For the past week I've made great strides in getting a lot of the essentials into the game engine. At this point I'm finally ready to jump into the battle system and see the gameplay start to form into something more concrete.
For RPGs of any kind, there's always going to be a lot of text to display. This was one of the things I took some time to properly consider. The main decision was around having support for Unicode. There's two schools of thought for this when building games. Either you:
I couldn't see a right way of doing it. There's a lot of bespoke solutions to supporting text in games, but nothing definitive. Some people use Freetype and build font caches dynamically at runtime, other people generate bitmapped font textures for each combination of font size / font style they need, and others like Valve go the extra mile and render fonts entirely with vertices on the GPU.
In previous unreleased games I've used the former two solutions. For this game, since it's going to be a 2D game that's designed to look like a GBA game, I wasn't too concerned with having uber high resolution fonts that scale dynamically, and I felt it would feel a bit jarring having very smooth fonts coupled with scaled up sprite art, so I opted for the simplest of all the options - bitmap fonts.
My aim is to support multiple languages easily. Initially though, my target languages are English and Japanese, so for my bitmapped fonts I've packed in all of the latin characters necessary to render English, and all of the Hiragana, Katakana and common Kanji I'd need. This could be massively reduced later by extrapolating all of the characters used in the localised text files and building a font from those. This also appears to be the strategy a lot of Chinese and Japanese games use in order to avoid packing in 30,000+ Kanji into a single texture.
It took a bit of time to figure out a decent way to render text. For now I've gone with a basic but highly inefficient solution of rendering each UTF-8 codepoint's glyph in a loop. I plan on going back later and instead generating vertices for each glyph in a loop and caching it from then on in order to avoid slowing down the render loop. The results are pretty cool nonetheless:
As you may have noticed from the borrowed test art, I want the game to have a very Earthbound-like feel to it. One of the things I love about the aesthetics of this series is it uses a very simple art style to great effect, rather than other games from the era like FF6 or Star Ocean that try to pack in a ton of detail into a very small space.
One of the benefits of this art style is it has a very light hearted feel about it, which can be used to great effect in the game when telling a dark and gritty story, as witnessed in all of the Earthbound games. I love that combination, and I definitely want to achieve similar results in my game.
With the UI side of things, keeping with the theme, I don't want it to be too detailed, rather I want it to take a backseat most of the time except when you actually need it. To this end I've pared down the UI widgets to the bare minimum - bordered windows, labels, a focus underline that shows you where your cursor is, and minimal use of icons where necessary.
Bordered windows are the only real special case as they involve some elements stretching (or tiling) dynamically, whilst others remain fixed. I've done this countless times when rendering on the CPU, but doing it on the GPU was all new to me, as evidenced by my many crazy renders along the way:
After a while I had my eureka moment and managed to get the shader to play ball. I had a realisation that I was basically working against the system which is why nothing seemed to work quite right. What I was attempting to do in the shader was map the appropriate portion of the texture depending on what vertex X & Y values were coming in. The problem with that is the X and Y values overlap, as shaders deal with points, not faces or triangles.
Once I realised this I tagged the coordinates I needed for a particular group of vertices in the vertex array itself, as that way I'd have all the data I needed without having to determine things on the shader. The results speak for themselves:
This was a really critical point for me as now I understand exactly how shaders work, and how to best utilise them in OpenGL, which will help me a lot when dealing with more complex rendering scenarios.
As shown in the above screenshot I also got tinting working for the UI fairly easily by applying a HSV adjustment inside the fragment shader. This lets me achieve dynamic UI colours, which Earthbound and Mother 3 both use to great effect. It's a nice touch of customisability in the game that I'm eager to replicate.
With my newfound knowledge, I had a lot more confidence in my ability to work with a lot of vertices in a single shader when rendering a chunk of the game, rather than just using two triangles like everything else has been done so far.
To this end, I really wanted to go back and revisit the map drawing routines. When profiling in Visual Studio I noticed that 60% of the render time was taken up by drawing the map by looping through the full tileset and rendering quad by quad. It was horribly inefficient and I couldn't see the game going live with map rendering in that state.
I spent a lot of time whilst getting text rendering to work on more optimal ways to do it, and on Reddit one person mentioned building up a list of vertices for the full string and caching it, to prevent looping every frame. I thought that the same approach could work for maps too, so I gave it a try. I was honestly shocked at how quick it was to get working - within half an hour I'd adapted the vertex shader to get all of the vertices in the right place, and as soon as that was done it worked. Instantly.
It felt like night and day with my previous attempts, where before I spent the best part of three days trying to get something reasonable displaying on-screen. I think my mistake before was trying to replicate what comes naturally to the GPU on the CPU. Before I was attempting to cull each individual tile outside the viewport and only render visible tiles, taking into account the camera being between tiles where you have to add an extra tile of padding, etc. While I'm sure I could've eventually got the maths right, it was still the wrong approach ultimately.
Now all I have to do is convert the tilemap into a vertex array, bind it, and let the GPU figure out what to display and what not to display. On the CPU side I can safely assume the full tilemap is rendered, which makes positioning the map taking into account the camera position trivial, and on the GPU side it happily crunches through all of the points in the vertex shader, making use of additional attributes passed in via the vertex array. Just like I'd done with the bordered windows.
There's no screenshot for this one since it's literally identical to the last blog post's screenshot. This one was all about performance. In terms of numbers - on my mid-2015 Macbook Pro the game used to use 30-35% CPU. Now it uses 10% CPU. That's a monumental drop in CPU usage, which makes this absolutely worth it in the end.
With this huge win I plan to revisit the text rendering later and apply a similar change in order to reap the same kind of performance benefits that this has gained me with map rendering, but for now I definitely want to move on to more gameplay-focused work to get something more substantial.
One of the things I wanted to do for this engine was build a flexible UI layer that could easily be tweaked to suit the needs of the game. I really liked working with ImGui for the debug windows, and wanted a similar experience for the normal UI present in-game.
I achieved this by building up a few POCOs holding the bare minimum
information for the UI -
Dialog structs that hold information on
the frame, and its contents, and
holding information on visual content present within a
This all then gets thrown at the renderer which uses its own APIs to render the low level elements like window borders, text and blocks of colour. They're essentially a 1:1 mapping, which makes rendering the UI trivial.
For the adaptibility side of things, I do plan on having Lua scripts be in charge of UI management in its entirety so that the UI can be treated as a resource in its own right, rather than being hardcoded into the engine. This allows for scenarios like overriding certain dialogs to operate differently on more bespoke platforms like consoles, and it also speeds up the development cycle by allowing for tiny changes to be tested without recompiling.
Wiring this all into the update loop was pretty straightforward. I'm forgoing actually building any kind of settings screens, etc, for now as I don't really need them yet, but I proved out the approach at least:
There were a ton of performance fixes, bug fixes, networking enhancements and other non-visible changes in the last week, and it definitely swayed significantly in that direction, so it felt at times like things were stagnating. This changed a lot towards the end of this week though, as I managed to improve a ton of things on the UI side and crossed another two features off of my list, which felt pretty good.
Over the course of the next week my aim is to start getting a decent chunk of the gameplay logic into the game and zone server, and start playing around with numbers and balancing to see if my game design holds up.
Til next time!