Optimising Convolutional Neural Networks Inference on Low-Powered GPUs
School of Informatics, University of Edinburgh, UK
12th International Workshop on Programmability and Architectures for Heterogeneous Multicores (MULTIPROG), 2019
In this paper we present effective optimisation techniques for accelerating convolutional neural networks inference on low-powered heterogeneous devices with OpenCL. Using LeNet and VGG-16 as test networks, we implement a custom neural network system in OpenCL and optimise it to minimise their inference times. Our baseline system shows a speedup of 17x for LeNet. We also outline two methods for fast convolution: an iterative vectorised approach and a Morton GEMM based approach. The two approaches demonstrate VGG-16 inference speeds up to 3x faster than current state-of-the-art systems and outperform other custom neural network systems by speedup factors of up to 1.82x.
February 10, 2019 by hgpu
