Jul 2026
Correct but Slow: An Empirical Study of the GPU Kernel Evaluation Gap in Modern Domain-Specific Languages
This study studies whether correctness-based evaluation identifies kernels unsuitable as library replacements, why such failures occur, and how they can be detected without exhaustive benchmark coverage.
Tingxi Li, Ravishka Rathnasuriya, Wei Yang
· arXiv.org · 0 citations