[SciPy-Dev] Accuracy of single-precision FFT

[Ralf Gommers]

On Sat, Jun 26, 2010 at 3:39 AM, Charles R Harris <charlesr.harris at gmail.com
> wrote:

>
>
> On Fri, Jun 25, 2010 at 12:50 PM, Pauli Virtanen <pav at iki.fi> wrote:
>
>> Fri, 25 Jun 2010 12:38:44 -0400, Anne Archibald wrote:
>> [clip]
>> > It seems to me that the right short-term solution is to switch to double
>> > precision for difficult sizes (any non-power-of-two is probably fine)
>> > and add a warning to the docstring (in particular mentioning the
>> > horrible performance as well as the roundoff error, which can be an
>> > issue for doubles too). This we can do for 0.8.
>>
>> Agreed. Done here:
>>
>>
>> http://github.com/pv/scipy-work/compare/0.8.x...disable-single-precision-fft
>>
>> Seems to work, I think it should go in to the 0.8.x branch ASAP.
>>
>> That looks good, I agree that can go in. For trunk I opened a ticket:
http://projects.scipy.org/scipy/ticket/1212.


> I'd also keep at least power of 5 also so that round decimal numbers work
> well.
>

Pauli's patch already kept 2, 3 and 5.

Cheers,
Ralf
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scipy-dev/attachments/20100627/ed39ff0c/attachment.html>