[Numpy-discussion] Caution about using intrisincs, and other 'advanced' optimizations

[Ralf Gommers]

On Fri, Nov 15, 2013 at 8:47 PM, Julian Taylor <
jtaylor.debian at googlemail.com> wrote:

> >
> > Will do, but the errors I am seeing only appear in the
> > simc.inc.src-based implementation of BOOL_logical_or (they disappear if
> > I disable the simd intrinsics manually in the numpy headers).
> >
>
> that is because the simd code always looks at the stride (as it only can
> run with unit strides) while the simple loop doesn't if the dimension is 1.
>
> GCC 4.1 is older than python2.5 which we do not support anymore in numpy
> >= 1.8.
> If you insist on using a buggy old compiler one could always use numpy 1.7.
>

Compiler age and Python version are not equivalent. The former is much
harder to upgrade, and much more depends on it for a user. OS X still ships
gcc 4.2, and the default compiler for Python 2.6 on OS X is gcc 4.0 which
we definitely still support. On Windows we even need gcc 3.4.5 until we
find the right way to get rid of it.

Ralf



> Also intrinsics are not more prone to compiler bugs than any other code,
> so I see no reason to special case them.
>
The code itself is more prone bugs due to its higher complexity in some
> parts, but I think it is reasonably well tested.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/numpy-discussion/attachments/20131116/7a71c186/attachment.html>