perception_pipeline

GPU targets

Arch	Hardware	Notes
sm_87	Jetson AGX Orin / Orin NX / Orin Nano	JetPack 6 userspace is Ubuntu 22.04
sm_120	Blackwell, RTX 50-series	Requires CUDA >= 12.8

CMAKE_CUDA_ARCHITECTURES defaults to 87-real;120-real;120-virtual: native machine code (SASS) for both targets so they load without JIT, plus PTX for Blackwell so the binary still runs on future architectures. Overriding it is how you trim build time while iterating — see below.

Camera access from the container (GigE)

Three things must be set on the host; a container cannot change them, and each shows up as dropped or incomplete frames rather than an obvious error:

# Receive buffers. The default ~200 KB is far too small for a GigE stream.
sudo sysctl -w net.core.rmem_max=10485760
sudo sysctl -w net.core.rmem_default=10485760

# Jumbo frames on the camera NIC (replace enp3s0). Without this you pay a
# per-packet overhead on ~1500-byte payloads and lose bandwidth headroom.
sudo ip link set enp3s0 mtu 9000

Make them permanent via /etc/sysctl.d/ and your netplan/NetworkManager config once the rig is settled.

Build the image

docker build --build-context spinnaker=/opt/spinnaker -t perception_pipeline:dev .

The first run pulls ~2 GB of CUDA base image. Compilation smoke test happens during the image build, so a successful docker build means the CUDA toolchain works.

--build-context spinnaker=... points at a host install of the Spinnaker SDK, which is baked into the image at /opt/spinnaker.

Useful overrides:

# Iterate faster by building for one arch only
docker build --build-arg CUDA_ARCHS=87-real ...

# Pin a different CUDA / Ubuntu base
docker build --build-arg CUDA_IMAGE=nvidia/cuda:12.6.3-devel-ubuntu22.04 ...

# See full compiler output instead of collapsed log lines
docker build --progress=plain ...

Run the smoke tests

smoke/ holds two toolchain checks — they verify that CUDA and Spinnaker are found, compile, link, and load at runtime. They test the environment, not pipeline behaviour. The image's default command runs both through ctest:

docker run --rm --network host perception_pipeline:dev

1/2 Test #1: cuda_smoke .......................   Passed
2/2 Test #2: spinnaker_smoke ..................   Passed

Or run them individually:

docker run --rm --network host perception_pipeline:dev ./build/bin/spinnaker_smoke
docker run --rm perception_pipeline:dev ./build/bin/cuda_smoke

Development loop

Rebuilding the image for every source edit is slow. Bind-mount the working tree into the container instead:

docker run --rm -it -h perception --network host \
  -v "$PWD":/workspace -w /workspace perception_pipeline:dev bash

Then, inside the container:

cmake -S . -B build -G Ninja \
  -DCMAKE_CUDA_ARCHITECTURES="87-real;120-real;120-virtual"
cmake --build build
ctest --test-dir build --output-on-failure