[SciPy-Dev] flapack and the GIL

Pauli Virtanen pav@iki...
Mon Feb 4 13:09:12 CST 2013


Hi,

04.02.2013 20:43, Sturla Molden kirjoitti:
[clip]
> But SAVE can be implicit too. AFAIK, it even used to be (or still is?) 
> the default compiler setting in many f77 compilers. Which means we have 
> to know how LAPACK was built to avoid the problem.
> 
> I don't think there is an implicit SAVE in MKL or Accelerate though ;-)
> 
> It's mainly an issue for those who build their own BLAS or LAPACK.

... and reading carefully Gfortran man page shows that it by default
puts local variables larger than 32768 bytes in static storage. Ouch.

Seems to have bit some people:
http://icl.cs.utk.edu/lapack-forum/viewtopic.php?t=1930

Recent LAPACK versions (>= 3.3) claim to be fully thread safe.
If we can detect that the LAPACK version (how?) is at least that, we
could enable the threadsafe declarations (and do so also for MKL et al.).

-- 
Pauli Virtanen



More information about the SciPy-Dev mailing list