Performance Comparison of Meshlet Generation Strategies

Mesh shaders were recently introduced for faster rendering of triangle meshes. Instead of pushing each individual triangle through the rasterization pipeline, we can create triangle clusters called meshlets and perform per-cluster culling operations. This is a great opportunity to efficiently render very large meshes. However, the performance of mesh shaders depends on how we create the meshlets. We tested rendering performance, on NVIDIA hardware, after the use of different methods for organizing triangle meshes into meshlets. To measure the performance of a method, we rendered meshes of different complexity from many randomly selected views and measured the render time per triangle. Based on our findings, we suggest guidelines for creation of meshlets. Using our guidelines we propose two simple methods for generating meshlets that result in good rendering performance, when combined with hardware manufactures best practices. Our objective is to make it easier for the graphics practitioner to organize a triangle mesh into high-performance meshlets. To support this we have uploaded our code to https://github.com/Senbyo/meshletmaker