This paper considers the implementation of a fast Fourier transform (FFT) structure using arrays of read-only memories. The arithmetic operations are based entirely on the residue number system. The most important aspect of the structure relates to the scaling arrays, which are required to prevent overflow. Because of the limitations of the number system, scaling factors have to be chosen on an a priori basis. This paper develops optimum procedures for choosing both scaling factors and the position of scaling arrays in the structure. Some examples are presented relating to the filtering of speech via a convolutional filter structure.

This paper considers an implementation of the FFT based upon the residue number system. This system offers the advantages of using integer based arithmetic operations and a simple hardware realization involving table look-up arrays. The proposed architecture is such that rapid evolutionary changes in read-only-memory technology can be easily incorporated into the hardware realization. In this paper, a generalized expression for predicting RMS relative error that includes A/D quantization, integer normalization and scaling rounding considerations, has been derived. This analysis leads to the optimal choice of several parameters, which are related to error minimization and simplification of the hardware realization. The generalized expression derived with respect to the FFT can also be used to predict errors in high-speed convolution filters.

Haralick has shown that the discrete cosine transform of N points can be computed more rapidly by taking two N-point fast Fourier transforms (FFT's) than by taking one 2N-point FFT as Ahmed had proposed. In this correspondence, we show that if Haralick had made use of the fact that the FFT's of real sequences can be computed more rapidly than general FFT's, the result would have been reversed. A modified algorithm is also presented.