Multimegapoint FFT´s running on workstation computers

The authors show that the data addressing required for an in-place fast Fourier transform (FFT) whose memory requirement significantly exceeds the physical memory size of a machine causes an unacceptable level of virtual memory disk access, commonly known as thrashing. In addition, a matrix-decomposition FFT is shown to replace the reordering of the large data set with matrix transpositions, using a two pass memory-efficient matrix transposition technique. The matrix FFT, combined with a file oriented matrix transpose, is shown to be an easy way to prevent thrashing. This method was applied to perform simulations of a 2²³ (8,388,608) point real-only FFT processor which will be used in NASA´s Search for Extraterrestrial Intelligence program