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

DOI10.1145/3728950MaRDI QIDQ7414905arXivFDO

Authors Ravishka Rathnasuriya, Nidhi Majoju, Zihe Song, Wei Yang

Publication date 3 November 2026

Published in arXiv

Tldr This paper investigates numerical bugs in general-purpose graphics processing unit (GPU) programs through a comprehensive analysis of real-world bug samples, identifies common root causes, symptoms, and fix strategies, and presents a preliminary detection tool called GPU-NBDetect that detects 226 bugs across four libraries, with 60 confirmed by developers.

Full work available at URL https://doi.org/10.1145/3728950

arXiv classification

cs.DCcs.PLcs.SE

Cites work

• Automatic detection of floating-point exceptions
• Handbook of floating-point arithmetic

This page was built for publication: An Investigation on Numerical Bugs in GPU Programs Towards Automated Bug Detection

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7414905)