So, on today's installment of "random rambles about development things"
But for real, it's a good time to do a new workblog and keep people in the loop for those not in the discord, or those that aren't spending every day in it
So, what's on the ol' discussion stuffs today?
Well, for the big one, the main Feature target for 4.1: Components.
Or, more specifically DOCs. What does DOCs mean? Well...
Directors, Objects, Components
You may have heard us discuss 'Entity-Components or Entity-Component-Systems(EC and ECS, respectively). For a brief refresher in concept, here's a simple breakdown:
Entity-Components as a paradigm can be described simply as having an Entity object, and then Component objects that contain and implement data. As in, if you have a component to render a shape, the component not only holds the info for what shape to render, but also the logic to render the shape. This is how, most other engines do components.
The reason for that is pretty simple. It's robust, and it's easy to work with. It's not the most efficient system, but it's pretty hard to screw up. You slap a component onto an object, set the properties on the component, and then the component does the thing.
When I did the main previous implementation of components, this was also the system I went with. The MAIN problem with this approach is that any given component is kinda...chonky. And you also have a lot of bloat on the Entity object in most cases. And with all the bits that have to cross-communicate to ensure dependencies work(you gotta have collisions for physics to work properly for example) as well as order of events(collisions are calculated then physics) as well as any deeper engine system dependencies. It can spiral quite a lot.
Beyond that, it's also very difficult to thread any of the component's workloads because everything cross-communicates in order to work. You can't easily punt a physics component into a thread if other threads need to talk to the same collision component or entity it uses, etc.
So, advancements to the theory of components implementations lead to ECS: Entity-Component-Systems.
Now, the confusing use of "Systems" aside, the main differentiator to EC is that Components now ONLY contain data. They don't implement any logic whatsoever. Likewise, ALL the burden of functionality is moved off of Entities. In a 'pure' ECS implementation, an Entity is nothing more than an ID for Components and Systems to reference. Instead, Systems implement all functionality logic. If you have a physics component, there's PhysicsSystem that implements the actual logic for it.
This is certainly more complex to implement. In fact, very few engines or games use ECS. Unity's new DOTS approach is based on ECS, and a few games like Overwatch have utilized it. But the innate complexity of the approach and how abstracted the data and implementation means that it's far less common.
So why use it?
Because it is MUCH easier to be cache-coherent and thread things. For the non-coders out there, cache-coherency is the idea of wanting to keep all the memory a given chunk of code in the engine uses all bunched together. Think of it like how if you're studying. Rather than getting a book, reading a paragraph, then walking back to the shelf, putting that book away, and getting a new book and reading the next paragraph, and so on - which would be very, very inefficient - instead you just get all the books you need, and can quickly reference between them.
In practice, memory in the computer works similarly. So if you can cram all the data you need to work into the same blob of memory. Performance is improved SIGNIFICANTLY. But it's not very 'human friendly'. Which is why you get stuff like ECS. All components of a type can be crammed into a dense set of data. So when a System goes to implement logic, you've got all the relevent components in a tight blob of memory, and the whole thing can be processed without having to go "get another book" as it were.
In addition to this, the data being more detached from implementing logic(and better managed in memory) makes it much easier to implement the logic in multiple threads. This allows the machine to crunch a lot of objects in parallel - which is especially good on modern CPUs that have sometimes dozens of threads.
But there's a good number of downsides to this approach as well. It is, as said, not very human friendly. Implementing new components and their associated systems is not how most code is implemented, so it can be difficult to work with. It also requires much more tracking of when things are added, removed, when things should run. Dependencies are still burdened on the components and systems to keep track of for when to implement things, and scripting it is very, very difficult.
All those and a bunch of other smaller inconveniences make it generally a pretty poor paradigm to work with in something as complex as a game engine. There's a LOT of ECS implementations out on the internet. But they're more academic than practical because of the inherent limitations of the approach. Cramming it into a game engine while still making it easy to work with from a scripter, designer or artist's perspective is pretty hard.
And both of these have various limitations in how to deal with it from a networking perspective. It's very difficult to have the server and client safely agree on the data the client has without trafficking a ton of data, which is bad for net performance.
So, between what I learned from implementing an EC style deal in the first pass of components, and a lot of tests and research into ECS. I settled on the fact that, both approaches just kinda aren't ideal.
So I did some work and fashioned up a - as far as I can tell - novel components implementation for Torque3D.
Directors, Objects, Components, natch.
So, what’s the deal then? Well, per the name, there’s 3 main components(heh) to the model, which we’ll cover here:
So what’s a director? Well, in practice a Director is a simple class that ‘directs’ when and where updates to components happen, hence the name. The idea is that we want to move the burden of when and why updates happen off the objects and components.
At it’s core, a Director is in charge of doing a particular thing, generally updating a specific component or set of components. Like, say, when we want our RenderMesh component to draw. The Director has a specific timing to it(aka, Rendering) that the rest of the engine can invoke to the DirectorManager, which is pretty much just a simple container class.
When we want anything with the Rendering timing to kick off, we tell that DirectorManager to run an update on said timing. And in turn, any Director with that timing is told to do it’s work. Simple enough.
So in our example of the RenderMesh components, the RenderMeshDirector has the Rendering timing, the engine, when it goes to draw objects, can tell the DirectorManager to run the Rendering timing, and our RenderMeshDirector gets told to update. When this happens, the Director loops over valid RenderMesh components and directs them to do their work.
And thus, our RenderMesh components have drawn their meshes.
Now this sounds like a lot of work compared to just looping over the objects or components directly, but there’s a bunch of benefits to this.
For example, as noted with EC and ECS implementations, one of the biggest tangle-up points is dependency management. Normally components have to track what dependencies they have, if they’re fulfilled, and if not then they are not enabled. Any time a new component is added to it’s owner, the component is in charge of validating its dependencies.
This is important, certainly, but it also can lead to a lot of complexity, spiraling dependency chains, and code bulk on the components themselves. So instead, we move that to the director.
Because ultimately, the director is in charge of a set of components, like our RenderMesh components, we can track which ones are valid in the Director. If an Object adds a new RenderMesh component, it naturally associates to our RenderMeshDirector, and it now knows if it’s valid or not. The component itself doesn’t have to care in the slightest.
This keeps the component code leaner and cleaner, so it’s easier to maintain.
It also means we can very much more explicitly control the timing of when things run and sequence in the engine. I used the example of the “Rendering” timing before, but it’s powered by a simple enum. So you have as many entries as you can cram into an enum for when to kick off updates. You can update just physics things, or just rendering things, or specifically objects that have client inputs because they’re controlled.
This gives a much more comprehensible order of operations about when and where stuff is executed in the engine, making it easier to track and debug when stuff kicks off.
Additionally, because the director has an explicit list of components it’s in charge of, and we are specifically working on that list of components at a specific time in the execution of the engine’s loop, it means that we have MUCH more control over the memory in play for the engine.
This ties back to the aforementioned cache coherency. We can keep a list of components, like RenderMesh components, and that list can be much more easily just shoved into memory as a straight shot, minimizing how much the CPU needs to jump around. The Director works on THESE objects, so the CPU can have all the data on hand.
It also means that, between the more tightly bound memory and the express execution timing, we can much more safely handle when things are threadable. Which is a big thing for game engines.
Even major engines are still predominantly single threaded. So when you’re busting out your brand new CPU with 36 cores. Most of them are sitting around doing nothing. With Directors controlling the memory and execution, we can spool up a bunch of tasks in the threadpool and split the workload across those cores/threads that aren’t doing anything, allowing the regular workloads to be processed way faster. And this should, in theory, scale well with object counts.
So yeah, Directors are kinda the MVP of the system, what with keeping a tight wrangle on memory, streamlining execution of parts of the engine, standardizing a lot of bits, and also making the engine significantly more threadable than before.
So, you know. A little bit of a thing there. Which takes us to the next bit of our paradigm:
Compared to everything that directors do and are, Objects are pretty simple in the end. These are your entities that you slap components onto. Unlike a full-fat ECS implementation, Entities are ultimately still full objects.
There’s a good reason for that, of course. The big one is that T3D has a scripting language, and that’s super useful. So an Entity can’t just be a ID that exists in the void, because we gotta have an object for the scripts to work through.
Additionally, T3D also has a very good networking system, and not maximizing that is just dumb. And rather than having each component be manually replicated, or ghosted to clients, we exploit the way T3D does networking streams and packing to go through our Entities.
Specifically, Entities keep a list of components they own. And if a component is marked to be networked, it has a separate list for that. When a networking event happens, such as a component is added, removed, or updated, the Entity is itself flagged for networking action.
Since the Entity is ghosted to clients as normal, we can then piggyback the Entity’s network updates. Each component that’s networked has it’s own mask bits for granularity - we only need to update what actually changes - and this is packed into the Entity’s network update.
This means we can fully network whatever number of components, but only 1 ghost per Entity. And what updates we DO traffic to the client is able to be as lean as possible. This keeps the traffic as thin as physically possible without giving up the very solid networking that T3D offers.
And lastly, we have the mainstay of any component system(duh):
In DOCs, Components behave very similarly to in the previously mentioned EC paradigm. They hold data and implement the functionality for that data. So a RenderMesh component holds what mesh we want to render, and also does the logic to render the mesh.
The main difference, as covered in the section on directors, is that the components are stripped down to JUST the data and implementing logic and all the surrounding boilerplate is largely standardized up into the Directors, as well as when the components are told when to kick off their logic.
This means that implementing new components can be relatively easy, as you have the basic data/implement setup, along with any networking pass-through logic as noted in the Objects section, and then a companion Director to manage when the whole shindig activates.
All in all, while a bit more complex than standard game object classes, you get a lot of flexibility and ability to quickly slap stuff together without completely shifting to a new conceptual paradigm like a pure ECS implementation.
It also keeps networking lean, keeps scripting on the table, but also opens up the door to massively thread workloads in the engine.
So more flexibility, cleaner structure, more performance, and without compromising the good bits the engine already offers.
Not too shabby a deal, eh?
Now, this update’s already quite a long one, pretty technical and tragically limited on pictures, so I’ll do a follow up post next weekend going into the front-end usage(which is realistically where most people will work with DOCs) as well as other development stuff going on or planned.
So I’ll see you all then!