All topics
#Kernel Correctness
Four entries on this site carry the Kernel Correctness tag: three papers and one post, dated 2026. Explore the full list of related work below.
Every entry here also appears under GPU Kernels, alongside one more, so that is the page to link to.
Papers
-
Operator-Aware Mixed-Precision Tolerance Calibration for Tensor Kernels
Paper · 2026 · arXiv · cited by 1 Derives per-operator, per-dtype tolerances from 8,076 test runs in the gpuemu corpus; calibrated tolerances are far stricter than hand-picked ones and raise bug-detection recall from 73.2% to 82.4%. -
Static PTX Metrics Track Structural Kernel Regressions but Miss Semantic Ones
Paper · 2026 · arXiv · cited by 0 Pairs static PTX metrics with measured runtime on five GPU classes: structural kernel bugs are visible in the static signal, semantic bugs that swap a constant compile to identical PTX. -
The Correctness Illusion in LLM-Generated GPU Kernels
Paper · 2026 · arXiv · cited by 6 Fixed-shape allclose checks used by LLM-kernel benchmarks certify buggy kernels as correct. Under op-schema-aware seeded fuzzing with an fp64 reference and per-(op, dtype) tolerances, a 26-op corpus shows 10 of 10 seeded illusions caught and 16 of 16 controls.
Posts
-
Why GPU Kernel Benchmarks Miss Real Bugs
Post · 2026 Passing a correctness benchmark and being correct in production are different claims for LLM-generated GPU kernels — the gap comes down to which inputs get sampled.