[SciPy-user] optimize.leastsq crashs my interpreter

[Hans Georg Krauthäuser]

Ed Schofield schrieb:
> On 14/07/2006, at 9:06 PM, weg werp wrote:
> 
>> Hi group,
>>
>> I can confirm some problems with an Athlon 1700+ processor. I just
>> downloaded and installed the latest enthought version
>> (enthon-python2.4-1.0.0.beta4)
>>
>> If you then do:
>> import numpy
>> numpy.test(10,10)
>>
>> you get:
>> ...lots of test results...
>> check_matmat (numpy.core.tests.test_numeric.test_dot) ... ok
>> check_matscalar (numpy.core.tests.test_numeric.test_dot) ... ok
>> check_matvec (numpy.core.tests.test_numeric.test_dot)
>> <crash>
> 
> 
> Thanks to all who have posted information about this.  It looks like  
> NumPy is built against an ATLAS library that uses SSE2 instructions,  
> and that Athlons don't support these.
> 
> I'm working with Travis to iron this out with the next NumPy release  
> (probably 1.0beta1).  In the meantime, if you need to use NumPy /  
> SciPy on an Athlon, you'll need to build your own NumPy binary using  
> the instructions at http://new.scipy.org/Wiki/Installing_SciPy.   
> Building SciPy too shouldn't be necessary, but I can't be sure.  To  
> simplify the build, I'd recommend using Pearu's ATLAS libraries built  
> for Pentium 2.  They were available on old.scipy.org, but it seems  
> this is no longer accessible.  I'll upload my local copy to the wiki  
> in the next few days.
> 
> -- Ed
Meanwhile, I compiled numpy and scipy (latest stable, not svn) on a 
Athlon (Duron) against the ATLAS library with SSE1. It is working for 
me, but I can't guarantee that it will work for someone else. But, if 
someone is interested in the windows installers, I will send it by email.

I didn't tried whether the official scipy works with my self-made numpy, 
but I can try if it helps.

Regards
Hans Georg