[SciPy-Dev] disabling Accelerate on OS X

[Jack Howarth]

On Wed, Jun 12, 2013 at 10:10:54AM -0500, Kyle Mandli wrote:
> Anybody out there at WWDC who could ask the Apple engineers?
> 
> I am trying to build SciPy and NumPy on 10.9 right now, will let you know
> if I get anywhere.

Kyle,
   I've convinced myself that the failures are indeed intrinsic to the Accelerate framework
and not due to newer compillers in 10.7 and 10.8. A build of scipy 0.12.0 on x86_64-apple-darwin10
produces...

OK (KNOWNFAIL=15, SKIP=36)

with...

PYTHONPATH=/sw/src/fink.build/root-scipy-py27-0.12.0-2/sw/lib/python2.7/site-packages /sw/bin/python2.7 -c "import scipy; scipy.test()"

whereas the same build and install directories transferred to a 10.8 machine produces...

FAILED (KNOWNFAIL=15, SKIP=36, errors=1, failures=72)

Just to be clear, have standalone test cases that exhibit these single and double precision failures been submitted?
This is what we really need to nudge Apple to fix the Accelerate framework. Perhaps the OpenBLAS testsuite or some other
blas/lapack testsuite can be built against the Accelerate framework and be used to provide some simple testcases (outside
of scipy) for these issues. This would also serve to decouple the Accelerate framework issues from any cmath problems in
python.
                    Jack

> 
> 
> On Wed, Jun 12, 2013 at 1:24 AM, Ralf Gommers <ralf.gommers at gmail.com>wrote:
> 
> >
> >
> >
> > On Wed, Jun 12, 2013 at 1:50 AM, Jack Howarth <howarth at bromo.med.uc.edu>wrote:
> >
> >> On Tue, Jun 11, 2013 at 08:34:31AM +0200, Ralf Gommers 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
> >>
> >> Ralf,
> >>    Has any attempt been made to open radar reports for these specific
> >> regressions
> >> in recent Accelerate framework releases? Also are these failures limited
> >> to single
> >> precision math? Lastly, were these failures introduced with Mac OS X
> >> 10.7? If so,
> >> perhaps they are related to the fact that llvm's compiler-rt was being
> >> used to
> >> build the Accelerate framework rather than the older libgcc derived math
> >> calls.
> >> If a few standalone test cases could be crafted for these Accelerate
> >> regressions
> >> detected by scipy, we might well be able to get those fixed for Mac OS X
> >> 10.9.
> >
> >
> > In 10.7 it looked like it was only single precision, but for 10.8 there
> > are also double precision tests that have started failing. llvm-gcc indeed
> > caused additional issues as well, could be a root cause. For 10.7 Clang and
> > plain gcc worked fine (except for the Accelerate bugs) while llvm-gcc was
> > crashing all over the place.
> >
> > I know there was a bug filed for SDOT in 2009 already,
> > http://www.macresearch.org/lapackblas-fortran-106#comment-17216, but it
> > hasn't been fixed. The Apple bug tracker isn't public, so I don't know what
> > else has been filed.
> >
> > Getting any fixes for 10.9 that are not already in sounds extremely
> > unlikely.
> >
> > 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