Thesis: Heuristics for Efficient Dynamic Verification of MPI and Thread Programs.

For my current research, I am collecting different classes of CUDA kernels and comparing the numerical precision loss between different kernels (with different underlying algorithms) in the same class. Then I will empirically conclude a programming guideline for developing precise CUDA programs. To achieve this study, I am also developing a numerical precision profiling framework for CUDA. Our ultimate goal of this project is providing a guideline and a tool that help programmers find a balance between performance and precision in their large CUDA programs such as CUDA libraries