Restoring get_num_threads function of BLAS

[usefulhyun]

BLAS.get_num_threads function have been deleted by this commit 1da2c9b.

But, I can not find why, I think, I and other people also need this function by #15008.
And It already implemented by @tkelman before.

[martinholters]

Note this:

julia/test/distributed_exec.jl

Lines 1172 to 1191 in e05a921

	const get_num_threads = function() # anonymous so it will be serialized when called
	blas = BLAS.vendor()
	# Wrap in a try to catch unsupported blas versions
	try
	if blas == :openblas
	return ccall((:openblas_get_num_threads, Base.libblas_name), Cint, ())
	elseif blas == :openblas64
	return ccall((:openblas_get_num_threads64_, Base.libblas_name), Cint, ())
	elseif blas == :mkl
	return ccall((:MKL_Get_Max_Num_Threads, Base.libblas_name), Cint, ())
	end
	# OSX BLAS looks at an environment variable
	if is_apple()
	return ENV["VECLIB_MAXIMUM_THREADS"]
	end
	end
	return nothing
	end

[tkelman]

IIRC the issue with this in openblas is that it returns the build time hard coded max number of threads, which is much less useful or expected than the run time dynamic number of threads setting.

[usefulhyun]

@tkelman , because in some case it returns the hard coded max number of threads, it have been deleted, is this right? Can I know anther option to replace this function?

[tkelman]

It would probably be more useful to look at the same environment variables that blas implentations do

[antoine-levitt]

@tkelman IIRC the issue with this in openblas is that it returns the build time hard coded max number of threads, which is much less useful or expected than the run time dynamic number of threads setting.

Has this not been fixed by OpenMathLib/OpenBLAS#762? At least here (0.6 and 0.7, linux openblas) it seems to work as expected. This PR should be re-opened as it seems like a useful functionality to have (if only for symmetry wrt set_num_threads)

[Jutho]

I am also in favour of reopening this. I think the MKL call needs to be
mkl_get_max_threads instead of MKL_Get_Max_Num_Threads.

One important use case I see, as multithreading in Julia is now more easy to use with @spawn, is that a package developer writes some code that implements an algorithm that uses multithreading at a higher level, and want to use single threaded matrix multiplication in the specific block of code ran on the different threads. So (s)he could do BLAS.set_num_threads(1) before the block of his own multithreaded code starts, but to restore BLAS.set_num_threads(originalvalue), (s)he needs to be able to find out what the originalvalue of number of BLAS threads is.