Algorithmic Skeletons within an Embedded Domain Specific Language for the CELL Processor

Tarik Saidani, Joel Falcou, Claude Tadonki, Lionel Lacassagne and Daniel Etiemble

Efficiently using the hardware capabilities of newarchitectures like the Cell processor, a heterogeneous chipmultiprocessor that uses several levels of parallelism todeliver high performance, and being able to reuse legacycode are real challenges for application developers. Wepropose to use Generative Programming and more preciselytemplate meta-programming to design an EmbeddedDomain Specific Language using algorithmic skeletonsto generate applications based on a high-level mappingdescription. The method is easy to use by developers anddelivers performance close to the performance of optimizedhand-written code, as shown on various benchmarksranging from simple BLAS kernels to image processingapplication.

Back to Program