Adding MKL_jll artifact

[Crown421]

This pull request add dependence on the MKL_jll artifact, and changes all references from the libmkl_vml_avx libraries to the libmkl_rt library.
This has the advantage that MKL chooses the correct instruction set on runtime, making this package and its precompile cache hardware independent.
(Thanks @KristofferC !)

Paradiso.jl currently checks for the existence of a local install of MKL and uses it if one is found. However, given that MKL_jll is now an explicit dependency, it will installed regardless. Given this, it might as well be used.
This will be revisited if it becomes useful to enable user choice for a specific version of MKL via a manual external install.

[Crown421]

This currently fails only on Mac with Julia 1.3, but not with nightly. According to the log IVM actually returns different results.
Will try tomorrow or so on my Mac to investigate.

[KristofferC]

Paradiso.jl currently checks for the existence of a local install of MKL and uses it if one is found.
This will be revisited if it becomes useful to enable user choice for a specific version of MKL via a manual external install.

In Pardiso.jl I thought about removing that check and just tell people to use Override.toml instead (https://julialang.github.io/Pkg.jl/v1/artifacts/#Overriding-artifact-locations-1).

[Crown421]

Reading about this I think I agree on the use of the Override.toml.

[KristofferC]

This currently fails only on Mac with Julia 1.3, but not with nightly. According to the log IVM actually returns different results.

Yeah, something is weird, the result changes a lot based on the length of the input vector:

julia> IntelVectorMath.gamma!(copy(INPUT[1:200]))[1:5]
5-element Array{Float64,1}:
     1.5191049997264277e21
     1.3775142112900882e11
     1.1814039832680319
     1.0624313147244564
 15581.002689052646

julia> IntelVectorMath.gamma!(copy(INPUT[1:800]))[1:5]
5-element Array{Float64,1}:
 Inf
 Inf
  0.9233209669728436
  0.9676118954476614
 Inf

I don't understand why it passes on nightly...

[KristofferC]

See JuliaPackaging/BinaryBuilder.jl#700 for some more info about the test failure.

[Crown421]

I have added a warning the readme, and an __init__ function, that warns if it detects that SpecialFunctions is already loaded.

[KristofferC]

This will not catch the case when someone gets SpecialFunctions loaded as a dependency, because then the name will not be available in Main.

To do it "properly", you need to look at Base.loaded_modules

So it would be something like

special_functions_uuid = Base.UUID("276daf66-3868-5448-9aa4-cd146d93841b")
if Sys.isapple() && haskey(Base.loaded_modules, Base.PkgId(special_functions_uuid, "SpecialFunctions"))
    ...
end

And in fact, the problem (as shown in JuliaPackaging/BinaryBuilder.jl#700) is not really SpecialFunctions but CompilerSupportLibraries_jll so maybe that is the package that should be looked if it is loaded.

[Crown421]

Would you in general say this is reasonable thing to do?

[KristofferC]

The N can be removed here.

[KristofferC]

What is this broadcast business about? That seems odd to me, it doesn't even call VML? Just remove?

[Crown421]

More legacy code, maybe something that was a thing before Julia 0.7? I will have a look

[Crown421]

It appears to have been some construct to maybe allow something like IVM.atan(rand(10,10), rand(100,1))
But guessing at the users intention here seems wrong, so instead now properly throwing a dimension mismatch error.

[aminya]

Tested on Windows 10, Julia 1.5.0-DEV74 with being already MKL.jl installed and works as expected.

[KristofferC]

Nice :)

[aminya]

Thank you for the work!