Re: [eigen] Parallel matrix multiplication causes heap allocation

[Benoit Steiner <benoit.steiner.goog@xxxxxxxxx>]

I've just started to merge some support for the libxsmm matrix multiplication kernels into Eigen. These kernels are extremely efficient and perform better than the Eigen ones on Haswell and above.

Note that there is more work to be done, in particular wrt to multi-threading. This will be done over the next few months.

On Mon, Dec 19, 2016 at 11:48 AM, Ilja Honkonen <ilja.j.honkonen@xxxxxxxx> wrote:

Hello

For the KNL, I have access to one of them but the performance is
awful compared to what I had this summer. So I would not trust it.

I could try to reproduce your results on KNL, especially if given copy&paste instructions for setting everything up. I did upload some results in http://eigen.tuxfamily.org/bz/show_bug.cgi?id=937 in case that's relevant here.
Ilja

--

Benoit