Efficient hardware implementation of data-flow parallel embedded systems

Abstract

Many modern computing systems deal with streams of data, which have to be processed in parallel in order to be handled in real-time. This is in particular the case for some kind of cyber physical systems, which process data provided by physical devices. We consider here an approach to generate efficient hardware for-a particular class of-such systems, which relies upon the polyhedral model. Flexible parallel components, described by the Alpha functional language, are modelled and assembled using a scheduling method which combines the synchronous data-flow principle of balance equations, and the polyhedral scheduling technique. The modelling of flexible components relies on a simple, affine-periodic, delayable and stretchable time model, which allows a full system to be assembled and synthesized by combining the component hardware descriptions with automatically generated wrappers. We illustrate this method on a simplified WCDMA system and we discuss the relationship of this approach with stream languages, latency-insensitive design, and multidimensional data-flow systems.