Re: [eigen] Re: recent improvements in the products

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

yes this is something like that but that's a bit more sophisticated especially because the blocks have to be appropriately copied into a packed fashion so that the product kernel can work with optimal performance.

To be honest I've followed Goto's paper:

http://www.cs.utexas.edu/users/flame/pubs/GotoTOMS2.pdf

with some improvements to improvement to optimize code reuse.

and the details of the kernel are here:

http://www.cs.utexas.edu/users/flame/pubs/GotoTOMS_final.pdf

cheers,
Gael.

On Tue, Jul 28, 2009 at 8:49 PM, Jitse Niesen <jitse@xxxxxxxxxxxxxxxxx> wrote:

On Tue, 28 Jul 2009, Gael Guennebaud wrote:

* all these new routines are just high level blocking algorithms built on
top of a single highly optimized product kernel.

With "blocking algorithm", do you mean that you divide the matrices up in blocks to reduce cache misses? For instance, to compute the product of two N-by-N matrices, partition them as an (N/n)-by-(N/n) block matrices of n-by-n blocks, and multiply the blocks. I had a quick look at the code and that does not seem what you're doing ...

Cheers,
Jitse

-- 
Gaël Guennebaud
Iparla - INRIA Bordeaux
(+33)5 40 00 37 95