bayertap

A passive conformance tap for bayerlink links.

In a streaming netlist, a tap is a sink that observes a stream without stalling it. This is that concept as bench equipment: point it at any V4L2 capture device carrying a bayerlink source and it tells you whether the bytes survive — judged with the same published codec the source encoded with, so the two ends cannot disagree about what conformance means.

bayertap probe                          # which byte-lane permutation is this rig?
bayertap --lane-map 2,1,0 check --pattern counting     # every sample, judged
bayertap --via tunnel check --pattern counting          # through a cheap dongle
bayertap --from-file frame.npy check    # no hardware at all

Capture tiers

The drawer tier deserves its sentence: a $10 dongle you already own becomes a working conformance receiver today, because bayerlink's luma tunnel carries the unchanged container as grey levels with a pilot line the decoder learns the channel from. Slow, and completely sufficient for proving bytes.

The two bring-up commands

probe resolves the byte-lane permutation — which captured channel holds container byte k — by trying all six against the header's magic and CRC. Its output is directly the lane_map for this tool's check and for the FPGA-side receiver (np2hw.bayerlink_in). One command instead of a scope.

check --pattern X regenerates the pattern locally (patterns are pure functions — nothing travels out of band) and compares every sample, watches frame_seq for gaps, tolerates scanout repeats (protocol-legal), and exits nonzero on any discrepancy. Point CI at it if you like.

For the TC358743 tier, remember the bridge needs an EDID loaded (v4l2-ctl --set-edid) before it captures — and that EDID is leverage: it is what the source reads, so advertise only RGB 4:4:4 at your one mode and the source is steered into the only format the protocol accepts.

Status

Everything except the ioctl layer is proven off-target: struct sizes pinned against the kernel uapi, both decode paths (direct with lane maps, tunnel with pilot learning) tested end-to-end from files. First runs against real capture silicon are the bench session; the known-good device table starts there.

Licence

MIT.