hi, i'm matthew

Using Numpy with Intel MKL on macOS

April 23, 2020 #python#numpy#blas

I’ve been working on a project that requires quite a few matrix operations using numpy, including dotting a 24000x2 and a 2x4096 (or larger) ndarray of floats. Clearly, optimization would be beneficial. There are several places in the code that I could cut down on multiplications or unnecessary regeneration of arrays using meshgrid and arange when I could do what I needed in fewer steps, but my through my online searches, I stumbled across the terms BLAS and LAPACK. TL;DR, these are very low level math libraries that can greatly improve matrix multiplication performance. The three major implementations are Intel’s MKL, Apple’s Accelerate, and OpenBLAS.

If you install numpy through conda, you’ll notice that once of the packages installed has openblas in the name. If I opened up a terminal and ran np.show_config(), I ended up with this output:

blas_mkl_info:
    libraries = ['blas'undefined 'cblas'undefined 'lapack'undefined 'pthread'undefined 'blas'undefined 'cblas'undefined 'lapack']
    library_dirs = ['/Users/matthew/miniconda3/lib']
    define_macros = [('SCIPY_MKL_H'undefined None)undefined ('HAVE_CBLAS'undefined None)]
    include_dirs = ['/Users/matthew/miniconda3/include']
blas_opt_info:
    libraries = ['blas'undefined 'cblas'undefined 'lapack'undefined 'pthread'undefined 'blas'undefined 'cblas'undefined 'lapack'undefined 'blas'undefined 'cblas'undefined 'lapack']
    library_dirs = ['/Users/matthew/miniconda3/lib']
    define_macros = [('SCIPY_MKL_H'undefined None)undefined ('HAVE_CBLAS'undefined None)]
    include_dirs = ['/Users/matthew/miniconda3/include']
lapack_mkl_info:
    libraries = ['blas'undefined 'cblas'undefined 'lapack'undefined 'pthread'undefined 'blas'undefined 'cblas'undefined 'lapack']
    library_dirs = ['/Users/matthew/miniconda3/lib']
    define_macros = [('SCIPY_MKL_H'undefined None)undefined ('HAVE_CBLAS'undefined None)]
    include_dirs = ['/Users/matthew/miniconda3/include']
lapack_opt_info:
    libraries = ['blas'undefined 'cblas'undefined 'lapack'undefined 'pthread'undefined 'blas'undefined 'cblas'undefined 'lapack'undefined 'blas'undefined 'cblas'undefined 'lapack']
    library_dirs = ['/Users/matthew/miniconda3/lib']
    define_macros = [('SCIPY_MKL_H'undefined None)undefined ('HAVE_CBLAS'undefined None)]
    include_dirs = ['/Users/matthew/miniconda3/include']

A bit cryptic, but it appeared that numpy was using MKL. I mean, it says bias_mkl_info, right?

Not so fast. Thanks to this StackOverflow answer, we need to check which library is actually being linked with numpy. According to the show_config() output, the BLAS library is linked in /Users/matthew/miniconda3/lib. I needed to run this bash command:

$ otool -L /Users/matthew/miniconda3/lib/libblas.3.dylib 

Alternatively, if the output is not clear, you can use

$ readlink /Users/matthew/miniconda3/lib/libblas.3.dylib 

Unfortunately, the output showed a link to openblas, not MKL, even though MKL is automatically included in conda. How interesting.

The solution, fortunately, is simple. Uninstall numpy, then reinstall it while only installing any BLAS packages that match mkl, avoiding any OpenBLAS linkage that might override the default MKL libs.

$ conda remove numpy
$ conda install numpy libblas=*=*mkl

Kudos to this Github issue for mentioning this solution.

Removing numpy also removes anything that depended on numpy, so you’ll need to also reinstall matplotlib, scipy, astropy, etc.

Now use the readlink command we see

$ readlink /Users/matthew/miniconda3/lib/libblas.3.dylib 
libmkl_rt.dylib

The output says libmkl_rt.dylib, so it looks like numpy will use MKL now!


Matthew Russell

Written by Matthew Russell who lives in Lexington, KY, studies machine learning at the University of Kentucky, and interns at NASA. Get to know him or check out his projects on GitHub.