A computer architecture for highly parallel signal processing

Abstract

A computer of unusual architecture is described that achieves highly parallel operation through use of a data-flow program representation. The machine is especially suited for signal processing computations such as waveform generation, modulation, and filtering, in which a group of operations must be performed once for each sample of the signals being processed. The difficulties of processor switching and memory/ processor interconnection arising in attempts to adapt Von Neuman type computers for parallel operation are avoided by an organization in which sections of the machine communicate through transmission of information packets, and delays in packet transmission do not compromise effective utilization of the hardware. The design concept is especially suited to implementation using asynchronous logic and large-scale integrated circuits. Application of the concepts to generalized data-flow program languages £8 under study.

Dynamic loop pipelining in data-driven architectures

“…One only triggers the execution when the control token and the two data tokens are ready, the other one triggers the execution as soon as the control token and the selected data token are ready (this type of evaluation is called lenient in [5]). Other special operators are specifically used to discard tokens, e.g., the T-and F-Gates used in some dataflow machines, which only copy input data to output when the control token has value "true" or when has value "false", respectively [12].…”

“…They resemble many of the concepts of processor arrays, introduced in the 80's [31] [12], namely wavefront [23] and data-driven arrays [22]. Those architectures devised a scalable and effective fashion to directly support the dataflow computational model and spatial computing.…”

Self-loop Pipelining and Reconfigurable Dataflow Arrays

“…Other special operators are specifically used to discard tokens, e.g., the T-and FGates used in some dataflow machines, which copy input data to output when the control token has value "true" or when has value "false", respectively [2].…”

“…Notice, however, that with respect to reconfiguration support several differences could be sketched. (4) WaveScalar [9] Dynamic C Yes (2) Yes (5) Besides work on using dataflow languages to program data-driven arrays [1], some efforts have been conducted to use imperative programming languages (e.g., [17]). However, lack of specific machine operations to effectively support high-level languages has been one of the major difficulties to attain more efficient compilation results.…”

“…Dataflow machines [1] have been promising to overcome the poor-support of parallelism of von Neumann architectures since the early 70's [2]. However, their envisaged use has been transcended by efforts on augmenting the parallel processing capabilities of traditional processors (e.g., VLIW).…”