Hyeonseok Jung

Today’s embedded systems operate under increasingly dynamic conditions. First, computational workloads can be either variable by nature or adjustable. Moreover, as many devices are battery-powered, it is common to have runtime power management technique, which results in dynamic power budget. This paper presents a design methodology for multi-core systems, based on dataflow specification, that can deal with various contexts. We optimize the original dataflow considering various working conditions, then, autonomously adapt it to a pre-defined optimal form in response to context changes. Such context changes at runtime might cause non-negligible delay or power budget violation. In order to overcome them, an efficient mode switching method is proposed. We show the effectiveness of the proposed technique with a real-life case study, stereo-vision, and synthetic benchmarks.

Today’s embedded systems often operate computer-vision applications, and are associated with timing and power constraints. Since it is not simple to capture the symmetry between the application and the model, the model-based design approach is generally not applicable to the optimization of computer-vision applications. Thus, in this paper, we propose a measurement-based optimization technique for an open-source computer-vision application library, OpenCV, on top of a heterogeneous multicore processor. The proposed technique consists of two sub-systems: the optimization engine running on a separate host PC, and the measurement library running on the target board. The effectiveness of the proposed optimization technique has been verified in the case study of latency-power co-optimization by using two OpenCV applications—canny edge detection and squeezeNet. It has been shown that the proposed technique not only enables broader design space exploration, but also improves optimality.

As recent mobile embedded systems are required to operate many compute-intensive applications, it is important to optimize the applications considering the underlying hardware architecture. In this paper, we propose to optimize ray-tracing application tailored for a given mobile computing platform with multiple heterogeneous processing elements. We take an open-source C/C++ implementation of ray-tracing, adapt it to OpenCL, and perform two optimizations. First, we find out the optimal degree of parallelism, work group size. Then, an optimal workload distribution over two OpenCL devices is examined. We obtain the speedup of 10× over the open-source implementation by using a single GPU device on Exynos 5422. Further, we show that the performance gain can be improved up to 17× by balancing the workload between two GPGPUs in heterogeneous multi-processor system on chips, considering the different compute capabilities thereof.