kernel-correctness
GitHub用于验证计算内核的正确性，检查数值精度、内存安全及输入变体。通过对比参考实现与中间值，隔离最小复现用例，诊断索引错误、同步缺失或竞态条件等问题。
触发场景
安装
npx skills add uw-syfi/vibesys --skill kernel-correctness -g -y
SKILL.md
Frontmatter
{
"name": "kernel-correctness",
"description": "Check a compute kernel against its task reference, including numerical tolerance, memory safety, and supported input variants. Use for kernel implementation or correctness failures."
}
Kernel correctness
Read the task's candidate contract and accuracy checker to identify the exact supported inputs and comparison rule. Translate the reference operation into invariants for every output element, including reductions, masks, aliasing, strides, and tails where applicable. Keep the evaluator-owned reference and checker unchanged.
When a case fails, isolate the smallest supported shape and input pattern that reproduces it. Compare intermediate values or one output tile against the reference, then distinguish indexing errors, missing synchronization, uninitialized memory, and allowed floating-point variation. Run the full checker after each fix; a passing benchmark shape is insufficient.
For parallel writes, establish which thread owns each address. For reductions, check accumulation dtype, order-sensitive tolerance, and behavior at zero or masked elements. If the contract requires repeatable output, verify repeated runs under the same input and reject race-dependent results.
版本历史
- 256ebe6 当前 2026-10-18 21:29
