PyTrilinos.Anasazi

The Anasazi Trilinos package makes heavy use of templates, which makes it nearly impossible for the python interface to track it identically. Anasazi classes are templated on one or more of the following parameters:

<ScalarType, MV, OP>

where ScalarType is a primitive data type specifying the floating point representation of the eigenproblem, MV is a multi-vector class and OP is an operator class. All Anasazi classes are templated on <ScalarType>. Some are templated on <ScalarType, MV>, and some are templated on all three parameters.

Currently, PyTrilinos.Anasazi supports only one interface, to Epetra. Thus, the corresponding concrete C++ instantiations are on:

<double, Epetra_MultiVector, Epetra_Operator>

These are indicated in python by using the Anasazi class name and adding the suffix Double for ScalarType-only templated classes, or the suffix Epetra for any of the other templated classes. So, for example:

>>> from PyTrilinos import Anasazi
>>> eig = Anasazi.ValueDouble()
>>> print eig
0+0j
>>> bSort = Anasazi.BasicSortEpetra()
>>> print bSort
<Anasazi.BasicSortEpetra; proxy of <Swig Object of type 'Anasazi::BasicSort<double,
Epetra_MultiVector,Epetra_Operator > *' at 0x6eef50> >

An additional layer of abstraction is provided so that these suffixes need not be remembered or used. Since there is currently only one interface, these suffix-less class names are essentially aliases for their longer-named counterparts:

>>> eig = Anasazi.Value()
>>> bSort = Anasazi.BasicSort()
>>> print type(eig), type(bSort)
<class 'Anasazi.ValueDouble'> <class 'Anasazi.BasicSortEpetra'>

In the future, if additional interfaces are supported (for example, a NumPy interface), then the requested interface will be inferred from the constructor arguments, if possible. For those cases where this is not possible, we will also provide type-specification capabilities that are similar to NumPy.

A second difference between Trilinos Anasazi and PyTrilinos.Anasazi is that any methods that expect a

Teuchos::RCP<object>

as an argument, where object is of any type, will take a python object of corresponding type, without the reference-counted pointer wrapper.

The Anasazi.Eigensolution class has been changed such that the attributes Evals, Evecs and Espace have been changed to methods Evals(), Evecs() and Espace(). Evals() returns a numpy.ndarray of type complex double rather than a wrapper to std::vector<Anasazi::Value<double> >. Also, the Evecs() and Espace() methods return Epetra.MultiVector objects. Currently, index is still an attribute, and is a python list-like object containing integers. (The other attributes must be implemented as methods in order to facilitate the type conversions.)

You can create an Anasazi.Value object if you wish, although it is not very functional. It has a __str__() method so that you can print its value.