Modern and future automotive systems incorporate several Advanced Driving Assistance Systems (ADAS). Those systems require significant performance that cannot be provided with traditional automotive processors and programming models. Multicore CPUs and Nvidia GPUs using CUDA are currently considered by both automotive industry and research community to provide the necessary computational power. However, despite several recent published works in this domain, there is an absolute lack of open implementations of GPU-based ADAS software, that can be used for benchmarking candidate platforms. In this work, we present a multi-CPU and GPU implementation of an open implementation of a pedestrian detection benchmark based on the Viola-Jones image recognition algorithm. We present our optimization strategies and evaluate our implementation on a multiprocessor system featuring multiple GPUs, showing an overall 88.5× speedup over the sequential version.

This work has been supported by the Spanish Ministry of Science and Innovation under grant TIN2015-65316P, the HiPEAC Network of Excellence and a Microsoft sponsored ACM SRC. The first two authors acknowledge Dr. Petrisor for her assistance in understanding and using the sequential version of the benchmark and dedicate this article to the memory of the late beloved advisor prof. Nacho Navarro, without whom this work would not have been possible.

Peer Reviewed