STM recently updated their AI layer from v6.1.0 to v6.1.1, so our v6.1.0
bbappend wasn't being applied anymore. Use a wildcard for the hotfix version
number in case this happens again in the future.
Signed-off-by: Gabriel Valcazar <gabriel.valcazar@digi.com>
This commit introduces the necessary changes in the Digi Embedded Yocto layer
to support the X-LINUX-AI v6.1.0 software package from the meta-st-x-linux-ai
layer.
https://onedigi.atlassian.net/browse/DEL-9734
Signed-off-by: Arturo Buzarra <arturo.buzarra@digi.com>
This commit introduces the necessary changes in the Digi Embedded Yocto layer
to support the X-LINUX-AI v6.0.1 software package from the meta-st-x-linux-ai
layer.
The update removes support that is now provided directly by the ST layer in
this new version and adapts existing recipes to align with the updated
mechanisms for AI application integration. These changes ensure compatibility
with platforms that feature hardware acceleration (NPU), as well as those that
rely solely on CPU-based inference.
Signed-off-by: Arturo Buzarra <arturo.buzarra@digi.com>
stai-mpu provides a common ML API for stm32mpu platforms, allowing applications
to use different inference frameworks without changing any code. Its packages
consist of pre-built libraries that are in charge of interfacing with the
three supported frameworks (tflite, onnx and openvx).
While it's possible to install only a subset of these libraries to limit the
supported frameworks on the target device, all libraries are packaged
unconditionally at build-time, pulling in the dependencies of all three
frameworks. In our case, this pulls in onnxruntime dependencies even though we
only install the tflite and openvx packages in our ccmp25 images. This tends to
cause out-of-memory errors when onnxruntime is built at the same time as other
large packages, such as opencv and wpewebkit.
Since most of these dependencies are already present at a higher level in the
x-linux-ai packagegroup recipes, remove them from the stai-mpu recipe and skip
the runtime dependency QA checks to be able to build the packages without their
respective frameworks. In the case of the ccmp25-dvk, this has no effect on the
final image contents, but it prevents onnxruntime from being built for no
reason.
Signed-off-by: Gabriel Valcazar <gabriel.valcazar@digi.com>
Gabriel Valcazar
Arturo Buzarra