Belos Package Browser (Single Doxygen Collection)  Development
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Groups Pages
Public Types | Public Member Functions | List of all members
Belos::InnerSolver< Scalar, MV, OP > Class Template Referenceabstract

Inner solver interface. More...

#include <BelosInnerSolver.hpp>

Public Types

typedef Scalar scalar_type
 
typedef Teuchos::ScalarTraits
< Scalar >::magnitudeType 
magnitude_type
 
typedef MV multivector_type
 
typedef OP operator_type
 

Public Member Functions

virtual ~InnerSolver ()
 Virtual destructor, for correctness. More...
 
virtual Teuchos::RCP< const
Teuchos::ParameterList
getCurrentParameters () const =0
 Current parameters for the inner solver implementation. More...
 
virtual InnerSolveResult solve (const Teuchos::RCP< MV > &X, const Teuchos::RCP< const MV > &B, const magnitude_type convTol, const int maxItersPerRestart, const int maxNumRestarts)=0
 Solve $AX=B$ for the given right-hand side(s) B. More...
 
virtual InnerSolveResult solve (const Teuchos::RCP< MV > &X, const Teuchos::RCP< const MV > &B)=0
 Solve $AX=B$ for the given right-hand side(s) B. More...
 

Detailed Description

template<class Scalar, class MV, class OP>
class Belos::InnerSolver< Scalar, MV, OP >

Inner solver interface.

Author
Mark Hoemmen
Warning
This is EXPERIMENTAL CODE. DO NOT RELY ON THIS CODE. The interface or implementation may change at any time.

An "inner solver" wraps an existing linear solver, especially an iterative solver (such as any of the iterative solvers implemented in Belos). InnerSolver is designed especially for implementing the inner solve in inner-outer iterations, such as the following:

InnerSolvers may be used directly as the "OP" template argument in Belos::OperatorTraits, but Belos' current architecture makes this hard to use in practice. The more typical use is via the wrapper defined in Belos::InnerSolveTraits. The InnerSolveTraits::makeInnerSolveOperator() method wraps an InnerSolve instance in an operator appropriate for your linear algebra library: Epetra_Operator for Epetra, Tpetra::Operator for Tpetra, or Thyra::LinearOpBase for Thyra. Then, you can mix and match InnerSolver instances with ordinary matrices or other kinds of operators in your Belos iterative solvers. The wrappers have an "envelope" capability, so your custom iterative solvers can receive the InnerSolver wrapped up in the wrapper, extract the InnerSolver (via Belos::InnerSolveTraits::getInnerSolver()), discard the wrapper if desired, and use the InnerSolver directly.

Definition at line 138 of file BelosInnerSolver.hpp.

Member Typedef Documentation

template<class Scalar, class MV, class OP>
typedef Scalar Belos::InnerSolver< Scalar, MV, OP >::scalar_type

Definition at line 140 of file BelosInnerSolver.hpp.

template<class Scalar, class MV, class OP>
typedef Teuchos::ScalarTraits<Scalar>::magnitudeType Belos::InnerSolver< Scalar, MV, OP >::magnitude_type

Definition at line 141 of file BelosInnerSolver.hpp.

template<class Scalar, class MV, class OP>
typedef MV Belos::InnerSolver< Scalar, MV, OP >::multivector_type

Definition at line 142 of file BelosInnerSolver.hpp.

template<class Scalar, class MV, class OP>
typedef OP Belos::InnerSolver< Scalar, MV, OP >::operator_type

Definition at line 143 of file BelosInnerSolver.hpp.

Constructor & Destructor Documentation

template<class Scalar, class MV, class OP>
virtual Belos::InnerSolver< Scalar, MV, OP >::~InnerSolver ( )
inlinevirtual

Virtual destructor, for correctness.

Definition at line 146 of file BelosInnerSolver.hpp.

Member Function Documentation

template<class Scalar, class MV, class OP>
virtual Teuchos::RCP<const Teuchos::ParameterList> Belos::InnerSolver< Scalar, MV, OP >::getCurrentParameters ( ) const
pure virtual

Current parameters for the inner solver implementation.

These parameters may change values in place, if the five-argument version of the solve() method is called. If you want to preserve the original parameter values, make a deep copy of the returned ParameterList.

template<class Scalar, class MV, class OP>
virtual InnerSolveResult Belos::InnerSolver< Scalar, MV, OP >::solve ( const Teuchos::RCP< MV > &  X,
const Teuchos::RCP< const MV > &  B,
const magnitude_type  convTol,
const int  maxItersPerRestart,
const int  maxNumRestarts 
)
pure virtual

Solve $AX=B$ for the given right-hand side(s) B.

Implementations are free to interpret the convergence tolerance and maximum iteration count parameters in any way they please. (They should interpret these parameters consistently between calls.) They are also free to preserve more state between calls than just the matrix and preconditioner(s); for example, they may choose to perform Krylov subspace recycling.

Parameters
X[in/out] On input: The initial guess for the inner solver, if the inner solver accepts an initial guess (it is not required to do so). On output: the approximate solution to Ax=B as computed by the inner solver. Whether or not the solver accepts an initial guess, X must be allocated to hold the output, and it must be in the range vector space of this solver.
B[in] Right-hand side(s) for which to solve. It must be in the domain vector space of this solver.
convTol[in] "Convergence tolerance," the meaning of which depends on the subclass
maxItersPerRestart[in] Maximum number of iterations per restart cycle in the inner solve.
maxNumRestarts[in] Maximum number of restart cycle(s) in the inner solve.
Returns
The result of the inner solve. It is a single result, aggregated over all right-hand side(s).
Note
Why isn't this method const? First, inner solves almost certainly involve side effects (other than just modifying the output argument X), and likely those side effects affect the state of the inner solver's implementation. Second, solve() may not be a pure function. For example, if the inner solve implementation is a recycling Krylov solver, it may compute a different (hopefully better) approximate solution if given the same inputs twice in a row, or at least it may converge in fewer iterations the second time.
template<class Scalar, class MV, class OP>
virtual InnerSolveResult Belos::InnerSolver< Scalar, MV, OP >::solve ( const Teuchos::RCP< MV > &  X,
const Teuchos::RCP< const MV > &  B 
)
pure virtual

Solve $AX=B$ for the given right-hand side(s) B.

This should do the same thing as the five-argument version of solve(), except it should pick reasonable defaults for the convergence tolerance, maximum number of iterations, and maximum number of restart cycles.

Note
Subclasses' implementations are welcome to change the convergence tolerance, maximum number of iterations, maximum number of restart cycles, and perhaps other solve parameters after every invocation of solve(). For example, solve() may implement an inexact Krylov method, in which the convergence tolerance is gradually increased according to a computed bound. Callers who want to override this behavior of a subclass' implementation should call the four-argument version of solve().
Parameters
X[in/out] On input: The initial guess for the inner solver, if the inner solver accepts an initial guess (it is not required to do so). On output: the approximate solution to Ax=B as computed by the inner solver. Whether or not the solver accepts an initial guess, X must be allocated to hold the output, and it must be in the range vector space of this solver.
B[in] Right-hand side(s) for which to solve. It must be in the domain vector space of this solver.
Returns
The result of the inner solve. It is a single result, aggregated over all right-hand side(s).
Note
Why isn't this method const? First, inner solves almost certainly involve side effects (other than just modifying the output argument X), and likely those side effects affect the state of the inner solver's implementation. Second, solve() may not be a pure function. For example, if the inner solve implementation is a recycling Krylov solver, it may compute a different (hopefully better) approximate solution if given the same inputs twice in a row, or at least it may converge in fewer iterations the second time. Third, the two-argument version of solve() reserves the right to modify the stopping criteria on each call.

The documentation for this class was generated from the following file: