[Question(s)] Can you build a C++ standalone shared-lib (no python dep)?

[BraynStorm]

Title.
Also, I see that you build with SSE too, does this mean you can build TF as a 32bit binary?

[fo40225]

I have not tried to build for x86, maybe I can try to build.

[fo40225]

Build fail, it seems to require some code fix.

"C:\Users\User\Source\Repos\tensorflow\tensorflow\contrib\cmake\build\tf_python_build_pip_package.vcxproj" (預設目標) (1) ->
"C:\Users\User\Source\Repos\tensorflow\tensorflow\contrib\cmake\build\pywrap_tensorflow_internal.vcxproj" (預設目標) (2) ->
"C:\Users\User\Source\Repos\tensorflow\tensorflow\contrib\cmake\build\pywrap_tensorflow_internal_static.vcxproj" (預設目標) (3) ->
"C:\Users\User\Source\Repos\tensorflow\tensorflow\contrib\cmake\build\tf_core_cpu.vcxproj" (預設目標) (126) ->
(ClCompile 目標) -> 
  C:\Users\User\Source\Repos\tensorflow\tensorflow/core/common_runtime/bfc_allocator.h(383): error C3861: '_BitScanReverse64': 找不到識別項 (正在編譯原始程式檔 C:\Users\User\Source\Repos\tensorflow\tensorflow\core\common_runtime\bfc_allocator.cc) [C:\Users\User\Source\Repos\tensorflow\tensorflow\contrib\cmake\build\tf_core_cpu.vcxproj]

[fo40225]

Because cc_test is broken, it cannot perform a full test.

I haven't used tensorflow's C++ API before, this built worked on the sample with target 32-bit exe.

Before using it, you should do a complete test yourself.

https://github.com/fo40225/tensorflow-windows-wheel/tree/master/1.7.0/cpp

---

Here is the way to build 32bit cpp lib.

env

win 10 x64
git 2.14.1 (in PATH)
cmake 3.9.6 (in PATH)
python (in PATH)
visual studio 2017

get source

cd %HOMEPATH%
git clone https://github.com/tensorflow/tensorflow.git -b v1.7.0
cd tensorflow/tensorflow/contrib/cmake
mkdir build
cd build

apply fix

add add_definitions(-DEIGEN_DEFAULT_DENSE_INDEX_TYPE=std::int64_t) into tensorflow\tensorflow\contrib\cmake\CMakeLists.txt

---

edit tensorflow\tensorflow\contrib\cmake\tools\create_def_file.py line 128
from
def_fp.write("\t ??1OpDef@tensorflow@@UEAA@XZ\n")
to
def_fp.write("\t ??1OpDef@tensorflow@@UAE@XZ\n")

---

edit tensorflow\tensorflow\core\common_runtime\bfc_allocator.h line 381
from

  inline int Log2FloorNonZero(uint64 n) {
#if defined(__GNUC__)
    return 63 ^ __builtin_clzll(n);
#elif defined(PLATFORM_WINDOWS)
    unsigned long index;
    _BitScanReverse64(&index, n);
    return index;
#else
    return Log2FloorNonZeroSlow(n);
#endif
  }

to

  inline int Log2FloorNonZero(uint64 n) {
#if defined(__GNUC__)
    return 63 ^ __builtin_clzll(n);
#elif defined(PLATFORM_WINDOWS) && defined(_WIN64)
    unsigned long index;
    _BitScanReverse64(&index, n);
    return index;
#else
    return Log2FloorNonZeroSlow(n);
#endif
  }

build

open VS2017 x64_x86 Cross Tools Command Prompt

cd %HOMEPATH%\tensorflow\tensorflow\contrib\cmake\build

cmake .. -G "Visual Studio 15 2017" -T host=x64 ^
-DCMAKE_BUILD_TYPE=Release ^
-Dtensorflow_BUILD_PYTHON_BINDINGS=OFF ^
-Dtensorflow_BUILD_SHARED_LIB=ON ^
-Dtensorflow_WIN_CPU_SIMD_OPTIONS="/arch:IA32"

cmake --build . --target tensorflow --config Release -- /fileLogger

You will get tensorflow libs in tensorflow\tensorflow\contrib\cmake\build\Release.

[BraynStorm]

Thanks for uploading.
Although, for these to be useful, you need to include the whole content of the "CMAKE_INSTALL_PREFIX" directory (usually C:\Program Files\ or something) as it contains the generated and properly placed header files.

[fo40225]

Updated, containing the properly placed header make it looks much better now.

[BraynStorm]

Perfect! Thank you very much for the quick response and going through the trouble of building it.

[miek0tube]

Applied all the fixed but still get the following error:

cmake .. -G "Visual Studio 15 2017" -T host=x64 -DCMAKE_BUILD_TYPE=Release -Dtensorflow_BUILD_PYTHON_BINDINGS=OFF -Dtensorflow_BUILD_SHARED_LIB=ON -Dtensorflow_WIN_CPU_SIMD_OPTIONS="/arch:IA32"
.............
CMake Error at tf_tools.cmake:52 (list):
list sub-command REMOVE_ITEM requires two or more arguments.
Call Stack (most recent call first):
CMakeLists.txt:463 (include)

Interestingly, running the build on TF 1.6.0.rc0 does not generate the error (but compilation fails anyway).
Could you please take a look at this?

[fo40225]

@miek0tube Try not turn tensorflow_BUILD_PYTHON_BINDINGS OFF, it has some reported issues.

[miek0tube]

Yeah, thanks! That worked!

[miek0tube]

To my disappointment, running in 32-bit mode is 6-8 times slower! Did you experience the same performance hit?

[sleeplessai]

@miek0tube
Out of the topic, I am curious about the scene your TF will apply to.
Could you share more?

[fo40225]

I don't know if this changes affect perf a lot. https://github.com/tensorflow/tensorflow/blob/v1.9.0/tensorflow/core/common_runtime/bfc_allocator.h#L382

You can help to figure out the root cause.
#20

[miek0tube]

I'm porting an image recognition software from Python to C++. It runs as a plugin to another software which is 32-bit C# application.
About the bfc allocator patch -- I don't thin this could cause such a dramatic slowdown. That piece of code is fast enough with or without the patch.

[miek0tube]

Looks like I was partially blind this day. This key -
-Dtensorflow_WIN_CPU_SIMD_OPTIONS="/arch:IA32"
wipes the enhanced CPU instructions support. Changed it to "/arch:SSE2" and everything is fine now.