[gareth-morgan-nv]

Summary

Adds SIPL camera support to the camera_streamer sender. Each ISP0 frame stays a native NV12 surface and is queued to the hardware HEVC encoder through its DMA-BUF: no RGBA conversion and no CPU pixel copy. The stream uses the existing AVS1/TCP link, so the unchanged receiver decodes it into its usual CUDA IPC buffers.

The branch is three layers, in this order:

1. A vendor-neutral capture seam, with Argus types removed from the generic encoder and timing code.
2. The SIPL capture backend and its config.
3. A shared, SDK-free sender loop that consumes any CameraCapture; SIPL uses it.

What changed

• CaptureFrame / CameraCapture: an SDK-free contract (camera descriptors, per-camera sequence, session identity, raw sensor time with an explicit clock domain, a native NV12 DMA-BUF surface, and a move-only lease that returns the buffer to its backend).
• capture_backend: auto | argus | sipl in the config, with a selection function that fails with a clear message for a backend that is not built in (and for auto when more than one is).
• ArgusFrameAdapter describes an Argus DMA-BUF capture as a CaptureFrame.
• DmaBuffer no longer carries Argus types: an owner hook (DmaBufferOwner, releaseToOwner) replaces the Argus stream/buffer fields; the Argus pairing and its acquire/release rules move to ArgusBufferBinding. DmaBuffer::createExternal wraps a DMA-BUF the sender does not own.
• FrameTiming and the sender clock readers move from StereoTiming.h into an SDK-free FrameTiming.h.
• SiplCapture implements CameraCapture: platform query, ICP and ISP0 buffer allocation, NvSci buffer and sync setup, tuning registration, frame and pipeline-notification threads. A frame replaced before the consumer takes it goes straight back to SIPL.
• CaptureSender (runCaptureSender): TCP accept and send, AVS1 stream templates and bundle assembly, one HEVC encoder per stream, and for each frame waitForFrame -> acquireLatest -> wrap the DMA-BUF -> bundle registration -> encoder queue, with the lease released when the encoder dequeues the input. A frame that arrives with no free encoder input is dropped rather than waited for. EncoderInputSlots holds the in-flight bookkeeping as a header-only template so it is unit-testable without the encoder SDK. SiplSender is now a thin wrapper that maps streams to SIPL pipelines.
• Config: a sipl: section (platform JSON and name, link masks, tuning directory, ISP0 buffer count); stream device is a SIPL pipeline index. network.send_buffer_kb (default 0, off) caps the kernel TCP send buffer.
• WITH_SIPL CMake option (off by default) builds SIPL as a separate C++17 library; the rest of the sender stays C++14 and Argus-only builds are unchanged.

Validation

• clang-format-14 and pre-commit (CI hook set, copyright-year hook skipped as in CI) clean on the changed files of each layer.
• aarch64 cross-compile: clang syntax check and a full link of argus_sender, with -DWITH_SIPL=ON and OFF.
• Unit tests that need no Jetson headers, built and run by hand on a PC: config, capture_backend, avs1_wire, bundle_assembler (stubbed clock readers) and the new encoder_input_slots (lease held while in flight, out-of-order completion, lease dropped before its wrapper, double and out-of-range completion rejected, release-all, release on destruction), including a run under address and undefined-behavior sanitizers.
• Native build with WITH_SIPL=ON and the camera-free unit tests (17 of 17) on a Jetson Orin (R39.2.1). That run was on a local build that also contained the two PRs stacked after this one, not on this branch alone.
• A real SIPL stream (one SHW5G camera, 2560x1984 at 30 FPS, HEVC) through the existing camera_receiver and CUDA IPC consumer: about 30 fps decoded with no decode errors at 4 Mbit/s over WiFi. Also run on that combined build: a receiver disconnect and reconnect rebuilt the pipeline and the stream decoded again.
• The Argus capture path on a camera. The test Orin has only the SIPL rig, so the DmaBuffer change that touches the Argus buffer lifetime has been compiled and linked but not run against an Argus camera.
• More than one camera: only SIPL pipeline 0 was streamed.

Known limitations / follow-ups

• The Argus loop (runConfiguredCapture) does not consume CameraCapture and is not changed: its buffers are client-owned encoder DMA-BUFs returned with releaseToOwner, it acquires per capture domain, one Argus bundle can span several streams and stereo pairs, and raw NV12 output has its own copy path. Only the SIPL half goes through the shared sender.
• The frame-drop decision and the drain, end-of-stream and shutdown order call the encoder directly, so a hardware run covers them, not a unit test.
• Throughput is link-bound: the test WiFi link carried about 8.4 Mbit/s, so 25 Mbit/s of HEVC backed up for seconds and 4 Mbit/s ran in real time; send_buffer_kb did not change that.
• Whether the encoder accepts the NvSci-backed block-linear surface through its DMA-BUF descriptor was established by running it on this Orin and driver; no comparison with a VIC copy was made.
• On a receiver disconnect the NVENC capture plane logs "Timed out waiting for dqthread" and a broken-pipe error before the pipeline rebuilds; the stream recovers.
• The component directory and binary are still named argus_sender.

Test plan

• Build with -DWITH_SIPL=ON on an Orin with the SIPL SDK and stream with a platform JSON for your rig: same startup lines, bundles decode, a clean Ctrl-C, and a receiver-disconnect rebuild.
• Run the Argus smoke test on an Argus-capable board.
• 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: teng/5ef2afa
• File: .coderabbit.yaml
• Commit: afdf8e8
• Layers: Repository YAML
• Review profile: CHILL
• Plan: Enterprise
• Run ID: a8c136a6-a6fa-4a38-a0f0-a46070b64984

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