PoC 3 — Wave NPU Runtime and Driver Trace

PoC 3 passed on 2026-07-28. Two independently captured, bounded traces of the firmware-matched stock CTC graph produced the same:

• eight probe lifecycle phases;
• loaded runtime-library set and order;
• driver open/probe events;
• 73-call normalized galcore command sequence;
• per-phase command assignment;
• syscall counts;
• 332,800-byte output with SHA-256 e621b2134934ab94ab83049b4b56e5bb410831de41d9727aaf95384d413b52d4;
• graph cleanup status zero;
• responsive post-run ADB and unchanged captured NPU state.

The normalized GAL sequence SHA-256 in both runs is:

33b93229efcf7258253d2124129ac24f107de63b1742d6666db2e674cbb1c226

This proves a stable, observable userspace-to-driver transaction for a known valid Wave graph. It does not unlock compilation, disclose command-buffer contents, or authorize ioctl replay.

Evidence

Tracked, pointer-free evidence:

File	Identity	Purpose
normalized-run-1.json	b694d88f778ecdc72b06afb24bd4c749824c358d9c5a4bcb9ada67d970b3f8b3	First successful normalized trace
normalized-run-2.json	ee0090152e5492dc6ac3f107a492610801e69500607b92ca8bf089f506f7cf7c	Second successful normalized trace
repeatability.json	207b70cdea1ecb89cbd2ad80e8a3d81c1fde4958d5369d02d378e563f36ea7fc	Machine comparison
galcore-interface.md	Analysis	Interface and command evidence
runtime-sequence.md	Analysis	Phase-by-phase interpretation
repeatability.md	Analysis	Acceptance and variance analysis

The full experiment record is 2026-07-28-poc3-runtime-driver-trace.md. Raw envelopes and streams remain under artifacts/npu/runs/ and are gitignored because they contain PIDs, addresses, timestamps, and complete system-call formatting.

What was observed

Exact live interface

Every GAL call used:

ioctl request:          30000 (0x7530 in raw strace)
driver envelope:        32 bytes
HAL interface bytes:    376
core index:             0
successful ioctl:       73 / 73
returned HAL status 0:  73 / 73

The 32-byte envelope and first 16-byte interface prefix are supported by the pinned public eCV5546 GAL 6.4.18.5 source. The live Wave library is 6.4.19.5.946407, so field names remain version-adjacent evidence; coherent live values across 146 completed call observations strongly corroborate the prefix on this unit. Details and evidence classes are in galcore-interface.md.

Driver lifecycle

The explicit probe check in phase 1 opened and closed /dev/galcore. During graph creation in phase 5, the runtime:

1. opened /dev/galcore;
2. probed /dev/galcore1, which returned ENOENT;
3. initialized the generic and VIP contexts;
4. allocated and locked graph memory;
5. kept the primary driver descriptor until process exit.

No explicit /dev/galcore close appears after runtime initialization. The probe reports graph cleanup zero, and normal process exit closes the descriptor. This describes the observed probe lifetime; it is not a recommendation for a long-lived production wrapper.

Graph lifecycle

Probe phase	GAL calls	Observed boundary
1 — driver reachability	0	Explicit open/close only
2 — read NPU status	0	Read-only sysfs
3 — read/validate model	0	Stock NBG file read
4 — query tensor metadata	0	Parser/runtime metadata path
5 — create graph	46	Version/chip discovery, VIP attach, signals, base address, 11 allocations and 11 locks
6 — submit zero input	0	No driver transaction visible at this boundary
7 — execute three times	9	Exactly three COMMIT → EVENT_COMMIT → USER_SIGNAL groups
8 — destroy/teardown	18	Seven unlock/release pairs, one final flush group, one final unlock

The absence of an ioctl in phase 6 indicates that the tested aml_module_input_set path updates already prepared userspace/runtime storage. It does not prove every tensor-binding mode is driver-free.

Dynamic runtime stack

The traced process successfully opened these relevant shared objects, in first-observed order:

libnnsdk.so
librt.so.1
libovxlib.so
libOpenVX.so
libVSC_Lite.so
libm.so.6
libVSC.so
libGAL.so
libArchModelSw.so
libNNArchPerf.so
libstdc++.so.6

The dynamic loader also opened the normal glibc support libraries libdl.so.2, libgcc_s.so.1, libpthread.so.0, and libc.so.6.

Loading libVSC.so, libVSC_Lite.so, or libArchModelSw.so is not proof that a stock NBG is compiled on-device. PoC 3 proves only that those DSOs are mapped during initialization. PoC 5 must find and exercise a legitimate Wave-targeted graph-production route separately.

Repeatability result

Invariant	Run 1	Run 2
Harness classification	success	success
Remote exit	0	0
Probe mean, descriptive	3.562 ms	3.474 ms
Output bytes	332,800	332,800
Output digest	e621b213…52d4	e621b213…52d4
Cleanup	0	0
GAL calls	73	73
Complete metadata/status pairs	73	73
Normalized sequence digest	33b93229…1c226	33b93229…1c226
Syscall counts identical	yes	yes
Loaded path observations identical	yes	yes
Driver events identical	yes	yes
Required hashes unchanged	yes	yes
NPU parameters unchanged	yes	yes

The raw stdout hashes differ because PIDs, addresses, timestamps, and durations are intentionally retained only in raw evidence. The normalizer removes those values before comparison.

Instrumentation

Source:

waved/tools/npu-trace/gal_trace.c
waved/tools/npu-trace/build.sh
waved/tools/npu-trace/stage.sh
waved/tools/npu-trace/stock-ctc-trace.sh
waved/scripts/npu/trace_normalize.py

Important artifact hashes:

Artifact	SHA-256
AArch64 interposer	dccaedd13812c161949905efecf7aa19fe74712b75de8d80211924e286e68601
Fixed device wrapper	7799626fcad221088a109740022e93ca5b7652d7589c9e79347ec92a1b1ac866
Live strace 5.9	2046787af2bec0edd5c3483b59bcbd870e5a0fb68cc3c618e9e464e61ee62aa2
Live libGAL.so	13b7d87eabb8924758d49fbd8f1fac41bb81f34069f6ceed949af764893061e6
Stock probe	fe8da5702032c397023acebd1a4bdf40ce086f6e86277a52c4f059f409552db0
Stock CTC NBG	5d2795be781af80593321cf0ae627f56f4f6f64e71e95d1d8b0a27f6192a9f66

The final experiment definition SHA-256 is:

dcc6efe65182a9a8ded2f1c1f96fbf403098a51d414f2f48da4f9f029ac19992

Reproduce

Run from the repository root with the authorized Wave connected:

waved/tools/npu-trace/build.sh
waved/tools/npu-trace/stage.sh

python3 waved/scripts/npu/npu_lab.py snapshot \
  --experiment stock-ctc-trace \
  --output artifacts/npu/poc3-snapshot.json

python3 waved/scripts/npu/npu_lab.py run stock-ctc-trace \
  --artifact-root artifacts/npu/runs \
  --run-id poc3-local-trace-1

python3 waved/scripts/npu/trace_normalize.py normalize \
  --run-directory artifacts/npu/runs/poc3-local-trace-1 \
  --output artifacts/npu/poc3-local-normalized-1.json

Use new paths and run IDs; the tools refuse to overwrite existing results. Repeat once, then compare:

python3 waved/scripts/npu/trace_normalize.py compare \
  --left artifacts/npu/poc3-local-normalized-1.json \
  --right artifacts/npu/poc3-local-normalized-2.json

The Wave’s /data filesystem was full during this PoC. The two disposable trace artifacts are therefore staged under the executable tmpfs directory /tmp/wave-npu-lab/. The harness allows exactly /tmp/wave-npu-lab/stock-ctc-trace, not arbitrary /tmp executables. Reboot clears the staged files. stage.sh verifies local hashes, refuses to overwrite an unexpected remote file, and reads both hashes back.

Safety and limits

The interposer:

• forwards the original ioctl once through the kernel syscall;
• records no pointer value;
• reads only the 32-byte envelope and 16-byte interface prefix;
• rejects interface sizes outside 16–65,536 bytes;
• never reads the command-specific union;
• records no model, coefficient, tensor, audio, or command-buffer payload;
• performs no mutation, replay, retry, or recovery.

PoC 3 does not prove:

• the meaning or layout of any command-specific union member;
• the NPU instruction set;
• how NBG command, coefficient, patch, or load-state sections map to hardware packets;
• that VSC can compile a custom graph for product 0x1000000e;
• that a noise-suppression topology is supported;
• that this ASR graph’s descriptive latency predicts audio-model latency.

Those are later gates. Raw ioctl replay and unverified graph execution remain prohibited.