[SciPy-Dev] disabling Accelerate on OS X

[Ralf Gommers]

On Tue, Jun 11, 2013 at 10:48 PM, Kyle Mandli <kyle.mandli at gmail.com> wrote:

> Thanks for the pointers on performance, sounds like it will be a mixed bag
> in my case as I use both the eigen and solver functionality.   Excuse my
> naivety on the topic but would it be possible to not remove support but
> encourage binary maintainers to use OpenBLAS instead?
>

The main issue is not binaries but people building from source. A simple
"python setup.py install" will result in ~70 test failures, plus whatever
is broken in scipy.linalg that's not caught by tests.

Maybe something else can be done - Pauli proposed something on
numpy-discussion just now that may help.

Ralf


>
>
> On Tue, Jun 11, 2013 at 3:09 PM, Ralf Gommers <ralf.gommers at gmail.com>wrote:
>
>>
>>
>>
>> On Tue, Jun 11, 2013 at 9:16 PM, Kyle Mandli <kyle.mandli at gmail.com>wrote:
>>
>>> Any idea what kind of performance decrease is this likely to cause
>>> (switching from accelerate to OpenBLAS)?  I currently rely on wrapping a
>>> Fortran routine that uses system_info to find the LAPACK libraries (via
>>> f2py --link-lapack_opt).
>>>
>>
>> Not necessarily a decrease - on average OpenBLAS seems to be comparable
>> to MKL and Accelerate, but it probably depends strongly on your application
>> and your CPU. For example in this thread (
>> https://github.com/Homebrew/homebrew-science/issues/7) someone finds
>> OpenBLAS is faster, but on eigenvalue decompositions it's slower.
>> http://home.uchicago.edu/~skrainka/pdfs/Talk.Eigen.pdf also has a few
>> benchmarks showing that OpenBLAS performance is similar to MKL.
>>
>> If you're interested, https://github.com/samueljohn/homebrew-pythonallows you to install Numpy and Scipy with OpenBLAS.
>>
>> Ralf
>>
>>
>>
>>>
>>> Kyle
>>>
>>>
>>> On Tue, Jun 11, 2013 at 1:34 AM, Ralf Gommers <ralf.gommers at gmail.com>wrote:
>>>
>>>> Hi,
>>>>
>>>> Given the issues we've been having with Accelerate on OS X 10.7 and
>>>> 10.8, we plan to disable support for it completely in scipy (and numpy,
>>>> will bring that up on numpy-discussion later) before the next release.
>>>>
>>>> Background:
>>>> https://github.com/scipy/scipy/issues/2248
>>>> https://github.com/scipy/scipy/issues/2547
>>>>
>>>> This will make compiling on OS X harder, but we can't leave
>>>> functionality giving incorrect results (and ~70 test errors) hang around
>>>> forever, so we have to do something. We'll do some testing and write up a
>>>> guide for how to build against other BLAS/LAPACK implementations. Homebrew
>>>> already supports OpenBLAS, so that's a good option for who doesn't want to
>>>> go through compiling their own libraries.
>>>>
>>>> Ralf
>>>>
>>>>
>>>> _______________________________________________
>>>> SciPy-Dev mailing list
>>>> SciPy-Dev at scipy.org
>>>> http://mail.scipy.org/mailman/listinfo/scipy-dev
>>>>
>>>>
>>>
>>> _______________________________________________
>>> SciPy-Dev mailing list
>>> SciPy-Dev at scipy.org
>>> http://mail.scipy.org/mailman/listinfo/scipy-dev
>>>
>>>
>>
>> _______________________________________________
>> SciPy-Dev mailing list
>> SciPy-Dev at scipy.org
>> http://mail.scipy.org/mailman/listinfo/scipy-dev
>>
>>
>
> _______________________________________________
> SciPy-Dev mailing list
> SciPy-Dev at scipy.org
> http://mail.scipy.org/mailman/listinfo/scipy-dev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scipy-dev/attachments/20130611/9203f756/attachment.html>