void pretty_print_matrix(std::ostream &ostr, MatrixT const &mat)
{
IndexType rows, cols;
mat.get_shape(rows, cols);
typename MatrixT::ScalarType zero(mat.get_zero());
for (IndexType row = 0; row < rows; ++row)
{
ostr << ((row == 0) ? "[[" : " [");
if (cols > 0)
{
auto val = mat.get_value_at(row, 0);
if (val == zero)
ostr << " ";
else
ostr << val;
}
for (IndexType col = 1; col < cols; ++col)
{
auto val = mat.get_value_at(row, col);
if (val == zero)
ostr << ", ";
else
ostr << ", " << val;
}
ostr << ((row == rows - 1) ? "]]\n" : "]\n");
}
}
void
test_transpose_readonly(MatrixT view)
{
typedef typename MatrixT::value_type T;
// Check that view is initialized
check_matrix(view, 0);
length_type const size1 = view.size(1);
typename MatrixT::const_transpose_type trans = view.transpose();
test_assert(trans.size(0) == view.size(1));
test_assert(trans.size(1) == view.size(0));
for (index_type idx0=0; idx0<trans.size(0); ++idx0)
for (index_type idx1=0; idx1<trans.size(1); ++idx1)
{
T expected = T(idx1 * size1 + idx0 + 0);
test_assert(equal(trans.get(idx0, idx1), expected));
test_assert(equal(trans.get(idx0, idx1),
view. get(idx1, idx0)));
}
// Check that view is unchanged
check_matrix(view, 0);
}
void LDLDecomposition<T>::getPseudoInverse(MatrixT& Ainv) const
{
Ainv.resize(LDL.n,LDL.n);
VectorT temp(LDL.n,Zero),y,x;
for(int i=0;i<LDL.n;i++) {
temp(i)=One;
LBackSub(temp,y);
for(int j=0;j<y.n;j++) {
if(!FuzzyZero(LDL(j,j),zeroTolerance))
y(j) = y(j)/LDL(j,j);
else
y(j) = 0.0;
}
LTBackSub(y,x);
//fill in a column
for(int j=0;j<LDL.n;j++)
Ainv(j,i)=x(j);
temp(i)=Zero;
}
T tol = Ainv.maxAbsElement()*Epsilon;
for(int i=0;i<LDL.n;i++)
for(int j=0;j<i;j++) {
if(!FuzzyEquals(Ainv(i,j),Ainv(j,i),tol))
LOG4CXX_INFO(KrisLibrary::logger(),Ainv);
Assert(FuzzyEquals(Ainv(i,j),Ainv(j,i),tol));
Ainv(i,j)=Ainv(j,i) = 0.5*(Ainv(i,j)+Ainv(j,i));
}
}
void operator()(LinPdeSysT const & pde_system,
SegmentT const & segment,
StorageType & storage,
MatrixT & system_matrix,
VectorT & load_vector)
{
typedef viennamath::equation equ_type;
typedef viennamath::expr expr_type;
typedef typename expr_type::interface_type interface_type;
typedef typename expr_type::numeric_type numeric_type;
typedef typename viennagrid::result_of::cell_tag<SegmentT>::type CellTag;
std::size_t map_index = viennafvm::create_mapping(pde_system, segment, storage);
system_matrix.clear();
system_matrix.resize(map_index, map_index, false);
load_vector.clear();
load_vector.resize(map_index);
for (std::size_t pde_index = 0; pde_index < pde_system.size(); ++pde_index)
{
#ifdef VIENNAFVM_DEBUG
std::cout << std::endl;
std::cout << "//" << std::endl;
std::cout << "// Equation " << pde_index << std::endl;
std::cout << "//" << std::endl;
#endif
assemble(pde_system, pde_index,
segment, storage,
system_matrix, load_vector);
} // for pde_index
} // functor
bool QRDecomposition<T>::set(const MatrixT& A)
{
QR.copy(A);
tau.resize(Min(A.m,A.n));
for (int i=0;i<Min(A.m,A.n);i++) {
/* Compute the Householder transformation to reduce the j-th
column of the matrix to a multiple of the j-th unit vector */
VectorT c_full,c;
QR.getColRef(i,c_full);
c.setRef(c_full,i);
T tau_i = HouseholderTransform (c);
tau(i)=tau_i;
/* Apply the transformation to the remaining columns and
update the norms */
if (i+1 < A.n) {
MatrixT m;
m.setRef(QR,i,i+1);
HouseholderPreMultiply (tau_i, c, m);
}
}
return true;
}
inline DCMatrix inv(const MatrixT &mat, double regularizationCoeff = 0.0) {
BOOST_ASSERT(mat.size1() == mat.size2());
unsigned int n = mat.size1();
DCMatrix inv = mat; // copy data, as it will be modified below
if (regularizationCoeff != 0.0)
inv += regularizationCoeff * ublas::identity_matrix<double>(n);
std::vector<int> ipiv(n); // pivot vector, is first filled by trf, then used by tri to inverse matrix
lapack::getrf(inv,ipiv); // inv and ipiv will both be modified
lapack::getri(inv,ipiv); // afterwards, "inv" is the inverse
return inv;
}
void DiagonalMatrixTemplate<T>::postMultiplyInverse(const MatrixT& a,MatrixT& x) const
{
Assert(this->n == a.n);
x.resize(a.m,this->n);
MyT xrow,arow;
for(int i=0;i<a.m;i++) {
x.getRowRef(i,xrow);
a.getRowRef(i,arow);
xrow.componentDiv(arow,*this);
}
}
void DiagonalMatrixTemplate<T>::postMultiplyTranspose(const MatrixT& a,MatrixT& x) const
{
Assert(this->n == a.m);
x.resize(a.n,this->n);
MyT xrow,acol;
for(int i=0;i<a.n;i++) {
x.getRowRef(i,xrow);
a.getColRef(i,acol);
xrow.componentMul(acol,*this);
}
}
void DiagonalMatrixTemplate<T>::preMultiplyInverse(const MatrixT& a,MatrixT& x) const
{
Assert(this->n == a.m);
x.resize(this->n,a.n);
ItT v=this->begin();
MyT xrow,arow;
for(int i=0;i<this->n;i++,v++) {
x.getRowRef(i,xrow);
a.getRowRef(i,arow);
xrow.div(arow,*v);
}
}
void DiagonalMatrixTemplate<T>::preMultiplyTranspose(const MatrixT& a,MatrixT& x) const
{
Assert(this->n == a.n);
x.resize(this->n,a.m);
ItT v=this->begin();
MyT xrow,acol;
for(int i=0;i<this->n;i++,v++) {
x.getRowRef(i,xrow);
a.getColRef(i,acol);
xrow.mul(acol,*v);
}
}
void NRQRDecomposition<T>::getQ(MatrixT& Q) const
{
int n=c.n;
Q.resize(n,n);
VectorT Qi;
Q.set(0);
for(int i=0;i<n;i++) {
Q.getRowRef(i,Qi);
Qi(i)=1;
QBackSub(Qi,Qi);
}
}
void backwardSolve(MatrixT const & R, VectorT const & y, VectorT & x)
{
for (long i2 = static_cast<long>(R.size2())-1; i2 >= 0; i2--)
{
vcl_size_t i = static_cast<vcl_size_t>(i2);
x(i) = y(i);
for (vcl_size_t j = static_cast<vcl_size_t>(i)+1; j < R.size2(); ++j)
x(i) -= R(i,j)*x(j);
x(i) /= R(i,i);
}
}
void CholeskyDecomposition<T>::getInverse(MatrixT& Ainv) const
{
Ainv.resize(L.n,L.n);
VectorT temp(L.n,Zero),y,x;
for(int i=0;i<L.n;i++) {
Ainv.getColRef(i,x); //x &= col i of A
temp(i)=One;
LBackSub(temp,y);
LTBackSub(y,x);
temp(i)=Zero;
}
}
void
fill_matrix(MatrixT view, int offset=0)
{
typedef typename MatrixT::value_type T;
length_type const size1 = view.size(1);
// Initialize view
for (index_type idx0=0; idx0<view.size(0); ++idx0)
for (index_type idx1=0; idx1<view.size(1); ++idx1)
view.put(idx0, idx1, T(idx0 * size1 + idx1 + offset));
}
inline DCMatrix invSym(const MatrixT &mat, double regularizationCoeff = 0.0) {
BOOST_ASSERT(mat.size1() == mat.size2());
unsigned int n = mat.size1();
DCMatrix inv = mat; // copy data, as it will be modified below
if (regularizationCoeff != 0.0)
inv += regularizationCoeff * ublas::identity_matrix<double>(n);
std::vector<int> ipiv(n); // pivot vector, is first filled by trf, then used by tri to inverse matrix
// TODO (9): use "po..." (po=positive definite matrix) instead if "sy..." (symmetric indefinite matrix) to make it faster
lapack::sytrf('U',inv,ipiv); // inv and ipiv will both be modified
lapack::sytri('U',inv,ipiv); // afterwards, "inv" is the real inverse, but only the upper elements are valid!!!
ublas::symmetric_adaptor<DCMatrix, ublas::upper> iSym(inv);
return iSym; // copies upper matrix to lower
}
void convolute_3d_out_of_place(MatrixT& _image, MatrixT& _kernel) {
if (_image.size() != _kernel.size()) {
std::cerr << "received image and kernel of mismatching size!\n";
return;
}
unsigned M, N, K;
M = _image.shape()[0];
N = _image.shape()[1];
K = _image.shape()[2];
unsigned fft_size = M * N * (K / 2 + 1);
// setup fourier space arrays
fftwf_complex* image_fourier = static_cast<fftwf_complex*>(
fftwf_malloc(sizeof(fftwf_complex) * fft_size));
fftwf_complex* kernel_fourier = static_cast<fftwf_complex*>(
fftwf_malloc(sizeof(fftwf_complex) * fft_size));
float scale = 1.0 / (M * N * K);
// define+run forward plans
fftwf_plan image_fwd_plan = fftwf_plan_dft_r2c_3d(
M, N, K, _image.data(), image_fourier, FFTW_ESTIMATE);
fftwf_execute(image_fwd_plan);
fftwf_plan kernel_fwd_plan = fftwf_plan_dft_r2c_3d(
M, N, K, _kernel.data(), kernel_fourier, FFTW_ESTIMATE);
fftwf_execute(kernel_fwd_plan);
// multiply
for (unsigned index = 0; index < fft_size; ++index) {
float real = image_fourier[index][0] * kernel_fourier[index][0] -
image_fourier[index][1] * kernel_fourier[index][1];
float imag = image_fourier[index][0] * kernel_fourier[index][1] +
image_fourier[index][1] * kernel_fourier[index][0];
image_fourier[index][0] = real;
image_fourier[index][1] = imag;
}
fftwf_destroy_plan(kernel_fwd_plan);
fftwf_destroy_plan(image_fwd_plan);
fftwf_plan image_rev_plan = fftwf_plan_dft_c2r_3d(
M, N, K, image_fourier, _image.data(), FFTW_ESTIMATE);
fftwf_execute(image_rev_plan);
for (unsigned index = 0; index < _image.num_elements(); ++index) {
_image.data()[index] *= scale;
}
fftwf_destroy_plan(image_rev_plan);
fftwf_free(image_fourier);
fftwf_free(kernel_fourier);
}
void MatrixT<ValueType>::transform_point(ValueType out[4], const MatrixT& m, const ValueType in[4])
{
out[0] = m.get(0, 0) * in[0] + m.get(0, 1) * in[1] + m.get(0, 2) * in[2] + m.get(0, 3) * in[3];
out[1] = m.get(1, 0) * in[0] + m.get(1, 1) * in[1] + m.get(1, 2) * in[2] + m.get(1, 3) * in[3];
out[2] = m.get(2, 0) * in[0] + m.get(2, 1) * in[1] + m.get(2, 2) * in[2] + m.get(2, 3) * in[3];
out[3] = m.get(3, 0) * in[0] + m.get(3, 1) * in[1] + m.get(3, 2) * in[2] + m.get(3, 3) * in[3];
}
void
check_matrix(MatrixT view, int offset=0)
{
typedef typename MatrixT::value_type T;
length_type const size1 = view.size(1);
for (index_type idx0=0; idx0<view.size(0); ++idx0)
for (index_type idx1=0; idx1<view.size(1); ++idx1)
{
test_assert(equal(view.get(idx0, idx1),
T(idx0 * size1 + idx1 + offset)));
}
}
/// <summary> Creates a shear matrix. </summary>
/// <param name="slope"> Strength of the shear. </param>
/// <param name="principalAxis"> Points are moved along this axis. </param>
/// <param name="modulatorAxis"> The displacement of points is proportional to this coordinate's value. </param>
/// <remarks> The formula for displacement along the pricipal axis is
/// <paramref name="slope"/>*pos[<paramref name="modulatorAxis"/>]. </remarks>
static MatrixT Shear(T slope, int principalAxis, int modulatorAxis) {
assert(principalAxis != modulatorAxis);
assert(principalAxis < Rows);
assert(modulatorAxis < Rows);
MatrixT ret;
ret.SetIdentity();
if (Order == eMatrixOrder::FOLLOW_VECTOR) {
ret(modulatorAxis, principalAxis) = slope;
}
else {
ret(principalAxis, modulatorAxis) = slope;
}
return ret;
}
void print_mismatching_items(MatrixT& _reference, MatrixT& _other) {
for (long x = 0; x < _reference.shape()[0]; ++x)
for (long y = 0; y < _reference.shape()[1]; ++y)
for (long z = 0; z < _reference.shape()[2]; ++z) {
float reference = _reference[x][y][z];
float to_compared = _other[x][y][z];
if (std::fabs(reference - to_compared) > (1e-3 * reference) &&
(std::fabs(reference) > 1e-4 || std::fabs(to_compared) > 1e-4)) {
std::cout << "[" << x << "][" << y << "][" << z
<< "] mismatch, ref: " << reference
<< " != to_compare: " << to_compared << "\n";
}
}
}
void CholeskyDecomposition<T>::backSub(const MatrixT& B, MatrixT& X) const
{
X.resize(B.m,B.n);
MatrixT temp(B.m,B.n);
if(!LBackSubstitute(L,B,temp)) FatalError("CholeskyDecomposition: LBackSubstitute failed!");
if(!LtBackSubstitute(L,temp,X)) FatalError("CholeskyDecomposition: LtBackSubstitute failed!");
}
ichol0_precond(MatrixT const & mat, ichol0_tag const & tag) : tag_(tag), LLT(mat.size1(), mat.size2(), viennacl::context(viennacl::MAIN_MEMORY))
{
//initialize preconditioner:
//std::cout << "Start CPU precond" << std::endl;
init(mat);
//std::cout << "End CPU precond" << std::endl;
}
void QRDecomposition<T>::getQ(MatrixT& Q) const
{
Assert(tau.n == Min(QR.m,QR.n));
Q.resize(QR.m,QR.m);
int i;
Q.setIdentity();
for (i=Min(QR.m,QR.n);i>0 && i--;) {
VectorT c,h;
QR.getColRef(i,c);
h.setRef(c,i);
MatrixT m;
m.setRef(Q,i,i);
HouseholderPreMultiply(tau(i),h,m);
}
}
ilut_precond(MatrixT const & mat, ilut_tag const & tag) : tag_(tag), LU_(mat.size1(), mat.size2())
{
//initialize preconditioner:
//std::cout << "Start CPU precond" << std::endl;
init(mat);
//std::cout << "End CPU precond" << std::endl;
}
void DiagonalMatrixTemplate<T>::copyDiagonal(const MatrixT& m)
{
if(!m.isSquare())
{
FatalError(MatrixError_ArgIncompatibleDimensions);
}
if(BaseT::n == 0)
{
resize(m.m);
}
else if(BaseT::n != m.m)
{
FatalError(MatrixError_DestIncompatibleDimensions);
}
m.getDiagCopy(0,*this);
}
void SparseMatrixTemplate_RM<T>::get(MatrixT& A) const
{
A.resize(m,n,Zero);
for(int i=0;i<m;i++) {
for(ConstRowIterator it=rows[i].begin();it!=rows[i].end();it++)
A(i,it->first) = it->second;
}
}
void row_index_of(MatrixT &mat)
{
graphblas::IndexType rows, cols;
mat.get_shape(rows, cols);
for (IndexType i = 0; i < rows; ++i)
{
for (IndexType j = 0; j < cols; ++j)
{
auto mat_ij = mat.get_value_at(i, j);
if (mat_ij != mat.get_zero())
{
mat.set_value_at(i, j, i);
}
}
}
}
typename viennacl::result_of::cpu_value_type<typename MatrixT::value_type>::type
eig(MatrixT const& A, power_iter_tag const & tag)
{
typedef typename viennacl::result_of::vector_for_matrix<MatrixT>::type VectorT;
VectorT eigenvec(A.size1());
return eig(A, tag, eigenvec);
}
void LDLDecomposition<T>::getA(MatrixT& A) const
{
MatrixT L,temp;
DiagonalMatrixTemplate<T> D;
getL(L);
getD(D);
D.postMultiply(L,temp);
A.mulTransposeB(temp,L);
}
void
banded_lu_decompose(MatrixT & _A)
{
size_t n = _A.rows();
for(size_t k = 0; k < n - 1; ++k) {
_A(k + 1, k) = _A(k + 1, k) / _A(k, k);
_A(k + 1, k + 1) -= _A(k + 1, k) * _A(k, k + 1);
}
}
