Distributed matricesΒΆ

The DistMatrix<T,U,V> class is meant to provide a distributed-memory analogue of the Matrix<T> class. In a manner similar to PLAPACK, roughly ten different matrix distributions are provided and it is trivial (in the sense that it requires a single command) to redistribute from one to another: in PLAPACK, one would simply call PLA_Copy, whereas, in Elemental, it is handled through overloading the \(=\) operator (or instead calling the Copy() function).

Since it is crucial to know not only how many processes to distribute the data over, but which processes, and in what manner they should be decomposed into a logical two-dimensional grid, an instance of the Grid class must be passed into the constructor of the DistMatrix<T,U,V> class.

Note

Since the DistMatrix<T,U,V> class makes use of MPI for message passing, custom interfaces must be written for nonstandard datatypes. As of now, the following datatypes are fully supported for DistMatrix<T,U,V>: int, float, double, Complex<float>, and Complex<double>.

The DistMatrix<T,U,V> class is implemented in a manner which attempts to expose as many symmetries within the various member functions and redistributions as possible. In particular, there are there is a hierarchy of three increasingly-specific distributed matrix classes: AbstractDistMatrix<T>, which contains every member function which whose prototype does not depend upon the exact matrix distribution, GeneralDistMatrix<T,U,V>, which contains every member function whose prototype depends upon the particular matrix distribution but can still be implemented in a generic way, and DistMatrix<T,U,V>, which contains implementations of member functions which are specific to a particular matrix distribution.