Comment by TeMPOraL
5 years ago
Thanks for the detailed evaluation. I'll start by reiterating that the project is a typical tile-based roguelike, so some of the concerns you mention in the second paragraph don't apply. Everything runs sequentially and deterministically - though the actual order of execution may not be apparent from the code itself. I mitigate it to an extent by adding introspection features, like e.g. code that dumps PlantUML graphs showing the actual order of execution of event handlers, or their relationship with events (e.g. which handlers can send what subsequent events).
I'll also add that this is an experimental hobby project, used to explore various programming techniques and architecture ideas, so I don't care about most constraints under which commercial game studios operate.
> The perceived behavior is intimately coupled to when its processing occurs and when the effects are "felt" elsewhere in the loop - everything's tied to some kind of clock, whether it's the CPU clock, the rendered frame, turn-taking, or an abstracted timer. These kinds of bugs are a matter of bad specification, rather than bad implementation, so they resist automated testing mightily.
Since day one of the project, the core feature was to be able to run headless automated gameplay tests. That is, input and output are isolated by design. Every "game feature" (GF) I develop comes with automated tests; each such test starts up a minimal game core with fake (or null) input and output, the GF under test, and all GFs on which it depends, and then executes faked scenarios. So far, at least for minor things, it works out OK. I expect I might hit a wall when there are enough interacting GFs that I won't be able to correctly map desired scenarios to actual event execution orders. We'll see what happens when I reach that point.
> that's the kind of thing where you could theoretically use SQLite and have a very flexible runtime data model with a robust query system - but fully exploiting it wouldn't have the level of performance that's expected for a game.
Funny you should mention that.
The other big weird thing about this project is that it uses SQLite for runtime game state. That is, entities are database rows, components are database tables, and the canonical gameplay state at any given point is stored in an in-memory SQLite database. This makes saving/loading a non-issue - I just use SQLite's Backup API to dump the game state to disk, and then read it back.
Performance-wise, I tested this approach extensively up front, by timing artificial reads and writes in expected patterns, including simulating a situation in which I pull map and entities data in a given range to render them on screen. SQLite turned out to be much faster than I expected. On my machine, I could easily get 60FPS out of that with minimum optimization work - but it did consume most of the frame time. Given that I'm writing a ASCII-style, turn(ish) roguelike, I don't actually need to query all that data 60 times per second, so this is quite acceptable performance - but I wouldn't try that with a real-time game.
> The other thing that might help is to have a language that actually understands that you want to do this decoupling and has the tooling built in to do constraint logic programming and enforce the "musts" and "cannots" at source level. I don't know of a language that really addresses this well for the use case of game loops - it entails having a whole general-purpose language already and then also this other feature. Big project.
Or a Lisp project. While I currently do constraint resolution at runtime, it's not hard to move it to compile time. I just didn't bother with it yet. Nice thing about Common Lisp is that the distinction between "compilation/loading" and "runtime" is somewhat arbitrary - any code I can execute in the latter, I can execute in the former. If I have a function that resolves constraints on some data structure and returns a sequence, and that data structure can be completely known at compile time, it's trivial to have the function execute during compilation instead.
> I've been taking the approach instead of aiming to develop "little languages" that compose well for certain kinds of features
I'm interested in learning more about the languages you developed - e.g. how your FSMs are encoded, and what that "programmable painter system" looks like. In my project, I do little languages too (in fact, the aforementioned "game features" are a DSL themselves) - Lisp makes it very easy to just create new DSLs on the fly, and to some extent they inherit the tooling used to power the "host" language.
Sounds like you may be getting close to an ideal result, at least for this project! :) Nice on the use of SQLite - I agree that it's right in the ballpark of usability if you're just occasionally editing or doing simple turn-taking.
When you create gameplay tests, one of the major limitations is in testing data. Many games end up with "playground" levels that validate the major game mechanics because they have no easier way of specifying what is, in essence, a data bug like "jump height is too short to cross gap". Now, of course you can engineer some kind of test, but it starts to become either a reiteration of the data (useless) or an AI programming problem that could be inverted into "give me the set of values that have solutions fitting these constraints" (which then isn't really a "test" but a redefinition of the medium, in the same way that a procedural level is a "solution" for a valid level).
It's this latter point that forms the basis of many of the "little languages". If you hardcode the constraints, then more of the data resides in a sweet spot by default and the runtime is dealing with less generality, so it also becomes easier to validate. One of my favorite examples of this is the light style language in Quake 1: https://quakewiki.org/wiki/lightstyle
It's just a short character string that sequences some brightness changes in a linear scale at a fixed rate. So it's "data," but it's not data encoded in something bulky like a bunch of floating point values. It's of precisely the granularity demanded by the problem, and much easier to edit as a result.
A short step up from that is something like MML: https://en.wikipedia.org/wiki/Music_Macro_Language - now there is a mostly-trivial parsing step involved, but again, it's "to the point" - it assumes features around scale and rhythm that allow it to be compact. You can actually do better than MML by encoding an assumption of "playing in key" and "key change" - then you can barf nearly any sequence of scale degrees into the keyboard and it'll be inoffensive, if not great music. Likewise, you could define rhythm in terms of rhythmic textures over time - sparse, held, arpeggiated, etc. - and so not really have to define the music note by note, making it easy to add new arrangements.
With AI, a similar thing can apply - define a tighter structure and the simpler thing falls out. A lot of game AI FSMs will follow a general pattern of "run this sequenced behavior unless something of a higher priority interrupts it". So encode the sequence, then hardcode the interruption modes, then figure out if they need to be parameterized into e.g. multiple sequences, if they need to retain a memory scratchpad and resume, etc. A lot of the headache of generalizing AI is in discovering needs for new scratchpads, if just to do something like a cooldown timer on a behavior or to retain a target destination. It means that your memory allocation per entity is dependent on how smart they have to be, which depends on the AI's program. It's not so bad if you are in something as dynamic as a Lisp, but problematic in the typical usages of ECS where part of the point is to systematize memory allocation.
With painting what you're looking for is a structuring metaphor for classes of images. Most systems of illustration have structuring metaphors of some kind specifically for defining proportions - they start with simple ratios and primitive shapes, and then use those as the construction lines for more detailed elements which subdivide the shapes again with another set of ratios. This is the conceptual basis of the common "6-8 heads of height" tip used in figure drawing - and there are systems of figure drawing which get really specific about what shapes to draw and how. If I encode such a system, I therefore have a method of automatic illustration that starts not with the actual "drawing" of anything, but with a proportion specification creating construction lines, which are then an input to a styling system that defines how to connect the lines or superimpose other shapes. Something I've been experimenting with to get those lines is a system that works by interpolation of coordinate transforms that aggregate a Cartesian and polar system together - e.g. I want to say "interpolate along this Cartesian grid, after it's been rotated 45 degrees". It can also perform interpolation between two entirely different coordinate systems(e.g. the same grid at two different scales). I haven't touched it in a while, but it generates interesting abstract animations, and I have a vision for turning that into a system for specifying character mannequins, textures, etc. Right now it's too complex to be a good one-liner system, but I could get there by building tighter abstractions on it in the same way as the music system.
My main thing this year has been a binary format that lets me break away from text encodings as the base medium, and instead have more precise, richer data types as the base cell type. This has gone through a lot of iteration to test various things I might want to encode and forms I could encode them in. The key thing I've hit on is to encode with a lot of "slack" in the system - each "cell" of data is 16 bytes; half of that is a header that contains information about how to render it, its ID in a listing of user named types, bitflags defined by the type, a "feature" value(an enumeration defined by the type), and a version field which could be used for various editing features. The other half is a value, which could be a 64-bit value, 8 bytes, a string fragment, etc. - the rendering information field indicates what it is in those general terms, but the definite meaning is named by the user type. The goal is to use this as a groundwork to define the little languages further - rather than relying on "just text" and sophisticated parsing, the parse is trivialized by being able to define richer symbols - and then I can provide more sophisticated editing and visualization more easily. Of course, I'm placing a bet on either having an general-purpose editor for it that's worthwhile, or being able to define custom editors that trivialize editing, neither of which might pan out; there's a case for either "just text" or "just spreadsheets" still beating my system. But I'd like to try it, since I think this way of structuring the bits is likely to be more long-run sustainable.
Thanks for the detailed explanation!
I don't think I can respond properly and ask the many more questions I have in a HN thread that's already this aged, so if you'd like to talk more, feel free to hit me up (my contact details are in my profile).