container m_sigma, m_vol;
    value_type m_jfactor;
    bool m_multiplyZ = true, m_addJumpZ = false;
    bool m_chi_weight_jump;
};
template< class G, class M, class V>
struct TensorTraits< Elliptic1d<G, M, V> >
{
    using value_type      = get_value_type<V>;
    using tensor_category = SelfMadeMatrixTag;
};
template< class G, class M, class V>
struct TensorTraits< Elliptic2d<G, M, V> >
{
    using value_type      = get_value_type<V>;
    using tensor_category = SelfMadeMatrixTag;
};
 
template< class G, class M, class V>
struct TensorTraits< Elliptic3d<G, M, V> >
{
    using value_type      = get_value_type<V>;
    using tensor_category = SelfMadeMatrixTag;
};
 
} //namespace dg