UnsatisfiedLinkError: No implementation found for void org.tensorflow.lite.NativeInterpreterWrapper.allowBufferHandleOutput(long, boolean)

[alexcohn]

System information

• OS Platform and Distribution: Linux 4.4.0-17134-Microsoft swig command not found #706-Microsoft x86_64 GNU/Linux (Ubuntu 1804)
• Mobile device: Android, arm64
• TensorFlow installed from (source or binary): source and binary
• TensorFlow version: 1.13.1
• Bazel version: 0.24.1
• GCC/Compiler version: NDK r18, clang

Describe the problem
Using the 'official' implementation 'org.tensorflow:tensorflow-lite:0.0.0-nightly', I tried

opt.setAllowBufferHandleOutput(true)

I get

E/aoyi.run.tflit: No implementation found for void org.tensorflow.lite.NativeInterpreterWrapper.allowBufferHandleOutput(long, boolean) (tried Java_org_tensorflow_lite_NativeInterpreterWrapper_allowBufferHandleOutput and Java_org_tensorflow_lite_NativeInterpreterWrapper_allowBufferHandleOutput__JZ)

Same error reproduced when I build the libtensorflowlite_jni.so locally with bazel.

Provide the exact sequence of commands / steps that you executed before running into the problem

To verify that this was a build issue, I run

> nm -D jniLibs/arm64-v8a/libtensorflowlite_jni.so | grep allow
0000000000010940 T Java_org_tensorflow_lite_NativeInterpreterWrapper_allowFp16PrecisionForFp32

Analysis

The root cause is that the build relies on nativeinterpreterwrapper_jni.h to declare the JNI functions as extern "C". Java_org_tensorflow_lite_NativeInterpreterWrapper_allowBufferHandleOutput was introduced in 18 Dec 2018 commit which did not update the .h file.

Proposed fix

The easy fix (tested here) is to tag the exported JNI functions as extern "C" in the .cc file. Relying on the header file is not necessary and (as we witness) error-prone.