Floating-to-fixed-point conversion for digital signal processors

Abstract

Digital signal processing applications are specified withfloating-point data types but they are usually implemented inembedded systems with fixed-point arithmetic to minimise cost andpower consumption. Thus, methodologies which establishautomatically the fixed-point specification are required to reducethe application time-to-market. In this paper, a new methodologyfor the floating-to-fixed point conversion is proposed forsoftware implementations. The aim of our approach is to determinethe fixed-point specification which minimises the code executiontime for a given accuracy constraint. Compared to previousmethodologies, our approach takes into account the DSParchitecture to optimise the fixed-point formats and thefloating-to-fixed-point conversion process is coupled with thecode generation process. The fixed-point data types and theposition of the scaling operations are optimised to reduce thecode execution time. To evaluate the fixed-point computationaccuracy, an analytical approach is used to reduce theoptimisation time compared to the existing methods based onsimulation. The methodology stages are described and severalexperiment results are presented to underline the efficiency ofthis approach.