C or Cython Extensions¶
Astropy supports using C extensions for wrapping C libraries and Cython for
speeding up computationally-intensive calculations. Both Cython and C extension
building can be customized using the get_extensions
function of the
setup_package.py
file. If defined, this function must return a list of
distutils.core.Extension
objects. The creation process is left to the
subpackage designer, and can be customized however is relevant for the
extensions in the subpackage.
While C extensions must always be defined through the get_extensions
mechanism, Cython files (ending in .pyx
) are automatically located and
loaded in separate extensions if they are not in get_extensions
. For
Cython extensions located in this way, headers for numpy C functions are
included in the build, but no other external headers are included. .pyx
files present in the extensions returned by get_extensions
are not
included in the list of extensions automatically generated extensions. Note
that this allows disabling a Cython file by providing an extension that
includes the Cython file, but giving it the special name
‘cython_skip’. Any
extension with this package name will not be built by setup.py
.
Note
If an Extension
object is provided for Cython
source files using the get_extensions
mechanism, it is very
important that the .pyx
files be given as the source
, rather than the
.c
files generated by Cython.
Using Numpy C headers¶
If your C or Cython extensions uses numpy
at the C level, you probably
need access to the numpy C headers. A common idiom you can find in the numpy
docs or other examples involves getting the include directory by calling
numpy.get_include()
. However, using this in setup_package.py
will not
work, because setup_package.py
needs to be able to import even when none of
the dependencies are present. To work around this need, simply include the
string 'numpy'
in the list that is passed to the include_dirs
argument
of distutils.core.Extension
. The astropy setup helpers will then use
numpy.get_include()
downstream once it is certain that the dependencies
have actually been processed. For example:
from distutils.extension import Extension
def get_extensions():
return Extension(name='myextension', sources=['myext.pyx'],
include_dirs=['numpy'])
Installing C header files¶
If your C extension needs to be linked from other third-party C code, you probably want to install its header files along side the Python module.
Create an
include
directory inside of your package for all of the header files.Use the
get_package_data
hook insetup_package.py
to install those header files. For example, theastropy.wcs
package has this:def get_package_data(): return {'astropy.wcs': ['include/*.h']}
Preventing importing at build time¶
In rare cases, some packages may need to be imported at build time.
Unfortunately, anything that requires a C or Cython extension will fail to
import until the build phase has completed. In this cases, the
_ASTROPY_SETUP_
variable can be used to determine if the package is being
imported as part of the build and choose to not import problematic modules.
_ASTROPY_SETUP_
is inserted into the builtins, and is True
when inside
of astropy’s setup.py
script, and False
otherwise.
For example, suppose there is a subpackage foo
that needs to
import a module called version.py
at build time in order to set
some version information, and also has a C extension, process
,
that will not be available in the source tree. In this case,
astropy/foo/__init__.py
would probably want to check the value of
_ASTROPY_SETUP_
before importing the C extension:
try:
from . import process
except ImportError:
if not _ASTROPY_SETUP_:
raise
from . import version
Speed up your builds with ccache¶
ccache is a tool that caches compiled sources so that they don’t have to be recompiled (so long as they are unchanged) even if the outputs have been deleted. This means that if you switch branches or clean your source checkout you can save a lot of time by avoiding the majority of re-compiles from scratch.
Because installation and configuration of ccache varies from platform to platform, please consult the ccache documentation and/or Google to set up ccache on your system–this is strongly encouraged for anyone doing significant development of Astropy or scientific programming in general.