tensorflow 1.11.0 gpu windows10下编译 python37 cuda10

系统：

win10 64bit

vs2017

硬件：

cpu：i7-6700hq

显卡：gtx970m

软件：

python3.7(64bit) https://www.python.org/downloads/release/python-370/

注意安装debugging symbols 和debug binraries

cuda10.10 cuda_10.0.130_411.31_win10.exe https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exelocal

cudnn7.3 cudnn-10.0-windows10-x64-v7.3.1.20.zip https://developer.nvidia.com/cudnn

tensorflow1.11 https://github.com/tensorflow/tensorflow.git c19e29306ce1777456b2dbb3a14f511edf7883a8

cmake   cmake-3.12.3-win64-x64.msi https://cmake.org/download/

swig swigwin-3.0.12 http://www.swig.org/download.html

numpy numpy-1.15.2-cp37-none-win_amd64.whl

环境变量：

PreferredToolArchitecture=x64     https://msdn.microsoft.com/zh-cn/library/ee662426.aspx

PreferredToolArchitecture 属性确定在生产过程中，是否使用了 32 位或 64 位编译器和工具。 此属性不影响输出平台体系结构或配置。 默认情况下，如果该属性未设置或设置的值不为 x64，MSBuild 就会使用编译器和工具的 x86 版本。

将 PreferredToolArchitecture 的属性设置为 x64 以使用 64 位编译器和工具生成应用程序。

msbuild myProject.vcxproj /p:PreferredToolArchitecture=x64

务必保证编译期间 Microsoft® C/C++ Compiler Driver 是64位，否则可能会爆出 fatal error C1002: 在第2遍中编译器的堆空间不足

Cmake配置

提前fix一下 tensorflow\tensorflow\contrib\cmake\CMakeLists.txt

https://stackoverflow.com/questions/42603407/how-to-compile-tensor-flow-with-sse-and-and-avx-instructions-on-windows

Tensorflow make a mistake on flag "tensorflow_WIN_CPU_SIMD_OPTIONS".

It is a Flag, not a Boolean.

"Tensorflow-github/tensorflow/contrib/cmake/CMakeLists.txt" 

option(tensorflow_WIN_CPU_SIMD_OPTIONS "Enables CPU SIMD instructions")

Replace it wtih

set(tensorflow_WIN_CPU_SIMD_OPTIONS "/arch:AVX" CACHE STRING "Enables CPU SIMD instructions" )

这里我们使用 /arch:AVX2 ,根据机器cpu选择

由于

• Starting from TensorFlow 1.11, Windows builds will use Bazel. Therefore, we will drop official support for cmake.

https://github.com/tensorflow/tensorflow/blob/master/RELEASE.md#breaking-changes

使用这里  https://github.com/tensorflow/tensorflow/issues/22047

FWIW, I've included a fix for CMake and Windows in this patch here:
https://github.com/bytedeco/javacpp-presets/blob/beb0a4a/tensorflow/tensorflow-windows.patch

对cmake相关文件进行修改

根据 https://medium.com/@amsokol.com/update-1-how-to-build-and-install-tensorflow-gpu-cpu-for-windows-from-source-code-using-bazel-and-c2e86fec9ef2

给eigen手动打上patch https://github.com/amsokol/tensorflow-windows-build-tutorial/blob/master/eigen_half.patch

cmake 应该是不支持bazel这种patch patch_file = clean_dep("//third_party:eigen_half.patch")

cmake GUI :

source code : tensorflow/tensorflow/contrib/cmake

configure 后

勾选 tensorflow_ENABLE_GPU

tensorflow_CUDA_VERSION 10.0

tensorflow_CUDNN_VERSION 7.3

新增 CUDNN_HOME 指向cudnn解压后目录

tensorflow_WIN_CPU_SIMD_OPTIONS /arch:AVX2

修改完后重新Configure,直到Configuring done

然后Generating done

 

开始build 参照 https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/cmake

 

命令行执行 set PreferredToolArchitecture=x64     设置环境变量（重要）

载入 VS 2017的开发人员命令提示符 

执行 MSBuild /p:Configuration=Release /p:PreferredToolArchitecture=x64 tf_python_build_pip_package.vcxproj 

可选参数 /p:CL_MPCount=4  同时运行的cl.exe数量，减少可有效降低内存占用，同时会降低编译速度

转载于:https://my.oschina.net/thinkwithlinc/blog/2223090