Add: integrate kernel-mode PRs with an end-to-end launch test

[sunkaixuan2018]

Summary

Integrates the in-flight kernel-mode PRs into one line on the current mainline, and proves it with a public-entry, caller-owned launch that computes real results on a2a3 hardware.

• Contains the contributions of Refactor: [H1] measure HBG graphs before device upload #2171, Add: [2B] declare HBG kernel resource capacity #2172, Add: [Context Prepare] prepare and freeze kernel resources #2173, Add: [H2] serialize immutable HBG packets for HostArgs #2174, Add: [H3] seal and validate HBG execution slots #2175, Add: persistent kernel-context execution resources (K2) #2176, Add: TMR kernel-mode resource contracts and init admission（2a） #2177, Add: TMR kernel invocation snapshots and native transport adapter（K4） #2180, Add: kernel-mode entry layer — give the four C entries callers (⑩a) #2185, Add: [Binder] standalone three-stream kernel submission #2187, Add: TMR kernel resident arguments and invocation isolation（K5） #2189, Add: bounded kernel callable registration #2190 and Fix: kernel-mode capacity refusals no longer free what they protect (⑤K3) #2193. Add: kernel-mode C ABI skeleton and wire headers (K1) #2064 (K1) is already on main, and this branch adopts its final contract.
• Launch reaches the device. On onboard tensormap_and_ringbuffer, simpler_kernel_mode_launch validates the context and callable, encodes the invocation and submits it through the three-stream binder. The AICPU dispatch entry hands the payload to the TMR executor. The kernel path commits the pooled arena before the runtime image is uploaded. host_build_graph and simulation report no kernel capability.
• End-to-end test. tests/ut/py/test_kernel_mode_c_api.py::test_kernel_eager_launch_executes_fresh_tensor_and_scalar_snapshots creates its own ACL state, device, stream and tensors, drives init, prepare_callable, launch and finalize, and compares all 16,384 FLOAT32 elements of two launches against x + scalar.
• Docs. docs/zh-cn/kernel-mode-integration-test.md describes the test, the launch sequence, and the current state of graph mode. The kernel docs are added to the mkdocs nav, which --strict requires.

Contract decisions

Every decision is recorded with its reasoning in INTEGRATION-LOG.md, entries D0 to D14.

Topic	Decision
Host error codes	Merged K1 numbering: INVALID_ARGUMENT is BASE - 4; CALLABLE_COUNT_EXCEEDED, CALLABLE_BYTES_EXCEEDED, CALLABLE_NOT_RESIDENT, CALLABLE_STALE, CAPACITY_EXCEEDED are BASE - 5 to BASE - 9
prepare_callable	Five parameters, keeping caller_stream
launch	Callable id, not a handle; generation is checked against the residency on host and device; an out-of-range id returns INVALID_ARGUMENT
Launch events	Fork from the caller's Start, AICore submitted before AICPU; KernelEventKind names the fifth event PrepareTail
AICPU transport	In-repo rtsLaunchCpuKernel path; the WithHostArgs native binder stays available for a later switch
Symbol resolution	Every runtime exports all four entries; ChipWorker resolves init, prepare and launch only when supported is nonzero
Kernel capacity	K3's shared static arena bank for onboard and simulation; the inline arena guard added to the onboard runner on main is not carried

Effect on the original PRs

• Each contribution is contained here. A PR that lands separately would need to adopt the decisions above: the four-parameter prepare_callable in Add: persistent kernel-context execution resources (K2) #2176, Add: TMR kernel invocation snapshots and native transport adapter（K4） #2180 and Add: TMR kernel resident arguments and invocation isolation（K5） #2189; the handle ABI in Add: bounded kernel callable registration #2190; the event enum in Add: persistent kernel-context execution resources (K2) #2176; the K1 variant carried by Refactor: [H1] measure HBG graphs before device upload #2171 to Add: [H3] seal and validate HBG execution slots #2175.
• Add: TMR kernel invocation snapshots and native transport adapter（K4） #2180 and Add: TMR kernel resident arguments and invocation isolation（K5） #2189 must not both land. This branch takes Add: TMR kernel invocation snapshots and native transport adapter（K4） #2180's K4 chain plus Add: TMR kernel resident arguments and invocation isolation（K5） #2189's K5 commit.

Testing

Hardware: a2a3, through onboard-arch-precheck and task-submit --device auto.

Suite	Result
C++ unit tests, no hardware	166/166 passed
C++ unit tests with SIMPLER_ENABLE_HARDWARE_TESTS=ON, no hardware	168/168 passed
C++ hardware tests, ^requires_hardware(_a2a3)?$	2/2 passed (test_comm_lifecycle, test_kernel_mode_entry)
Python unit tests, tests/ut -m "not requires_hardware"	2335 passed
Python hardware unit tests, tests/ut -m requires_hardware --platform a2a3	30 passed
a2a3 hardware: kernel C API and capture probe	13/13 passed
Invocation transport probe, tools/cann-examples/tmr-invocation-snapshot	24 gated asynchronous snapshots passed
a2a3 onboard scenes, -m "not sdma" --exclude-level 4	167 passed, 1 skipped
a2a3 SDMA scenes	3 passed
a2a3sim scenes	82 passed, 8 skipped
a5sim scenes	78 passed
clang-tidy 18 on changed C++ files	passed, zero diagnostics
Other pre-commit hooks on changed files	passed
mkdocs build --strict	passed

The 13 hardware cases are the eager numerical launch, seven lifecycle retry scenarios, two device-query rejections, kernel init on a borrowed device for both runtimes, and the K2 capture probe with 100 replays.

The full Python hardware run found tests/ut/py/test_worker/test_kernel_mode_entry.py still asserting that every runtime refuses kernel init, which failed three jobs against TMR. The second commit points its refusal cases at host_build_graph and adds init_claims_borrowed_stream, matching the C++ twin; the 30-passed result is from the fixed file. Details are in docs/kernel-integration-validation.md, sections "Mainline merge revalidation" and "Full sweeps on the merged tree".

ctest --test-dir tests/ut/cpp/build -LE requires_hardware
pytest tests/ut -m "not requires_hardware"
task-submit --device auto --device-num 2 --run "pytest tests/ut -m requires_hardware --platform a2a3 --device \$TASK_DEVICE"
task-submit --device auto --device-num 4 --run "pytest examples tests/st -m 'not sdma' --platform a2a3 --exclude-level 4 --device \$TASK_DEVICE --require-pto-isa --pto-session-timeout 1200"
task-submit --device auto --device-num 2 --run "pytest examples tests/st -m sdma --platform a2a3 --device \$TASK_DEVICE --require-pto-isa --pto-session-timeout 600"
pytest examples tests/st --platform a2a3sim --device 0-15 --require-pto-isa --pto-session-timeout 600
pytest examples tests/st --platform a5sim --device 0-15 --require-pto-isa --pto-session-timeout 600
mkdocs build --strict

Not covered

• Public host_build_graph kernel execution: H4 has no submitted PR, so HBG reports no kernel capability.
• A5 hardware: only builds, unit tests and simulation.
• A public launch inside an ACLGraph capture window: the capture probe replays the three-stream primitives 100 times, and the public launch is verified eagerly, but the two have not run together.
• Shapes, multi-operator graphs and concurrent launches beyond the single fixed-shape operator above.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

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: Advanced

Run ID: e00f8456-077c-4987-a239-21b883fe35a3

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

---

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.}