Just a quick question/possibility.

What about just parallelizing ufunc with only 1 inputs that is c or fortran contiguous like trigonometric function? Is there a fast path in the ufunc mechanism when the input is fortran/c contig? If that is the case, it would be relatively easy to add an openmp pragma to parallelize that loop, with a condition to a minimum number of element.

Anyway, I won't do it. I'm just outlining what I think is the most easy case(depending of NumPy internal that I don't now enough) to implement and I think the most frequent (so possible a quick fix for someone with the knowledge of that code).

In Theano, we found in a few CPUs for the addition we need a minimum of 200k element for the parallelization of elemwise to be useful. We use that number by default for all operation to make it easy. This is user configurable. This warenty that with current generation, the threading don't slow thing down. I think that this is more important, don't show user slow down by default with a new version.

Fred




On Wed, May 7, 2014 at 2:27 PM, Julian Taylor <jtaylor.debian@googlemail.com> wrote:
On 07.05.2014 20:11, Sturla Molden wrote:
> On 03/05/14 23:56, Siegfried Gonzi wrote:
>
> A more technical answer is that NumPy's internals does not play very
> nicely with multithreading. For examples the array iterators used in
> ufuncs store an internal state. Multithreading would imply an excessive
> contention for this state, as well as induce false sharing of the
> iterator object. Therefore, a multithreaded NumPy would have performance
> problems due to synchronization as well as hierachical memory
> collisions. Adding multithreading support to the current NumPy core
> would just degrade the performance. NumPy will not be able to use
> multithreading efficiently unless we redesign the iterators in NumPy
> core. That is a massive undertaking which prbably means rewriting most
> of NumPy's core C code. A better strategy would be to monkey-patch some
> of the more common ufuncs with multithreaded versions.


I wouldn't say that the iterator is a problem, the important iterator
functions are threadsafe and there is support for multithreaded
iteration using NpyIter_Copy so no data is shared between threads.

I'd say the main issue is that there simply aren't many functions worth
parallelizing in numpy. Most the commonly used stuff is already memory
bandwidth bound with only one or two threads.
The only things I can think of that would profit is sorting/partition
and the special functions like sqrt, exp, log, etc.

Generic efficient parallelization would require merging of operations
improve the FLOPS/loads ratio. E.g. numexpr and theano are able to do so
and thus also has builtin support for multithreading.

That being said you can use Python threads with numpy as (especially in
1.9) most expensive functions release the GIL. But unless you are doing
very flop intensive stuff you will probably have to manually block your
operations to the last level cache size if you want to scale beyond one
or two threads.
_______________________________________________
NumPy-Discussion mailing list
NumPy-Discussion@scipy.org
http://mail.scipy.org/mailman/listinfo/numpy-discussion