kernel-ncu-analysis
GitHub用于分析NVIDIA Nsight Compute捕获的CUDA内核性能数据，诊断瓶颈如占用率、内存吞吐和Warp停滞，并提供优化假设与后续实验建议。
触发场景
安装
npx skills add uw-syfi/vibesys --skill kernel-ncu-analysis -g -y
SKILL.md
Frontmatter
{
"name": "kernel-ncu-analysis",
"description": "Interpret NVIDIA Nsight Compute evidence for a CUDA kernel task. Use when an NCU capture is available and a measured bottleneck needs diagnosis."
}
Kernel NCU analysis
Match each captured launch to the benchmarked candidate, input shape, dtype, device, and warmup state. Multiple launches with similar names may represent different phases; inspect launch parameters and source correlation before attributing cost. NCU replay and counter collection can perturb timing, so use the task benchmark for before and after performance decisions.
Form one hypothesis from the relevant counters. Low achieved occupancy can follow register or shared-memory pressure, but low occupancy alone does not prove it limits throughput. Memory throughput and cache hit rates need byte counts and access patterns to be meaningful. Warp stalls identify where warps wait; relate them to the instruction mix, dependency chain, and memory traffic before choosing a change.
Report the capture identity, dominant launches, metric values, uncertainty or missing counters, and the next experiment. Confirm any optimization with the accuracy checker and scored benchmark.
版本历史
- 256ebe6 当前 2026-10-18 21:29
