[WIP] FEA Support Veclib/Accelerate

[jeremiedbb]

Fixes #135.

[jeremiedbb]

I don't think it's possible to support Accelerate:

• when installed by conda-forge, there's only a libblas.dylib (whereas openblas for instance comes with libblas.dylib and libopenblas.dylib) so threadpoolctl can't rely on the lib name only to guess which one it is.
• this libblas.dylib seems to not expose any symbol that allows to identify accelerate. It only exposes the generic blas symbols. I say that based on running nm -a libblas.dylib, where the only non generic symbol _veclib which is a local (non-external) bss (block starting symbol) section symbol which is not accessible in python using hasattr.

[ogrisel]

When using otool I can see that it's actually linked to libvecLibFort-ng via an @rpath entry:

otool -L /Users/ogrisel/mambaforge-disabled/envs/dev/lib/libblas.3.dylib
/Users/ogrisel/mambaforge-disabled/envs/dev/lib/libblas.3.dylib:
        @rpath/libvecLibFort-ng.dylib (compatibility version 0.0.0, current version 0.0.0)
        /System/Library/Frameworks/Accelerate.framework/Versions/A/Frameworks/vecLib.framework/Versions/A/libBLAS.dylib (compatibility version 1.0.0, current version 1.0.0, reexport)
        @rpath/liblapack-netlib.3.9.0.dylib (compatibility version 0.0.0, current version 0.0.0, reexport)
        @rpath/liblapacke-netlib.3.9.0.dylib (compatibility version 0.0.0, current version 0.0.0, reexport)
        /usr/lib/libSystem.B.dylib (compatibility version 1.0.0, current version 1292.0.0)

[ogrisel]

Here are known cases of numpy being linked accelerate in the wild:

• numpy installed from conda-forge with conda install -c conda-forge numpy libblas=*=*accelerate"
• the official numpy wheels for macOS on pypi.org as discussed in https://github.com/joblib/threadpoolctl/issues/135#issuecomment-1782572823 (not the case at the time of writing though, see the linked discussion);
• the macports numpy package: https://github.com/joblib/threadpoolctl/issues/162#issuecomment-1932138190

Note that the various packaging methods my differ in how vecLib is linked to the Python program. This might introduce additional difficulty in finding a robust heuristic to detect accelerate/vecLib given the fact that there are no vecLib-specific exported symbols, as far as I know.

[jeremiedbb]

So irl we discussed the possibility to check if the detected libblas.dylib has a dependency over libvecLibFort-ng.dylib, but I didn't manage to do that easily.

Although on linux it's seems doable using dlinfo, this is gnu only, and I didn't an equivalent for macos. The dyld functions seems too limited (https://developer.apple.com/library/archive/documentation/System/Conceptual/ManPages_iPhoneOS/man3/dyld.3.html).

I tried to look at the source code of otool to find how they do it but the solution seems way to complicated...

I have an alternative that I don't particularly like. When we find a libblas we create a subprocess where we only dlopen this libblas and check if _find_libraries_with_dyld finds libvecLibFort-ng.dylib. It would not add that much complexity to the code base but the cost of creating a new process is clearly not negligible.

[jeremiedbb]

Another alternative that we discussed irl but I'm not a fan either is to directly check for libvecLibFort-ng.dylib but it's not the real dylib containing the blas, and the info dict would not show the path to libblas either.

[jeremiedbb]

Anyway, let's not delay the release any longer for that.