Can one of the admins verify this patch?

Thanks for your pull request. It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

📝 Please visit https://cla.developers.google.com/ to sign.

Once you've signed, please reply here (e.g. I signed it!) and we'll verify. Thanks.

• If you've already signed a CLA, it's possible we don't have your GitHub username or you're using a different email address. Check your existing CLA data and verify that your email is set on your git commits.
• If your company signed a CLA, they designated a Point of Contact who decides which employees are authorized to participate. You may need to contact the Point of Contact for your company and ask to be added to the group of authorized contributors. If you don't know who your Point of Contact is, direct the project maintainer to go/cla#troubleshoot.
• In order to pass this check, please resolve this problem and have the pull request author add another comment and the bot will run again.

I signed it!

CLAs look good, thanks!

Also note that AVX and AVX2 are not enabled by default in CMakeLists.txt, even when native arch optimization is enabled. I will place another pull request to fix this in the future, along with a number of other CMake fixes.

Jenkins, test this please.

Jenkins, test this please.

Looks like the CI server is having some unrelated failures? Is that common?

Unfortunately more common than we would like.
Retrying tests.
Jenkins, test this please.

The change looks good to me, but I'll defer to @benoitsteiner since it's in Eigen code. (I'm not sure how or whether we pull in Eigen code from other repositories, and whether it would be better to make the change upstream first.)

@scottmudge, thanks for your job about avx2, great!
I pulled the latest tensorflow project which included your commits about avx2 and compiled successfully by command:
cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release -DSWIG_EXECUTABLE=C:\D\tools\swigwin-3.0.12\swig.exe -DPYTHON_EXECUTABLE=C:\D\tools\Anaconda3\python.exe -DPYTHON_LIBRARIES=C:\D\tools\Anaconda3\python35.lib -Dtensorflow_ENABLE_GPU=ON -DCUDNN_HOME="C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0" -Dtensorflow_WIN_CPU_SIMD_OPTIONS=/arch:AVX

but it still hints when running,
Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2

What else should I do to enable AVX2 for tensorflow on WINDOWS? Thanks so much.

Frank

Hey Frank,

Yes the CMakeLists.txt file in TensorFlow needs some modifications; it does not properly set the AVX/AVX2 flags.

Find this line in the CMakeLists.txt file in ./tensorflow/contrib/cmake/:

if (tensorflow_OPTIMIZE_FOR_NATIVE_ARCH)
  include(CheckCXXCompilerFlag)
  CHECK_CXX_COMPILER_FLAG("-march=native" COMPILER_OPT_ARCH_NATIVE_SUPPORTED)
  if (COMPILER_OPT_ARCH_NATIVE_SUPPORTED)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=native")
  endif()
endif()

And change it to:

if (tensorflow_OPTIMIZE_FOR_NATIVE_ARCH)
  include(CheckCXXCompilerFlag)
  CHECK_CXX_COMPILER_FLAG("-march=native" COMPILER_OPT_ARCH_NATIVE_SUPPORTED)
  if (WIN32)
	  if (COMPILER_OPT_ARCH_NATIVE_SUPPORTED)
		set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX /arch:AVX2")
	  endif()
  else()
	if (COMPILER_OPT_ARCH_NATIVE_SUPPORTED)
		set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=native")
	endif()
  endif()
endif()

Sort of a hacky way to force enable it, but it'll do the job for now. I need to do another pull request with a better fix.

If you get error C1001 when compiling with GPU support, take a look at this thread:

#9470

I'm not sure if it does it on the master branch, but I had the problem as of v1.3.1

@scottmudge

I tried your hacky way, but it still indicated the issue of “was not compiled to use: AVX2”. I noticed the information in configuration stage on my platform,
-- Performing Test COMPILER_OPT_ARCH_NATIVE_SUPPORTED
-- Performing Test COMPILER_OPT_ARCH_NATIVE_SUPPORTED - Failed
-- Performing Test COMPILER_OPT_WIN_CPU_SIMD_SUPPORTED
-- Performing Test COMPILER_OPT_WIN_CPU_SIMD_SUPPORTED - Success

I got the code from tensorflow:master with the latest commit as follow,
commit 10c871e
Merge: 87ac990 188297f
Author: Shanqing Cai cais@google.com
Date: Mon Oct 9 09:34:12 2017 -0400

Finally I modified one line in CMakeLists.txt with adding "/arch:AVX2", and it seems to work. Haha, it is another hacky way.

if (tensorflow_WIN_CPU_SIMD_OPTIONS)
if (WIN32)
CHECK_CXX_COMPILER_FLAG("${tensorflow_WIN_CPU_SIMD_OPTIONS}" COMPILER_OPT_WIN_CPU_SIMD_SUPPORTED)
if(COMPILER_OPT_WIN_CPU_SIMD_SUPPORTED)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${tensorflow_WIN_CPU_SIMD_OPTIONS} /arch:AVX2")
else()
message(FATAL_ERROR "${tensorflow_WIN_CPU_SIMD_OPTIONS} not supported")
endif()
endif()
endif()

Thanks so much for your help.

@yang0773
Can you please share python wheels for Windows with AVX2 support?
I've been trying to build tensorflow for couple of days, but still get stupid fatal error C1002...

@iNomaD Here is one I compiled, v1.3.1 w/ AVX2, GPU (up to CUDA 6.1), x64 for Windows:

https://github.com/scottmudge/tensorflow/releases/download/v1.3.1_mod/tensorflow_gpu-1.3.1-cp36-cp36m-win_amd64.whl

Actually tf on windows can be built like this:

cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release ^
-DSWIG_EXECUTABLE=C:\local\swigwin-3.0.10\swig.exe ^
-DPYTHON_EXECUTABLE=C:\local\Anaconda3-4.1.1-Windows-x86_64\python.exe ^
-DPYTHON_LIBRARIES=C:\local\Anaconda3-4.1.1-Windows-x86_64\libs\python35.lib ^
-Dtensorflow_ENABLE_GPU=ON ^
-DCUDNN_HOME="C:\local\cudnn-8.0-v5.1\cuda" ^
-Dtensorflow_WIN_CPU_SIMD_OPTIONS=/arch:AVX2

The last option can be: =/arch:AVX2, then log "Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2" is disappeared.

can you build it using AVX2 and AVX at the same time?

https://docs.microsoft.com/en-us/cpp/preprocessor/predefined-macros

In Visual C++, when you set /arch:AVX2 both AVX and AVX2 are used.