Пример #1
0
    /// \brief Compute QR factorization [Q,R] = qr(A,0).
    ///
    /// \param A [in/out] On input: the multivector to factor.
    ///   Overwritten with garbage on output.
    ///
    /// \param Q [out] On output: the (explicitly stored) Q factor in
    ///   the QR factorization of the (input) multivector A.
    ///
    /// \param R [out] On output: the R factor in the QR factorization
    ///   of the (input) multivector A.
    ///
    /// \param forceNonnegativeDiagonal [in] If true, then (if
    ///   necessary) do extra work (modifying both the Q and R
    ///   factors) in order to force the R factor to have a
    ///   nonnegative diagonal.
    ///
    /// \warning Currently, this method only works if A and Q have the
    ///   same communicator and row distribution ("Map," in Petra
    ///   terms) as those of the multivector given to this adapter
    ///   instance's constructor.  Otherwise, the result of this
    ///   method is undefined.
    void
    factorExplicit (MV& A,
                    MV& Q,
                    dense_matrix_type& R,
                    const bool forceNonnegativeDiagonal=false)
    {
      TEUCHOS_TEST_FOR_EXCEPTION
        (! A.isConstantStride (), std::invalid_argument, "TsqrAdaptor::"
         "factorExplicit: Input MultiVector A must have constant stride.");
      TEUCHOS_TEST_FOR_EXCEPTION
        (! Q.isConstantStride (), std::invalid_argument, "TsqrAdaptor::"
         "factorExplicit: Input MultiVector Q must have constant stride.");
      prepareTsqr (Q); // Finish initializing TSQR.

      // FIXME (mfh 16 Jan 2016) Currently, TSQR is a host-only
      // implementation.
      A.template sync<Kokkos::HostSpace> ();
      A.template modify<Kokkos::HostSpace> ();
      Q.template sync<Kokkos::HostSpace> ();
      Q.template modify<Kokkos::HostSpace> ();
      auto A_view = A.template getLocalView<Kokkos::HostSpace> ();
      auto Q_view = Q.template getLocalView<Kokkos::HostSpace> ();
      scalar_type* const A_ptr =
        reinterpret_cast<scalar_type*> (A_view.ptr_on_device ());
      scalar_type* const Q_ptr =
        reinterpret_cast<scalar_type*> (Q_view.ptr_on_device ());
      const bool contiguousCacheBlocks = false;
      tsqr_->factorExplicitRaw (A_view.dimension_0 (),
                                A_view.dimension_1 (),
                                A_ptr, A.getStride (),
                                Q_ptr, Q.getStride (),
                                R.values (), R.stride (),
                                contiguousCacheBlocks,
                                forceNonnegativeDiagonal);
    }
Пример #2
0
    /// \brief Rank-revealing decomposition
    ///
    /// Using the R factor and explicit Q factor from
    /// factorExplicit(), compute the singular value decomposition
    /// (SVD) of R: \f$R = U \Sigma V^*\f$.  If R is full rank (with
    /// respect to the given relative tolerance \c tol), do not modify
    /// Q or R.  Otherwise, compute \f$Q := Q \cdot U\f$ and \f$R :=
    /// \Sigma V^*\f$ in place.  If R was modified, then it may not
    /// necessarily be upper triangular on output.
    ///
    /// \param Q [in/out] On input: explicit Q factor computed by
    ///   factorExplicit().  (Must be an orthogonal resp. unitary
    ///   matrix.)  On output: If R is of full numerical rank with
    ///   respect to the tolerance tol, Q is unmodified.  Otherwise, Q
    ///   is updated so that the first \c rank columns of Q are a
    ///   basis for the column space of A (the original matrix whose
    ///   QR factorization was computed by factorExplicit()).  The
    ///   remaining columns of Q are a basis for the null space of A.
    ///
    /// \param R [in/out] On input: N by N upper triangular matrix
    ///   with leading dimension LDR >= N.  On output: if input is
    ///   full rank, R is unchanged on output.  Otherwise, if \f$R = U
    ///   \Sigma V^*\f$ is the SVD of R, on output R is overwritten
    ///   with \f$\Sigma \cdot V^*\f$.  This is also an N by N matrix,
    ///   but it may not necessarily be upper triangular.
    ///
    /// \param tol [in] Relative tolerance for computing the numerical
    ///   rank of the matrix R.
    ///
    /// \return Rank \f$r\f$ of R: \f$ 0 \leq r \leq N\f$.
    int
    revealRank (MV& Q,
                dense_matrix_type& R,
                const magnitude_type& tol)
    {
      TEUCHOS_TEST_FOR_EXCEPTION
        (! Q.isConstantStride (), std::invalid_argument, "TsqrAdaptor::"
         "revealRank: Input MultiVector Q must have constant stride.");
      prepareTsqr (Q); // Finish initializing TSQR.
      // FIXME (mfh 18 Oct 2010) Check Teuchos::Comm<int> object in Q
      // to make sure it is the same communicator as the one we are
      // using in our dist_tsqr_type implementation.

      Q.template sync<Kokkos::HostSpace> ();
      Q.template modify<Kokkos::HostSpace> ();
      auto Q_view = Q.template getLocalView<Kokkos::HostSpace> ();
      scalar_type* const Q_ptr =
        reinterpret_cast<scalar_type*> (Q_view.ptr_on_device ());
      const bool contiguousCacheBlocks = false;
      return tsqr_->revealRankRaw (Q_view.dimension_0 (),
                                   Q_view.dimension_1 (),
                                   Q_ptr, Q.getStride (),
                                   R.values (), R.stride (),
                                   tol, contiguousCacheBlocks);
    }