Parallel processors have long been viewed as potential general purpose signal processing computers, and there are many examples of algorithms that run efficiently on parallel computers. However, because software has become a dominant factor in signal-processing R&D, issues of software reusability and probability have become as important for productivity as the development of efficient parallel algorithms. Several problems have constrained progress in software engineering for parallel signal processing: First, most applications are written originally for serial machines -- "porting" to a parallel computer often means a full redevelopment (not just the revision of a signal processing kernel). Second, a programming method suitable for one parallel architecture is often unsuitable of other architectures. This is compounded by a third problem: there are many alternative parallel architecture under development; there is no clear winner, and there may be many winners. In Phase I, we designed abstractions and interfaces to solve these problems by allowing programming much of the application in a uniform way regardless of whether the underlying implementation runs on a serial machine or any of various parallel architectures. In Phase II we will develop test implementations to demonstrate the feasibility of our approach on a variety of parallel architectures, to evaluate its performance, and to adjust the design. A full library will be built in (optional) Phase II-A. Anticipated
Benefits: The Parallel Signal-Processing Library (PSPL) project will benefit the Federal Government because of the considerable signal-processing and parallel processing R&D work it supports. The PSPL will facilitate the widespread use of parallel processing in signal processing R&D. Because the PSPL will be developed in a manner compatible with an existing commercial product already in wide use in signal-processing laboratories, the probability of commercial success is high.
Keywords: Signal Processing Parallel Computers Parallel Processing Multiprocessors Multicomputers