OpenDNN: An Open-source, cuDNN-like Deep Learning Primitive Library
Graduate School of Seoul National University
Seoul National University, 2019
@phdthesis{kim2019opendnn,
title={OpenDNN: An Open-source, cuDNN-like Deep Learning Primitive Library},
author={Daeyeon Kim},
year={2019},
school={Graduate School of Seoul National University}
}
Deep neural networks (DNNs) are a key enabler of today’s intelligent applications and services. cuDNN is the de-facto standard library of deep learning primitives, which makes it easy to develop sophisticated DNN models. However, cuDNN is a propriatary software from NVIDIA, and thus does not allow the user to customize it based on her needs. Furthermore, it only targets NVIDIA GPUs and cannot support other hardware devices such as manycore CPUs and FPGAs. In this thesis we propose OpenDNN, an open-source, cuDNN-like DNN primitive library that can flexibly support multiple hardware devices. In particular, we demonstrate the portability and flexibility of OpenDNN by porting it to multiple popular DNN frameworks and hardware devices, including GPUs, CPUs, and FPGAs.
May 19, 2019 by hgpu
