Resource-aware Computer Vision Application on Heterogeneous Multi-tile Architecture

I. DESCRIPTION The ever growing demand for autonomous systems has pushed algorithms to new levels of flexibility, which assures a resourceaware implementation on fast reconfigurable architectures capable to adapt at run-time according to changing requirements and constraints. Here, of great importance are computer vision algorithms used to obtain important information about the spatial position of objects in a scene. However, the efficient exploration of their capability to adapt to different environment conditions in real-time is still a challenge. Therefore, in order to provide applications more flexibility, a self-organizing computing paradigm called invasive computing can be used [1]. In invasive computing, applications running on an MPSoC and competing for resources have the ability to explore and dynamically spread their computations to neighborhood processors. Here, an application may dynamically claim and reserve resources (invade), employ them for parallel execution (infect), and finally release them (retreat). Our demonstration will present the benefits of invasive computing by showing the efficiency and utilization improvements in a resource-aware manner by algorithmic selection of different invasive resources, such as TCPA (tightly-coupled processor array), and RISC processors. More specific we present a dynamic load balancing between multiple RISC cores and a TCPA, based on invasive mechanisms supported by our operating system and the agent system. By exploiting information from the invasive run-time system we consider two variations of the well-known Harris Corner Detector both optimized for RISC [2] and TCPA architectures [3].