Thanks for tracking that one down.
Summary: the FFTW people claim they weren’t trying to be malicious, and after DJB called them out publicly (when his private emails didn’t work) they eventually entirely removed djbfft from their benchmark summary pages.
* * *
For many use cases, it is possible (and helpful) to stick to power-of-2 sizes of FFTs, because these are significantly faster than arbitrary sizes. In some cases it is fine to not re-order the terms at one of the ends because it only matters that the order is consistent internally, not with some external spec; this can also save a bunch of time.
But for other contexts it is of course important to handle any arbitrary lengths including those that aren’t products of small primes, and input/output terms in a canonical order. FFTW tries to be reasonably fast in every possible context by doing fancy runtime code specialization.
I’m curious if anyone has a page with updated benchmarks for various sizes of FFTs on modern CPUs and maybe GPUs, testing whatever FFT libraries out there are still being actively maintained (including those which limit themselves to power-of-2 sizes). CPUs have changed some in the last 10–15 years (more SIMD, more important to optimize for cache, etc.), and GPUs are a whole lot faster now.
The FFTW benchmark site still doesn’t have anything more recent than ~2005 as far as I can tell.
* * *
One more thing: if anyone knows a simple and fast WebAssembly FFT implementation (ideally with benchmarks shown for common web browsers / CPUs) I would love to know about it. Just power-of-2 sizes would be fine.