[Numpy-discussion] Problem with numpy.linalg.eig?

[David Cournapeau]

Keith Goodman wrote:
> On Nov 13, 2007 8:42 PM, David Cournapeau <david at ar.media.kyoto-u.ac.jp> wrote:
>>
>> Here we are:
>>
>> http://www.ar.media.kyoto-u.ac.jp/members/david/archives/numpy-1.0.4.win32-py2.4.exe
>
> Thank you. He said it worked. He didn't even notice a slow down
> without ATLAS. On some calculations the results were different
> (between 1.0.2 and 1.0.4) in the last three decimal places. But that's
> to be expected, right? 
I don't think there is any chance to have exactly the same results 
(compiler/OS/CPU/BLAS all enter the equation). ATLAS will not change 
much the general performances of numpy: this only enter the equation for 
some functions (numpy.dot) and linear algebra of course, for relatively 
big numbers. For example, in my use case (linear algebra with maximum a 
few tens dimensions), ATLAS does not give much outside numpy.dot. And 
anyway, if you want good performance from atlas, you should compile it 
by yourself (ATLAS performance seems to really depend on the size of L1 
cache, for example).

So all in all, I think it worths considering just using netlib 
BLAS/LAPACK instead of ATLAS for binaries, at least on windows (I don't 
know who is responsible for the windows binaries); note that we still do 
not know why the official binaries hang, which is bothering.

cheers,

David