strange numpy fft performance

Question

During testing I have noticed something strange.

I’m FFT’ing a lot of vectors, and from time to time the numpy FFT function seemed to crash.

I briefly debu

Answer 1

Divide-and-conquer FFT algorithms, such as Cooley-Tukey, work much better the more factors the input length has. Powers of 2 work especially well, whereas primes (like 165037) require alternate, slower implementations. If you can pad your input to a power-of-2 length, you may be able to drastically speed up slow FFTs.