Add mkl_path preference to support using "system" MKL

[carstenbauer]

Addressing #82, this PR adds the option to use a "system" MKL, i.e. an external libmkl_rt, instead of the one provided by MKL_jll.jl.

Specifically, a user may specify a mkl_path via a preference, or the environment variable JULIA_MKL_PROVIDER. Supported are the values mkl_jll (default) and system. If system is chosen, we check that libmkl_rt is available on the linker search path, i.e. Libdl.find_library is able to find it, or otherwise throw an error. The user may also specify the path to libmkl_rt explicitly, either via the environment variable JULIA_MKL_PATH or the preference mkl_path. The function set_mkl_path can be used to change the MKL provider preference (it retriggers compilation in contrast to changing the environment variable, which won't have any effect).

Note that with this PR we only call using MKL_jll if it is really needed. Combined with the laziness of the MKL_jll artifact this ensures that the ~1.5 GB (on linux) MKL dependency is only downloaded if mkl_path = "mkl_jll" (default) but not when mkl_path = "system".

Credit: This PR is largely inspired by a draft implementation by @staticfloat which originated out of a Slack discussion.

TODOs (OLD):

• Update README.md
• I commented out https://github.com/crstnbr/MKL.jl/blob/systemmkl2/src/MKL.jl#L85 in this PR. I think this can be dropped?
• Check that Julia 1.6 (no LBT) still works correctly
• Drop Julia < 1.6 support or set preference defaults (Preferences.jl requires 1.6 to function)

Closes #82

[carstenbauer]

Would be great if someone could approve CI for this to test 1.6 behavior (which hopefully shouldn't have been effected) etc.

[carstenbauer]

I included the comments and, in particular, dropped support for Julia versions < 1.6. I also adjusted the CI accordingly.

[carstenbauer]

Still need to check that 1.6 (no LBT) still works but maybe CI can try that for me?

[staticfloat]

Looks pretty good to me!

[carstenbauer]

The 1.6 CI runners seem to fail for different reasons...?

[staticfloat]

It looks to me like it's failing because it can't find the MKL_Set_Threading_Layer symbol; I'm guessing that during precompilation, libmkl_rt doesn't point to where we think it should, and so it can't find that symbol within whatever library it's pointing at.

[jishnub]

Gentle bump. Looks like julia v1.6 support is the holdup here. Since MKL.jl now requires Julia v1.8+, perhaps we may go ahead with this?

[jishnub]

Gentle bump @carstenbauer

[carstenbauer]

Alright, I've resurrected this PR. Should be close to the finish line now (it works fine for me locally). Let's see if CI fails (but it seems to be broken anyways 😞).

I've dropped the environment variable option as the preference option is generally superior. Let's see if we get complaints.

The only thing I'm not super sure about is if !(mkl_path == "mkl_jll" && !MKL_jll.is_available()) in __init__. If a user provides a path to a system MKL we (currently) don't know anything about it and can't set the interface layer etc. correctly. I guess we could also add preferences for the interface etc. but maybe the current option is just fine? (@staticfloat)

[staticfloat]

I guess we could also add preferences for the interface etc. but maybe the current option is just fine?

You don't need to add a preference for this, because LBT automatically figures out the interface through the magic of scary calling convention hacks.

[carstenbauer]

Kind bump.

[carstenbauer]

From a code POV, the PR can be merged. I don't know why CI is failing on Windows/macOS though. For Windows, there are lots of these errors

The `LinearAlgebra/qr` test set mutated ENV and did not restore the original values

which I don't think originate from this PR. For macOS, we have three numerical test errors. Again, I don't see how this could be related to the changes in this PR.

[ViralBShah]

@carstenbauer Please rebase and merge if you think this is good. We have disabled macOS CI anyways, since it is now unsupported and agree it shouldn't hold it back.

Do you have commit access to this repo?