[Numpy-discussion] Scipy dot
Fri Nov 9 16:13:51 CST 2012
I too encourage users to use scipy.linalg for speed and robustness
(hence calling this scipy.dot), but it just brings so much confusion!
When using the scipy + numpy ecosystem, you'd almost want everything
be done with scipy so that you get the best implementation in all
cases: scipy.zeros(), scipy.array(), scipy.dot(), scipy.linalg.inv().
Anyway this is indeed for another thread, the confusion we'd like to
fix here is that users shouldn't have to understand the C/F contiguous
concepts to get the maximum speed for np.dot()
- The python snippet I posted is still valid and can speed up your
code if you can change all your dot() calls.
- The change in dotblas.c is a bit more problematic because it's very
core. I'm having issues right now to replicate the timings, I've got
better timing for a.dot(a.T) than for a.dot(a). There might be a bug.
It's a pain to test because I cannot do the test in a single python session.
I'm going to try to integrate most of your suggestions, I cannot
guarantee I'll have time to do them all though.
On Fri, Nov 9, 2012 at 8:56 AM, Nathaniel Smith <email@example.com> wrote:
> On Fri, Nov 9, 2012 at 4:25 PM, Gael Varoquaux
> <firstname.lastname@example.org> wrote:
>> On Fri, Nov 09, 2012 at 03:12:42PM +0000, Nathaniel Smith wrote:
>>> But what if someone compiles numpy against an optimized blas (mkl,
>>> say) and then compiles SciPy against the reference blas? What do you
>>> do then!? ;-)
>> This could happen. But the converse happens very often. What happens is
>> that users (eg on shared computing resource) ask for a scientific python
>> environment. The administrator than installs the package starting from
>> the most basic one, to the most advanced one, thus starting with numpy
>> that can very well build without any external blas. When he gets to scipy
>> he hits the problem that the build system does not detect properly the
>> blas, and he solves that problem.
>> Also, it used to be that on the major linux distributions, numpy would not
>> be build with an optimize lapack because numpy was in the 'base' set of
>> packages, but not lapack. On the contrary, scipy being in the 'contrib'
>> set, it could depend on lapack. I just checked, and this has been fixed
>> in the major distributions (Fedora, Debian, Ubuntu).
>> Now we can discuss with such problems should not happen, and put the
>> blame on the users/administrators, the fact is that they happen often. I
>> keep seeing environments in which np.linalg is unreasonnably slow.
> If this is something that's been a problem for you, maybe we should
> start another thread on things we could do to fix it directly? Improve
> build instructions, advertise build systems that set up the whole
> environment (and thus do the right thing), make numpy's setup.py
> scream and yell if blas isn't available...?
> NumPy-Discussion mailing list
More information about the NumPy-Discussion