A computational framework and SIMD algorithms for low-level support of intermediate level vision processing

Computation on and among data sets mapped to irregular, non-uniform, aggregates of processing elements (PEs) is a very important, but largely ignored, problem in parallel vision processing. Associative processing 11] is an eeective means of applying parallel processing to these computations 33], but is often restricted to operating on one data set at a time. What we propose is an additional level of parallelism we call multi-associativity as a framework for performing associative computation on these data sets simultaneously. In this paper we introduce algorithms developed for the Content Addressable Array Parallel Processor (CAAPP) 35] to simulate eeciently within aggregates of PEs simultaneously the associative algorithms typically supported in hardware at the array level. Some of the results are: the eecient application of existing associative algorithms (e.g. 10, 11]) to arbitrary aggregates of PEs in parallel, and the development of new multi-associative algorithms, among them parallel preex and convex hull. The multi-associative framework also extends the associative paradigm by allowing operation on and among aggregates themselves, operations not deened when the entity in question is always an entire array. Two consequences are: support of divide-and-conquer algorithms within aggregates, and communication among aggregates. The rest of the paper describes a mapping of multi-associativity onto the CAAPP, and numerous multi-associative algorithms.