Chainer-XP: A Flexible Framework for ANNs Run on the Intel ® Xeon Phi TM Coprocessor

Chainer is a well-known deep learning framework facilitating the quick and efficient establishment of Artificial Neural Networks. Chainer can be deployed on systems consisting of Central Processing Units and Graphics Processing Units efficiently. In addition, it is possible to run Chainer on systems containing Intel Xeon Phi coprocessors. Nonetheless, Chainer can only be deployed on Intel Xeon Phi Knights Landing, not Knights Corner. There are many existing systems, such as Tiane2 (MilkyWay-2), Thunder, Cascade, SuperMUC, and so on, including Knights Corner only. For that reason, Chainer cannot fully exploit the computing power of such systems, which leads to the demand for supporting Chainer run on them. It becomes more challenging in the situation where deep learning applications are written in Python while the Xeon Phi processor is only capable of interpreting C/C\(++\) or Fortran. Fortunately, there is an offloading module called pyMIC which helps port Python applications into the Intel Xeon Phi Knights Corner coprocessor. In this paper, we present Chainer-XP as a deep learning framework assisting applications to run on the systems containing the Intel Xeon Phi Knights Corner coprocessor. Chainer-XP is an extension of Chainer by integrating pyMIC into Chainer. The experimental findings show that Chainer-XP can help to move the core computation (matrix multiplication) to the Intel Xeon Phi Knights Corner coprocessor with acceptable performance in comparison with Chainer.