docs(benchmarks): add baseline-gemmini-rocket skeleton for Phase 2 #101

Merged

navigator merged 1 commit from auto/issue-98-20260526T120712Z_issue98 into main

2026-05-26 09:20:48 -03:00

navigator commented

2026-05-26 09:10:34 -03:00

Owner

Summary

Promote docs/benchmarks/baseline-gemmini-rocket.md from Stub to Draft skeleton so Phase 2 (PLAN.md §7) measurements have a stable landing spot and report-v1.md has a deterministic anchor for its speedup denominators.

Sections added per the issue acceptance criteria:

Overview — scope: upstream CONFIG=GemminiRocketConfig, pre-FluidPop
Build environment — Chipyard hash (TBD until §6 Week 4 lands), Verilator, toolchain commit; placeholder pin table
Workloads — bare-metal Gemmini self-tests + ResNet-50 FP16 batch=1
Metrics collected — placeholder tables (header rows only) for cycles, scratchpad utilisation, branch predictor, IPC
Methodology notes — single-threaded sim, warm-up TBD, reset state, determinism / seeds
Caveats — sim-only, no power, no area (synthesis report scope), upstream-only configuration
Reproduction recipe — placeholder; command lines TBD until Phase 2 lands

Honest-reporting boundary held: no fabricated cycle counts, utilisation percentages, or branch-predictor values committed. Per PLAN.md §9.3, every result row will be added only when a Verilator run produces it and the producing commit hash + harness invocation are recorded alongside.

docs/benchmarks/README.md status column for this doc updated Stub → Draft skeleton to keep the index in sync.

Plan refs

§7 (Phase 2 baseline measurements) — primary owner of these numbers
§6 Week 4 (Chipyard pin in rtl/chipyard.hash) — flagged as TBD-blocker for the build-environment row
§10 deliverable #5 (report-v1.md) — consumer of the baseline as speedup denominator
§9.3 (honest reporting) — bounds every TODO in the file

Test plan

File present at docs/benchmarks/baseline-gemmini-rocket.md
Status: Draft skeleton + Owner: TBD header
All 7 sections present with TODO markers and 2–3 line intent
Placeholder tables carry header rows only — no fabricated data
Chipyard hash explicitly noted TBD until §6 Week 4 lands
PLAN.md §§7, 6 Week 4, 10 referenced
docs/benchmarks/README.md updated to reflect new status

Closes #98

## Summary Promote `docs/benchmarks/baseline-gemmini-rocket.md` from `Stub` to `Draft skeleton` so Phase 2 (PLAN.md §7) measurements have a stable landing spot and `report-v1.md` has a deterministic anchor for its speedup denominators. Sections added per the issue acceptance criteria: 1. Overview — scope: upstream `CONFIG=GemminiRocketConfig`, pre-FluidPop 2. Build environment — Chipyard hash (TBD until §6 Week 4 lands), Verilator, toolchain commit; placeholder pin table 3. Workloads — bare-metal Gemmini self-tests + ResNet-50 FP16 batch=1 4. Metrics collected — placeholder tables (header rows only) for cycles, scratchpad utilisation, branch predictor, IPC 5. Methodology notes — single-threaded sim, warm-up TBD, reset state, determinism / seeds 6. Caveats — sim-only, no power, no area (synthesis report scope), upstream-only configuration 7. Reproduction recipe — placeholder; command lines TBD until Phase 2 lands Honest-reporting boundary held: no fabricated cycle counts, utilisation percentages, or branch-predictor values committed. Per PLAN.md §9.3, every result row will be added only when a Verilator run produces it and the producing commit hash + harness invocation are recorded alongside. `docs/benchmarks/README.md` status column for this doc updated `Stub` → `Draft skeleton` to keep the index in sync. ## Plan refs - §7 (Phase 2 baseline measurements) — primary owner of these numbers - §6 Week 4 (Chipyard pin in `rtl/chipyard.hash`) — flagged as TBD-blocker for the build-environment row - §10 deliverable #5 (`report-v1.md`) — consumer of the baseline as speedup denominator - §9.3 (honest reporting) — bounds every TODO in the file ## Test plan - [x] File present at `docs/benchmarks/baseline-gemmini-rocket.md` - [x] `Status: Draft skeleton` + `Owner: TBD` header - [x] All 7 sections present with TODO markers and 2–3 line intent - [x] Placeholder tables carry header rows only — no fabricated data - [x] Chipyard hash explicitly noted TBD until §6 Week 4 lands - [x] PLAN.md §§7, 6 Week 4, 10 referenced - [x] `docs/benchmarks/README.md` updated to reflect new status Closes #98

navigator added 1 commit

2026-05-26 09:10:34 -03:00

docs(benchmarks): add baseline-gemmini-rocket skeleton for Phase 2

build / scalafmt-check (pull_request) Successful in 4s

Details

build / sbt-compile (pull_request) Successful in 3s

Details

build / shell-lint (pull_request) Successful in 20s

Details

2b5d5521e7

Promote the baseline reference doc from Stub to Draft skeleton so Phase 2
(PLAN.md §7) measurements have a stable landing spot and the head-to-head
report-v1.md has a deterministic anchor for its speedup denominators.

The doc is intentionally numbers-free per PLAN.md §9.3 (honest reporting):
sections, placeholder tables (header rows only), and TODO markers are in
place; cycle counts, scratchpad / branch-predictor / IPC values arrive
only with their producing sim runs cited (Chipyard hash + Verilator
version + harness invocation). Chipyard hash is explicitly flagged TBD
until PLAN.md §6 Week 4 lands and rtl/chipyard.hash is populated.

README.md status column for this doc updated Stub -> Draft skeleton to
match.

navigator added the

ready-for-review

label

2026-05-26 09:10:43 -03:00

navigator commented

2026-05-26 09:18:42 -03:00

Author

Owner

VERDICT: PASS

PR #101 promotes docs/benchmarks/baseline-gemmini-rocket.md from Stub to Draft skeleton, alongside a minimal docs/benchmarks/README.md status-column sync. Scope is cleanly bounded to Issue #98 and consistent with the structure established by PR #90 (which created the docs/benchmarks/ tree). Mechanical guardrails pass: existing CC-BY-SA-4.0 SPDX headers preserved on both files (diff starts at line 4 on the baseline doc, header untouched on the README), no AI/Anthropic attribution, no off-limits paths, no Chisel module so rules 4–6 don't apply. Rule 7 is handled meticulously: every result-table cell is uniformly TBD, every section is a TODO block, and the top-of-file callout plus the closing paragraph both restate that "no cycle count, utilisation percentage, or branch predictor statistic may be fabricated, extrapolated, or back-filled from the analytic model" — directly anchored to PLAN.md §9.3 honest reporting. The §6 Caveats list (sim-only, no power, no area, upstream-only) reinforces the boundary at the right scope, mirroring PR #90's same-section pattern. Every numerical claim that does appear is attributed to a specific PLAN.md section: "Verilator >= 5.012 floor per Section 5.3", "Tag 1.13.0 per Section 6" (Chipyard), "~10 kHz sim-clock ceiling on a 16-core host called out in Section 8.3" — these match prior approved PRs (PR #90 used the same Verilator/Chipyard/§8.3 references). The Reproduction recipe in §7 uses a code block but the content is honestly TBD — populated when PLAN.md Section 7 (Phase 2) lands with a non-fabricated description of what the recipe will contain. The pinned-toolchain table in §2 carries TBD in every "Pinned value" cell, with the Chipyard row explicitly noting it's awaiting rtl/chipyard.hash from §6 Week 4. The README status-column update from "Stub" to "Draft skeleton" is the right minimal companion edit — keeping the index in sync rather than a drive-by — and matches what PR #90 already had for report-v1.md. The doc correctly positions itself as the speedup denominator for report-v1.md without committing the numerator side, exactly the separation PR #90 set up.

Findings

None.

**VERDICT:** PASS PR #101 promotes `docs/benchmarks/baseline-gemmini-rocket.md` from Stub to Draft skeleton, alongside a minimal `docs/benchmarks/README.md` status-column sync. Scope is cleanly bounded to Issue #98 and consistent with the structure established by PR #90 (which created the `docs/benchmarks/` tree). Mechanical guardrails pass: existing `CC-BY-SA-4.0` SPDX headers preserved on both files (diff starts at line 4 on the baseline doc, header untouched on the README), no AI/Anthropic attribution, no off-limits paths, no Chisel module so rules 4–6 don't apply. Rule 7 is handled meticulously: every result-table cell is uniformly `TBD`, every section is a `TODO` block, and the top-of-file callout plus the closing paragraph both restate that "no cycle count, utilisation percentage, or branch predictor statistic may be fabricated, extrapolated, or back-filled from the analytic model" — directly anchored to PLAN.md §9.3 honest reporting. The §6 Caveats list (sim-only, no power, no area, upstream-only) reinforces the boundary at the right scope, mirroring PR #90's same-section pattern. Every numerical claim that does appear is attributed to a specific PLAN.md section: "Verilator `>= 5.012` floor per Section 5.3", "Tag `1.13.0` per Section 6" (Chipyard), "~10 kHz sim-clock ceiling on a 16-core host called out in Section 8.3" — these match prior approved PRs (PR #90 used the same Verilator/Chipyard/§8.3 references). The Reproduction recipe in §7 uses a code block but the content is honestly `TBD — populated when PLAN.md Section 7 (Phase 2) lands` with a non-fabricated description of what the recipe will contain. The pinned-toolchain table in §2 carries `TBD` in every "Pinned value" cell, with the Chipyard row explicitly noting it's awaiting `rtl/chipyard.hash` from §6 Week 4. The README status-column update from "Stub" to "Draft skeleton" is the right minimal companion edit — keeping the index in sync rather than a drive-by — and matches what PR #90 already had for `report-v1.md`. The doc correctly positions itself as the speedup *denominator* for `report-v1.md` without committing the numerator side, exactly the separation PR #90 set up. ## Findings _None._

navigator added the

review:pass

label

2026-05-26 09:18:42 -03:00