cutile-autotuning

What problem does it solve?

This Skill prevents slow or incorrect performance tuning by giving a structured, architecture-aware workflow for adding CuTile autotuning to kernels.

Core Features & Use Cases

• Tune-once/cache/launch pattern: Uses exhaustive_search to find the best config once, then reuses a cached tuned kernel for fast repeated launches.
• Search space design for CuTile: Builds a small, precise config set (≤ 30) using occupancy-only tuning for memory/bandwidth-bound kernels or full tile searches for compute-bound kernels.
• Safety for in-place kernels: Applies the split-buffer strategy during exhaustive_search to avoid data corruption across trial runs.
• DISABLE_AUTOTUNE fallback: Supports CI and profiling determinism by bypassing tuning when DISABLE_AUTOTUNE=1.
• Common pitfall prevention: Includes guardrails for empty search spaces, compilation timeouts, and avoiding replace_hints on the hot path.

Quick Start

Use it to add autotuning to a new CuTile kernel by first classifying it with the decision tree, then generating the smallest relevant search space, then implementing the tune-once/cache/launch wrapper with a DISABLE_AUTOTUNE-safe fallback.