[SciPy-user] Compile scipy from SVN on OS X: no clapack ?

David Cournapeau david@ar.media.kyoto-u.ac...
Wed Apr 8 02:29:58 CDT 2009


Pierre GM wrote:
> On Apr 7, 2009, at 10:41 PM, David Cournapeau wrote:
>
>   
>> Pierre GM wrote:
>>     
>>> All,
>>> I'm working on a piece of code that requires numerous calls to
>>> scipy.linalg.eigvals. I was surprised by the slowness of the code, so
>>> I figured I'd check whether my installation of scipy was correct.
>>>
>>>       
>> When you said surprised by the slowness, is it comparatively to  
>> scipy on
>> another OS on the same/comparable machine ?
>>     
>
> Oh no, I don't have any particular reference, it's just that ~3s to  
> compute the eigenvalues of a 365x365 matrix looked a bit long  
> (especially as I need to compute a lot of those)... I was just afraid  
> I had missed something somewhere.
>   

Sounds a bit long,  but not that much. Finding N eigenvalues for a NxN
matrix is N**3. I have just tested on my machine, it takes ~ 1.5 on a
random 400x400 matrix, and the machine is a bit old (P4 @ 3ghz), with
the debian atlas (that is not very fast).

Recent atlas have included quite a few more optimized LAPACK functions,
maybe the one used by eigvals is one of them.

> But I don;t know if I do! Do I need to compile lapack from source,  
> when I have an atlas from macports and the Accelerate framework ?
>   

Oh, sorry: I meant that the accelerate framework includes optimized blas
(and hopefully some lapack). You did against the accelerate framework,
not with your own ATLAS, right ?

cheers,

David


More information about the SciPy-user mailing list