specialized-path
How to discover, validate, build, and honestly verify a SPECIALIZED path for ANY function — a cheaper route for a recognizable SUBSET of an operation's inputs, sitting behind the general implementation with a gate + fallback so the general case never regresses and semantics stay bit-identical. The cheaper route can be a different KERNEL (SIMD vs scalar, gemv/syrk/stencil), a different KERNEL CONTRACT (one call per block instead of per row), a REPRESENTATION trick (bit-level f16, byte-lane bool, widen-compute-narrow), a different ALGORITHM (hash vs sort, with an observed-growth bailout), a VIEW instead of a copy (O(1) stride tricks), an EARLY EXIT (all-false / empty / k==0 / provably-constant output), a FIXED-COST BYPASS (skip iterator/broadcast setup for the trivial subset), a MEMORY strategy (cpblk, prefetch, write-once, privatized accumulators, retained scratch), a FUSED pass, a composition of EXISTING validated kernels, wrapper-glue removal on a parity-bound op, or a native BACKEND seam. Use this whenever
ソース情報
- リポジトリ
- SciSharp/NumSharp
- ソースの最終更新活動
- 2026年10月31日 20:47
- 検出された SKILL.md の言語
- 英語
- スター
- 1,483
- フォーク
- 206
インストール方法
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
ソースファイルを確認
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
SKILL.md を表示中
| name | specialized-path |
|---|---|
| description | How to discover, validate, build, and honestly verify a SPECIALIZED path for ANY function — a cheaper route for a recognizable SUBSET of an operation's inputs, sitting behind the general implementation with a gate + fallback so the general case never regresses and semantics stay bit-identical. The cheaper route can be a different KERNEL (SIMD vs scalar, gemv/syrk/stencil), a different KERNEL CONTRACT (one call per block instead of per row), a REPRESENTATION trick (bit-level f16, byte-lane bool, widen-compute-narrow), a different ALGORITHM (hash vs sort, with an observed-growth bailout), a VIEW instead of a copy (O(1) stride tricks), an EARLY EXIT (all-false / empty / k==0 / provably-constant output), a FIXED-COST BYPASS (skip iterator/broadcast setup for the trivial subset), a MEMORY strategy (cpblk, prefetch, write-once, privatized accumulators, retained scratch), a FUSED pass, a composition of EXISTING validated kernels, wrapper-glue removal on a parity-bound op, or a native BACKEND seam. Use this whenever |