AKO4ALL — Agentic Kernel Optimization

What it does

AKO4ALL runs an agentic optimize→benchmark→log→commit workflow for GPU kernels. Given a kernel (CUDA, Triton, TileLang, C++, Python), it bootstraps a workspace, profiles the code, runs iterative micro-optimizations, verifies correctness, and records each iteration. The goal is measurable runtime improvement over a provided reference while preserving correctness and reproducible commits.

When to use it

Use AKO4ALL when you have an existing GPU kernel you want to speed up and measure: 'optimize this kernel', 'run AKO on my CUDA kernel', or when you need repeatable iteration with profiling (ncu) and git-backed experiment logging. It is not for writing a kernel from scratch or for generic non-kernel performance advice.

What's included

• Scripts: scaffolded bench-wrapper and scripts/bench.sh patterns (bench/kernelbench integration) when present in the repo; the skill provides guidance to render and run bench commands.
• References: guidance to use ncu profiling and KernelBench-style inputs; the skill explains HINTS.md and ITERATIONS.md usage for persistent directives and experiment logs.
• Instructions: detailed procedural workflow covering workspace inventory, branch/solution initialization, bench command generation, iteration protocol (bench → ITERATIONS.md entry → commit), profiling guidance, stall handling and finalization steps.

Compatible agents

Best used by agents with code execution and shell capabilities (Copilot/Codex-style or CLI-capable agents that can run Python, shell, and profiling tools). The skill expects the environment to run benchmarks, invoke ncu, and commit to git.