Refactor: host-build trb runtime arena

[poursoul]

Summary

Move the trb runtime arena's layout + data initialization from AICPU boot
onto host, so each AICPU launch reduces to a cheap pointer-fixup pass plus
the SM reset that can't run off-device. The pooled prebuilt image lives in
a per-Worker DeviceRunner pool and is reused across runs via a single
rtMemcpy — multi-launch boot cost drops from O(task_window_size) per
worker to a constant.

Two related cleanups ride along:

• RingSchedState::init's O(task_window_size) slot-bind loop is lifted
  into per-submit orch::prepare_task, making startup independent of
  window size. The two extra stores hit the same 64B cache line that
  prepare_task already dirties, so the per-submit cost is essentially
  free.
• AICPU SM reset (per-slot bind_ring + reset_for_reuse +
  fanin_count/active_mask zero) consolidated into
  PTO2SharedMemoryHandle::init_header_per_ring so the host-build path
  doesn't dereference SM.

Covers both a2a3 and a5 trb runtimes (platform layer onboard + sim).
hbg is unaffected by the runtime-arena split — its
setup_static_arena(...,0) keeps the third region unreserved.

Mechanism

• runtime_create_from_sm split into four phases that run on either side:
  • runtime_reserve_layout — pure arithmetic; host computes sub-region
    offsets on a libc-backed DeviceArena.
  • runtime_init_data_from_layout — writes standalone fields, memset's
    arena regions, and stores SM device pointers (only stores, no
    dereferences).
  • runtime_wire_arena_pointers — walks every arena-internal pointer
    field and binds it to arena.base() + offset. Idempotent: host runs
    once with the host mirror, AICPU runs once after attach with device
    addresses.
  • runtime_finalize_after_wire — AICPU-only fixup for s_runtime_ops
    (device-side file-local global) and the SPMD core counts from the
    SchedulerContext.
• DeviceArena::attach() wraps an externally-owned buffer with no
  per-attach allocation; re-attach is permitted so each AICPU boot can
  reuse the same pooled image. Pre-alignment / non-null / power-of-two
  checks std::abort() instead of assert() so release builds still
  trap on contract violations.
• pto2_sm_layout namespace computes SM device-side field addresses by
  pure offset arithmetic so host init never reads SM. Takes a per-ring
  task_window_sizes[] array (mirroring the SM API) and asserts
  ring_id in range — structurally prevents the host-built image from
  silently disagreeing with the SM layout.
• New runtime/shared/pto_runtime2_init.cpp holds the host-pluggable
  cold-path lifted from pto_runtime2.cpp / pto_orchestrator.cpp /
  scheduler/pto_scheduler.cpp. AICPU-only ops table / submit_task /
  dispatch / business logic stay in their original files.
• DeviceRunner now owns three independent pooled arenas —
  gm_heap_arena_, gm_sm_arena_, runtime_arena_pool_ — one
  device_malloc each. Split out from a single backing allocation
  because the combined size can exceed the device allocator's largest
  contiguous block. setup_static_arena(gm_heap_size, gm_sm_size, runtime_arena_size) commits each region independently;
  acquire_pooled_runtime_arena() returns nullptr when the region is
  unreserved (hbg's setup_static_arena(...,0) path) so misuse is loud,
  not undefined.
• bind_prepared_to_runtime_impl (host runtime_maker) does the full
  reserve_layout → init_data → wire on a host arena, stashes the layout
  inside the PTO2Runtime image at prebuilt_layout, then rtMemcpys
  the whole arena into the pooled device region.
• Dead fields and parameters dropped: PTO2TensorMap::orch back-pointer
  (never dereferenced), PTO2Runtime::prebuilt_arena_base mirror (host
  Runtime::prebuilt_arena_base_ is the real source of truth), unused
  task_window_size / dep_pool_capacity from
  PTO2SchedulerState::init_data_from_layout and
  RingSchedState::init_data_from_layout (scheduler only needs SM base
  • ring index, both window-size-independent).

Test plan

• cpput: 25/25 pass. ready_queue / spsc_queue / scheduler_state /
  task_state / wiring / tensormap UTs migrated to the data+wire API.
  task_allocator.init grew an optional initial_local_task_id
  (default 0) so the near-INT32_MAX corner case is still exercised
  without reading the SM.
• a5sim: L2 trb 21/21 + L2 host_build_graph 6/6 pass.
• a2a3sim: L2 trb 29/29 + L2 host_build_graph 9/9 pass.
• a2a3 hardware: tests/st/.../paged_attention_unroll passes on
  device 9 (--build, pto-isa commit pinned to CI).

[gemini-code-assist]

Code Review

This pull request implements a prebuilt-arena fast path for the PTO2 runtime, allowing the host to pre-compute the runtime arena image and upload it to the device. This optimization reduces AICPU boot time by replacing full initialization with a simple attachment and pointer "wiring" phase. Key changes include refactoring the initialization logic for the runtime, orchestrator, and scheduler into separate data-population and pointer-wiring stages, extending the DeviceRunner to manage a pooled runtime arena, and adding an attach method to DeviceArena for externally-owned buffers. Review feedback correctly identified potential undefined behavior in the new acquire_pooled_runtime_arena methods when the arena is not provisioned, suggesting defensive checks against SIZE_MAX offsets.

[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

Run ID: 82ac7cfe-a395-49b9-8341-50c746d76ed5

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

This PR introduces a prebuilt runtime arena fast path for PTO2 initialization. It refactors device memory management from a single monolithic arena into three independent pooled regions (GM heap, GM SM, runtime) and restructures the runtime lifecycle from a single creation call into four distinct phases: layout reservation, data initialization, pointer wiring, and finalization. The changes span platform layers, runtime initialization, orchestrator/scheduler/tensormap components, and all supporting tests.

Changes

Device Arena Infrastructure & Three-Region Pooled Refactoring

Layer / File(s)	Summary
DeviceArena attach() API src/common/device_comm/device_arena.h	DeviceArena gains attach() to bind externally owned, pre-aligned buffers without allocation, updating release behavior and tracking attached state.
a2a3 onboard three-region split src/a2a3/platform/onboard/host/device_runner.h, device_runner.cpp	DeviceRunner refactored from single static_arena_ to three independent DeviceArena pools (GM heap, GM SM, runtime) with idempotent commit_region helper, cached size tracking, and updated public API to accept runtime_arena_size.
a2a3 sim three-region split src/a2a3/platform/sim/host/device_runner.h, device_runner.cpp	Mirrors a2a3 onboard refactoring with per-region idempotent commit/release logic and acquire_pooled_runtime_arena() accessor.
a5 onboard three-region split src/a5/platform/onboard/host/device_runner.h, device_runner.cpp	a5 onboard implementation of three pooled arenas using shared commit_region helper, leaving runtime arena uncommitted when size is zero.
a5 sim three-region split src/a5/platform/sim/host/device_runner.h, device_runner.cpp	a5 sim variant with per-region commit logic and cached size tracking for reuse checks.
Host API wiring src/a2a3/platform/onboard/host/pto_runtime_c_api.cpp, src/a2a3/platform/sim/host/pto_runtime_c_api.cpp, src/a5/platform/onboard/host/pto_runtime_c_api.cpp, src/a5/platform/sim/host/pto_runtime_c_api.cpp, src/a2a3/runtime/host_build_graph/runtime/runtime.h, src/a5/runtime/host_build_graph/runtime/runtime.h	All platform host API interfaces updated to accept runtime_arena_size in setup_static_arena and wire new acquire_pooled_runtime_arena() function pointer into Runtime::host_api.

Runtime Lifecycle Phase Split & Layout Descriptors

Layer / File(s)	Summary
PTO2RuntimeArenaLayout & multi-phase API src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2.h, pto_runtime2.cpp, src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2.h, pto_runtime2.cpp, src/a2a3/runtime/tensormap_and_ringbuffer/runtime/runtime.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/runtime.h, src/a2a3/runtime/tensormap_and_ringbuffer/runtime/shared/runtime.cpp, src/a5/runtime/tensormap_and_ringbuffer/runtime/shared/runtime.cpp	Introduces PTO2RuntimeArenaLayout descriptor to cache prebuilt runtime arena offsets. Replaces single runtime_create_from_sm with four-phase API: runtime_reserve_layout, runtime_init_data_from_layout, runtime_wire_arena_pointers, runtime_finalize_after_wire. Updates Runtime to store prebuilt arena base and offset.
Scheduler & ready-queue phase split src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.h	Splits scheduler and ready-queue initialization from single init_from_layout into init_data_from_layout (write data without pointer binding) and wire_arena_pointers (idempotent pointer wiring). Updates RingSchedState to use device base pointer instead of SM header pointer.
Orchestrator & tensormap phase split src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.h, src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_tensormap.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_tensormap.h	Refactors orchestrator and tensormap initialization APIs to split data initialization from pointer wiring. Removes PTO2OrchestratorState forward declaration from tensormap header and updates method signatures to use device base pointer.
Task slot binding refactor src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2_types.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2_types.h, src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_ring_buffer.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_ring_buffer.h	Splits PTO2TaskSlotState::bind(payload, task, ring_id) into bind_ring(ring_id) for one-time slot-invariant binding and bind_buffers(payload, task) for per-submit rebinding. Updates PTO2TaskAllocator::init to accept optional initial_local_task_id seed parameter.
Shared memory layout address helpers src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_shared_memory.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_shared_memory.h	Introduces pto2_sm_layout namespace with inline noexcept helpers computing shared-memory field addresses via offset arithmetic. Aligns current_task_index to 64-byte boundary and clarifies per-boot SM reset coupling with allocator initialization.

Centralized Shared Runtime Arena Initialization Implementation

Layer / File(s)	Summary
pto_runtime2_init.cpp implementation src/a2a3/runtime/tensormap_and_ringbuffer/runtime/shared/pto_runtime2_init.cpp, src/a5/runtime/tensormap_and_ringbuffer/runtime/shared/pto_runtime2_init.cpp	Complete implementation of ready-queue, scheduler, and orchestrator layout reservation, data initialization, pointer wiring, and teardown. Allocates SM handle, computes sub-region layout, initializes runtime/SM/orchestrator/scheduler data, wires top-level pointers, and provides destroy routines that null pointers without freeing arena.
Remove orchestrator/scheduler init from old locations src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.cpp, src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.cpp, src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.cpp, src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.cpp	Removes orchestrator and scheduler initialization method definitions from pto_orchestrator.cpp and pto_scheduler.cpp. Logic now centralized in pto_runtime2_init.cpp.
Tensormap & shared memory init updates src/a2a3/runtime/tensormap_and_ringbuffer/runtime/shared/pto_tensormap.cpp, src/a5/runtime/tensormap_and_ringbuffer/runtime/shared/pto_tensormap.cpp, src/a2a3/runtime/tensormap_and_ringbuffer/runtime/shared/pto_shared_memory.cpp, src/a5/runtime/tensormap_and_ringbuffer/runtime/shared/pto_shared_memory.cpp, src/a2a3/runtime/tensormap_and_ringbuffer/host/dep_gen_replay.cpp	Refactors PTO2TensorMap::init_from_layout into split init_data_from_layout and wire_arena_pointers methods. Updates PTO2SharedMemoryHandle::init_header_per_ring to initialize per-ring slot_states. Updates dep_gen_replay.cpp to use split initialization API.

Host-Side Prebuilt Arena Construction & Device-Side Attachment

Layer / File(s)	Summary
Host-side prebuilt arena image construction src/a2a3/runtime/tensormap_and_ringbuffer/host/runtime_maker.cpp, src/a5/runtime/tensormap_and_ringbuffer/host/runtime_maker.cpp	Updates bind_prepared_to_runtime_impl to compute runtime arena layout on host, reserve via runtime_reserve_layout, commit host arena, initialize prebuilt image using split methods, upload image to device, and register via set_prebuilt_arena. Adds timing instrumentation for prebuilt arena upload.
AICPU device-side prebuilt arena attachment src/a2a3/runtime/tensormap_and_ringbuffer/aicpu/aicpu_executor.cpp, src/a5/runtime/tensormap_and_ringbuffer/aicpu/aicpu_executor.cpp	Updates AicpuExecutor::run to fetch prebuilt arena base and runtime offset, attach runtime_arena_ to external memory, reinterpret prebuilt PTO2Runtime, wire arena-internal pointers, reset SM and mailbox, and finalize using runtime_finalize_after_wire. Removes runtime_create_from_sm usage.
Runtime accessors & orchestrator per-task slot rebinding src/a2a3/runtime/tensormap_and_ringbuffer/runtime/runtime.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/runtime.h, src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.cpp, src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.cpp, src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.h, src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.h	Adds Runtime::set_prebuilt_arena, get_prebuilt_arena_base, get_prebuilt_runtime_offset accessors. Updates prepare_task to explicitly rebind slot_state buffer pointers via bind_buffers() before scheduler queue wiring. Also removes per-submit fanin insertion guard and increases RUNTIME_MAX_ORCH_SO_SIZE from 1MB to 4MB.

Unit Test Updates for Two-Phase Initialization APIs

Layer / File(s)	Summary
a2a3 test suite updates tests/ut/cpp/a2a3/test_ready_queue.cpp, test_scheduler_state.cpp, test_spsc_queue.cpp, test_task_allocator.cpp, test_task_state.cpp, test_tensormap.cpp, test_wiring.cpp	Updates all a2a3 runtime tests to initialize scheduler/queue data via init_data_from_layout followed by wire_arena_pointers. Updates task allocator test to pass initial_local_task_id seed.
a5 test suite updates tests/ut/cpp/a5/test_ready_queue.cpp, test_scheduler_state.cpp, test_spsc_queue.cpp, test_task_allocator.cpp, test_task_state.cpp, test_tensormap.cpp, test_wiring.cpp	Updates a5 tests to use arena-backed layout reservation with two-phase initialization. Adds DeviceArena members to fixtures and updates all initialization calls to split data initialization from pointer wiring.
CMakeLists.txt build system updates tests/ut/cpp/CMakeLists.txt	Adds pto_runtime2_init.cpp to a2a3_rt_objs sources. Creates new a5_rt_objs OBJECT library with a5 runtime sources. Adds add_a5_runtime_test helper and registers all a5 runtime unit test targets.

🎯 4 (Complex) | ⏱️ ~60 minutes

🐰 Three arenas bloom in springtime fair,
Host constructs, device attaches with care,
Pointers wired in phases precise,
Prebuilt images—no need to build twice!
Tests now dance in sync so bright,
Ring buffers bound both left and right. 🌱

[coderabbitai]

Actionable comments posted: 6

🧹 Nitpick comments (2)

src/a2a3/platform/onboard/host/device_runner.h (1)

210-220: ⚡ Quick win

The pointer-stability contract is stronger than the implementation.

setup_static_arena() can re-commit a region when a later request grows, so that region's base() is not guaranteed to stay stable until finalize(). Please document that callers must reacquire any region that was successfully resized.

🤖 Prompt for 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.

In `@src/a2a3/platform/onboard/host/device_runner.h` around lines 210 - 220, The
current comment claiming that pointers from acquire_pooled_runtime_arena()/the
pooled GM heap/PTO2 SM/runtime arena are stable until finalize() is incorrect
because setup_static_arena() may re-commit (resize) a region and change its
base(); update the documentation in device_runner.h so it clearly states that
pointers returned by
acquire_pooled_runtime_arena()/acquire_pooled_gm_heap()/acquire_pooled_pto2_sm()
are stable only until a subsequent call to setup_static_arena() that grows or
re-commits the region, and that callers must re-acquire the region (call the
appropriate acquire_* function and use the returned base()) after any resize;
mention the specific symbols setup_static_arena, acquire_pooled_runtime_arena,
finalize, and base() to make the requirement obvious.

src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_shared_memory.h (1)

214-257: ⚡ Quick win

Prefer uintptr_t for opaque device-address arithmetic.

Lines 214-257 are doing raw device-address math via char *. That works on today's targets, but it treats an opaque device pointer token as if it were a real host object pointer. Casting through uintptr_t keeps this as pure address arithmetic and avoids UB/sanitizer surprises in host builds.

Suggested change

+inline std::uintptr_t sm_byte_addr(void *sm_dev_base, std::size_t off) noexcept {
+    return reinterpret_cast<std::uintptr_t>(sm_dev_base) + off;
+}
+
 inline std::atomic<int32_t> *orch_error_code_addr(void *sm_dev_base) noexcept {
     return reinterpret_cast<std::atomic<int32_t> *>(
-        static_cast<char *>(sm_dev_base) + offsetof(PTO2SharedMemoryHeader, orch_error_code)
+        sm_byte_addr(sm_dev_base, offsetof(PTO2SharedMemoryHeader, orch_error_code))
     );
 }

🤖 Prompt for 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.

In `@src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_shared_memory.h` around
lines 214 - 257, The code does pointer arithmetic on sm_dev_base via char* which
can be UB; change all address computations in orch_error_code_addr,
ring_header_addr, ring_current_task_index_addr, ring_last_task_alive_addr, and
ring_task_descriptors_addr to perform arithmetic on a uintptr_t base (e.g.
uintptr_t base = reinterpret_cast<uintptr_t>(sm_dev_base)), add offsetof/size
offsets to that integer, and only at the end reinterpret_cast the final integer
(cast to void* first if needed) to the desired pointer type; ensure all uses of
static_cast<char*> are replaced with integer arithmetic and final
reinterpret_cast to preserve opaque device-address semantics.

🤖 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/platform/onboard/host/device_runner.cpp`:
- Around line 290-300: The failure path in the device runner currently releases
previously committed peer arenas (gm_heap_arena_.release(),
gm_sm_arena_.release()) when commit_region(runtime_arena_pool_, ...) fails,
which can invalidate pooled pointers; change the error handling so that on a
failed commit of runtime_arena_pool_ you only undo or reset state related to
runtime_arena_pool_ and its cached_runtime_arena_size_ (do not release
gm_heap_arena_ or gm_sm_arena_ or touch
cached_gm_heap_size_/cached_gm_sm_size_), and conversely if
commit_region(gm_sm_arena_, ...) fails only undo gm_sm_arena_ related state;
keep commit_region successes independent so callers holding pooled pointers are
not invalidated by later failures.

In `@src/a2a3/platform/sim/host/device_runner.cpp`:
- Around line 1056-1065: The device buffer pointed to by device_wall_dev_ptr_
must be freed via free_tensor() (which uses mem_alloc_.free()) before calling
mem_alloc_.finalize(); move the free_tensor(device_wall_dev_ptr_) call so it
executes prior to mem_alloc_.finalize(), then reset device_wall_dev_ptr_ to
nullptr after freeing to avoid dangling pointers; update the same ordering
wherever the block around gm_heap_arena_.release()/runtime_arena_pool_.release()
and mem_alloc_.finalize() appears (the other 1067-1075 occurrence) to ensure no
free goes through the allocator after finalize.

In `@src/a2a3/runtime/tensormap_and_ringbuffer/host/runtime_maker.cpp`:
- Around line 284-285: The code narrows runtime->dep_pool_size (uint64_t) to
int32_t in eff_dep_pool_capacity without checking upper bounds; clamp the
selected value to int32_t range before casting. Replace the ternary cast with:
determine uint64_t chosen = runtime->dep_pool_size ? runtime->dep_pool_size :
PTO2_DEP_LIST_POOL_SIZE; clamp chosen = std::min(chosen,
static_cast<uint64_t>(std::numeric_limits<int32_t>::max())); then set
eff_dep_pool_capacity = static_cast<int32_t>(chosen); (ensure <limits> is
included). Use the symbols runtime->dep_pool_size, PTO2_DEP_LIST_POOL_SIZE, and
eff_dep_pool_capacity to locate the change.

In `@src/a5/platform/onboard/host/device_runner.cpp`:
- Around line 1006-1015: Move the release of device_wall_dev_ptr_ so it occurs
before the allocator shutdown (mem_alloc_.finalize()) and before the device
reset; specifically, locate device_wall_dev_ptr_ and call its release/cleanup
prior to the gm_heap_arena_.release()/mem_alloc_.finalize() sequence and before
the device reset call, and ensure free_tensor() invocations happen while the
device context and allocator are still valid (i.e., call free_tensor() or other
tensor cleanup before mem_alloc_.finalize() and device reset). Reference
symbols: device_wall_dev_ptr_, mem_alloc_.finalize(),
gm_heap_arena_.release()/gm_sm_arena_.release()/runtime_arena_pool_.release(),
free_tensor(), and the device reset invocation to reorder cleanup safely.

In `@src/a5/platform/sim/host/device_runner.cpp`:
- Around line 154-164: The second failure branch currently tears down
previously-committed peer arenas (gm_heap_arena_, gm_sm_arena_) which can leave
pooled pointers dangling; change the runtime_arena_pool_ failure handling so it
does NOT release gm_heap_arena_ or gm_sm_arena_. Specifically, in the
commit_region(runtime_arena_pool_, cached_runtime_arena_size_,
runtime_arena_size) != 0 branch, only undo/cleanup the runtime_arena_pool_
allocation (release or reset whatever was allocated for runtime_arena_pool_ and
set cached_runtime_arena_size_ = 0) and return -1, leaving gm_heap_arena_,
cached_gm_heap_size_, gm_sm_arena_, and cached_gm_sm_size_ intact; keep the
existing behavior in the first branch (commit_region(gm_sm_arena_, ...) failure)
as-is.

In `@tests/ut/cpp/a5/test_tensormap.cpp`:
- Around line 108-119: The test InitRequiresPowerOfTwoBuckets currently uses a
power-of-two bucket count (8) so it doesn't verify the rejection path; change it
to either (A) become an explicit success-path smoke test by renaming the test
(e.g., InitWithValidBucketsSucceeds) and keep the current assertions around
PTO2TensorMap::reserve_layout, bad.init_data_from_layout and
bad.wire_arena_pointers, or (B) restore the rejection check by using a
non-power-of-two bucket count (e.g., 12) and gate a death/assertion expectation
around the failing call (wrap ASSERT_DEATH or equivalent around
PTO2TensorMap::reserve_layout/init_data_from_layout) so it only runs when
assertions are enabled (use preprocessor guard for debug builds). Ensure
references remain to PTO2TensorMap::reserve_layout,
PTO2TensorMap::init_data_from_layout, bad.wire_arena_pointers, and the test name
InitRequiresPowerOfTwoBuckets when making the change.

---

Nitpick comments:
In `@src/a2a3/platform/onboard/host/device_runner.h`:
- Around line 210-220: The current comment claiming that pointers from
acquire_pooled_runtime_arena()/the pooled GM heap/PTO2 SM/runtime arena are
stable until finalize() is incorrect because setup_static_arena() may re-commit
(resize) a region and change its base(); update the documentation in
device_runner.h so it clearly states that pointers returned by
acquire_pooled_runtime_arena()/acquire_pooled_gm_heap()/acquire_pooled_pto2_sm()
are stable only until a subsequent call to setup_static_arena() that grows or
re-commits the region, and that callers must re-acquire the region (call the
appropriate acquire_* function and use the returned base()) after any resize;
mention the specific symbols setup_static_arena, acquire_pooled_runtime_arena,
finalize, and base() to make the requirement obvious.

In `@src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_shared_memory.h`:
- Around line 214-257: The code does pointer arithmetic on sm_dev_base via char*
which can be UB; change all address computations in orch_error_code_addr,
ring_header_addr, ring_current_task_index_addr, ring_last_task_alive_addr, and
ring_task_descriptors_addr to perform arithmetic on a uintptr_t base (e.g.
uintptr_t base = reinterpret_cast<uintptr_t>(sm_dev_base)), add offsetof/size
offsets to that integer, and only at the end reinterpret_cast the final integer
(cast to void* first if needed) to the desired pointer type; ensure all uses of
static_cast<char*> are replaced with integer arithmetic and final
reinterpret_cast to preserve opaque device-address semantics.

🪄 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

Run ID: afb4c479-cd23-4812-8bd2-97bc8283d600

📥 Commits

Reviewing files that changed from the base of the PR and between 324df3d and c22ca95.

📒 Files selected for processing (65)

• src/a2a3/platform/onboard/host/device_runner.cpp
• src/a2a3/platform/onboard/host/device_runner.h
• src/a2a3/platform/onboard/host/pto_runtime_c_api.cpp
• src/a2a3/platform/sim/host/device_runner.cpp
• src/a2a3/platform/sim/host/device_runner.h
• src/a2a3/platform/sim/host/pto_runtime_c_api.cpp
• src/a2a3/runtime/host_build_graph/runtime/runtime.h
• src/a2a3/runtime/tensormap_and_ringbuffer/aicpu/aicpu_executor.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/host/dep_gen_replay.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/host/runtime_maker.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_ring_buffer.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2_types.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_shared_memory.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/pto_tensormap.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/runtime.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.h
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/shared/pto_runtime2_init.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/shared/pto_shared_memory.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/shared/pto_tensormap.cpp
• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/shared/runtime.cpp
• src/a5/platform/onboard/host/device_runner.cpp
• src/a5/platform/onboard/host/device_runner.h
• src/a5/platform/onboard/host/pto_runtime_c_api.cpp
• src/a5/platform/sim/host/device_runner.cpp
• src/a5/platform/sim/host/device_runner.h
• src/a5/platform/sim/host/pto_runtime_c_api.cpp
• src/a5/runtime/host_build_graph/runtime/runtime.h
• src/a5/runtime/tensormap_and_ringbuffer/aicpu/aicpu_executor.cpp
• src/a5/runtime/tensormap_and_ringbuffer/host/runtime_maker.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_orchestrator.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_ring_buffer.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_runtime2_types.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_shared_memory.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/pto_tensormap.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/runtime.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.h
• src/a5/runtime/tensormap_and_ringbuffer/runtime/shared/pto_runtime2_init.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/shared/pto_shared_memory.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/shared/pto_tensormap.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/shared/runtime.cpp
• src/common/device_comm/device_arena.h
• tests/ut/cpp/CMakeLists.txt
• tests/ut/cpp/a2a3/test_ready_queue.cpp
• tests/ut/cpp/a2a3/test_scheduler_state.cpp
• tests/ut/cpp/a2a3/test_spsc_queue.cpp
• tests/ut/cpp/a2a3/test_task_allocator.cpp
• tests/ut/cpp/a2a3/test_task_state.cpp
• tests/ut/cpp/a2a3/test_tensormap.cpp
• tests/ut/cpp/a2a3/test_wiring.cpp
• tests/ut/cpp/a5/test_ready_queue.cpp
• tests/ut/cpp/a5/test_scheduler_state.cpp
• tests/ut/cpp/a5/test_spsc_queue.cpp
• tests/ut/cpp/a5/test_task_allocator.cpp
• tests/ut/cpp/a5/test_task_state.cpp
• tests/ut/cpp/a5/test_tensormap.cpp
• tests/ut/cpp/a5/test_wiring.cpp

💤 Files with no reviewable changes (2)

• src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.cpp
• src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/pto_scheduler.cpp