Gluon Kernel Development Skill

A Codex skill for writing, validating, profiling, and improving GPU kernels with Triton's experimental Gluon language.

Start from a trusted reference or the actual production implementation; a Triton intermediate and translation are optional. Define numerical and measurement contracts, use Proton for GPU timing, and test algorithm, layout, pipeline, and scheduling hypotheses with constrained autotuning. Detailed counters and intra-kernel instrumentation are optional diagnostic tools. Numerically screen candidates, sanitize new memory/synchronization designs early, and independently qualify promoted artifacts.

The guidance is general: workload-specific winners, tile sizes, clocks, and tolerances belong in project records, not in the skill's defaults. NVIDIA and AMD mechanisms remain conditional on the installed tools and target.

Install

Ask Codex to install the skill from this repository:

$skill-installer Install the skill from https://github.com/Deep-Learning-Profiling-Tools/gluon-kernel-development/tree/main/skills/gluon-kernel-development

Restart Codex after installation so the new skill is discovered.

Contents

• SKILL.md: the kernel-development workflow.
• references/api-usage-reference.md: concise Gluon API and layout examples distilled from the tutorials.
• references/performance-reference.md: optimization patterns distilled from Gluon tutorials and examples.
• references/profiling-benchmarking-autotuning.md: Proton-first benchmarking, selective NCU analysis, and autotuning guidance.
• references/proton-instrumentation.md: PC attribution, partition scopes, cross-partition events, and observation-window limits.
• references/correctness-checks.md: staged numerical, sanitizer, memory-lifetime, and deployment checks.
• MEMORY.md: an intentionally empty task-memory template for recording evidence when requested or authorized by the host's memory policy.

The skill is self-contained and does not assume that a Triton checkout is present in the workspace.

Contributing

Keep MEMORY.md as a template. Do not publish task-specific measurements, machine details, or user code recorded by an installed copy of the skill.