icon fullempty.sh

A Few Thoughts on hobbyist Python Game Development


Let's start this off properly: I'm not a professional game developer, but I've been writing unfinished games since childhood, and almost none of them end up in a playable state, let alone anything fun. But, over the years, they've served as a practice for improving my programming skills.

Games have proven to be the hardest challenge for those three needs, as I tend to end up with multi-hundred line main loops and become frustrated and bored. I get intensely excited about an idea for a game, but if I can't strike while the iron's hot, the choices for code paths and game design patterns are hard to maintain focus on.

Over the last couple of months, I made more progress than usual, especially with Python and its various inefficiences. The goal I set for myself was to make a simple SimCity-alike with some weird stuff in it. I did not finish, but I did build a fast, efficient framework with an infinitely growing map.


I initially started with pygame, but as with every other experience I've had with it, the combination of badly performing core features and general bugginess prevented me from going much further. Pygame can be very good for getting started, making very simple toys, and many of the interfaces are easy to understand. But by the time you need to do something more complicated, there are no real intuitive tools for doing so, so you end up rolling your own. One exception: The Rect class. I used the shit out of these, to the point that I reimplemented it in the next iteration of this project.

Once the limits of pygame prevented me from moving further, I settled on pyglet. Pyglet's main advantage over pygame is that it is hardware-accelerated by default, instead of bolted on. in Pygame's OpenGL acceleration is also intensely buggy, and limits the feature set arbitrarily, instead of augmenting it.

Pyglet is less intuitive to use than pygame, but far more powerful if you spend a little more time with the docs and tutorials. You will be furiously searching for additional references the whole time, but at a certain point the entire architecture of pyglet clicks. Pyglet's sprite class is incredibly powerful, and for a time I did not think so, until I realized it was my own code working against it.

For the pygame version of the game, once I started maxing out time spent within frame loops, I had to resort to implementing a quadtree from scratch, mostly because I wanted the challenge, and because I wanted to tightly couple some features specific to this game. Once I switched to Pyglet, it became unnecessary for performance and I went back to a simple multidimensional dict, indexing sections of the map using keys formed from coordinates modulused by an arbitrary chunk size, as it was simpler to maintain and far faster than what I'd done previously.

Getting somewhere

Pyglet provides a fairly shallow abstraction of OpenGL's interface, making it far less stateful, and allowing you to use more object-oriented principles while working on your project. However, I ended up using some of pyglet's OpenGL bindings directly, which I don't think is unexpected. Pyglet also makes use of a batch system, wherein you attach individual sprites to a given batch object, and draw that batch, which is a common, efficient pattern and present in MonoGame as well.

Shoving thousands of sprites into a single batch did not scale well, but simply having a 1:1 batch-to-chunk design proved to be incredibly fast. For a while I thought it was not performing well, and ended up reimplementing a lot of my code as Python C API classes, before realizing my problem was a single stray, expensive loop within in my main game loop. Ooooops

C classes were a fun digression. I'd never written Python classes in C from scratch, and I kept their concerns focused. In most cases I simply translated the Python version to C. I left the classes in Pythin in places where I would have had to do complex memory management. Once exception: I adapted a Perlin noise implementation in C, because it is entirely cpu bound code that benefits greatly from being compiled. I use this for generating tesselated backgrounds for each chunk.

I also have some sprite types (which I subclass as Cell) that use a spritesheet to auto-tesselate. I considered this necessary because in the original SimCity, this is how roads and rails appear when you add them to your city. These are somewhat hard to do right. It is fine detail work, involving both your sprite sheet and your code. I ended up with a slightly overcomplicated system of mask-like structures and texture name suffixes. I am not sure I'd do it this way again.

Shelving it

I made reasonable progress before I became overwhelmed by both the game and code design decisions. I did not finish it, but I largely finished what I set out to do. I had a functioning simulation loop, even if it was very simple. I have a very efficient framework for adding any cell type to the game state, and a great deal of fast, reusable code. I have uploaded the entire works to Github, named Birch, because I started with simple tree textures to test game features.

My next goal is to rework any tightly coupled game features and genericize the project, so I can use the framework for some kind of roguelike, maybe with city-building features not unlike Rimworld or even strategy games such as Age of Empires. Having an efficient game state that can handle very large maps without transitions is a great tool to have, one I shouldn't have to write again, even if I probably will out of boredom.

« Back
© 2019 Derek Arnold
This page modified on Tue Apr 30 18:12:20 CDT 2019