Actually, the NPU demos were not making use of the NPU, just the CPU. To enable the NPU usage,
it is required to add the external delegate library 'tflite-vx-delegate'to the system. This is
done by forcing 'tensorflow-lite' to depend on it for the CCMP2 platform, which also requires
'tim-vx' library as a build dependency.
Both the 'tflite-vx-delegate' and 'tim-vx' recipes have been verbatim copied from the ST Beta
branch, using commit with SHA256: 68686850c75061f1c7c4e756a313a41ca810f6ae
Signed-off-by: David Escalona <david.escalona@digi.com>
Customize the NPU samples provided by ST so they can be executed in our DEY images:
- Add a generic launch script to pre-configure environment variables before launching demos.
- Remove 'weston' user check from demos launch scripts.
- Reduce font size for big screens.
- Set camera resolution to 640x480px and center the image.
- Adapt home folder to 'root' user instead of 'weston'.
- Adapt package group to include the new samples and remove the old ones.
- Fix a couple of errors with models download and include them manually.
- Add a runtime dependency with 'libdrm-tests' as the 'modetest' binary is required.
https://onedigi.atlassian.net/browse/CCS-12
Signed-off-by: David Escalona <david.escalona@digi.com>
Current samples were not working properly due to problems with libraries and
the runtime frameworks.
This commit brings the updated samples recipes with the new folder structure
that have been proved to work as they were in the st beta branch, using commit
with SHA256: 68686850c75061f1c7c4e756a313a41ca810f6ae
https://onedigi.atlassian.net/browse/CCS-12
Signed-off-by: David Escalona <david.escalona@digi.com>
Current recipe build was failing due to dependencies with cpuinfo header files and
libraries. This commit backports the recipe that is currently being built in the
ST main branch in GitHub which clears these dependencies.
The recipe was backported as it was in commit with SHA256:
fc56201c67adf899080cff5b45ed33824e427609
https://onedigi.atlassian.net/browse/CCS-12
Signed-off-by: David Escalona <david.escalona@digi.com>
Current recipe build was failing due to dependencies with cpuinfo header files and
libraries. This commit backports the recipe that is currently being built in the
ST main branch in GitHub which clears these dependencies.
The recipe was backported as it was in commit with SHA256:
fc56201c67adf899080cff5b45ed33824e427609
https://onedigi.atlassian.net/browse/CCS-12
Signed-off-by: David Escalona <david.escalona@digi.com>
Backport a new version of cmake-native from Poky's mickledore release. This
is required by the new version of the onnxruntime package (backported in
the following commit).
https://onedigi.atlassian.net/browse/CCS-12
Signed-off-by: David Escalona <david.escalona@digi.com>
These examples have a gtk+3 dependency that can't be met on the ccmp13 due to
it being a headless device. Remove these examples from the packagegroups so we
can at least build the rest of the ML packages.
Note that all of ST's ML examples are GUI-based and will not be usable on the
ccmp13, but since the remaining examples don't have an explicit gtk+3
dependency, at least they won't trigger build errors when included in the
image.
https://onedigi.atlassian.net/browse/DEL-8308
Signed-off-by: Gabriel Valcazar <gabriel.valcazar@digi.com>
ST's onnxruntime recipe moves a file in a way that triggers a QA error due to
file ownership issues. Copy the do_install() function as-is and modify the
offending line to copy the file instead. Create a dynamic layer for stm32mpu-ai
to include the bbappend.
https://onedigi.atlassian.net/browse/DEL-8308
Signed-off-by: Gabriel Valcazar <gabriel.valcazar@digi.com>
David Escalona
Isaac Hermida
Gabriel Valcazar