RocmKernelWiki — AMD CDNA3 / CDNA4 Kernel Optimization Knowledge Base

Knowledge cutoff: 2026-05-25. See data/refresh-cutoff.yaml.

A structured knowledge base of AMD GPU kernel optimization for MI300X (CDNA3, gfx942) and MI350 / MI355 (CDNA4, gfx950), packaged as a Claude Code skill. The repository root is the skill directory.

Architecture and conventions mirror KernelWiki (Blackwell/Hopper). Initial content distilled from Apex/tools — skills/, jsons/{hip_sheet,triton_sheet,rocm}.json, and mcps/.

Install as a Claude Code Skill

git clone <repo-url> ~/.claude/skills/RocmKernelWiki
pip install -r ~/.claude/skills/RocmKernelWiki/requirements.txt

Smoke test:

cd ~/.claude/skills/RocmKernelWiki
python3 scripts/query.py --tag mfma --type hardware --compact
python3 scripts/get_page.py hw-mfma-cdna3 --frontmatter-only

Optional override:

export ROCM_WIKI_ROOT=/path/to/RocmKernelWiki

What's Here

• Hardware pages — MFMA (CDNA3 + CDNA4), SMFMAC + 4:2 sparsity, AccVGPR / dual register files, LDS + bank conflicts, GWS, buffer resource ops, async-copy direct-to-LDS, wave64, FP8/BF8/MX-FP4
• Technique pages — coalesced/vectorized access, grid-stride loops, software pipelining, LDS double-buffering, swizzling, register budgeting, occupancy tuning, persistent + stream-K kernels, kernel/epilogue fusion, autotune knobs (num_stages, num_warps, matrix_instr_nonkdim, kpack, waves_per_eu)
• Pattern pages — memory-bound / compute-bound / register-pressure / low-CU-utilization / lds-bank-conflict / lane-conflict diagnostics with candidate techniques
• Kernel case studies — GEMM (HIP MFMA, Triton-AMD, CK), FlashAttention on MI300/MI350, fused MoE, paged attention, RMSNorm
• Language guides — HIP C++, Triton ROCm backend, Composable Kernel, Gluon-AMD, FlyDSL
• Migration guides — CUDA → HIP, Triton-NVIDIA → Triton-AMD, WGMMA → MFMA, TMA → buffer-load + async-copy, MI300 → MI350 (gfx942 → gfx950)
• Source ledgers under candidates/ — triton-amd, vllm-rocm, sglang-rocm, aiter, magpie, flydsl
• Auto-generated query indices under queries/

Query Tools

Tool	Purpose
scripts/query.py	Unified search across all pages (keywords + filters + alias-aware)
scripts/get_page.py	Fetch any page by id or path; --follow-sources expands cited sources
scripts/grep_wiki.py	Regex text search across wiki bodies and source pages
scripts/validate.py	Schema/tag/source-id validator
scripts/generate-indices.py	Rebuilds queries/*.md from frontmatter

Companion Docs

• SKILL.md — Skill entry point: when to engage, query paths, output contract.
• CLAUDE.md — Extended schema + navigation reference.
• references/primer.md — Topic map.
• references/schema.md — Frontmatter schema.
• references/examples.md — Worked query patterns.
• index.md — Curated top-level index.

Architecture

Three layers (the Karpathy LLM Wiki pattern, same as KernelWiki):

1. sources/ — raw data: PRs, vendor docs, blogs, contests.
2. wiki/ — synthesized pages cross-referenced by id. All have YAML frontmatter.
3. queries/ — auto-generated indices. Do not edit by hand.

Supporting files:

• data/schemas.yaml — required/optional fields per page type
• data/tags.yaml — controlled vocabulary
• data/aliases.yaml — canonical → synonym mappings (e.g. gfx950 ← MI350, mfma ← matrix-fma)
• data/refresh-cutoff.yaml / data/tool-versions.yaml — versioning

Scope Rules

• CDNA3/CDNA4-first. gfx942 + gfx950 are primary. gfx90a (CDNA2) only with explicit cdna_relevance. RDNA out of scope.
• Kernel-only. No RCCL / distributed-training / scheduler topics.
• First-class DSLs. HIP, Triton-AMD, Composable Kernel, Gluon-AMD, FlyDSL.
• English canonical.

License

Summaries and wiki syntheses are derivative works citing upstream PRs, docs, and the Apex/KernelWiki projects. Tooling (scripts/, data/, references/) is MIT-style.