On Sun, 07 Jul 2019 05:18:49 -0700 Mo Zhou wrote:
> On 2019-07-07 12:11, Benda Xu wrote:
> > Mo Zhou <lumin@debian.org> writes:
> > 
> >> Next I'm going to integrate intel's math kernel
> >> library into this mechanism. By the mid-July
> >> we would be able to present the document to
> >> more audience and have the mechanism more widely
> >> tested.
> > 
> > FYI, a relevant bug is at
> > 
> >   https://bugs.gentoo.org/672466
> 
> I think it's worthwhile to invest a bit more time
> to the MKL integration from ebuild side. After all
> when comparing all the freely available CPU-based
> BLAS/LAPACK implementations, MKL is always ridiculously
> fast. (Although it's another story on AMD CPU).

Not always. I made tests on Gentoo-based HPC cluster and OpenBLAS
was a bit faster on LINPACK test. This "a bit" was about
0.50 ± 0.05% within 95% probability range. Not much, but
statistically significant. OpenBLAS was built with -march=native
mode. Though this result is about 6 years old now.

Anyway integrating MKL in the runtime switching scheme is a good
idea. Thank you for your work.

Best regards,
Andrew Savchenko