Wepresent an object-oriented framework for constructing parallel implementations of stencil algorithms. This framework simplifres the development process by encapsulating the commonaspects of stencil algorithms in a base stencil class so that application-specifre derived classes can be easily defined via inheritance and overloading. In addition, the stencil base class contains mechanismsfor parallel execution. The result is a highperformance, parallel, application-specifre stencil class. We present the design rationale for the base class and illustrate the derivation process by defrning two subclasses, an image convolution class ana' a PDE solver. The classes have been implemented in Mentat, an objectoriented parallel programming system that is available on a variety ofplatforms. Performance results are givenfor a networkof Sun SPARCstationIPCS'.
John F. Karpovich, Matthew Judd, W. Timothy Straye