Idea: Temporary Workaround for Native C++ 16KB Memory Page Compatibility

[yong1020]

Hi everyone!

I’d like to share an idea that might help others who are running into issues with the recent 16KB memory page policy.

I experimented with rebuilding libtensorflowlite.so for arm64-v8a devices so that it aligns with a 16KB memory page size.

If you’d like to try it out, you can replace the following files in your project with the rebuilt versions:

• cpp/tf-lite-api/generated-libs/arm64-v8a/libtensorflowlite.so libtensorflowlite.zip
• cpp/tf-lite-api/tensorflow_src/tensorflow/ tensorflow.zip

---

🔧 Build Details

• TensorFlow version: 2.12.0
  I couldn’t determine the exact TensorFlow version used in the original project, so I explored compatibility with TFLiteEngine and found that 2.12.0 worked.
• OS: Ubuntu 20.04
• Build tool: Bazel 5.3.0 (compatible with TF 2.12.0)
• NDK version: 21.4.7075529
• Build command:

  $ bazel build -c opt \
    --config=android_arm64 \
    --linkopt='-Wl,-z,max-page-size=0x4000' \
    //tensorflow/lite:libtensorflowlite.so
  

• Build Result: If you see 0x4000 in the alignment field, it means that the build successfully applied the 16KB memory page size:

  $ readelf -lW bazel-bin/tensorflow/lite/libtensorflowlite.so | grep LOAD
    LOAD           0x000000 0x0000000000000000 0x0000000000000000 0x462ffc 0x462ffc R E 0x4000
    LOAD           0x4632c8 0x00000000004672c8 0x00000000004672c8 0x00eb28 0x0160a8 RW  0x4000
  

[vilassn]

@yong1020 Thanks for your efforts

[KiwiWilkinson]

Thanks Yong, that looks like a great step forward.

However, when I compile it into whisper_native I get:

Has something changed in TensorFlow than requires changes in whisper_native?

[yong1020]

@KiwiWilkinson

Did you replace both the attached in the 'libtensorflow.so' and the 'tensorflow/lite' source specified path?

• cpp/tf-lite-api/generated-libs/arm64-v8a/libtensorflowlite.so -> libtensorflowlite.zip
• cpp/tf-lite-api/tensorflow_src/tensorflow/ -> tensorflow.zip

It should work without any other changes.

[KiwiWilkinson]

Thanks Yong. You were quite right. I'd copied the source code to the wrong place in the directory hierarchy!

It's now working properly. Thank you very much.

[KiwiWilkinson]

Do I also need a libtensorflowlite.so for other ABIs: armabi-v7a and x86_64? I'm getting a linking error when I try to build an apk.

[KiwiWilkinson]

@yong1020 , thanks to your bazel build instruction above, I've created the other two myself. I'm not familiar with Ubuntu or Bazel, but I created a new Virtual Machine and installed the necessary bits.

This is the file for armeabi-v7a:

libtensorflowlite.zip

and this is the file for x86_64:

libtensorflowlite.zip

@vilassn do you want to incorporate these in a new version of Whisper?