Fast, accurate static analysis for fixed-point finite-precision effects in DSP designs

Abstract

Translating digital signal processing (DSP) software into its finite-precision hardware implementation is often a time-consuming task. We describe a new static analysis technique that can accurately analyze finite-precision effects arising from fixed-point implementations of DSP algorithms. The technique is based on recent interval representation methods from affine arithmetic, and the use of new probabilistic bounds. The resulting numerical error estimates are comparable to detailed statistical simulation, but achieve speedups of four to five orders of magnitude by avoiding actual bit-true simulation. We show error analysis results on both feed forward and feedback DSP kernels.