GPU delegate linker errors when building TFLite 2.16.1 for Android with CMake about tensorflow

[LakshmiKalaKadali]

Hi @eltimen,

Enabling GPU support requires additional libraries and dependencies related to OpenCL. This might be the reason for
the error tflite::gpu::gl and tflite::gpu::OptionalAndroidHardwareBuffer. Please try to add the following to the Android manifest in order to detect GPU delegate.

<uses-library android:name="libOpenCL.so"
           android:required="false"/>

<uses-library android:name="libOpenCL-pixel.so"
           android:required="false"/>

Also Verify that the NDK version (r25b) is compatible with the TensorFlow Lite version you are trying to build.

Thank You

[eltimen]

Hi @LakshmiKalaKadali,

I'm not using Gradle or Android Studio. I just tried to build libtensorflow-lite.so for Android using CMake following the documentation and I got the linker errors above.

I tried to build it with Bazel as mentioned here and build was successful. According to this page NDK r25b should be compatible with this TFLite version.

According to the documentation, TensorFlow Lite supports both ways to build it for Android: with CMake and with Bazel. Looks like CMake building is broken with enabled GPU delegate in this version.

[pkgoogle]

This looks like a duplicate of #65744, investigating.

[pkgoogle]

Hi @eltimen, I was able to build fine with these steps:

git clone https://github.com/tensorflow/tensorflow.git tensorflow_src
cd tensorflow_src
git switch r2.16
cd ..
mkdir tflite_build
cd tflite_build
cmake -S ../tensorflow_src/tensorflow/lite -B build -D CMAKE_BUILD_TYPE=Release -D CMAKE_SYSTEM_PROCESSOR=aarch64 -D BUILD_SHARED_LIBS=ON -D TFLITE_ENABLE_GPU=ON -D CMAKE_SYSTEM_NAME=Linux -D CMAKE_SYSTEM_VERSION=29 -D ANDROID_PLATFORM=29 -D CMAKE_ANDROID_ARCH_ABI=arm64-v8a -D ANDROID_ABI=arm64-v8a -D CMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake

Please note $ANDROID_NDK is the path to my NDK, which is 25b.

I am building on MacOS, you didn't specify an OS so that could be an issue.

Can you please let us know what OS you are using? Additionally please review my reproduce steps and let me know if I did something different than what you are doing? Thanks.

[eltimen]

Hi @pkgoogle, I use Ubuntu 22.04.
I'm building TFLite from the tag v2.16.1, but you are building it from the branch r2.16.
Can you try to reproduce this using the release 2.16.1 version?

[pkgoogle]

Hi @eltimen, I am using a Debian system, and went to that tag, but am using NDK=r25c, I am still not reproducing the error:

git clone https://github.com/tensorflow/tensorflow.git tensorflow_src
cd tensorflow_src
git checkout v2.16.1
cd ..
mkdir tflite_build
cd tflite_build
cmake -S ../tensorflow_src/tensorflow/lite -B build -D CMAKE_BUILD_TYPE=Release -D CMAKE_SYSTEM_PROCESSOR=aarch64 -D BUILD_SHARED_LIBS=ON -D TFLITE_ENABLE_GPU=ON -D CMAKE_SYSTEM_NAME=Linux -D CMAKE_SYSTEM_VERSION=29 -D ANDROID_PLATFORM=29 -D CMAKE_ANDROID_ARCH_ABI=arm64-v8a -D ANDROID_ABI=arm64-v8a -D CMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake

It looks like my Clang is different (14.0.7) probably from the different r25c.

Can you try with a fresh repo, perhaps fresh NDK, and ensure you follow the same steps exactly? Let me know if somehow that resolves your issue.