vector_type solve(const matrix_type& A, const vector_type& y)
{
typedef typename matrix_type::size_type size_type;
typedef typename matrix_type::value_type value_type;
namespace ublas = boost::numeric::ublas;
matrix_type Q(A.size1(), A.size2()), R(A.size1(), A.size2());
qr (A, Q, R);
vector_type b = prod(trans(Q), y);
vector_type result;
if (R.size1() > R.size2())
{
size_type min = (R.size1() < R.size2() ? R.size1() : R.size2());
result = ublas::solve(subrange(R, 0, min, 0, min),
subrange(b, 0, min),
ublas::upper_tag());
}
else
{
result = ublas::solve(R, b, ublas::upper_tag());
}
return result;
}
/**
* 13.
* @brief Copy the contents of a matrix.
* @param[out] A A is overwritten with B.
* @param[in] B The matrix to be copied.
*/
static void copy (matrix_type &A, const matrix_type &B) {
/** Result matrices should always be of the right size */
A.Resize (B.Height(), B.Width());
/** This one is pretty simple, but the order is different */
elem::Copy (B, A);
}
/**
* 15.
* @brief Extract the diagonal elements of a matrix.
* @param[in] A The (square) matrix whose diagonal is to be extracted.
* @param[out] B A diagonal matrix containing entries from A.
*/
static void diag(const matrix_type& A,
matrix_type& B) {
/* create a zeros matrix of the right dimension and assign to B */
const int n = A.Width();
B = zeros(n, n);
/* Set the diagonals of B */
for (int i=0; i<n; ++i) B.Set(i, i, A.Get(i,i));
}
/**
* 6.
* @brief Compute the matrix transpose.
* @param[in] A The matrix to be transposed.
* @param[out] B B is overwritten with A^{T}
*/
static void transpose (const matrix_type& A,
matrix_type& B) {
/** Result matrices should always have sufficient space */
B.Resize(A.Width(), A.Height());
/** Compute transpose */
elem::Transpose (A, B);
}
/**
* 5.
* @brief Multiply one matrix with another.
* @param[in] A The first matrix
* @param[in] B The second matrix
* @param[out] C C is overwritten with (A*B)
*/
static void multiply (const matrix_type& A,
const matrix_type& B,
matrix_type& C) {
/** Result matrices should always have sufficient space */
C.Resize(A.Height(), B.Width());
/** We have to do a Gemm */
elem::Gemm (elem::NORMAL, elem::NORMAL, 1.0, A, B, 0.0, C);
}
KOKKOS_INLINE_FUNCTION
void operator()( size_type inode ) const
{
// Apply a dirichlet boundary condition to 'irow'
// to maintain the symmetry of the original
// global stiffness matrix, zero out the columns
// that correspond to boundary conditions, and
// adjust the load vector accordingly
const size_type iBeg = matrix.graph.row_map[inode];
const size_type iEnd = matrix.graph.row_map[inode+1];
const ScalarCoordType z = node_coords(inode,2);
const bool bc_lower = z <= bc_lower_z ;
const bool bc_upper = bc_upper_z <= z ;
if ( bc_lower || bc_upper ) {
const ScalarType bc_value = bc_lower ? bc_lower_value
: bc_upper_value ;
rhs(inode) = bc_value ; // set the rhs vector
// zero each value on the row, and leave a one
// on the diagonal
for( size_type i = iBeg ; i < iEnd ; i++) {
matrix.coefficients(i) =
(int) inode == matrix.graph.entries(i) ? 1 : 0 ;
}
}
else {
// Find any columns that are boundary conditions.
// Clear them and adjust the load vector
for( size_type i = iBeg ; i < iEnd ; i++ ) {
const size_type cnode = matrix.graph.entries(i) ;
const ScalarCoordType zc = node_coords(cnode,2);
const bool c_bc_lower = zc <= bc_lower_z ;
const bool c_bc_upper = bc_upper_z <= zc ;
if ( c_bc_lower || c_bc_upper ) {
const ScalarType c_bc_value = c_bc_lower ? bc_lower_value
: bc_upper_value ;
rhs( inode ) -= c_bc_value * matrix.coefficients(i);
matrix.coefficients(i) = 0 ;
}
}
}
}
/** @brief Constructs a tensor with a matrix
*
* \note Initially the tensor will be two-dimensional.
*
* @param v matrix to be copied.
*/
BOOST_UBLAS_INLINE
tensor (const matrix_type &v)
: tensor_expression_type<self_type>()
, extents_ ()
, strides_ ()
, data_ (v.data())
{
if(!data_.empty()){
extents_ = extents_type{v.size1(),v.size2()};
strides_ = strides_type(extents_);
}
}
// Vectorizes (stacks columns of) an input matrix.
state_type Vec(const matrix_type &A)
{
state_type vectorized(A.size1()*A.size2());
for (int n = 0; n < A.size1(); n++) {
for (int m = 0; m < A.size2(); m++) {
vectorized(n*A.size1()+m) = A(m,n);
}
}
return vectorized;
}
/**
* 7.
* @brief Compute element-wise negation of a matrix.
* @param[in] A The matrix to be negated.
* @param[out] B B is overwritten with -1.0*A
*/
static void negation (const matrix_type& A,
matrix_type& B) {
/** Result matrices should always have sufficient space */
B.Resize(A.Height(), A.Width());
/** Copy over the matrix */
elem::Copy (A, B);
/** Multiply by -1.0 */
elem::Scal(-1.0, B);
}
static void calc(matrix_type& a, vector_type& v, const char& jobz) {
namespace impl = boost::numeric::bindings::lapack::detail;
//impl::syev(jobz,a,v);
char uplo = 'L';
integer_t n = a.size2();
integer_t lda = a.size1();
integer_t lwork =-1;
integer_t info = 0;
value_type dlwork;
impl::syev(jobz, uplo, n, mtraits::data(a), lda, vtraits::data(v), &dlwork, lwork, info);
}
/**
* Apply columnwise the sketching transform that is described by the
* the transform with output sketch_of_A.
*/
void apply (const matrix_type& A,
output_matrix_type& sketch_of_A,
columnwise_tag dimension) const {
const value_type *a = A.LockedBuffer();
El::Int lda = A.LDim();
value_type *sa = sketch_of_A.Buffer();
El::Int ldsa = sketch_of_A.LDim();
for (El::Int j = 0; j < A.Width(); j++)
for (El::Int i = 0; i < data_type::_S; i++)
sa[j * ldsa + i] = a[j * lda + data_type::_samples[i]];
}
/**
* Apply rowwise the sketching transform that is described by the
* the transform with output sketch_of_A.
*/
void apply (const matrix_type& A,
output_matrix_type& sketch_of_A,
rowwise_tag dimension) const {
const value_type *a = A.LockedBuffer();
El::Int lda = A.LDim();
value_type *sa = sketch_of_A.Buffer();
El::Int ldsa = sketch_of_A.LDim();
for (El::Int j = 0; j < data_type::_S; j++)
for (El::Int i = 0; i < A.Height(); i++)
sa[j * ldsa + i] = a[data_type::_samples[j] * lda + i];
}
/**
* 4.
* @brief Subtract one matrix from another.
* @param[in] A The first matrix
* @param[in] B The second matrix
* @param[out] C C is overwritten with (A-B)
*/
static void minus (const matrix_type& A,
const matrix_type& B,
matrix_type& C) {
/** Result matrices should always have sufficient space */
C.Resize(A.Height(), A.Width());
/** first copy the matrix over */
elem::Copy (A, C);
/** now, subtract the other matrix in */
elem::Axpy (-1.0, B, C);
}
bool is_positive_definite( const matrix<T,N,A>& m )
{
typedef matrix<T,N,A> matrix_type;
typedef typename matrix_type::range_type range_type;
if ( m.row() != m.col() ) return false;
for ( std::size_t i = 1; i != m.row(); ++i )
{
const matrix_type a{ m, range_type{0,i}, range_type{0,i} };
if ( a.det() <= T(0) ) return false;
}
return true;
}
// --------------------------------------------------------------------
static matrix_type _compute_rooted_fa(const matrix_type & fa)
{
const size_t K = fa.get_height();
const size_t J = fa.get_width();
assert(K > 1);
matrix_type rooted_fa (K - 1, J);
for (size_t k = 0; k + 1 < K; k++)
for (size_t j = 0; j < J; j++)
rooted_fa(k, j) = fa(k + 1, j) - fa(0, j);
return rooted_fa;
}
/**
* Apply the sketching transform on A and write to sketch_of_A.
* Implementation for columnwise.
*/
void apply_impl_vdist(const matrix_type& A,
output_matrix_type& sketch_of_A,
skylark::sketch::columnwise_tag tag) const {
// Just a local operation on the Matrix
_local.apply(A.LockedMatrix(), sketch_of_A.Matrix(), tag);
}
// Tensor product. Probably a better way of doing this.
matrix_type Kronecker(const matrix_type &A, const matrix_type &B)
{
matrix_type C(A.size1()*B.size1(), A.size2()*B.size2());
for (int i=0; i < A.size1(); i++) {
for (int j=0; j < A.size2(); j++) {
for (int k=0; k < B.size1(); k++) {
for (int l=0; l < B.size2(); l++) {
C(i*B.size1()+k, j*B.size2()+l) = A(i,j)*B(k,l);
}
}
}
}
return C;
}
static scalar_type r(const matrix_type& m)
{
BOOST_STATIC_ASSERT(Row >= 0);
BOOST_STATIC_ASSERT(Row < rows);
BOOST_STATIC_ASSERT(Col >= 0);
BOOST_STATIC_ASSERT(Col < cols);
return m.at(Col, Row);
}
static scalar_type ir(int row, int col, const matrix_type& m)
{
BOOST_ASSERT(row >= 0);
BOOST_ASSERT(row < rows);
BOOST_ASSERT(col >= 0);
BOOST_ASSERT(col < cols);
return m.at(col, row);
}
void apply (const matrix_type& A,
output_matrix_type& sketch_of_A,
Dimension dimension) const {
// TODO do we allow different communicators and different roots?
// If on root: Just a local operation on the Matrix
if (skylark::utility::get_communicator(A).rank() == 0)
_local.apply(A.LockedMatrix(), sketch_of_A.Matrix(), dimension);
}
void fill_from_file(matrix_type& tens, const std::string& str)
{
std::ifstream ifs(str);
for(int j = 0 ; j < tens.extent<1>(); ++j)
for(int i = 0 ; i < tens.extent<0>(); ++i)
{
ifs >> tens.at(i,j);
}
}
// ----------------------------------------------------------------
explicit q_data(const matrix_type & d)
: d_ij ()
, d_ik ()
, d_jk ()
, i ()
, j ()
{
static const auto k_0_5 = value_type(0.5);
const auto n = d.get_height();
const auto k_n_2 = value_type(n - 2);
//
// Cache the row sums; column sums would work equally well
// because the matrix is symmetric.
//
std::vector<value_type> sigma;
for (size_t c = 0; c < n; c++)
sigma.push_back(d.get_row_sum(c));
//
// Compute the values of the Q matrix.
//
matrix_type q (n, n);
for (size_t r = 0; r < n; r++)
for (size_t c = 0; c < r; c++)
q(r, c) = k_n_2 * d(r, c) - sigma[r] - sigma[c];
//
// Find the cell with the minimum value.
//
i = 1, j = 0;
for (size_t r = 2; r < n; r++)
for (size_t c = 0; c < r; c++)
if (q(r, c) < q(i, j))
i = r, j = c;
//
// Compute distances between the new nodes.
//
d_ij = d(i, j);
d_ik = k_0_5 * (d_ij + ((sigma[i] - sigma[j]) / k_n_2));
d_jk = d_ij - d_ik;
}
int operator()()
{
if ( make_preprocessing() ) { assert( !"Failed to make preprocessing." ); return 1; }
if( make_initialization_preprocessing() ) { assert( !"Failed to make initialization preprocessing." ); return 1; }
unknown_parameters = make_unknown_parameters(); if( !unknown_parameters ) { assert( !"Failed to make unknown parameters." ); return 1; }
fitting_array = make_fitting_array_initial_guess(); if ( fitting_array.size() != unknown_parameters ) { assert( !"Failed to make initial guess." ); return 1; }
max_iteration = make_max_iteration(); if( !max_iteration ) { assert( !"Failed to make max iteration." ); return 1; }
eps = make_eps();if( eps < value_type{0} ) { assert( !"Failed to make eps." ); return 1; }
if ( make_initialization_postprocessing() ) { assert( !"Failed to make initialization postprocessing." ); return 1; }
merit_function = make_merit_function(); if( !merit_function ) { assert( !"Failed to make merit function." ); return 1; }
jacobian_matrix_function.resize( 1, unknown_parameters );
for ( size_type index = 0; index != unknown_parameters; ++index )
{
jacobian_matrix_function[0][index] = make_jacobian_matrix_function( index );
if ( !jacobian_matrix_function[0][index] ) { assert( "Failed to make jacobian matrix function." ); return 1; }
}
hessian_matrix_function.resize( unknown_parameters, unknown_parameters );
for ( size_type index = 0; index != unknown_parameters; ++index )
for ( size_type jndex = 0; jndex <= index; ++jndex )
{
hessian_matrix_function[index][jndex] = make_hessian_matrix_function( index, jndex );
if ( !hessian_matrix_function[index][jndex] ) { assert( "Failed to make jacobian matrix function." ); return 1; }
if ( index != jndex )
{
hessian_matrix_function[jndex][index] = hessian_matrix_function[index][jndex];
}
}
if ( make_iteration_preprocessing() ) { assert( !"Failed to make iteration preprocessing." ); return 1; }
for ( size_type step_index = 0; step_index != max_iteration; ++step_index )
{
if ( make_every_iteration_preprocessing() ) { assert( !"Failed to make every iteration preprocessing." ); return 1; }
if ( make_iteration() ) { assert( !"Failed to make iteration" ); return 1; }
if ( ! make_fitting_flag() ) { /*eps reached*/break;}
if ( make_every_iteration_postprocessing() ) { assert( !"Failed to make every iteration postprocessing." ); return 1; }
}
if ( make_iteration_postprocessing() ) { assert( !"Failed to make iteration postprocessing." ); return 1; }
if ( make_postprocessing() ) { assert( !"Failed to make postprocessing. " ); return 1; }
return 0;
}
/**
* Parametrized constructor.
* \param aBelief The belief-state of the gaussian distribution.
*/
gaussian_pdf(const BeliefState& aBelief) : mean_state(aBelief.get_mean_state()), E_inv(aBelief.get_covariance().get_inverse_matrix()), factor(-1) {
factor = determinant_Cholesky(E_inv);
if(fabs(factor) < std::numeric_limits< scalar_type >::epsilon()) {
factor = scalar_type(-1);
} else {
factor = scalar_type(1) / factor;
for(size_type i = 0; i < E_inv.get_row_count(); ++i)
factor *= scalar_type(6.28318530718);
};
};
box<dimension>::box(matrix_type const& edges)
{
if (edges.size1() != dimension || edges.size2() != dimension) {
throw std::invalid_argument("edge vectors have invalid dimensionality");
}
edges_ = zero_matrix_type(dimension, dimension);
for (unsigned int i = 0; i < dimension; ++i) {
edges_(i, i) = edges(i, i);
}
if (norm_inf(edges_ - edges) != 0) {
throw std::invalid_argument("non-cuboid geomtries are not implemented");
}
for (unsigned int i = 0; i < dimension; ++i) {
length_[i] = edges_(i, i);
}
length_half_ = 0.5 * length_;
LOG("edge lengths of simulation domain: " << length_);
}
constexpr auto operator*(matrix_type<S1> m1, matrix_type<S2> m2) {
auto storage = fmap(
[=](auto row) {
return zip_with(
scalar_prod,
repeat_n(m2.ncolumns(), row),
columns(m2));
}
, rows(m1));
return matrix_type<decltype(storage)>{storage};
}
void qr(const matrix_type& A, matrix_type& Q, matrix_type& R)
{
using namespace boost::numeric::ublas;
typedef typename matrix_type::size_type size_type;
typedef typename matrix_type::value_type value_type;
// TODO resize Q and R to match the needed size.
int m=A.size1();
int n=A.size2();
identity_matrix<value_type> ident(m);
if (Q.size1() != ident.size1() || Q.size2() != ident.size2())
Q = matrix_type(m, m);
Q.assign(ident);
R.clear();
R = A;
for (size_type k=0; k< R.size1() && k<R.size2(); k++)
{
slice s1(k, 1, m - k);
slice s2(k, 0, m - k);
unit_vector<value_type> e1(m - k, 0);
// x = A(k:m, k);
matrix_vector_slice<matrix_type> x(R, s1, s2);
matrix_type F(x.size(), x.size());
Reflector(x, F);
matrix_type temp = subrange(R, k, m, k, n);
//F = prod(F, temp);
subrange(R, k, m, k, n) = prod(F, temp);
// <<---------------------------------------------->>
// forming Q
identity_matrix<value_type> iqk(A.size1());
matrix_type Q_k(iqk);
subrange(Q_k, Q_k.size1() - F.size1(), Q_k.size1(),
Q_k.size2() - F.size2(), Q_k.size2()) = F;
Q = prod(Q, Q_k);
}
}
/**
* 8.
* @brief Compute the inverse of a matrix.
* @param[in] A The (non-singular and square) matrix to be inverted.
* @param[out] B B is overwritten with A's inverse.
*/
static void inv(const matrix_type& A,
matrix_type& B) {
/** Assume that the matrix has an inverse */
const int n = A.Height();
/** First, copy over an identity as the right hand side */
B = eye(n);
/** Solve the linear system using LU */
elem::lu::SolveAfter(elem::NORMAL, A, B);
}
// --------------------------------------------------------------------
static matrix_type _create_b_vec(
const matrix_type & current_values,
const size_t row_padding)
{
assert(current_values.is_vector());
const auto K = current_values.get_length();
matrix_type b_vec (K + K + row_padding, 1);
for (size_t k = 0; k < K; k++)
{
b_vec[k] = current_values[k];
b_vec[k + K] = value_type(1) - current_values[k];
}
for (size_t k = K + K; k < b_vec.get_height(); k++)
b_vec[k] = value_type(0);
return b_vec;
}
value_type iterate( matrix_type const& initial_matrix, matrix_type& result_matrix )
{
triple_homotopy_fitting<value_type> thf{ug_size};
size_type const tilt_number = diag_matrix.row();
matrix_type intensity{ intensity_matrix.col(), 1 };
for ( size_type index = 0; index != tilt_number; ++index )
{
std::copy( intensity_matrix.row_begin(index), intensity_matrix.row_end(index), intensity.col_begin(0) );
//TODO -- optimizaton here
thf.register_entry( ar,
//C1 approximation
alpha(progress_ratio), make_coefficient_matrix( thickness, diag_matrix.row_begin(index), diag_matrix.row_end(index), column_index ),
//C/2 * C/2 approximation
beta(progress_ratio), make_coefficient_matrix( thickness/2.0, diag_matrix.row_begin(index), diag_matrix.row_end(index) ), expm( make_structure_matrix(ar, initial_matrix, diag_matrix.row_begin(index), diag_matrix.row_end(index) ), thickness/2.0, column_index ),
//standard expm
gamma(progress_ratio), make_scattering_matrix( ar, initial_matrix, diag_matrix.row_begin(index), diag_matrix.row_end(index), thickness, column_index ),
intensity, column_index );
}
result_matrix.resize( ug_size, 1 );
value_type const residual = thf.output( result_matrix.begin() );
/*
std::cout << "\n current residual is " << residual << "\n";
std::cout << "\n current ug is \n" << result_matrix.transpose() << "\n";
*/
return residual;
}