An efficient parallel structure for evaluating the triple matrix product algorithm

A parallel structure that can implement a matrix-matrix and a triple-matrix filtering algorithm efficiently is introduced. The architecture is a specialized form of a class of multiple single-instruction multiple-data machines (MSIMDs). It consists of systolic nodes of fixed dimension connected via a system network similar to the mesh interconnection network. The nodes have their own local memory and can access a common global memory. Communication among the nodes is performed by message passing. The structure is called the m ²-array (MSA). A technique is developed that analyzes the evolution of partial results in the MSA. The approach enables the designer to use the compute time as an added degree of freedom. This results in generating a family of low-complexity high-concurrence algorithms in the structure. This class of algorithms offer an attractive alternative for increasing the response time of the dedicated architecture