Fix: a2a3 onboard AICPU never enables dep_gen capture by hw-native-sys-bot · Pull Request #742 · hw-native-sys/simpler

hw-native-sys-bot · 2026-05-12T01:06:57Z

Summary

src/a2a3/platform/onboard/aicpu/kernel.cpp propagates dump_tensor / l2_swimlane / pmu from k_args->enable_profiling_flag on every kernel entry, but the two equivalent setters for dep_gen were missing — so on hardware is_dep_gen_enabled() stayed false, dep_gen_aicpu_record_submit early-returned for every submit, and the host replay wrote {\"version\":1,\"edges\":[]} for every run. Sim works because its host device_runner already dlsym's set_platform_dep_gen_base + set_dep_gen_enabled into the AICPU .so; onboard had no equivalent wiring.
The PR Add: dep_gen replay → deps.json + swimlane integration + fanout ⊆ deps gate #737 fanout ⊆ deps gate was therefore vacuously passing on the only path where it ran (sim) and would silently FAIL with the entire fanout set on hardware. dep_gen smoke st was sim-only (CASES[0].platforms == [\"a2a3sim\"]), so this never surfaced in CI.

Changes

aicpu/kernel.cpp: include dep_gen_collector_aicpu.h and add set_platform_dep_gen_base(k_args->dep_gen_data_base) + set_dep_gen_enabled(GET_PROFILING_FLAG(k_args->enable_profiling_flag, PROFILING_FLAG_DEP_GEN)) next to the existing dump_tensor / l2_swimlane / pmu propagation.
tests/st/a2a3/tensormap_and_ringbuffer/dep_gen_capture/test_dep_gen_capture.py: add \"a2a3\" to the case's platforms list and override test_run so the pytest path invokes _post_validate when --enable-dep-gen is passed. Without the override the pytest collection ran orchestration but never asserted deps.json — a future regression of the same shape would still slip through.
.github/workflows/ci.yml: add a dedicated dep_gen smoke step to the st-onboard-a2a3 job that runs the capture test with --enable-dep-gen --enable-l2-swimlane, so the fanout ⊆ deps gate exercises hardware on every PR touching a2a3.

Test plan

Local rebuild of a2a3 onboard runtime is clean (build_runtimes.py --platform a2a3).
python -m pytest tests/st/a2a3/tensormap_and_ringbuffer/dep_gen_capture --platform a2a3 --enable-dep-gen --enable-l2-swimlane passes on a2a3 hardware.
Resulting outputs/TestDepGenCapture_default_*/deps.json contains the six expected edges (t0→t1, t0→t2, t1→t3, t2→t3, t0→t4, t3→t4) — matches example_orchestration.cpp exactly.
CI st-onboard-a2a3 job's new "Run dep_gen smoke" step passes once this PR runs through CI.
CI st-sim-a2a3 continues to pass — only the test case's platforms list changes for sim (no behavior change on sim).

Context

Found while running a sweep to look for hardware cases where |deps - fanout| > 0 (the kind of race-window evidence dep_gen was designed to surface). Every onboard deps.json came back empty, which traced to the missing wiring in aicpu/kernel.cpp. The sweep needs to be re-run after this lands.

🤖 Generated with Claude Code

`src/a2a3/platform/onboard/aicpu/kernel.cpp` propagates dump_tensor / l2_swimlane / pmu from `k_args->enable_profiling_flag` to AICPU globals on every kernel entry, but the two equivalent calls for dep_gen were missing — so `is_dep_gen_enabled()` stayed false on hardware, the orchestrator's `dep_gen_aicpu_record_submit` early-returned for every submit, the ring stayed empty, and the host replay wrote `{"version":1,"edges":[]}` for every run. The fanout ⊆ deps gate introduced in PR hw-native-sys#737 then ran against an empty deps set — vacuously passing only when fanout was also empty, and silent-FAILing otherwise. Sim was unaffected: its host device_runner already dlsym's `set_platform_dep_gen_base` and `set_dep_gen_enabled` into the AICPU .so. Onboard had no equivalent wiring. Changes - `aicpu/kernel.cpp`: include `dep_gen_collector_aicpu.h` and add the two missing setters next to the existing dump_tensor / l2_swimlane / pmu propagation. - `dep_gen_capture/test_dep_gen_capture.py`: add `"a2a3"` to the case's platforms list (was sim-only — the reason this never surfaced in CI) and override `test_run` so the pytest path actually invokes `_post_validate` when `--enable-dep-gen` is passed. Without that override, pytest collected the test and ran orchestration but never asserted anything about deps.json, so a future regression of the same shape would still slip through. - `.github/workflows/ci.yml`: add a dedicated dep_gen smoke step to the st-onboard-a2a3 job that runs the capture test with `--enable-dep-gen --enable-l2-swimlane`, so the fanout ⊆ deps gate exercises hardware. Verified on a2a3 onboard: `deps.json` now contains the six edges documented in `example_orchestration.cpp` (t0->t1, t0->t2, t1->t3, t2->t3, t0->t4, t3->t4), the smoke test passes with the gate flags, and the broader scene-test sweep is unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

gemini-code-assist

Code Review

This pull request enables the dep_gen profiling feature in the AICPU kernel by integrating the collector header and initializing the base address and enablement flags from kernel arguments. Additionally, the test suite is updated to support the a2a3 platform and includes an override for test_run to ensure that dep_gen output is validated when enabled. Feedback suggests verifying that profiling is correctly handled when the number of execution rounds is greater than one, as per profiling guidelines.

Addresses gemini-code-assist on PR hw-native-sys#742: the override in TestDepGenCapture.test_run re-read --enable-dep-gen directly from request.config and ignored the framework's --rounds > 1 disable rule, so `pytest --enable-dep-gen --rounds 2` would run super().test_run() without dep_gen (no deps.json produced), then call _post_validate which asserts deps.json exists — turning a silent regression into a confusing AssertionError. Add SceneTestCase._effective_enable_dep_gen(request, *, warn=False) as the single source of truth for the resolved flag (CLI value AND the rounds-guard). The framework's test_run consumes it with warn=True (it owns the user-facing "disabled because rounds > 1" message); the subclass override consumes it with warn=False (super() already warned). The other three profiling flags (l2_swimlane / dump_tensor / pmu) keep their inline rounds-guard — no subclass needs them yet, so factoring those out would be premature. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Follow-up to hw-native-sys#737 addressing post-merge review findings: 1. **Dedup replay fanin per-successor** ``dep_gen_replay_emit_deps_json`` now mirrors the runtime's ``PTO2FaninBuilder::append_fanin_or_fail`` semantics: an ``std::unordered_set<uint64_t>`` tracks predecessor task ids seen so far for the current successor, and both STEP 1 (``explicit_deps``) and STEP 3 (creator retention + tensormap lookup) push through a single ``emit_unique`` lambda. Previously an ``explicit_dep`` that the tensormap also surfaced (via ``owner_task_id`` or an overlap hit) emitted two edges, which double-counted ``deps.json`` and made ``swimlane_converter.py`` draw duplicate flow events. 2. **Document the OUTPUT-slot safety contract at the capture site** ``dep_gen_replay.cpp`` sets ``tref_buf[i].ptr`` for every captured tensor slot including OUTPUT — the on-disk blob for OUTPUT is zeroed by the AICPU writer. Added an inline comment pointing at ``pto_dep_compute.h``'s per-tag dispatch (which is what makes the never-dereferenced-on-OUTPUT contract hold) so the next reader of the arg_types width-fix area doesn't have to re-derive it. 3. **``make_deps_json_path`` helper** Both onboard + sim device_runner used to build ``deps.json`` with ``output_prefix_ + "/deps.json"`` inline — out of step with the ``make_<feature>_path()`` convention shared by PMU and (previously) submit_trace. Added ``make_deps_json_path`` in ``dep_gen_collector.h``; both call sites now go through it, and the helper also handles ``create_directories`` so the path is safe even when the output dir hasn't been touched by anything else yet. 4. **``_task_id(ring, local)`` helper in the validation test** The 6-edge expected-set in ``test_dep_gen_capture.py`` was open- coded ``1 << 32`` arithmetic at every call. One helper, layout stated once. 5. **``docs/dep_gen.md``** First user-facing doc for the feature. Covers motivation (links hw-native-sys#599), enable flags, ``deps.json`` format, ``deps_to_graph.py`` usage + node shape/color legend (AIC cube box, AIV vector ellipse, mix diamond, alloc dashed note), the ``fanout ⊆ deps`` validation gate, and the architecture touchpoints. 6. **CI smoke step for dep_gen** (``.github/workflows/ci.yml``) The default ``pytest tests/st`` invocation does not pass any DFX flag, so the dep_gen capture path never executed in CI. Added a second pytest step in ``st-sim-a2a3`` that re-runs only ``test_dep_gen_capture`` with ``--enable-dep-gen --enable-l2-swimlane``, which forces the full capture → replay → deps.json → fanout ⊆ deps gate path to execute. Same pattern is the model for future PMU / tensor_dump / swimlane smoke steps once those grow dedicated tests. 7. **UT roundtrip for ``enable_dep_gen``** (``tests/ut/py/test_chip_worker.py``) Adds the missing nanobind setter / getter / repr roundtrip for the ``enable_dep_gen`` ``CallConfig`` field, alongside the existing round-trips for ``enable_l2_swimlane`` / ``enable_dump_tensor`` / ``enable_pmu``. Catches binding-ABI regressions at the UT layer (e.g. the ``bool`` vs ``int32`` Python-wrapper bug that broke 9 st jobs earlier in the PR series — 10 s ut would have caught it). 8. **Pytest-friendly test hook via ``test_run`` override** instead of a framework-level ``_post_validate`` callback. ``test_dep_gen_capture`` overrides the inherited ``SceneTestCase.test_run`` to call ``super()`` then walk its cases and assert the dep_gen artifacts. Keeps the framework unchanged (no implicit ``hasattr`` check on every SceneTestCase), localises the test-specific behavior, and the ``--enable-dep-gen`` flag gate prevents the assertion from firing in default invocations. 9. **Unified DFX smoke directory + symmetric a2a3sim/a2a3 coverage** (``tests/st/a2a3/tensormap_and_ringbuffer/dfx/{dep_gen,l2_swimlane,pmu,tensor_dump}/``) Moved ``dep_gen_capture/`` → ``dfx/dep_gen/`` and added three sibling per-feature subdirs matching the repo's "one feature, one directory" convention (cf. ``spmd_sync_start/``, ``paged_attention_unroll/`` etc.). Each subdir holds one ``test_<feature>.py``: - ``dep_gen/test_dep_gen.py`` — asserts ``deps.json`` edges + fanout ⊆ deps gate - ``l2_swimlane/test_l2_swimlane.py`` — asserts ``l2_perf_records.json`` shape - ``pmu/test_pmu.py`` — asserts ``pmu.csv`` header + row count - ``tensor_dump/test_tensor_dump.py`` — asserts ``tensor_dump/`` manifest + bin Each test uses ``vector_example`` as a fixed 5-task workload, opens only its own flag, asserts only its own artifact. ``platforms`` field is ``["a2a3sim", "a2a3"]`` on all four — DFX features matter most on real hardware (the race window in dep_gen never fires on sim's deterministic timing, and post-hw-native-sys#742 we have empirical evidence the gate catches missed edges only on hardware), so both st-sim-a2a3 and st-onboard-a2a3 get four matching smoke steps. Each smoke is its own pytest invocation because L2PerfCollector::initialize trips its dup-init guard (``shm_host_ != nullptr``) when the L2 worker pool reuses the DeviceRunner across tests in the same session — fresh process per smoke side-steps that state leak; making the runtime init re-init-safe is a separate concern. 10. **Sibling helper consistency** (``make_pmu_csv_path``) Converted ``make_pmu_csv_path`` in ``pmu_collector.h`` to use ``std::filesystem::path`` operator/ instead of bare string concat, matching the new ``make_deps_json_path`` convention. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Follow-up to #737 addressing post-merge review findings: 1. **Dedup replay fanin per-successor** ``dep_gen_replay_emit_deps_json`` now mirrors the runtime's ``PTO2FaninBuilder::append_fanin_or_fail`` semantics: an ``std::unordered_set<uint64_t>`` tracks predecessor task ids seen so far for the current successor, and both STEP 1 (``explicit_deps``) and STEP 3 (creator retention + tensormap lookup) push through a single ``emit_unique`` lambda. Previously an ``explicit_dep`` that the tensormap also surfaced (via ``owner_task_id`` or an overlap hit) emitted two edges, which double-counted ``deps.json`` and made ``swimlane_converter.py`` draw duplicate flow events. 2. **Document the OUTPUT-slot safety contract at the capture site** ``dep_gen_replay.cpp`` sets ``tref_buf[i].ptr`` for every captured tensor slot including OUTPUT — the on-disk blob for OUTPUT is zeroed by the AICPU writer. Added an inline comment pointing at ``pto_dep_compute.h``'s per-tag dispatch (which is what makes the never-dereferenced-on-OUTPUT contract hold) so the next reader of the arg_types width-fix area doesn't have to re-derive it. 3. **``make_deps_json_path`` helper** Both onboard + sim device_runner used to build ``deps.json`` with ``output_prefix_ + "/deps.json"`` inline — out of step with the ``make_<feature>_path()`` convention shared by PMU and (previously) submit_trace. Added ``make_deps_json_path`` in ``dep_gen_collector.h``; both call sites now go through it, and the helper also handles ``create_directories`` so the path is safe even when the output dir hasn't been touched by anything else yet. 4. **``_task_id(ring, local)`` helper in the validation test** The 6-edge expected-set in ``test_dep_gen_capture.py`` was open- coded ``1 << 32`` arithmetic at every call. One helper, layout stated once. 5. **``docs/dep_gen.md``** First user-facing doc for the feature. Covers motivation (links #599), enable flags, ``deps.json`` format, ``deps_to_graph.py`` usage + node shape/color legend (AIC cube box, AIV vector ellipse, mix diamond, alloc dashed note), the ``fanout ⊆ deps`` validation gate, and the architecture touchpoints. 6. **CI smoke step for dep_gen** (``.github/workflows/ci.yml``) The default ``pytest tests/st`` invocation does not pass any DFX flag, so the dep_gen capture path never executed in CI. Added a second pytest step in ``st-sim-a2a3`` that re-runs only ``test_dep_gen_capture`` with ``--enable-dep-gen --enable-l2-swimlane``, which forces the full capture → replay → deps.json → fanout ⊆ deps gate path to execute. Same pattern is the model for future PMU / tensor_dump / swimlane smoke steps once those grow dedicated tests. 7. **UT roundtrip for ``enable_dep_gen``** (``tests/ut/py/test_chip_worker.py``) Adds the missing nanobind setter / getter / repr roundtrip for the ``enable_dep_gen`` ``CallConfig`` field, alongside the existing round-trips for ``enable_l2_swimlane`` / ``enable_dump_tensor`` / ``enable_pmu``. Catches binding-ABI regressions at the UT layer (e.g. the ``bool`` vs ``int32`` Python-wrapper bug that broke 9 st jobs earlier in the PR series — 10 s ut would have caught it). 8. **Pytest-friendly test hook via ``test_run`` override** instead of a framework-level ``_post_validate`` callback. ``test_dep_gen_capture`` overrides the inherited ``SceneTestCase.test_run`` to call ``super()`` then walk its cases and assert the dep_gen artifacts. Keeps the framework unchanged (no implicit ``hasattr`` check on every SceneTestCase), localises the test-specific behavior, and the ``--enable-dep-gen`` flag gate prevents the assertion from firing in default invocations. 9. **Unified DFX smoke directory + symmetric a2a3sim/a2a3 coverage** (``tests/st/a2a3/tensormap_and_ringbuffer/dfx/{dep_gen,l2_swimlane,pmu,tensor_dump}/``) Moved ``dep_gen_capture/`` → ``dfx/dep_gen/`` and added three sibling per-feature subdirs matching the repo's "one feature, one directory" convention (cf. ``spmd_sync_start/``, ``paged_attention_unroll/`` etc.). Each subdir holds one ``test_<feature>.py``: - ``dep_gen/test_dep_gen.py`` — asserts ``deps.json`` edges + fanout ⊆ deps gate - ``l2_swimlane/test_l2_swimlane.py`` — asserts ``l2_perf_records.json`` shape - ``pmu/test_pmu.py`` — asserts ``pmu.csv`` header + row count - ``tensor_dump/test_tensor_dump.py`` — asserts ``tensor_dump/`` manifest + bin Each test uses ``vector_example`` as a fixed 5-task workload, opens only its own flag, asserts only its own artifact. ``platforms`` field is ``["a2a3sim", "a2a3"]`` on all four — DFX features matter most on real hardware (the race window in dep_gen never fires on sim's deterministic timing, and post-#742 we have empirical evidence the gate catches missed edges only on hardware), so both st-sim-a2a3 and st-onboard-a2a3 get four matching smoke steps. Each smoke is its own pytest invocation because L2PerfCollector::initialize trips its dup-init guard (``shm_host_ != nullptr``) when the L2 worker pool reuses the DeviceRunner across tests in the same session — fresh process per smoke side-steps that state leak; making the runtime init re-init-safe is a separate concern. 10. **Sibling helper consistency** (``make_pmu_csv_path``) Converted ``make_pmu_csv_path`` in ``pmu_collector.h`` to use ``std::filesystem::path`` operator/ instead of bare string concat, matching the new ``make_deps_json_path`` convention.

Port the dep_gen (SubmitTrace) feature from a2a3 to a5 so the tensormap_and_ringbuffer runtime on a5 can produce deps.json and feed flow events into swimlane_converter.py. Without this, --enable-dep-gen was a no-op on a5 and merged_swimlane_*.json had no dependency arrows. Reused from a2a3 (code-identical; the .cpp files are byte-for-byte copies, the headers differ only in include-guard names and L2Swimlane->L2Perf / a2a3->a5 comment refs): - Shared-memory ABI: common/dep_gen.h (DepGenRecord 2624 B, overflow chain, SPSC free_queue, per-thread ready_queue) — layout byte-identical - AICPU writer: aicpu/dep_gen_collector_aicpu.{h,cpp} (.cpp identical) - Runtime replay: runtime/tensormap_and_ringbuffer/host/dep_gen_replay (.cpp identical) - Orchestrator capture point + aicpu_executor lifecycle hooks - 5 platform_config constants + PROFILING_FLAG_DEP_GEN bit Specialized for a5 (no SVM, see profiling_common diff vs a2a3): - dep_gen_collector.cpp uses alloc_single_buffer (malloc shadow + profiling_copy_to_device) instead of identity-mapping when register_cb is null — matches a5's PMU/L2Perf/Dump collectors. - Two-phase set_memory_context: callbacks first, then shm pointers once the region is committed, so start(tf) gates correctly. - reconcile_counters explicitly copy_from_device's the BufferState + current_buf before reading (mgmt thread is stopped by then). - finalize lets BufferPoolManager::clear_mappings() be the single source of truth for host-shadow lifetime — no per-collector dedup. Sim path: dlsym set_platform_dep_gen_base / set_dep_gen_enabled out of the AICPU .so and forward kernel_args.dep_gen_data_base + enable flag at boot, mirroring the existing pmu / dump / l2_perf setters. Onboard kernel.cpp adds two lines to forward dep_gen_data_base + PROFILING_FLAG_DEP_GEN into the AICPU writer's globals, mirroring the existing PMU / L2 / Dump setters. c_api: a5's DeviceRunner now overrides set_dep_gen_enabled() (was the base no-op), so run_prepared's already-present enable_dep_gen call is honored on a5 instead of being silently dropped. The shared c_api_shared.cpp / device_runner_base.h changes are comment-only — refreshing stale "a2a3-only" notes to reflect both arches now wire it. Tests: - tests/st/a5/.../dfx/dep_gen/test_dep_gen.py: 6-edge validation against vector_example orchestration (byte-identical to a2a3 — same expected edge set). - tests/st/a5/.../dfx/dep_gen/test_dep_gen_chain.py (+ kernels/orchestration/chain_barrier_orch.cpp): overflow chain regression for >64 explicit deps. - tests/st/a2a3/.../dfx/dep_gen/test_dep_gen.py: harden the existing a2a3 gate — a missing deps.json under --enable-dep-gen now fails loudly (assert) instead of silently skipping, so a capture->reconcile->replay regression can no longer pass green (hw-native-sys#742). Docs: - docs/dfx/dep_gen.md: §8 Architecture Touchpoints now lists both platforms; "Currently a2a3 only" line removed. - src/a5/runtime/.../docs/profiling_levels.md: Code Locations point at src/a5/ (was stale src/a2a3/ refs from PR hw-native-sys#777 cleanup) and add a dep_gen entry.

Port the dep_gen (SubmitTrace) feature from a2a3 to a5 so the tensormap_and_ringbuffer runtime on a5 can produce deps.json and feed flow events into swimlane_converter.py. Without this, --enable-dep-gen was a no-op on a5 and merged_swimlane_*.json had no dependency arrows. Reused from a2a3 (code-identical; the .cpp files are byte-for-byte copies, the headers differ only in include-guard names and L2Swimlane->L2Perf / a2a3->a5 comment refs): - Shared-memory ABI: common/dep_gen.h (DepGenRecord 2624 B, overflow chain, SPSC free_queue, per-thread ready_queue) — layout byte-identical - AICPU writer: aicpu/dep_gen_collector_aicpu.{h,cpp} (.cpp identical) - Runtime replay: runtime/tensormap_and_ringbuffer/host/dep_gen_replay (.cpp identical) - Orchestrator capture point + aicpu_executor lifecycle hooks - 5 platform_config constants + PROFILING_FLAG_DEP_GEN bit Specialized for a5 (no SVM, see profiling_common diff vs a2a3): - dep_gen_collector.cpp uses alloc_single_buffer (malloc shadow + profiling_copy_to_device) instead of identity-mapping when register_cb is null — matches a5's PMU/L2Perf/Dump collectors. - Two-phase set_memory_context: callbacks first, then shm pointers once the region is committed, so start(tf) gates correctly. - reconcile_counters explicitly copy_from_device's the BufferState + current_buf before reading (mgmt thread is stopped by then). - finalize lets BufferPoolManager::clear_mappings() be the single source of truth for host-shadow lifetime — no per-collector dedup. Sim path: dlsym set_platform_dep_gen_base / set_dep_gen_enabled out of the AICPU .so and forward kernel_args.dep_gen_data_base + enable flag at boot, mirroring the existing pmu / dump / l2_perf setters. Onboard kernel.cpp adds two lines to forward dep_gen_data_base + PROFILING_FLAG_DEP_GEN into the AICPU writer's globals, mirroring the existing PMU / L2 / Dump setters. c_api: a5's DeviceRunner now overrides set_dep_gen_enabled() (was the base no-op), so run_prepared's already-present enable_dep_gen call is honored on a5 instead of being silently dropped. The shared c_api_shared.cpp / device_runner_base.h changes are comment-only — refreshing stale "a2a3-only" notes to reflect both arches now wire it. Tests: - tests/st/a5/.../dfx/dep_gen/test_dep_gen.py: 6-edge validation against vector_example orchestration (byte-identical to a2a3 — same expected edge set). - tests/st/a5/.../dfx/dep_gen/test_dep_gen_chain.py (+ kernels/orchestration/chain_barrier_orch.cpp): overflow chain regression for >64 explicit deps. - tests/st/a2a3/.../dfx/dep_gen/test_dep_gen.py: harden the existing a2a3 gate — a missing deps.json under --enable-dep-gen now fails loudly (assert) instead of silently skipping, so a capture->reconcile->replay regression can no longer pass green ([hw-native-sys#742](hw-native-sys#742)). Docs: - docs/dfx/dep_gen.md: §8 Architecture Touchpoints now lists both platforms; "Currently a2a3 only" line removed. - src/a5/runtime/.../docs/profiling_levels.md: Code Locations point at src/a5/ (was stale src/a2a3/ refs from PR [hw-native-sys#777](hw-native-sys#777) cleanup) and add a dep_gen entry.

Port the dep_gen (SubmitTrace) feature from a2a3 to a5 so the tensormap_and_ringbuffer runtime on a5 can produce deps.json and feed flow events into swimlane_converter.py. Without this, --enable-dep-gen was a no-op on a5 and merged_swimlane_*.json had no dependency arrows. Reused from a2a3 (code-identical; the .cpp files are byte-for-byte copies, the headers differ only in include-guard names and L2Swimlane->L2Perf / a2a3->a5 comment refs): - Shared-memory ABI: common/dep_gen.h (DepGenRecord 2624 B, overflow chain, SPSC free_queue, per-thread ready_queue) — layout byte-identical - AICPU writer: aicpu/dep_gen_collector_aicpu.{h,cpp} (.cpp identical) - Runtime replay: runtime/tensormap_and_ringbuffer/host/dep_gen_replay (.cpp identical) - Orchestrator capture point + aicpu_executor lifecycle hooks - 5 platform_config constants + PROFILING_FLAG_DEP_GEN bit Specialized for a5 (no SVM, see profiling_common diff vs a2a3): - dep_gen_collector.cpp uses alloc_single_buffer (malloc shadow + profiling_copy_to_device) instead of identity-mapping when register_cb is null — matches a5's PMU/L2Perf/Dump collectors. - Two-phase set_memory_context: callbacks first, then shm pointers once the region is committed, so start(tf) gates correctly. - reconcile_counters explicitly copy_from_device's the BufferState + current_buf before reading (mgmt thread is stopped by then). - finalize lets BufferPoolManager::clear_mappings() be the single source of truth for host-shadow lifetime — no per-collector dedup. Sim path: dlsym set_platform_dep_gen_base / set_dep_gen_enabled out of the AICPU .so and forward kernel_args.dep_gen_data_base + enable flag at boot, mirroring the existing pmu / dump / l2_perf setters. Onboard kernel.cpp adds two lines to forward dep_gen_data_base + PROFILING_FLAG_DEP_GEN into the AICPU writer's globals, mirroring the existing PMU / L2 / Dump setters. c_api: a5's DeviceRunner now overrides set_dep_gen_enabled() (was the base no-op), so run_prepared's already-present enable_dep_gen call is honored on a5 instead of being silently dropped. The shared c_api_shared.cpp / device_runner_base.h changes are comment-only — refreshing stale "a2a3-only" notes to reflect both arches now wire it. Tests: - tests/st/a5/.../dfx/dep_gen/test_dep_gen.py: 6-edge validation against vector_example orchestration (byte-identical to a2a3 — same expected edge set). - tests/st/a5/.../dfx/dep_gen/test_dep_gen_chain.py (+ kernels/orchestration/chain_barrier_orch.cpp): overflow chain regression for >64 explicit deps. - tests/st/a2a3/.../dfx/dep_gen/test_dep_gen.py: harden the existing a2a3 gate — a missing deps.json under --enable-dep-gen now fails loudly (assert) instead of silently skipping, so a capture->reconcile->replay regression can no longer pass green ([[hw-native-sys#742](https://github.com/hw-native-sys/simpler/issues/742)](https://github.com/hw-native-sys/simpler/issues/742)). Docs: - docs/dfx/dep_gen.md: §8 Architecture Touchpoints now lists both platforms; "Currently a2a3 only" line removed. - src/a5/runtime/.../docs/profiling_levels.md: Code Locations point

Port the dep_gen (SubmitTrace) feature from a2a3 to a5 so the tensormap_and_ringbuffer runtime on a5 can produce deps.json and feed flow events into swimlane_converter.py. Without this, --enable-dep-gen was a no-op on a5 and merged_swimlane_*.json had no dependency arrows. Reused from a2a3 (code-identical; the .cpp files are byte-for-byte copies, the headers differ only in include-guard names and L2Swimlane->L2Perf / a2a3->a5 comment refs): - Shared-memory ABI: common/dep_gen.h (DepGenRecord 2624 B, overflow chain, SPSC free_queue, per-thread ready_queue) — layout byte-identical - AICPU writer: aicpu/dep_gen_collector_aicpu.{h,cpp} (.cpp identical) - Runtime replay: runtime/tensormap_and_ringbuffer/host/dep_gen_replay (.cpp identical) - Orchestrator capture point + aicpu_executor lifecycle hooks - 5 platform_config constants + PROFILING_FLAG_DEP_GEN bit Specialized for a5 (no SVM, see profiling_common diff vs a2a3): - dep_gen_collector.cpp uses alloc_single_buffer (malloc shadow + profiling_copy_to_device) instead of identity-mapping when register_cb is null — matches a5's PMU/L2Perf/Dump collectors. - Two-phase set_memory_context: callbacks first, then shm pointers once the region is committed, so start(tf) gates correctly. - reconcile_counters explicitly copy_from_device's the BufferState + current_buf before reading (mgmt thread is stopped by then). - finalize lets BufferPoolManager::clear_mappings() be the single source of truth for host-shadow lifetime — no per-collector dedup. Sim path: dlsym set_platform_dep_gen_base / set_dep_gen_enabled out of the AICPU .so and forward kernel_args.dep_gen_data_base + enable flag at boot, mirroring the existing pmu / dump / l2_perf setters. Onboard kernel.cpp adds two lines to forward dep_gen_data_base + PROFILING_FLAG_DEP_GEN into the AICPU writer's globals, mirroring the existing PMU / L2 / Dump setters. c_api: a5's DeviceRunner now overrides set_dep_gen_enabled() (was the base no-op), so run_prepared's already-present enable_dep_gen call is honored on a5 instead of being silently dropped. The shared c_api_shared.cpp / device_runner_base.h changes are comment-only — refreshing stale "a2a3-only" notes to reflect both arches now wire it. Tests: - tests/st/a5/.../dfx/dep_gen/test_dep_gen.py: 6-edge validation against vector_example orchestration (byte-identical to a2a3 — same expected edge set). - tests/st/a5/.../dfx/dep_gen/test_dep_gen_chain.py (+ kernels/orchestration/chain_barrier_orch.cpp): overflow chain regression for >64 explicit deps. - tests/st/a2a3/.../dfx/dep_gen/test_dep_gen.py: harden the existing a2a3 gate — a missing deps.json under --enable-dep-gen now fails loudly (assert) instead of silently skipping, so a capture->reconcile->replay regression can no longer pass green ([[#742](https://github.com/hw-native-sys/simpler/issues/742)](https://github.com/hw-native-sys/simpler/issues/742)). Docs: - docs/dfx/dep_gen.md: §8 Architecture Touchpoints now lists both platforms; "Currently a2a3 only" line removed. - src/a5/runtime/.../docs/profiling_levels.md: Code Locations point

) `src/a2a3/platform/onboard/aicpu/kernel.cpp` propagates dump_tensor / l2_swimlane / pmu from `k_args->enable_profiling_flag` to AICPU globals on every kernel entry, but the two equivalent calls for dep_gen were missing — so `is_dep_gen_enabled()` stayed false on hardware, the orchestrator's `dep_gen_aicpu_record_submit` early-returned for every submit, the ring stayed empty, and the host replay wrote `{"version":1,"edges":[]}` for every run. The fanout ⊆ deps gate introduced in PR hw-native-sys#737 then ran against an empty deps set — vacuously passing only when fanout was also empty, and silent-FAILing otherwise. Sim was unaffected: its host device_runner already dlsym's `set_platform_dep_gen_base` and `set_dep_gen_enabled` into the AICPU .so. Onboard had no equivalent wiring. Changes - `aicpu/kernel.cpp`: include `dep_gen_collector_aicpu.h` and add the two missing setters next to the existing dump_tensor / l2_swimlane / pmu propagation. - `scene_test.py`: add `SceneTestCase._effective_enable_dep_gen(request, *, warn=False)` as the single source of truth for the CLI flag after the `--rounds > 1` disable rule. The framework's `test_run` consumes it with `warn=True` (it owns the user-facing "disabled because rounds > 1" message); subclass overrides consume it with `warn=False` so the gating logic can't drift and the warning fires once. - `dep_gen_capture/test_dep_gen_capture.py`: add `"a2a3"` to the case's platforms list (was sim-only — the reason this never surfaced in CI) and override `test_run` so the pytest path actually invokes `_post_validate` when `--enable-dep-gen` is passed, via the new helper. Without that override, pytest collected the test and ran orchestration but never asserted anything about deps.json, so a future regression of the same shape would still slip through. - `.github/workflows/ci.yml`: add a dedicated dep_gen smoke step to the st-onboard-a2a3 job that runs the capture test with `--enable-dep-gen --enable-l2-swimlane`, so the fanout ⊆ deps gate exercises hardware on every PR touching a2a3. Verified on a2a3 onboard: `deps.json` now contains the six edges documented in `example_orchestration.cpp` (t0->t1, t0->t2, t1->t3, t2->t3, t0->t4, t3->t4), the smoke test passes with the gate flags, and the broader scene-test sweep is unchanged. Co-authored-by: wcwxy <26245345+ChaoWao@users.noreply.github.com>

…ve-sys#740) Follow-up to hw-native-sys#737 addressing post-merge review findings: 1. **Dedup replay fanin per-successor** ``dep_gen_replay_emit_deps_json`` now mirrors the runtime's ``PTO2FaninBuilder::append_fanin_or_fail`` semantics: an ``std::unordered_set<uint64_t>`` tracks predecessor task ids seen so far for the current successor, and both STEP 1 (``explicit_deps``) and STEP 3 (creator retention + tensormap lookup) push through a single ``emit_unique`` lambda. Previously an ``explicit_dep`` that the tensormap also surfaced (via ``owner_task_id`` or an overlap hit) emitted two edges, which double-counted ``deps.json`` and made ``swimlane_converter.py`` draw duplicate flow events. 2. **Document the OUTPUT-slot safety contract at the capture site** ``dep_gen_replay.cpp`` sets ``tref_buf[i].ptr`` for every captured tensor slot including OUTPUT — the on-disk blob for OUTPUT is zeroed by the AICPU writer. Added an inline comment pointing at ``pto_dep_compute.h``'s per-tag dispatch (which is what makes the never-dereferenced-on-OUTPUT contract hold) so the next reader of the arg_types width-fix area doesn't have to re-derive it. 3. **``make_deps_json_path`` helper** Both onboard + sim device_runner used to build ``deps.json`` with ``output_prefix_ + "/deps.json"`` inline — out of step with the ``make_<feature>_path()`` convention shared by PMU and (previously) submit_trace. Added ``make_deps_json_path`` in ``dep_gen_collector.h``; both call sites now go through it, and the helper also handles ``create_directories`` so the path is safe even when the output dir hasn't been touched by anything else yet. 4. **``_task_id(ring, local)`` helper in the validation test** The 6-edge expected-set in ``test_dep_gen_capture.py`` was open- coded ``1 << 32`` arithmetic at every call. One helper, layout stated once. 5. **``docs/dep_gen.md``** First user-facing doc for the feature. Covers motivation (links hw-native-sys#599), enable flags, ``deps.json`` format, ``deps_to_graph.py`` usage + node shape/color legend (AIC cube box, AIV vector ellipse, mix diamond, alloc dashed note), the ``fanout ⊆ deps`` validation gate, and the architecture touchpoints. 6. **CI smoke step for dep_gen** (``.github/workflows/ci.yml``) The default ``pytest tests/st`` invocation does not pass any DFX flag, so the dep_gen capture path never executed in CI. Added a second pytest step in ``st-sim-a2a3`` that re-runs only ``test_dep_gen_capture`` with ``--enable-dep-gen --enable-l2-swimlane``, which forces the full capture → replay → deps.json → fanout ⊆ deps gate path to execute. Same pattern is the model for future PMU / tensor_dump / swimlane smoke steps once those grow dedicated tests. 7. **UT roundtrip for ``enable_dep_gen``** (``tests/ut/py/test_chip_worker.py``) Adds the missing nanobind setter / getter / repr roundtrip for the ``enable_dep_gen`` ``CallConfig`` field, alongside the existing round-trips for ``enable_l2_swimlane`` / ``enable_dump_tensor`` / ``enable_pmu``. Catches binding-ABI regressions at the UT layer (e.g. the ``bool`` vs ``int32`` Python-wrapper bug that broke 9 st jobs earlier in the PR series — 10 s ut would have caught it). 8. **Pytest-friendly test hook via ``test_run`` override** instead of a framework-level ``_post_validate`` callback. ``test_dep_gen_capture`` overrides the inherited ``SceneTestCase.test_run`` to call ``super()`` then walk its cases and assert the dep_gen artifacts. Keeps the framework unchanged (no implicit ``hasattr`` check on every SceneTestCase), localises the test-specific behavior, and the ``--enable-dep-gen`` flag gate prevents the assertion from firing in default invocations. 9. **Unified DFX smoke directory + symmetric a2a3sim/a2a3 coverage** (``tests/st/a2a3/tensormap_and_ringbuffer/dfx/{dep_gen,l2_swimlane,pmu,tensor_dump}/``) Moved ``dep_gen_capture/`` → ``dfx/dep_gen/`` and added three sibling per-feature subdirs matching the repo's "one feature, one directory" convention (cf. ``spmd_sync_start/``, ``paged_attention_unroll/`` etc.). Each subdir holds one ``test_<feature>.py``: - ``dep_gen/test_dep_gen.py`` — asserts ``deps.json`` edges + fanout ⊆ deps gate - ``l2_swimlane/test_l2_swimlane.py`` — asserts ``l2_perf_records.json`` shape - ``pmu/test_pmu.py`` — asserts ``pmu.csv`` header + row count - ``tensor_dump/test_tensor_dump.py`` — asserts ``tensor_dump/`` manifest + bin Each test uses ``vector_example`` as a fixed 5-task workload, opens only its own flag, asserts only its own artifact. ``platforms`` field is ``["a2a3sim", "a2a3"]`` on all four — DFX features matter most on real hardware (the race window in dep_gen never fires on sim's deterministic timing, and post-hw-native-sys#742 we have empirical evidence the gate catches missed edges only on hardware), so both st-sim-a2a3 and st-onboard-a2a3 get four matching smoke steps. Each smoke is its own pytest invocation because L2PerfCollector::initialize trips its dup-init guard (``shm_host_ != nullptr``) when the L2 worker pool reuses the DeviceRunner across tests in the same session — fresh process per smoke side-steps that state leak; making the runtime init re-init-safe is a separate concern. 10. **Sibling helper consistency** (``make_pmu_csv_path``) Converted ``make_pmu_csv_path`` in ``pmu_collector.h`` to use ``std::filesystem::path`` operator/ instead of bare string concat, matching the new ``make_deps_json_path`` convention.

Port the dep_gen (SubmitTrace) feature from a2a3 to a5 so the tensormap_and_ringbuffer runtime on a5 can produce deps.json and feed flow events into swimlane_converter.py. Without this, --enable-dep-gen was a no-op on a5 and merged_swimlane_*.json had no dependency arrows. Reused from a2a3 (code-identical; the .cpp files are byte-for-byte copies, the headers differ only in include-guard names and L2Swimlane->L2Perf / a2a3->a5 comment refs): - Shared-memory ABI: common/dep_gen.h (DepGenRecord 2624 B, overflow chain, SPSC free_queue, per-thread ready_queue) — layout byte-identical - AICPU writer: aicpu/dep_gen_collector_aicpu.{h,cpp} (.cpp identical) - Runtime replay: runtime/tensormap_and_ringbuffer/host/dep_gen_replay (.cpp identical) - Orchestrator capture point + aicpu_executor lifecycle hooks - 5 platform_config constants + PROFILING_FLAG_DEP_GEN bit Specialized for a5 (no SVM, see profiling_common diff vs a2a3): - dep_gen_collector.cpp uses alloc_single_buffer (malloc shadow + profiling_copy_to_device) instead of identity-mapping when register_cb is null — matches a5's PMU/L2Perf/Dump collectors. - Two-phase set_memory_context: callbacks first, then shm pointers once the region is committed, so start(tf) gates correctly. - reconcile_counters explicitly copy_from_device's the BufferState + current_buf before reading (mgmt thread is stopped by then). - finalize lets BufferPoolManager::clear_mappings() be the single source of truth for host-shadow lifetime — no per-collector dedup. Sim path: dlsym set_platform_dep_gen_base / set_dep_gen_enabled out of the AICPU .so and forward kernel_args.dep_gen_data_base + enable flag at boot, mirroring the existing pmu / dump / l2_perf setters. Onboard kernel.cpp adds two lines to forward dep_gen_data_base + PROFILING_FLAG_DEP_GEN into the AICPU writer's globals, mirroring the existing PMU / L2 / Dump setters. c_api: a5's DeviceRunner now overrides set_dep_gen_enabled() (was the base no-op), so run_prepared's already-present enable_dep_gen call is honored on a5 instead of being silently dropped. The shared c_api_shared.cpp / device_runner_base.h changes are comment-only — refreshing stale "a2a3-only" notes to reflect both arches now wire it. Tests: - tests/st/a5/.../dfx/dep_gen/test_dep_gen.py: 6-edge validation against vector_example orchestration (byte-identical to a2a3 — same expected edge set). - tests/st/a5/.../dfx/dep_gen/test_dep_gen_chain.py (+ kernels/orchestration/chain_barrier_orch.cpp): overflow chain regression for >64 explicit deps. - tests/st/a2a3/.../dfx/dep_gen/test_dep_gen.py: harden the existing a2a3 gate — a missing deps.json under --enable-dep-gen now fails loudly (assert) instead of silently skipping, so a capture->reconcile->replay regression can no longer pass green ([[hw-native-sys#742](https://github.com/hw-native-sys/simpler/issues/742)](https://github.com/hw-native-sys/simpler/issues/742)). Docs: - docs/dfx/dep_gen.md: §8 Architecture Touchpoints now lists both platforms; "Currently a2a3 only" line removed. - src/a5/runtime/.../docs/profiling_levels.md: Code Locations point

gemini-code-assist Bot reviewed May 12, 2026

View reviewed changes

Comment thread tests/st/a2a3/tensormap_and_ringbuffer/dep_gen_capture/test_dep_gen_capture.py

ChaoWao approved these changes May 12, 2026

View reviewed changes

ChaoWao merged commit 9fd6c19 into hw-native-sys:main May 12, 2026
26 of 28 checks passed

ChaoWao deleted the fix/a2a3-onboard-aicpu-dep-gen-wiring branch May 12, 2026 01:45

indigo1973 mentioned this pull request Jun 1, 2026

Add: dep_gen capture+replay support on a5 #886

Merged

ChaoWao mentioned this pull request Jul 26, 2026

Fix: capture dep_gen on the host orchestrator for host_build_graph #1492

Merged

Navigation Menu

Fix: a2a3 onboard AICPU never enables dep_gen capture - #742

Fix: a2a3 onboard AICPU never enables dep_gen capture#742
ChaoWao merged 2 commits into
hw-native-sys:mainfrom
hw-native-sys-bot:fix/a2a3-onboard-aicpu-dep-gen-wiring

hw-native-sys-bot commented May 12, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

hw-native-sys-bot commented May 12, 2026

Summary

Changes

Test plan

Context

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants