diff options
Diffstat (limited to 'third_party/eigen3/Eigen/src/SVD')
| -rw-r--r-- | third_party/eigen3/Eigen/src/SVD/JacobiSVD.h | 960 | ||||
| -rw-r--r-- | third_party/eigen3/Eigen/src/SVD/JacobiSVD_MKL.h | 92 | ||||
| -rw-r--r-- | third_party/eigen3/Eigen/src/SVD/UpperBidiagonalization.h | 396 |
3 files changed, 1448 insertions, 0 deletions
diff --git a/third_party/eigen3/Eigen/src/SVD/JacobiSVD.h b/third_party/eigen3/Eigen/src/SVD/JacobiSVD.h new file mode 100644 index 0000000000..d17d3a667d --- /dev/null +++ b/third_party/eigen3/Eigen/src/SVD/JacobiSVD.h @@ -0,0 +1,960 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2009-2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_JACOBISVD_H +#define EIGEN_JACOBISVD_H + +namespace Eigen { + +namespace internal { +// forward declaration (needed by ICC) +// the empty body is required by MSVC +template<typename MatrixType, int QRPreconditioner, + bool IsComplex = NumTraits<typename MatrixType::Scalar>::IsComplex> +struct svd_precondition_2x2_block_to_be_real {}; + +/*** QR preconditioners (R-SVD) + *** + *** Their role is to reduce the problem of computing the SVD to the case of a square matrix. + *** This approach, known as R-SVD, is an optimization for rectangular-enough matrices, and is a requirement for + *** JacobiSVD which by itself is only able to work on square matrices. + ***/ + +enum { PreconditionIfMoreColsThanRows, PreconditionIfMoreRowsThanCols }; + +template<typename MatrixType, int QRPreconditioner, int Case> +struct qr_preconditioner_should_do_anything +{ + enum { a = MatrixType::RowsAtCompileTime != Dynamic && + MatrixType::ColsAtCompileTime != Dynamic && + MatrixType::ColsAtCompileTime <= MatrixType::RowsAtCompileTime, + b = MatrixType::RowsAtCompileTime != Dynamic && + MatrixType::ColsAtCompileTime != Dynamic && + MatrixType::RowsAtCompileTime <= MatrixType::ColsAtCompileTime, + ret = !( (QRPreconditioner == NoQRPreconditioner) || + (Case == PreconditionIfMoreColsThanRows && bool(a)) || + (Case == PreconditionIfMoreRowsThanCols && bool(b)) ) + }; +}; + +template<typename MatrixType, int QRPreconditioner, int Case, + bool DoAnything = qr_preconditioner_should_do_anything<MatrixType, QRPreconditioner, Case>::ret +> struct qr_preconditioner_impl {}; + +template<typename MatrixType, int QRPreconditioner, int Case> +class qr_preconditioner_impl<MatrixType, QRPreconditioner, Case, false> +{ +public: + typedef typename MatrixType::Index Index; + void allocate(const JacobiSVD<MatrixType, QRPreconditioner>&) {} + bool run(JacobiSVD<MatrixType, QRPreconditioner>&, const MatrixType&) + { + return false; + } +}; + +/*** preconditioner using FullPivHouseholderQR ***/ + +template<typename MatrixType> +class qr_preconditioner_impl<MatrixType, FullPivHouseholderQRPreconditioner, PreconditionIfMoreRowsThanCols, true> +{ +public: + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + enum + { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime + }; + typedef Matrix<Scalar, 1, RowsAtCompileTime, RowMajor, 1, MaxRowsAtCompileTime> WorkspaceType; + + void allocate(const JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd) + { + if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols()) + { + m_qr.~QRType(); + ::new (&m_qr) QRType(svd.rows(), svd.cols()); + } + if (svd.m_computeFullU) m_workspace.resize(svd.rows()); + } + + bool run(JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix) + { + if(matrix.rows() > matrix.cols()) + { + m_qr.compute(matrix); + svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView<Upper>(); + if(svd.m_computeFullU) m_qr.matrixQ().evalTo(svd.m_matrixU, m_workspace); + if(svd.computeV()) svd.m_matrixV = m_qr.colsPermutation(); + return true; + } + return false; + } +private: + typedef FullPivHouseholderQR<MatrixType> QRType; + QRType m_qr; + WorkspaceType m_workspace; +}; + +template<typename MatrixType> +class qr_preconditioner_impl<MatrixType, FullPivHouseholderQRPreconditioner, PreconditionIfMoreColsThanRows, true> +{ +public: + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + enum + { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + ColsAtCompileTime = MatrixType::ColsAtCompileTime, + MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, + Options = MatrixType::Options + }; + typedef Matrix<Scalar, ColsAtCompileTime, RowsAtCompileTime, Options, MaxColsAtCompileTime, MaxRowsAtCompileTime> + TransposeTypeWithSameStorageOrder; + + void allocate(const JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd) + { + if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols()) + { + m_qr.~QRType(); + ::new (&m_qr) QRType(svd.cols(), svd.rows()); + } + m_adjoint.resize(svd.cols(), svd.rows()); + if (svd.m_computeFullV) m_workspace.resize(svd.cols()); + } + + bool run(JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix) + { + if(matrix.cols() > matrix.rows()) + { + m_adjoint = matrix.adjoint(); + m_qr.compute(m_adjoint); + svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView<Upper>().adjoint(); + if(svd.m_computeFullV) m_qr.matrixQ().evalTo(svd.m_matrixV, m_workspace); + if(svd.computeU()) svd.m_matrixU = m_qr.colsPermutation(); + return true; + } + else return false; + } +private: + typedef FullPivHouseholderQR<TransposeTypeWithSameStorageOrder> QRType; + QRType m_qr; + TransposeTypeWithSameStorageOrder m_adjoint; + typename internal::plain_row_type<MatrixType>::type m_workspace; +}; + +/*** preconditioner using ColPivHouseholderQR ***/ + +template<typename MatrixType> +class qr_preconditioner_impl<MatrixType, ColPivHouseholderQRPreconditioner, PreconditionIfMoreRowsThanCols, true> +{ +public: + typedef typename MatrixType::Index Index; + + void allocate(const JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd) + { + if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols()) + { + m_qr.~QRType(); + ::new (&m_qr) QRType(svd.rows(), svd.cols()); + } + if (svd.m_computeFullU) m_workspace.resize(svd.rows()); + else if (svd.m_computeThinU) m_workspace.resize(svd.cols()); + } + + bool run(JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix) + { + if(matrix.rows() > matrix.cols()) + { + m_qr.compute(matrix); + svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView<Upper>(); + if(svd.m_computeFullU) m_qr.householderQ().evalTo(svd.m_matrixU, m_workspace); + else if(svd.m_computeThinU) + { + svd.m_matrixU.setIdentity(matrix.rows(), matrix.cols()); + m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixU, m_workspace); + } + if(svd.computeV()) svd.m_matrixV = m_qr.colsPermutation(); + return true; + } + return false; + } + +private: + typedef ColPivHouseholderQR<MatrixType> QRType; + QRType m_qr; + typename internal::plain_col_type<MatrixType>::type m_workspace; +}; + +template<typename MatrixType> +class qr_preconditioner_impl<MatrixType, ColPivHouseholderQRPreconditioner, PreconditionIfMoreColsThanRows, true> +{ +public: + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + enum + { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + ColsAtCompileTime = MatrixType::ColsAtCompileTime, + MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, + Options = MatrixType::Options + }; + + typedef Matrix<Scalar, ColsAtCompileTime, RowsAtCompileTime, Options, MaxColsAtCompileTime, MaxRowsAtCompileTime> + TransposeTypeWithSameStorageOrder; + + void allocate(const JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd) + { + if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols()) + { + m_qr.~QRType(); + ::new (&m_qr) QRType(svd.cols(), svd.rows()); + } + if (svd.m_computeFullV) m_workspace.resize(svd.cols()); + else if (svd.m_computeThinV) m_workspace.resize(svd.rows()); + m_adjoint.resize(svd.cols(), svd.rows()); + } + + bool run(JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix) + { + if(matrix.cols() > matrix.rows()) + { + m_adjoint = matrix.adjoint(); + m_qr.compute(m_adjoint); + + svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView<Upper>().adjoint(); + if(svd.m_computeFullV) m_qr.householderQ().evalTo(svd.m_matrixV, m_workspace); + else if(svd.m_computeThinV) + { + svd.m_matrixV.setIdentity(matrix.cols(), matrix.rows()); + m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixV, m_workspace); + } + if(svd.computeU()) svd.m_matrixU = m_qr.colsPermutation(); + return true; + } + else return false; + } + +private: + typedef ColPivHouseholderQR<TransposeTypeWithSameStorageOrder> QRType; + QRType m_qr; + TransposeTypeWithSameStorageOrder m_adjoint; + typename internal::plain_row_type<MatrixType>::type m_workspace; +}; + +/*** preconditioner using HouseholderQR ***/ + +template<typename MatrixType> +class qr_preconditioner_impl<MatrixType, HouseholderQRPreconditioner, PreconditionIfMoreRowsThanCols, true> +{ +public: + typedef typename MatrixType::Index Index; + + void allocate(const JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd) + { + if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols()) + { + m_qr.~QRType(); + ::new (&m_qr) QRType(svd.rows(), svd.cols()); + } + if (svd.m_computeFullU) m_workspace.resize(svd.rows()); + else if (svd.m_computeThinU) m_workspace.resize(svd.cols()); + } + + bool run(JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd, const MatrixType& matrix) + { + if(matrix.rows() > matrix.cols()) + { + m_qr.compute(matrix); + svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView<Upper>(); + if(svd.m_computeFullU) m_qr.householderQ().evalTo(svd.m_matrixU, m_workspace); + else if(svd.m_computeThinU) + { + svd.m_matrixU.setIdentity(matrix.rows(), matrix.cols()); + m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixU, m_workspace); + } + if(svd.computeV()) svd.m_matrixV.setIdentity(matrix.cols(), matrix.cols()); + return true; + } + return false; + } +private: + typedef HouseholderQR<MatrixType> QRType; + QRType m_qr; + typename internal::plain_col_type<MatrixType>::type m_workspace; +}; + +template<typename MatrixType> +class qr_preconditioner_impl<MatrixType, HouseholderQRPreconditioner, PreconditionIfMoreColsThanRows, true> +{ +public: + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + enum + { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + ColsAtCompileTime = MatrixType::ColsAtCompileTime, + MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, + Options = MatrixType::Options + }; + + typedef Matrix<Scalar, ColsAtCompileTime, RowsAtCompileTime, Options, MaxColsAtCompileTime, MaxRowsAtCompileTime> + TransposeTypeWithSameStorageOrder; + + void allocate(const JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd) + { + if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols()) + { + m_qr.~QRType(); + ::new (&m_qr) QRType(svd.cols(), svd.rows()); + } + if (svd.m_computeFullV) m_workspace.resize(svd.cols()); + else if (svd.m_computeThinV) m_workspace.resize(svd.rows()); + m_adjoint.resize(svd.cols(), svd.rows()); + } + + bool run(JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd, const MatrixType& matrix) + { + if(matrix.cols() > matrix.rows()) + { + m_adjoint = matrix.adjoint(); + m_qr.compute(m_adjoint); + + svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView<Upper>().adjoint(); + if(svd.m_computeFullV) m_qr.householderQ().evalTo(svd.m_matrixV, m_workspace); + else if(svd.m_computeThinV) + { + svd.m_matrixV.setIdentity(matrix.cols(), matrix.rows()); + m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixV, m_workspace); + } + if(svd.computeU()) svd.m_matrixU.setIdentity(matrix.rows(), matrix.rows()); + return true; + } + else return false; + } + +private: + typedef HouseholderQR<TransposeTypeWithSameStorageOrder> QRType; + QRType m_qr; + TransposeTypeWithSameStorageOrder m_adjoint; + typename internal::plain_row_type<MatrixType>::type m_workspace; +}; + +/*** 2x2 SVD implementation + *** + *** JacobiSVD consists in performing a series of 2x2 SVD subproblems + ***/ + +template<typename MatrixType, int QRPreconditioner> +struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, false> +{ + typedef JacobiSVD<MatrixType, QRPreconditioner> SVD; + typedef typename SVD::Index Index; + static void run(typename SVD::WorkMatrixType&, SVD&, Index, Index) {} +}; + +template<typename MatrixType, int QRPreconditioner> +struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, true> +{ + typedef JacobiSVD<MatrixType, QRPreconditioner> SVD; + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::RealScalar RealScalar; + typedef typename SVD::Index Index; + static void run(typename SVD::WorkMatrixType& work_matrix, SVD& svd, Index p, Index q) + { + using std::sqrt; + Scalar z; + JacobiRotation<Scalar> rot; + RealScalar n = sqrt(numext::abs2(work_matrix.coeff(p,p)) + numext::abs2(work_matrix.coeff(q,p))); + if(n==0) + { + z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q); + work_matrix.row(p) *= z; + if(svd.computeU()) svd.m_matrixU.col(p) *= conj(z); + if(work_matrix.coeff(q,q)!=Scalar(0)) + z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q); + else + z = Scalar(0); + work_matrix.row(q) *= z; + if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z); + } + else + { + rot.c() = conj(work_matrix.coeff(p,p)) / n; + rot.s() = work_matrix.coeff(q,p) / n; + work_matrix.applyOnTheLeft(p,q,rot); + if(svd.computeU()) svd.m_matrixU.applyOnTheRight(p,q,rot.adjoint()); + if(work_matrix.coeff(p,q) != Scalar(0)) + { + Scalar z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q); + work_matrix.col(q) *= z; + if(svd.computeV()) svd.m_matrixV.col(q) *= z; + } + if(work_matrix.coeff(q,q) != Scalar(0)) + { + z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q); + work_matrix.row(q) *= z; + if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z); + } + } + } +}; + +template<typename MatrixType, typename RealScalar, typename Index> +void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q, + JacobiRotation<RealScalar> *j_left, + JacobiRotation<RealScalar> *j_right) +{ + using std::sqrt; + using std::abs; + Matrix<RealScalar,2,2> m; + m << numext::real(matrix.coeff(p,p)), numext::real(matrix.coeff(p,q)), + numext::real(matrix.coeff(q,p)), numext::real(matrix.coeff(q,q)); + JacobiRotation<RealScalar> rot1; + RealScalar t = m.coeff(0,0) + m.coeff(1,1); + RealScalar d = m.coeff(1,0) - m.coeff(0,1); + if(t == RealScalar(0)) + { + rot1.c() = RealScalar(0); + rot1.s() = d > RealScalar(0) ? RealScalar(1) : RealScalar(-1); + } + else + { + RealScalar t2d2 = numext::hypot(t,d); + rot1.c() = abs(t)/t2d2; + rot1.s() = d/t2d2; + if(t<RealScalar(0)) + rot1.s() = -rot1.s(); + } + m.applyOnTheLeft(0,1,rot1); + j_right->makeJacobi(m,0,1); + *j_left = rot1 * j_right->transpose(); +} + +} // end namespace internal + +/** \ingroup SVD_Module + * + * + * \class JacobiSVD + * + * \brief Two-sided Jacobi SVD decomposition of a rectangular matrix + * + * \param MatrixType the type of the matrix of which we are computing the SVD decomposition + * \param QRPreconditioner this optional parameter allows to specify the type of QR decomposition that will be used internally + * for the R-SVD step for non-square matrices. See discussion of possible values below. + * + * SVD decomposition consists in decomposing any n-by-p matrix \a A as a product + * \f[ A = U S V^* \f] + * where \a U is a n-by-n unitary, \a V is a p-by-p unitary, and \a S is a n-by-p real positive matrix which is zero outside of its main diagonal; + * the diagonal entries of S are known as the \em singular \em values of \a A and the columns of \a U and \a V are known as the left + * and right \em singular \em vectors of \a A respectively. + * + * Singular values are always sorted in decreasing order. + * + * This JacobiSVD decomposition computes only the singular values by default. If you want \a U or \a V, you need to ask for them explicitly. + * + * You can ask for only \em thin \a U or \a V to be computed, meaning the following. In case of a rectangular n-by-p matrix, letting \a m be the + * smaller value among \a n and \a p, there are only \a m singular vectors; the remaining columns of \a U and \a V do not correspond to actual + * singular vectors. Asking for \em thin \a U or \a V means asking for only their \a m first columns to be formed. So \a U is then a n-by-m matrix, + * and \a V is then a p-by-m matrix. Notice that thin \a U and \a V are all you need for (least squares) solving. + * + * Here's an example demonstrating basic usage: + * \include JacobiSVD_basic.cpp + * Output: \verbinclude JacobiSVD_basic.out + * + * This JacobiSVD class is a two-sided Jacobi R-SVD decomposition, ensuring optimal reliability and accuracy. The downside is that it's slower than + * bidiagonalizing SVD algorithms for large square matrices; however its complexity is still \f$ O(n^2p) \f$ where \a n is the smaller dimension and + * \a p is the greater dimension, meaning that it is still of the same order of complexity as the faster bidiagonalizing R-SVD algorithms. + * In particular, like any R-SVD, it takes advantage of non-squareness in that its complexity is only linear in the greater dimension. + * + * If the input matrix has inf or nan coefficients, the result of the computation is undefined, but the computation is guaranteed to + * terminate in finite (and reasonable) time. + * + * The possible values for QRPreconditioner are: + * \li ColPivHouseholderQRPreconditioner is the default. In practice it's very safe. It uses column-pivoting QR. + * \li FullPivHouseholderQRPreconditioner, is the safest and slowest. It uses full-pivoting QR. + * Contrary to other QRs, it doesn't allow computing thin unitaries. + * \li HouseholderQRPreconditioner is the fastest, and less safe and accurate than the pivoting variants. It uses non-pivoting QR. + * This is very similar in safety and accuracy to the bidiagonalization process used by bidiagonalizing SVD algorithms (since bidiagonalization + * is inherently non-pivoting). However the resulting SVD is still more reliable than bidiagonalizing SVDs because the Jacobi-based iterarive + * process is more reliable than the optimized bidiagonal SVD iterations. + * \li NoQRPreconditioner allows not to use a QR preconditioner at all. This is useful if you know that you will only be computing + * JacobiSVD decompositions of square matrices. Non-square matrices require a QR preconditioner. Using this option will result in + * faster compilation and smaller executable code. It won't significantly speed up computation, since JacobiSVD is always checking + * if QR preconditioning is needed before applying it anyway. + * + * \sa MatrixBase::jacobiSvd() + */ +template<typename _MatrixType, int QRPreconditioner> class JacobiSVD +{ + public: + + typedef _MatrixType MatrixType; + typedef typename MatrixType::Scalar Scalar; + typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar; + typedef typename MatrixType::Index Index; + enum { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + ColsAtCompileTime = MatrixType::ColsAtCompileTime, + DiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime,ColsAtCompileTime), + MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, + MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, + MaxDiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(MaxRowsAtCompileTime,MaxColsAtCompileTime), + MatrixOptions = MatrixType::Options + }; + + typedef Matrix<Scalar, RowsAtCompileTime, RowsAtCompileTime, + MatrixOptions, MaxRowsAtCompileTime, MaxRowsAtCompileTime> + MatrixUType; + typedef Matrix<Scalar, ColsAtCompileTime, ColsAtCompileTime, + MatrixOptions, MaxColsAtCompileTime, MaxColsAtCompileTime> + MatrixVType; + typedef typename internal::plain_diag_type<MatrixType, RealScalar>::type SingularValuesType; + typedef typename internal::plain_row_type<MatrixType>::type RowType; + typedef typename internal::plain_col_type<MatrixType>::type ColType; + typedef Matrix<Scalar, DiagSizeAtCompileTime, DiagSizeAtCompileTime, + MatrixOptions, MaxDiagSizeAtCompileTime, MaxDiagSizeAtCompileTime> + WorkMatrixType; + + /** \brief Default Constructor. + * + * The default constructor is useful in cases in which the user intends to + * perform decompositions via JacobiSVD::compute(const MatrixType&). + */ + JacobiSVD() + : m_isInitialized(false), + m_isAllocated(false), + m_usePrescribedThreshold(false), + m_computationOptions(0), + m_rows(-1), m_cols(-1), m_diagSize(0) + {} + + + /** \brief Default Constructor with memory preallocation + * + * Like the default constructor but with preallocation of the internal data + * according to the specified problem size. + * \sa JacobiSVD() + */ + JacobiSVD(Index rows, Index cols, unsigned int computationOptions = 0) + : m_isInitialized(false), + m_isAllocated(false), + m_usePrescribedThreshold(false), + m_computationOptions(0), + m_rows(-1), m_cols(-1) + { + allocate(rows, cols, computationOptions); + } + + /** \brief Constructor performing the decomposition of given matrix. + * + * \param matrix the matrix to decompose + * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed. + * By default, none is computed. This is a bit-field, the possible bits are #ComputeFullU, #ComputeThinU, + * #ComputeFullV, #ComputeThinV. + * + * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not + * available with the (non-default) FullPivHouseholderQR preconditioner. + */ + JacobiSVD(const MatrixType& matrix, unsigned int computationOptions = 0) + : m_isInitialized(false), + m_isAllocated(false), + m_usePrescribedThreshold(false), + m_computationOptions(0), + m_rows(-1), m_cols(-1) + { + compute(matrix, computationOptions); + } + + /** \brief Method performing the decomposition of given matrix using custom options. + * + * \param matrix the matrix to decompose + * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed. + * By default, none is computed. This is a bit-field, the possible bits are #ComputeFullU, #ComputeThinU, + * #ComputeFullV, #ComputeThinV. + * + * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not + * available with the (non-default) FullPivHouseholderQR preconditioner. + */ + JacobiSVD& compute(const MatrixType& matrix, unsigned int computationOptions); + + /** \brief Method performing the decomposition of given matrix using current options. + * + * \param matrix the matrix to decompose + * + * This method uses the current \a computationOptions, as already passed to the constructor or to compute(const MatrixType&, unsigned int). + */ + JacobiSVD& compute(const MatrixType& matrix) + { + return compute(matrix, m_computationOptions); + } + + /** \returns the \a U matrix. + * + * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p, + * the U matrix is n-by-n if you asked for #ComputeFullU, and is n-by-m if you asked for #ComputeThinU. + * + * The \a m first columns of \a U are the left singular vectors of the matrix being decomposed. + * + * This method asserts that you asked for \a U to be computed. + */ + const MatrixUType& matrixU() const + { + eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); + eigen_assert(computeU() && "This JacobiSVD decomposition didn't compute U. Did you ask for it?"); + return m_matrixU; + } + + /** \returns the \a V matrix. + * + * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p, + * the V matrix is p-by-p if you asked for #ComputeFullV, and is p-by-m if you asked for ComputeThinV. + * + * The \a m first columns of \a V are the right singular vectors of the matrix being decomposed. + * + * This method asserts that you asked for \a V to be computed. + */ + const MatrixVType& matrixV() const + { + eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); + eigen_assert(computeV() && "This JacobiSVD decomposition didn't compute V. Did you ask for it?"); + return m_matrixV; + } + + /** \returns the vector of singular values. + * + * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p, the + * returned vector has size \a m. Singular values are always sorted in decreasing order. + */ + const SingularValuesType& singularValues() const + { + eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); + return m_singularValues; + } + + /** \returns true if \a U (full or thin) is asked for in this SVD decomposition */ + inline bool computeU() const { return m_computeFullU || m_computeThinU; } + /** \returns true if \a V (full or thin) is asked for in this SVD decomposition */ + inline bool computeV() const { return m_computeFullV || m_computeThinV; } + + /** \returns a (least squares) solution of \f$ A x = b \f$ using the current SVD decomposition of A. + * + * \param b the right-hand-side of the equation to solve. + * + * \note Solving requires both U and V to be computed. Thin U and V are enough, there is no need for full U or V. + * + * \note SVD solving is implicitly least-squares. Thus, this method serves both purposes of exact solving and least-squares solving. + * In other words, the returned solution is guaranteed to minimize the Euclidean norm \f$ \Vert A x - b \Vert \f$. + */ + template<typename Rhs> + inline const internal::solve_retval<JacobiSVD, Rhs> + solve(const MatrixBase<Rhs>& b) const + { + eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); + eigen_assert(computeU() && computeV() && "JacobiSVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice)."); + return internal::solve_retval<JacobiSVD, Rhs>(*this, b.derived()); + } + + /** \returns the number of singular values that are not exactly 0 */ + Index nonzeroSingularValues() const + { + eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); + return m_nonzeroSingularValues; + } + + /** \returns the rank of the matrix of which \c *this is the SVD. + * + * \note This method has to determine which singular values should be considered nonzero. + * For that, it uses the threshold value that you can control by calling + * setThreshold(const RealScalar&). + */ + inline Index rank() const + { + using std::abs; + eigen_assert(m_isInitialized && "JacobiSVD is not initialized."); + if(m_singularValues.size()==0) return 0; + RealScalar premultiplied_threshold = m_singularValues.coeff(0) * threshold(); + Index i = m_nonzeroSingularValues-1; + while(i>=0 && m_singularValues.coeff(i) < premultiplied_threshold) --i; + return i+1; + } + + /** Allows to prescribe a threshold to be used by certain methods, such as rank() and solve(), + * which need to determine when singular values are to be considered nonzero. + * This is not used for the SVD decomposition itself. + * + * When it needs to get the threshold value, Eigen calls threshold(). + * The default is \c NumTraits<Scalar>::epsilon() + * + * \param threshold The new value to use as the threshold. + * + * A singular value will be considered nonzero if its value is strictly greater than + * \f$ \vert singular value \vert \leqslant threshold \times \vert max singular value \vert \f$. + * + * If you want to come back to the default behavior, call setThreshold(Default_t) + */ + JacobiSVD& setThreshold(const RealScalar& threshold) + { + m_usePrescribedThreshold = true; + m_prescribedThreshold = threshold; + return *this; + } + + /** Allows to come back to the default behavior, letting Eigen use its default formula for + * determining the threshold. + * + * You should pass the special object Eigen::Default as parameter here. + * \code svd.setThreshold(Eigen::Default); \endcode + * + * See the documentation of setThreshold(const RealScalar&). + */ + JacobiSVD& setThreshold(Default_t) + { + m_usePrescribedThreshold = false; + return *this; + } + + /** Returns the threshold that will be used by certain methods such as rank(). + * + * See the documentation of setThreshold(const RealScalar&). + */ + RealScalar threshold() const + { + eigen_assert(m_isInitialized || m_usePrescribedThreshold); + return m_usePrescribedThreshold ? m_prescribedThreshold + : (std::max<Index>)(1,m_diagSize)*NumTraits<Scalar>::epsilon(); + } + + inline Index rows() const { return m_rows; } + inline Index cols() const { return m_cols; } + + private: + void allocate(Index rows, Index cols, unsigned int computationOptions); + + protected: + MatrixUType m_matrixU; + MatrixVType m_matrixV; + SingularValuesType m_singularValues; + WorkMatrixType m_workMatrix; + bool m_isInitialized, m_isAllocated, m_usePrescribedThreshold; + bool m_computeFullU, m_computeThinU; + bool m_computeFullV, m_computeThinV; + unsigned int m_computationOptions; + Index m_nonzeroSingularValues, m_rows, m_cols, m_diagSize; + RealScalar m_prescribedThreshold; + + template<typename __MatrixType, int _QRPreconditioner, bool _IsComplex> + friend struct internal::svd_precondition_2x2_block_to_be_real; + template<typename __MatrixType, int _QRPreconditioner, int _Case, bool _DoAnything> + friend struct internal::qr_preconditioner_impl; + + internal::qr_preconditioner_impl<MatrixType, QRPreconditioner, internal::PreconditionIfMoreColsThanRows> m_qr_precond_morecols; + internal::qr_preconditioner_impl<MatrixType, QRPreconditioner, internal::PreconditionIfMoreRowsThanCols> m_qr_precond_morerows; +}; + +template<typename MatrixType, int QRPreconditioner> +void JacobiSVD<MatrixType, QRPreconditioner>::allocate(Index rows, Index cols, unsigned int computationOptions) +{ + eigen_assert(rows >= 0 && cols >= 0); + + if (m_isAllocated && + rows == m_rows && + cols == m_cols && + computationOptions == m_computationOptions) + { + return; + } + + m_rows = rows; + m_cols = cols; + m_isInitialized = false; + m_isAllocated = true; + m_computationOptions = computationOptions; + m_computeFullU = (computationOptions & ComputeFullU) != 0; + m_computeThinU = (computationOptions & ComputeThinU) != 0; + m_computeFullV = (computationOptions & ComputeFullV) != 0; + m_computeThinV = (computationOptions & ComputeThinV) != 0; + eigen_assert(!(m_computeFullU && m_computeThinU) && "JacobiSVD: you can't ask for both full and thin U"); + eigen_assert(!(m_computeFullV && m_computeThinV) && "JacobiSVD: you can't ask for both full and thin V"); + eigen_assert(EIGEN_IMPLIES(m_computeThinU || m_computeThinV, MatrixType::ColsAtCompileTime==Dynamic) && + "JacobiSVD: thin U and V are only available when your matrix has a dynamic number of columns."); + if (QRPreconditioner == FullPivHouseholderQRPreconditioner) + { + eigen_assert(!(m_computeThinU || m_computeThinV) && + "JacobiSVD: can't compute thin U or thin V with the FullPivHouseholderQR preconditioner. " + "Use the ColPivHouseholderQR preconditioner instead."); + } + m_diagSize = (std::min)(m_rows, m_cols); + m_singularValues.resize(m_diagSize); + if(RowsAtCompileTime==Dynamic) + m_matrixU.resize(m_rows, m_computeFullU ? m_rows + : m_computeThinU ? m_diagSize + : 0); + if(ColsAtCompileTime==Dynamic) + m_matrixV.resize(m_cols, m_computeFullV ? m_cols + : m_computeThinV ? m_diagSize + : 0); + m_workMatrix.resize(m_diagSize, m_diagSize); + + if(m_cols>m_rows) m_qr_precond_morecols.allocate(*this); + if(m_rows>m_cols) m_qr_precond_morerows.allocate(*this); +} + +template<typename MatrixType, int QRPreconditioner> +JacobiSVD<MatrixType, QRPreconditioner>& +JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsigned int computationOptions) +{ + using std::abs; + allocate(matrix.rows(), matrix.cols(), computationOptions); + + // currently we stop when we reach precision 2*epsilon as the last bit of precision can require an unreasonable number of iterations, + // only worsening the precision of U and V as we accumulate more rotations + const RealScalar precision = RealScalar(2) * NumTraits<Scalar>::epsilon(); + + // limit for very small denormal numbers to be considered zero in order to avoid infinite loops (see bug 286) + const RealScalar considerAsZero = RealScalar(2) * std::numeric_limits<RealScalar>::denorm_min(); + + /*** step 1. The R-SVD step: we use a QR decomposition to reduce to the case of a square matrix */ + + if(!m_qr_precond_morecols.run(*this, matrix) && !m_qr_precond_morerows.run(*this, matrix)) + { + m_workMatrix = matrix.block(0,0,m_diagSize,m_diagSize); + if(m_computeFullU) m_matrixU.setIdentity(m_rows,m_rows); + if(m_computeThinU) m_matrixU.setIdentity(m_rows,m_diagSize); + if(m_computeFullV) m_matrixV.setIdentity(m_cols,m_cols); + if(m_computeThinV) m_matrixV.setIdentity(m_cols, m_diagSize); + } + + // Scaling factor to reducover/under-flows + RealScalar scale = m_workMatrix.cwiseAbs().maxCoeff(); + if(scale==RealScalar(0)) scale = RealScalar(1); + m_workMatrix /= scale; + + /*** step 2. The main Jacobi SVD iteration. ***/ + + bool finished = false; + while(!finished) + { + finished = true; + + // do a sweep: for all index pairs (p,q), perform SVD of the corresponding 2x2 sub-matrix + + for(Index p = 1; p < m_diagSize; ++p) + { + for(Index q = 0; q < p; ++q) + { + // if this 2x2 sub-matrix is not diagonal already... + // notice that this comparison will evaluate to false if any NaN is involved, ensuring that NaN's don't + // keep us iterating forever. Similarly, small denormal numbers are considered zero. + RealScalar threshold = numext::maxi(considerAsZero, precision * numext::maxi(abs(m_workMatrix.coeff(p,p)), + abs(m_workMatrix.coeff(q,q)))); + if(numext::maxi(abs(m_workMatrix.coeff(p,q)),abs(m_workMatrix.coeff(q,p))) > threshold) + { + finished = false; + + // perform SVD decomposition of 2x2 sub-matrix corresponding to indices p,q to make it diagonal + internal::svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner>::run(m_workMatrix, *this, p, q); + JacobiRotation<RealScalar> j_left, j_right; + internal::real_2x2_jacobi_svd(m_workMatrix, p, q, &j_left, &j_right); + + // accumulate resulting Jacobi rotations + m_workMatrix.applyOnTheLeft(p,q,j_left); + if(computeU()) m_matrixU.applyOnTheRight(p,q,j_left.transpose()); + + m_workMatrix.applyOnTheRight(p,q,j_right); + if(computeV()) m_matrixV.applyOnTheRight(p,q,j_right); + } + } + } + } + + /*** step 3. The work matrix is now diagonal, so ensure it's positive so its diagonal entries are the singular values ***/ + + for(Index i = 0; i < m_diagSize; ++i) + { + RealScalar a = abs(m_workMatrix.coeff(i,i)); + m_singularValues.coeffRef(i) = a; + if(computeU() && (a!=RealScalar(0))) m_matrixU.col(i) *= m_workMatrix.coeff(i,i)/a; + } + + m_singularValues *= scale; + + /*** step 4. Sort singular values in descending order and compute the number of nonzero singular values ***/ + + m_nonzeroSingularValues = m_diagSize; + for(Index i = 0; i < m_diagSize; i++) + { + Index pos; + RealScalar maxRemainingSingularValue = m_singularValues.tail(m_diagSize-i).maxCoeff(&pos); + if(maxRemainingSingularValue == RealScalar(0)) + { + m_nonzeroSingularValues = i; + break; + } + if(pos) + { + pos += i; + std::swap(m_singularValues.coeffRef(i), m_singularValues.coeffRef(pos)); + if(computeU()) m_matrixU.col(pos).swap(m_matrixU.col(i)); + if(computeV()) m_matrixV.col(pos).swap(m_matrixV.col(i)); + } + } + + m_isInitialized = true; + return *this; +} + +namespace internal { +template<typename _MatrixType, int QRPreconditioner, typename Rhs> +struct solve_retval<JacobiSVD<_MatrixType, QRPreconditioner>, Rhs> + : solve_retval_base<JacobiSVD<_MatrixType, QRPreconditioner>, Rhs> +{ + typedef JacobiSVD<_MatrixType, QRPreconditioner> JacobiSVDType; + EIGEN_MAKE_SOLVE_HELPERS(JacobiSVDType,Rhs) + + template<typename Dest> void evalTo(Dest& dst) const + { + eigen_assert(rhs().rows() == dec().rows()); + + // A = U S V^* + // So A^{-1} = V S^{-1} U^* + + Matrix<Scalar, Dynamic, Rhs::ColsAtCompileTime, 0, _MatrixType::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime> tmp; + Index rank = dec().rank(); + + tmp.noalias() = dec().matrixU().leftCols(rank).adjoint() * rhs(); + tmp = dec().singularValues().head(rank).asDiagonal().inverse() * tmp; + dst = dec().matrixV().leftCols(rank) * tmp; + } +}; +} // end namespace internal + +#ifndef __CUDACC__ +/** \svd_module + * + * \return the singular value decomposition of \c *this computed by two-sided + * Jacobi transformations. + * + * \sa class JacobiSVD + */ +template<typename Derived> +JacobiSVD<typename MatrixBase<Derived>::PlainObject> +MatrixBase<Derived>::jacobiSvd(unsigned int computationOptions) const +{ + return JacobiSVD<PlainObject>(*this, computationOptions); +} +#endif // __CUDACC__ + +} // end namespace Eigen + +#endif // EIGEN_JACOBISVD_H diff --git a/third_party/eigen3/Eigen/src/SVD/JacobiSVD_MKL.h b/third_party/eigen3/Eigen/src/SVD/JacobiSVD_MKL.h new file mode 100644 index 0000000000..decda75405 --- /dev/null +++ b/third_party/eigen3/Eigen/src/SVD/JacobiSVD_MKL.h @@ -0,0 +1,92 @@ +/* + Copyright (c) 2011, Intel Corporation. All rights reserved. + + Redistribution and use in source and binary forms, with or without modification, + are permitted provided that the following conditions are met: + + * Redistributions of source code must retain the above copyright notice, this + list of conditions and the following disclaimer. + * Redistributions in binary form must reproduce the above copyright notice, + this list of conditions and the following disclaimer in the documentation + and/or other materials provided with the distribution. + * Neither the name of Intel Corporation nor the names of its contributors may + be used to endorse or promote products derived from this software without + specific prior written permission. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND + ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED + WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE + DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR + ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES + (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; + LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON + ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + + ******************************************************************************** + * Content : Eigen bindings to Intel(R) MKL + * Singular Value Decomposition - SVD. + ******************************************************************************** +*/ + +#ifndef EIGEN_JACOBISVD_MKL_H +#define EIGEN_JACOBISVD_MKL_H + +#include "Eigen/src/Core/util/MKL_support.h" + +namespace Eigen { + +/** \internal Specialization for the data types supported by MKL */ + +#define EIGEN_MKL_SVD(EIGTYPE, MKLTYPE, MKLRTYPE, MKLPREFIX, EIGCOLROW, MKLCOLROW) \ +template<> inline \ +JacobiSVD<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>, ColPivHouseholderQRPreconditioner>& \ +JacobiSVD<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>, ColPivHouseholderQRPreconditioner>::compute(const Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>& matrix, unsigned int computationOptions) \ +{ \ + typedef Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic> MatrixType; \ + typedef MatrixType::Scalar Scalar; \ + typedef MatrixType::RealScalar RealScalar; \ + allocate(matrix.rows(), matrix.cols(), computationOptions); \ +\ + /*const RealScalar precision = RealScalar(2) * NumTraits<Scalar>::epsilon();*/ \ + m_nonzeroSingularValues = m_diagSize; \ +\ + lapack_int lda = matrix.outerStride(), ldu, ldvt; \ + lapack_int matrix_order = MKLCOLROW; \ + char jobu, jobvt; \ + MKLTYPE *u, *vt, dummy; \ + jobu = (m_computeFullU) ? 'A' : (m_computeThinU) ? 'S' : 'N'; \ + jobvt = (m_computeFullV) ? 'A' : (m_computeThinV) ? 'S' : 'N'; \ + if (computeU()) { \ + ldu = m_matrixU.outerStride(); \ + u = (MKLTYPE*)m_matrixU.data(); \ + } else { ldu=1; u=&dummy; }\ + MatrixType localV; \ + ldvt = (m_computeFullV) ? m_cols : (m_computeThinV) ? m_diagSize : 1; \ + if (computeV()) { \ + localV.resize(ldvt, m_cols); \ + vt = (MKLTYPE*)localV.data(); \ + } else { ldvt=1; vt=&dummy; }\ + Matrix<MKLRTYPE, Dynamic, Dynamic> superb; superb.resize(m_diagSize, 1); \ + MatrixType m_temp; m_temp = matrix; \ + LAPACKE_##MKLPREFIX##gesvd( matrix_order, jobu, jobvt, m_rows, m_cols, (MKLTYPE*)m_temp.data(), lda, (MKLRTYPE*)m_singularValues.data(), u, ldu, vt, ldvt, superb.data()); \ + if (computeV()) m_matrixV = localV.adjoint(); \ + /* for(int i=0;i<m_diagSize;i++) if (m_singularValues.coeffRef(i) < precision) { m_nonzeroSingularValues--; m_singularValues.coeffRef(i)=RealScalar(0);}*/ \ + m_isInitialized = true; \ + return *this; \ +} + +EIGEN_MKL_SVD(double, double, double, d, ColMajor, LAPACK_COL_MAJOR) +EIGEN_MKL_SVD(float, float, float , s, ColMajor, LAPACK_COL_MAJOR) +EIGEN_MKL_SVD(dcomplex, MKL_Complex16, double, z, ColMajor, LAPACK_COL_MAJOR) +EIGEN_MKL_SVD(scomplex, MKL_Complex8, float , c, ColMajor, LAPACK_COL_MAJOR) + +EIGEN_MKL_SVD(double, double, double, d, RowMajor, LAPACK_ROW_MAJOR) +EIGEN_MKL_SVD(float, float, float , s, RowMajor, LAPACK_ROW_MAJOR) +EIGEN_MKL_SVD(dcomplex, MKL_Complex16, double, z, RowMajor, LAPACK_ROW_MAJOR) +EIGEN_MKL_SVD(scomplex, MKL_Complex8, float , c, RowMajor, LAPACK_ROW_MAJOR) + +} // end namespace Eigen + +#endif // EIGEN_JACOBISVD_MKL_H diff --git a/third_party/eigen3/Eigen/src/SVD/UpperBidiagonalization.h b/third_party/eigen3/Eigen/src/SVD/UpperBidiagonalization.h new file mode 100644 index 0000000000..40067682c9 --- /dev/null +++ b/third_party/eigen3/Eigen/src/SVD/UpperBidiagonalization.h @@ -0,0 +1,396 @@ +// This file is part of Eigen, a lightweight C++ template library +// for linear algebra. +// +// Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com> +// +// This Source Code Form is subject to the terms of the Mozilla +// Public License v. 2.0. If a copy of the MPL was not distributed +// with this file, You can obtain one at http://mozilla.org/MPL/2.0/. + +#ifndef EIGEN_BIDIAGONALIZATION_H +#define EIGEN_BIDIAGONALIZATION_H + +namespace Eigen { + +namespace internal { +// UpperBidiagonalization will probably be replaced by a Bidiagonalization class, don't want to make it stable API. +// At the same time, it's useful to keep for now as it's about the only thing that is testing the BandMatrix class. + +template<typename _MatrixType> class UpperBidiagonalization +{ + public: + + typedef _MatrixType MatrixType; + enum { + RowsAtCompileTime = MatrixType::RowsAtCompileTime, + ColsAtCompileTime = MatrixType::ColsAtCompileTime, + ColsAtCompileTimeMinusOne = internal::decrement_size<ColsAtCompileTime>::ret + }; + typedef typename MatrixType::Scalar Scalar; + typedef typename MatrixType::RealScalar RealScalar; + typedef typename MatrixType::Index Index; + typedef Matrix<Scalar, 1, ColsAtCompileTime> RowVectorType; + typedef Matrix<Scalar, RowsAtCompileTime, 1> ColVectorType; + typedef BandMatrix<RealScalar, ColsAtCompileTime, ColsAtCompileTime, 1, 0, RowMajor> BidiagonalType; + typedef Matrix<Scalar, ColsAtCompileTime, 1> DiagVectorType; + typedef Matrix<Scalar, ColsAtCompileTimeMinusOne, 1> SuperDiagVectorType; + typedef HouseholderSequence< + const MatrixType, + CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, const Diagonal<const MatrixType,0> > + > HouseholderUSequenceType; + typedef HouseholderSequence< + const typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type, + Diagonal<const MatrixType,1>, + OnTheRight + > HouseholderVSequenceType; + + /** + * \brief Default Constructor. + * + * The default constructor is useful in cases in which the user intends to + * perform decompositions via Bidiagonalization::compute(const MatrixType&). + */ + UpperBidiagonalization() : m_householder(), m_bidiagonal(), m_isInitialized(false) {} + + UpperBidiagonalization(const MatrixType& matrix) + : m_householder(matrix.rows(), matrix.cols()), + m_bidiagonal(matrix.cols(), matrix.cols()), + m_isInitialized(false) + { + compute(matrix); + } + + UpperBidiagonalization& compute(const MatrixType& matrix); + UpperBidiagonalization& computeUnblocked(const MatrixType& matrix); + + const MatrixType& householder() const { return m_householder; } + const BidiagonalType& bidiagonal() const { return m_bidiagonal; } + + const HouseholderUSequenceType householderU() const + { + eigen_assert(m_isInitialized && "UpperBidiagonalization is not initialized."); + return HouseholderUSequenceType(m_householder, m_householder.diagonal().conjugate()); + } + + const HouseholderVSequenceType householderV() // const here gives nasty errors and i'm lazy + { + eigen_assert(m_isInitialized && "UpperBidiagonalization is not initialized."); + return HouseholderVSequenceType(m_householder.conjugate(), m_householder.const_derived().template diagonal<1>()) + .setLength(m_householder.cols()-1) + .setShift(1); + } + + protected: + MatrixType m_householder; + BidiagonalType m_bidiagonal; + bool m_isInitialized; +}; + +// Standard upper bidiagonalization without fancy optimizations +// This version should be faster for small matrix size +template<typename MatrixType> +void upperbidiagonalization_inplace_unblocked(MatrixType& mat, + typename MatrixType::RealScalar *diagonal, + typename MatrixType::RealScalar *upper_diagonal, + typename MatrixType::Scalar* tempData = 0) +{ + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + + Index rows = mat.rows(); + Index cols = mat.cols(); + + typedef Matrix<Scalar,Dynamic,1,ColMajor,MatrixType::MaxRowsAtCompileTime,1> TempType; + TempType tempVector; + if(tempData==0) + { + tempVector.resize(rows); + tempData = tempVector.data(); + } + + for (Index k = 0; /* breaks at k==cols-1 below */ ; ++k) + { + Index remainingRows = rows - k; + Index remainingCols = cols - k - 1; + + // construct left householder transform in-place in A + mat.col(k).tail(remainingRows) + .makeHouseholderInPlace(mat.coeffRef(k,k), diagonal[k]); + // apply householder transform to remaining part of A on the left + mat.bottomRightCorner(remainingRows, remainingCols) + .applyHouseholderOnTheLeft(mat.col(k).tail(remainingRows-1), mat.coeff(k,k), tempData); + + if(k == cols-1) break; + + // construct right householder transform in-place in mat + mat.row(k).tail(remainingCols) + .makeHouseholderInPlace(mat.coeffRef(k,k+1), upper_diagonal[k]); + // apply householder transform to remaining part of mat on the left + mat.bottomRightCorner(remainingRows-1, remainingCols) + .applyHouseholderOnTheRight(mat.row(k).tail(remainingCols-1).transpose(), mat.coeff(k,k+1), tempData); + } +} + +/** \internal + * Helper routine for the block reduction to upper bidiagonal form. + * + * Let's partition the matrix A: + * + * | A00 A01 | + * A = | | + * | A10 A11 | + * + * This function reduces to bidiagonal form the left \c rows x \a blockSize vertical panel [A00/A10] + * and the \a blockSize x \c cols horizontal panel [A00 A01] of the matrix \a A. The bottom-right block A11 + * is updated using matrix-matrix products: + * A22 -= V * Y^T - X * U^T + * where V and U contains the left and right Householder vectors. U and V are stored in A10, and A01 + * respectively, and the update matrices X and Y are computed during the reduction. + * + */ +template<typename MatrixType> +void upperbidiagonalization_blocked_helper(MatrixType& A, + typename MatrixType::RealScalar *diagonal, + typename MatrixType::RealScalar *upper_diagonal, + typename MatrixType::Index bs, + Ref<Matrix<typename MatrixType::Scalar, Dynamic, Dynamic> > X, + Ref<Matrix<typename MatrixType::Scalar, Dynamic, Dynamic> > Y) +{ + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + typedef Ref<Matrix<Scalar, Dynamic, 1> > SubColumnType; + typedef Ref<Matrix<Scalar, 1, Dynamic>, 0, InnerStride<> > SubRowType; + typedef Ref<Matrix<Scalar, Dynamic, Dynamic> > SubMatType; + + Index brows = A.rows(); + Index bcols = A.cols(); + + Scalar tau_u, tau_u_prev(0), tau_v; + + for(Index k = 0; k < bs; ++k) + { + Index remainingRows = brows - k; + Index remainingCols = bcols - k - 1; + + SubMatType X_k1( X.block(k,0, remainingRows,k) ); + SubMatType V_k1( A.block(k,0, remainingRows,k) ); + + // 1 - update the k-th column of A + SubColumnType v_k = A.col(k).tail(remainingRows); + v_k -= V_k1 * Y.row(k).head(k).adjoint(); + if(k) v_k -= X_k1 * A.col(k).head(k); + + // 2 - construct left Householder transform in-place + v_k.makeHouseholderInPlace(tau_v, diagonal[k]); + + if(k+1<bcols) + { + SubMatType Y_k ( Y.block(k+1,0, remainingCols, k+1) ); + SubMatType U_k1 ( A.block(0,k+1, k,remainingCols) ); + + // this eases the application of Householder transforAions + // A(k,k) will store tau_v later + A(k,k) = Scalar(1); + + // 3 - Compute y_k^T = tau_v * ( A^T*v_k - Y_k-1*V_k-1^T*v_k - U_k-1*X_k-1^T*v_k ) + { + SubColumnType y_k( Y.col(k).tail(remainingCols) ); + + // let's use the begining of column k of Y as a temporary vector + SubColumnType tmp( Y.col(k).head(k) ); + y_k.noalias() = A.block(k,k+1, remainingRows,remainingCols).adjoint() * v_k; // bottleneck + tmp.noalias() = V_k1.adjoint() * v_k; + y_k.noalias() -= Y_k.leftCols(k) * tmp; + tmp.noalias() = X_k1.adjoint() * v_k; + y_k.noalias() -= U_k1.adjoint() * tmp; + y_k *= numext::conj(tau_v); + } + + // 4 - update k-th row of A (it will become u_k) + SubRowType u_k( A.row(k).tail(remainingCols) ); + u_k = u_k.conjugate(); + { + u_k -= Y_k * A.row(k).head(k+1).adjoint(); + if(k) u_k -= U_k1.adjoint() * X.row(k).head(k).adjoint(); + } + + // 5 - construct right Householder transform in-placecols + u_k.makeHouseholderInPlace(tau_u, upper_diagonal[k]); + + // this eases the application of Householder transforAions + // A(k,k+1) will store tau_u later + A(k,k+1) = Scalar(1); + + // 6 - Compute x_k = tau_u * ( A*u_k - X_k-1*U_k-1^T*u_k - V_k*Y_k^T*u_k ) + { + SubColumnType x_k ( X.col(k).tail(remainingRows-1) ); + + // let's use the begining of column k of X as a temporary vectors + // note that tmp0 and tmp1 overlaps + SubColumnType tmp0 ( X.col(k).head(k) ), + tmp1 ( X.col(k).head(k+1) ); + + x_k.noalias() = A.block(k+1,k+1, remainingRows-1,remainingCols) * u_k.transpose(); // bottleneck + tmp0.noalias() = U_k1 * u_k.transpose(); + x_k.noalias() -= X_k1.bottomRows(remainingRows-1) * tmp0; + tmp1.noalias() = Y_k.adjoint() * u_k.transpose(); + x_k.noalias() -= A.block(k+1,0, remainingRows-1,k+1) * tmp1; + x_k *= numext::conj(tau_u); + tau_u = numext::conj(tau_u); + u_k = u_k.conjugate(); + } + + if(k>0) A.coeffRef(k-1,k) = tau_u_prev; + tau_u_prev = tau_u; + } + else + A.coeffRef(k-1,k) = tau_u_prev; + + A.coeffRef(k,k) = tau_v; + } + + if(bs<bcols) + A.coeffRef(bs-1,bs) = tau_u_prev; + + // update A22 + if(bcols>bs && brows>bs) + { + SubMatType A11( A.bottomRightCorner(brows-bs,bcols-bs) ); + SubMatType A10( A.block(bs,0, brows-bs,bs) ); + SubMatType A01( A.block(0,bs, bs,bcols-bs) ); + Scalar tmp = A01(bs-1,0); + A01(bs-1,0) = 1; + A11.noalias() -= A10 * Y.topLeftCorner(bcols,bs).bottomRows(bcols-bs).adjoint(); + A11.noalias() -= X.topLeftCorner(brows,bs).bottomRows(brows-bs) * A01; + A01(bs-1,0) = tmp; + } +} + +/** \internal + * + * Implementation of a block-bidiagonal reduction. + * It is based on the following paper: + * The Design of a Parallel Dense Linear Algebra Software Library: Reduction to Hessenberg, Tridiagonal, and Bidiagonal Form. + * by Jaeyoung Choi, Jack J. Dongarra, David W. Walker. (1995) + * section 3.3 + */ +template<typename MatrixType, typename BidiagType> +void upperbidiagonalization_inplace_blocked(MatrixType& A, BidiagType& bidiagonal, + typename MatrixType::Index maxBlockSize=32, + typename MatrixType::Scalar* /*tempData*/ = 0) +{ + typedef typename MatrixType::Index Index; + typedef typename MatrixType::Scalar Scalar; + typedef Block<MatrixType,Dynamic,Dynamic> BlockType; + + Index rows = A.rows(); + Index cols = A.cols(); + Index size = (std::min)(rows, cols); + + Matrix<Scalar,MatrixType::RowsAtCompileTime,Dynamic,ColMajor,MatrixType::MaxRowsAtCompileTime> X(rows,maxBlockSize); + Matrix<Scalar,MatrixType::ColsAtCompileTime,Dynamic,ColMajor,MatrixType::MaxColsAtCompileTime> Y(cols,maxBlockSize); + Index blockSize = (std::min)(maxBlockSize,size); + + Index k = 0; + for(k = 0; k < size; k += blockSize) + { + Index bs = (std::min)(size-k,blockSize); // actual size of the block + Index brows = rows - k; // rows of the block + Index bcols = cols - k; // columns of the block + + // partition the matrix A: + // + // | A00 A01 A02 | + // | | + // A = | A10 A11 A12 | + // | | + // | A20 A21 A22 | + // + // where A11 is a bs x bs diagonal block, + // and let: + // | A11 A12 | + // B = | | + // | A21 A22 | + + BlockType B = A.block(k,k,brows,bcols); + + // This stage performs the bidiagonalization of A11, A21, A12, and updating of A22. + // Finally, the algorithm continue on the updated A22. + // + // However, if B is too small, or A22 empty, then let's use an unblocked strategy + if(k+bs==cols || bcols<48) // somewhat arbitrary threshold + { + upperbidiagonalization_inplace_unblocked(B, + &(bidiagonal.template diagonal<0>().coeffRef(k)), + &(bidiagonal.template diagonal<1>().coeffRef(k)), + X.data() + ); + break; // We're done + } + else + { + upperbidiagonalization_blocked_helper<BlockType>( B, + &(bidiagonal.template diagonal<0>().coeffRef(k)), + &(bidiagonal.template diagonal<1>().coeffRef(k)), + bs, + X.topLeftCorner(brows,bs), + Y.topLeftCorner(bcols,bs) + ); + } + } +} + +template<typename _MatrixType> +UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::computeUnblocked(const _MatrixType& matrix) +{ + Index rows = matrix.rows(); + Index cols = matrix.cols(); + + eigen_assert(rows >= cols && "UpperBidiagonalization is only for Arices satisfying rows>=cols."); + + m_householder = matrix; + + ColVectorType temp(rows); + + upperbidiagonalization_inplace_unblocked(m_householder, + &(m_bidiagonal.template diagonal<0>().coeffRef(0)), + &(m_bidiagonal.template diagonal<1>().coeffRef(0)), + temp.data()); + + m_isInitialized = true; + return *this; +} + +template<typename _MatrixType> +UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::compute(const _MatrixType& matrix) +{ + Index rows = matrix.rows(); + Index cols = matrix.cols(); + + eigen_assert(rows >= cols && "UpperBidiagonalization is only for Arices satisfying rows>=cols."); + + m_householder = matrix; + upperbidiagonalization_inplace_blocked(m_householder, m_bidiagonal); + + m_isInitialized = true; + return *this; +} + +#if 0 +/** \return the Householder QR decomposition of \c *this. + * + * \sa class Bidiagonalization + */ +template<typename Derived> +const UpperBidiagonalization<typename MatrixBase<Derived>::PlainObject> +MatrixBase<Derived>::bidiagonalization() const +{ + return UpperBidiagonalization<PlainObject>(eval()); +} +#endif + +} // end namespace internal + +} // end namespace Eigen + +#endif // EIGEN_BIDIAGONALIZATION_H |