The effective use of SIMD processor arrays

Many computationally intensive problems have a very high degree of structured, fine-grain parallelism and benefit substantially from highly parallel execution. Such problems map efficiently on to truly parallel SIMD (Single Instruction stream, Multiple Data stream) processor arrays. This is essentially because the control, synchronisation and communication associated with the constituent parallel processes are implicit in the simplicity and regularity of the control and communication structure of such an architecture. To effectively use a processor array requires a mapping of the regular data structures and operations implied by the algorithms on to the regular structure of the hardware. The paper investigates this general mapping problem with particular emphasis on an `indirect´ approach known as parallel data transforms