Outline of a fast hardware implementation of Winograd's DFT algorithm

Abstract

The main characteristics of the discrete Fourier transform (DFT) algorithm considered by Winograd (1976) is a significant reduction in the number of multiplications. Its primary disadvantage is a higher structural complexity. It is, therefore, difficult to translate the reduced number of multiplications into faster execution of the DFT by means of a software implementation of the algorithm. For this reason, a hardware implementation is considered in the current study, taking into account a design based on the algorithm prescription discussed by Zohar (1979). The hardware implementation of a FORTRAN subroutine is proposed, giving attention to a pipelining scheme in which 5 consecutive data batches are being operated on simultaneously, each batch undergoing one of 5 processing phases.