基本参照我的这篇文章：《Windows下编译带CUDA 11.2的TensorFlow 2.4.1（Python3.9.1，cuDNN 8.1.0，兼容性3.5 - 8.6，附编译结果下载）》，有些地方有所改动。
环境准备
1. 内存要求
在8个并行任务下（默认并行数为CPU线程数），应有不小于10G的内存，否则会产生编译器堆空间不足的错误。
2. Python & Pip
首先Python需要安装一些包：six、numpy、wheel、setuptools、keras_applications和keras_preprocessing，使用管理员权限打开命令提示符：
pip install six numpy wheel setuptools
pip install Keras_applications Keras_preprocessing --no-deps
注意，Python路径中不能出现空格，即Windows下默认安装路径C:\Program Files\Python39会在编译时报错，因此如果装到了这个路径，需要在一个没有空格的目录下创建一个链接（不是快捷方式），用mklink命令。
3. CUDA
这里选的CUDA 11.3，CUDA官网下载安装，没什么好说的。
4. Bazel
然后是Bazel，bazel很简单，就一个exe，需要设置环境变量给到Path下，我偷懒直接放到CUDA的bin目录下。我选的版本是3.7.2。
5. MSYS2
再安装MSYS2，同样需要给msys64\usr\bin目录设置环境变量。
装好后再安装一些包，用的是pacman，由于默认源极慢极慢，所以建议国内换源。
进到msys64\etc\pacman.d目录下，修改所有mirrolist，分别在各自所有Server行前加一行，把下面清华/中科大/北邮的随便选一个复制上来就行。
打开msys64命令行，官方教程少提了一个zip包，因此安装命令如下：
pacman -S git patch unzip zip
6. Visual Studio 2019
然后是VS，下载VS安装器，为避免麻烦，装到C盘默认路径（这次我没有尝试非C盘路径，不知道找不到编译器的bug还在不在）。如果非VS用户，只需安装除必选组件外的MSVC v142 - VS 2019 C++ x64/x86生成工具（随便一个，我选的最新版本）和Windows 10 SDK（同样随便，我选的最新的）。
编译
配置编译
下载TensorFlow 2.5.0源码，进入解压后的根目录，执行
D:\tensorflow-2.5.0>python configure.py
You have bazel 3.7.2 installed.
Please specify the location of python. [Default is C:\Python39\python.exe]:
Found possible Python library paths:
C:\Python39\lib\site-packages
Please input the desired Python library path to use. Default is [C:\Python39\lib\site-packages]
Do you wish to build TensorFlow with ROCm support? [y/N]:
No ROCm support will be enabled for TensorFlow.
Do you wish to build TensorFlow with CUDA support? [y/N]: y
CUDA support will be enabled for TensorFlow.
Found CUDA 11.3 in:
D:/CUDA/lib/x64
D:/CUDA/include
Found cuDNN 8 in:
D:/CUDA/lib/x64
D:/CUDA/include
Please specify a list of comma-separated CUDA compute capabilities you want to build with.
You can find the compute capability of your device at: https://developer.nvidia.com/cuda-gpus. Each capability can be specified as "x.y" or "compute_xy" to include both virtual and binary GPU code, or as "sm_xy" to only include the binary code.
Please note that each additional compute capability significantly increases your build time and binary size, and that TensorFlow only supports compute capabilities >= 3.5 [Default is: 3.5,7.0]: 3.5,3.7,5.0,5.2,6.0,6.1,7.0,7.5,8.0,8.6
Please specify optimization flags to use during compilation when bazel option "--config=opt" is specified [Default is /arch:AVX]: /arch:AVX2
Would you like to override eigen strong inline for some C++ compilation to reduce the compilation time? [Y/n]:
Eigen strong inline overridden.
Would you like to interactively configure ./WORKSPACE for Android builds? [y/N]:
Not configuring the WORKSPACE for 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=v2 # Build TensorFlow 2.x instead of 1.x.
Preconfigured Bazel build configs to DISABLE default on features:
--config=noaws # Disable AWS S3 filesystem support.
--config=nogcp # Disable GCP support.
--config=nohdfs # Disable HDFS support.
--config=nonccl # Disable NVIDIA NCCL support.
这版TensorFlow编译SM 3.5会报错，查了一下貌似TensorRT不支持这么低的版本。
代码修改
启用MKL时
从2.4开始TensorFlow将MKL用到的OpenMP从直接下载二进制可执行文件变为了从LLVM项目下载开源代码并编译，带来了一系列问题，这里建议官网下载安装程序（Pre-Built Binaries）安装LLVM（记住安装中写入系统环境变量），作用有二：
- 用现成的libiomp5md.lib和libiomp5md.dll取代编译过程，MSVC编译LLVM的OpenMP会出错。
- 将LLVM用到的DLL（在LLVM安装目录\bin下）复制到msys安装目录\usr\bin下，否则在后续编译步骤中使用msys里的bash时没有复制系统环境变量的Path会报错（比如编译Lite相关内容时会报找不到api-ms-win-crt-locale-l1-1-0.dll的错）。
将libiomp5md.lib放到third_party\mkl目录下，并在third_party\mkl\mkl.BUILD的第75行后插入：
cc_import(
name = "iomp5",
interface_library = "libiomp5md.lib",
system_provided = 1,
)
然后将下方Windows编译配置修改为：
cc_library(
name = "mkl_libs_windows",
deps = [
"iomp5"
],
visibility = ["//visibility:public"],
)
将libiomp5md.dll放到系统环境变量里。
将third_party\llvm_openmp\BUILD第74行修改为0，取消强制使用MSVC：
omp_vars_win = {
"MSVC": 0,
}
没有这一步会报类似于如下错误（这段报错复制自2.4.1版，但2.5.0也会报同样的错误）：
ERROR: D:/output_base/external/llvm_openmp/BUILD.bazel:176:10: C++ compilation of rule '@llvm_openmp//:libiomp5md.dll' failed (Exit 1): ml64.exe failed: error executing command
cd D:/output_base/execroot/org_tensorflow
SET CUDA_TOOLKIT_PATH=D:/CUDA
SET INCLUDE=C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\ATLMFC\include;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\include;C:\Program Files (x86)\Windows Kits\10\include\10.0.18362.0\ucrt;C:\Program Files (x86)\Windows Kits\10\include\10.0.18362.0\shared;C:\Program Files (x86)\Windows Kits\10\include\10.0.18362.0\um;C:\Program Files (x86)\Windows Kits\10\include\10.0.18362.0\winrt;C:\Program Files (x86)\Windows Kits\10\include\10.0.18362.0\cppwinrt
SET LIB=C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\ATLMFC\lib\x64;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\lib\x64;C:\Program Files (x86)\Windows Kits\10\lib\10.0.18362.0\ucrt\x64;C:\Program Files (x86)\Windows Kits\10\lib\10.0.18362.0\um\x64
SET PATH=C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\\Extensions\Microsoft\IntelliCode\CLI;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\bin\HostX64\x64;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\VC\VCPackages;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\TestWindow;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\TeamFoundation\Team Explorer;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\MSBuild\Current\bin\Roslyn;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Team Tools\Performance Tools\x64;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Team Tools\Performance Tools;C:\Program Files (x86)\Microsoft Visual Studio\Shared\Common\VSPerfCollectionTools\vs2019\\x64;C:\Program Files (x86)\Microsoft Visual Studio\Shared\Common\VSPerfCollectionTools\vs2019\;C:\Program Files (x86)\Windows Kits\10\bin\10.0.18362.0\x64;C:\Program Files (x86)\Windows Kits\10\bin\x64;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\\MSBuild\Current\Bin;C:\Windows\Microsoft.NET\Framework64\v4.0.30319;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\Tools\;;C:\Windows\system32;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\CommonExtensions\Microsoft\CMake\Ninja;C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\Common7\IDE\VC\Linux\bin\ConnectionManagerExe
SET PWD=/proc/self/cwd
SET PYTHON_BIN_PATH=C:/Python39/python.exe
SET PYTHON_LIB_PATH=C:/Python39/lib/site-packages
SET RUNFILES_MANIFEST_ONLY=1
SET TEMP=D:/tmp
SET TF2_BEHAVIOR=1
SET TF_CONFIGURE_IOS=0
SET TF_CUDA_COMPUTE_CAPABILITIES=3.5,3.7,5.0,5.2,6.0,6.1,7.0,7.5,8.0,8.6
SET TF_NEED_CUDA=1
SET TMP=D:/tmp
C:/Program Files (x86)/Microsoft Visual Studio/2019/Community/VC/Tools/MSVC/14.27.29110/bin/HostX64/x64/ml64.exe -B external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py /nologo /DCOMPILER_MSVC /DNOMINMAX /D_WIN32_WINNT=0x0600 /D_CRT_SECURE_NO_DEPRECATE /D_CRT_SECURE_NO_WARNINGS /D_SILENCE_STDEXT_HASH_DEPRECATION_WARNINGS /bigobj /Zm500 /J /Gy /GF /EHsc /wd4351 /wd4291 /wd4250 /wd4996 /Iexternal/llvm_openmp /Ibazel-out/x64_windows-opt/bin/external/llvm_openmp /Iexternal/bazel_tools /Ibazel-out/x64_windows-opt/bin/external/bazel_tools /Iexternal/llvm_openmp/runtime/src /Ibazel-out/x64_windows-opt/bin/external/llvm_openmp/runtime/src /Iexternal/llvm_openmp/include /Ibazel-out/x64_windows-opt/bin/external/llvm_openmp/include /showIncludes /MD /O2 /DNDEBUG /W0 /D_USE_MATH_DEFINES -DWIN32_LEAN_AND_MEAN -DNOGDI /experimental:preprocessor -DTHRUST_IGNORE_CUB_VERSION_CHECK /Domp_EXPORTS /D_M_AMD64 /DOMPT_SUPPORT=0 /D_WINDOWS /D_WINNT /D_USRDLL /Fobazel-out/x64_windows-opt/bin/external/llvm_openmp/_objs/libiomp5md.dll/z_Windows_NT-586_asm.obj /c bazel-out/x64_windows-opt/bin/external/llvm_openmp/z_Windows_NT-586_asm.S
Execution platform: @local_execution_config_platform//:platform
Microsoft (R) Macro Assembler (x64) Version 14.27.29111.0
Copyright (C) Microsoft Corporation. All rights reserved.
MASM : warning A4018:invalid command-line option : -B
Assembling: external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(1) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(2) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(3) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(4) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(5) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(6) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(7) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(8) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(9) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(10) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(11) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(12) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(13) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(14) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(15) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(16) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(17) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(18) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(19) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(20) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(21) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(22) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(23) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(24) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(25) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(26) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(27) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(28) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(29) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(30) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(31) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(32) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(33) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(34) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(35) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(36) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(37) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(38) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(39) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(40) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(41) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(42) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(43) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(44) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(45) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(46) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(47) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(48) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(49) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(50) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(51) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(52) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(53) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(54) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(55) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(56) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(57) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(58) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(59) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(60) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(61) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(62) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(63) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(64) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(65) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(66) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(67) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(68) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(69) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(70) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(71) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(72) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(73) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(74) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(75) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(76) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(77) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(78) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(79) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(80) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(81) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(82) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(83) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(84) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(85) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(86) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(87) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(88) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(89) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(90) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(91) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(92) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(93) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(94) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(95) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(96) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(97) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(98) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(99) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(100) : error A2044:invalid character in file
external/local_config_cuda/crosstool/windows/msvc_wrapper_for_nvcc.py(101) : fatal error A1012:error count exceeds 100; stopping assembly
Target //tensorflow/tools/pip_package:build_pip_package failed to build
INFO: Elapsed time: 229.166s, Critical Path: 30.31s
INFO: 966 processes: 239 internal, 727 local.
FAILED: Build did NOT complete successfully
此外，最后链接时找不到DLL，需要将LLVM的libiomp5md.dll放到python.exe同目录下，否则会出现以下错误：
ERROR: D:/tensorflow-2.4.1/tensorflow/python/keras/api/BUILD:111:19: Executing genrule //tensorflow/python/keras/api:keras_python_api_gen failed (Exit 1): bash.exe failed: error executing command
cd D:/output_base/execroot/org_tensorflow
SET CUDA_TOOLKIT_PATH=D:/CUDA
SET PATH=C:\msys64\usr\bin;C:\msys64\bin;C:\Windows;C:\Windows\System32;C:\Windows\System32\WindowsPowerShell\v1.0;C:\NASM;D:\;D:\CUDA\bin;D:\CUDA\libnvvp;C:\Python39\Scripts\;C:\Python39\;C:\Windows\system32;C:\Windows;C:\Windows\System32\Wbem;C:\Windows\System32\WindowsPowerShell\v1.0\;C:\Windows\System32\OpenSSH\;C:\Program Files\dotnet\;C:\Program Files\Microsoft SQL Server\130\Tools\Binn\;C:\Program Files\NVIDIA Corporation\Nsight Compute 2020.1.2\;C:\Program Files\Microsoft VS Code\bin;C:\Program Files (x86)\NVIDIA Corporation\PhysX\Common;C:\Program Files\NVIDIA Corporation\Nsight Compute 2020.2.0\;C:\Program Files\CMake\bin;C:\LLVM\bin;C:\Git\cmd;C:\msys64\usr\bin;C:\Users\用户名\AppData\Local\Microsoft\WindowsApps;C:\Users\用户名\.dotnet\tools
SET PYTHON_BIN_PATH=C:/Python39/python.exe
SET PYTHON_LIB_PATH=C:/Python39/lib/site-packages
SET RUNFILES_MANIFEST_ONLY=1
SET TF2_BEHAVIOR=1
SET TF_CONFIGURE_IOS=0
SET TF_CUDA_COMPUTE_CAPABILITIES=3.5,3.7,5.0,5.2,6.0,6.1,7.0,7.5,8.0,8.6
SET TF_NEED_CUDA=1
C:/msys64/usr/bin/bash.exe bazel-out/x64_windows-opt/bin/tensorflow/python/keras/api/keras_python_api_gen.genrule_script.sh
Execution platform: @local_execution_config_platform//:platform
Traceback (most recent call last):
File "\\?\D:\tmp\Bazel.runfiles_q5x32q4w\runfiles\org_tensorflow\tensorflow\python\pywrap_tensorflow.py", line 64, in <module>
from tensorflow.python._pywrap_tensorflow_internal import *
ImportError: DLL load failed while importing _pywrap_tensorflow_internal: 找不到指定的模块。
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "\\?\D:\tmp\Bazel.runfiles_q5x32q4w\runfiles\org_tensorflow\tensorflow\python\tools\api\generator\create_python_api.py", line 26, in <module>
from tensorflow.python.tools.api.generator import doc_srcs
File "\\?\D:\tmp\Bazel.runfiles_q5x32q4w\runfiles\org_tensorflow\tensorflow\python\__init__.py", line 39, in <module>
from tensorflow.python import pywrap_tensorflow as _pywrap_tensorflow
File "\\?\D:\tmp\Bazel.runfiles_q5x32q4w\runfiles\org_tensorflow\tensorflow\python\pywrap_tensorflow.py", line 83, in <module>
raise ImportError(msg)
ImportError: Traceback (most recent call last):
File "\\?\D:\tmp\Bazel.runfiles_q5x32q4w\runfiles\org_tensorflow\tensorflow\python\pywrap_tensorflow.py", line 64, in <module>
from tensorflow.python._pywrap_tensorflow_internal import *
ImportError: DLL load failed while importing _pywrap_tensorflow_internal: 找不到指定的模块。
Failed to load the native TensorFlow runtime.
See https://www.tensorflow.org/install/errors
for some common reasons and solutions. Include the entire stack trace
above this error message when asking for help.
Target //tensorflow/tools/pip_package:build_pip_package failed to build
ERROR: D:/tensorflow-2.4.1/tensorflow/tools/pip_package/BUILD:165:10 Executing genrule //tensorflow/python/keras/api:keras_python_api_gen_compat_v2 failed (Exit 1): bash.exe failed: error executing command
cd D:/output_base/execroot/org_tensorflow
SET CUDA_TOOLKIT_PATH=D:/CUDA
SET PATH=C:\msys64\usr\bin;C:\msys64\bin;C:\Windows;C:\Windows\System32;C:\Windows\System32\WindowsPowerShell\v1.0;C:\NASM;D:\;D:\CUDA\bin;D:\CUDA\libnvvp;C:\Python39\Scripts\;C:\Python39\;C:\Windows\system32;C:\Windows;C:\Windows\System32\Wbem;C:\Windows\System32\WindowsPowerShell\v1.0\;C:\Windows\System32\OpenSSH\;C:\Program Files\dotnet\;C:\Program Files\Microsoft SQL Server\130\Tools\Binn\;C:\Program Files\NVIDIA Corporation\Nsight Compute 2020.1.2\;C:\Program Files\Microsoft VS Code\bin;C:\Program Files (x86)\NVIDIA Corporation\PhysX\Common;C:\Program Files\NVIDIA Corporation\Nsight Compute 2020.2.0\;C:\Program Files\CMake\bin;C:\LLVM\bin;C:\Git\cmd;C:\msys64\usr\bin;C:\Users\用户名\AppData\Local\Microsoft\WindowsApps;C:\Users\用户名\.dotnet\tools
SET PYTHON_BIN_PATH=C:/Python39/python.exe
SET PYTHON_LIB_PATH=C:/Python39/lib/site-packages
SET RUNFILES_MANIFEST_ONLY=1
SET TF2_BEHAVIOR=1
SET TF_CONFIGURE_IOS=0
SET TF_CUDA_COMPUTE_CAPABILITIES=3.5,3.7,5.0,5.2,6.0,6.1,7.0,7.5,8.0,8.6
SET TF_NEED_CUDA=1
C:/msys64/usr/bin/bash.exe bazel-out/x64_windows-opt/bin/tensorflow/python/keras/api/keras_python_api_gen_compat_v2.genrule_script.sh
Execution platform: @local_execution_config_platform//:platform
INFO: Elapsed time: 19969.544s, Critical Path: 4113.50s
INFO: 10487 processes: 127 internal, 10360 local.
FAILED: Build did NOT complete successfully
启用Release时
修改.bazelrc的第559行：
build:release_gpu_base --repo_env=TF_CUDA_COMPUTE_CAPABILITIES="sm_35,sm_50,sm_60,sm_70,sm_75,compute_80,compute_86"
也就是加了个“compute_86”。
启动编译
如果需要代理（bazel需要下载大量直连非常慢的资源），在命令提示符中输入：
git config --global http.proxy http://ip:port
git config --global https.proxy https://ip:port
set http_proxy=http://ip:port
set https_proxy=https://ip:port
如果C盘空间不够，可以单独指定临时文件夹位置（实测一次编译下来，临时文件夹需要50GB左右的空间）：
set TMP=D:/tmp
set TEMP=D:/tmp
构建pip包：
bazel --output_user_root=D:/output_user_root --output_base=D:/output_base build --config=mkl --config=numa --config=monolithic --define=no_tensorflow_py_deps=true --//tensorflow/core/kernels/mlir_generated:enable_gpu=false --experimental_strict_action_env=false --config=release_gpu_windows //tensorflow/tools/pip_package:build_pip_package
这一行编译，耗时极长。
- 下载的bazel会因为权限问题无法写入用户文件夹，指定另外的文件夹是比较方便的操作。
- 编译输出最好也指定另外的文件夹，一次编译下来会占用50GB左右的空间。
- 之前我说可以从GitHub提前下载bazel本体文件夹的压缩包再解压以免额外时间，即指定--distdir=D:/bazel，但现在发现可能有些包的校验码对不上，还是建议去掉这个选项在线下载。
- --config=mkl用于启用MKL（依赖libiomp5md.dll），可以不写以不启用MKL。
- --experimental_strict_action_env=false用于启用系统环境变量，以防链接时找不到DLL。
- --config=release_gpu_windows用于启用Release编译，之前我没启用文件体积很大，不知道启用了会不会影响debug，未测试。
- --//tensorflow/core/kernels/mlir_generated:enable_gpu=false用于关闭MLIR的GPU编译，否则编译会出错，类似于（关键词：xxx_kernel_generator_kernel.o failed (Exit -1073741515): tf_to_kernel.exe failed）：
ERROR: D:/tensorflow-2.5.0/tensorflow/core/kernels/mlir_generated/BUILD:893:23: compile tensorflow/core/kernels/mlir_generated/neg_gpu_f16_f16_kernel_generator_kernel.o failed (Exit -1073741515): tf_to_kernel.exe failed: error executing command
cd D:/output_base/execroot/org_tensorflow
bazel-out/x64_windows-opt/bin/tensorflow/compiler/mlir/tools/kernel_gen/tf_to_kernel.exe --unroll_factors=4 --tile_sizes=256 --arch=compute_50,compute_52,compute_60,compute_61,compute_70,compute_75,compute_80,compute_86 --input=bazel-out/x64_windows-opt/bin/tensorflow/core/kernels/mlir_generated/neg_gpu_f16_f16.mlir --output=bazel-out/x64_windows-opt/bin/tensorflow/core/kernels/mlir_generated/neg_gpu_f16_f16_kernel_generator_kernel.o --enable_ftz=False --cpu_codegen=False
Execution platform: @local_execution_config_platform//:platform
Target //tensorflow/tools/pip_package:build_pip_package failed to build
然后构建包，会在源码根目录下创建tfpy文件夹，成功后whl包就会放到这个文件夹下，文件夹名字随意：
bazel-bin\tensorflow\tools\pip_package\build_pip_package.exe tfpy
构建C++库
执行：
bazel --output_user_root=D:/output_user_root --output_base=D:/output_base build --config=mkl --config=numa --config=monolithic --//tensorflow/core/kernels/mlir_generated:enable_gpu=false --experimental_strict_action_env=false --config=release_gpu_windows //tensorflow:tensorflow_cc.dll
bazel --output_user_root=D:/output_user_root --output_base=D:/output_base build --config=mkl --config=numa --config=monolithic --//tensorflow/core/kernels/mlir_generated:enable_gpu=false --experimental_strict_action_env=false --config=release_gpu_windows //tensorflow:tensorflow_cc_dll_import_lib
即在bazel-bin\tensorflow目录下得到tensorflow_cc.dll和tensorflow_cc.lib。
再执行：
bazel --output_user_root=D:/output_user_root --output_base=D:/output_base build --config=mkl --config=numa --config=monolithic --//tensorflow/core/kernels/mlir_generated:enable_gpu=false --experimental_strict_action_env=false --config=release_gpu_windows //tensorflow:install_headers
即在bazel-bin\tensorflow目录下得到include文件夹，里面是需要的头文件。
安装须知
编译得到whl包后，由于源码中依赖包版本限制得比较死，pip安装过程中会各种报错（尤其是grpcio，TensorFlow指定的版本无法在Windows下成功编译），因此需要先去掉依赖安装TensorFlow本体：
pip install tensorflow-2.5.0-cp39-cp39-win_amd64.whl --no-deps
然后再安装其依赖的包，如：
pip install absl-py astunparse flatbuffers google_pasta grpcio h5py keras_preprocessing numpy opt_einsum protobuf six termcolor typing_extensions wheel wrapt gast tensorboard tensorflow_estimator
如有缺失，后续导入的时候会有明确提示。
如果启用MKL，则需要另外将libomp.dll放到系统python.exe同目录下（即便是虚拟环境也要放到系统的Python目录下）。
结果获取（还没来得及测试，有问题请评论或私信）
pip安装包：tensorflow-2.5.0-cp39-cp39-win_amd64.whl（394.34MB）
pip安装包压缩版：tensorflow-2.5.0-cp39-cp39-win_amd64.rar（250.15MB）
C++库：tensorflow-cpp-2.5.0.rar（218.10MB）
pip安装包（MKL）：tensorflow-mkl-2.5.0-cp39-cp39-win_amd64.whl（393.64MB）
pip安装包压缩版（MKL）：tensorflow-mkl-2.5.0-cp39-cp39-win_amd64.rar（249.58MB）
C++库（MKL）：tensorflow-cpp-mkl-2.5.0.rar（217.57MB）
libomp.dll：libomp-tensorflow-2.5.0.rar（378.75KB）
pip安装包的压缩版，解压后再用ZIP打包，改后缀为“whl”即可正常安装。
本文内容由网友自发贡献，版权归原作者所有，本站不承担相应法律责任。如您发现有涉嫌抄袭侵权的内容，请联系:hwhale#tublm.com(使用前将#替换为@)
