This project is a good starting point for learning how to write compute shaders in Unreal. It implements a boid simulation the GPU. It achieves 0.5 million boids at 45 fps on a GTX 1080.

URzQs0i6RtC1ys7q.mp4

Tweet of the above

This project is for Unreal 4.24. It might work in later versions, but you might need to fix a few compiler errors around shader binding. I would be suprised if my FIBMInstanceBuffer works in 4.26.

The compute shader magic happens in ComputeShaderTestComponent.cpp. To work with compute shaders in Unreal you need a few things:

• the compute shader/kernel source that runs on the GPU (a .usf file)
• a way to map/bind parameters and buffers to the compute kernel (subclass a FGlobalShader)
• something that will dispatch the shader to the GPU (FComputeShaderUtils::Dispatch)

This project was my way of learning to write computer shaders in Unreal, so not everything is perfect, and some things are written for clarity (you should create your TShaderMapRefs once and cache them, I recreate them every frame).

This project is also quite heavily optimized from the theoretical side of things. Some important bits:

• I'm using a hashed grid (powered by the bitonic GPU sort algorithm) to massively accelerate boid neighbourhood queries on the GPU
• I sort the boids by their grid cell every frame to greatly improve cache coherence (it really makes a big difference, try disabling it)
• I rewrote UInstanceStaticMeshComponent to work directly with GPU side position buffers

The hashed grid implementation was inspired by the one from Wicked Engine.