[SciPy-User] SciPy+NumPy on 4 major linux distributions

[David Cournapeau]

On Tue, Sep 8, 2009 at 10:48 PM, Lev Givon<lev at columbia.edu> wrote:
> Received from David Cournapeau on Sun, Sep 06, 2009 at 04:59:10AM EDT:
>> On Fri, Sep 4, 2009 at 2:59 AM, Lev Givon<lev at columbia.edu> wrote:
>>
>> > I'm not sure whether the current prebuilt libraries are built with
>> > thread support. I'm also not sure whether the current (3.8.3) prebuilt
>> > libraries consistently provide any improved performance compared to
>> > the netlib blas/lapack.
>>
>> The netlib blas/lapack built with gfortran has quite poor performance.
>> Using SSE/SSE2 alone gives a significant boost. Of course, building
>> your own will give improved performance.
>>
>> cheers,
>>
>> David
>
> When I recently tried comparing the timing of some numpy functions
> that invoke lapack routines when run against the netlib libraries and
> the prebuilt atlas 3.8.3 libraries on Mandriva 2009.1 (32 bit), I
> found that the prebuilt atlas libraries provide no speedup compared to
> the netlib libraries (at least on the various machines to which I have
> access).

Maybe something is wrong in the mandriva package. I have a factor 2
difference between using netlib and atlas on ubuntu for a simple
inversion.

numpy.dot is also much faster if you use ATLAS (one order of magnitude
faster), but that's partly a limitation of numpy build system,

cheers,

David