Graph-based Shader Code Analysis to Assist in GPU's Performance Improvement [Demo]

A pipeline regarding how to analyze a new game to efficently improve GPU performance from shader perspective

This is a joint work between Lin Zhao (NTU Singapore), Arijit Khan (AAU Denmark) and Robby Luo (AMD)

This paper is accepted by "Joint Workshop on Graph Data Management Experiences & Systems (GRADES) and Network Data Analytics (NDA) (GRADES-NDA 2022)

Link to the demonstration video[Video]



Analyze a video game's basic information


Given a new game, identify its group

Identify which existing game group the new game belongs to

If it can be grouped with existing games, use analysis of existing game group

If not, identify its low efficiency frequent sub-patterns


Given a new game, identify its low efficiency shaders

Predict scene of an unknow frame

Given an unknow frame number shader codes, check which scene it belongs to

If it belongs to known scenes, use representative frame data from that group to tune GPU performance

If it doesnt belong to any scene, treat it as a separate scene to tune GPU performance

ShaderNet's Architecture