GPU Wavefront Splitting for Safety-Critical Systems

Abstract

Graphics processing units (GPUs) are compute platforms that are ideal for highly parallel workloads due to their high degree of hardware parallelism. Parallelism offered by GPUs lends itself well to machine learning and computer vision applications, including in safety-critical systems. Safety-critical systems require a guarantee of timing predictability. Guaranteeing timing predictability means being able to statically analyze the worst-case execution time (WCET) of the GPU program. Unfortunately, existing GPUs are designed for average-case performance and are thus not designed for timing predictability. Consequently, there is potential for research effort to provide these guarantees.

Prior research works have proposed several new techniques to improve performance. One such technique is wavefront splitting, which reduces the number of idle threads on the GPU and increase utilization. However, no prior work addresses the WCET of this technique. The purpose of this thesis is to develop a GPU implementation for safety-critical systems that leverages wavefront splitting and to enable analysis of the WCET in such an implementation.