| Summary: | Heterogeneity has become the cornerstone of modern embedded System-on-Chips (SoCs), used in latest smart-phones and edge computing platforms to achieve high performance under tight power budgets. Alongside the general purpose multi-core CPUs, such SoCs typically integrate several specialized processing elements such as GPU and DSP to ensure power-efficient execution of specific workloads. In the past, many computer vision algorithms and their kernels have been shown to benefit from execution on GPUs, both in terms of performance and power consumption. Existing work has also demonstrated the benefit of accelerating them simultaneously on multi-core CPUs and discrete as well as integrated GPUs found in PCs and workstations. Recently, authors have also focused on accelerating such applications on heterogeneous embedded SoCs with integrated CPU and GPU. In this paper, we first present an extensive literature review of such efforts and highlight their strengths and limitations. Next, we use the latest state-of-the-art edge computing platform, Odroid-N2, and the older Odroid-XU3 platform, both of which use heterogeneous embedded SoCs, to explore the acceleration of the convolution kernel with different filter sizes and its impact on the KLT tracking algorithm. Lastly, we discuss the challenges and opportunities in leveraging such SoCs for computer vision and other AI applications.
|
|---|