Abstract
This paper presents a paradigm for the design of multiple wordlength parallel processing systems for DSP applications based on varying the wordlength and scaling of each signal in a DSP block diagram. A technique for estimating the observable effects of truncation and roundoff error is illustrated, and used to form the basis of an optimization algorithm to automate the design of such multiple wordlength systems. Results from implementation on a reconfigurable computing platform show that significant logic usage savings and increased clock rates can be obtained by customizing the datapath precision to the algorithm according to the techniques described in this paper. On selected DSP benchmarks, we obtain up to 45% area reduction and up to 39% speed increase over standard design techniques.