[PYTHON] How to install mkl numpy

The Math Kernel Library (mkl), a high-performance matrix library made by Intel, is now available for free. I will explain how to make this available from numpy. The target is Linux (I am Ubuntu 14.04).

MKL installation

Click "click here now to register and download" on the following site. https://software.intel.com/en-us/articles/free_mkl Enter your name, email address, company name, etc. You will receive a registration email, which you can follow to download. There is no Mac version. It's about 1GB. Make a note of the registration key that you will need during installation. I downloaded l_mkl_11.3.0.109.tgz

Extract the downloaded file. There is `ʻinstall.shin it, so execute it. The installation destination is / opt / intel`` by default, and you will be prompted to enter the registration key on the way, so enter it. (

If you don't want to use interactive, modify the following part of the silent.cfg file in the unzipped directory.

ACCEPT_EULA=accept
ACTIVATION_SERIAL_NUMBER=XXXX-XXXXXXXX
ACTIVATION_TYPE=serial_number

You can install it from the command line only by running ʻinstall.sh -s silent.cfg`.

Put LD_LIBRARY_PATH in / opt / intel / mkl / lib / intel64.

numpy installation

For the settings when installing numpy, go to ~ / .numpy-site.cfg. The template for this file is in the source. https://github.com/numpy/numpy/blob/master/site.cfg.example Modify the file and keep it in your home directory. The contents are as follows.

[mkl]
library_dirs = /opt/intel/mkl/lib/intel64/
include_dirs = /opt/intel/mkl/include
mkl_libs = mkl_rt
lapack_libs =

For details, refer to the following site. https://software.intel.com/en-us/articles/numpyscipy-with-intel-mkl

In this state, pip install --no-binary: all: numpy will install numpy using mkl without permission. If numpy is originally included, uninstall it once.

Operation check

You can see which matrix library you are using by looking at numpy.show_config (). If you run Chainer as an example, it seems that it is running at about 1.5 times faster than OpenBLAS.

$ python
>>> import numpy
>>> numpy.show_config()
lapack_opt_info:
    libraries = ['mkl_rt', 'pthread']
    library_dirs = ['/opt/intel/mkl/lib/intel64/']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/mkl/include']
blas_opt_info:
    libraries = ['mkl_rt', 'pthread']
    library_dirs = ['/opt/intel/mkl/lib/intel64/']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/mkl/include']
openblas_lapack_info:
  NOT AVAILABLE
lapack_mkl_info:
    libraries = ['mkl_rt', 'pthread']
    library_dirs = ['/opt/intel/mkl/lib/intel64/']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/mkl/include']
blas_mkl_info:
    libraries = ['mkl_rt', 'pthread']
    library_dirs = ['/opt/intel/mkl/lib/intel64/']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/mkl/include']
mkl_info:
    libraries = ['mkl_rt', 'pthread']
    library_dirs = ['/opt/intel/mkl/lib/intel64/']
    define_macros = [('SCIPY_MKL_H', None)]
    include_dirs = ['/opt/intel/mkl/include']