Moving SciPy to the Meson build system

Let's start with an announcement: SciPy now builds with Meson on Linux, and the full test suite passes!

This is a pretty exciting milestone, and good news for SciPy maintainers and contributors - they can look forward to much faster builds and a more pleasant development experience. So how fast is it? Currently the build takes about 1min 50s (a ~4x improvement) on my 3 year old 12-core Intel CPU (i9-7920X @ 2.90GHz):

Profiling result of a parallel build of SciPy with Meson

Profiling result of a parallel build (12 jobs) of SciPy with Meson. Visualization created with ninjatracing and Perfetto.

As you can see from the tracing results, building a single C++ file (bsr.cxx, which is one of SciPy's sparse matrix formats) takes over 90 seconds. So the 1min 50 sec build time is close to optimal - the only ways to improve it are major surgery on that C++ code, or buying a faster CPU.

Why move to Meson, and why now?

In Python 3.10 distutils will be deprecated and in Python 3.12 it will be removed (see PEP 632). SciPy uses numpy.distutils and setuptools as its build system, so the removal of distutils from the Python standard library would have a major impact on SciPy - and even more on numpy.distutils, which directly extends distutils. That PEP was written almost a year ago, and at the time my first instinct was to wait until setuptools integrated distutils (which the setuptools maintainers still plan to do, with a cleaned up API) and then update numpy.distutils for that change. That would also require moving parts of numpy.distutils into setuptools, like Fortran support (see setuptools/issues/2372 for more details). It has become clear though that this will be a really slow and painful process - after almost a year, the vendored distutils still hasn't been re-enabled inside setuptools.

The other driver was that SciPy development has become more painful over time. SciPy contains C, C++, Fortran and Cython code, exports C and Cython APIs, has complex dependencies (see diagram below) and does a lot of code generation. The growing amount of C++ and Cython code has increased build times. CI systems frequently time out because build plus test suite run takes close to 60 minutes (the limit on several free-for-open-source CI offerings). Working on compiled code is cumbersome and slow, to the point where a couple of other maintainers have told me it's a blocker for them to work on new features. Finally, debugging build issues is quite difficult - distutils doesn't really have a design, so every other extension just monkey patches things as needed. Earlier in the year I spent half a day hunting a build problem through four (!) different projects that all modified how extensions are built.

Those things together made me realize that it's time to move to a better build system - if we are forced to do a lot of work because of PEP 632 anyway, then now is the time. That left the question: which build system? Really there were only two viable candidates: CMake and Meson. After some experimentation, I had a strong preference for Meson for two main reasons: it has great documentation (unlike CMake), and it's easy to contribute to (it is ~25,000 lines of clean Python code, compared to ~1,000,000 lines of C/C++ for CMake). Given that there are very few Python projects using either CMake or Meson, good docs and an easy to understand code base are very important - we're going to have to contribute! CMake has a larger user base, as well as better Python integration than Meson via scikit-build right now. But to me that was less important. Also, scikit-build still depends on setuptools and one of the goals of this exercise is to get away from distutils/setuptools completely and use only a modern build system and Python packaging standards (i.e. pyproject.toml-based builds) to interact with pip and other Python packaging tools.

So in February I wrote an RFC titled "switch to Meson as a build system", and after positive feedback started working on this project.

Diagram of SciPy's build and runtime dependencies

SciPy's external build and runtime dependencies. Vendored dependencies (e.g., SuperLU, ARPACK, large parts of Boost, Uarray, HiGHS, Qhull, PocketFFT, etc.) aren't shown.

Some of the benefits of building with Meson

Let's start with the biggest benefit: Meson is very fast.

$ time meson setup build
...
# Output: ~50 lines of useful configure info

real    0m3.765s
user    0m5.718s
sys     0m4.456s

$ time ninja -C build -j 12
...
# Output: ~25 lines of build output, only relevant compiler warnings
[1311/1311] Linking target scipy/sparse/sparsetools/_sparsetools.cpython-39-x86_64-linux-gnu.so

real    1m48.058s
user    19m58.007s
sys     2m4.746s

Compare this with the numpy.distutils based build:

$ export NPY_NUM_BUILD_JOBS=12  # parallelize over files for a single .so
$ time python setup.py build_ext --inplace -j 12  # parallelize .pyx -> .c/cxx
...
# Output: ~17,000 lines of build output, mostly noise.
#         Relevant warnings are effectively impossible to spot.

real    6m56.364s
user    13m15.175s
sys     0m39.141s

So the Meson build is about 4x faster. The main reason numpy.distutils is slow is that parallelism is limited - there are hard to fix race conditions that prevent running build_ext completely in parallel. Another reason is that it just goes around invoking compilers directly in a fairly ad-hoc fashion, while Meson uses Ninja as a backend. Ninja is about as fast as it gets.

What we can also see is that we now get a clean build log. This is part of the reason it took a while to get to this point - every single compiler warning was either silenced in the meson.build files or fixed in SciPy's master branch.

Better debugging of build issues

One very nice improvement is that you don't actually have to run the build anymore to figure out how an extension is built exactly. To look at compiler flags, include paths, and other information relevant to how a target is built, simple run meson introspect build/ -i --targets to obtain a JSON representation like this:

{
    "name": "_sparsetools.cpython-39-x86_64-linux-gnu",
    "id": "[email protected]@[email protected]",
    "type": "shared module",
    "defined_in": "/home/rgommers/code/bldscipy/scipy/sparse/sparsetools/meson.build",
    "filename": [
        "/home/rgommers/code/bldscipy/build/scipy/sparse/sparsetools/_sparsetools.cpython-39-x86_64-linux-gnu.so"
    ],
    "build_by_default": true,
    "target_sources": [
        {
            "language": "cpp",
            "compiler": [
                "/home/rgommers/anaconda3/envs/scipy-meson/bin/x86_64-conda-linux-gnu-c++"
            ],
            "parameters": [
                "-I/home/rgommers/code/bldscipy/build/scipy/sparse/sparsetools/_sparsetools.cpython-39-x86_64-linux-gnu.so.p",
                "-I/home/rgommers/code/bldscipy/build/scipy/sparse/sparsetools",
                "-I/home/rgommers/code/bldscipy/scipy/sparse/sparsetools",
                "-I/home/rgommers/anaconda3/envs/scipy-meson/lib/python3.9/site-packages/numpy/core/include",
                "-I/home/rgommers/anaconda3/envs/scipy-meson/include/python3.9",
                "-fdiagnostics-color=always",
                "-D_FILE_OFFSET_BITS=64",
                "-Wall",
                "-Winvalid-pch",
                "-Wnon-virtual-dtor",
                "-std=c++14",
                "-O2",
                "-g",
                "-fvisibility-inlines-hidden",
                "-std=c++17",
                "-fmessage-length=0",
                "-march=nocona",
                "-mtune=haswell",
                "-ftree-vectorize",
                "-fPIC",
                "-fstack-protector-strong",
                "-fno-plt",
                "-O2",
                "-ffunction-sections",
                "-pipe",
                "-isystem",
                "/home/rgommers/anaconda3/envs/scipy-meson/include",
                "-DNDEBUG",
                "-D_FORTIFY_SOURCE=2",
                "-O2",
                "-isystem",
                "/home/rgommers/anaconda3/envs/scipy-meson/include",
                "-fPIC",
                "-DNPY_NO_DEPRECATED_API=NPY_1_9_API_VERSION"
            ],
            "sources": [
                "/home/rgommers/code/bldscipy/scipy/sparse/sparsetools/bsr.cxx",
                "/home/rgommers/code/bldscipy/scipy/sparse/sparsetools/csc.cxx",
                "/home/rgommers/code/bldscipy/scipy/sparse/sparsetools/csr.cxx",
                "/home/rgommers/code/bldscipy/scipy/sparse/sparsetools/other.cxx",
                "/home/rgommers/code/bldscipy/scipy/sparse/sparsetools/sparsetools.cxx"
            ],
            "generated_sources": []
        }
    ],
    "extra_files": [],
    "subproject": null,
    "installed": true,
    "install_filename": [
        "/usr/local/lib/python3.9/site-packages/scipy/sparse/_sparsetools.cpython-39-x86_64-linux-gnu.so"
    ]
}

This helps quickly pinpointing mistakes in your meson.build files. In addition, debugging potential issues in Meson itself, the generated ninja.build file is fairly readable too - its syntax is simple enough that missing dependencies can be found easily (e.g., scipy.cluster extensions depend on scipy.linalg.cython_linalg, and that dependency must be declared correctly).

Other benefits

Other significant benefits include:

  1. Cross-compiling will become possible. For years we've told people "sorry, distutils wasn't really made for cross-compiling, let us know if you have any luck". As a result we've completely ignored users on some exotic platforms, and also spent a lot of time fighting with different CI systems to do native builds. For example, we will likely want to cross-compile to aarch64 rather than struggle with underpowered ARM hardware on Travis CI.
  2. Developers can use multiple builds at the same time. Because Meson builds are out-of-tree by design, it is now easy to have for example a GCC build, a Clang build, and a Python debug build in parallel in the same repo.
  3. More development tools work out of the box, e.g. ccache will just be picked up if it's installed, no configuration needed. I also managed to get AddressSanitizer to work with only a few small tweaks.
  4. Build definitions are easier to understand and modify. Not everything is easier, but common tasks like setting a compiler flag depending on some condition (like "the compiler supports this flag") certainly are:
# meson.build
thread_dep = dependency('threads', required : false)
if thread_dep.found()
  pocketfft_threads = '-DPOCKETFFT_PTHREADS'
endif

py3.extension_module('pypocketfft',
  'pypocketfft.cxx',
  cpp_args: pocketfft_threads,
  include_directories: inc_pybind11,
  dependencies: [py3_dep, thread_dep],
  gnu_symbol_visibility: 'hidden',
  install: true,
  subdir: 'scipy/fft/_pocketfft',
)

compared with:

# setup.py
def pre_build_hook(build_ext, ext):
    from scipy._build_utils.compiler_helper import (
        set_cxx_flags_hook, try_add_flag, try_compile, has_flag)
    cc = build_ext._cxx_compiler
    args = ext.extra_compile_args

    set_cxx_flags_hook(build_ext, ext)

    if cc.compiler_type == 'msvc':
        args.append('/EHsc')
    else:
        # Use pthreads if available
        has_pthreads = try_compile(cc, code='#include <pthread.h>\n'
                                   'int main(int argc, char **argv) {}')
        if has_pthreads:
            ext.define_macros.append(('POCKETFFT_PTHREADS', None))
            if has_flag(cc, '-pthread'):
                args.append('-pthread')
                ext.extra_link_args.append('-pthread')
            else:
                raise RuntimeError("Build failed: System has pthreads header "
                                   "but could not compile with -pthread option")

        # Don't export library symbols
        try_add_flag(args, cc, '-fvisibility=hidden')

def configuration(parent_package='', top_path=None):
    ...
    config = Configuration('_pocketfft', parent_package, top_path)
    ext = config.add_extension('pypocketfft',
                               sources=['pypocketfft.cxx'],
                               depends=['pocketfft_hdronly.h'],
                               include_dirs=include_dirs,
                               language='c++')
    ext._pre_build_hook = pre_build_hook

Some key Meson design principles

Meson is a well-designed build system, and both the docs and a number of talks by Meson devs on YouTube do a great job at explaining that design. So I won't try and give a full picture here. However, there are a few things that are particularly important and would have helped me to fully grasp when I started on this project. So let's have a look at those.

Builds must be out-of-tree. Meson will not allow building inside your existing source tree - and that includes code generation. There are good reasons for this, but when coming from distutils it may bite you. For example, if you have a script to generate Cython .pyx and .pxd files (quite common in SciPy), those files cannot be placed next to the template, they must go into a build directory (the one we chose with meson setup <builddir>).

Meson is not Turing-complete, and not extensible via APIs. What this means is that if something is not supported in Meson itself, you cannot just write a bit of Python code to hack around the problem. Instead, you must add it in Meson itself. If that takes too long, just fork Meson and use your fork until your feature is merged upstream. This seems painful, but it guarantees that people don't just copy around changes from project to project and long-term maintainability deteriorates. Instead, the philosophy is to fix things once for all users.

All source files and targets must be listed explicitly. This means that if you build a single Python extension from 50 Fortran files, you must list all 50 files names. This is not really a problem in practice, but it can be a little verbose. Using a few helper snippets to generate file listings in IPython saves time:

>>> for f in sorted(os.listdir('.')):
...     if f.endswith('.f'):
...         print(f"  '{f}',")
...
  'dgamln.f',
  'zabs.f',
  ...

Meson has an imperative DSL with immutable objects. The syntax of the DSL is inspired by Python, and is easy to read. Objects - which can be dependencies, generated files, built extensions, etc. - being immutable makes things easy to debug, but in some cases it can restrict what you can do. For example, this code for building all scipy.sparse.csgraph extensions is quite elegant, but the foreach pattern will not work if the .pyx files are generated. This is because they'd then be objects created by custom_target, and there is no syntax to give them unique names within the foreach loop - this cost me some time to figure out:

pyx_files = [
  ['_flow', '_flow.pyx'],
  ['_matching', '_matching.pyx'],
  ['_min_spanning_tree', '_min_spanning_tree.pyx'],
  ['_reordering', '_reordering.pyx'],
  ['_shortest_path', '_shortest_path.pyx'],
  ['_tools', '_tools.pyx'],
  ['_traversal', '_traversal.pyx'],
]

foreach pyx_file : pyx_files
  py3.extension_module(pyx_file[0],
    pyx_file[1],
    include_directories : inc_np,
    dependencies : py3_dep,
    install : true,
    subdir : 'scipy/sparse/csgraph')
endforeach

Meson uses pkg-config for discovering dependencies. This means certain things "just work". E.g., blas = dependency('openblas') followed by using dependencies: blas to build the scipy.linalg._fblas extension worked on the first try for me - definitely did not expect that.

There is one escape hatch to do things Meson does not natively support: custom_target. As long you list its source files and outputs, you can invoke Python scripts through custom_target. This is how the SciPy build invokes numpy.f2py, Pythran and cython --cplus (Meson's Cython support is brand new, and only supports targeting C if you give a .pyx file to py3.extension_module right now).

The sharp bits

Not everything was smooth sailing. Here are a few of the most important issues I ran into:

  1. Cython + generated sources. This was by far the most time-consuming topic. Cython's cimport mechanism relies on the complete source tree. For example, not having a __init__.py file two directories up from where the .pyx file that you are building lives changes the C code that Cython generates. The full source tree layout matters for Cython. Cython was largely designed around in-tree builds, and Meson only allows you to generate files out-of-tree. This necessitates hacks like writing scripts or invoking cp to copy over many __init__.py and .pxi|.pxd files. And then more hacks to ensure those files are respected as dependencies by Ninja.
  2. SciPy is a tangled mess - there are 17 scipy.xxx submodules and almost all of them depend on each other. So we could only start running tests after the Meson build was close to 100% complete. We have had issues in SciPy with import cycles before, and I'm now surprised we didn't have more issues in the past.
  3. One feature I missed in Meson: installing generated files with py3.install_sources is not allowed. One can specify install: true in custom_target, but it is a bit hacky to make that recognize the correct Python-specific install directory and it's then not possible to generate say 10 files and only install two of those (I ran into this multiple times, typically with Cython again where we need the .pyx and .pxi files at build time, and the .pxd files also at runtime).
  4. You must use --prefix when using meson setup. If you don't, Meson just ignores with which Python interpreter you built and simply installs to the default /usr/local/lib/, asking for elevated permissions if needed. Meson doesn't know it's building a Python package (we've just told it in the project definition that we're building C, C++, Cython and Fortran code) so from the Meson perspective this is normal. It's not a major issue, however I have forgotten to add --prefix=$PWD/installdir often enough that it's a point of friction. The plan is to solve this by writing a dev.py CLI wrapper, similar to SciPy's runtests.py.

Next steps for SciPy and beyond

Right now the Meson support lives in my fork of SciPy. If you'd like to play with it, see rgommers/scipy/MESON_BUILD.md. There's a lot left to do before we can declary victory and make Meson the default build system. The most important topics are:

  • support for macOS (should be easy) and Windows (probably tricky),
  • implementing support for MKL and other BLAS/LAPACK libraries in Meson,
  • making the BLAS/LAPACK library to use user-configurable, and
  • building wheels and sdists.

For more details, see the tracking issue and my proposal on the scipy-dev list to use some SciPy project funds to help accelerate build and CI improvements.

When I started this endeavour, I wasn't yet completely sure using Meson was going to work out. However I did know that if building SciPy was successful, Meson will work for pretty much any other scientific Python project. We have about two years left before distutils disappears, and my hope is that we can retire numpy.distutils over that time frame. Projects which don't have complex build requirements (i.e. just Cython, or maybe a few C extensions) should be fine to simply use setuptools. And those projects which do have complex builds can move to Meson, or to scikit-build if they prefer CMake.

There are a lot of technical reasons to argue for or against using any build system. However, why I'm convinced Meson is a great choice for both SciPy and other projects is: Meson is a pleasure to use. It "feels right". That's a rare thing to say about a build system - I've certainly not heard anyone say this about numpy.distutils or setuptools. Nor about CMake. We have had two other alternative build systems before in NumPy and SciPy: first NumScons (based on Scons) and then Bento (based on Waf). Both were created by David Cournapeau; NumScons in 2008 and Bento in 2011. Bento/Waf was most similar to Meson, and it made working on SciPy so much better that I maintained it until 2018 - but it never made the transition to Python 3. Yes, I stayed with Python 2.7 until 2018 specifically so I could use Bento rather than distutils - nothing in Python 3.x was interesting enough that it was worth dealing with distutils on a daily basis. Unfortunately Waf was not a stable enough project to build on, and Bento was a one-person project. Meson is even better than Waf and is well-maintained, so we can finally have a single nice build system.

This project has consumed a lot of nights and weekends to get it to this point. I wouldn't have gotten this far in five months without some very important contributions though. Dylan Baker, one of the Meson maintainers, implemented Cython support in Meson impressively quickly after my "Treating Cython like a language?" proposal. That was a real life saver - using custom_target for every Cython extension would have been very tedious. My colleague Gagandeep Singh added Meson support for some of the largest SciPy modules (optimize, stats, spatial, interpolate, sparse.linalg). Smit Lunagariya, who started an internship with SciPy earlier this month, added support for the signal and misc submodules, and is fixing build warnings and bugs in SciPy that the Meson build made visible. Jussi Pakkanen (the Meson project leader), offered encouraging advice early on. Thibault Saunier implemented preliminary support for PEP 621 in mesonpep517. Frederik Rietdijk sent me a PR to demonstrate building SciPy with Meson and Nix. And many fellow SciPy maintainers and others gave positive feedback on my proposals to try adopting Meson. In short, it takes a village to switch build systems - thanks to everyone who is participating in this effort!

Comments