Refactor: derive tmr scheduler thread/core dims from platform_config

[ChaoZheng109]

Fixes #1734

Three hardware-derived dimensions in the tensormap_and_ringbuffer scheduler were written as bare numeric literals even though platform_config.h already defines the exact named constant they equal. The literals obscured that a2a3 and a5 share one rule, read as divergences when they are not, and would go silently wrong if a platform constant changed. Landed across both trees in one commit per codestyle.md rule 10.

No behavior change — every derived value equals the current literal on both arches today (PLATFORM_MAX_AICPU_THREADS = 4/5, PLATFORM_CORES_PER_BLOCKDIM = 3, PLATFORM_MAX_CORES = 72/108).

Changes

G — dummy-drain thread gate (scheduler_dispatch.cpp)
thread_idx < 3 (a2a3) / thread_idx < 4 (a5) → thread_idx < PLATFORM_MAX_AICPU_THREADS - 1 — "every scheduler thread except the last".

M1 — RUNTIME_MAX_WORKER (runtime.h + scheduler_context.h)
72 / 108 → PLATFORM_MAX_CORES. Both the authoritative unguarded define in runtime.h and the #ifndef fallback in scheduler_context.h are changed: the runtime.h define is unguarded, so leaving it a literal while the fallback expands to the constant is a mismatched macro redefinition that -Werror rejects. (The issue's Location listed only scheduler_context.h; runtime.h had to move with it for the tree to compile.)

M2 — cores-per-cluster 3 (scheduler_types.h + scheduler_cold_path.cpp)
The semantic cluster_count_ * 3 (core_num()), the two cores_total_num_ / 3, and the % 3 != 0 validity check → PLATFORM_CORES_PER_BLOCKDIM; the "multiple of 3" error text now interpolates the constant with %d. The self-documenting bit(i*3+k) / core_id_map_[cluster_idx*3+k] bitmask-layout arithmetic is left as-is, per the issue's scope note.

Both trees now read identically at every changed site.

Testing

Rebuilt both runtimes (pip install --no-build-isolation -e .) and ran tmr sim scene tests on both arches:

Test	a2a3sim	a5sim	Exercises
dummy_task	✅	✅	G dummy-drain gate
available_aicore_counts	✅	✅	M2 core-count / multiple-of-3 validation
mixed_example	✅	✅	general scheduling / dispatch

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 787c851e-f836-404b-be27-062c803b0466

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

The scheduler runtime now derives worker capacity, core-layout calculations, validation, and dummy-drain thread eligibility from platform configuration constants in both the a2a3 and a5 trees.

Changes

Scheduler dimension alignment

Layer / File(s)	Summary
Platform-derived worker capacity src/a2a3/.../runtime.h, src/a2a3/.../scheduler/scheduler_context.h, src/a5/.../runtime.h, src/a5/.../scheduler/scheduler_context.h	RUNTIME_MAX_WORKER now uses PLATFORM_MAX_CORES instead of fixed platform-specific values.
Platform-derived core layout src/{a2a3,a5}/.../scheduler/scheduler_cold_path.cpp, src/{a2a3,a5}/.../scheduler/scheduler_types.h	Core counts, cluster sizing, and divisibility validation now use PLATFORM_CORES_PER_BLOCKDIM.
Platform-derived dummy-drain gate src/{a2a3,a5}/.../scheduler/scheduler_dispatch.cpp	Dummy completion draining now runs on every AICPU scheduler thread except the final platform thread.

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

Possibly related PRs

• hw-native-sys/simpler#1679: Modified the a5 dummy-drain scheduling logic that this PR generalizes.
• hw-native-sys/simpler#1477: Updated related CoreTracker and scheduler sizing logic.
• hw-native-sys/simpler#1618: Changed scheduler thread and core-count assumptions in the runtime.

Poem

A rabbit hops through cores in line,
With platform constants keeping time.
Three becomes a named decree,
Threads drain neatly, wild and free.
Worker bounds now fit the design—
Thump, thump, the schedulers align!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the refactor to derive scheduler thread and core dimensions from platform configuration.
Description check	✅ Passed	The description directly explains the platform-constant refactor, affected files, preserved behavior, and validation results.
Linked Issues check	✅ Passed	The changes satisfy issue #1734 by replacing the specified scheduler magic literals across both architecture trees while preserving bitmask arithmetic.
Out of Scope Changes check	✅ Passed	All changes support issue #1734; the runtime.h updates are necessary to keep the authoritative worker-limit definitions consistent.
Docstring Coverage	✅ Passed	Docstring coverage is 85.71% which is sufficient. The required threshold is 80.00%.

---

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: 1

🤖 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/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_cold_path.cpp`:
- Line 903: Make semantic cluster-width handling consistent with
PLATFORM_CORES_PER_BLOCKDIM in assign_own_clusters() and
assign_cores_to_threads() across both files:
src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_cold_path.cpp
at lines 903, 803, and 1154-1158, and
src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_cold_path.cpp
at lines 900, 800, and 1135-1139. Replace remaining c * 3, sub < 3, N/3, and
multiple-of-3 logic and diagnostics with the platform-derived width, or isolate
any intentional fixed bitmask offsets behind a named constant with an explicit
invariant; keep AIC-count division and downstream traversal aligned.

🪄 Autofix

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: 441f9e08-fd2d-4706-a164-d54d3d15babd

📥 Commits

Reviewing files that changed from the base of the PR and between 2990441 and eaa120d.

📒 Files selected for processing (10)

• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/runtime.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_cold_path.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_context.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_dispatch.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_types.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/runtime.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_cold_path.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_context.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_dispatch.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_types.h