Delegate creation failure

General Discussion
,
#1

Hi,
I’m using tensor flow 2.1.0 and I would like to add a GPU implementation.
Without GPU, everything works fine. However, I have some issues with creating a delegate.

Delegate = TfLiteGpuDelegateCreate(nullptr);
auto res = Interperter->ModifyGraphWithDelegate(Delegate);

After that - res != kTfLiteOk (res equals 1 instead).

Can you please support?

Thanks!

#3

Do you have a specific reason for using TF 2.1? If not, please use the latest version TF 2.10 to
create the delegate with TfLiteGpuDelegateV2Create()

// NEW: Prepare GPU delegate.
auto* delegate = TfLiteGpuDelegateV2Create(/*default options=*/nullptr);
if (interpreter->ModifyGraphWithDelegate(delegate) != kTfLiteOk) return false;

Verify if your device indeed has a supported SoC. Run adb shell cat /proc/cpuinfo | grep Hardware

Enable GPU support by rooting adb shell setenforce 0

Thank you

#7

Please refer to Build from Source and GPU acceleration delegate. Thank you

#11

In TF 2.10, the commands to build GPU delegates are as shown below for android arm64.

bazel build -c opt --config android_arm64 tensorflow/lite/delegates/gpu:delegate        # for static library
bazel build -c opt --config android_arm64 tensorflow/lite/delegates/gpu:libtensorflowlite_gpu_delegate.so  # for dynamic library

You can choose other flags armv7, armv8 as per your android architecture.

Once you build a delegate you have to add the delegate through the interpreter option. Thank you!