Opaque objects that share a mesh and a look are drawn as one instanced draw call. The renderer sorts the frame’s objects by a key (RenderView.MakeKey): the mesh, then the material if the renderer has one, otherwise its texture, flat shading, unlit and glow. Each run of equal keys is one draw.
- Metallic and roughness don’t split batches. Renderers without a material carry their own metallic and roughness in each instance. They go in the bottom row of the instance’s world matrix, which is always (0, 0, 0, 1); the shaders rebuild it (
vs_mesh.sc,vs_shadow.sc,vs_depth.sc). So 10,000 cubes with 10,000 different metallic/roughness values draw like one material: 1.5 ms, down from 14.9. - See-through objects are drawn back to front. Neighbours in that order that share a key are one instanced draw. Instances blend in order, so the result is the same. 1,000 see-through cubes: 1.6 ms to 0.3 ms.
- Culling first: objects out of view that cast no shadow are skipped before their reflection probe and batch key are worked out.
- A different texture, material, reflection probe, lightmap or mirroring (negative scale) starts a new batch.
A custom shader built from the fs_mesh.sc / vs_mesh.sc templates gets the instance values in v_surface (x metallic, y roughness, z 1 when present). Draws with a material always use the material’s uniforms.
This guide is docs/rendering-batching.md in the engine’s repository. Spotted a mistake? Open an issue.