GPU's support massive parallel computing natively. This is a very powerful mechanism, and can achieve large speedups for many (scientific) applications. However, the parallel nature of the programs also introduces the risk of bugs.

 

At the eScienceCenter, Kernel Tuner has been developed as a tool to automatically support many advanced use-cases and optimization strategies that speed up the auto-tuning process, including automatic support to execute tests. However, test cases at the moment are mainly created manually. In this project, we would like to investigate how existing test theory can be used and adapted to the specific use case of GPU kernels. In particular, we are interested in investigating what would be useful coverage criteria to estimate the quality of test suites. Also, it is interesting to consider the generation of test cases based on either code inspection or user-defined properties.