Image based visualisation on massively parallel architectures: design considerations for general purpose and application specific arrays

Massively parallel architectures are well suited to a class of SIMD image processing and visualisation algorithms based on separable transformations, for which multi-dimensional operations can be decomposed into independent one-dimensional operations. This allows regularisation of data movements as parallel mappings between computational operations. This paper outlines a generic class of separable visualisation and image processing algorithms, their data movement requirements, and their implementation on a general purpose massively parallel array, the MasPar MP-1. Design and performance issues for a custom ASIC array and an interconnected DSP array are compared against the general purpose implementation. A software framework to support this class of algorithms on parallel platforms, and its extension to include generalisation to tiled data algorithms, are described. Design implications for special purpose hardware are discussed.<>