A specification invariant technique for regularity improvement between flow-graph clusters

In this paper, we present a new technique that improves the regularity between two or more flow-graph clusters (partitions) by means of algebraic transformations, operating at the word-level. Regularity improvement is aimed at decreasing the sharing overhead incurred by sharing clusters of flow-graph on the same resource. It is modelled as an operator cost minimization problem. The technique consists of a steepest descent optimisation step, preceded by a normalisation step. Both steps use optimising algebraic transformations that feature a limited look-ahead. The technique is invariant to structural changes in the specification of the clusters, as long as their functionality is not affected. Our main target domain consists of lowly-multiplexed realisations of real-time DSP applications. The power of our approach will be substantiated on several real-life applications from the video and image processing domain.