Improving the performance of ADAS application in heterogeneous context: A case of lane detection

Abstract

This paper investigates the optimization of OpenCLbased ADAS applications in heterogeneous context. In particular, we take the widely-used lane detection algorithm (LDA) as a case study. The application is profiled to identify the performance bottlenecks and then three optimization strategies are adopted. On the kernel side, the parallel granularity is regulated via compute unit replication and loop unrolling. On the host side, the kernel API function calls are scheduled in an interleaved manner to overlap the accelerator execution time. Moreover, the computation workload of the algorithm is tuned by dynamically adjusting the processed image ROI size. Experimental results reveal that the optimized implementation can achieve an average 2.27x speedup when compared with the naive parallel application.

Performance Optimisation of Parallelized ADAS Applications in FPGA-GPU Heterogeneous Systems: A Case Study With Lane Detection

“…During the runtime evaluation of p-LDA, we use 2 12 good lines and 2 13 candidate lines to detect 2 lane markings. As for the parameters of r-LDA, we set the initial value of the number of RANSAC iterations to 300 and the observed optimal value is 40.…”

“…This article presents a substantial extended work of our previous study in [12]. We also provide a customized heterogeneous design of the work in [13] and thereupon apply the proposed procedure to obtain the optimal execution.…”