Graph Mining and Machine Learning for Shader Codes Analysis to Accelerate GPU Tuning

Abstract

The graphics processing unit (GPU) has become one of the most important computing technologies. Disassembly shader codes, which are machine-level codes, are important for GPU designers (e.g., AMD, Intel, NVIDIA) to tune the hardware, including customization of clock speeds and voltages. Due to many use-cases of modern GPUs, engineers generally find it difficult to manually inspect a large number of shader codes emerging from these applications. To this end, we develop a framework that converts shader codes into graphs, and employs sophisticated graph mining and machine learning techniques over a number of applications to simplify shader graphs analysis in an effective and explainable manner, aiming at accelerating the whole debugging process and improving the overall hardware performance. We study shader codes’ evolution via temporal graph analysis and structure mining with frequent subgraphs. Using them as the underlying tools, we conduct a frame’s scene detection and representative frames selection. We group the scenes (applications) to identify the representative scenes, and predict a new application’s inefficient shaders. We empirically demonstrate the effectiveness of our solution and discuss future directions.