kernel-performance
GitHub用于优化计算内核性能，通过基准测试定位瓶颈，指导启动几何、分块、内存布局等调整。需验证精度并对比指标。
触发场景
安装
npx skills add uw-syfi/vibesys --skill kernel-performance -g -y
SKILL.md
Frontmatter
{
"name": "kernel-performance",
"description": "Optimize a compute kernel using its scored benchmark and measured bottlenecks. Use when choosing launch geometry, tiling, memory layout, fusion, or precision changes."
}
Kernel performance
Start with a correct baseline and the task's scored metric. Record the device, shape, dtype, warmup, and timing method for each comparison. Separate compile, setup, launch, and device execution costs when the benchmark permits it; only the scored path decides whether a change helps.
Build a performance hypothesis from evidence: bytes moved, arithmetic work, parallelism, and measured launch count. Change one variable at a time, such as tile shape, warp count, layout, fusion boundary, or memory staging. Re-run the accuracy checker before accepting a faster result. Compare medians or another task-approved robust statistic over repeated runs, accounting for measurement noise near the claimed improvement.
Use a profiler when timing alone cannot locate the cost. A counter suggests a cause only when the captured input and launch match the scored workload; validate the proposed fix with the original benchmark.
版本历史
- 3a50e8b 当前 2026-11-02 13:56
