KernelGenBench

Compile, validate, and benchmark a CUDA, CUTLASS, or Triton operator against an optional Python reference by driving `skills/optimized-skill/kernel-benchmark/scripts/benchmark.py`. Use when Claude needs to 验证算子正确性、做 baseline 性能测试、比较 operator 与 reference 的 speedup、对 CUDA/CUTLASS 读取 `extern "C" void solve(...)` 签名推断参数，或在进入后续分析前先得到稳定 benchmark 结果。

Profile a CUDA, CUTLASS, or Triton operator with Nsight Compute or analyze an existing `.ncu-rep` report to diagnose bottlenecks and produce actionable optimization guidance. Use when Claude needs to 解释 NCU 指标、定位 kernel 为什么慢、生成 fresh `.ncu-rep`、判断 memory/latency/compute/occupancy 瓶颈，或把报告结论整理成下一轮算子优化可直接使用的建议。

Run an operator optimization loop for CUDA, CUTLASS, or Triton implementations. The loop enforces correctness validation, benchmark, backend-aware profiling/artifact capture, strategy memory (positive/negative/rejected), and final best-version selection. Use when Claude needs to iterate on a `.cu` or Triton `.py` operator for a user-specified number of rounds, compare versions by benchmark results, keep full Nsight Compute evidence for winners, and prepare the next optimization step from the latest reports.