[SciPy-User] Single precision FFT insufficiently accurate.

David Cournapeau cournape@gmail....
Sun Jun 27 21:30:55 CDT 2010


On Mon, Jun 28, 2010 at 8:29 AM, Pauli Virtanen <pav@iki.fi> wrote:
> Sun, 27 Jun 2010 23:13:44 +0200, Sebastian Haase wrote:
>> this workaround seems in-acceptable if single-precision was used (by the
>> user of SciPy) because of memory constrains .... !!!
>
> Suggest a better one, then.
>
> We cannot just return incorrect results, which FFTPACK seems to produce,
> especially as this easily occurs in the size range where memory
> constraints would be important.
>
> Moreover, single-precision FFT is a new feature in 0.8.x, so it probably
> does not have an exceedingly large number of users yet who rely on it.

Nevertheless, Sebasiten remark just made me realize that using double
instead of single in some cases which are input-dependent is not that
great. It means that some program may work in some cases, but will not
in other (because of memory issues).

Maybe it is better to remove it altogether in 0.8.0 - I will try to
implement Bluestein algo during euroscipy,

cheers,

David


More information about the SciPy-User mailing list