TensorFlow is a free and open-source AI/ML framework from Google used for training and inferencing of neural networks. Tensorflow Lite is designed as a smaller subset of Tensorflow aimed at deployment on mobile and edge devices.

Requirements

• Clonecpuinfo from GitHub - pytorch/cpuinfo: CPU INFOrmation library (x86/x86-64/ARM/ARM64, Linux/Windows/Android/macOS/iOS)

• Clone pthreadpool from GitHub - Maratyszcza/pthreadpool: Portable (POSIX/Windows/Emscripten) thread pool for C/C++

• Clone XNNPACK from GitHub - everton1984/XNNPACK: High-efficiency floating-point neural network inference operators for mobile, server, and Web and use branch woa_enablement

• Clone tensorflow from GitHub - everton1984/tensorflow: An Open Source Machine Learning Framework for Everyone and use branch woa_enablement2

• Download LLVM for WoA at least 18.1.0

• Download bazel for WoA version 6.5.0

• Make sure to take a look into XNNPACK’s and tensorflow’s workspace and update the addresses for cpuinfo, pthreadpool and XNNPACK to your respective local clones

Steps

• Make sure to set the environment variables BAZEL_LLVM, BAZEL_VC and TF_PYTHON_VERSION to the respective proper directories on your system as well as the Python version you are using.

• Go to the XNNPACK folder and open the WORKSPACE file. Fix the directory of the local repositories to match to your own clones of pthreadpool and cpuinfo. Do the same for Tensorflow under tensorflow\workspace2.bzl

• Now run

  bazel.exe build //tensorflow/lite:tensorflowlite --cpu=arm64_windows --compiler=clang-cl --copt="/clang:-march=armv8-a+dotprod+fp16+i8mm" --cxxopt="/clang:-march=armv8-a+dotprod+fp16+i8mm"

• You will find tensorflow.dll inside the bazel-bin\tensorflow directory