Skip to content

Latest commit

 

History

History
758 lines (552 loc) · 27.5 KB

tutorial_buildsystems.rst

File metadata and controls

758 lines (552 loc) · 27.5 KB
Raw

Spack Package Build Systems

You may begin to notice after writing a couple of package template files a pattern emerge for some packages. For example, you may find yourself writing an install() method that invokes: configure, cmake, make, make install. You may also find yourself writing "prefix=" + prefix as an argument to configure or cmake. Rather than having you repeat these lines for all packages, Spack has classes that can take care of these patterns. In addition, these package files allow for finer grained control of these build systems. In this section, we will describe each build system and give examples on how these can be manipulated to install a package.

Package Class Hierarchy

.. graphviz::

    digraph G {

        node [
            shape = "record"
        ]
        edge [
            arrowhead = "empty"
        ]

        PackageBase -> Package [dir=back]
        PackageBase -> MakefilePackage [dir=back]
        PackageBase -> AutotoolsPackage [dir=back]
        PackageBase -> CMakePackage [dir=back]
        PackageBase -> PythonPackage [dir=back]
    }

The above diagram gives a high level view of the class hierarchy and how each package relates. Each subclass inherits from the PackageBaseClass super class. The bulk of the work is done in this super class which includes fetching, extracting to a staging directory and installing. Each subclass then adds additional build-system-specific functionality. In the following sections, we will go over examples of how to utilize each subclass and to see how powerful these abstractions are when packaging.

Package

We've already seen examples of a Package class in our walkthrough for writing package files, so we won't be spending much time with them here. Briefly, the Package class allows for arbitrary control over the build process, whereas subclasses rely on certain patterns (e.g. configure make make install) to be useful. Package classes are particularly useful for packages that have a non-conventional way of being built since the packager can utilize some of Spack's helper functions to customize the building and installing of a package.

Autotools

As we have seen earlier, packages using Autotools use configure, make and make install commands to execute the build and install process. In our Package class, your typical build incantation will consist of the following:

def install(self, spec, prefix):
    configure("--prefix=" + prefix)
    make()
    make("install")

You'll see that this looks similar to what we wrote in our packaging tutorial.

The Autotools subclass aims to simplify writing package files and provides convenience methods to manipulate each of the different phases for a Autotools build system.

Autotools packages consist of four phases:

  1. autoreconf()
  2. configure()
  3. build()
  4. install()

Each of these phases have sensible defaults. Let's take a quick look at some the internals of the Autotools class:

$ spack edit --build-system autotools

This will open the AutotoolsPackage file in your text editor.

Note

The examples showing code for these classes is abridged to avoid having long examples. We only show what is relevant to the packager.

.. literalinclude:: _spack_root/lib/spack/spack/build_systems/autotools.py
    :emphasize-lines: 2,4,28-37
    :lines: 138-158,589-617
    :linenos:

Important to note are the highlighted lines. These properties allow the packager to set what build targets and install targets they want for their package. If, for example, we wanted to add as our build target foo then we can append to our build_targets property:

build_targets = ["foo"]

Which is similar to invoking make in our Package

make("foo")

This is useful if we have packages that ignore environment variables and need a command-line argument.

Another thing to take note of is in the configure() method. Here we see that the prefix argument is already included since it is a common pattern amongst packages using Autotools. We then only have to override configure_args(), which will then return it's output to to configure(). Then, configure() will append the common arguments

Packagers also have the option to run autoreconf in case a package needs to update the build system and generate a new configure. Though, for the most part this will be unnecessary.

Let's look at the mpileaks package.py file that we worked on earlier:

$ spack edit mpileaks

Notice that mpileaks is a Package class but uses the Autotools build system. Although this package is acceptable let's make this into an AutotoolsPackage class and simplify it further.

.. literalinclude:: tutorial/examples/Autotools/0.package.py
   :language: python
   :emphasize-lines: 9
   :linenos:

We first inherit from the AutotoolsPackage class.

Although we could keep the install() method, most of it can be handled by the AutotoolsPackage base class. In fact, the only thing that needs to be overridden is configure_args().

.. literalinclude:: tutorial/examples/Autotools/1.package.py
   :language: python
   :emphasize-lines: 25,26,27,28,29,30,31,32
   :linenos:

Since Spack takes care of setting the prefix for us we can exclude that as an argument to configure. Our packages look simpler, and the packager does not need to worry about whether they have properly included configure and make.

This version of the mpileaks package installs the same as the previous, but the AutotoolsPackage class lets us do it with a cleaner looking package file.

Makefile

Packages that utilize Make or a Makefile usually require you to edit a Makefile to set up platform and compiler specific variables. These packages are handled by the Makefile subclass which provides convenience methods to help write these types of packages.

A MakefilePackage class has three phases that can be overridden. These include:

  1. edit()
  2. build()
  3. install()

Packagers then have the ability to control how a Makefile is edited, and what targets to include for the build phase or install phase.

Let's also take a look inside the MakefilePackage class:

$ spack edit --build-system makefile

Take note of the following:

.. literalinclude:: _spack_root/lib/spack/spack/build_systems/makefile.py
   :language: python
   :emphasize-lines: 60,64,69
   :lines: 40-111
   :linenos:

Similar to Autotools, MakefilePackage class has properties that can be set by the packager. We can also override the different methods highlighted.

Let's try to recreate the Bowtie package:

$ spack create -f https://downloads.sourceforge.net/project/bowtie-bio/bowtie/1.2.1.1/bowtie-1.2.1.1-src.zip
==> This looks like a URL for bowtie
==> Found 1 version of bowtie:

1.2.1.1  https://downloads.sourceforge.net/project/bowtie-bio/bowtie/1.2.1.1/bowtie-1.2.1.1-src.zip

==> How many would you like to checksum? (default is 1, q to abort) 1
==> Downloading...
==> Fetching https://downloads.sourceforge.net/project/bowtie-bio/bowtie/1.2.1.1/bowtie-1.2.1.1-src.zip
######################################################################## 100.0%
==> Checksummed 1 version of bowtie
==> This package looks like it uses the makefile build system
==> Created template for bowtie package
==> Created package file: /Users/mamelara/spack/var/spack/repos/builtin/packages/bowtie/package.py

Once the fetching is completed, Spack will open up your text editor in the usual fashion and create a template of a MakefilePackage package.py.

.. literalinclude:: tutorial/examples/Makefile/0.package.py
   :language: python
   :linenos:

Spack was successfully able to detect that Bowtie uses Make. Let's add in the rest of our details for our package:

.. literalinclude:: tutorial/examples/Makefile/1.package.py
   :language: python
   :emphasize-lines: 10,11,13,14,18,20
   :linenos:

As we mentioned earlier, most packages using a Makefile have hard-coded variables that must be edited. These variables are fine if you happen to not care about setup or types of compilers used but Spack is designed to work with any compiler. The MakefilePackage subclass makes it easy to edit these Makefiles by having an edit() method that can be overridden.

Let's take a look at the default Makefile that Bowtie provides. If we look inside, we see that CC and CXX point to our GNU compiler:

$ spack stage bowtie

Note

As usual make sure you have shell support activated with spack:
source /path/to/spack_root/spack/share/spack/setup-env.sh
$ spack cd -s bowtie
$ cd spack-src
$ vim Makefile
CPP = g++ -w
CXX = $(CPP)
CC = gcc
LIBS = $(LDFLAGS) -lz
HEADERS = $(wildcard *.h)

To fix this, we need to use the edit() method to write our custom Makefile.

.. literalinclude:: tutorial/examples/Makefile/2.package.py
   :language: python
   :emphasize-lines: 23,24,25
   :linenos:

Here we use a FileFilter object to edit our Makefile. It takes in a regular expression and then replaces CC and CXX to whatever Spack sets CC and CXX environment variables to. This allows us to build Bowtie with whatever compiler we specify through Spack's spec syntax.

Let's change the build and install phases of our package:

.. literalinclude:: tutorial/examples/Makefile/3.package.py
   :language: python
   :emphasize-lines: 28,29,30,31,32,35,36
   :linenos:

Here demonstrate another strategy that we can use to manipulate our package We can provide command-line arguments to make(). Since Bowtie can use tbb we can either add NO_TBB=1 as a argument to prevent tbb support or we can just invoke make with no arguments.

Bowtie requires our install_target to provide a path to the install directory. We can do this by providing prefix= as a command line argument to make().

Let's look at a couple of other examples and go through them:

$ spack edit esmf

Some packages allow environment variables to be set and will honor them. Packages that use ?= for assignment in their Makefile can be set using environment variables. In our esmf example we set two environment variables in our edit() method:

def edit(self, spec, prefix):
    for var in os.environ:
        if var.startswith('ESMF_'):
            os.environ.pop(var)

    # More code ...

    if self.compiler.name == 'gcc':
        os.environ['ESMF_COMPILER'] = 'gfortran'
    elif self.compiler.name == 'intel':
        os.environ['ESMF_COMPILER'] = 'intel'
    elif self.compiler.name == 'clang':
        os.environ['ESMF_COMPILER'] = 'gfortranclang'
    elif self.compiler.name == 'nag':
        os.environ['ESMF_COMPILER'] = 'nag'
    elif self.compiler.name == 'pgi':
        os.environ['ESMF_COMPILER'] = 'pgi'
    else:
        msg  = "The compiler you are building with, "
        msg += "'{0}', is not supported by ESMF."
        raise InstallError(msg.format(self.compiler.name))

As you may have noticed, we didn't really write anything to the Makefile but rather we set environment variables that will override variables set in the Makefile.

Some packages include a configuration file that sets certain compiler variables, platform specific variables, and the location of dependencies or libraries. If the file is simple and only requires a couple of changes, we can overwrite those entries with our FileFilter object. If the configuration involves complex changes, we can write a new configuration file from scratch.

Let's look at an example of this in the elk package:

$ spack edit elk
def edit(self, spec, prefix):
    # Dictionary of configuration options
    config = {
        'MAKE': 'make',
        'AR':   'ar'
    }

    # Compiler-specific flags
    flags = ''
    if self.compiler.name == 'intel':
        flags = '-O3 -ip -unroll -no-prec-div'
    elif self.compiler.name == 'gcc':
        flags = '-O3 -ffast-math -funroll-loops'
    elif self.compiler.name == 'pgi':
        flags = '-O3 -lpthread'
    elif self.compiler.name == 'g95':
        flags = '-O3 -fno-second-underscore'
    elif self.compiler.name == 'nag':
        flags = '-O4 -kind=byte -dusty -dcfuns'
    elif self.compiler.name == 'xl':
        flags = '-O3'
    config['F90_OPTS'] = flags
    config['F77_OPTS'] = flags

    # BLAS/LAPACK support
    # Note: BLAS/LAPACK must be compiled with OpenMP support
    # if the +openmp variant is chosen
    blas = 'blas.a'
    lapack = 'lapack.a'
    if '+blas' in spec:
        blas = spec['blas'].libs.joined()
    if '+lapack' in spec:
        lapack = spec['lapack'].libs.joined()
    # lapack must come before blas
    config['LIB_LPK'] = ' '.join([lapack, blas])

    # FFT support
    if '+fft' in spec:
        config['LIB_FFT'] = join_path(spec['fftw'].prefix.lib,
                                    'libfftw3.so')
        config['SRC_FFT'] = 'zfftifc_fftw.f90'
    else:
        config['LIB_FFT'] = 'fftlib.a'
        config['SRC_FFT'] = 'zfftifc.f90'

    # MPI support
    if '+mpi' in spec:
        config['F90'] = spec['mpi'].mpifc
        config['F77'] = spec['mpi'].mpif77
    else:
        config['F90'] = spack_fc
        config['F77'] = spack_f77
        config['SRC_MPI'] = 'mpi_stub.f90'

    # OpenMP support
    if '+openmp' in spec:
        config['F90_OPTS'] += ' ' + self.compiler.openmp_flag
        config['F77_OPTS'] += ' ' + self.compiler.openmp_flag
    else:
        config['SRC_OMP'] = 'omp_stub.f90'

    # Libxc support
    if '+libxc' in spec:
        config['LIB_libxc'] = ' '.join([
            join_path(spec['libxc'].prefix.lib, 'libxcf90.so'),
            join_path(spec['libxc'].prefix.lib, 'libxc.so')
        ])
        config['SRC_libxc'] = ' '.join([
            'libxc_funcs.f90',
            'libxc.f90',
            'libxcifc.f90'
        ])
    else:
        config['SRC_libxc'] = 'libxcifc_stub.f90'

    # Write configuration options to include file
    with open('make.inc', 'w') as inc:
        for key in config:
            inc.write('{0} = {1}\n'.format(key, config[key]))

config is just a dictionary that we can add key-value pairs to. By the end of the edit() method we write the contents of our dictionary to make.inc.

CMake

CMake is another common build system that has been gaining popularity. It works in a similar manner to Autotools but with differences in variable names, the number of configuration options available, and the handling of shared libraries. Typical build incantations look like this:

def install(self, spec, prefix):
    cmake("-DCMAKE_INSTALL_PREFIX:PATH=/path/to/install_dir ..")
    make()
    make("install")

As you can see from the example above, it's very similar to invoking configure and make in an Autotools build system. However, the variable names and options differ. Most options in CMake are prefixed with a '-D' flag to indicate a configuration setting.

In the CMakePackage class we can override the following phases:

  1. cmake()
  2. build()
  3. install()

The CMakePackage class also provides sensible defaults so we only need to override cmake_args().

Let's look at these defaults in the CMakePackage class in the _std_args() method:

$ spack edit --build-system cmake
.. literalinclude:: _spack_root/lib/spack/spack/build_systems/cmake.py
   :language: python
   :lines: 167-300
   :emphasize-lines: 87,96
   :linenos:

Some CMake packages use different generators. Spack is able to support Unix-Makefile generators as well as Ninja generators.

If no generator is specified Spack will default to Unix Makefiles.

Next we setup the build type. In CMake you can specify the build type that you want. Options include:

  1. empty
  2. Debug
  3. Release
  4. RelWithDebInfo
  5. MinSizeRel

With these options you can specify whether you want your executable to have the debug version only, release version or the release with debug information. Release executables tend to be more optimized than Debug. In Spack, we set the default as Release unless otherwise specified through a variant.

Spack then automatically sets up the -DCMAKE_INSTALL_PREFIX path, appends the build type (RelWithDebInfo default), and then specifies a verbose Makefile.

Next we add the rpaths to -DCMAKE_INSTALL_RPATH:STRING.

Finally we add to -DCMAKE_PREFIX_PATH:STRING the locations of all our dependencies so that CMake can find them.

In the end our cmake line will look like this (example is xrootd):

$ cmake $HOME/spack/var/spack/stage/xrootd-4.6.0-4ydm74kbrp4xmcgda5upn33co5pwddyk/xrootd-4.6.0 -G Unix Makefiles -DCMAKE_INSTALL_PREFIX:PATH=$HOME/spack/opt/spack/darwin-sierra-x86_64/clang-9.0.0-apple/xrootd-4.6.0-4ydm74kbrp4xmcgda5upn33co5pwddyk -DCMAKE_BUILD_TYPE:STRING=RelWithDebInfo -DCMAKE_VERBOSE_MAKEFILE:BOOL=ON -DCMAKE_FIND_FRAMEWORK:STRING=LAST -DCMAKE_INSTALL_RPATH_USE_LINK_PATH:BOOL=FALSE -DCMAKE_INSTALL_RPATH:STRING=$HOME/spack/opt/spack/darwin-sierra-x86_64/clang-9.0.0-apple/xrootd-4.6.0-4ydm74kbrp4xmcgda5upn33co5pwddyk/lib:$HOME/spack/opt/spack/darwin-sierra-x86_64/clang-9.0.0-apple/xrootd-4.6.0-4ydm74kbrp4xmcgda5upn33co5pwddyk/lib64 -DCMAKE_PREFIX_PATH:STRING=$HOME/spack/opt/spack/darwin-sierra-x86_64/clang-9.0.0-apple/cmake-3.9.4-hally3vnbzydiwl3skxcxcbzsscaasx5

We can see now how CMake takes care of a lot of the boilerplate code that would have to be otherwise typed in.

Let's try to recreate callpath:

$ spack create -f https://github.com/llnl/callpath/archive/v1.0.3.tar.gz
==> This looks like a URL for callpath
==> Found 4 versions of callpath:

1.0.3  https://github.com/LLNL/callpath/archive/v1.0.3.tar.gz
1.0.2  https://github.com/LLNL/callpath/archive/v1.0.2.tar.gz
1.0.1  https://github.com/LLNL/callpath/archive/v1.0.1.tar.gz
1.0    https://github.com/LLNL/callpath/archive/v1.0.tar.gz

==> How many would you like to checksum? (default is 1, q to abort) 1
==> Downloading...
==> Fetching https://github.com/LLNL/callpath/archive/v1.0.3.tar.gz
######################################################################## 100.0%
==> Checksummed 1 version of callpath
==> This package looks like it uses the cmake build system
==> Created template for callpath package
==> Created package file: /Users/mamelara/spack/var/spack/repos/builtin/packages/callpath/package.py

which then produces the following template:

.. literalinclude:: tutorial/examples/Cmake/0.package.py
   :language: python
   :linenos:

Again we fill in the details:

.. literalinclude:: tutorial/examples/Cmake/1.package.py
   :language: python
   :linenos:
   :emphasize-lines: 9,13,14,18,19,20,21,22,23

As mentioned earlier, Spack will use sensible defaults to prevent repeated code and to make writing CMake package files simpler.

In callpath, we want to add options to CALLPATH_WALKER as well as add compiler flags. We add the following options like so:

.. literalinclude:: tutorial/examples/Cmake/2.package.py
   :language: python
   :linenos:
   :emphasize-lines: 26,30,31

Now we can control our build options using cmake_args(). If defaults are sufficient enough for the package, we can leave this method out.

CMakePackage classes allow for control of other features in the build system. For example, you can specify the path to the "out of source" build directory and also point to the root of the CMakeLists.txt file if it is placed in a non-standard location.

A good example of a package that has its CMakeLists.txt file located at a different location is found in spades.

$ spack edit spades
root_cmakelists_dir = "src"

Here root_cmakelists_dir will tell Spack where to find the location of CMakeLists.txt. In this example, it is located a directory level below in the src directory.

Some CMake packages also require the install phase to be overridden. For example, let's take a look at sniffles.

$ spack edit sniffles

In the install() method, we have to manually install our targets so we override the install() method to do it for us:

# the build process doesn't actually install anything, do it by hand
def install(self, spec, prefix):
    mkdir(prefix.bin)
    src = "bin/sniffles-core-{0}".format(spec.version.dotted)
    binaries = ['sniffles', 'sniffles-debug']
    for b in binaries:
        install(join_path(src, b), join_path(prefix.bin, b))

PythonPackage

Python extensions and modules are built differently from source than most applications. These modules are usually installed using the following line:

$ pip install .

We can write package files for Python packages using the Package class, but the class brings with it a lot of methods that are useless for Python packages. Instead, Spack has a PythonPackage subclass that allows packagers of Python modules to be able to invoke pip.

We will write a package file for Pandas:

$ spack create -f https://pypi.io/packages/source/p/pandas/pandas-0.19.0.tar.gz
==> This looks like a URL for pandas
==> Warning: Spack was unable to fetch url list due to a certificate verification problem. You can try running spack -k, which will not check SSL certificates. Use this at your own risk.
==> Found 1 version of pandas:

0.19.0  https://pypi.io/packages/source/p/pandas/pandas-0.19.0.tar.gz

==> How many would you like to checksum? (default is 1, q to abort) 1
==> Downloading...
==> Fetching https://pypi.io/packages/source/p/pandas/pandas-0.19.0.tar.gz
######################################################################## 100.0%
==> Checksummed 1 version of pandas
==> This package looks like it uses the python build system
==> Changing package name from pandas to py-pandas
==> Created template for py-pandas package
==> Created package file: /Users/mamelara/spack/var/spack/repos/builtin/packages/py-pandas/package.py

And we are left with the following template:

.. literalinclude:: tutorial/examples/PyPackage/0.package.py
   :language: python
   :linenos:

As you can see this is not any different than any package template that we have written. We have the choice of providing build options or using the sensible defaults

Luckily for us, there is no need to provide build args.

Next we need to find the dependencies of a package. Dependencies are usually listed in setup.py. You can find the dependencies by searching for install_requires keyword in that file. Here it is for Pandas:

# ... code
if sys.version_info[0] >= 3:

setuptools_kwargs = {
                     'zip_safe': False,
                     'install_requires': ['python-dateutil >= 2',
                                          'pytz >= 2011k',
                                          'numpy >= %s' % min_numpy_ver],
                     'setup_requires': ['numpy >= %s' % min_numpy_ver],
                     }
if not _have_setuptools:
    sys.exit("need setuptools/distribute for Py3k"
             "\n$ pip install distribute")

# ... more code

You can find a more comprehensive list at the Pandas documentation.

By reading the documentation and setup.py we found that Pandas depends on python-dateutil, pytz, and numpy, numexpr, and finally bottleneck.

Here is the completed Pandas script:

.. literalinclude:: tutorial/examples/PyPackage/1.package.py
   :language: python
   :linenos:

It is quite important to declare all the dependencies of a Python package. Spack can "activate" Python packages to prevent the user from having to load each dependency module explicitly. If a dependency is missed, Spack will be unable to properly activate the package and it will cause an issue. To learn more about extensions go to spack extensions.

From this example, you can see that building Python modules is made easy through the PythonPackage class.

Other Build Systems

Although we won't get in depth with any of the other build systems that Spack supports, it is worth mentioning that Spack does provide subclasses for the following build systems:

  1. IntelPackage
  2. SconsPackage
  3. WafPackage
  4. RPackage
  5. PerlPackage
  6. QMakePackage

Each of these classes have their own abstractions to help assist in writing package files. For whatever doesn't fit nicely into the other build-systems, you can use the Package class.

Hopefully by now you can see how we aim to make packaging simple and robust through these classes. If you want to learn more about these build systems, check out Implementing the installation procedure in the Packaging Guide.