[vtkusers] VTK and Python on OS X El Capitan

[Sturla Molden]

Eric Nodwell <eric.nodwell at numerics88.com> wrote:
> I don't think Apple has been entirely consistent here. For example, I
> am 99.999% sure that the operating system never needs to use scipy,
> but they have included that in the system python for a couple of
> versions now. Furthermore, they link it against their Accelerate
> Framework, which is wicked fast - and it is very far from trivial to
> compile scipy with the Accerelate Framework if you want to do it
> yourself,

Actually it is very trivial. Just run setup.py and it will pick Accelerate
by default on OSX. You don't have to specify anything and it will just
work. NumPy does the same. If you want to use a different BLAS/LAPACK it
quickly becomes more complicated and you must start to specify
configuration files for distutils. In void of BLAS information SciPy and
NumPy will use a reference implementation on Linux and Windows, but on OSX
it just picks Accelerate unless you say otherwise. OSX used to be the most
troublesome platform on which to build NumPy and SciPy, but now it is the
easiest. 

Accelerate used to be a bit slow, roughly similar to a generic ATLAS, but
from OSX 10.9 (Mavericks) it suddenly became the fastest BLAS on the market
(even faster than Intel MKL in my tests). Apple has been quite about this
but they really improved the performance. Unfortunaltely they also
introduced a bug in the subroutine SGEMV, which we fixed by calling SGEMM
in the affected cases, but it was fixed by Apple in Yosemite. The patch is
still in NumPy and SciPy, but we will take it away when Mavericks is no
longer relevant. 


Sturla