fix(camera): blur CLI/self-verify calls undefined edge_detect -> NameError on NPU

[Vyacheslav-Tomashevskiy]

Problem

camera/blur_pipeline.py defines its NPU design as blur (def blur(...), and both camera/npu_camera_daemon.py and camera/test_blur.py do from blur_pipeline import blur). But the module's own runtime paths were still calling edge_detect — a leftover name from AMD's edge_detect IRON example this file was adapted from. edge_detect is never defined or imported anywhere in blur_pipeline.py:

• _run_and_verify() — edge_detect(in_t, b_t, out_t, **_compile_kwargs(opts))
• main() — run_design_cli(edge_detect, ...)

Why CI stays green but real hardware crashes

blur_pipeline.py imports aie.iron at the top, so on a host without the NPU toolchain the module fails at that import first — which is exactly the boundary ci.yml's import loop expects, so CI is green. On a real NPU box the import succeeds, python camera/blur_pipeline.py reaches main(), and it dies with:

NameError: name 'edge_detect' is not defined

before running the blur design or its self-verify. The module's CLI/self-verify is unusable on the one platform it's meant for.

Fix

Rename the two runtime references from edge_detect to blur (the function that actually exists and that the daemon/test already import). Two-line change; no behavior change on the happy path.

Regression test

Added camera/test_blur_pipeline_cli_symbol.py — a pure-AST test (no aie/numpy, runs anywhere) asserting the design symbol the CLI hands to the NPU is one the module actually defines. Mutation-guarded: it fails on the pre-fix code (AssertionError: ... references undefined edge_detect ...) and passes after the fix. python -m compileall -q camera stays clean.

---

Verified bug — claimed under the RustChain verified-bug channel (rustchain-bounties#71).
RTC: RTCd1554f0f35576faf01d386a6be1c947f560dd0b7

[Vyacheslav-Tomashevskiy]

Good catch, thank you -- pushed a revision. Added _blur_ref(), a numpy reference that actually mirrors what blur() computes: rgba2gray -> 3x3 Gaussian blur (the unity-gain [[1,2,1],[2,4,2],[1,2,1]]/16 kernel filter_kernel_buff holds, per the module's own "GAIN FIX" docstring) -> gray2rgba, combined via add_weighted with alpha=1/beta=0 (register values 16384/0 -- the original RGBA is not blended back in, unlike edge_detect's highlighted-edges-over-original composite). _run_and_verify() now calls _blur_ref instead of _edge_detect_ref.

Added camera/test_blur_pipeline_golden_ref.py: an AST guard on the _run_and_verify call target, an impulse-response test pinning the exact kernel weights and normalization shift, a flat-field test for BORDER_REPLICATE at the corners, and a checkerboard-smoothing test mirroring test_blur.py's real-hardware smoke assertion (which also catches a beta!=0 regression). No aie/NPU toolchain needed -- minimal import-time stand-ins only, the real design/verify code is never executed, and blur_pipeline.py still fails only at the aie import boundary per this repo's CI convention. Mutation-checked: reverting the _run_and_verify call target fails the AST guard, corrupting the kernel weights fails the impulse-response test.

[Scottcjn]

Reviewed properly, not bulk-triaged. The core diagnosis is confirmed on main: blur_pipeline.py has no def edge_detect and no import of it, yet _run_and_verify() (line 320) and main() (line 342) both reference it, so python blur_pipeline.py on a real NPU dies with NameError before it ever reaches the design. I also confirmed the second half of your claim by reading the design body: only two workers are wired (rgba2gray_fn, gray2rgba_add_weight_fn around the filter_fn), threshold_line_kernel and of_intermediates[2] are never used, filter_kernel_buff really is [[256,512,256],[512,1024,512],[256,512,256]], and alpha, beta, gamma = 16384, 0, 0 — so _blur_ref’s _add_weighted_cv_ref(mask_rgba, rgba_uint8, 1, 0, 0) is the right shape and _edge_detect_ref was genuinely the wrong golden. I ran pytest camera/test_blur_pipeline_cli_symbol.py camera/test_blur_pipeline_golden_ref.py — 6 passed.

Two things before this goes in.

1. The header comment of test_blur_pipeline_golden_ref.py contradicts its own PR:

# _run_and_verify() used to compare the blur() design's output against
# _edge_detect_ref() -- ...
# #2 fixed the CLI NameError crash (edge_detect -> blur) but left
# that mismatch in place, so on real NPU hardware --verify would go from
# "crashes" to "runs and reports the blur output as wrong"

This is #2, and it fixes both in the same diff — _run_and_verify gets blur(...) and _blur_ref(...) in one hunk. gh pr list --state all shows only #1 (the path/CI fix) and this one, so there is no earlier PR the note can be pointing at. As written it sends a future reader hunting for a #3 that closed a gap that was never actually left open. Please reword to describe the single change.

2. Not a defect, just the limit of what I could check: I have no NPU box, so every assertion I ran was against the pure-numpy _..._ref() helpers. What I could not exercise is the thing this PR changes the meaning of — whether _blur_ref matches what the AIE filter2d kernel actually emits, given the fixed-point path (int16 coeffs >>8, then >>4 for SRS_SHIFT-8) versus _gaussian_blur3x3_ref’s float np.round before the _EPSILON = 2.0 comparison. Since --verify currently cannot run at all, this cannot regress anything, but it does mean nobody has yet seen it pass. A single python camera/blur_pipeline.py --verify on the NPU machine would settle it, and would be worth noting in the PR body so the claim rests on a run rather than on inspection.

Thanks — this is a good catch on a path CI structurally cannot reach (the aie.iron import fails first on hosts without the toolchain, which is exactly why it stayed green while being broken). Happy to merge once (1) is reworded.