Generalised parallelism for embedded vision applications

Parallel architectures have yet to achieve the generality of sequential processing, particularly in areas such as image processing and vision. The main reason why parallel processing does not exhibit the same generality as sequential processing is because a large number of new degrees of freedom have been introduced for both the compiler writer and programmer. They include class of architecture, topology, programming paradigm reconfiguration, and the number of processors to be used. The complex relationship between these factors in typical real-time vision applications, makes performance estimation almost impossible and thus encourages a heuristic design philosophy. The authors approach is to fix some of these degrees of freedom to constrain the parallel architecture so that performance estimation is feasible. In this paper it is shown that, although not optimal in all areas, tree topologies of distributed processors with point-to-point communication links (e.g. the transputer) offer the best compromise between generality, performance and cost