Investigation of the Indirect Hypercube as Natural Architecture for Parallel Algorithms of a Transpose Type for FFT and Other Fourier-Related Transforms

Abstract

Abstract The natural architectures are architectures, derived from the signal graph of the corresponding algorithm. That is why they are considered to be the most appropriate architectures for parallel realization of this algorithm. For Fast Fourier Transform algorithm (FFT) two types of natural architectures are known – the direct and the indirect hypercube. The direct hypercube has been investigated and analyzed a long time ago. The development of the concept of Indirect Hypercube, although quite old, is too difficult, controversal and still unfinished. Fast Hartley transform (FHT)/Real-valued Fast Fourier transform (RFFT) algorithms are important Fourier-related transforms, because they lower twice the operational and memory requirements when the input data is real-valued. These types of algorithms, however, have an irregular computational structure, which makes their parallel implementation a very difficult task. The aim of this paper is, based on the results achieved so far, to present further development of the concept Indirect Hypercube. A method of parametric synthesis of an indirect hypercube is described as a model of parallel FFT algorithms of a transpose type with different granularity/radix. This method is generalized for relevant RFFT/FHT and FCT algorithms. Two types of SIMD array architectures are described (radix-2 and radix-4), based on the indirect hypercube concept. These architectures are implemented as fast FFT/RFFT/FHT processors for real time applications. The performance estimation, as well as the estimation of resource utilization is carried out.