Targeting System-Level and Kernel-Level Optimizations of Computer Vision Applications on Embedded Systems

Accelerating computer vision applications for Advanced Driver Assistance Systems (ADAS) is not a trivial task. Indeed, ADAS require real-time update rates at high image resolution. The task of accelerating computer vision applications falls into two major axes: system-level optimizations and kernel-level optimizations. The formal accelerate the whole use-case while the latter act at kernel or function granularity. In this paper, we propose an approach to target both system-level and kernel-level optimizations on different hardware architectures. While system-level optimizations cover for instance memory bandwidth loading and inter-processor communication, kernel-level optimizations accelerate single kernels. Our approach consists in merging OpenVX framework and Numerical Template Toolbox (NT ²) Library. OpenVX gives a close attention to system-level optimizations and enables hardware vendors to implement customized accelerated imaging and vision algorithms. NT ² library accelerates kernels on different architectures with a minimal cost of rewriting, based on generative programming model. We test our approach on different computer vision kernels employed in ADAS. We target different acceleration techniques such as vectorization and shared memory parallelization. We perform our experiments on ×86 architecture as well as on Nvidia Tegra X1 ARM cores. We manage to execute OpenVX hardware customized kernels in both architectures with zero rewriting cost thanks to NT ². Also, it is worth noting that we get the same performances on both architectures.