Scientific DCU Porting

Version: 0.4.1

A reusable, agent-neutral skill and engineering playbook for porting CFD, PDE, and other scientific-computing codes to Hygon DCU/DTK/HIP. It works with any agent that supports the standard skill convention (Codex, Claude Code, and others).

This is a porting/migration capability, not an application entry point. Application skills are authored with scnet-app, use scnet-hpc as the foundation, and remain the single skill users load for normal workload execution.

Contents

• SKILL.md — the skill entry point.
• references/porting-playbook.md — the numerical, runtime, memory, testing, evidence, and scope checklist, including mature-application regression checks folded in from the retired OneFLOW workflow.
• references/incremental-migration.md — how to migrate a mature application to a new version without treating it as a from-scratch port.
• scripts/compare_oracle.py — offline CPU-oracle vs device comparison (text or .npy) with finite/positivity/conservation invariants and a JSON evidence report.
• scripts/check_evidence.py — lints porting case cards in assets/cases/ against the evidence contract.
• scripts/validate.py — offline repository self-checks.
• scripts/probe_compute_node.sbatch — target-compute-node probe template (toolchain, gfx architecture, minimal HIP kernel).
• scripts/install.sh — idempotent installer for Codex and Claude Code skill directories (--codex, --claude, --project, --link).
• assets/cases/ — measured porting case cards; see its README and _template.json.
• agents/openai.yaml — Codex UI metadata; .claude-plugin/ — Claude Code plugin manifests.
• tests/ — unit tests; .github/workflows/validate.yml — cross-platform CI.

Use

Install with scripts/install.sh (Codex or Claude Code), add the repository as a Claude Code marketplace plugin via .claude-plugin/, or copy the repository into any agent's skill directory. Invoke it explicitly with $scientific-dcu-porting or the agent's equivalent. Use the scnet-hpc skill alongside it for cluster profiles, Slurm, SSH, modules, and compute-node operations.

Validate a clone offline with:

python3 scripts/validate.py && python3 scripts/check_evidence.py && python3 -m unittest discover -s tests

The guidance is intentionally evidence-driven: preserve a CPU numerical oracle, validate on the target compute node, separate diagnostic and production data paths, and keep unverified accelerator capabilities clearly marked.