[SciPy-Dev] Accuracy of single-precision FFT

Sturla Molden sturla@molden...
Fri Jun 25 08:07:28 CDT 2010


Den 25.06.2010 11:02, skrev Pauli Virtanen:
> If we are unsure that it's not severe, I think we should disable the
> support for single precision FFT in the 0.8.x branch, and remove the
> corresponding mention from the release notes.
>
>    

That is the safest option. If the single precision FFT is inaccurate, it 
makes sence to coerce to double precision. Double precision is not 
slower on modern architectures anyway. The FPU on x86 uses a 80 bit 
exended double natively, and exceeds both single and double resolution. 
We only save some memory by using single-precision.

Another question is whether we should enforce strict IEEE 754 arthmetics 
for single and double precision FFTs, or or allow 80 bit extended double 
(native on x86). It is both faster and more accurate, but less 
reproducible on other platforms. I am not sure what SciPy currently does.

Sturla



More information about the SciPy-Dev mailing list