High-performance FFT algorithms for the Convex C4/XA supercomputer

Some implementations of a power-of-two one-dimensional fast Fourier transform (FFT) on vector computers use radix-4 Stockham autosort kernels with a separate transpose step. This paper describes an algorithm that performs well on a Convex C4/XA vector supercomputer on large FFTs by using higher-radix kernels and moving the transpose step into the computational steps. For short transforms a different algorithm is used that calculates the FFT without storing any intermediate results to memory. Performance results using these techniques are given.