This is my old single-file double-precision SSE2 FFT.
You can copy the header part into a separate file, or you
can include the C-file with DUSTFFT_HEADER_ONLY defined.
While it compiles as C++ the performance is known to be better at least with some compilers when compiled as regular C.
It also has a scalar version as a fall-back for unaligned, but if you want compile it on non-x86 platforms, then you'll have to edit the SSE versions out (or include something like sse2neon).
This version is unchanged from the 2014 version, except for the additional real/half-complex wrappers.
While it does power-of-two only, it should be reasonably fast and works perfectly fine at least up to a few million points and has been written to be somewhat cache-aware.