Belos
Version of the Day
|
Interface of callback invoked by TsqrOrthoManager on reorthogonalization. More...
#include <BelosTsqrOrthoManagerImpl.hpp>
Inherits binary_function< Teuchos::ArrayView< Teuchos::ScalarTraits< Scalar >::magnitudeType >, Teuchos::ArrayView< Teuchos::ScalarTraits< Scalar >::magnitudeType >, void >.
Public Types | |
typedef Scalar | scalar_type |
The template parameter of this class; the type of an inner product result. More... | |
typedef Teuchos::ScalarTraits < Scalar >::magnitudeType | magnitude_type |
The type of a norm result. More... | |
Public Member Functions | |
virtual | ~ReorthogonalizationCallback () |
Destructor (virtual for memory safety of derived classes) More... | |
virtual void | operator() (Teuchos::ArrayView< magnitude_type > normsBeforeFirstPass, Teuchos::ArrayView< magnitude_type > normsAfterFirstPass)=0 |
Callback invoked by TsqrOrthoManager on reorthogonalization. More... | |
Interface of callback invoked by TsqrOrthoManager on reorthogonalization.
Scalar | The same type as the template parameter of TsqrOrthoManagerImpl. |
This callback's operator() is invoked by TsqrOrthoManagerImpl, and therefore by TsqrOrthoManager. It is invoked right after discovering the need to reorthogonalize (for the first time), but before actually reorthogonalizing. It is only invoked if reorthogonalization is necessary. You can define your own callback by implementing this interface.
This callback lets you collect metrics on reorthogonalization. For example, you might want to measure how often it occurs, or by how much the norms of the vectors drop each time. You can use this information in order to adjust parameters (such as the reorthogonalization parameters) dynamically for your desired balance of accuracy and performance. You might also use it as a numerical debugging aid.
Why a reorthgonalization callback, but not other kinds of callbacks? Reorthogonalization is an event that affects performance, and happens in a data-driven way. Even if you have enabled reorthogonalization, it may not happen at all, or only infrequently. Other kinds of data-driven events (such as a normalization discovering numerical rank deficiency) immediately return to the user with useful diagnostics. Reorthogonalization does not; it happens silently. We could have the orthogonalization method itself gather metrics on reorthogonalization, but the callback lets you define what metrics you want to collect and how you want to display them yourself.
Definition at line 131 of file BelosTsqrOrthoManagerImpl.hpp.
typedef Scalar Belos::ReorthogonalizationCallback< Scalar >::scalar_type |
The template parameter of this class; the type of an inner product result.
Definition at line 138 of file BelosTsqrOrthoManagerImpl.hpp.
typedef Teuchos::ScalarTraits<Scalar>::magnitudeType Belos::ReorthogonalizationCallback< Scalar >::magnitude_type |
The type of a norm result.
This may differ from scalar_type. For example, if scalar_type is complex, magnitude_type will be real.
Definition at line 143 of file BelosTsqrOrthoManagerImpl.hpp.
|
virtual |
Destructor (virtual for memory safety of derived classes)
Definition at line 158 of file BelosTsqrOrthoManagerImpl.hpp.
|
pure virtual |
Callback invoked by TsqrOrthoManager on reorthogonalization.