Signal Processing Computational Needs
Jeffrey M. Speiser · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1986
Previous reviews of signal processing computational needs and their systolic implementation have emphasized the need for a small set of matrix operations, primarily matrix multiplication, orthogonal triangularization, triangular backsolve, singular value decomposition, and the generalized singular value decomposition. Algorithms and architectures for these tasks are sufficiently well understood to begin transitioning from research to exploratory development. Substantial progress has also been reported on parallel algorithms for updating symmetric eigensystems and the singular value decomposition. Another problem which has proved to be easier than expected is inner product computation for high-speed high resolution predictive analog-to-digital conversion. Although inner product computation in a general setting will require 0(log n) time via a tree, the special structure of the prediction problem permits the use of a systolic transversal filter, producing a new predicted value in time 0(1). Problem areas which are still in an early stage of study include parallel algorithms for the Wigner-Ville Distribution function, L1 norm approximation, inequality constrained least squares, and the total least squares problem.