Add an experimental docker image

[kazuki0824]

Run CPU only tasks on any hosts (ex. Builds, Tests)

DOCKER_BUILDKIT=1 docker build . -t perception_build_test -f Dockerfile.perceptioncamtest
docker run --rm -it --platform linux/arm64/v8  --network host -e DISPLAY=$DISPLAY -v /tmp/.X11-unix/:/tmp/.X11-unix perception_build_test

and the run this in the container.

# colcon build --continue-on-error

Run with GPU(still partway, only works on a real Jetson host)

docker build . -t perception_build_test -f Dockerfile.perceptioncamtest
docker run --rm -it --gpus all --platform linux/arm64/v8  --network host -e DISPLAY=$DISPLAY -v /tmp/.X11-unix/:/tmp/.X11-unix perception_build_test

and, for example, the run this in the container.

cp -r /usr/local/cuda/samples /tmp
cd /tmp/samples/1_Utilities/deviceQuery
make -j`nproc`

/tmp/samples/bin/aarch64/linux/release/deviceQuery 

If you use GUI, run xhost +local:root and xhost -local:root as non-root.

[kazuki0824]

Known issues when using with x86 + GPU

If you run gpu codes in the container on x86 host, it won't work. I suspect CUDA in L4T is ONLY compatible with Tegra chips.

Run,

cp -r /usr/local/cuda/samples /tmp
cd /tmp/samples/1_Utilities/deviceQuery
make -j`nproc`

/tmp/samples/bin/aarch64/linux/release/deviceQuery 

then it emits an error below,

root@devenv:/tmp/samples/1_Utilities/deviceQuery# /tmp/samples/bin/aarch64/linux/release/deviceQuery 
/tmp/samples/bin/aarch64/linux/release/deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

cudaGetDeviceCount returned 35
-> CUDA driver version is insufficient for CUDA runtime version
Result = FAIL

.

Next, run apt install nvidia-container-csv* -y and run /tmp/samples/bin/aarch64/linux/release/deviceQuery . It will get

root@devenv:/tmp/samples/1_Utilities/deviceQuery# /tmp/samples/bin/aarch64/linux/release/deviceQuery 
/tmp/samples/bin/aarch64/linux/release/deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

NvRmPrivGetChipIdLimited: Could not read Tegra chip id/rev 
Expected on kernels without fuse support, using Tegra K1
NvRmPrivGetChipPlatform: Could not read platform information 
Expected on kernels without fuse support, using silicon
libnvrm_gpu.so: NvRmGpuLibOpen failed
cudaGetDeviceCount returned 999
-> unknown error
Result = FAIL

But if you uncomment

ouxt_automation/docker/docker_testbench/Dockerfile.perceptioncamtest

Lines 58 to 59 in 632ca64

	#RUN apt install nvidia-container -y
	#RUN apt install nvidia-container-csv-* -y

and rebuild , and then run /tmp/samples/bin/aarch64/linux/release/deviceQuery in the container, the error code is 35, not 999.

[kazuki0824]

@hakuturu583 please ensure it can "colcon build" and "colcon test" on x86 correctly. If successful, tensorrt_yolox will emit compilation errors and others can be built and tested properly

[hakuturu583]

Can you add documentation here?
https://github.com/OUXT-Polaris/ouxt_automation/tree/master/docs

[hakuturu583]

Please add document.