TREE ALGORITHM FOR TWO-DIMENSIONAL CONVOLUTION.

In this paper the authors present an algorithm for 2-D convolution that can be efficiently executed on a general model of parallel architectures. The algorithm is based on the concept of systolic processing employing pipelining and parallel execution to obtain speedups in execution time. However, unlike systolic algorithms which work only on the regular architectures for which they were designed, the present algorithm is flexible with regard to the underlying architectures. This fact has two advantages, namely, it can be executed on a large class of parallel machines like Pyramids, CHiP, MPP etc. Secondly, the independence of the present algorithm from any specific architecture makes the algorithm robust to the failure of processors, switches and communication links in the underlying architecture.