1 |
On Mon, 26 Mar 2018 08:03:32 +0900 Benda Xu wrote: |
2 |
> Hi Andrew, |
3 |
> |
4 |
> Andrew Savchenko <bircoph@g.o> writes: |
5 |
> |
6 |
> >> I’ll then start the most challenge part, porting and packaging |
7 |
> >> Intel Tools. |
8 |
> > |
9 |
> > While Intel stuff may rightfully be a part of your project, I do not |
10 |
> > recommend to focus on them too much, since this is a proprietary |
11 |
> > software and GSoC is all about Free/Libre software. |
12 |
> > |
13 |
> >> I’m also interested to port Intel’s python distribution |
14 |
> > |
15 |
> > I've discussed this project with Intel devs on one of the |
16 |
> > conferences. There is nothing special about it: it is a normal |
17 |
> > Python linked with Intel libraries and with some math libs replaced |
18 |
> > with more optimized free software solutions. So everyone can do the |
19 |
> > same with Intel MKL without need to obtain Intel Python. They |
20 |
> > created this project mostly due to marketing issues, since python |
21 |
> > is a popular language and management want to establish Intel's |
22 |
> > presence in this area. |
23 |
> > |
24 |
> > If you want to pursue this task, I recommend to build on FLOSS |
25 |
> > solutions as described above, packaging Intel Python itself is |
26 |
> > quite useless. |
27 |
> |
28 |
> +1 |
29 |
> |
30 |
> With a more rubost blas/lapack framework/eclass, an optimized |
31 |
> scipy/numpy linked with OpenBLAS or Intel MKL will be on par if not |
32 |
> overtake the Intel python binaries. |
33 |
|
34 |
[OT] |
35 |
I made some tests with linpack (HPL) with MKL vs OpenBLAS on |
36 |
Gentoo-based HPC cluster at my previous job. OpenBLAS was beating |
37 |
MKL by 0.5 ± 0.1%. Not large, but statistically significant result. |
38 |
But of course MKL is not just BLAS/LAPACK... |
39 |
[/OT] |
40 |
|
41 |
Best regards, |
42 |
Andrew Savchenko |