Can MKL be made non-mandatory?
No. Could you describe your scenario? Why is this a problem?
Just don't want to use a proprietary library taking more than a GiB.
It’s only 15MB in binaries (our total size). Sorry, we have no plans to do any non-MKL mode. It gives us a good performance!
It’s only 15MB in binaries (our total size).
The following additional packages will be installed:
libmkl-avx libmkl-avx2 libmkl-avx512 libmkl-avx512-mic libmkl-computational-dev libmkl-core libmkl-def libmkl-dev libmkl-gf-ilp64 libmkl-gf-lp64 libmkl-gnu-thread libmkl-intel-ilp64 libmkl-intel-lp64
libmkl-intel-thread libmkl-interface-dev libmkl-locale libmkl-mc libmkl-mc3 libmkl-meta-computational libmkl-meta-interface libmkl-meta-threading libmkl-pgi-thread libmkl-rt libmkl-sequential
libmkl-tbb-thread libmkl-threading-dev libmkl-vml-avx libmkl-vml-avx2 libmkl-vml-avx512 libmkl-vml-avx512-mic libmkl-vml-cmpt libmkl-vml-def libmkl-vml-mc libmkl-vml-mc2 libmkl-vml-mc3 libomp-10-dev
libomp-dev libomp5-10
Suggested packages:
intel-mkl-doc libomp-10-doc
The following packages will be REMOVED:
libomp-11-dev libomp5-11
The following NEW packages will be installed:
intel-mkl libmkl-avx libmkl-avx2 libmkl-avx512 libmkl-avx512-mic libmkl-computational-dev libmkl-core libmkl-def libmkl-dev libmkl-gf-ilp64 libmkl-gf-lp64 libmkl-gnu-thread libmkl-intel-ilp64
libmkl-intel-lp64 libmkl-intel-thread libmkl-interface-dev libmkl-locale libmkl-mc libmkl-mc3 libmkl-meta-computational libmkl-meta-interface libmkl-meta-threading libmkl-pgi-thread libmkl-rt
libmkl-sequential libmkl-tbb-thread libmkl-threading-dev libmkl-vml-avx libmkl-vml-avx2 libmkl-vml-avx512 libmkl-vml-avx512-mic libmkl-vml-cmpt libmkl-vml-def libmkl-vml-mc libmkl-vml-mc2 libmkl-vml-mc3
libomp-10-dev libomp-dev libomp5-10
0 upgraded, 39 newly installed, 2 to remove and 0 not upgraded.
Need to get 271 MB of archives.
After this operation, 1713 MB of additional disk space will be used.
Which of them do I need? Also I wanna know if you are ready to accept a PR implementing an MKL-less configuration.
@KOLANICH, NeoML requires the following libraries: mkl-core, mkl-sequential, mkl-intel and so you can try to run the command to get them (Debian Linux for example):
sudo apt install libmkl-core libmkl-sequential libmkl-intel-lp64
It's about 30 MB.
@KOLANICH Sorry, this is the wrong way to install MKL. Unfortunately true way is:
sudo apt install libmkl-dev
It will require downloading 250 megabytes from repo :(.
We are ready to accept your PR but it's difficult to match MKL in performance.
but it's difficult to match MKL in performance.
... if one is using a chip by Intel (I use one by AMD). Thanks for the instructions though.