Add Gpu Op
Developmentadd-gpu-opAdd Gpu Op is an Agent Skill for development published in the openvinotoolkit/openvino repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.
Add a new operation to the OpenVINO GPU plugin — OpenCL kernel design, oneDNN-backed paths, sub-group/LWS tuning, and functional tests.
How to use this skill
Bring this guide into your coding agent with a prompt tailored to the tool you use.
- Open your project in Codex.
- Copy and paste the prompt below.
- Review files and risks before approving any action.
Inspect the Agent Skill at https://github.com/openvinotoolkit/openvino/blob/HEAD/.github/agents-prototype/skills/add-gpu-op/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/add-gpu-op, but wait for my approval before writing files or running scripts.
Copying this prompt does not install or run the skill. Review third-party files before use.
Skill: Add GPU Op
When to use
When the Core OpSpec agent has produced a new op spec and you need to implement GPU plugin support: kernel implementation, registration, and testing.
Steps
Execute in order — each step produces artifacts consumed by the next.
| Step | File | Purpose |
|---|---|---|
| 0a | step0-plan.md | Read op spec, build implementation plan, decide kernel vs oneDNN path |
| 0b | step0-parse-spec.md | Parse op spec JSON/MD into GPU-readable format |
| 1 | step1-hardware-analysis.md | Identify hardware constraints, sub-group size, memory layout requirements |
| 2 | step2-file-structure.md | Create kernel and primitive files, register in factory |
| 3a | step3-kernel-development.md | Write OpenCL kernel (blocked reads, sub-groups, LWS tuning) |
| 3b | step3-write-tests.md | Write layer tests |
| 3c | step3-run-tests.md | Run and verify tests |
| 3d | step3-profiling.md | Profile and tune kernel |
| 4 | step4-onednn-integration.md | Add oneDNN-backed primitive path where applicable |
| 5 | step5-optimize.md | Final performance optimizations |
Entry point for orchestrated use
When invoked by the GPU Agent or Enable Operator Agent, start from orchestrator.md — it selects the relevant subset of steps based on the op type and available hardware.