OpenBLAS vs. MKL performance

[ViralBShah]

We did some performance comparison of OpenBLAS and MKL here and created some plots: JuliaLang/julia#3965

1. OpenBLAS is actually faster than MKL in all the level-1 tests for 1,2, and 4 threads. The difference is larger for smaller problems. In small level-2 and level-3 instances, MKL does better.
2. Specifically in the case of matrix-vector products, MKL seems to do much better, since OpenBLAS does not seem to have different kernels for different matrix sizes, as noted by @xianyi in the above mentioned julia issue.
3. On various linear algebra tests with LAPACK, MKL is faster for smaller problem sizes, whereas OpenBLAS is faster for larger problems.

I am capturing this here, so that it is available for the OpenBLAS team.

Cc: @blakejohnson @andreasnoackjensen @staticfloat

[ViralBShah]

Is this closed due to improved performance of OpenBLAS now?

[wernsaar]

We know, that we need better algorithmens for small and medium sizes.
I reopen this issue as a feature request

Werner

[ViralBShah]

Thanks.

[wernsaar]

Hi,

@ViralBShah. Can you please provide some examples with performance numbers.

Regards

Werner

[blakejohnson]

See the referenced Julia issue. Specifically:
JuliaLang/julia#3965 (comment)
and
JuliaLang/julia#3965 (comment)

[ViralBShah]

I'll open specific issues as necessary. This is too old and a lot has changed.