feat(a5): route DFX collectors through MemoryAllocator so committed_device_memory counts them

[ChaoZheng109]

What

Route the a5 onboard DFX collectors (l2_swimlane, args_dump, pmu,
scope_stats, dep_gen) through the per-runner MemoryAllocator instead of
raw rtMalloc / rtFree.

Why

committed_device_memory() (added in #1592) reports the device HBM tracked by
MemoryAllocator. On a2a3, the DFX collectors already allocate their device
buffers through mem_alloc_ (init_pmu / init_dep_gen / init_scope_stats),
so their HBM is counted. On a5 they used prof_alloc_cb / prof_free_cb —
raw rtMalloc(RT_MEMORY_HBM) — bypassing mem_alloc_, so a5 profiling buffers
were invisible to committed_device_memory().

That made the #1592 docstring ("excludes HCCL/VMM comm windows", which implies
profiling buffers are counted) correct on a2a3 but wrong on a5. This PR
closes that platform gap.

How

Replace the two file-static callbacks with capturing alloc_cb / free_cb
lambdas (mem_alloc_.alloc / mem_alloc_.free) at each collector init and in
finalize_collectors, matching a2a3's style. Both paths call the same
rtMalloc(RT_MEMORY_HBM) / rtFree underneath, so device-buffer properties are
unchanged — the only effect is that a5 DFX buffers are now tracked and counted,
including the runtime watermark replenishment.

Scope / testing

• Completes the a5 coverage deferred from feat: add committed_device_memory() API (L1 byte counter + L2/L3 query) #1592 to keep that PR single-purpose.
  Not part of [Code Health] Reconcile a2a3 vs a5 tensormap_and_ringbuffer runtime implementation divergence #1582 (that issue tracks tensormap_and_ringbuffer runtime-tree
  divergence and treats profiling instrumentation as already aligned); this is a
  platform-host-tree consistency fix stemming from feat: add committed_device_memory() API (L1 byte counter + L2/L3 query) #1592.
• a5 onboard DFX scene tests (test_pmu / test_dep_gen / test_scope_stats /
  test_args_dump) exercise the new alloc path.
• Verified the a5 onboard host_runtime.so compiles against this change.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: c1ed3041-0a55-44f8-be5d-edb1ac34b9bc

📥 Commits

Reviewing files that changed from the base of the PR and between ce81c97 and 834293e.

📒 Files selected for processing (1)

• src/a5/platform/onboard/host/device_runner.cpp

---

📝 Walkthrough

Walkthrough

Profiling collector initialization and finalization now use per-run allocation callbacks backed by mem_alloc_, replacing shared profiling allocator wrappers across the affected collectors.

Changes

Profiling allocator ownership

Layer / File(s)	Summary
Collector initialization callbacks src/a5/platform/onboard/host/device_runner.cpp	L2 swimlane, argument dump, PMU, scope stats, and dependency-generation collectors use mem_alloc_-backed allocation and free lambdas.
Collector finalization callbacks src/a5/platform/onboard/host/device_runner.cpp	Initialized collectors are finalized with a mem_alloc_.free callback, while scope-stats state cleanup remains in place.

Estimated code review effort: 2 (Simple) | ~10 minutes

Possibly related PRs

• hw-native-sys/simpler#1592: Adds MemoryAllocator bookkeeping used by the allocator path updated here.

Poem

A bunny hops through buffers bright,
With mem_alloc_ guiding every byte.
Collectors gather, then safely free,
Scope stats clear as neat as can be.
No shared wrappers in sight—
Hop, hop, memory right!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main change: routing a5 DFX collectors through MemoryAllocator for committed memory tracking.
Description check	✅ Passed	The description matches the change set and explains the motivation, implementation, and testing for the allocator routing.
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.

---

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.}