智慧型行動裝置OpenCL運算移轉框架

Abstract

Nowadays, mobile devices are suffering from limited computational resource. To increase capabilities of mobile devices, many efforts have been made to offload computation from mobile devices to resourceful servers. However, most of the approaches are only capable of offloading computation to CPUs. Recently, GPUs have received a lot of attention from the scientific community. Indeed, the exceptional parallel computing capabilities of GPUs can be used to accelerate different types of applications. In this thesis, we propose a computation offloading framework based on OpenCL – a standard for GPU computing. Our framework transparently offloads OpenCL workloads from mobile devices to an available OpenCL compatible device. We deployed our framework on real machines and conducted evaluation experiments using various OpenCL programs including basic matrix computations, an image processing Android app and benchmarks from AMD. The program achieves up to 50.3X speedup by remote offloading compared to the local execution using CPU.