Making a game (engine) with ECS and in OCaml

Hi!๐Ÿ‘‹๐Ÿป๏ธ My name's Sergio, and I had kind of a dream: how would a game engine look if it were made in a functional style? ๐Ÿค”๏ธ Also, every project implementing a framework needs something to do, so I went with the game style of Megaman Battle Network.

Language

The language I wanted to use was OCaml, the โญ๏ธ star of the show in the ML family. I know F# exists (and have used it too), but wanted to try OCaml in itself, functors, object model and their venerable compiler.

Unfortunately, that is hard ๐Ÿ˜ฅ๏ธ... Entering OCaml means you need to understand:

  • OPAM: ๐Ÿ“ฆ๏ธ The package manager
  • Dune: ๐Ÿ”จ๏ธ The build system
  • OCaml: ๐Ÿ“‹๏ธ The language itself (which is unique and somewhat quirky)
  • Standard library: ๐Ÿ“š๏ธ The included one is pretty basic, and according to some, aging. The problem is that there are three alternatives out there and they are incompatible with each other.

As you see OCaml has opted to have a tool for each need. And while each of those tools are ๐Ÿ‘Œ๐Ÿผ๏ธ top-notch as they are fast and will cover almost any case you can imagine, they also do very little in terms of integrations and do require some setup.

Truth being told, saying this is just unfair. This is all true, but also is true that the community is making great ammends in fixing OCaml holes. Which brings me to...

One program to rule them all

After some tries of using OCaml and having little succes I found drom. You see, drom is to OCaml what cargo is to Rust, that is an unified interface for the whole set of tools. Install it with OPAM:

opam install drom

You can make a new program with just:

drom new my_program --skeleton program

That command will also tell which files to edit.

๐Ÿ™‹โ€โ™‚๏ธ๏ธ Hey, I've changed some of those configuration files, how I make the project react to it? Easy, in the root of your project do:

drom project

And to see you program in action ๐Ÿƒโ€โ™€๏ธ๏ธ:

drom run

Which will build it and run it.

Not everything are roses though, there some thorns as documentation is scarce and some topics like loading a debugger or profiling are not covered yet. But it is a marked improvement over having to learn opam, opam's switches, dune, and using all of them. Also, it feels more integrated and less intrusive than esy.

The game

As I said, the test project I made was a Megaman Battle Network. Though made is a little big as there's little more than 4 sprites, but already gives an impression of how building the project would feel.

For those who don't know what Megaman Battle Network is, let met tell you that is an action RPG (since there aren't traditional turns in it) where you control what could be called a blue Alexa deleting viruses and other badies that want to take over the internet. For what I've made is enogh to say that battles are played in a grid. Here, let me show you:

The engine

As I wanted to make this from scratch I wanted to keep it modern and use an ECS. ECS stands for Entity, Component, System and is all the rage these days. The premise is that instead of following the tradition of making an object which contains a Sprite, it's position, it's health and other parameters, all the data of one kind is held together (there's a table with all the positions from all objects) this brings huge performance wins for a large quantity of objects (100 times better cache usage and possible usage of SIMD) and improves extensibility.

As for the technology underneath I used Raylib. Raylib is a library with pretty much everything you need to make a game, think of Allegro, SFML or SDL, but:

  • is available on anything you can think of
  • it covers everything, yeah, even 3D, fonts, networking ...
  • is dead simple to use, and have really good documentation
  • there are bindings for any language in existance, really.

The code

As you know ECS is divided into Entities, Components and Systems.
Entities are here nothing more than an ID, a key that we will use to access the hash tables where all the data broken into components is stored, then the systems access any number of components they might need.

Components

To get a better example: to render a sprite we need a position and a sprite to render:

module Position = struct
  type s = {
    x: int;
    y: int
  }

  include(val (Component.create ()): Component.Sig with type t = s)
end;;
module Sprite = struct
  type s = Raylib.Texture2D.t' Raylib.ctyp

  let load path = (
    let tex = Raylib.load_texture path in
    Gc.finalise Raylib.unload_texture tex;
    tex
  )

  include(val (Component.create ()): Component.Sig with type t = s)
end;;

As you can see, we include a base module here (using first-class modules) which actually does much of the heavy lifting. Altough the include gets a bit convoluted it means that the same module is a component now (it even responds to Component.Sig the interface to components). Here, you can see the functions that our new modules have:

module type Sig = sig
    type t

    val set: t -> key -> key
    val s: t -> key -> key
    val fold: (key -> t -> 'a -> 'a) -> 'a -> 'a
    val iter: (key -> t -> unit) -> unit
    val get: key -> t
    val get_opt: key -> t option
    val remove: key-> unit
  end;;

Systems

In order to render the sprite we need to get info from both of those components:

module Rendering = struct
  let on_process _key (pos: Position.t) spr =
    Raylib.draw_texture spr pos.x pos.y Raylib.Color.white

  include (val (System2R.create on_process Position.iter Sprite.get_opt))
end;;

Again, I'm using first-class modules to fill in all the boilerplate and make us fill like we are working with just one instance (easier to think about too ๐Ÿ˜š๏ธ).

What happens if some entity lacks either Position or Sprite that entity will be ignored and nothing will be rendered for it.

Actually, my Component system is quite basic, and I'm ignoring things like hierachies, incredibily useful for position where something will always be at an offset of something else (like a sword a character is carraying and swinging around).

Input

An special case is input. Every game has input and is the main way through which the game state changes, it also shouldn't be directly tied to keys (so that it can either be ported easily or be remapped by the player in case they want to).

Here I borrowed from FRP frameworks like Elm:

module Commands = struct
type command =
  | GoUp
  | GoDown
  | GoLeft
  | GoRight
  | Attack
end;;

module Input = InputFrom(Commands)

The input componnent, which holds data and registers an entity as accepting input.

module InputHandling = struct  
  let on_input cmd (pos: Position.t) = 
    let x_size = 128 in
    let y_size = 60 in
    match cmd with
    | Commands.GoDown -> {pos with y=(Stdlib.min (pos.y + y_size) ((2 * y_size)))}
    | Commands.GoUp -> {pos with y=(Stdlib.max (pos.y - y_size) 0)}
    | Commands.GoLeft -> {pos with x=(Stdlib.max (pos.x - x_size) 32)}
    | Commands.GoRight -> {pos with x=(Stdlib.min (pos.x +  x_size) ((2 * x_size)+32)) }
    | Attack -> let key = Entity.by_name "Enemy" in Health.set ( (Health.get key) - 1 ) key|> ignore; pos

  let on_process key pos inp =
    Array.iter (
      fun (kb_key, cmd) -> 
        if Raylib.is_key_pressed kb_key then 
          Position.set (on_input cmd pos) key |> ignore 
        else ()
    )
    inp

  include (val (System2R.create on_process Position.iter Input.get_opt ))
end;;

The input system, which does all the processing, some sugar could be added to make it more amenable (maybe putting each command in a function and making on_process dissapear), but that could be added in the future. When we make an entity with input we get to link each key to each action:

Entity.by_name "Player"
    |> Position.s {x=32;y=0}
    |> Sprite.s (Sprite.load "assets/player.png")
    |> Health.s 100
    |> Script.s (can_die)
    |> Input.s [|
      (Raylib.Key.Up, Commands.GoUp);
      (Raylib.Key.Down, Commands.GoDown);
      (Raylib.Key.Left, Commands.GoLeft);
      (Raylib.Key.Right, Commands.GoRight);
      (Raylib.Key.Space, Commands.Attack);
    |] |> ignore;

See that the argument to Input.s is an array of tuples with a key and what command does it execute. About the rest of the syntax we'll talk more about it in a minute.

So now, we can assign each entity different keys and we can filter entities (in the InputHandling system we could say the we needed another component which registered this entity as a player, for example).

Entities

Thanks to the fact that now entities are nothing but an ID (a number in my implementation) and everything is in the component, nothing restrains me of:

a) Using OCaml's power to make a mini DSL
b) Adding any characteristics to any entity in a super simple way:

Entity.by_name "Enemy"
    |> Position.s {x=32+384;y=0}
    |> Sprite.s (Sprite.load "assets/enemy.png")
    |> Health.s 10
    |> Enemy.s ()
    |> Script.s (can_die)
    |> ignore;

See? I've easily marked the entity as:

  • Having a position
  • Rendering a sprite
  • Having health
  • Being an enemy
  • Being able to die

Now, while everything from there is easy to understand if you know that Component.s is nothing more than a shortcut to set and that it returns the ID to be used again later (that's why we need the ignore) the Script part may be weird to you.

And yes, ECS has an incredible limitation, it is as verbose as it can get. A simple game with both imperative old-style direct programming and OOP-like game (with something like nodes) gets to have an order of magnitude less of lines than it's full-ECS counterpart (maybe of about 200 to 400, and without taking into account all the engine behind the scenes), when a lot of what you want in the game is just a one time behavior (like the UI which there's only one of which, maybe an attack, lots of small things ...) ECS is just plain overkill, you need either another system or a escape hatch.

My scape hatch is Script a Component which will trigger a function each frame (most probably it can improved and filtered for performance) but for the time being it proved a great addition (which was even better thanks to OCaml partial application).

My impressions

This was a great project, it gave me some interesting impressions on all technologies.

OCaml

In one word.

Quirky*

Don't get me wrong, it has some really strong points to it: marlin (the autocompletion engine) is a beast, it is fast as hell, the compiler (and the build system) is so ๐Ÿ”ฅ๏ธ blazing fast that it felt more as if I was working with an scripting language ๐Ÿ๏ธ, this makes iterating a joy.

But as I hinted not so subtly earlier, the integration is ... subpar. I have only two files, everytime I modified engine.ml the whole project needed to be recompiled so that the new definitions were available to marlin for main.ml, not just that but at first it seems that drom kinda mixed versions and got a nasty error saying that I was using OCaml "4.10" while the version the tools expected was "4.11", I had to clear all caches and packages storages so and upgrade to "4.11" in the project files (just upgrading the version in the files wasn't enough). This is isn't my first time dabbling with OCaml either, and everytime I come seems like something new pops up (either that I have forgotten the kinda complicated set of instructions that I need to do).

And finally, we have the language itself, using VS Code and OCaml's plugin means having something that tells me which types have been inferred whic is ๐Ÿ‘Œ๐Ÿผ๏ธ just the best, but OCaml's fixation in not making it dependent on indentation means that the syntax gets weird at times (after all of this I still don't know 100% where to use ;; and sometimes it seems it is not needed, though that is good, actually).

My OCaml tips

So, you want to give OCaml a try? First, let me remind you of F#, which is pretty close to OCaml, is based on .NET and you can mix it with C# projects, which means you have a ton of libraries out there, you have a ton of documentation available (most of it with C#, but with some expertise you'll be able to understand both and translate) and you'll have a better time convincing your boss to use it on work.

What's worse is that one of the most compelling reasons up until now to learn OCaml, Reason, has now transformed into ReScript and most definetely they see like they want to get rid of OCaml from it. That sucks.

But this language is still something to see by yourself, and learning functors made my head go...

If you still feel like giving OCaml a try then here are some resources and tips:

  • OCaml guide: https://dev.realworldocaml.org/ -> ๐ŸŽŠ๏ธ Hands-down best resource to learn OCaml structures
  • IDE: VSCode + OCaml Platform -> There used to be several plugins (including the one from Reason) but nowadays the best is this one, made by the OCaml Platform itself (a group making OCaml more amenable)
  • Standard library: Batteries or Jane Street's Core: You know that OCaml's standard library is a bit... ๐Ÿ˜‘๏ธ lacking, and old, so there are several attempts to replace it. On the one hand you have the battle-proven Core by the folks on Jane Street, a big brand using exclusively OCaml but is incompatible with OCaml's standard library and the other hand you have Batteries a community effort pretending to extend OCaml's standard library.

ECS

Finally, I understand what I've read about using ECS. While it feels awesome defining an entity and giving it any property I feel like, trying to implement the a small label that showed our health (and another for the enemy) meant that I needed to make a Component and a System just for that, at that time I understood that trying to shove everything inside the ECS is going to make your code verbose as hell and you may end with several times the number of lines you might had otherwise.

In my case, I made the Script Component as an escape hatch, and honestly it felt ๐Ÿ˜€๏ธ good. OCaml's partially applied functions really helped too. Honestly, I can see this design getting some sugar and being used more.

Functional programming

The code here is not super functional, I acknowledge that, yet the part where it is more functional, that is the Input, while still verbose feels pretty good, the Command idea taken from FRP and Elm is a great way of abstracting input hardware.

But I can see shoving everything inside like an object called World and passing that down on all functions for them to do wathever they feel like, something like what the Nu game engine does.

Raylib

If I take something really good from here is Raylib. As I said you can use Raylib pretty much wherever you use SDL, but is really easy, with functions with understandable names, modern and easy-to-parse web with tons and tons of examples. But what does it have? Well, in Raylib you'll find: 3d (with it's own abstraction layer), audio (no 3d audio, but it does support streaming), physics, texts, 2d, GUI and even networking seems to be on the works: https://github.com/raysan5/raylib/issues/753

Basically, is an engine but as a library.

My only gripe is that the 3d part lacks Vulkan, and while raysan5 (the person behind Raylib) is not against it, does not have a lot interest in it.

I would take raylib for a ride if your project needs low-level control.

Final thoughts

TL;DR: ECS is good but tedious and verbose, OCaml is quirky but with people putting some real effort on it, and raylib has chances of becoming my go-to low level game lib.

I might further develop this architecture on the future, most probably not to use it by itself, but as research and to port it to other frameworks later.

Some frameworks that I'm interested in and that I might test are:

If you have had experiences on those, let me know in the comment. Also, have you had any experience on functional games and/or ECS?

Links

31