Skip to content

Latest commit

 

History

History
155 lines (104 loc) · 5.15 KB

INSTALL.rst.txt

File metadata and controls

155 lines (104 loc) · 5.15 KB

Building and installing NumPy

IMPORTANT: the below notes are about building Numpy, which for most users is not the recommended way to install Numpy. Instead, use either a complete scientific Python distribution (recommended) or a binary installer - see http://scipy.org/install.html.

Building NumPy requires the following software installed:

  1. For Python 2, Python__ 2.7.x or newer. For Python 3, Python__ 3.4.x or newer.

    On Debian and derivative (Ubuntu): python python-dev

    On Windows: the official python installer on Python__ is enough

    Make sure that the Python package distutils is installed before continuing. For example, in Debian GNU/Linux, distutils is included in the python-dev package.

    Python must also be compiled with the zlib module enabled.

  2. Cython >= 0.19 (for development versions of numpy, not for released

    versions)

  3. nose__ (optional) 1.0 or later

    This is required for testing numpy, but not for using it.

Python__ http://www.python.org nose__ http://somethingaboutorange.com/mrl/projects/nose/

Note

If you want to build Numpy in order to work on Numpy itself, use runtests.py. For more details, see http://docs.scipy.org/doc/numpy-dev/dev/development_environment.html

Note

More extensive information on building Numpy (and Scipy) is maintained at http://scipy.org/scipylib/building/index.html

To install numpy run:

python setup.py build -j 4 install --prefix $HOME/.local

This will compile numpy on 4 CPUs and install it into the specified prefix. To perform an inplace build that can be run from the source folder run:

python setup.py build_ext --inplace -j 4

The number of build jobs can also be specified via the environment variable NPY_NUM_BUILD_JOBS.

Numpy needs a C compiler, and for development versions also Cython. A Fortran compiler isn't needed to build Numpy itself; the numpy.f2py tests will be skipped when running the test suite if no Fortran compiler is available. For building Scipy a Fortran compiler is needed though, so we include some details on Fortran compilers in the rest of this section.

On OS X and Linux, all common compilers will work. Note that for Fortran, gfortran is strongly preferred over g77, but if you happen to have both installed then g77 will be detected and used first. To explicitly select gfortran in that case, do:

python setup.py build --fcompiler=gnu95

On Windows, building from source can be difficult. Currently the most robust option is to use the Intel compilers, or alternatively MSVC (the same version as used to build Python itself) with Intel ifort. Intel itself maintains a good application note on this.

If you want to use a free compiler toolchain, the recommended compiler is MingwPy__. The older MinGW32 compiler set used to produce older .exe installers for Numpy itself is still available at https://github.com/numpy/numpy-vendor, but not recommended for use anymore.

MingwPy__ http://mingwpy.github.io

Configuring which BLAS/LAPACK is used if you have multiple libraries installed, or you have only one installed but in a non-standard location, is done via a site.cfg file. See the site.cfg.example shipped with Numpy for more details.

The Intel compilers work with Intel MKL, see the application note linked above. MingwPy__ works with OpenBLAS. For an overview of the state of BLAS/LAPACK libraries on Windows, see here.

OS X ships the Accelerate framework, which Numpy can build against without any manual configuration. Other BLAS/LAPACK implementations (OpenBLAS, Intel MKL, ATLAS) will also work.

For best performance a development package providing BLAS and CBLAS should be installed. Some of the options available are:

  • libblas-dev: reference BLAS (not very optimized)
  • libatlas-base-dev: generic tuned ATLAS, it is recommended to tune it to the available hardware, see /usr/share/doc/libatlas3-base/README.Debian for instructions
  • libopenblas-base: fast and runtime detected so no tuning required but a very recent version is needed (>=0.2.15 is recommended). Older versions of OpenBLAS suffered from correctness issues on some CPUs.

The package linked to when numpy is loaded can be chosen after installation via the alternatives mechanism:

update-alternatives --config libblas.so.3
update-alternatives --config liblapack.so.3

Or by preloading a specific BLAS library with:

LD_PRELOAD=/usr/lib/atlas-base/atlas/libblas.so.3 python ...

If you run into build issues and need help, the Numpy mailing list is the best place to ask. If the issue is clearly a bug in Numpy, please file an issue (or even better, a pull request) at https://github.com/numpy/numpy.