[SciPy-Dev] ANN: SciPy 0.9.0 release candidate 2

[Christoph Gohlke]

On 2/2/2011 6:50 AM, Pauli Virtanen wrote:
> Wed, 02 Feb 2011 09:43:04 -0500, josef.pktd wrote:
> [clip]
>>  From this it looks like the problem is more general, but until now we
>> have seen it only on Win32.
>
> If the matter is data alignment, perhaps this is because of memory
> allocators or the BLAS library functioning differently on different
> platforms.
>
> To get a definite answer if the issue is data alignment, one can perhaps
> check if the different results are correlated with
>
> 	input_array.__array_interface__['data'][0] % 16
>
> (assuming no copies are made inside dot()).
>

Yes, the 'K' array used in the test is not always aligned optimally. If 
K is aligned, the number of iterations is as expected (116) and the test 
passes. If K is not aligned, the number of iterations is > 116. It does 
not matter to use numpy.core.multiarray.dot instead of numpy.dot in the 
func() function. The 'params' array returned by optimize.fmin_powell() 
varies when K is unaligned but it always passes the 'abs(func(params) - 
func(solution)) < 1e-6' test so it should be safe to increase the test 
tolerance (self.funccalls <= 128).

Christoph