This application demonstrates how to import existing GXF Codelets and Components as Holoscan Operators/Resources. The example code is adapted from the Tensor interoperability example (examples/tensor_interop) and modified to illustrate the use of GXFCodeletOp and GXFComponentResource classes to import GXF Codelet/Component and customize the setup() and initialize() methods.

The main components include a set of custom operators that encapsulate GXF Codelets for sending and receiving tensors, and a resource class for managing block memory pools.

• GXFSendTensorOp: An operator that wraps a GXF Codelet responsible for sending tensors.
• GXFReceiveTensorOp: Extends GXFCodeletOp to wrap a GXF Codelet that receives tensors, with customizable setup and initialization.
• MyBlockMemoryPool: Represents a resource wrapping the nvidia::gxf::BlockMemoryPool GXF component, used to manage memory allocation.

• using deb package install or NGC container:

  /opt/nvidia/holoscan/examples/import_gxf_components/cpp/import_gxf_components

• source (dev container):

  ./run launch # optional: append `install` for install tree
  ./examples/import_gxf_components/cpp/import_gxf_components

• source (local env):

  ${BUILD_OR_INSTALL_DIR}/examples/import_gxf_components/cpp/import_gxf_components

Python example shows how to utilize the GXFCodeletOp and GXFComponentResource in a Holoscan application, focusing on video processing using custom and built-in operators.

• ToDeviceMemoryOp: A derived class from GXFCodeletOp for copying tensors to device memory.
• ImageProcessingOp: Processes video frames using Gaussian filtering with CuPy.
• DeviceMemoryPool: Manages device memory allocations.

• using python wheel:

  # [Prerequisite] Download NGC dataset above to `DATA_DIR`
  export HOLOSCAN_INPUT_PATH=<DATA_DIR>
  # [Prerequisite] Download example .py and .yaml file below to `APP_DIR`
  # [Optional] Start the virtualenv where holoscan is installed
  python3 -m pip install cupy-cuda12x
  python3 <APP_DIR>/import_gxf_components.py

• from NGC container:

  python3 /opt/nvidia/holoscan/examples/import_gxf_components/python/import_gxf_components.py

• source (dev container):

  ./run launch # optional: append `install` for install tree
  python3 ./examples/import_gxf_components/python/import_gxf_components.py

• source (local env):

  python3 -m pip install cupy-cuda12x
  export PYTHONPATH=${BUILD_OR_INSTALL_DIR}/python/lib
  export HOLOSCAN_INPUT_PATH=${SRC_DIR}/data
  python3 ${BUILD_OR_INSTALL_DIR}/examples/import_gxf_components/python/import_gxf_components.py