libcamera-apps doesn't build with tensorflow-lite

[vivien]

With the meta-tensorflow layer added to bblayers, enabling TFLITE in the libcamera-apps build will result in:

fatal error: tensorflow/lite/builtin_op_data.h: No such file or directory
|    16 | #include "tensorflow/lite/builtin_op_data.h"
|       |          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
| compilation terminated.

This happens regardless of having tensorflow-lite listed as a build dependency or not, as fixed by 2cdb97a.

[vivien]

As of 76263ba, libcamera-apps builds for raspberrypi3 and raspberrypi4, but tensorflow-lite fails for raspberrypi with:

Linking tensorflow/lite/libtensorflowlite.so failed: (Exit 1): arm-poky-linux-gnueabi-gcc failed: error executing command

[vivien]

Even though libcamera-apps now sets the dependency correctly with:

PACKAGECONFIG[tflite] = "-Denable_tflite=true, -Denable_tflite=false, tensorflow-lite"

the configure stage for raspberrypi4 ends up with:

Run-time dependency tensorflow-lite found: NO (tried pkgconfig and cmake)
...
Build configuration
...
    TFLite postprocessing: NO
...
  User defined options
...
    enable_tflite        : true

Any idea why the libcamera-apps recipe doesn't detect the tensorflow-lite library?

[kraj]

Run-time dependency tensorflow-lite found: NO (tried pkgconfig and cmake)

there will be some .txt files for errors generated by cmake in builddir. So please share those.

[vivien]

The package is now using meson instead of cmake. Here's the related snippet from /build/tmp/work/cortexa7t2hf-neon-vfpv4-poky-linux-gnueabi/libcamera-apps/1.2.1+gitAUTOINC+1c1d1c1a2a-r0/build/meson-logs/meson-log.txt:

Determining dependency 'tensorflow-lite' with pkg-config executable '/build/tmp/work/cortexa7t2hf-neon-vfpv4-poky-linux-gnueabi/libcamera-apps/1.2.1+gitAUTOINC+1c1d1c1a2a-r0/recipe-sysroot-native/usr/bin/pkg-config'
env[PKG_CONFIG_DISABLE_UNINSTALLED]: yes
env[PKG_CONFIG_SYSTEM_LIBRARY_PATH]: /lib:/usr/lib
env[PKG_CONFIG_PATH]: /build/tmp/work/cortexa7t2hf-neon-vfpv4-poky-linux-gnueabi/libcamera-apps/1.2.1+gitAUTOINC+1c1d1c1a2a-r0/recipe-sysroot/usr/lib/pkgconfig:/build/tmp/work/cortexa7t2hf-neon-vfpv4-poky-linux-gnueabi/libcamera-apps/1.2.1+gitAUTOINC+1c1d1c1a2a-r0/recipe-sysroot/usr/share/pkgconfig
env[PKG_CONFIG_DIR]: /build/tmp/work/cortexa7t2hf-neon-vfpv4-poky-linux-gnueabi/libcamera-apps/1.2.1+gitAUTOINC+1c1d1c1a2a-r0/recipe-sysroot/usr/lib/pkgconfig
env[PKG_CONFIG_SYSROOT_DIR]: /build/tmp/work/cortexa7t2hf-neon-vfpv4-poky-linux-gnueabi/libcamera-apps/1.2.1+gitAUTOINC+1c1d1c1a2a-r0/recipe-sysroot
env[PKG_CONFIG_LIBDIR]: /build/tmp/work/cortexa7t2hf-neon-vfpv4-poky-linux-gnueabi/libcamera-apps/1.2.1+gitAUTOINC+1c1d1c1a2a-r0/recipe-sysroot/usr/lib/pkgconfig
env[PKG_CONFIG_SYSTEM_INCLUDE_PATH]: /usr/include
Called `/build/tmp/work/cortexa7t2hf-neon-vfpv4-poky-linux-gnueabi/libcamera-apps/1.2.1+gitAUTOINC+1c1d1c1a2a-r0/recipe-sysroot-native/usr/bin/pkg-config --modversion tensorflow-lite` -> 1
stderr:
Package tensorflow-lite was not found in the pkg-config search path.
Perhaps you should add the directory containing `tensorflow-lite.pc'
to the PKG_CONFIG_PATH environment variable
No package 'tensorflow-lite' found

Adding inherit pkgconfig to the tensorflow-lite recipe results in the same error.

[kraj]

you might want to look at tensorflow-lite recipe and see why it is not generating and/or packaging the .pc file

[vivien]

So tensorflow-lite from meta-tensorflow is not installing any tensorflow-lite.pc file, it isn't installing the header files neither. While this can be fixed, it doesn't provide the missing flatbuffers files too. I switched to tensorflow-lite from meta-neural-network, which works with a few fixups. I now have libcamera-detect on raspberrypi4-64.

[kraj]

interesting, I dont know which layer should be considered more official some indicators I use is activity and on that meta-tensorflow is on https://git.yoctoproject.org/meta-tensorflow so it should be the one people usually will default to and therefore should be fixed. I would suggest to inform the meta-tensorflow maintainers about this issue. meta-neural-network is for NNstreamer and they might maintain tensorflow-lite in a narrow sense of just NNstreamer which makes sense for what the layer is about.

btw. there is https://github.com/NobuoTsukamoto/meta-tensorflow-lite as well so you have choices like in a fish market :)

[agherzan]

So this doesn't look RPi-specific, does it?

[vivien]

Indeed, these are issues in meta-tensorflow and compilation issue on 32-bit ARM. I'm closing this issue for now but I'll push a few tweaks for libcamera-detect integration soon though.