Edit

Setup BuildMklDnn VS 2017

  • Article

Creating a compiled MKL-DNN 0.14 for CNTK

If you want to build the MKL-DNN 0.14 for CNTK yourself, follow these instructions. The compiled MKL-DNN version will be located in c:\local\mklml-mkldnn-0.14

The following steps assume that the MKL-DNN sources have been unzipped to the directory c:\local\src\mkl-dnn-0.14. The instructions below will create the MKL-DNN libraries and header files in the directory c:\local\mklml-mkldnn-0.14. If this location differs in your case, please adapt the steps below to match your directory structure.

Note that MKL-DNN headers and libraries are stored in the same location as MKLML to simpilify setup, since their file names are different. mkl_dnn* files are for dnn functions in MKL which is not open source, while mkldnn* files are built from open source MKL-DNN.

There are two ways to build the CNTK MKL-DNN library, either through a batch file or manually.

Building through a batch file

In the directory Tools\devInstall\Windows you find the batch file buildMklDnnVS17.bat. This batch file takes two parameters to build the CNTK MKL-DNN library. The first parameter is the directory of the MKL-DNN source files you unpacked above, the second parameter is the destination directory for the created MKL-DNN library. In our case follow these steps from a standard Windows command prompt:

c:
cd \repos\cntk\Tools\devInstall\Windows
buildMklDnnVS17 C:\local\src\mkl-dnn-0.14 C:\local\mklml-mkldnn-0.14

Set the environment variable MKL_PATH to the MKL-DNN location (this is the second parameter in the command invocation above, not the sources location!

setx MKL_PATH c:\local\mklml-mkldnn-0.14

Manual build steps for a CNTK MKL-DNN library

Note

If VS2017INSTALLDIR environment variable was not set by the Visual Studio installer, set it manually. The exact installation directory depends on both the version (2017 in this case) and offering (Community, Enterprise, etc) of Visual Studio, according to the following pattern: C:\Program Files (x86)\Microsoft Visual Studio\<version>\<offering>.

For Visual Studio 2017 Enterprise edition, run:

setx VS2017INSTALLDIR "C:\Program Files (x86)\Microsoft Visual Studio\2017\Enterprise"

For Visual Studio 2017 Community edition, run:

setx VS2017INSTALLDIR "C:\Program Files (x86)\Microsoft Visual Studio\2017\Community"
  1. Launch a 64-bit Visual Studio 2017 development environment. The most convenient way is to execute the batch file vcvarsall.bat in the Visual Studio directory with the required parameters from a standard windows command shell:
"%VS2017INSTALLDIR%\VC\Auxiliary\Build\vcvarsall.bat" amd64 --vcvars_ver=14.11

If you can't find this batch file on your system, please revisit the installation of VS2017 and make sure you have the 'VC++ version 15.4 v14.11 toolset' option selected. 2. Create a build directory, and set MKLROOT for MKL-DNN cmake to locate MKLML library:

cd c:\local\src\mkl-dnn-0.14
set MKLROOT=c:\local\mklml-mkldnn-0.14
  1. Create a release build using CMake, and copy built binary to MKL_PATH (user may replace release with debug to create a debug build):
cd cmake
cmake .. -G "Visual Studio 15" -DCMAKE_BUILD_TYPE=Release
msbuild "Intel(R) MKL-DNN.sln" /t:Rebuild /p:Configuration=Release /m
copy ..\include\* c:\local\mklml-mkldnn-0.14\include
copy src\Release\*.lib c:\local\mklml-mkldnn-0.14\lib
copy src\Release\*.dll c:\local\mklml-mkldnn-0.14\lib