lanczos.cpp

Go to the documentation of this file.

/* =========================================================================
   Copyright (c) 2010-2016, Institute for Microelectronics,
                            Institute for Analysis and Scientific Computing,
                            TU Wien.
   Portions of this software are copyright by UChicago Argonne, LLC.

                            -----------------
                  ViennaCL - The Vienna Computing Library
                            -----------------

   Project Head:    Karl Rupp                   rupp@iue.tuwien.ac.at

   (A list of authors and contributors can be found in the PDF manual)

   License:         MIT (X11), see file LICENSE in the base directory
============================================================================= */

// include necessary system headers
#include <iostream>

//include basic scalar and vector types of ViennaCL
#include "viennacl/scalar.hpp"
#include "viennacl/vector.hpp"
#include "viennacl/matrix.hpp"
#include "viennacl/matrix_proxy.hpp"
#include "viennacl/compressed_matrix.hpp"

#include "viennacl/linalg/lanczos.hpp"
#include "viennacl/io/matrix_market.hpp"

// Some helper functions for this tutorial:
#include <iostream>
#include <string>
#include <iomanip>

int main()
{
  // If you GPU does not support double precision, use `float` instead of `double`:
  typedef double     ScalarType;

  std::vector< std::map<unsigned int, ScalarType> > host_A;
  if (!viennacl::io::read_matrix_market_file(host_A, "../examples/testdata/mat65k.mtx"))
  {
    std::cout << "Error reading Matrix file" << std::endl;
    return EXIT_FAILURE;
  }

  viennacl::compressed_matrix<ScalarType> A;
  viennacl::copy(host_A, A);

  viennacl::linalg::lanczos_tag ltag(0.75,    // Select a power of 0.75 as the tolerance for the machine precision.
                                     10,      // Compute (approximations to) the 10 largest eigenvalues
                                     viennacl::linalg::lanczos_tag::partial_reorthogonalization, // use partial reorthogonalization
                                     30);   // Maximum size of the Krylov space

  std::cout << "Running Lanczos algorithm (eigenvalues only). This might take a while..." << std::endl;
  std::vector<ScalarType> lanczos_eigenvalues = viennacl::linalg::eig(A, ltag);

  std::cout << "Running Lanczos algorithm (with eigenvectors). This might take a while..." << std::endl;
  viennacl::matrix<ScalarType> approx_eigenvectors_A(A.size1(), ltag.num_eigenvalues());
  lanczos_eigenvalues = viennacl::linalg::eig(A, approx_eigenvectors_A, ltag);

  for (std::size_t i = 0; i< lanczos_eigenvalues.size(); i++)
  {
    std::cout << "Approx. eigenvalue " << std::setprecision(7) << lanczos_eigenvalues[i];

    // test approximated eigenvector by computing A*v:
    viennacl::vector<ScalarType> approx_eigenvector = viennacl::column(approx_eigenvectors_A, static_cast<unsigned int>(i));
    viennacl::vector<ScalarType> Aq = viennacl::linalg::prod(A, approx_eigenvector);
    std::cout << " (" << viennacl::linalg::inner_prod(Aq, approx_eigenvector) << " for <Av,v> with approx. eigenvector v)" << std::endl;
  }

  return EXIT_SUCCESS;
}