gameoflife.mp4
This project was made during the dots-challenge https://itch.io/jam/dots-challenge-1.
The core concept I was trying to go for was to render quads per cell. I could have opted to put everything into textures but I wanted to go for the individual mesh approach.
My initial attempt can be found in the MeshInstancedRenderingSystem which used the old unity API Graphics.DrawMeshInstanced, but that bottlenecked at around 1024x1024 cells on my machine. The main culprit for that is the transfer of render matrices and mesh data on a per-frame basis. Also, this API was only able to render about 500 instances per batch so it also required a lot of batches for big grids which also slowed down framerate by a lot. At this point, I thought that this would be the maximally achievable performance, but then I thought to myself there must be a way to skip the transfer of the localToWorld matrices every frame and to render more than 500 instances.
I searched the internet and asked some people on Discord how to improve the rendering performance. A person on Discord immediately came to my help and showed me that there is an API called BatchRendererGroup (BRG) that enables you to have more low-level access to rendering and is even compatible with Burst. This API allows you to persistently upload GPU data and access data per instance. I even found a project by Unity that used it (https://github.com/Unity-Technologies/brg-shooter). Using this API I was able to remove the GPU bandwidth bottleneck completely by only uploading the changing cell state and keeping everything else constant. This enabled me to double the grid size to around 2048x2048. At this point, my GPU became a bottleneck once again, but this time it seemed to be the amount of vertices that needed to be rendered. For this grid size, it is required to draw 16 million vertices using the quad-per-cell approach, which apparently was too much. the code for it can be found in BRGMeshInstancedRenderingSystem.
The current approach uses 4 bytes to represent the cell state which only essentially requires 1 bit. This is a lot of wasted space thus I decided to pack 32 cells inside these 4 bytes. To make it work it required creating a custom shader that unpacks the data in the fragment shader. I achieved that by adding a custom shader graph node (see MultiCell.hlsl) which uses the UV coordinates of the mesh to subdivide the quad into multiple smaller ones.
This greatly helped boost the rendering performance since it can now render 32 cells per quad, essentially reaching grid sizes of 4096x4096.
I feel like I could be getting to about 8Kx8K, but this would require rewriting the chunk data access to be more efficient and investigating the auto-vectorization of Burst. Being quite satisfied with the current result and the amount of work that has gone into it I will leave it at that for now.
Final performance tested with a Ryzen 7800X3D with an RTX 3070 achieving about 30fps on a grid of size 4096x4096