[SciPy-Dev] Accuracy of single-precision FFT

David david@silveregg.co...
Thu Jun 24 20:47:32 CDT 2010


On 06/25/2010 02:41 AM, Anne Archibald wrote:

>
> We can't use FFTW, unfortunately, but it's clear that prime-size FFTs
> can do much better than we are. (Though what exactly is the definition
> of the errors, below?) FFTW does this, IIRC, by not actually doing
> prime-size FFTs; they have some scheme by which the prime-size FFT is
> viewed as a convolution, then padded and computed with a
> nicely-divisible-size FFT.

Yep, the bluestein algorithm. I thought fftpack in scipy did this, but 
it actually does not (I tested size 4096 vs 4093, one is 100 times 
slower than the other).

So the solution would be to implement this (for both numpy and scipy). I 
won't be able to do it for 0.8, though.

cheers,

David


More information about the SciPy-Dev mailing list