a5: batch sync_to_sm every K=16 local advances

[yanghaoran29]

Summary

• Publish RingSchedState::last_task_alive to shared memory every 16 local advances while work remains.
• Force-publish the drained tail so a final partial batch cannot stall reclamation.
• Reset the publication shadow when a ring is initialized or reused.
• Add A5 unit coverage for 1, 15, 16, and 17 advances in active and drained states.

Validation

• Local pre-commit: passed.
• C++ unit tests (non-hardware): 66/66 passed.

Extract from PR #906.

[coderabbitai]

📝 Walkthrough

Walkthrough

The PR adds an OCCUPY-based AICPU topology fallback, throttles ring scheduler updates to shared memory, and introduces A5-specific benchmark mappings with a single-example benchmark wrapper.

Changes

AICPU topology fallback

Layer / File(s)	Summary
OCCUPY topology synthesis and probing src/a5/platform/onboard/host/aicpu_topology_probe.cpp	CPU topology probing now intersects CPU_TOPO results with OCCUPY data and synthesizes entries from occupied CPU IDs when no usable CPU_TOPO entries exist.

Ring scheduler synchronization

Layer / File(s)	Summary
Throttled SM publication src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.h	RingSchedState tracks the last published task counter and updates shared memory only after a 16-task advance.

A5 benchmark selection

Layer / File(s)	Summary
A5 example configuration tools/benchmark_rounds.sh	Adds A5 example cases and ordering, selects them for the A5 architecture, and supports validating a single requested example.
Case1 benchmark wrapper tools/benchmark_a5_case1.sh	Adds strict option parsing and runs the A5 tensormap-and-ringbuffer benchmark for paged_attention_unroll.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Sequence Diagram(s)

sequenceDiagram
  participant Probe as probe_aicpu_topology_uncached
  participant CPU_TOPO
  participant OCCUPY as OCCUPY bitmap
  Probe->>CPU_TOPO: query_cpu_topo
  CPU_TOPO-->>Probe: entries or failure
  Probe->>OCCUPY: intersect reported CPU IDs
  OCCUPY-->>Probe: usable IDs or empty result
  Probe->>OCCUPY: synthesize pool when needed

Loading

Possibly related PRs

• hw-native-sys/simpler#1521: Related AICPU thread resolution and clamping based on usable probed cores.
• hw-native-sys/simpler#1552: Related A5 OCCUPY-based CPU topology fallback.

Poem

A rabbit found CPUs in a bitmap bright,
And taught the scheduler to hop just right.
Sixteen steps, then memory sings,
A5 benchmarks flutter on speedy wings.
paged_attention_unroll leads the way—
“Binky!” says Bunny, “let’s test today!”

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	It clearly summarizes the sync_to_sm batching change and the K=16 interval.
Description check	✅ Passed	It is directly related to the scheduler batching change and the added validation coverage.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.h`:
- Line 475: Update advance_ring_pointers() and its sync_to_sm() ordering so
ring->fc.last_task_alive is stored with last_task_alive once PUBLISH_INTERVAL_K
slots have been consumed, avoiding the extra-call throttle; when reuse resets
last_published_to_sm to 0, publish immediately after the local lifetime advance.
- Around line 467-478: Update the scheduler’s terminal, idle, and drained-work
paths around advance_ring_pointers() to force-publish the current
last_task_alive watermark, including sub-interval advances of 1–15 steps.
Preserve batched sync_to_sm() behavior for ongoing work, and add coverage for 1,
15, 16, and 17 advances to verify final publication and threshold batching.

In `@tools/benchmark_a5_case1.sh`:
- Around line 17-21: Update the option handling case for -d|--device and
-n|--rounds in the argument parser to validate that a following value exists
before reading $2 or executing shift 2. When either option is missing its value,
emit the script’s usage error and terminate consistently; preserve the existing
assignments and shifts for valid values.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: f3f19174-72e0-4e19-bda6-6ddebea50fd0

📥 Commits

Reviewing files that changed from the base of the PR and between b079a23 and 511a165.

📒 Files selected for processing (4)

• src/a5/platform/onboard/host/aicpu_topology_probe.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.h
• tools/benchmark_a5_case1.sh
• tools/benchmark_rounds.sh

[yanghaoran29]

A5 performance re-test

Re-tested paged_attention_unroll Case1 with 100 rounds on the same device (device 0), running the main baseline first and PR #1575 second under one task-submit lock.

Metric	Main baseline	PR #1575	Delta	Change
Effective	2061.2 us	1965.6 us	-95.6 us	-4.64%
Orch	1906.7 us	1799.1 us	-107.6 us	-5.64%
Sched	2056.8 us	1960.3 us	-96.5 us	-4.69%
Device	2092.9 us	1996.9 us	-96.0 us	-4.59%
Host	148042.1 us	148592.8 us	+550.7 us	+0.37%

Test details:

• Baseline: cb32999c (the main parent of the PR commit)
• PR: 526eb8bb
• PTO-ISA pin: 83d01313d9bfc247c4b7c8bcf969d1019f0d106f
• Device/Effective coefficient of variation: approximately 1.5%-1.9%, so the ~4.6% device-side improvement is above the observed run-to-run noise.

On this host, both CPU topology query paths return 65534, so unmodified main and the current PR cannot pass the AICPU affinity gate. For measurement only, the same OCCUPY-only topology fallback was applied to temporary baseline and PR worktrees. It was therefore a shared control variable; the only A/B code difference was the K=16 scheduler publication change. The fallback was not added back to this PR.