Increasing Testing Robustness of GPU Software in Embedded Critical Systems

Critical autonomous systems build on massively parallel software (e.g., AI software) requiring accelerators like GPUs. However, code coverage requirements imposed by safety standards intended for sequential code are challenged by parallel GPU execution. This paper illustrates how parallel execution challenges test coverage and proposes an approach to explicitly address the new dimensions of the test space so that coverage results on GPUs have comparable meaning to those of conventional tools on sequential devices.1