(Mostly) Pure Python Physics

Continuing my work on the Pure Python port, I did a small amount of profiling with the outstanding runsnakerun to determine what was bogging it down. As expected, the slow speed of Python's math operations -- and especially the matrix and vector operations -- were the major bottleneck.

Having said that, it would appear that there's very little I can do for the Pure Python port to speed it up. If you followed the previous post, you'll know that I tried PyPy (it works after a large number of iterations) and Cython (faster, but still not so great). I considered using some library like NumPy to get its matrix and vector functionality, but including such a massive library just for this tiny physics library just felt wrong.

I stumbled upon quite a few matrix and vector C extensions for Python, notably Casey Duncan's excellent Planar project. However, there were always issues with all of the ones I found: not supporting tuples in place of vector and matrix types, immutable vectors (though this can be a good thing for other projects which don't expect it), and just the general headache of requiring restructuring all of the pypybox2d code that already relies on vector and matrix classes with certain properties and functions. [side-note: were it not for this project, I would definitely be using the Planar library.]

In short, and for better or worse, I re-invented the wheel, creating pure C extensions for the Vec2, Mat22, Transform, and AABB types (should be compatible with Python 2.5~3.2, but let me know your results). Performance significantly improved (sorry, I don't have any benchmarks right now -- but run any of the testbed tests and you'll feel it). Most of the testbed tests run "fast enough". Exceptions include the pyramid and vertical stack ones, which have altogether too many contacts, and the solver gets very slow.

After adding pickling support for the above classes, I then really wanted to give multiprocessing a go. I made a simple test with a set of processes dedicated to doing shape distance calculations (i.e., consumers). I passed in thousands and thousands of shapes and transforms, and it properly generated the results. Unfortunately, comparing it to the single process version, regardless of the number of input sets, the multi-process version was on the order of 10-50x (!) slower -- not even including the time it took to start the processes. All of the overhead, presumably from synchronization and pickling the objects to pass between processes is just too expensive. It looks as if this will not be a viable option for a real-time physics engine.

Now, where to go from here? I'm stumped -- err, I mean -- I'm open to suggestions!

Pure Python physics?

Recently I tried to get pybox2d compiled and running with PyPy, but it failed for whatever reason. I had it in my head to just give up and port the whole bloody thing to pure Python and see how that would go. Not wanting to waste inspiration, I gave that a shot. Fast forwarding to today, and it's pretty much done.

So, what's in the port? (almost) All features of Box2D up to r175:

• All shape types (polygons, circles, edges, and loops)
• All joint types (distance, revolute, friction, prismatic, weld, rope, wheel, mouse, pulley, gear)
• A dynamic tree class for AABBs
• Collisions, contacts, shape distance calculations, etc.
• A mostly Pythonic interface to all of the useful classes
• The testbed, in simplified form

Before you jump for joy, there are many reasons to not use this:

• It's slow.
• It's really slow.
• It has a bad name (pypybox2d -- changing it is on the TODO list)
• It's still early, though almost everything has worked well enough in my tests. I'm positive there are plenty of typos and minor bugs.
• Some source files are much too big, because I haven't gotten around to splitting them up
• It uses __slots__ for my own debugging purposes. I have no qualms about (almost wholly) removing their usage eventually.

"Why do this, then, if it's so slow and unusable?" Do you really have to ask? Why not?

So, what of PyPy after all that hard work? Unfortunately, in my few tests, it's been slower than CPython **. Perhaps one day it will be sped up by another means. Or maybe I'll figure out a way to get Cython working with it such that it won't totally screw everything up. If you have any ideas, please do get in touch with me.

Find the source here. The testbed requires pygame, but there's a (very) simple graphic-less hello.py in examples/ if you simply must run something.

** edit May 20th: But those tests were superficial, only testing for a few iterations. Increase the number of iterations, and PyPy wins, hands-down. See Antonio Cuni's helpful comment below for more information.

Simple examples? oh yeah, and an alpha release for you...

So, not everyone likes the testbed. It's big, C++-ish, and not so pretty. And we all hate searching around in multiple files for function definitions.

Making some simple examples has been on the back-burner since I started on pybox2d a few years back. I've finally started on some examples, but I need your input on how to make them easily digestible for everyone else:

simple_01.py

simple_02.py

Is it worth going down this path?

To hopefully rouse up some interest, I will provide another alpha release. This time, I'll make it available to a larger audience by including builds for the platforms I have access to (sorry, OSX users -- you might not get any installer release for 2.2 at all).

So, read the manual, download from below, and try it out. Without your help, this version will never see daylight, I assure you!

win32 2.6, 2.7, 3.0, 3.1
linux-i686 2.6, 2.7, 3.1

Edit 4/7/2011: OK, so -- not a single comment in a month, and I recently fixed a major leak affecting all versions since 2.0.x, so I just removed those versions from the svn. 2.0.x has also seen its first update in a few years.

2.1 on the move?

test_Car.py

Hi,

So, I've moved the 2.1 branch into the trunk. I've also spent more than half my Saturday fixing and writing documentation.

Here's the 2.1 manual on the wiki! If you have some time, please read through it. It's a lot of work -- hopefully someone will appreciate it. :)

I also re-uploaded the 2.0.2 bugs-n-stuff that were on the Box2D wiki, here. A couple of other wiki entries got updated, too. Notice anything strange? Please do let me know.

There's a few changes upstream that I'll get around to including soon, but I'm a bit burnt out after the above. Nothing huge, though.

Warm wishes,

kne

Found in the wild: Bricks Knock Off

I just happened upon a decently polished little game called Bricks Knock Off (English translation here) that's actually using the 2.1 branch of pybox2d. Cool!

As you might have guessed, it's a game about bricks. Shooting them off some platform to get a decent enough score to continue to the next level. It features music, sound effects, a few levels, and even a scrolling background.

I only have one comment from the pybox2d-end of things, and it's that the position/velocity iterations need to be tweaked to get the bricks to stack stably. If you check it out, change main.py around line 115ish to be something like:

    vel_iters, pos_iters = 15,15 

Progress

Yes, there has been some progress. Only a bit though. I know you love it when I make bullet points, so here you go:

• Most importantly, I fixed up a few things thanks to issues pointed out by users. Thanks mikkelin, Fahri, dabski, and anyone else that I might have forgotten!
• Updated to the latest Box2D SVN r141 -- there were some changes upstream, including:
• A new RopeJoint (maximum distance constraint)
• Allocator, joint and some other fixes and improvements
• Working b2LoopShapes -- featuring smooth collision both inside and out (you can see these in the character collision test)
• Added the rope joint test
• PyQt4 framework fixes
• Some properties are changeable now (*)
• Fixed updated circle radii between calls
• Epydoc documentation for the 2.1 branch was generated by using the Doxygen documentation for the C++ side (which, although it is included in the SVN, is usually not at all necessary for pybox2d users).
• An up-to-date version of the compiled library is included on the SVN
• It's here. For those of you who would like to help test, is this convenient enough? Or is it "give me an archive or test the damn thing yourself"? :)

I think it's almost time that the trunk from early last year to be finally overwritten. I don't think there are too many project requirements with the old library anymore. Thoughts?

(*) Any idea how to properly size a QListWidget in a QTreeWidget cell? I think sizing widgets appropriately is what most frustrates me about Qt.

Not quite what you were hoping for, but...

Recently, I haven't had much time to work on pybox2d. And when I say 'not much', I mean just about zero.

In any case, I know that there are probably a couple people out there that are anxious to try the new version but don't want to go through installing SWIG, compiling, etc. It's a pain, I realize.

The other day Fahri, creator of Mekanimo --- the only commercial product to use pybox2d that I know of -- asked me if I had a binary he could test of the new branch. I threw a simple Python 2.6 (win32) package together, and thought that the rest of you might like to give it a go also.

The library itself is contained in the archive, so there's no need to install it. The redesigned testbed examples are in there, and you can experiment with the new syntax to your heart's content. For any of the tests, simply run: python test_Web.py . You can also specify the (incomplete) pyglet or pyqt4 back-ends with: python test_Web.py --backend pyglet or '--backend pyqt4' for example.

I only ask that if you give it a try, please at least run python run_unit_tests.py and see if you get any errors. Also, please note that this is alpha-quality and does not have much documentation.

Download here.