cg.hpp

Go to the documentation of this file.

#ifndef VIENNACL_LINALG_CG_HPP_
#define VIENNACL_LINALG_CG_HPP_

/* =========================================================================
   Copyright (c) 2010-2015, 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 manual)

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

#include <vector>
#include <map>
#include <cmath>
#include <numeric>

#include "viennacl/forwards.h"
#include "viennacl/tools/tools.hpp"
#include "viennacl/linalg/ilu.hpp"
#include "viennacl/linalg/prod.hpp"
#include "viennacl/linalg/inner_prod.hpp"
#include "viennacl/linalg/norm_2.hpp"
#include "viennacl/traits/clear.hpp"
#include "viennacl/traits/size.hpp"
#include "viennacl/meta/result_of.hpp"
#include "viennacl/linalg/iterative_operations.hpp"

namespace viennacl
{
namespace linalg
{

class cg_tag
{
public:
  cg_tag(double tol = 1e-8, unsigned int max_iterations = 300) : tol_(tol), abs_tol_(0), iterations_(max_iterations) {}

  double tolerance() const { return tol_; }

  double abs_tolerance() const { return abs_tol_; }
  void abs_tolerance(double new_tol) { if (new_tol >= 0) abs_tol_ = new_tol; }

  unsigned int max_iterations() const { return iterations_; }

  unsigned int iters() const { return iters_taken_; }
  void iters(unsigned int i) const { iters_taken_ = i; }

  double error() const { return last_error_; }
  void error(double e) const { last_error_ = e; }


private:
  double tol_;
  double abs_tol_;
  unsigned int iterations_;

  //return values from solver
  mutable unsigned int iters_taken_;
  mutable double last_error_;
};

namespace detail
{

  template<typename VectorT, typename PreconditionerT>
  class z_handler{
  public:
    z_handler(VectorT & residual) : z_(residual){ }
    VectorT & get() { return z_; }
  private:
    VectorT z_;
  };

  template<typename VectorT>
  class z_handler<VectorT, viennacl::linalg::no_precond>{
  public:
    z_handler(VectorT & residual) : presidual_(&residual){ }
    VectorT & get() { return *presidual_; }
  private:
    VectorT * presidual_;
  };

}

namespace detail
{

  //template<typename MatrixType, typename ScalarType>
  template<typename MatrixT, typename NumericT>
  viennacl::vector<NumericT> pipelined_solve(MatrixT const & A, //MatrixType const & A,
                                             viennacl::vector<NumericT> const & rhs,
                                             cg_tag const & tag,
                                             viennacl::linalg::no_precond,
                                             bool (*monitor)(viennacl::vector<NumericT> const &, NumericT, void*) = NULL,
                                             void *monitor_data = NULL)
  {
    typedef typename viennacl::vector<NumericT>::difference_type   difference_type;

    viennacl::vector<NumericT> result(rhs);
    viennacl::traits::clear(result);

    viennacl::vector<NumericT> residual(rhs);
    viennacl::vector<NumericT> p(rhs);
    viennacl::vector<NumericT> Ap = viennacl::linalg::prod(A, p);
    viennacl::vector<NumericT> inner_prod_buffer = viennacl::zero_vector<NumericT>(3*256, viennacl::traits::context(rhs)); // temporary buffer
    std::vector<NumericT>      host_inner_prod_buffer(inner_prod_buffer.size());
    vcl_size_t                 buffer_size_per_vector = inner_prod_buffer.size() / 3;
    difference_type            buffer_offset_per_vector = static_cast<difference_type>(buffer_size_per_vector);

    NumericT norm_rhs_squared = viennacl::linalg::norm_2(residual); norm_rhs_squared *= norm_rhs_squared;

    if (norm_rhs_squared <= tag.abs_tolerance() * tag.abs_tolerance()) //check for early convergence of A*x = 0
      return result;

    NumericT inner_prod_rr = norm_rhs_squared;
    NumericT alpha = inner_prod_rr / viennacl::linalg::inner_prod(p, Ap);
    NumericT beta  = viennacl::linalg::norm_2(Ap); beta = (alpha * alpha * beta * beta - inner_prod_rr) / inner_prod_rr;
    NumericT inner_prod_ApAp = 0;
    NumericT inner_prod_pAp  = 0;

    for (unsigned int i = 0; i < tag.max_iterations(); ++i)
    {
      tag.iters(i+1);

      viennacl::linalg::pipelined_cg_vector_update(result, alpha, p, residual, Ap, beta, inner_prod_buffer);
      viennacl::linalg::pipelined_cg_prod(A, p, Ap, inner_prod_buffer);

      // bring back the partial results to the host:
      viennacl::fast_copy(inner_prod_buffer.begin(), inner_prod_buffer.end(), host_inner_prod_buffer.begin());

      inner_prod_rr   = std::accumulate(host_inner_prod_buffer.begin(),                                host_inner_prod_buffer.begin() +     buffer_offset_per_vector, NumericT(0));
      inner_prod_ApAp = std::accumulate(host_inner_prod_buffer.begin() +     buffer_offset_per_vector, host_inner_prod_buffer.begin() + 2 * buffer_offset_per_vector, NumericT(0));
      inner_prod_pAp  = std::accumulate(host_inner_prod_buffer.begin() + 2 * buffer_offset_per_vector, host_inner_prod_buffer.begin() + 3 * buffer_offset_per_vector, NumericT(0));

      if (monitor && monitor(result, std::sqrt(std::fabs(inner_prod_rr / norm_rhs_squared)), monitor_data))
        break;
      if (std::fabs(inner_prod_rr / norm_rhs_squared) < tag.tolerance() *  tag.tolerance() || std::fabs(inner_prod_rr) < tag.abs_tolerance() * tag.abs_tolerance())    //squared norms involved here
        break;

      alpha = inner_prod_rr / inner_prod_pAp;
      beta  = (alpha*alpha*inner_prod_ApAp - inner_prod_rr) / inner_prod_rr;
    }

    //store last error estimate:
    tag.error(std::sqrt(std::fabs(inner_prod_rr) / norm_rhs_squared));

    return result;
  }


  template<typename NumericT>
  viennacl::vector<NumericT> solve_impl(viennacl::compressed_matrix<NumericT> const & A,
                                        viennacl::vector<NumericT> const & rhs,
                                        cg_tag const & tag,
                                        viennacl::linalg::no_precond,
                                        bool (*monitor)(viennacl::vector<NumericT> const &, NumericT, void*) = NULL,
                                        void *monitor_data = NULL)
  {
    return pipelined_solve(A, rhs, tag, viennacl::linalg::no_precond(), monitor, monitor_data);
  }


  template<typename NumericT>
  viennacl::vector<NumericT> solve_impl(viennacl::coordinate_matrix<NumericT> const & A,
                                        viennacl::vector<NumericT> const & rhs,
                                        cg_tag const & tag,
                                        viennacl::linalg::no_precond,
                                        bool (*monitor)(viennacl::vector<NumericT> const &, NumericT, void*) = NULL,
                                        void *monitor_data = NULL)
  {
    return detail::pipelined_solve(A, rhs, tag, viennacl::linalg::no_precond(), monitor, monitor_data);
  }



  template<typename NumericT>
  viennacl::vector<NumericT> solve_impl(viennacl::ell_matrix<NumericT> const & A,
                                        viennacl::vector<NumericT> const & rhs,
                                        cg_tag const & tag,
                                        viennacl::linalg::no_precond,
                                        bool (*monitor)(viennacl::vector<NumericT> const &, NumericT, void*) = NULL,
                                        void *monitor_data = NULL)
  {
    return detail::pipelined_solve(A, rhs, tag, viennacl::linalg::no_precond(), monitor, monitor_data);
  }



  template<typename NumericT>
  viennacl::vector<NumericT> solve_impl(viennacl::sliced_ell_matrix<NumericT> const & A,
                                        viennacl::vector<NumericT> const & rhs,
                                        cg_tag const & tag,
                                        viennacl::linalg::no_precond,
                                        bool (*monitor)(viennacl::vector<NumericT> const &, NumericT, void*) = NULL,
                                        void *monitor_data = NULL)
  {
    return detail::pipelined_solve(A, rhs, tag, viennacl::linalg::no_precond(), monitor, monitor_data);
  }


  template<typename NumericT>
  viennacl::vector<NumericT> solve_impl(viennacl::hyb_matrix<NumericT> const & A,
                                        viennacl::vector<NumericT> const & rhs,
                                        cg_tag const & tag,
                                        viennacl::linalg::no_precond,
                                        bool (*monitor)(viennacl::vector<NumericT> const &, NumericT, void*) = NULL,
                                        void *monitor_data = NULL)
  {
    return detail::pipelined_solve(A, rhs, tag, viennacl::linalg::no_precond(), monitor, monitor_data);
  }


  template<typename MatrixT, typename VectorT, typename PreconditionerT>
  VectorT solve_impl(MatrixT const & matrix,
                     VectorT const & rhs,
                     cg_tag const & tag,
                     PreconditionerT const & precond,
                     bool (*monitor)(VectorT const &, typename viennacl::result_of::cpu_value_type<typename viennacl::result_of::value_type<VectorT>::type>::type, void*) = NULL,
                     void *monitor_data = NULL)
  {
    typedef typename viennacl::result_of::value_type<VectorT>::type           NumericType;
    typedef typename viennacl::result_of::cpu_value_type<NumericType>::type   CPU_NumericType;

    VectorT result = rhs;
    viennacl::traits::clear(result);

    VectorT residual = rhs;
    VectorT tmp = rhs;
    detail::z_handler<VectorT, PreconditionerT> zhandler(residual);
    VectorT & z = zhandler.get();

    precond.apply(z);
    VectorT p = z;

    CPU_NumericType ip_rr = viennacl::linalg::inner_prod(residual, z);
    CPU_NumericType alpha;
    CPU_NumericType new_ip_rr = 0;
    CPU_NumericType beta;
    CPU_NumericType norm_rhs_squared = ip_rr;
    CPU_NumericType new_ipp_rr_over_norm_rhs;

    if (norm_rhs_squared <= tag.abs_tolerance()) //solution is zero if RHS norm is zero
      return result;

    for (unsigned int i = 0; i < tag.max_iterations(); ++i)
    {
      tag.iters(i+1);
      tmp = viennacl::linalg::prod(matrix, p);

      alpha = ip_rr / viennacl::linalg::inner_prod(tmp, p);

      result += alpha * p;
      residual -= alpha * tmp;
      z = residual;
      precond.apply(z);

      if (static_cast<VectorT*>(&residual)==static_cast<VectorT*>(&z))
        new_ip_rr = std::pow(viennacl::linalg::norm_2(residual),2);
      else
        new_ip_rr = viennacl::linalg::inner_prod(residual, z);

      new_ipp_rr_over_norm_rhs = new_ip_rr / norm_rhs_squared;
      if (monitor && monitor(result, std::sqrt(std::fabs(new_ipp_rr_over_norm_rhs)), monitor_data))
        break;
      if (std::fabs(new_ipp_rr_over_norm_rhs) < tag.tolerance() *  tag.tolerance() || std::fabs(new_ip_rr) < tag.abs_tolerance() * tag.abs_tolerance())    //squared norms involved here
        break;

      beta = new_ip_rr / ip_rr;
      ip_rr = new_ip_rr;

      p = z + beta*p;
    }

    //store last error estimate:
    tag.error(std::sqrt(std::fabs(new_ip_rr / norm_rhs_squared)));

    return result;
  }

}



template<typename MatrixT, typename VectorT, typename PreconditionerT>
VectorT solve(MatrixT const & matrix, VectorT const & rhs, cg_tag const & tag, PreconditionerT const & precond)
{
  return detail::solve_impl(matrix, rhs, tag, precond);
}

template<typename IndexT, typename NumericT, typename PreconditionerT>
std::vector<NumericT> solve(std::vector< std::map<IndexT, NumericT> > const & A, std::vector<NumericT> const & rhs, cg_tag const & tag, PreconditionerT const & precond)
{
  viennacl::compressed_matrix<NumericT> vcl_A;
  viennacl::copy(A, vcl_A);

  viennacl::vector<NumericT> vcl_rhs(rhs.size());
  viennacl::copy(rhs, vcl_rhs);

  viennacl::vector<NumericT> vcl_result = solve(vcl_A, vcl_rhs, tag, precond);

  std::vector<NumericT> result(vcl_result.size());
  viennacl::copy(vcl_result, result);
  return result;
}

template<typename MatrixT, typename VectorT>
VectorT solve(MatrixT const & matrix, VectorT const & rhs, cg_tag const & tag)
{
  return solve(matrix, rhs, tag, viennacl::linalg::no_precond());
}



template<typename VectorT>
class cg_solver
{
public:
  typedef typename viennacl::result_of::cpu_value_type<VectorT>::type   numeric_type;

  cg_solver(cg_tag const & tag) : tag_(tag), monitor_callback_(NULL), user_data_(NULL) {}

  template<typename MatrixT, typename PreconditionerT>
  VectorT operator()(MatrixT const & A, VectorT const & b, PreconditionerT const & precond) const
  {
    if (viennacl::traits::size(init_guess_) > 0) // take initial guess into account
    {
      VectorT mod_rhs = viennacl::linalg::prod(A, init_guess_);
      mod_rhs = b - mod_rhs;
      VectorT y = detail::solve_impl(A, mod_rhs, tag_, precond, monitor_callback_, user_data_);
      return init_guess_ + y;
    }
    return detail::solve_impl(A, b, tag_, precond, monitor_callback_, user_data_);
  }


  template<typename MatrixT>
  VectorT operator()(MatrixT const & A, VectorT const & b) const
  {
    return operator()(A, b, viennacl::linalg::no_precond());
  }

  void set_initial_guess(VectorT const & x) { init_guess_ = x; }

  void set_monitor(bool (*monitor_fun)(VectorT const &, numeric_type, void *), void *user_data)
  {
    monitor_callback_ = monitor_fun;
    user_data_ = user_data;
  }

  cg_tag const & tag() const { return tag_; }

private:
  cg_tag   tag_;
  VectorT  init_guess_;
  bool     (*monitor_callback_)(VectorT const &, numeric_type, void *);
  void     *user_data_;
};


}
}

#endif