Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
ca52fdf
feat(hipdnn): report kernel ingestor runtime impact in deep review
GarbyGilbertAMD Oct 22, 2026
db2145e
fix(hipdnn): close false-safe paths in the deep review impact lane
GarbyGilbertAMD Oct 23, 2026
ad029a8
fix(hipdnn): close more false-safe paths in the deep review impact lane
GarbyGilbertAMD Oct 23, 2026
f019d5c
feat(hipdnn): add the hipdnn-ingestor-impact skill
GarbyGilbertAMD Oct 27, 2026
34b5b4a
fix(hipdnn): pair ingestor impact checks with multi-launch and method…
GarbyGilbertAMD Oct 27, 2026
7edb059
fix(hipdnn): resolve launch surface symbols as launch_surface.py does
GarbyGilbertAMD Oct 28, 2026
706d00f
Merge remote-tracking branch 'origin/develop' into users/ggilbert/ALM…
GarbyGilbertAMD Oct 28, 2026
232251c
feat(hipdnn): test ingestor support predicates and C++ restatements i…
GarbyGilbertAMD Nov 2, 2026
1b47edb
fix(hipdnn): close ingestor impact gaps that read a break as no impact
GarbyGilbertAMD Nov 3, 2026
423c836
fix(hipdnn): harden ingestor impact probes, parsing and lane wiring
GarbyGilbertAMD Nov 3, 2026
7284649
chore(hipdnn): merge develop into the ingestor impact branch
GarbyGilbertAMD Nov 3, 2026
bb5fc75
fix(hipdnn): follow develop's flavor and descriptor layout in ingesto…
GarbyGilbertAMD Nov 3, 2026
cd19d0c
Merge remote-tracking branch 'origin/develop' into users/ggilbert/ALM…
GarbyGilbertAMD Nov 3, 2026
b2a5fda
fix(hipdnn): keep ingestor impact folds and readers sound across targets
GarbyGilbertAMD Nov 4, 2026
3219fac
fix(hipdnn): keep ingestor impact fallbacks and test the fold guard
GarbyGilbertAMD Nov 4, 2026
bfc7f13
fix(hipdnn): read unresolved repository imports as ingestor readers
GarbyGilbertAMD Nov 5, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions AGENTS.md
Original file line number Diff line number Diff line change
Expand Up @@ -77,7 +77,9 @@ All under [`projects/hipdnn/tools/ai/skills/`](projects/hipdnn/tools/ai/skills/)
| `hipdnn-superbuild-test` | Running tests against an existing hipDNN superbuild. |
| `hipdnn-tidy` | Running clang-tidy on hipDNN, scoped to the files you changed. |
| `hipdnn-ingestor-engine` | Creating or extending a generic-kernel-ingestor integration. |
| `hipdnn-ingestor-impact` | Checking whether a kernel ingestor change breaks kernel ABI or dispatch at run time while the build stays green, without a full review. |
| `hipdnn-uhd` | Taking a hipDNN engine through the UHD heuristic lifecycle on the local GPU: generating, training, judging, validating, installing, and triaging heuristics and their training data. |
| `hipdnn-deep-review` | Reviewing a hipDNN or provider change in depth with parallel agents until its findings close. Its `IMPACT` lane runs the `hipdnn-ingestor-impact` scripts. |
| `hipdnn-codegen` (at [`projects/hipdnn/tools/DescriptorGenerator/.claude/skills/`](projects/hipdnn/tools/DescriptorGenerator/.claude/skills/hipdnn-codegen/SKILL.md)) | Generating operation boilerplate from a YAML config. |
| `rfc-review` | Reviewing a technical RFC across the structured perspectives. Follow-ups: `rfc-review-compatibility`, `rfc-review-ops`, `rfc-review-security`. |
| `rfc-backlog` | Turning an RFC into a backlog of Jira stories and tasks. |
Expand Down
6 changes: 5 additions & 1 deletion projects/hipdnn/docs/ai-rules.md
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,8 @@ When a user asks for a workflow covered by a project skill, tell them the projec
- Deprecated stub that redirects to `hipdnn-pr-quality` (review assist). Same behavior: announce deprecation, then defer.
- `tools/ai/skills/hipdnn-deep-review/SKILL.md`
- Runbook for a multi-agent deep review of a hipDNN or provider PR, branch or commit range: scouts map each affected area, focused reviewer lanes run in parallel, every finding is verified by source proof, reproduction or mutation, and a ledger carries findings across rounds until no blocking or important one stays open. Owns its review policy; the review is advisory and never posts without approval.
- Suggest this skill when the user asks for a deep, thorough or multi-agent review, or wants to iterate a review until all findings close. Use `hipdnn-pr-quality` for drafting PR text or a single-pass review. Requires subagents and its sibling skills `hipdnn-superbuild` and `hipdnn-superbuild-test`.
- Suggest this skill when the user asks for a deep, thorough or multi-agent review, or wants to iterate a review until all findings close. Use `hipdnn-pr-quality` for drafting PR text or a single-pass review. Requires subagents and its sibling skills `hipdnn-superbuild`, `hipdnn-superbuild-test` and `hipdnn-ingestor-impact`.
- Its `IMPACT` lane runs the `hipdnn-ingestor-impact` scripts and turns their rows into review findings.
- `tools/ai/skills/hipdnn-superbuild/SKILL.md`
- Builds hipDNN together with one or more providers via the repository-root superbuild presets (`hipdnn-providers`, `miopen-provider`, `hipblaslt-provider`, `hip-kernel-provider`, `hipdnn-samples`, etc.), in a single CMake invocation. On Windows it auto-runs the wheel-based ROCm setup when no SDK path is supplied.
- Suggest this skill when the user asks to build hipDNN with providers, run a superbuild preset, rebuild after a rebase or merge, or set up a fresh build from the repo root. Prefer it over the standalone build whenever providers are involved.
Expand All @@ -47,6 +48,9 @@ When a user asks for a workflow covered by a project skill, tell them the projec
- `tools/ai/skills/hipdnn-ingestor-engine/SKILL.md`
- Creates or extends a hipDNN generic-kernel-ingestor integration: graph and kernel contracts, descriptors, native hooks, packaging, host checks, and device correctness.
- Suggest this skill when the user asks to add or extend an ingestor-based engine.
- `tools/ai/skills/hipdnn-ingestor-impact/SKILL.md`
- Reports kernel ingestor changes that break at run time while the build stays green: kernel argument lists out of step with their C++ and test copies (also against the target branch's tip, which catches two PRs that each pass alone), dispatch predicates, descriptors and generic targets, removed Python symbols still read, and which shipped variants' emitted code changed. Reads git objects; its optional emitter runs rocKE on the CPU. Advisory: it never builds, tests or posts.
- Suggest this skill for a quick answer on a hip-kernel-provider, rocKE or ingestor change without a full review, and before turning the ingestor flags ON by default. Use `hipdnn-deep-review` for a full review; its `IMPACT` lane runs the same scripts.
- `tools/ai/skills/rfc-review/SKILL.md`, with the focused follow-ups `rfc-review-compatibility`, `rfc-review-ops`, and `rfc-review-security`
- Review a technical RFC or design document and produce a written review report. They are not for implementation PRs.
- Suggest these skills when the user asks to review an RFC, design doc, ADR, or proposal.
Expand Down
18 changes: 12 additions & 6 deletions projects/hipdnn/tools/ai/skills/hipdnn-deep-review/RUNBOOK.md
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,7 @@ needs subagents; without them, stop and tell the user to run `hipdnn-pr-quality`

`<pr>` is the PR number, `<N>` the round. `<owner>/<repo>` comes from the PR URL, else
`origin` of `$CLONE`, else `ROCm/rocm-libraries`; pass `-R <owner>/<repo>` to every `gh pr` and
`gh issue` call. `pin.py`, `partition.py`, `ledger.py` and `comments.py` are in `$SKILL/scripts/`;
`gh issue` call. `pin.py`, `partition.py`, `ledger.py` and `comments.py` are in `$SKILL/scripts/` (the `IMPACT` lane's scripts are in the sibling skill `hipdnn-ingestor-impact`);
a script name below stands for `python3 "$SKILL/scripts/<name>"` (`python` or `py -3` on Windows, as in SKILL.md *Platforms*). Each prints terse results and
exits non-zero with a one-line reason. After pinning, commands run from `$EVIDENCE/r<N>/`
(sourcing `env.sh` changes to it); GNU coreutils, grep and sed are assumed.
Expand Down Expand Up @@ -151,6 +151,7 @@ its trigger holds, else mark it `N/A`; when unsure, run it.
| `DOCS` | [docs](lanes/docs.md) | Accuracy and concision of docs and comments, including edited public headers | A documentation or source file changes |
| `DRIFT` | [drift](lanes/drift.md) | The change merged onto the target's current tip: conflicts, broken target-side contracts, tests on the merge | `drift.txt` says `Trigger: yes` |
| `RUN` | [build-validate](lanes/build-validate.md) | Building the head and running the relevant tests | Always |
| `IMPACT` | [runtime-impact](lanes/runtime-impact.md) | Kernel ABI and dispatch contracts of the kernel ingestor (kernarg lists, launch surfaces, matchers, descriptors, generic targets) at base, head and the merged tree; which shipped variants change and which device runs would settle it | A file under `dnn-providers/hip-kernel-provider/`, `projects/hipdnn/plugin_sdk/include/hipdnn_plugin_sdk/` (any plugin SDK header: the report works out which ones the descriptor loader includes), `projects/hipdnn/tools/IngestorGenerator/`, `projects/hipdnn/plugin_sdk/data/gpu_generic_targets.json` or `projects/hipdnn/plugin_sdk/cmake/GpuGenericTargets*` changes |

`ORCH` (process findings you file), `SCOUT<k>`, `SYNTH` ([orient](lanes/orient.md)) and `VER`
([verify](lanes/verify.md)) are not partition choices. Merge two lanes into one agent when the
Expand All @@ -167,7 +168,7 @@ ledger.py seed "$EVIDENCE" <N> --lane 'DEPS=N/A (no manifest or source changes)'

`packets` writes `packets/<LANE>.md` ([lanes/common.md](lanes/common.md) *Context packet*) for
each lane of `owners.tsv` and each concern lane in `--lanes` (changed files in scope with their
`numstat` rows, areas and Orient paths; BUILD and DEPS scope to their trigger files and print
`numstat` rows, areas and Orient paths; BUILD, DEPS and IMPACT scope to their trigger files and print
`<LANE>: no changed file in scope, N/A` when none; `DRIFT` adds `drift.txt` and prints
`DRIFT: N/A (<reason>)` when its trigger is `no` (give that reason to `seed` with `--lane`); every other concern lane takes all changed
files, so name only lanes whose trigger holds; `--parent` is the stack parent). In round 2 on,
Expand Down Expand Up @@ -209,10 +210,12 @@ Scope: <user's scope notes, e.g. areas being deprecated; "do not file findin
writer as its own fresh agent with shell and write access; fill this template as its first
message. Agents see only their dispatch: each begins by reading `dispatch-context.md`.

**Queue.** Run at most `Agents:` agents at once, in this priority order: `RUN` first, then `DRIFT`, the
**Queue.** Run at most `Agents:` agents at once, in this priority order: `RUN` first, then `DRIFT`, then
`IMPACT` (its emitter builds every shipped variant twice), then the
longest lanes (largest `CORR<k>`, then `TESTQ`), `DUP` and `DOCS` last. `RUN`, `DRIFT`, an approved
`ROCK` build, `ROCK`'s configure diff (`Build host: yes`), or any other lane that builds or mutates, run concurrently in their own `$SCRATCH` trees and
split `Cores:` between them while they overlap, `RUN` taking the larger share.
split `Cores:` between them while they overlap, `RUN` taking the larger share. `IMPACT`'s emitter does not
read `Cores:`: it builds the two sides on two concurrent workers, so count two cores against `Cores:` while it runs.

```markdown
Role: <LANE> | SCOUT<k> | SYNTH | VER-<batch> | WRITER; round <N>; follow-up f<k> if any
Expand All @@ -225,7 +228,7 @@ Inputs: under $EVIDENCE/r<N>/: PIN.md; pr/body.md or intent.md; CONTEXT.md an
Scout: area, slug, changed paths, prior area file or none. SYNTH: orient/*.md.
Verifier: finding IDs and report paths. WRITER: LEDGER.md, verify/, PIN.md,
pr/pr.json (title), CONTEXT.md, lanes/*.md (M-rule lines, RUN execution counts,
ROCK not-run items). `RUN`, `ROCK`, `AUTH`: the **Options and guards** and
ROCK not-run items). `RUN`, `ROCK`, `AUTH`, `IMPACT`: the **Options and guards** and
**Consumers** sections of `CONTEXT.md` and the area files. Follow-up: the lane report
path plus the approved request, decision or Handoff.
Output: <exact file>; scratch $SCRATCH/<ID>/; first provisional ID <LANE>-p<n> (a follow-up
Expand All @@ -245,7 +248,7 @@ waits; the answer resumes the round (*Iterate*, same head).
A follow-up dispatch (approved request, decision, Handoff) carries the lane's report path and
the item only, and writes `lanes/<LANE>-f<k>.md`, numbering IDs after the lane's last `p<n>`.

`AUTH`, `TESTQ` and `DRIFT` run in the first fan-out without waiting for `RUN`. When `RUN` returns, dispatch
`AUTH`, `TESTQ`, `DRIFT` and `IMPACT` run in the first fan-out without waiting for `RUN`. When `RUN` returns, dispatch
`AUTH-f1` with the `AUTH` and `RUN` report paths: it reconciles the claims and the
**Consumers** list against `RUN`'s executed counts ([lanes/author-testing.md](lanes/author-testing.md)).
When `TESTQ` wrote `M1: UNVERIFIED`, also dispatch `TESTQ-f1` with the `TESTQ` and `RUN` report paths: it settles M1 from
Expand Down Expand Up @@ -352,6 +355,9 @@ Findings whose `Where` is in `affected.txt`, and every `process` finding, are af
each lane the delta touches, every lane `FAILED` earlier (closing its `ORCH` finding when it
completes), and always `AUTH` and `LEAK`: a new head brings new commits, PR text and CI runs.
Re-run `DRIFT` whenever the round's `drift.txt` says `Trigger: yes`: the target moves between rounds.
Re-run `IMPACT` whenever the round's pinned `TARGET_TIP` differs from the tip of its last run: its
merged view is computed against that tip and goes stale with it. Only *Pin and intake* records a
tip: a round reopened on the same head keeps its pinned tip, so its `IMPACT` re-run does too.
Re-verify every affected `OPEN` or `UNVERIFIED` finding. A lane not re-run carries forward
(Coverage and Lanes rows copied, marked `carried from r<M>`).

Expand Down
11 changes: 8 additions & 3 deletions projects/hipdnn/tools/ai/skills/hipdnn-deep-review/SKILL.md
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
---
name: hipdnn-deep-review
description: "Multi-agent deep review of a hipDNN or dnn-providers PR, branch or commit range. Scouts map each affected area, focused reviewer lanes run in parallel, every finding is verified by source proof, reproduction or mutation, and a ledger carries findings across rounds until no blocking or important one stays open. For a PR body or a quick single-pass review, use hipdnn-pr-quality."
description: "Multi-agent deep review of a hipDNN or dnn-providers PR, branch or commit range. Scouts map each affected area, focused reviewer lanes run in parallel, every finding is verified by source proof, reproduction or mutation, and a ledger carries findings across rounds until no blocking or important one stays open. Its IMPACT lane reports kernel ingestor changes that break at run time while the build stays green (kernel ABI, dispatch, descriptors) with the scripts of hipdnn-ingestor-impact, which answers that question without a review. For a PR body or a quick single-pass review, use hipdnn-pr-quality."
argument-hint: "[<PR URL or number> | branch:<name> | <base>..<head>] evidence:<dir> scratch:<dir> [clone:<dir>] [intent:<file>] [terms:<file>] [knowledge:<dir>] [round:<N>] [prior:<dir>]"
allowed-tools: Bash, Read, Write, Edit, Grep, Glob, Task
---
Expand All @@ -13,10 +13,12 @@ scouts, lanes and verifiers, merge their output and write the report; you do not
code. Your formats are in [report.md](report.md); [RUNBOOK.md](RUNBOOK.md) *Fan out* says
what each agent reads, and `lanes/` holds the briefs. `RUN` builds with
[hipdnn-superbuild](../hipdnn-superbuild/SKILL.md) and tests with
[hipdnn-superbuild-test](../hipdnn-superbuild-test/SKILL.md); install those skills alongside this one.
[hipdnn-superbuild-test](../hipdnn-superbuild-test/SKILL.md); `IMPACT` runs the scripts of
[hipdnn-ingestor-impact](../hipdnn-ingestor-impact/SKILL.md). Install those skills alongside this one.

Prerequisites: network access, an authenticated `gh` with read access, `git` 2.38 or later (`pin.py` uses `git merge-tree --write-tree`), and Python 3 (it runs the scripts under
`$SKILL/scripts/`); `pre-commit` and `pytest` for the lanes that run validators. `RUN` and verifiers also need
`$SKILL/scripts/`); `pre-commit` and `pytest` for the lanes that run validators. The `IMPACT` lane needs
Python 3.10 or newer and PyYAML; its emitter also needs Jinja2 ([hipdnn-ingestor-impact](../hipdnn-ingestor-impact/SKILL.md) *Prerequisites*). `RUN` and verifiers also need
a ROCm install (Linux: `/opt/rocm` or `ROCM_PATH`; Windows: the wheel SDK that [hipdnn-superbuild](../hipdnn-superbuild/SKILL.md) provisions), CMake, Ninja and about 25 GB free in `$SCRATCH`; without them `defect` and
`test-gap` findings stay `UNVERIFIED` and the review closes only through human
`ACCEPTED-RISK` decisions, recorded with `ledger.py set` (`PIN.md` records `Build host`).
Expand All @@ -25,6 +27,9 @@ Platforms: Linux, macOS and Windows. On Windows run the skill from Git Bash (Git
`bash`, `awk`, `sed`, `grep` and `comm`. Blocks write the interpreter as `python3`; on Windows that name is the Microsoft Store
stub, so use `python` or `py -3` instead. Scripts read and write UTF-8 and LF whatever the host's locale.

For a quick answer without a review, use [hipdnn-ingestor-impact](../hipdnn-ingestor-impact/SKILL.md): it runs the
same scripts as the `IMPACT` lane ([lanes/runtime-impact.md](lanes/runtime-impact.md)) and ends in a `Runtime impact:` line.

## Entry contract

Ask for what is missing; record it in `PIN.md`:
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -183,7 +183,7 @@ load_output() {
load_output || { echo "STOP: sweep input missing, not clean" >&2; exit 1; }
sed -i.bak 's#users/[A-Za-z0-9._-]*/#users/#g' "$L/pub.content" && rm "$L/pub.content.bak" # the expected branch prefix
cat > "$L/sweep-allow.txt" <<'EOF2'
^(CORR|TESTQ|TESTB|AUTH|BUILD|DEPS|API|ROCK|LEAK|DUP|DOCS|RUN|DRIFT|ORCH|VER|SCOUT|SYNTH)[0-9]*-p?[0-9]+$
^(CORR|TESTQ|TESTB|AUTH|BUILD|DEPS|API|ROCK|LEAK|DUP|DOCS|RUN|DRIFT|IMPACT|ORCH|VER|SCOUT|SYNTH)[0-9]*-p?[0-9]+$
^(handoffs?|worktrees?)$
^https?://(github\.com/ROCm/[A-Za-z0-9._-]+|raw\.githubusercontent\.com/ROCm/[A-Za-z0-9._-]+|rocm\.docs\.amd\.com|docs\.amd\.com|www\.amd\.com|example\.com)([/:?#].*)?$
EOF2
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -29,10 +29,10 @@ where, what it covers, how to select it); **Options and guards**; **Consumers**.

**Options and guards.** For the area's changed files: each `option()` and cache variable with its default, each
`#ifdef`, `#if` and CMake `if(...)` guard, and for each change the flag it depends on and that flag's default, so a change
that is dead code with the flag OFF shows. `RUN`, `ROCK` and `AUTH` read this section.
that is dead code with the flag OFF shows. `RUN`, `ROCK`, `AUTH` and `IMPACT` read this section.

**Consumers.** The in-repo includers of the area's changed public headers (`git -C "$CLONE" grep -l <header name> "$BASE"
-- <pathspec>`) and which project or provider tests exercise them. `RUN` and `AUTH` read this section.
-- <pathspec>`) and which project or provider tests exercise them. `RUN`, `AUTH` and `IMPACT` read this section.

Sources, in order: the area's `README.md` and `docs/`; `projects/hipdnn/docs/`; RFCs under
`projects/hipdnn/docs/rfcs/`; then the code and tests. Check a document against the code
Expand Down
Loading
Loading