warning about BLAS.vendor being deprecated on Julia v1.7

[jishnub]

I see the following warning on Julia 1.7 on using PyPlot (which seems to use PyCall internally):

┌ Warning: `vendor()` is deprecated, use `BLAS.get_config()` and inspect the output instead
│   caller = npyinitialize() at numpy.jl:67
└ @ PyCall ~/.julia/packages/PyCall/tqyST/src/numpy.jl:67
PyObject <matplotlib.legend.Legend object at 0x7f8a5b854e50>

Sorry I don't have a repro immediately as the warning doesn't tell me where in my code this is happening. However this is clearly arising because BLAS.vendor is deprecated.

[stevengj]

This is from here:

PyCall.jl/src/numpy.jl

Line 67 in 5d227fc

if LinearAlgebra.BLAS.vendor() === :mkl &&

Hopefully, this warning can just be removed in Julia 1.7, since I think #443 may finally be fixed by @staticfloat's new libtrampoline magic?

[staticfloat]

Mmm, it's unclear to me what the root cause of #443 is; is it that Julia configures MKL to be 64-bit and Numpy expects it to be 32-bit? If so, I guess someone needs to experiment with whether it's possible to load MKL as 32-bit and 64-bit at the same time with LBT.

[stevengj]

is it that Julia configures MKL to be 64-bit and Numpy expects it to be 32-bit

Yes.

It's why we were bugging Intel about getting us an ILP64 version of MKL with prefixed symbols, as you'll recall. (Are we able to use that yet?)

[carstenbauer]

Should we at least replace the BLAS.vendor() call with something like any(contains("mkl"), getfield.(BLAS.get_config().loaded_libs, :libname)) to avoid the deprecation message before this is checked / worked out more carefully?

[devmotion]

It would be good if the deprecation warning could be fixed (possibly as suggested by @carstenbauer) since the deprecation warnings break e.g. the doctests in PythonOT (JuliaOptimalTransport/PythonOT.jl#18).