An Investigation on Numerical Bugs in GPU Programs Towards Automated Bug Detection

Ravishka Rathnasuriya, Nidhi Majoju, Zihe Song, Wei Yang

Summary

This work studies 397 real-world numerical bugs in GPU programs, characterizing their causes, symptoms, triggering inputs, test oracles, and fixes. It introduces GPU-NBDetect, which targets six bug categories and detects 226 bugs across 186 mathematical functions in four libraries; 60 were confirmed by developers.