SciPy

Building from source on Windows

Overview

Compared to OSX and Linux, building NumPy and SciPy on Windows is more difficult, largely due to the lack of compatible, open-source libraries like BLAS/LAPACK and open-source compilers that are necessary to build both libraries and have them perform relatively well. It is not possible to just call a one-liner on the command prompt as you would on other platforms via sudo apt install machinery.

This document describes one option to build OpenBLAS and SciPy from source that was validated for SciPy 1.6.0. However, in light of all the work currently being done, please do not expect these instructions to be accurate in the long-run and be sure to check up on any of the open source projects mentioned for the most up-to-date information. For more information on all of these projects, the Mingwpy website is an excellent source of in-depth information than this document will provide.

Building the Released SciPy

This section provides the step-by-step process to build the released SciPy. If you want to build completely from source, you should estimate at least three hours to build all libraries and compile SciPy. Feel free to stop and inspect any step at any time, but for this section, we’ll just mention the steps without providing an in-depth explanation for the reasons behind them. If you have further questions about what we’re doing, more in-depth documentation is provided in the sections below. Also, please make sure to read this section before proceeding, as the presence or absence of error messages in general is not a good indication of whether you’ve completed a step correctly. Each step creates particular files, and what ultimately matters is whether you have built the required files rather than whether error messages appeared in your terminal.

Building OpenBLAS

First, we need to install the software required to build OpenBLAS, which is the BLAS library that we’re going to use. Because the software to build OpenBLAS is different than that required to build SciPy and because OpenBLAS takes a long time to build, we’re going to start building OpenBLAS first and then explain what to do next while the OpenBLAS build is running. Alternatively, if you’d rather download a pre-built OpenBLAS, download the one of the pre-built zip files and skip to the Installing OpenBLAS section below.. However it is also likely that your version of Windows and the compiler you wish to use won’t be compatible with what these prebuilt binaries produced. This is still one of the main pain points of building for Windows. That’s why we will attempt to build our own OpenBLAS.

We start by installing the MSYS2 platform, on which the OpenBLAS build will take place. First, download the MSYS2 installer from msysintaller via choosing 32 or 64 bit. Make sure to install the correct architecture for the SciPy that you want to build (e.g., 32 or 64 bit). If you are not sure which one to use, proceed with 64bit. Please follow the installation instructions carefully, especially step 6 and 7 to update all components.

Note

Occasionally, during the updates, the terminal might ask you to close the terminal but then might refuse to be closed and hang. If this happens, you can kill it via Task Manager and continue with the instructions.

Now, the next step is to install some more package bundles that we will need. Open a MSYS2 MinGW (64 or 32 bit) terminal and type the following depending on the architecture of your choice; run the following for the common 64-bit build

pacman -S --needed base-devel mingw-w64-x86_64-toolchain mingw-w64-x86_64-cmake git

and for 32-bit run instead

pacman -S --needed base-devel mingw-w64-i686-toolchain mingw-w64-i686-cmake git

Again, if you are not sure which one you want, choose 64-bit option in every step.

It will prompt to whether install everything in these packages and you can simply accept all via hitting enter key at each step which also takes some time to complete. Once you install everything, close and reopen the MSYS2 MinGW terminal.

If you already have a GitHub repository folder where you keep your own repos, it is better to use that location to keep things nice and tidy since we are going to clone yet another repository to obtain the source code. It should be somewhere convenient and with write permissions. If this is your first time then you can pick “DocumentsGitHub” as a viable option. We will assume that you picked this folder in the rest of this document. You can create a folder in “My Documents” using Windows Explorer. To make sure that we’re ready to build, type the following in the terminal one-by-one:

make
gfortran
gcc
git

Each of these commands should fail as we have not provided any arguments to them. However, an explicit failure from the program rather than from the command prompt implies that the program is accessible on the path, which is what we wanted to test. In turn, if an error about the command being not found is returned, then installation of the packages didn’t complete successfully. If any of these are missing, you’re not ready to build. Go back and make sure that MSYS2 is installed correctly and has the required packages enabled.

Now it’s time to clone the OpenBLAS repository somewhere convenient. Run the following line-by-line separately, modifying the path to your GitHub repo folder as appropriate.

cd C:\Users\<user name>\Documents\GitHub
git clone https://github.com/xianyi/OpenBLAS.git
cd OpenBLAS
git submodule update --init --recursive
git fetch --all --tags --prune

Now we are going to switch to a release of our choice. At the time of writing, the newest OpenBLAS release version is 0.3.7, hence we will use that.

git checkout tags/v0.3.7 -b v0.3.7

You can see all available options via

git tag

Now change the directory one level up via cd .. to get out of the directory and create a file named build_openblas.sh. The easiest way is to type

touch build_openblas.sh

Of course, you can still also use Windows Explorer to create a new txt file at that location and then rename it. So the resulting structure would be

my repo folder
    ├─── build_openblas.sh
    ├─── OpenBLAS
            ├─── ...

Then open this file in any text editor, like Notepad++, and paste the following content in this empty file:

# You may adjust to your preferred output directory
OPENBLAS_ROOT=/c/opt

# Adjust to match the MSYS2 version you installed
BUILD_BITS=64

# Print some gcc info that MSYS2 discovered in the path
which gcc
gcc --version

# Get into the repository that we cloned
cd OpenBLAS

# The following two lines clean up in case we make a mistake and need
# to run the script again
git clean -fxd
git reset --hard
rm -rf $OPENBLAS_ROOT/$BUILD_BITS

# Set architecture flags
march="x86-64"
extra="-fno-asynchronous-unwind-tables"
vc_arch="X64"
cflags="-O2 -march=$march -mtune=generic $extra"
fflags="$cflags -frecursive -ffpe-summary=invalid,zero"

# Build name for output library from gcc version and OpenBLAS commit.
GCC_TAG="gcc_$(gcc -dumpversion | tr .- _)"
OPENBLAS_VERSION=$(git describe --tags)
# Build OpenBLAS
# Variable used in creating output libraries
export LIBNAMESUFFIX=${OPENBLAS_VERSION}-${GCC_TAG}
make BINARY=$BUILD_BITS DYNAMIC_ARCH=1 USE_THREAD=1 USE_OPENMP=0 \
    NO_WARMUP=1 BUILD_LAPACK_DEPRECATED=1 \
    COMMON_OPT="$cflags" FCOMMON_OPT="$fflags"
make install PREFIX=$OPENBLAS_ROOT/$BUILD_BITS

This is the automation script that will make sure the right variables are used in the right place. Linux users are very familiar to such scripts, but for Windows users it might be a bit awkward. You can think of these as .bat files. The script should work as-in for MSYS2 64-bit, but you can change the variables to your situation as needed. After you’ve created this file and you are one directory up the OpenBLAS repo of that, start the OpenBLAS build with:

./build_openblas.sh

Building OpenBLAS is challenging and time-consuming. The build may fail with an error after a few hours but may also fail silently and produce an incorrect binary. Please, if you have any issues, report them so that we can save the next person’s time.

One of the known issues is the following; if you, by any chance, receive the following error

<command-line>:0:4: error: expected identifier or '(' before numeric constant

that means you have some header file definition clash and you have to downgrade certain items. This is not related to SciPy but let’s attempt to provide a solution. See this OpenBLASwiki page to read on which packages to downgrade and how to do it. Basically, it involves downloading three files. Then in the MSYS terminal change the directory to the place where you downloaded the files and run the commands given in the wiki link. Then come back to the script directory where /build_openblas.sh lives and try again. This should be sufficient for you to build OpenBLAS.

While you’re waiting on OpenBLAS to finish building, go ahead and install build tools from Microsoft, since these take a while to install and you’ll need them later.

After the build_openblas.sh script has completed, there should be an libopenblas.....a as a resulting artifact. If OPENBLAS_ROOT was set to C:\\opt, then you might see a line like this in the MSYS2 terminal:

Copying the static library to /c/opt/64/lib

This is very good news: you have successfully built OpenBLAS!

Installing OpenBLAS

Look for the lib folder in the folder you used as a parameter to OPENBLAS_ROOT. (It’s /c/opt/64/lib if you didn’t change anything in the script.) You will find three .a files such as (the names can differ):

libopenblas_v0.2.20-2-g5f998efd-gcc_9_2_0.a
libopenblas_v0.2.20-2-g5f998efd-gcc_9_2_0.dll.a
libopenblas_v0.2.20-2-g5f998efd-gcc_9_2_0.p-r0.2.20.a

From these three we are interested only in the first one. Just make a copy and rename it to openblas.a.

If you don’t have that file, you’ll probably need to find out what happened and then build OpenBLAS again. We know this is very annoying, but unfortunately we have no other alternatives. The first place to look for is inside the OpenBLAS directory. If the build succeeds but (for some reason) auto-moving files to OPENBLAS_ROOT fails, the artifacts will stay inside the OpenBLAS repo folder. But if you have that file, that’s great and we’ll assume that you’ve completed this step correctly. Proceeding on that assumption, let’s build SciPy.

Before continuing, make sure that you don’t have other copies of either openblas.a or libopenblas.a from previous attempts or via previous downloads. Multiple copies could result in later build errors that will be difficult to debug. If this is the first attempt, you don’t need to worry about this step.

Building SciPy

Once you have built OpenBLAS, it’s time to build SciPy. Before continuing, make sure to install the following software for building on the latest Python version. For building on other Python versions, see the WindowsCompilers page. We are also assuming that your Python is on the system path. That is to say, when you type python in the Windows command prompt the correct Python is executed.

Install Microsoft Visual Studio 2017 or 2019 Community Edition (use the build tools from Microsoft). If you feel that it is too bloated to install everything in that bundle (which we do feel a bit so) then here are a subset which are tested during the build of SciPy 1.6.0 and VS 2019. You can switch to the individual items view at the top and select only the following

C++ Core Features
Windows Universal C Runtime
MSVC v142 - VS 2019 C++ x64/x86 build tools (...)
Windows 10 SDK (10.0.18362.0)
C++ 2019 Redistributable Update
C++ Clang-cl for 142 build tools (x64/x86)
C++ Clang Compiler for Windows (8.0.1)

Just like before, pick a convenient place to clone SciPy. Next to OpenBLAS is often a convenient option (note: not inside OpenBLAS folder but next to). Continuing the example from above

my repo folder
    ├─── build_openblas.sh
    ├─── OpenBLAS
    ├─── SciPy
            ├─── ...

Again using the same generic example folder from above

cd C:\Users\<user name>\Documents\GitHub
git clone https://github.com/scipy/scipy.git
cd scipy

Now we need to copy the openblas.a file that we’ve built earlier to the correct location. If your Python is installed somewhere like the following:

C:\Users\<user name>\AppData\Local\Programs\Python\Python38\python.exe

then you’ll need to put the openblas.a file that we previously copied and renamed somewhere like the following:

C:\Users\<user name>\AppData\Local\Programs\Python\Python38\Lib

Adjust the location accordingly based on where python.exe is located.

At this stage, we are done with the OpenBLAS part and hopefully we will not need to build OpenBLAS anytime soon. But we tend to build SciPy more often as it is on a quicker release cycle. Hence it makes sense to use Windows cmd or Powershell for the the build as it is a more native tool. This requires placing the MinGW compilers on the path. Hence, make sure that the following folder (or the folder you have installed MSYS to) is on the system path variable sufficiently close to the top.

C:\MSYS64\MINGW64\BIN

For a sanity check, restart cmd or Powershell and type:

gfortran

If you see a missing command error with the above, gfortran is not correctly installed or is still not on the path. However, we assume that it is now on the path and accessible.

Now install the dependencies that we need to build and test SciPy.

python -m pip install wheel setuptools numpy>=1.16.5 Cython>=0.29.18 pybind11>=2.4.3 pytest pytest-xdist

Note

These instructions use pip as the package manager. You can also use conda according to the instructions in the Development environment quickstart guide (Ubuntu) with minimal modifications (e.g. you don’t need to install gfortran and git because you already have them).

The last two are for using SciPy’s test suite, which is handy if you want to test some new change locally.

Please note that this is a simpler procedure than what is used for the official binaries. Your binaries will only work with the latest NumPy. For building against older NumPy versions, see Building Against an Older NumPy Version.

Assuming that you are in the top of the SciPy repository directory where setup.py is and assuming that you have set up everything correctly, you are ready to build. Run the following commands:

python setup.py build

You may verify that the OpenBLAS library was correctly picked up by looking for the following in your build log:

FOUND:
   libraries = ['openblas']
   library_dirs = ['C:\...........\lib']
   language = f77
   define_macros = [('HAVE_CBLAS', None)]

Notice that there will be multiple lines similar to these. You only need to track the OpenBLAS one.

When everything finishes without an error, congratulations! You’ve built SciPy!

You can further install the built SciPy via

python setup.py install

Just make sure that you uninstalled the existing installation of other SciPy if there were any (by the regular pip uninstall scipy machinery).

Building Against an Older NumPy Version

If you want to build SciPy to work with an older NumPy version, then you will need to replace the NumPy “distutils” folder with the folder from the latest numpy. The following Powershell snippet can upgrade NumPy distutils while retaining an older NumPy ABI.

$NumpyDir = $((python -c 'import os; import numpy; print(os.path.dirname(numpy.__file__))') | Out-String).Trim()
rm -r -Force "$NumpyDir\distutils"
$tmpdir = New-TemporaryFile | %{ rm $_; mkdir $_ }
git clone -q --depth=1 -b master https://github.com/numpy/numpy.git $tmpdir
mv $tmpdir\numpy\distutils $NumpyDir

Additional Resources

As discussed in the overview, this document is not meant to provide extremely detailed explanations on how to build NumPy and SciPy on Windows. This is largely because currently, there is no single superior way to do so and because the process for building these libraries on Windows is under development. It is likely that any information will go out of date relatively soon. If you wish to receive more assistance, please reach out to the NumPy and SciPy mailing lists, which can be found here. There are many developers out there working on this issue right now, and they would certainly be happy to help you out! Google is also a good resource, as there are many people out there who use NumPy and SciPy on Windows, so it would not be surprising if your question or problem has already been addressed.