* meta-ml was renamed to meta-imx-ml, and its BBFILE_COLLECTIONS was renamed to
imx-machine-learning
* meta-st-stm32mpu-ai was renamed to meta-st-x-linux-ai
https://onedigi.atlassian.net/browse/DEL-9011
Signed-off-by: Gabriel Valcazar <gabriel.valcazar@digi.com>
The recipe level DEPENDS on ethos-u-vela-native is insufficient to ensure
the native vela tool population into the eiq-examples recipe sysroot
native directory before use. The new download_transform_models task uses
that tool, so add the dependence on the task to ensure it is available
before use.
While on it, configure the working directory of the new task using the
Yocto way (using task[dirs]).
Signed-off-by: Javier Viguera <javier.viguera@digi.com>
* do not run the demos in background, if not they can not be ctrl+c.
* fix service dependencies: required camera and weston.
* patch the examples to:
** set camera resolution to 1280x720
** set default format to ov5640 camera (YUYV)
** detect the "x" event when the demo application is running in windowed mode (not fullscreen).
Signed-off-by: Isaac Hermida <isaac.hermida@digi.com>
Until now, the recipe only installed the binaries and some scripts in the target device. Then,
user had to execute an script in the running system to download and transform the models for
the demo, a process that could take up to 30 minutes. Also, a custom script was required to
acomodate the environment settings before executing any demo. All of this was preventing us from
creating out-of-the-box images including NPU demos for exhibitions and clients.
With the new changes, the recipe now downloads and transforms all the models in the host machine,
saving around 27 minutes and leaving the demos in the target ready to be executed. To ease the
process, a launch script is included for all the demos to configure the environment as well as an
auto-start service.
https://onedigi.atlassian.net/browse/CCS-10
Co-authored-by: Isaac Hermida <isaac.hermida@digi.com>
Signed-off-by: David Escalona <david.escalona@digi.com>
Rework commit c5c9838e54 to only limit ML
packages for our ccimx93 and not for other imx93-based devices.
Signed-off-by: Javier Viguera <javier.viguera@digi.com>
This patch fixes the hang issue with EiQ demos using multiple tflite files,
for instance the gesture_detection demo.
https://onedigi.atlassian.net/browse/DEL-8949
Signed-off-by: Isaac Hermida <isaac.hermida@digi.com>
Not installing all the ML packages but just tensorflow-lite saves space.
We are not including onnxruntime and torchvision, which are not supported
by the i.mx93 (see NXP user guide for details).
The ext4.gz size is decreased from 430MB to 217MB.
Signed-off-by: Isaac Hermida <isaac.hermida@digi.com>
As part of the integration of the new ML package, also update the
ethos-u-firmware binary built from Stash:
Repo: emp/ethos_u_firmware.git
Revision: bd5506ddba364ad04602d5009b77077f78450b97
Source: NXP's MCUXpresso SDK_2.14.2_MIMX9352xxxxM
Co-authored-by: Javier Viguera <javier.viguera@digi.com>
Signed-off-by: Isaac Hermida <isaac.hermida@digi.com>
Signed-off-by: Javier Viguera <javier.viguera@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 (updated in
the following commit).
Signed-off-by: Javier Viguera <javier.viguera@digi.com>
As part of the integration of the new ML package, also update the
ethos-u-firmware binary built from Stash:
Repo: emp/ethos_u_firmware.git
Revision: a0352b0cf7267c896b55980dedd08daddc780733
Source: NXP's MCUXpresso SDK_2.14.0_MIMX9352xxxxK
And, drop deepview-rt support for ccimx93, as NXP has done the same in
6.1.36-2.1.0 release.
Signed-off-by: Javier Viguera <javier.viguera@digi.com>
The default ethosu_firmware in github reconfigures the uart2 to be used as
debug port. Those pins are used to manage the HW flow control for the Bluetooth
chip.
Build a custom ethosu_firmware that does not reconfigure the uart2 pins.
Compiled from internal "ethos_u_firmware.git" repository on Stash,
with revision 48a4b9e6a0770212aac529fb7f81ed2e6ff51cbd
md5sum: 0bf8686fcabec7378483755106dc6433
https://onedigi.atlassian.net/browse/DEL-8587
Signed-off-by: Isaac Hermida <isaac.hermida@digi.com>
Backport from NXP's lf-6.1.1-1.0.0 release for the ccimx93 (meta-ml
layer).
This version of flatbuffers is needed for ethos-u-vela version 3.6.0
(backported in following commit).
Signed-off-by: Javier Viguera <javier.viguera@digi.com>
Commit d2c1494bbf36b6392e47ffd4a75307d29681d190 in poky adds this variable
to EXTRA_OESCONS, which breaks the do_compile() task of this recipe. Remove
the variable to be able to build the package.
Signed-off-by: Gabriel Valcazar <gabriel.valcazar@digi.com>