Fix: parallelize A5 dummy task resolution

[yanghaoran29]

Summary

• let all four default A5 scheduler threads drain dependency-only tasks
• reduce the per-thread dummy batch from 16 to 8 to distribute queue work
• rely on the MPMC ready queue and per-slot atomic fanout path for safe concurrent resolution

Performance

Measured sequentially on the same Ascend950PR device through task-submit; both variants use batch 8, default aicpu_thread_num=5, independent worktrees/venvs, and 100 rounds per workload.

Workload	Metric	1 scheduler	4 schedulers	Change
4096 dummy, shared producer	Effective	4369.7 us	4254.3 us	-2.64%
4096 dummy, shared producer	Sched	4363.7 us	4247.0 us	-2.67%
4096 immediately-ready dummy	Effective	4558.5 us	4533.4 us	-0.55%
4096 immediately-ready dummy	Sched	4551.6 us	4526.4 us	-0.55%

The shared-producer workload shows a modest improvement. The immediately-ready result is within the +/-2% noise margin; shared queue and per-slot atomic contention limit scaling.

Testing

• targeted pre-commit on scheduler_dispatch.cpp: passed
• A5Sim dummy scene test after rebasing to latest main: passed
• Ascend950PR onboard dummy scene test through task-submit: passed
• two 4096-dummy hardware benchmarks, 100 rounds per single/four-scheduler variant

Addresses A3 of #1582.

[coderabbitai]

📝 Walkthrough

Walkthrough

The scheduler now drains dummy dependency-only tasks on scheduler threads 0–3 in batches of eight. The dummy-task test matrix removes per-case AICPU thread settings and adds a dense fanout/fanin case.

Changes

Dummy-task scheduler execution

Layer / File(s)	Summary
Scheduler ready-queue draining src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_dispatch.cpp	Dummy dependency-only tasks are drained on scheduler threads 0–3. The per-iteration batch size changes from 16 to 8.
Dummy-task test coverage tests/st/a5/tensormap_and_ringbuffer/dummy_task/test_dummy_task.py	Existing cases remove per-case aicpu_thread_num settings. Case 4 adds DenseFanoutFanin.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• hw-native-sys/simpler#1618: Both distribute dummy-task ready-queue draining across multiple scheduler threads.
• hw-native-sys/simpler#1619: Both use four scheduler threads for dummy-task processing.
• hw-native-sys/simpler#1298: Both cover dense fanout/fanin behavior and scheduler ready-queue processing.

Poem

A rabbit watched four schedulers run,
While eight-task batches moved as one.
Dense fanout joined the queue,
With fanin waiting neatly too.
“Hop!” said Bunny, “the tests are done.”

🚥 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	The title clearly and concisely describes the main change: parallelizing A5 dummy task resolution.
Description check	✅ Passed	The description directly explains the scheduler changes, queue behavior, benchmarks, and tests.

---

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.}