See #10238 (comment)

scipy.fft currently lacks any plan caching. For repeated transforms, this does a significant amount of duplicate work and makes scipy.fft slower than scipy.fftpack for repeated regular sized ffts. (For one off ffts, pocketfft is still much faster)

Things to look into/discuss:

• How much faster is pocketfft after adding a cache?
• Can we have users pass in the plans instead of caching?
• Should the user have control over the cache itself?