Concepts for execution time prediction of 3D GPU rendering

The relevance of graphical functions in vehicular applications has increased significantly during the last years. Modern cars are equipped with multiple displays used by different applications such as speedometer, navigation system, or media players. The recent trend towards hardware consolidation to reduce hardware cost, installation space, and energy consumption, causes graphical 3D applications of different safety-criticality to share a single GPU. This requires effective real-time GPU scheduling concepts to ensure safety and isolation for 3D rendering. Since current GPUs are not preemptible, a deadline-based scheduler must know the GPU execution time of GPU commands in advance. In this work, we present a novel framework to measure and predict the execution time of GPU commands using OpenGL ES 2.0. We present prediction models for the main GPU commands relevant for 3D rendering, namely, FLUSH, CLEAR, and DRAW. For the DRAW command we propose to use the 3D bounding box of the rendered model and apply the vertex shader projection to heuristically estimate the number of fragments rendered. We finally present the implementation and evaluation of our framework, which demonstrates its feasibility and shows that good prediction accuracy can be achieved. In our evaluation using realistic scenarios the absolute prediction error of the DRAW command did not exceed 260 μs.