Releasing (or not) 32-bit Windows wheels

Hi all,

I meant to write this message a while ago already; now got triggered by a question in response to the SciPy 1.8.1 release announcement and some discussion that followed that on the SciPy issue tracker.

There’s a question about whether we should still support 32-bit Python wheels for Windows. Or more importantly: we already don’t anymore for SciPy at least on Python 3.10, so the question is more: can we keep that status quo, and if that’s okay then will other projects drop 32-bit wheels soon too (saves on packaging & release effort)?

Context:

  • When Python 3.10 was released, Azure DevOps - which is where we build wheels for releases and nightlies - did not provide 32-bit Python in its images.
  • As a result the first NumPy and SciPy releases that supported Python 3.10 did not have 32-bit Python wheels for 3.10 (building from source was, and still is, possible of course). The same is probably true for other projects.
  • There weren’t many (or any?) complaints. However, later Azure did make 32-bit Python available, and therefore @charris updated the NumPy wheels later. 1.21.6 is the lowest version number with wheels; 1.22.x releases have them too.
  • Note that there’s actually no way to distinguish between 32-bit and 64-bit Python on 64-bit Windows (see No 1.21.4 wheel available for cp310-win32 · Issue #45 · scipy/oldest-supported-numpy · GitHub for details), which is potentially an issue - users will get builds from source.
  • It’s actually a little surprising that we haven’t had more complaints, because building from source on Windows if users do pip install scipy will fail for the vast majority of users, unless they are prepared for it and have installed the right compilers and BLAS/LAPACK. The explanation is probably that Python 3.10 adoption is still low, and those adopters are unlikely to run a 32-bit Python; that’s more for the laggards.
  • So the current status for SciPy is: there are no 32-bit wheels for Python 3.10 in any of the releases that support Python 3.10 (= scipy 1.7.2 - 1.8.1).
  • For SciPy 1.9.0 we are switching to Meson as the default build system, with the numpy.distutils based build still available as a backup. Currently there is no compiler toolchain which works with the Meson build and allows us to build 32-bit wheels for distribution.
    • It is possible to build 32-bit wheels from source, given a suitable compiler toolchain. That exists: MSVC + Intel Fortran. We just cannot redistribute those wheels.
    • 32-bit Windows wheels are available at Archived: Python Extension Packages for Windows - Christoph Gohlke
    • The free Anaconda distribution is also shipped as 32-bit and 64-bit for Windows, and the defaults channel supports 32-bit Windows (conda-forge does not).
    • The compiler toolchain we use is a Mingw-w64 based one, used with some flags to ensure that long double is 64-bits (compatible with MSVC) rather than the Mingw default of 80-bits. See e.g. scipy/windows.yml at 743c283bbe79473a03ca2eddaa537661846d8a19 · scipy/scipy · GitHub for the compiler we’re using in CI.
    • Carl Kleffner, who is our expert for this compiler toolchain, opened https://github.com/scipy/scipy/issues/16286 to gauge interest in whether it’s worth to invest effort in updating the Mingw-w64 compiler toolchain to support 32-bit Windows. If you have a use case, please describe it on that issue (in addition to replies you may send here).
    • As for why we don’t have a compiler toolchain: basically MSVC + gfortran is what we used via a hacky (if surprisingly robust) solution in numpy.distutils. We did not attempt to replicate that hack inside Meson, because (a) it’d be a lot of work, (b) the Meson devs don’t really like hacks, and (c) I don’t like hacks. See https://pav.iki.fi/blog/2017-10-08/pywingfortran.html#building-python-wheels-with-fortran-for-windows for details on what the hack entailed.
  • It is also possible that LFortran may mature, and then we have a full Clang-based toolchain which would solve the problem. That will take a while though, and it may not happen at all (it didn’t with Flang, and 32-bit support is probably not a priority for LFortran).

So that’s where we are right now. Usually we use download stats to help inform us with such decisions, so here they are: https://pypistats.org/packages/numpy has downloads per OS, but doesn’t distinguish between 32-bit and 64-bit Python on Windows. Downloads are dominated by Linux (86%) and the total for Windows is 3%, so 32-bit Python on Windows is probably well below 1%. That’s not super informative though.

Longer-term, 32-bit Python is on the way out it looks like - but it could take a long time to disappear. Microsoft was quite slow to provide a 32-bit build of Python 3.10 in Azure DevOps, but it materialized in the end (side note: we could have gotten it through Nuget, I learned later - link). At some point Windows on ARM will probably become a thing that we have to support too (unclear to me when though).
EDIT: Matti started a conversation on the Python packaging Discourse here, and from that it turns out that 32-bit packages are actually useful for Windows on ARM users (that was the main reason people were inquiring about 32-bit binaries).

My personal opinion:

  • It’d be quite nice to have a suitable compiler toolchain, because then we wouldn’t have to have this discussion.
  • It’s a lot of work to create and maintain though. So if there’s not much demand, that effort could perhaps be spent better on something else. And either way, Carl is the only person around in our community who has both the knowledge and (potentially) interest to work on this. If he would decide not to spend time on this, I’d be completely supportive of that choice.
  • I’m personally not really interested to even propose to the Meson devs to replicate the numpy.distutils hack. One thing I like about Meson’s philosophy is doing things “the right way” - that’s going to help us in the long run - and glueing together technically incompatible compilers just doesn’t seem like the right thing to do here. Plus this’d be a lot of work too.
  • It is possible to build SciPy from source, and there are alternative binaries (both wheels and conda packages) as noted above. It doesn’t seem unfair to say that 32-bit Python is niche, and users should go to a third-party vendor for binaries.
    • One practical issue with that is that the number of users complaining may increase after the 1.9.0 release, because we won’t have 32-bit Python wheels for Python 3.8 and 3.9 anymore, so pip install scipy is going to fail more often. We’ll see soon enough.
  • It would be nice to keep one CI job around that tests 32-bit Python on Windows, even if we don’t release wheels. It’s actually unlikely to break badly if 64-bit Windows and 32-bit Linux jobs pass, and we don’t test many other platforms in CI either, so it’s not a must-have, but still.

Thoughts? Use cases / reasons for not being able to drop 32-bit Python?

Cheers,
Ralf

Philosophical musing: how many person-years would it take to get rid of Fortran in SciPy completely? That would be a better outcome; the most work but it’d get rid of a ton of outdated and badly written code that no one wants to maintain. Many birds with one stone …

2 Likes

The 32-windows part is not so much interesting to me since Microsoft is dropping the ball in their own typical way (not telling anyone about it but slowly letting it rot). So I think the matter is a “when” rather than “if”. I don’t see any problems dropping earlier than MS legitimizes the intention.

But! if we are serious about this part

we can start looking into this with eigen/BLIS/(insert your favorite) framework to get the bulk out very quickly (which is essentially linalg and sparse) and and look for grants to replicate the other less critical dependencies in convenient languages.

To be honest, most of those old Fortran libs have been replaced with comparable “modern” code. I know the risks of sidetracking the discussion when I use the word “Rust”. So I won’t do that until we absolutely decide to go in this direction.

1 Like

We are working very hard to get LFortran to compile all Fortran codes, including SciPy’s Fortran. We are always looking for more contributors, if anybody is interested to help out, please let us know!

Besides the default LLVM backend in LFortran, we also have a C++ backend that generates readable C++ code for any Fortran code. Once LFortran can compile all of SciPy using the LLVM backend, it would not be hard to ensure the C++ backend can also compile it, thus if you wanted to migrate away from Fortran, you could simply use the C++ backend in LFortran to translate all of the Fortran code to an equivalent C++.

Regarding 32bit, we currently test our LLVM backend with 64bit Linux, macOS and Windows. (We have a very preliminary prototype direct x86 backend that is 32 bit but we are first focusing on LLVM.) If there is significant interest in 32bit output, we can have a look how to support it (whether via LLVM or separate backend).

3 Likes

As an evening project, I started a straight-forward, procedural port of PROPACK to C++ because of the frustrations of integrating old Fortran code into SciPy. It’s not at all functional (I kind of forgot about it :slight_smile: ), but as far as time estimates go, it takes me an evening to get a few hundred lines ported from Fortran → C++ (and find this is similar for porting for MATLAB scripts to Python)

1 Like

Thanks for the detailed write-up, Ralf.

This is becoming somewhat of a niche platform now, and as long as users have recourse to get the packages installed (which it sounds like they do), I think it’s OK. We have to pick our battles.

I’d be hesitant: we use several well-tested Fortran packages, and those numerical codes are tricky. Sure, they may not be the best quality, but by now we know they work. It also does not really help to port this to C++ (Fortran is likely easier to parse for most contributors?). If we do decide to go this route, we’d need a more thorough set of unit tests, and those are hard to construct after-the-fact. And, who knows, perhaps with LFortran our problems will eventually disappear!

2 Likes

It’s just a small piece of the whole, but I’d be happy to follow up my ongoing work on hyp2f1 by trying to replace all of the Fortran in special with clean, documented, bug-free Cython. Also writing thorough tests like I did for hyp2f1. If it was possible for me to work on it full time, I think I could power through in about a month. Realistically with my schedule, I think I could get through everything in about two years. At the moment I don’t have any spare bandwidth, but starting in August I should be able to put in about 2 hours per week.

1 Like

Porting from a standard, well-defined language to Cython, which is only supporting by a niche tool without a lot of developer bandwidth, seems like a step in the wrong direction. My argument above was that “bug-free” is a very hard thing to guarantee. The most robust solution may be Ondrej’s suggestion: automatically generate C++ code, and then tidy it up and write more test cases.

That assumes that the original code is bug-free in Fortran too which is often not and requires intricate “old” Fortran knowledge. Cython at this point for SciPy is quite established. So not that esoteric at all and certainly more maintainer-friendly. Also Fortran code often forces f2py conversion and quite a lot of copies of arrays that hinders performance. As long as the code is clear in functionality it is much more readable and fixable, say an array is overflowing, finding that in Fortran is orders of magnitude to figure out than a Cython code which is much more debuggable even though you might not know what the code is doing.

But the part that is worrying is the new release is not visible in the horizon and many Cython features and bugs are fixed in the 3.0alpha which we cannot switch to.

2 Likes

ISTM “we do not provide these prebuilt binaries, here are third-party alternatives” is a prefectly suitable answer.

Re: fortran. We can chisel bits and pieces, and in fact we have several almost complete replacements . E.g. much of quadpack can be replaced by quad_vec. So a nice project could be to look at what is not available and try estimating the effort to port the missing bits. That would at least help extrapolating the total effort :-).

That said, I actually find clean fortran code often easier to deal with than almost anything else. The key here is “clean”, which many fortran’66 ports are not quite. So if lfortan support does materialize, maybe we should think about improving the glue (f2py or Cython or whatnot) instead.

I agree for Fortran code that is relatively clean and without clear issues. special is its own beast though. Much of the code in specfun.f is clearly of very low quality. The Fortran implementation of hyp2f1 had many outstanding issues which were never fixed due to the difficulty of working with and reviewing the dense Fortran code. With the Cython implementation, I was able to clean up issues that were open for nearly 5 years. The other hypergeometric functions also have many problems and I plan to replace these as well. I haven’t looked into anything beyond hypergeometric functions though. Perhaps many of these are already OK.

Please compare the new Cython implementation with the old Fortran implementation before judging whether this was a step in the wrong direction. See also the suite of benchmarks that I wrote at special/_precompute/hyp2f1_data.py (unfortunately I’m only able to post 2 links).

Automatic code generation seems like a great idea for old Fortran which is already at production grade, but this isn’t true for much of specfun.f. Perhaps no one can promise perfectly bug-free code, but I can almost guarantee that the new code will be an improvement for these particular cases.

The impetus for rewriting here wasn’t to remove Fortran for the sake of it, but because much of specfun.f is a bug riddled mess which probably shouldn’t be trusted in production code. When I initially inquired about fixing hyp2f1, I was asked to use Cython for ease of review and maintenance. If there were consensus around using something like C++ instead, I think I could handle that, but it would likely slow down the pace of development.

1 Like

Thanks for the data points @mckib2 and @steppi, very interesting. And thanks @certik for weighing in. I’m following LFortran development from a distance, with great interest!

This is indeed a key part, and I’m afraid it wouldn’t be that quick. Redoing BLAS/LAPACK support in particular is challenging, and we’d need to keep supporting cython_blas and cython_lapack as a service to the rest of the ecosystem. That said, while not “quick”, I do think it’s a feasible task in principle.

They don’t work (well), unfortunately. @steppi’s qualification of specfun applies to many other vendored libraries, like ARPACK, FITPACK, QUADPACK, interpolative, and so on. They’re not well-tested libraries with a few bugs - many of them deserve to be thrown away and completely replaced; they have segfaults and correctness issues that we just aren’t able to address, in addition to the impossibility to extend them with new code in practice.

Looking at experiences over the past few years, this is clearly not true:

  • Fortran: few hard bugs get fixed, and zero new code gets written. The only significant addition of Fortran code in the past 5 (or more?) years was PROPACK I think, and as @mckib2 describes it was so frustrating that he started a port to C++. We typically also don’t have more than 1-2 maintainers who even want to review Fortran PRs.
  • Several newer maintainers and contributors are quite enthusiastic about C++. New features or rewrites of code focusing on performance do happen (e.g., scipy.spatial.cKDTree, scipy.fft, scipy.spatial.distance, scipy.special.logit/expit, earlier also the scipy.sparse matrix data structures). Some of the most significant new functionality we added recently is based on Boost and HiGHS, both high-quality C++ libraries. And it’s a lot easier to find new folks with C++ skills willing to work on high-performance numerical Python libraries than it is to find folks with Fortran skills. Finally, for existing maintainers who don’t know a language, learning C++ potentially makes sense from a career perspective. Learning how to deal with old Fortran code, not so much.

C++ has its issues and can be complex, but the reality is that we attract new talented maintainers that want to use it. For Fortran, it’s zero. And those folks that are enthusiastic about Fortran are talking about modern Fortran (!= F77/F90), which can indeed be nice - but only has a good story for HPC / on Linux. Fortran on Windows is just never-ending pain, and responsible for our worst packaging issues. It was also the worst problem for getting things to work on macOS M1. In terms of negative externalities, it is also a problem for Pyodide for example, while something like Cython or Pythran isn’t even though those tools are more niche overall (because they are transpilers, they basically work wherever C/C++ work).

There’s a whole bunch of reasons pro/con for any other language too:

  • C is most portable and simple to integrate, but limited because of no templating, and not many people write new code in it,
  • C++ is most popular with people who like writing native code and feature-full, but harder to understand than C
  • Cython is the most approachable to write new code in for the largest number of maintainers and is nice for binding generation too, but is a pain for build system integration, creates binaries that are too large, and there are long-term maintenance worries because it relies on 1-2 maintainers only,
  • Modern Fortran: nice language for array-based algorithms and fast, but no good support for interfacing with Python, still niche, lack of compilers, and we lack maintainers/reviewers,

For Fortran as we have it in SciPy though (F77 mostly), there’s just no pros at all beyond “we already have the code”, and many cons.

Yes, that is a problem. It would depend on the component whether a line-by-line translation (or auto-translation) would make sense, or a rewrite from scratch.

Yes, that would be nice if anyone is looking for a potentially high-impact project :slight_smile:

2 Likes

Thanks! That is 3 votes for “yeah no 32-bit wheels is okay” for now, and no dissenters. So we can get back to the more fun part, Fortran :wink:

(or auto-translation) would make sense

f2c, anyone? :-).
(Shudders)

I can see how it would be easier to find C++ maintainers. Fortran is, per definition, a simpler language. and the readability of C++ can vary dramatically, depending on how it is written / what features get used. Octave, e.g., is a good example of clean C++ that is relatively easy to parse; but when done wrong C++ can be a nightmare. So, its use requires appropriate restraint.

Cython is approachable for simple things, but can get hairy as complexity increases—without the advantage of being able to factor out to a standard, self-contained library. I like Cython, but worry about the longevity of code written in a specialized language.

For hyp2f1, and also hyp1f1, hyperu and hyp0f1 whose Cython rewrites were already initiated by @person142, I think it would be a waste of effort to rewrite yet again in a new language but I would be willing to have a decision made for me on what language to use for newly initiated specfun rewrites. I’d be comfortable with either of C, C++ or Cython. I think @person142 should probably have most say in the choice of language since he would be the one tasked with reviewing the PRs.

Old Fortran should be modernized. As an example, here is a modernized minpack, to which 6 people contributed so far, and an issue to make SciPy use it:

It looks like we would be able to maintain this library for you, including the Python wrappers, and SciPy could just use it. I know you commented on that issue also.

So it is definitely possible to organize the community to maintain a Fortran library.

I agree with this, and that is one of the motivations for LFortran. We are working hard to finish most semantics and to compile codes like SciPy. We already support macOS (including M1, my main development machine, everything works) and Windows (you can link things with MSVC), as well as we compile to WebAssembly, so Pyodide could use it; we also support transpiling to C++, if people wanted, and so that you are not locked into Fortran. I will be sure to let you know once we can compile SciPy, or at least parts of it, and we can investigate more.

Just wanted to let people know, that while we do not have a production solution that you can use today, we have a very solid plan how to fix all these issues (I think), and are well on the way of implementing it.

1 Like

Thanks, that is a good point. minpack is a nice test case I guess, it certainly would be a nice upgrade for scipy.optimize. I’m not sure if that’s what actually modern Fortran looks like - I had expected not, because there’s still a lot of goto's in the code (which is bad).

Sounds promising - thanks for pushing hard on that :raised_hands:

You’re right, rewriting again doesn’t make sense, that code looks good now. I think our preference of languages hasn’t really changed. @stefanv does have a point about Cython becoming hairy when the code gets more complex.

I think there’s some personal judgement that’s fine to make here. Assuming you are coming from a “more Python-like is easier”, then:

  • The simplest use cases: Pythran
  • If Pythran isn’t enough: Cython
  • If Cython code grows more complex and starts to look a lot like C code: then using C or C++ is better
  • If the code is simple(-ish) or can use NumPy’s C API for ufuncs for example: C
  • If you need to support multiple dtypes and start to use macros or homegrown templating to do so: C++ (with templates, but otherwise relatively simple) is preferred

Thanks! These are good rules of thumb.

Good point, I just created an issue to fix this: Remove goto · Issue #74 · fortran-lang/minpack · GitHub. We took the old F77 style minpack, and carefully modernized it step by step, without breaking functionality. We still have to replace all goto. Besides that, if you see anything else you don’t like, please let us know!

2 Likes

Just a heads up that the latest main of Minpack now does not contain any goto. If you find anything else that you don’t like, feel free to just open up an issue and we’ll try to fix it.