-
Notifications
You must be signed in to change notification settings - Fork 278
Windows Development
The NuPIC Core C++ library can be built with MinGWPy GCC or Microsoft Visual C++ compilers. The associated Python SWIG nupic.bindings can ONLY be built with the MinGWPy GCC compiler. To use the C++ library with the NuPIC Python project you must build the library with the MinGW-Py GCC tools.
NuPIC's dependency on the Core C++ library can be found here - https://github.com/numenta/nupic/wiki/NuPIC's-Dependency-on-nupic.core
It describes how the C++ library is linked to other languages using SWIG, such as 64-bit Python (x86-64 MSI). The NuPIC.Core repository consist of two parts;
- Main C++ core library, and
- Python SWIG based bindings.
The NuPIC Core C++ library file is split into two files;
-
%NUPIC_CORE%/release/lib/nupic_core_solo.lib
contains only the core library -
%NUPIC_CORE%/release/lib/nupic_core.lib
contains the C++ core and external support libraries
Where NUPIC_CORE
is an environment variable that points to the git cloned directory.
CMake based build files are used to define the entire build process. The nupic.bindings SWIG side uses Python distutil and setuptools. The CMake-GUI application can be used to generate MinGW Makefiles files.
The %NUPIC_CORE%\appveyor.yml
script file shows how the AppVeyor automated build system progresses through building, packaging, and deployment using MinGWPy GCC.
If you need to make changes to the nupic.core C++ code and/or Python SWIG bindings, follow the remaining sections to see how to rebuild it.
If you only intend to use the C++ library, and not use Python NuPIC, this README.md file describes how to setup and build nupic.core with Microsoft Visual Studio 2015 -
The following applications are required when rebuilding the core C++ library, Python SWIG bindings, and if required the external support libraries;
- CMake - version 3.1+
- Python 2.7.9+ - x86-64 version
-
7-Zip - Or any archive manager that can extract
.7z
files -
MinGW GCC for Python -
mingwpy_amd64_vc90.7z
andlibpython-cp27-none-win_amd64.7z
-
NumPy C++ headers - pip install
numpy-1.9.3+vanilla-cp27-none-win_amd64.whl
or via MSys2 commandpacman -S mingw64/mingw-w64-x86_64-python2-numpy
- MSYS2 - 64 bit version. Only needed when rebuilding external support libraries with MinGWPy GCC or installing alternative support libraries.
-
Git for Windows - Includes patch.exe required when building external libraries using Visual Studio. Make sure
C:\Program Files\Git\usr\bin
is added to your PATH environment variable.
The following table shows example CMake common settings;
Name | Value |
---|---|
Source code | %NUPIC_CORE% |
Binaries | %NUPIC_CORE%/build/scripts |
CMAKE_INSTALL_PREFIX | %NUPIC_CORE%/build/release |
PY_EXTENSIONS_DIR | %NUPIC_CORE%/bindings/py/nupic/bindings |
For rebuilding nupic.core and nupic.bindings you need to use the MinGW Makefiles
generator. For example;
rem Clone the repo
git clone https://github.com/numenta/nupic.core.git
cd nupic.core
rem Setup Python x64 to use MinGWPy GCC compilers
copy nupic.core\external\windows64-gcc\bin\distutils.cfg C:\Python27-x64\Lib\distutils
rem Setup nupic.core and a place to store build files
cd nupic.core
set NUPIC_CORE=%CD%
mkdir build\scripts
cd build\scripts
rem Run cmake to generator MinGW Makefiles
cmake -G "MinGW Makefiles"
-DCMAKE_BUILD_TYPE=Debug
-DCMAKE_INSTALL_PREFIX=..\release
-DPY_EXTENSIONS_DIR=..\..\bindings\py\nupic\bindings
..\..
rem Build and install NuPIC.core, and build SWIG binding libraries
mingw32-make -f Makefile install
rem Install Python SWIG nupic.bindings
cd %NUPIC_CORE%
python setup.py install
After building and installing, the following can run C++ related tests;
cd %NUPIC_CORE%\build\release\bin
cpp_region_test
unit_tests
You can run the nupic.bindings tests with py.test;
cd %NUPIC_CORE%
py.test --pyargs nupic.bindings
The %NUPIC_CORE%\.gitignore
file has a rule that ignores any directory called build/
from Git. Making that directory a handy place to store build dependencies.
- You must use the
mingwpy_amd64_vc90.7z
toolchain. Other MinGW x64 GCC toolchains do not work. It also must be the amd64 vc90 version to match up with the Python 2.7.9+ x64 version. - The
libpython-cp27-none-win_amd64.7z
has to be extracted into your Python 2.7 x64 directory before running cmake (two.a
library files placed intoC:\Python27-x64\libs
directory). - It is convenient to extract the
mingwpy_amd64_vc90.7z
into aC:\mingwpy
directory. - After extracting the MinGW GCC 7z archive your
PATH
environment variable will needC:\mingwpy\bin;
adding to it. - Make sure to copy
%NUPIC_CORE%\external\windows64-gcc\bin\distutils.cfg
into theC:\Python27-x64\Lib\distutils
directory. -
mingw32-make
is used with the CMake generated Makefiles.
The C++ Core library depends on a handful of support libraries that are distributed within the GitHub repository. The following sub-sections can be used as a guide if you require changes to these pre-built libraries.
The MinGWPy GCC versions can be found in %NUPIC_CORE%\external\windows64-gcc\lib
.
The following libraries are embedded into the %NUPIC_CORE%/lib/nupic_core
library.
Library | Version | Filename | Website |
---|---|---|---|
APR | 1.5.1 | apr-1.5.1-win32-src.zip | https://apr.apache.org/ |
Apr Iconv | 1.2.1 | apr-iconv-1.2.1-win32-src-r2.zip | https://apr.apache.org/ |
Apr Util | 1.5.4 | apr-util-1.5.4-win32-src.zip | https://apr.apache.org/ |
CapnProto | 0.5.3 | capnproto-c++-win32-0.5.3.zip | https://capnproto.org |
Yaml | 0.1.5 | yaml-0.1.5.tar.gz | http://pyyaml.org/wiki/LibYAML |
Yaml Cpp | 0.3.0 | yaml-cpp-0.3.0.tar.gz | https://code.google.com/p/yaml-cpp/ |
Z Lib | 1.2.8 | zlib-1.2.8.tar.gz | http://www.zlib.net/ |
Extract them all into %NUPIC_CORE%/build
directory. You may need to un-nest some of the directories.
Note that APR expects certain directory names. Rename the following directories;
apr-1.5.1
to apr
apr-iconv-1.2.1
to apr-iconv
apr-util-1.5.4
to apr-util
Most of the libraries have a CMakeList.txt file. A CMakeList.txt file allows the use of the CMake-GUI application. Which can make it easy to tweak build variables, e.g. the required update of CMAKE_INSTALL_PREFIX
for each library; and advanced options such as compiler settings.
Install the 64 bit version of MSys2 - http://sourceforge.net/p/msys2/wiki/MSYS2%20installation/
From the MSys2 command line run the update-core
script followed by pacman -Su
, to make sure MSys2 is up to date.
Building the Apr-iconv library is not required and can be skipped.
All support libraries are need to be built for x64.
If a solution contains an INSTALL project, the install scripts are placed inside a file called cmake_install.cmake
A manual install can be performed by execute the following command;
cmake.exe -DBUILD_TYPE=Release -P cmake_install.cmake
This implies that your PATH
environment variable has a directory to the cmake.exe (typically "C:\Program Files (x86)\CMake\bin").
The C++ headers from NumPy can be installed via a pre-built NumPy Python package. For example; from http://www.lfd.uci.edu/~gohlke/pythonlibs/ or via MSys2 pacman -S mingw64/mingw-w64-x86_64-python2-numpy
command.
Using pacman -Ss apr
searches for APR packages. Use the following commands to install the required APR libraries; pacman -S mingw64/mingw-w64-x86_64-apr --noconfirm
and pacman -S mingw64/mingw-w64-x86_64-apr-util --noconfirm
This installs (assuming a default installation directory for MSys2) the libraries into C:/msys64/mingw64/lib
. The two APR libraries (libapr-1.a
and libaprutil-1.a
) can then be copied into the %NUPIC_CORE%\external\windows64-gcc-\lib
directory. And used with the CMake build command as described above to rebuild the core library.
This is part of the main nupic.core CMake build process. The appropriate version is git cloned, as well as the associated Cap'n Proto compiler tools. It is built using a MinGW Makefile
and ends up in the %NUPIC_CORE%\build\scripts\ThirdParty\Install
directories.
A CMake command, or the CMake-GUI program, can be used to generate MinGW Makefiles. These can then be used with the mingw32-make -f Makefile install
command to rebuild these libraries. Make sure that the CMAKE_BUILD_TYPE
and CMAKE_INSTALL_PREFIX
are set appropriately.