Overview

What is Holoscan OpenEmbedded Builder?

The Holoscan OpenEmbedded Builder is a containerized build environment based on the Yocto Project and OpenEmbedded framework, specifically tailored for developing and building Linux-based software for NVIDIA Holoscan Developer Kits. It leverages the OpenEmbedded build system to create custom Board Support Package (BSP) images, allowing developers to cross-compile and integrate applications optimized for Holoscan's edge AI and medical computing platforms.

This builder simplifies the setup process by packaging all required dependencies, tools, scripts, and NVIDIA-proprietary binaries into a single, reproducible Docker container—ideal for embedded Linux development targeting Holoscan hardware.

OpenEmbedded/Yocto Build Container for NVIDIA Holoscan

This container image contains runtime dependencies, scripts, and the NVIDIA-proprietary binary packages that are required to build an OpenEmbedded BSP image for NVIDIA Holoscan Developer Kits.

The following documentation provides information specific to the usage of the Holoscan Build Container, and may be missing information from the main documentation that may be useful to know when configuring or using the BSP. Please see the main README file for additional documentation.

Note: the main README file can be found at meta-tegra-holoscan/README.md after following the 1. Setting up the Local Development Environment section, below.

Also note that building a BSP for NVIDIA Holoscan requires a significant amount of resources, and at least 300GB of free disk space is required to build. See the System Requirements section in the main README for more details.

1. Setting up the Local Development Environment

While it would be possible to build an OE image directly from source that is stored within a container, doing so would mean that any additions or modifications to the build recipes would also only live inside the running container and so would be lost whenever the container terminates. Instead, this container operates by initially setting up a local host volume with all of the recipes, dependencies, and initial configuration that is needed for the BSP build such that all of the recipes, configuration, and build cache is stored in persistent storage on the host and thus is not limited to the lifespan of a single container runtime.

In order to perform this initial setup navigate to the directory into which you would like to initialize the development environment and run the following (making sure IMAGE matches the name and tag of this container image):

$ export IMAGE=nvcr.io/nvidia/clara-holoscan/holoscan-oe-builder:v3.9.0
$ docker run -it --rm -v $(pwd):/workspace --network host ${IMAGE} setup.sh ${IMAGE} $(id -u) $(id -g)

This setup processes initializes the following:

1. The OE recipes and dependencies in these folders:

   poky
   meta-openembedded
   meta-virtualization
   meta-tegra
   meta-tegra-holoscan
   

2. A sample build configuration in the build folder.

3. Wrapper script bitbake.sh, which runs a build container and passes the arguments to the script to the container's bitbake command.

4. A flash.sh script to flash the device with the built image.

5. A .buildimage file which contains the name of the container image. This is used by the bitbake.sh script and prevents the need to export an IMAGE environment variable anytime a build is performed.

2. Configure the Image

The OE image configuration file is created by the previous step and is written to build/conf/local.conf. This file is based on the default local.conf that is created by the Poky environment setup script (oe-init-build-env) and has various NVIDIA configuration defaults and samples added to it. For example, the MACHINE configuration in this template file is set to igx-orin-devkit; the GPU configration is set to use the dGPU; and CUDA, TensorRT, Holoscan SDK, and the HoloHub sample applications are installed by default. This configuration can be used as-is to build a BSP for the IGX Orin Developer Kit using the A6000 dGPU, but it may be neccessary to change this configuration to use the iGPU or to add additional components like Rivermax or support for third-party hardware such as AJA video capture cards or Emergent high-speed cameras. See the Build Configuration section in the main README for more details.

To see the additional configuration that is added to this file relative to the standard OpenEmbedded local.conf, as well as some documentation as to what additional components offered by this meta-tegra-holoscan layer may be enabled, scroll down to the "BEGIN NVIDIA CONFIGURATION" section in this file.

3. Build the Image

Once the image has been configured in the local host development tree, the container image is used again for the actual bitbake build process. This can be done using the bitbake.sh build wrapper that is written to the root of the development directory. This script simply runs the bitbake process in the container and passes the arguments to the script to this process. For example, to build a Holoscan reference image, use the following:

$ ./bitbake.sh core-image-holoscan

This build is expected to take at least an hour with build times of 3 or 4 hours being expected on machines with slower hardware or internet connections.

Note: For a list of different image targets that are available to build, see the Yocto Project Images List.

Note: If the build fails due to unavailable resource errors, try the build again. Builds are extremely resource-intensive, and having a number of particularly large tasks running in parallel can exceed even 32GB of system memory usage. Repeating the build can often reschedule the tasks so that they can succeed. If errors are still encountered, try lowering the value of BB_NUMBER_THREADS in build/conf/local.conf to reduce the maximum number of tasks that BitBake should run in parallel at any one time.

Using the default configuration, the above script will build the BSP image and write the final output to:

build/tmp/deploy/images/igx-orin-devkit/core-image-holoscan-igx-orin-devkit.tegraflash.tar.gz

4. Flash the Image

The flash.sh script can be used to flash the BSP image that is output by the previous step onto the Holoscan Developer Kit hardware. For example, to flash the core-image-holoscan image that was produced by the previous step, connect the developer kit to the host via the USB-C debug port, You should also have these commands available on your build host:

sgdisk (from the gdisk/gptfdisk package)
udisksctl (part of the udisks2 package)

You should disable automatic mounting of removable media in your desktop settings. On recent Ubuntu (GNOME), go to Settings -> Removable Media, and check the box next to "Never prompt or start programs on media insertion." You may also need to update the /org/gnome/desktop/media-handling/automount setting via dconf. Check the setting with:

$ dconf read /org/gnome/desktop/media-handling/automount

If it reports true, set it with:

$ dconf write /org/gnome/desktop/media-handling/automount false

Put it into recovery mode, ensure the developer kit is visible to the host using lsusb then run:

$ ./flash.sh core-image-holoscan

Note: If the initrd-flash command fails due to a No such file: 'dtc' error, install the device tree compiler (dtc) using the following:

$ sudo apt-get install device-tree-compiler

For instructions on how to put the developer kit into recovery mode and how to check that it is visible using lsusb, see the developer kit user guide:

• IGX Orin Developer Kit User Guide.

Note that flashing the device will require root privileges and so you may be asked for a sudo password by this script.

Once flashed, the Holoscan Developer Kit can then be disconnected from the host system and booted. A display, keyboard, and mouse should be attached to the developer kit before it is booted. The display connection depends on the GPU configuration that was used for the build: the iGPU configuration uses the onboard Tegra display connection while the dGPU configuration uses one of the connections on the discrete GPU. Please refer to the developer kit user guide for diagrams showing the locations of these display connections. During boot you will see a black screen with only a cursor for a few moments before an X11 terminal or GUI appears (depending on your image type).

Running the Holoscan SDK and HoloHub Applications

When the core-image-holoscan reference image is used, the Holoscan SDK and Holohub apps are built into the image, including some tweaks to make running the samples even easier. Upon boot, the core-image-holoscan image presents a Matchbox UI with icons for a variety of Holoscan SDK and Holohob sample applications, all of which can be run with just a single click.

Note that the first execution of these samples will rebuild the model engine files and it will take a few minutes before the application fully loads. These engine files are then cached and will significantly reduce launch times for successive executions. Check the console windows with the application logs for additional information.

While a handful of graphical Holoscan applications have icons installed on the desktop, many more are console-only and must be launched from a console.

When the holoscan-sdk component is installed, the Holoscan SDK is installed into the image in the /opt/nvidia/holoscan directory, with examples present in the examples subdirectory. Due to relative data paths being used by the apps, these examples should be run from the /opt/nvidia/holoscan directory. To run the C++ version of an example, simply run the executable in the example's cpp subdirectory:

$ cd /opt/nvidia/holoscan
$ ./examples/hello_world/cpp/hello_world

To run the Python version of an example, run the application in the example's python subdirectory using python3:

$ cd /opt/nvidia/holoscan
$ python3 ./examples/hello_world/python/hello_world.py

When the holohub-apps component is installed, the HoloHub sample applications are installed into the image in the /opt/nvidia/holohub directory, with the applications present in the applications subdirectory. Due to relative data paths being used by the apps, these applications should be run from the /opt/nvidia/holohub directory. To run the C++ version of an application, simply run the executable in the applications's cpp subdirectory:

$ cd /opt/nvidia/holohub
$ ./applications/endoscopy_tool_tracking/cpp/endoscopy_tool_tracking

To run the Python version of an application, run the application in the python subdirectory using python3:

$ cd /opt/nvidia/holohub
$ python3 ./applications/endoscopy_tool_tracking/python/endoscopy_tool_tracking.py

License

By pulling and using the container, you accept the terms and conditions of this End User License Agreement.