The computational regularity of convolution can be exploited for efficient design of high-throughput architecture able considerably to decrease the calculation times with respect to those typical of full-software solutions. We describe a fast coprocessor able to perform one- and two-dimensional convolutions, the latter case being applicable provided that the convolving bidimensional filter is separable. The arithmetic unit is based on a semisystolic architecture whose high modularity feature is also useful for an efficient very-large-scale-integration custom implementation. The main novelty here consists in some architectural features permitting convolutions with filters having dimensions greater then the number of cells in the semisystolic array either by paralleling a suitable number of identical arrays or by iterative calculations with a single array. Details concerning the arithmetic unit are given together with a general description of the coprocessor. The overall performances of the actual prototype are also presented.
© 1986 Optical Society of AmericaFull Article | PDF Article
OSA Recommended Articles
Taiwei Lu, Shudong Wu, Xin Xu, and Francis T. S. Yu
Appl. Opt. 28(22) 4908-4913 (1989)
J. Opt. Soc. Am. A 28(3) 410-419 (2011)
Miles J. Murdocca, Alan Huang, Jurgen Jahns, and Norbert Streibl
Appl. Opt. 27(9) 1651-1660 (1988)