[SOLVED] Possibility to use Python 3.6 with Intel MKL 2017 and a Xeon Phi KNC Card

Possibility to use Python 3.6 with Intel MKL 2017 and a Xeon Phi KNC Card

I am experimenting with an Intel Xeon Phi 3120A card and automatic offloading using Python. I got it running using Intel Python 2017 with the help of this post. By that I found out that the card unfortunately is only supported by the 2017 version of MKL for automatic offloading. As Intel Python 2017 is based on Python 3.5 I am wondering if there is some way to use, e.g., Python 3.6 with MKL 2017 to preserve the automatic offloading functionality of the Xeon Phi card. Can I somehow set up a Miniconda environment with Python 3.6 still using MKL 2017? Or is it possible to compile numpy and scipy with the old MKL library?

Thanks!

Solution

You need to build scipy and numpy from source while linking it against MKL BLAS and MKL LAPACK. Use this to get you started: https://docs.scipy.org/doc/numpy/user/building.html