Go back to the benchmark results.

Fourier Transform Errors:

The errors were the roughly the same on all the machines, so we only present the results for one platform, the Sun UltraSPARC.

You will notice that the errors for CWP are several orders of magnitude worse than those of the other codes. This is not due to any inherent weakness in the prime-factor algorithm, but rather to the fact that the constants hard-coded in the CWP source only have nine decimal places (the code was originally intended for single precision calculations).

The essential pattern seems to be that the codes which use pre-computed trig. tables (e.g. FFTW, FFTPACK) are more accurate than the codes that build up their trig. values on the fly using iterative algorithms (e.g. GO, NR, Mayer).

Errors for 1D Transforms

Errors for 1D Transforms, Zoomed

Errors for 3D Transforms

Go back to the benchmark results.