Multi - processor FFT

Computing the Fast Fourier Transform on a distributed memory architecture by a direct pipelined radix-2 algorithm, a bi-section or multi-section algorithm, all yield the same communications requirement, if communication for all FFT stages can be performed concurrently, the input data is in normal order, and the data allocation consecutive. With a cyclic data allocation, or bit-reversed input data and a consecutive allocation, multi-sectioning o ers a reduced communications requirement by approximately a factor of two. For a consecutive data allocation, normal input order, a decimation-in-time FFT requires that P N + d 2 twiddle factors be stored for P elements distributed evenly over N processors, and the axis subject to transformation distributed over 2 d processors. No communication of twiddle factors is required. The same storage requirements hold for a decimation-in-frequency FFT, bit-reversed input order, and consecutive data allocation. The opposite combination of FFT type and data ordering requires a factor of log 2 N more storage for N processors. The peak performance for a Connection Machine system CM-200 implementation is 12.9 G ops/s in 32-bit precision, and 10.7 G ops/s in 64-bit precision for unordered transforms local to each processor. The corresponding execution rates for ordered transforms are 11.1 G ops/s and 8.5 G ops/s, respectively. For distributed oneand two-dimensional transforms the peak performance for unordered transforms exceeds 5 G ops/s in 32-bit precision, and 3 G ops/s in 64-bit precision. Three-dimensional transforms executes at a slightly lower rate. Distributed ordered transforms executes at a rate of about 1 2 to 2 3 of the unordered transforms.