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

[Pauli Virtanen]

Wed, 02 Feb 2011 12:53:11 +0000, Pauli Virtanen wrote:
[clip]
> @Cristoph: Have you been able to reproduce random variation in results
> from numpy.dot? If yes, does the same variations occur in results from
> numpy.core.multiarray.dot (which never uses BLAS)? I would be much
> happier if we could point a finger towards MKL as the culprit for non-
> reproducible FP results.

OK, seems to be MKL

http://software.intel.com/sites/products/documentation/hpc/mkl/lin/
MKL_UG_coding_tips/Aligning_Data_for_Numerical_Stability.htm

I wouldn't be too surprised if data alignment and threading also had a 
similar effect with ATLAS.

The correct solution is probably just to bump the test tolerances so that 
they pass with MKL.

-- 
Pauli Virtanen