[Numpy-discussion] Ufunc memory access optimization
Charles R Harris
charlesr.harris@gmail....
Tue Jun 15 10:51:33 CDT 2010
On Tue, Jun 15, 2010 at 9:37 AM, David Cournapeau <cournape@gmail.com>wrote:
> On Wed, Jun 16, 2010 at 12:16 AM, Pauli Virtanen <pav@iki.fi> wrote:
> > ti, 2010-06-15 kello 10:10 -0400, Anne Archibald kirjoitti:
> >> Correct me if I'm wrong, but this code still doesn't seem to make the
> >> optimization of flattening arrays as much as possible. The array you
> >> get out of np.zeros((100,100)) can be iterated over as an array of
> >> shape (10000,), which should yield very substantial speedups. Since
> >> most arrays one operates on are like this, there's potentially a large
> >> speedup here. (On the other hand, if this optimization is being done,
> >> then these tests are somewhat deceptive.)
> >
> > It does perform this optimization, and unravels the loop as much as
> > possible. If all arrays are wholly contiguous, iterators are not even
> > used in the ufunc loop. Check the part after
> >
> > /* Determine how many of the trailing dimensions are contiguous
> > */
> >
> > However, in practice it seems that this typically is not a significant
> > win -- I don't get speedups over the unoptimized numpy code even for
> > shapes
> >
> > (2,)*20
> >
> > where you'd think that the iterator overhead could be important:
>
> I unfortunately don't have much time to look into the code ATM, but
> tests should be run with different CPU. When I implemented the
> neighborhood iterator, I observed significant (somtimes several tens
> of %) differences - the gcc version also matters,
>
>
That's a common problem with trying to optimize at that level, things become
architecture and compiler dependent. Reminds me a bit of an experiment where
the experimenter was using genetic optimization to design a circuit on a
chip and the optimal design ended up taking advantage of some stray
capacitance.
Chuck
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/numpy-discussion/attachments/20100615/4655dcd6/attachment.html
More information about the NumPy-Discussion
mailing list