SUMMARY
I want to add tensor flow to a c++ app in Visual Studio, to read and write TFRecords and ideally execute models trained in Python. I’m, not interested in GPU (I know there is no support for it). I only need to be able to deploy a c++ app in windows with some tensorflow usability.
I need, a .dll, and the corresponding .lib and .h files to link in my software.
First I figured out, I need to compile tensorflow_cpp to do this. But for the life of me, I am unable to do this. This is what I tried, and what I obtained:
VERSIONS TESTED:
tensorflow = 2.16.1 , 2.15.1, 2.15.0
bazel = 7.1.1 , 6.5.0 , 6.4.0 , 6.1.0
python = 3.12.3 , 3.11.9
msys2 = 20240113
compiler = MSVC 2019 (Visual Studio 2022 version 17.9.0, MSVC toolset version 14.39.33519)
STEPS:
- Install Python and add to path
pip3 install -U pip
pip3 install -U six numpy wheel packaging
pip3 install -U keras_applications --no-deps
pip3 install -U keras_preprocessing --no-deps
- Install Bazel and add to path
- Install MSYS2 and add to path
- Download and uncompress Tensorflow
- Some Configuration
pacman -Syu
pacman -S git patch unzip
pacman -S git patch unzip rsync
python .\configure.py
- No ROCm support
- Default optimization flags (/arch:AVX)
- Not override eigen strong inline (tried both)
- Not configure Android builds
Preconfigured Bazel build configs. You can use any of the below by adding “–config=<>” to your build command. See .bazelrc for more details.
–config=mkl # Build with MKL support.
–config=mkl_aarch64 # Build with oneDNN and Compute Library for the Arm Architecture (ACL).
–config=monolithic # Config for mostly static monolithic build.
–config=numa # Build with NUMA support.
–config=dynamic_kernels # (Experimental) Build kernels into separate shared objects.
–config=v1 # Build with TensorFlow 1 API instead of TF 2 API.
Preconfigured Bazel build configs to DISABLE default on features:
–config=nogcp # Disable GCP support.
–config=nonccl # Disable NVIDIA NCCL support.
RESULTS: ABLE TO BUILD
bazel build -c opt --define=no_tensorflow_py_deps=true tensorflow/lite:tensorflowlite
bazel build -c opt --define=no_tensorflow_py_deps=true tensorflow.dll
RESULTS: UNABLE TO BUILD
bazel build -c opt --define=no_tensorflow_py_deps=true tensorflow:tensorflow_cc.dll
bazel build -c opt --define=no_tensorflow_py_deps=true tensorflow:libtensorflow_cc.so
ERRORS
Im just adding the meaningful part in the wall of text.
ERROR: C:/tensorflow-2.16.1/tensorflow/compiler/mlir/quantization/tensorflow/calibrator/BUILD:97:11: Compiling tensorflow/compiler/mlir/quantization/tensorflow/calibrator/calibration_statistics_collector_average_min_max.cc failed: (Exit 2): cl.exe failed: error executing command (from target //tensorflow/compiler/mlir/quantization/tensorflow/calibrator:calibration_statistics_collector_average_min_max)
[…]
cl : command line error D8022 : cannot open ‘bazel-out/x64_windows-opt/bin/tensorflow/compiler/mlir/quantization/tensorflow/calibrator/_objs/calibration_statistics_collector_average_min_max/calibration_statistics_collector_average_min_max.obj.params’
Target //tensorflow:tensorflow_cc.dll failed to build
Some notes:
- I turned Developer mode on (to allow bazel to create symlinks).
- Some bazel versions did not work with MSVC 2019, but otherwise I obtain the same results despite the versions of python, bazel and tensorflow.
- This happens around the 15000/18000 tasks (So it takes around an hour to get here).
- The file
C:/tensorflow-2.16.1/tensorflow/compiler/mlir/quantization/tensorflow/calibrator/calibration_statistics_collector_average_min_max.ccexists and can be opened. - I tried to use clang (LLVM-17.0.6-win64) to compile the library with it instead of MSVC, since it appears to be the recommended option with tf-2.16.1, I was unable to do it (there is NO documentation online on how to configure LLVM to do this, and I wasted hours without being able to tell bazel to use properly clang (even after adding the bin to the path, and using set
BAZEL_LLVMto the proper path, bazel recognized the installation but kept telling me did not find the files). Also Bazel documentation seems to be for the version 7, which is unsupported. - There is no prebuilt distribution of any recent version of
tensorflow_cclibtensorflow_ccas far as I know.
QUESTIONS
- I unsure of the difference between
tensorflow.dllcompiled in c++ (which I am able to build) andtensorflow_cc.dllandlibtensorflow_cc.so(which im not able to build). As I understand thetensorflow_ccis the c++ API detailed in TensorFlow C++ API Reference | TensorFlow v2.16.1 , but can I usetensorflow.dll? if so, how? what is the difference? - Can somebody explain me how to configure and use LLVM in bazel for testing in my case?
- I’m unsure on how to get the proper header files, I know I can build the target
tensorflow :install_headers, can I get from this target the headers of tensorflowlite or libtensorflow_cc? - I have seen that the header files must be edited manually after compile the .dll, adding
TF_EXPORTis some undocumented way manually and then recompile the library. Does this affect only the :tensorflow target? Or also tensorflow_cc? - Does anyone know how to get any somewhat recent precompiled cpu distribution of
tensor_flow_cc - AND MOST IMPORTANTLY: Any Idea on how to solve this issue?