技能 2026-09-28 15:57 UTC
intel-performance-skillslatest
版本【latest】由提交人 fenrus75 提交,主要更新类型为:功能特性新增与扩展。提交说明:performance-patterns: trim redundant generic explanations to cut token cost。为项目引入了新的能力与模块支持。
Commit提交信息原文
performance-patterns: trim redundant generic explanations to cut token cost
Feedback indicated skill content has a high token cost relative to value
because larger models already know a lot of the generic domain/CS
knowledge being spelled out. Audit found this concentrated in
triggers/from-source.md, whose "Pattern descriptions" section (77% of the
file) duplicated generic mechanism explanations already present in each
patterns/*.md file's "Why this is slow" section, and in the prose length
of those "Why this is slow" sections themselves.
- triggers/from-source.md: removed the generic descriptions section,
replaced with a compact "Additional detection notes" block containing
only facts not already in the quick-match table (name-variant lists,
caveats). 1664 -> 580 words.
- triggers/from-profile.md: lightly trimmed 2 sentences that duplicated
pattern-file content; left mostly intact since it is already
signal-specific (perf event names, instruction mnemonics, thresholds)
rather than generic.
- patterns/{ttas,missing-restrict,cv-thundering-herd,missing-vzeroupper,
mutex-to-rwlock,parallel-accumulator,cold-path-annotation,
false-sharing,per-cpu-stats}.md: shortened "why this is slow" sections
to 1 short paragraph, keeping concrete numbers/formulas and cutting
narrative explanation of well-known mechanisms.
linux-perf and phoronix-test-suite SKILL.md files were left untouched --
they are dense with project-specific facts (exact perf event names, PTS
JSON schema, file paths) that models don't already know.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>