kernel-performance
GitHub用于优化计算内核性能，指导选择启动几何、分块、内存布局、融合或精度调整。通过基准测试识别瓶颈，假设验证变更，确保准确性与性能提升。
Trigger Scenarios
Install
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.
Version History
- 256ebe6 Current 2026-10-18 21:29
