[SciPy-dev] All newscipy tests pass, except on 64-bit system

[Travis Oliphant]

Stephen Walton wrote:

>Pearu Peterson wrote:
>
>  
>
>>Hi,
>>Yes! Now all newscipy tests pass on my 32-bit system
>>
>>    
>>
>Sorry, but this is not true here.  I traced a failure in scipy.stats to 
>a bug which the following code demonstrates (the BIG array is cribbed 
>from scipy.stats.tests.test_stats.py):
>
>    from scipy import stats
>    from scipy.base import array,Float
>    mean = 99999995.
>    BIG=array([99999991,99999992,99999993,99999994,99999995,99999996,99999997,99999998,99999999],Float)
>    print BIG-mean
>    print BIG-array(mean)
>    print BIG-array([mean])
>
>  
>
We are getting errors in stats too.  He said this before he added back 
the tests for the untested packages.

>The last line is, I think, not quite correct.  A Fedora Core 4 system 
>with Absoft Fortran gives similarly wacky results.
>  
>
Good catch.  This was a bug introduced by some optimizations I did a 
while ago.   It was executing the fastest section of code but with 
incorrect step sizes.  This should be fixed in SVN.

-Travis