TFFTDP.c is by Dave Edelblute (edelblut@cod.nosc.mil).
The program (tfftdp.c) and latest Double Precision (DP) table of results (tfftdp.tbl) are available via anonymous ftp from 'ftp.nosc.mil' in directory 'pub/aburto'. The ftp.nosc.mil IP address is: 128.49.192.51
I'd appreciate any new results of any kind (new machines, compilers, compiler options). I will also periodically post results to 'comp.benchmarks'. Send results to: aburto@nosc.mil.
TFFTDP.c uses the Duhamel-Hollman split-radix Fast Fourier Transform (FFT) algorithm (Electronic Letters, 5 Jan 1984) to compute double precision FFT's from 32 points up to 262,144 points. This requires at least 4 MBytes of free memory. Although each individual FFT conducted is timed, the overall measure of performance is the run time for the whole program referenced to the run time on a VAX 8650 system. Timing the individual FFT's is difficult due to a lack of precision in the available timer routines.
The Duhamel-Hollman split radix FFT implemented here was not intended to represent the fastest FFT algorithm possible.
In addition IEEE Double Precision is assumed so that performance results reported here will not have an absolute worst error greater than approximately 3.2e-10.
Reference: VAX 8650 using gcc 1.37.1 with gcc -DUNIX -O -lm for a run time of 140.658 seconds (= 1 VAX_FFT).
Results as of 11 Sep 1997: