Analysis of GPGPU Programs for Data-race and Barrier Divergence

Santonu Sarkar, Prateek Kandelwal, Soumyadip Bandyopadhyay, Holger Giese
2018 Proceedings of the 13th International Conference on Software Technologies  
Todays business and scientific applications have a high computing demand due to the increasing data size and the demand for responsiveness. Many such applications have a high degree of parallelism and GPGPUs emerge as a fit candidate for the demand. GPGPUs can offer an extremely high degree of data parallelism owing to its architecture that has many computing cores. However, unless the programs written to exploit the architecture are correct, the potential gain in performance cannot be
more » ... In this paper, we focus on the two important properties of the programs written for GPGPUs, namely i) the data-race conditions and ii) the barrier divergence. We present a technique to identify the existence of these properties in a CUDA program using a static property verification method. The proposed approach can be utilized in tandem with normal application development process to help the programmer to remove the bugs that can have an impact on the performance and improve the safety of a CUDA program.
doi:10.5220/0006834904940505 dblp:conf/icsoft/SarkarKBG18 fatcat:5hr4xlpyrvg4tpx7eksp7liqsy