michaelgloff / meta-tegra

BSP layer for NVIDIA Tegra, based on L4T

OpenEmbedded/Yocto BSP layer for NVIDIA Jetson TX1/TX2/AGX Xavier/Nano

Boards supported:

• Jetson-TX1 development kit (Linux4Tegra R32.3.1, JetPack 4.3)
• Jetson-TX2 development kit (Linux4Tegra R32.3.1, JetPack 4.3)
• Jetson AGX Xavier development kit (Linux4Tegra R32.3.1, JetPack 4.3)
• Jetson Nano development kit (Linux4Tegra R32.3.1, JetPack 4.3)

Also supported thanks to community support:

• Jetson-TX2i module (Linux4Tegra R32.3.1, JetPack 4.3)
• Jetson-TX2 4GB module (Linux4Tegra R32.3.1, JetPack 4.3)
• Jetson AGX Xavier 8GB module (Linux4Tegra R32.3.1, JetPack 4.3)

This layer depends on: URI: git://git.openembedded.org/openembedded-core branch: master LAYERSERIES_COMPAT: zeus

PLEASE NOTE

• Starting with JetPack 4.2, packages outside the L4T BSP can only be downloaded with an NVIDIA Developer Network login. So to use CUDA 10, cuDNN, and any other packages that require a Devnet login, you must create a Devnet account and download the JetPack packages you need for your builds using NVIDIA SDK Manager.

  You must then set the variable NVIDIA_DEVNET_MIRROR to "file://path/to/the/downloads" in your build configuration (e.g., local.conf) to make them available to your bitbake builds. This can be the NVIDIA SDK Manager downloads directory, /home/$USER/Downloads/nvidia/sdkm_downloads

  Note Starting with L4T R32.3.1 and JetPack 4.3, The Tegra Multimedia API kit has moved to JetPack, so all builds now require you to set up an SDK Manager downloads area.

• The SDK Manager downloads a different package of CUDA host-side tools depending on whether you are running Ubuntu 16.04 or 18.04. If you downloaded the Ubuntu 16.04 package, you should add

  CUDA_BINARIES_NATIVE = "cuda-binaries-ubuntu1604-native"
  

  to your build configuration so the CUDA recipes can find them. Otherwise, the recipes will default to looking for the Ubuntu 18.04 package.

• The TensorRT 6.0.1 packages for Xavier are different from those for TX1/TX2, even though the deb files have the same name. To prevent mixups during the build, the recipe here expects to find the Xavier packages in a DLA subdirectory under ${NVIDIA_DEVNET_MIRROR}, and non-Xavier packages in a NoDLA subdirectory.

  If you need to include TensorRT in your builds, you must create the subdirectory and move all of the TensorRT packages downloaded by the SDK Manager there. Xavier example:

  $ cd ~/Downloads/nvidia/sdkm_downloads
  $ mkdir DLA
  $ mv tensorrt*.deb libnvinfer*.deb DLA/
  

• CUDA 10 supports up through gcc 7 only, and some NVIDIA-provided binary libraries appear to be compiled with g++ 7 and cause linker failures when building applications with g++ 6, so only gcc 7 should be used if you intend to use CUDA. See the following wiki pages for instructions on including gcc 7 in your builds:

  Using gcc7 from the contrib layer Using linaro gcc7 for CUDA support

Contributing

Please use GitHub (https://github.com/madisongh/meta-tegra) to submit issues or pull requests, or add to the documentation on the wiki. Contributions are welcome!