godot – The Roohr Way

My vow to stop making EV-likes has had mixed results. Here, I’ll discuss some of the more interesting aspects of a recent EV-Like, Galactic Night.

Perspective on perspectives

While I did make a serious run at doing a vehicle game in Godot, it was hard to get around just how excellent the affordances of what I’d already built were, and what was lacking in Godot 3. But it didn’t start there. It started with an observation:

First Let’s Play google gives you for “EV Nova.” I haven’t listened to the whole thing, so I can’t vouch for it. But it’s got footage of the game.

Ships in EV Nova (and MPEVMVP, for that matter) are sprites rendered from 3d models. They are rendered with perspective so that when the front of a ship faces the screen, it appears larger than the back. It’s subtle on some ships, obvious on others.

Perspective camera isn’t the norm for sprite games. Most games with pre-rendered sprites use an isometric camera; here’s Starcraft as an example:

First result for broodwar gameplay. The Terran Siege Tanks are the best example of isometry.

For games where the units are in a rich world with lots of points of reference, rendering them in perspective might look wonky. Too much perspective might look wonky under any circumstances! However, the decision to use perspective in the sparse world of Nova lends the ships a sense of scale.

2.5d games that use a full 3d world instead of sprites tend to either use a perspective or isometric camera. It’s easy to implement either in Godot (and other engines) by changing the camera’s mode. But I was curious: could I implement something with a vertex shader that would deform a mesh such that it appeared to have perspective, even while rendered in an isometric view? Perspective, I should note, not truly relative to the camera, but relative to an imaginary camera ‘rendering’ the imaginary ‘sprite’?

This is difficult to convey with words, so I built a demo:

What’s interesting about this, at least to me, is how difficult it is to notice what’s going on. Your brain seems to happily accept the very unnatural perspective. At least mine does. But if you look closely, you can see that no matter where the spaceship (and thus actual iso camera) move, the spinning cubes maintain their fixed perspective, just like sprites in an old school video game.

Flying around was fun, I wanted to do more stuff. It got out of hand. It’s a whole game now.

More Ship Shader Fun

Using a 3d pipeline offered the possibility of using some nice 3d features, exploiting the fact that the textures and effects were in-engine rather than an external pipeline. Nova has engine glows and weapon glows; I implemented those too.

Nova also has an (unused) feature where you can stain your ship a different color. I wanted to try something similar, but nicer:

You can check out the full ship shader here. Note that in GD4 you can split shaders into multiple files, so the important parts are the perspective transform here, and the color swap logic here.

Union Bytes Painter allows for multiple texture layers, so it’s easy-ish to make separate textures for lights, engines, weapons, paint, and a base. It’s a little clunky for exporting though – I need to show/hide layers and export a few times. Still miles better than the Blender workflow I used while working on Flythrough.Space.

Flat Space

I wanted to create a nice set of planets and nail the EV Nova look, so I followed tradition and picked up the venerable LunarCell, which was also used to render the planets for EV Nova. This creates (gorgeous) 2d planet sprites:

Only problem? Lunar Cell doesn’t create masks, just planets on a black background. Solution? This remarkably simple shader:

https://github.com/EamonnMR/galactic-night/blob/main/entities/spobs/Spob.gdshader

You’ll notice though that that’s a 2d shader, and the game is 3d. That’s the neat part – the background is a 2d canvas! We use this class to make sure the sprite in 2d space tracks the correct position in 2.5d space:

https://github.com/EamonnMR/galactic-night/blob/main/component/FollowerSprite.gd

The background itself is basically the same shader I used in Survive Space but with additional special effects you can cue up to make hyperspace more exciting:

I wanted to visually represent folding space, and the looped background made for a unique opportunity to do it. The shader is of course just dropping specific images (Screaming Brain’s awesome Nebula backgrounds) on a couple of layers, but it would look better looped if I did something with Perlin noise.

Procedural Generation

Flythrough.Space used a handmade universe I originally drew on graph paper. While charming, that took a whole lot of work for very little gain at the end of the day. For MPEVMVP I used universe maps generated by Mag Steel Glass’s spreadsheet. Which is fine by all accounts, but I wanted the player to be able to generate new universes on the fly. That’s where I ran into my new best friend: Delauny Triangulation.

Compare:

(source: https://preterhuman.net/software/escape-velocity-macintosh/)

(wikipedia)

Ok, you need to look a bit closely, but in both cases, you’ve got a point cloud that is stitched up by non-overlapping lines. So to generate an EV map, because there’s a library function for that all you need to do is feed it a set of points and get your EV map’s hyperlanes!

https://github.com/EamonnMR/SurviveSpace/blob/main/procgen/Procgen.gd#L87

That same strategy carried over to Galactic Night, with some modifications, namely the division of space into quadrants to provide a difficulty curve, and the separation of growing biomes from growing faction influence. In Valheim, the game I imitated most on Survive Space, all three are effectively the same; a biome is a difficulty level is a spawn location for various monsters. In Galactic Night, spawns can depend on biome (for asteroid types) faction (for where to spawn enemies or allies) and quadrant (more difficult NPCs.)

Codex

The codex window loads up a folder tree of bbcode (weirdly, Godot’s rich text boxes support this) files and displays them. Item fluff that would usually be embedded in the interface in an EV goes here, a much more conventional placement for a modern game. I never did get around to writing a system for unlocking entries.

Retrospective

I’m proud of the technology and writing that went into Galactic Night, and some of it will certainly be recycled for future projects. I’d be happy if people get any use out of the code or design concepts tested out. I’ll close out with some final thoughts on how the project went.

I think the perspective looked really good, but some people definitely found it jarring. In future projects, it should definitely be an optional feature.

Doing the textures in Union Bytes painter became a slog. Small UI issues compounded over the course of several objects, and it stopped being fun.

Upgrading from GD3 to GD4 mid project was a big hiccup, but ultimately worth it because the syntax of GDscript is far nicer.

Though the procedural universe was really fun to tweak and play with, I don’t think it provided enough stuff for an exploration focused game. Games like Minecraft and Valheim exploit the inherent human desire to play around in a virtual space, and an EV map doesn’t really provide that; in a game like this the contours of the explored world are unique spaceships you find, well written place descriptions, missions, graphics, and the like. I never found a way to work in that level of diversity. Also, I never found an effective way to hint to players that, for example, getting lostech will allow them to increase their tech level and make more options available. Sure, it’s there in the manual, but who’s gonna read that! The upshot was that it felt like a big empty map with nothing particularly interesting going on in it, and no motivation to strike out and try to do stuff. Oh well.

Godot’s High Level Networking tutorial (and book chapter) do a great job of explaining a peer-to-peer networking setup. However, many games want to be server-authoritative, and to set up a multi-area game, you sort of need to. Luckily, you can use the same primitives to build a server/client game. This is just sort of a brain dump of my notes that may be helpful in navigating the MPEVMVP source. Feel free to post any questions you have about it; I love to talk about this stuff.

The technical requirements that deviate from the basic tutorial they hand out are as follows:

Server-authoritative client/server game
Multiple levels that players can switch between
Players are able to join a match in progress

In case you haven’t tried it, the project is here: https://github.com/eamonnmr/mpevmvp. The gameplay is similar to Asteroids (and, well, similar to Flythrough.Space): it uses RigidBodies to simulate movement, you can shoot other players to force them to respawn, and you can press “J” to initiate a hyperjump after you’ve pressed “m” to select a star system from the map.

Godot’s Networking Golden Rule

A remote call (rpc, rset) is going to be called on the remote node with the same node path. That includes name of the node, the name of the parent, etc, all the way to the root node. You can set a child node’s name with set_name but if a similarly named child already exists, an @number will be added to your node. This is a major source of bugs for networked projects, so be careful. You cannot set a node to a name with @ in it. So in effect, you cannot use Godot’s auto-incremented names (since they append @increment) to sync between the client and the server. For this reason, I used a uuid library for generating node names.

The need for the same structure on client and server despite the fact that the server held a whole universe while the client only cared about one system is a major source of complexity in the codebase.

Server Authoritative

I implemented the Client and Server as singletons that exist on both sides. The reason for this is that it makes it really readable; when the client needs to send something to the server you write Server.rpc and when the server needs to send something to the client, you use Client.rpc. Besides this line of communication, I’ve used rpc and rset within various nodes to update and alter the client versions of themselves. This always happens inside an is_network_master block, to ensure that it’s being called on the server version of the game.

The one thing we do need to gather from clients is input, and for that each client has a PlayerInput node, network_master’d to them. They rset_id input up to the server. The server copy of the player nodes looks at this remote copy of the player_input to determine it’s behavior, then pushes its state back down to all clients.

Multi Level

Suppose you want to have multiple different levels/worlds/rooms in your game. The trick is that you need to mesh the following two constraints:

Node paths need to be the same between client and server (see discussion above)
The server needs to have different node paths for each level

The trick, then, is to send the client to the new level by loading the appropriate level and making sure it’s named appropriately. There’s one wrinkle though: different 2d physics worlds need to use a Viewport class, but you don’t want to add additional viewports on the client for performance reasons (tried it, it was way too slow.) My solution was to add an extra layer of nodes called “universes” in the comments which are a Node2d on the client but a Viewport in the server. They are managed on the server by ServerMultiverse and on the client by ClientUniverse. Client Universe cares about a single “universe” child whereas ServerMultiverse handles multiple children.

Sending Entities/Switching Levels

In order to join a match in progress or switch to a level that already has stuff in it (or, indeed, switch a player into a level) we need a robust way for the server to move fully-formed nodes to the client. Godot does not provide a generic “send this node” method, so we need to write one for ourselves. The approach I took was:

Assume that the ent will be instanced from the same scene that the ent on the server is
Write a “serialize” and “deserialize” method that allows it to dump a dictionary with all of its important state and recreate itself from that dictionary.
Make sure the node on the client side is assigned the same name as the one on the server side.

So when a client’s level is switched, they dump the current level and recreate the new one by loading in the scene and then, node by node, deserializing the state sent from the server. As it turns out, this means that any state can be synced freely, and it enables us to dispense with the lobby entirely and let players join and leave at will.

Sync In Multiple Levels

Each level owns the process of syncing its state to clients in that level. Scenes within the level can declare a method that serializes their state for a net frame, and the level gathers all of those up and sends them. The client receives the frames from the server, and the individual scenes use the content of those frames to interpolate their position. To save bandwidth, I only do this for ships and guided projectiles, everything else moves deterministically. This whole process is derived from the teaching of (and better described by) this video.

Player Lifecycle

Because we’re using the player’s presence in a level to determine if we should update entities in that level for that player, when a player’s avatar is destroyed the world seems to stop for that player until they respawn. This might be desirable in some games, but I think most games at least want to show the player the immediate seconds after their own demise. So what we do is replace the player with a ghost.

Tooling & Workflow

GDScript is based on Python 3 and the most basic syntax resembles it. ‘func’ in place of ‘def’, no ‘self’ and no list or dict comprehensions or other advanced features. It feels like the python of basic python tutorials. However in exchange for this, you get one thing python 3 does not offer: true static typing. In Python 3 you can offer a type hint for a function argument but it is not truly enforced. GDscript enforces the types at compile time and boy does it ever catch a lot of bugs. You also get the ability to define classes with typed members, similar to what you get in Python with Pydantic. Overall I like gdscript and the things I miss aren’t dealbreakers.

The lack of a vi mode for the editor hurts. I know I could just use the godot plugin for vi, but I actually really like staying inside Godot’s editor. Despite the slightly crowded layout, it’s really very nice. The autocomplete to node paths is a killer feature that makes dealing with complicated trees less of a pain.

Conclusion

Building a multiplayer game in Godot is possible but by no means easy. The high level multiplayer API despite its gotchas works well, but you need to implement the sync logic yourself. This shouldn’t be too surprising as games have very different needs when it comes to net code. I hear that a similar mechanism for frame based sync and lerping is going to be more built-in in godot 4, which would be sweet. The puppet system (which I initially used; you can see a version in the tutorial) seems to be useful only for very low latency situations, so I regret that it’s a first-class feature because it may lead people down a rabbit hole of net code that will not work for most games.

Tag: godot

(Re) Introducing Galactic Night