As embedded systems, such as smartphones with limited resources, have become increasingly popular, active research has recently been conducted on performing on-device deep learning in such systems. Therefore, in this study, we propose a deep learning framework that is specialized for embedded systems with limited resources, the operation processing structure of which differs from that of standard PCs. The proposed framework supports an OpenCL-based accelerator engine for accelerator deep learning operations in various embedded systems. Moreover, the parallel processing performance of OpenCL is maximized through an OpenCL kernel that is optimized for embedded GPUs, and the structural characteristics of embedded systems, such as unified memory. Furthermore, an on-device optimizer for optimizing the performance in on-device environments, and model converters for compatibility with conventional frameworks, are provided. The results of a performance evaluation show that the proposed on-device framework outperformed conventional methods.
Design and Performance Evaluation of Configurable Component Middleware for End-to-End Adaptation of Distributed Real-Time Embedded Systems
