Three-dimensional FFTs on a digital-signal parallel processor, with no interprocessor communication
Hon Keung Kwan, R.L. Nelson, Edward J. Powers, Earl E. Swartzlander · 2002
This paper presents an efficient method of computing the three-dimensional DFT on the AT&T DSP-3 parallel processor. The method eliminates interprocessor communication in all stages of computation. The DSP-3 parallel processor contains 16 processor elements, each of which performs part of the transform computation. Individual results are sent to the host processor to be accumulated to give the final answer. Properties of the twiddle factors are exploited to simplify the implementation. In particular, the last two stages involve no nontrivial multiplications since the multiplications at the last step are all by 1 and -1.