I found this excellent talk to be complementary to my talk[1] on data-oriented GUI in Rust, not to plug my own work too much.
I found a lot of common ground:
* Trying to write object-oriented code in Rust doesn't work well. Writing data-oriented code does.
* "Fighting the borrow checker" is a sign you might be doing it wrong. Try data-oriented approaches instead.
* Use a structure-of-arrays rather than an array-of-structures.
* Use what I call "state splitting" to borrow mutable references to just the state you need; this works well with the structure-of-arrays approach and is a powerful motivator to use it.
* To build graph-like structures, reach for a Vec of components, and indexes into the Vec for relationships, as your first choice. Self-references, arena allocators, and Rc are viable alternatives, but you should have a good reason to use them instead of Vec.
* Some form of dynamic typing is useful to keep your system loosely coupled (though we ended up with very different forms of dynamic typing, see below).
* Data-oriented approaches have taken root in the C++ gaming community, mostly motivated by performance, but adapt well to Rust, and some of the ideas may be useful in domains beyond gaming.
There were some other points that went beyond what I talked about, but could fit in well:
* A "generational index" is a good way to avoid use-after-free style errors that result from the use of a stale index. The slotmap crate can help.
And now for the things that are different.
* The exact nature of dynamic typing is different. An ECS usually uses a registry of the different component types (anymap is a useful crate) and is quite open-ended in adding new types of components. My xi-win-ui, by contrast, has two main component types, Widget and listener, and does dynamic dispatch on a Widget trait.
This naturally raises the question, which is better? From my perspective, both are valid. The pure ECS approach definitely makes sense when the components are interacting with each other in diverse ways (collisions, damage, etc). In a GUI, the components are mostly interacting with the framework, and have diverse behaviors within standardized interfaces (input, layout, paint).
Thus, one point of my talk is that you don't have to reject all vestiges of object oriented programming, it's perfectly reasonable to hybridize, using data-oriented approaches for state splitting and graph structure, and dynamic dispatch for the behavior variation.
My talk also went deeper into the ideas of "data flow rather than control flow" and the use of an event or command data structure rather than method calls. I'm not sure how deeply these ideas apply to games, but wouldn't be surprised if they do.
This sounds like it is just going to trade one set of problems for another. It makes it impossible to write generic container types. What if elements of the same collection need to have different structure? What benefits justify this extremely tight coupling?
It's frequently done that way for performance. Imagine a game of Starcraft, with 1000 zerglings rushing your base. The game has to repeatedly loop over all the zerglings to move them. Since there are lots of other fields tracking all of the other data about each zergling, the normal AOS approach has poor data locality; you load a cache line and then you only touch a few bytes of it. With the SOA approach you're looping over an array of positions, so every byte that you fetch from memory ends up being used.
I agree with you but this I am not sure it is the right example. In this case, the position of the zergs would be indirectly a component. You would have arrays of struct Zergs, each would only have a ref (or an index) to an arrays of positions. this array would be updated efficiently.
> What if elements of the same collection need to have different structure?
They don't. If you ever end up in a situation where you feel they do, the correct solution is typically instead to split the component into multiple, different components, only some of which will be used for any given entity. This is basically the same as defining a schema for an SQL table. A component is just a set of state, there is no requirement for it to map 1-to-1 to a specific functionality.
> What benefits justify this extremely tight coupling?
The most commonly stated one is speed. ECS was adopted first in game design because it is just so much faster. On modern OoO cpus it's typically something like 5-10x faster than traversing an object graph. On the previous generation consoles (PS3, XB360) with their in-order CPUs and crappy load/store subsystems, it could easily be 20x-50x faster. I don't know how relevant this is to your typical GUI, though; the speedup in an ECS comes from linear memory access, which means that the prefetchers make sure every memory access is an L1 access, which is great when you have a game that has thousands of entities, which don't fit into any cache. But just how many GUIs have enough state to overflow the L1 anyway?
However, speed is not the only benefit. This is somewhat subjective, but having implemented similar logic for ECS and OO based games, I feel that the logic is almost always much clearer, more understandable and less buggy in the ECS versions. Basically, in OO doing things that have cross-cutting concerns tends to get split into many small parts done in multiple places, and it's hard to understand the whole system at once. In an ECS, the logic for one system is implemented in one place, it is always just a transformation that reads in some data, does some computation on it, and writes out some data, without complex control flow. It's so much easier to understand and test.
> It makes it impossible to write generic container types.
An example of a Generic container type is AnyMap, which holds one value of each type (and each value will typically be either a straight Vec for small/common components, or some kind of more complex set for components that hold a lot of data.)
(edit: looked up old numbers and found that 100x was pushing it, even on Xenon. 50x ought to be realistic.)
I mean, that's sorta the point of the talk. Did you watch it?
> This sounds like it is just going to trade one set of problems for another.
Sure, that's exactly what a tradeoff is.
> what substantiates the claim that it is generally superior?
I don't think the claim is that it is always superior.
Oh, and I would see this refactor of a decoupling. The issue is that, if you’re trying to process Players in certain ways, the fact that the name and health are coupled together in a single struct is an issue. This pulls them apart.
As steveklabnik says, it absolutely is about tradeoffs. The anymap crate may provide enough of "generic container types" to be useful, and can avoid a lot of repetition of per-type code. A graph with heterogeneous node types is definitely possible, Box<Any> is one solution and there are others.
Doen't using indexes into arrays introduce, more or less, the same problems with C pointers? after all, a C pointer is an index into a huge array, the current process' memory space.
I found a lot of common ground:
* Trying to write object-oriented code in Rust doesn't work well. Writing data-oriented code does.
* "Fighting the borrow checker" is a sign you might be doing it wrong. Try data-oriented approaches instead.
* Use a structure-of-arrays rather than an array-of-structures.
* Use what I call "state splitting" to borrow mutable references to just the state you need; this works well with the structure-of-arrays approach and is a powerful motivator to use it.
* To build graph-like structures, reach for a Vec of components, and indexes into the Vec for relationships, as your first choice. Self-references, arena allocators, and Rc are viable alternatives, but you should have a good reason to use them instead of Vec.
* Some form of dynamic typing is useful to keep your system loosely coupled (though we ended up with very different forms of dynamic typing, see below).
* Data-oriented approaches have taken root in the C++ gaming community, mostly motivated by performance, but adapt well to Rust, and some of the ideas may be useful in domains beyond gaming.
There were some other points that went beyond what I talked about, but could fit in well:
* A "generational index" is a good way to avoid use-after-free style errors that result from the use of a stale index. The slotmap crate can help.
And now for the things that are different.
* The exact nature of dynamic typing is different. An ECS usually uses a registry of the different component types (anymap is a useful crate) and is quite open-ended in adding new types of components. My xi-win-ui, by contrast, has two main component types, Widget and listener, and does dynamic dispatch on a Widget trait.
This naturally raises the question, which is better? From my perspective, both are valid. The pure ECS approach definitely makes sense when the components are interacting with each other in diverse ways (collisions, damage, etc). In a GUI, the components are mostly interacting with the framework, and have diverse behaviors within standardized interfaces (input, layout, paint).
Thus, one point of my talk is that you don't have to reject all vestiges of object oriented programming, it's perfectly reasonable to hybridize, using data-oriented approaches for state splitting and graph structure, and dynamic dispatch for the behavior variation.
My talk also went deeper into the ideas of "data flow rather than control flow" and the use of an event or command data structure rather than method calls. I'm not sure how deeply these ideas apply to games, but wouldn't be surprised if they do.
[1] video: https://www.youtube.com/watch?v=4YTfxresvS8 , slides: https://docs.google.com/presentation/d/1aDTRl5R-icAF38Di-qJ4...