Re: [eigen] AVX support

[ Thread Index | Date Index | More Archives ]

On 05.11.2013 04:50, Gael Guennebaud wrote:
On Tue, Nov 5, 2013 at 2:51 AM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx>wrote:

Regarding AVX support, I'm just watching Eigen from the sidelines anymore,
but I'll say this: the best thing that could happen to Eigen is broadening
of its user base, so it would be great to see features like AVX being
contributed by new/non-core contributors.

there are already two starts which seems to be at about the same stage:

I have just glanced over those forks (please correct me if I state something incorrect). It appears to me that the first merely implemented all the p***-functions, but has not done anything w.r.t to integration. The second appears to have "hacked in" (most likely losing ABI compatibility at some points) its integration into the Core and I would assume it is in a testable state (I can't check, since I don't have an AVX machine at hand).

Maybe someone could try out that fork and compare how computation-times are with it (just for basic operations). Perhaps even compare them with blaze -- I assume their bigger FLOPS are mostly due to AVX usage. I think just testing that does not require deep understanding of Eigen or AVX, so anyone interested in AVX is welcome to do so. If someone gets a benchmark running, it would also be nice to have it run on different architectures.

For a clean integration of AVX, however, I think we indeed need to enhance/rethink the packet logic (e.g. introduce meta-packets where the size is specified by another template parameter). And that is unlikely doable without good internal knowledge.

Do you happen to plan another Eigen-Meeting in the forseable future? This (as well as the Expression-Evaluator) would be nice things which could be boosted at such an event.


Dipl.-Inf., Dipl.-Math. Christoph Hertzberg
Cartesium 0.049
Universität Bremen
Enrique-Schmidt-Straße 5
28359 Bremen

Tel: +49 (421) 218-64252

Mail converted by MHonArc 2.6.19+