[Numpy-discussion] numarray and ATLAS

Todd Miller jmiller at stsci.edu
Mon Oct 18 04:53:01 CDT 2004


On Mon, 2004-10-18 at 04:29, Francesc Alted wrote:
> Hi Todd,
> 
> A Diumenge 17 Octubre 2004 03:50, Todd Miller va escriure:
> > I don't think numarray dot() will pick up any boost at all from ATLAS
> > because it's not written to do it.   Besides that,  there are two
> > performance problems I know of with numarray's dot() which may dominate
> > or dilute any ATLAS benefits:
> > 
> > 1. dot() requires array creation.
> 
> Yes, but my guess is that for large arrays, this time should be negligible
> compared with the multiplication time.
> 

Probably true.  I should measure this.  For small computations,  it's an
issue.

> > 2. dot() requires array copies.
> 
> Mmm, you mean even for well-behaved arrays? Sorry, but I don't understand
> why.

I looked at this some this morning,  trying to figure out why this is a
problem only for numarray.  It turns out that Numeric strides its arrays
to get around the copy.   When I implemented numarray,  I chose not to
stride because I thought it would be too slow...  Recently I realized
that one input array to dot() is *always* transposed and therefore
likely noncontiguous and therefore copied.  I think it's now possible to
simply port the Numeric code so I'll look into that.

> May I ask if there is any plan to complete a better integration of external
> LAPACK libraries in numarray or this is considered low priority?

Perry may answer this.  I have no immediate plans for it...  it does
sound like enough people need this that it should be done.

Regards,
Todd






More information about the Numpy-discussion mailing list