The Trees The Fork Maple Day3 - Rust Gpu

Reviving an old project to experiment with a new Neovide renderer

2023-03-25

A branch splits off the main trunk of the Maple and stops abruptly at a stump. A rough bench is carved into the space for climbers to rest at before climbing on. Sitting on the bench is a pile of papers containing research on the nature of the trio and speculation about their origins. Although maybe plausible, the ideas are likely unfounded.

Today I looked into reviving an old project of mine to start playing with Rust-GPU for a Neovide front end replacement.

Skia

When I first started working on Neovide, I did a survey of the available 2d renderers in Rust. At the time, the only options really were druid, pathfinder, and skia. Druid was interesting because the Piet renderer worked in so many places and the promise of Xi editor meant that it seemed like it had some legs. I ended up leaving druid because I was unable to get the window events I needed to make a good neovim front end. Pathfinder was also very exciting as a next gen, rust first solution, but I struggled to get it building on my machine and gave up.

That left Skia which is a 2d renderer maintained by Google and used heavily in parts of Android and Chrome. Its pretty fast, very cross platform, and tested very well on many platforms and in tons of apps including Sublime Text. Right around when I started working on Neovide, a crate was released called Skulpin which wrapped Skia-Safe and handled the creation of the window and vulkan boilerplate making it easy to get something on the screen. This worked great to get things going.

Fast forward a couple years and I think the benefits of Skia are not outweighing the drawbacks. Although the api is relatively simple, the actual performance is hard to predict. The build process for skia is also less than ideal because it isn't well integrated into Cargo. Some platforms struggle to build it because a prebuilt binary isn't available. Even when it builds correctly, the compile time isn't ideal. All of the above issues have lead me to look for another solution.

Rust-gpu

During my time at Zed I grew to appreciate the fact that we owned the entire stack from windowing library to renderer. When something wasn't working, it was always tractable because it was written in relatively understandable rust. Rather than having to dig down a layer or create a PR to some other codebase with unknown constraints, all the moving parts were available for us to look at and consider.

I also discovered that a relatively simple architecture doing pretty basic graphics programming could get you a long way before hitting some performance limit. Until working at Zed I assumed that the advanced techniques described by folks like Raph Levien were necessary to achieve good 2d graphics performance. In reality though nice, if you are smart about how you select what to render and don't rerender unnecessarily, you can get more than good enough performance without that complexity. And then if you need more of those techniques, they are available after the fact.

Enter rust-gpu the successor to an earlier project called rlsl which works to compile rust code directly to a target called spirv. In turn, spirv is a binary compile target for higher level shader languages like glsl and hlsl. By compiling to it, the shaders can be run on many platforms.

I like rust-gpu not only because writing shaders in rust is super cool, but because it enables sharing code between the shader and the application. Rather than writing a type on the app side and an equivalent type in the shader which must be kept in sync, we can write a single type on the shader side and reference that same type in the application. Because the source code for the referenced type is literally the same in both places, we don't have to worry about them getting out of sync. Write once, use on cpu and gpu.

A couple years ago I wrote a ray marcher that explored this idea further. The scene was a complex combination of signed distance fields which described the world and was relatively easily rendered in a full screen pixel shader. The player controller then sampled the exact same code on the cpu in order to let the player walk around the scene and bump into objects in it without phasing through them. The code for the basic game version of that can be found here. I later took that system and tried to create a simple modeling application based on that idea and marching squares to 3d print the sdfs, but I eventually abandoned that effort due to accuracy issues.

Where I'm At So Far

So I spent today working on updating the dependencies of my rust-gpu demo and investigating what it would take to make a simple renderer that was suitable for Neovide to replace Skia with. My conclusion was that we could use the struct replication strategy but instead of using PushConstants like my demo does for the uniforms, use a Uniform Buffer filled with a flat array of instance structs. That Buffer is then bound in the shader and indexed into using the instance id. This way we define the shape of the instance in one place and use it both in the shader and the app code.

The next step with this effort is to create a simple scene object containing the list of quads and potentially layers. As the scene is built up, new quads are pushed into the list and when the scene is complete, they can all be marshalled over to the uniform buffer together and rendered with a single draw call. This pattern can then be extended with glyph atlases to render text, and layers to handle blurs or shadows as necessary. The goal here is simplicity. We aren't going for the fastest possible solution. Just something that can get us off Skia. Then we can profile to figure out what needs more attention from there.

Till tomorrow,
Kaylee