Re: MKL is redistributable; Should we package it?
Hi,
Further investigation: MKL is already available via apt[1] and pip[2],
no registration is needed to download MKL in both ways.
APT
-----
apt sources.list:
```
deb https://apt.repos.intel.com/mkl all main
```
A series of packages named intel-mkl-* can be installed with the above source.
Although there are obvious lintian errors at the Architecture section, e.g.
intel-mkl-core-2018.2-199/all 2018.2-199 amd64
PIP
-----
simply `pip3 install mkl`. The core libs will be pulled.
MKL version of numpy is also available `pip3 install intel-numpy`
MKL as BLAS/LAPACK alternative
-------------------------------------------
The candidate might be libmkl_rt.so . BLAS's dgemm and LAPACK's dgesv
are found in this blob. But I didn't test it yet.
~/.l/lib ❯❯❯ readelf -s libmkl_rt.so | grep ' dgemm$' | grep FUNC
9333: 000000000014d2d0 288 FUNC GLOBAL DEFAULT 12 dgemm
33031: 000000000014d2d0 288 FUNC GLOBAL DEFAULT 12 dgemm
~/.l/lib ❯❯❯ readelf -s libmkl_rt.so | grep ' dgesv$' | grep FUNC
9632: 00000000001855d0 192 FUNC GLOBAL DEFAULT 12 dgesv
30008: 00000000001855d0 192 FUNC GLOBAL DEFAULT 12 dgesv
Well ... I remember several months ago when I was trying to download
MKL, registration is required and what I got is a big all-in-one tarball.
Now things has changed...
[1] https://software.intel.com/en-us/articles/installing-intel-free-libs-and-python-apt-repo
[2] https://software.intel.com/en-us/articles/installing-the-intel-distribution-for-python-and-intel-performance-libraries-with-pip-and
On 30 March 2018 at 07:43, Ghislain Vaillant <ghisvail@gmail.com> wrote:
>
> Le ven. 30 mars 2018 à 08:18, Lumin <cdluminate@gmail.com> a écrit :
>>
>> Hello d-science folks,
>>
>> I noticed that intel's math kernel library is redistributable[1].
>> Since it is widely used in various fields, an mkl package
>> in Debian archive my be beneficial to at least science
>> software users.
>>
>> However despite of the explicit declaration that MKL
>> is redistributable, MKL itself is a proprietary and
>> its source is not available. That means mkl must enter
>> nonfree, and the packaging is to simply repack a pile
>> of binary blobs.
>>
>> As a co-maintainer of CUDA I really dislike working on
>> huge binary blobs that are permitted to redistribute.
>> But CUDA is still uploaded and maintained due to
>> its usefullness.
>
>
> The effort is justified for CUDA due to the lack of established free
> alternatives (apart from OpenCL which Nvidia does not support fully).
>
> For BLAS and FFT, there are already performant and established alternatives
> (OpenBLAS and FFTW), so the appeal to provide a non-free alternative is
> lower, imo.
>
> That being said, I'd be happy if someone were to commit to provide MKL with
> the same level of quality as CUDA in Debian. You guys are doing a great job,
> afaic.
>
>>
>> So, what's your opinion about the mkl package?
>>
>> MKL itself contains BLAS routines, FFT, etc. Scientific
>> packages may possibly gain a performance boost
>> from mkl if they can be linked against mkl.
>>
>> [1] https://software.intel.com/en-us/mkl/license-faq
>> --
>> Best,
--
Best,
Reply to: