[CongMa13]

Fixes #14474

Summary: For human reviewers

The golden file llvm_gfx_all.json is stale. The checks rocke_differential_golden_ll and rocke_differential_golden_canonical fail on develop.

This PR records the golden again on f37a7056c79. The PR changes only llvm_gfx_all.json.

What changed

• target_intrinsics: feat(rocke): make the C++ engine consult the arch-domain table, and add error mode #13723 changed the IR. The golden changes for llvm20, llvm22, llvm23 and canonical.
• moe_fused_mega_fp8: fix(rocke): one valid K=128 f8f6f4 MFMA lowering for fp8, fp6 and fp4 #13976 changed the IR. The golden changes for llvm20, llvm22 and llvm23.
• gfx942_attention_tiled_2d: The golden also changes for this family. The issue does not list it. The cause is not confirmed.

Review request

1. Confirm the authors of feat(rocke): make the C++ engine consult the arch-domain table, and add error mode #13723 and fix(rocke): one valid K=128 f8f6f4 MFMA lowering for fp8, fp6 and fp4 #13976 meant the IR changes.
2. Confirm the author of perf(rocke): build one K/V resource per page on the gfx942 2D i64 path #14349 meant the gfx942_attention_tiled_2d change. The probable source is perf(rocke): build one K/V resource per page on the gfx942 2D i64 path #14349.

Verification

• Before the change, the golden check failed on develop for llvm20 and llvm22.
• After the change, run_diff.py --mode ll --python-only --check-golden --flavor llvm20 --only target_intrinsics,moe_fused_mega_fp8 prints GOLDEN OK.
• The full check for each flavor, the canonical check and the byte-identity gate did not run. CI runs them.

Summary: For AI reviewers

• The diff changes one JSON file. No emitter code changes.
• The command run_diff.py --record-golden wrote the file with --python-only.
• Check the changed hashes only against the three families above.

🤖 Generated with Claude Code

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	PR does not contain code files — Unit Test auto-passed
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[cgmillette]

Automated review: Model - qwen3.6:35b-a3b-coding-mtp-q4_K_M, effort: max

PR #14501 — re-bless the differential golden on develop (CongMa13) — 🟢 ACCEPT: exact, fully attributed re-bless

Prior item	Status
Reviewers	🟢 None yet: no reviews or threads, only the policy bot
Issue #14474 (BrianHarrisonAMD, 10-27): #13723 left target_intrinsics[0,1,2,4,7,9,10,11] (ll) and [7,9,10,11] (canonical) unblessed; #13976 left moe_fused_mega_fp8[1..5] (ll) unblessed	🟢 Matches exactly (ran). The PR changes 63 anchors and nothing else: • target_intrinsics: 8 changed + 1 added per ll flavor, 4 + 1 in canonical; • moe_fused_mega_fp8: 5 changed per ll flavor; • gfx942_attention_tiled_2d: 4 added per section. No family is added or removed
Author's request 1: confirm that #13723 and #13976 meant the IR changes	🟢 Each change appears exactly at its PR's merge and matches that PR's stated intent (ran: every config emitted at each develop commit from before #13723 to today, llvm22; today's hashes equal the PR's anchors). • #13723: each config now lowers for a target that has its intrinsics (the new ARCHES table). Configs 0, 1, 2 and 7 move to gfx1250 and config 4 to gfx1201, so their amdgpu-flat-work-group-size minimum drops from 64 to 32, as wave32 needs. Config 0 also gains gfx1250's s.wait.asynccnt. s.wait.event moves out of config 7 into a new config inserted at index 9. That shifts old configs 9–11 to 10–12 with identical hashes, so the real output changes are 0, 1, 2, 4, 7 and the new 9. • #13976: in moe_fused_mega_fp8 configs 1–5, the only difference is that the K=128 scale MFMA is now called by its overload-mangled name (…16x16x128.f8f6f4.v8i32.v8i32). Undoing that suffix leaves identical IR. Config 0 has no K=128 MFMA and is unchanged
Author's request 2: confirm that the gfx942_attention_tiled_2d change came from #14349	🟢 Confirmed. The change only adds configs: 0–6 hash identically at every commit, and 7–10 first appear at #14349's merge (ran). They are #14349's new i64 K/V parity configs, which our follow-up on #14349 found byte-identical across the two engines at llvm20/22/23
Bots / CI	🟡 The policy bot, pre-commit, clang-tidy and Component CI pass. TheRock Multi-Arch is still building, so the hipkernelprovider shards that run the two golden CTests have not run yet (read)
Our checks	🟢 • The PR merges cleanly with develop (5 behind) (ran). • We ran the exact CTest commands (run_diff.py … --python-only --check-golden: ll at llvm20/22/23, then canonical). On today's develop both fail, reproducing #14474. With the PR merged into today's develop, both print GOLDEN OK, well inside the 300 s CTest timeout (ran). • Byte identity is unaffected: the PR changes only the golden

#	Sev	file:line	Finding
N1	🟡	golden/llvm_gfx_all.json (merge order)	develop broke because #13974's golden was recorded before #13723 and #13976 landed: develop does not require an up-to-date base. Any other rocKE PR that changes emitter output and merges before this one breaks the check again → land this next. If develop's rocKE tree moves first, re-run the two --check-golden commands on the merge before merging

Utility: 🟢 Positive — unblocks develop's golden checks, red on develop's hipkernelprovider runs since 05:14 UTC; exact and fully attributed.

---

Legend: 🔴 blocking · 🟠 should-fix · 🟡 minor nit · 🟢 info/good · ⚪ retracted

First-round review: a local qwen3.6 model made the independent first pass. The orchestrator (Claude Opus 5.5) verified it. The checks were:

• every anchor the PR changes, old against new;
• the CI golden commands on today's develop and on the merge;
• each changed config emitted at every develop commit between feat(rocke): make the C++ engine consult the arch-domain table, and add error mode #13723 and today, to find where its output changes;
• the IR difference behind each change;
• CI at head.

The orchestrator owns this merge-worthiness call. It flags whether a closer look is needed; it is not an approval or a rejection.

[adickin-amd]

issueing override merge

[assistant-librarian]

TheRock Submodule Bump Activity

Newest first

• 2026-10-30 — Included in TheRock via ROCm/TheRock#9594.