[gareth-morgan-nv]

Summary

Places the SIPL sensor timestamp on the sender's CLOCK_MONOTONIC with a checked mapping and an explicit uncertainty, and carries it to receivers in an optional AVS1 timing record. Frames whose timestamp fails the checks carry no capture time; nothing is replaced by a host-side stamp. The record is off by default.

What changed

• SensorClockMap: pairs the platform counter with CLOCK_MONOTONIC once a second and converts each frame's counter timestamp from the difference. It refuses a capture after the frame was ready, an implausibly old one, an unreadable counter, an unknown frequency, or a step between two anchors, and it reports the reason when it changes. A mapped frame carries an uncertainty (the anchor's read bracket plus 500 ppm of the time between the anchor and the sensor event, the bound on how far CLOCK_MONOTONIC can be slewed against the counter) and the anchor's age as a separate freshness value.
• SIPL start-of-frame is reported as a checked platform TSC in nanoseconds in the existing AVS1 sample and bundle TSC fields (the same domain the Argus start-of-frame uses), only while the check passes. CAMERA_SENDER_SIPL_TSC_PROBE=1 logs the raw SIPL timestamps beside the counter for repeating the measurement.
• AVS1 timing extension (bundle flag 0x0008): one 24-byte record per sample after the sample descriptors, with the capture instant on CLOCK_MONOTONIC, the uncertainty, the anchor age, a VALID flag and the clock domain. A record without VALID is all zero; the encoder and parser reject the other combinations, unknown flags and nonzero reserved fields; values that do not fit 32 bits are left unqualified, never wrapped. WIRE_FORMAT.md documents it with a golden byte vector that a unit test compares against.
• network.timing_extension (default false) turns the extension on. A receiver that predates it rejects those bundles, so keep it off until the receiver parses it.

Validation

• clang-format-14 and pre-commit clean on the changed files (and pre-commit run --all-files clean on this head).
• aarch64 cross-compile with -DWITH_SIPL=ON and OFF: syntax check and full link.
• Unit tests, built and run by hand on a PC: sensor_clock (wrong units, capture after ready, implausible age, refresh schedule, a deliberate clock step, reset, unreadable counter, and the uncertainty over the anchor-to-event distance), avs1_wire (extension round trip, rejections, golden vector), bundle_assembler (extension fill, 32-bit overflow left unqualified).
• Native build and unit tests on a Jetson Orin (R39.2.1): 17 of 17 camera-free tests, on a local build that also contained the clock-exchange PR.
• Measured on the Orin with the SIPL backend: the start-of-frame value counts CNTVCT_EL0 ticks at 31.25 MHz, advances at the same rate as CLOCK_MONOTONIC (1.00368 s against 1.00369 s over 30 frames), leads the dequeue by about 22 ms, and the end-of-frame value is about 15.5 ms later.
• A real SIPL stream with the extension on, read by an independent AVS1 parser: 897 of 897 samples with a VALID record, no sequence gaps, uncertainty median about 0.23 ms and maximum 0.5 ms, and after a forced pipeline rebuild 896 of 896 again with a new session id.
• The same stream with the extension on through a receiver that parses it: 901 of 901 records valid, no sequence gaps, and a different session id after a receiver reconnect. That receiver is not part of this PR.
• Agreement of the capture instant with an independent physical measurement: not measured.

Known limitations / follow-ups

• The counter units and the start-of-frame offset were measured on one Orin and one camera; they are not taken from vendor documentation, and how the event relates to the physical exposure start and to tracking data is not established.
• Only the SIPL backend fills the record; the Argus path does not emit the extension.
• A receiver has to parse the extension before it is enabled; that is a change on the receiver side.

Test plan

• Stream with network.timing_extension: true into a receiver that parses the extension and check the per-frame capture times.
• CI.

🤖 Generated with Claude Code

[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

• Source mode: Auto
• Repository: NVIDIA/IsaacCapture
• Branch: gmorgan/CameraStreamingSender
• File: .coderabbit.yaml
• Commit: 42a5c25
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 5a39268f-2141-493d-be45-d118876a154e

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

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts