high performance computing on graphics processing units: hgpu.org

The increasing GPU performance and maturing computational platform make it possible to handle general-purpose computing jobs traditionally computed by the CPU. Also, just like what we did in the CPU program, we use testing to verify the correctness of the GPU program. However, the quality of the tests may remain unknown, which inspires us to use mutation testing, a fault-based testing technique, to measure the effectiveness of a test. In this research, we conduct a feasibility study on applying testing and mutation testing to GPU programming, adapting existing research methodology, hypotheses, experiments, and optimization methods to the specific use case of the GPU kernel, and exploring the GPU-native mutation testing theories and techniques. A mutation testing tool is developed to validate and evaluate the theoretical analysis above, as a module of a kernel-tuning tool to offer an out-of-the-box mutation testing workflow for kernel developers. Our results reveal that testing can benefit the quality of kernels, and more test cases will lead to a higher possibility of detecting faults. We also discovered that mutation testing is able to quantify the testing quality for GPU kernels.