[SciPy-Dev] Use vecLibFort instead of Accelerate?

[Matthew Brett]

Hi,

On Sat, Oct 11, 2014 at 11:42 AM, Sturla Molden <sturla.molden at gmail.com> wrote:
> Matthew Brett <matthew.brett at gmail.com> wrote:
>
>> That sounds like an excellent idea.  Sturla - would you consider
>> making a PR for this?  I would be very happy to change the numpy and
>> scipy OSX automated builds ...
>
> I can look into it, I was just going to ask your opinion first. :-)
>
> It is actually just a matter of adding one C file and two header files to
> the build process, and then we link against Accelerate as before.
> vecLibFort is actually a very tiny library. The reason this is different
> from the current ABI wrappers is some magic it does with dlload and dlsym
> to replace the patched BLAS and LAPACK symbols. For the proposed Cython
> BLAS and LAPACK layer it should silence the fortranname bug in f2py, which
> causes a problem with the current ABI wrappers.
>
> I don't think the numpy.dot problem with Accelerate would be fixed though,
> but the parallel problem with scipy.linalg.lapack.sgemv would go away.
> vecLibFort only patches the Fortran subroutine it seems.
>
> Presumably one could also add a fix for cblas_sgemv the same way in a
> modified vecLibFort. That way we could fix dotblas for the AVX segfault
> without actually modifying the code. But that is a NumPy issue and this is
> the SciPy list. So we should keep this separate I think.

I think the discussion is OK here because it relates to numpy and
scipy OSX binaries - maybe Cc the numpy list though?

Is that correct - that using vecLibFort will _not_ fix this issue of
segfaults for float32 with numpy.dot? :

https://github.com/numpy/numpy/issues/4007

That would make it much less attractive...

Cheers,

Matthew