Not by hand, but the FFT I wrote for the spectral analyzer is compiled with -ffast-math and gcc vectorizes it (verified with objdump). Same thing with the functions that change the filter parameters.
Automatic vectorization is extremely limited, in all compilers.
Check this out, feel free to merge if it works in reality not just in the test: https://github.com/Const-me/prettyeq/tree/master
Did you test against pulseeffects cpu usage?
If OP did that, I probably wouldn’t bother to improve their code: https://github.com/Const-me/prettyeq/tree/master
While my primary motivation was “for lulz”, possibility to reuse the code whenever I need a fast 1D FFT also contributed.