Concepts and metrics for measurement and prediction of the execution time of GPU rendering commands

Abstract

Graphic Processing Units (GPUs), wtih their highly parallel structure, is a powerful tool to manipulate computer graphics and process large amounts of data in parallel. GPUs are becoming increasingly popular in today's embedded system world. One typical example is the modern vehicle. Nowadays, modern vehicles are often required to run graphic applications such as instrumentation cluster, navigation system, media player or games with higher graphic quality concurrently on multiple displays. Compared to the CPU, GPU is more suitable to fulfil these requirements. In order to reduce the hardware cost and the power consumption at the same time, a good solution is to share a single GPU for multiple applications. Thus emerges the need of GPU real time scheduling. Since current GPUs do not support preemption, accurate prediction of the GPU command execution time becomes an indispensable precondition for scheduling. In this work, the driver model and hardware features of a typical 3D embedded GPU is analysed. Concepts to measure GPU execution time and other statistics is present and implemented. The measured execution time and the relevant statistics of the GPU command group allow us to explore the factors and metrics that affect the execution time of GPU rendering commands, more specifically, OpenGL ES 2.0 commands/APIs. Different models are developed to predict the important time-consuming rendering commands: Flush, Swap, Clear, Draw and Texture Loading. As for the Draw command, a recent-history-based approach is proposed to estimate the number of fragments. Finally, the accuracy of the measurement and prediction is evaluated.