Implementation of a morphological image processing algorithm on an FPS T-20 hypercube

Efficient use of the distributed architecture of a hypercube requires balancing tasks among the processor nodes, of which there are sixteen in the FPS T-20. Since each of the FPS T-20 nodes is a vector processor, algorithms which have a natural vectorization are easier to implement. Morphological image-processing algorithms can be decomposed into the elementary morphological operations of dilation and erosion which, for binary representations of the images, can be realized as vector shifts and vector AND/ORs. Several decompositions of tasks for load balancing are discussed, including different masks for different nodes, different structuring elements, and different intensity thresholds. The tradeoffs between computational costs and communication costs for each decomposition are of particular interest.<>