Double reported verts without OptimizeReorderVertexBuffer

I’m seeing the following problem in my project:

I’m combining multiple meshes together in order to improve performance. That in fact reduces draw calls and improves FPS but if the mesh has multiple submeshes then the stats panel shows double number of vertices (but the correct number of tris).

If I use OptimizeReorderVertexBuffer, the problem goes away but the method has a relatively high performance overhead.

So I would like to understand what causes the doubled number of vertices in the stats window and is there anything else I can do in how I combine meshes to prevent this from happening.

My suspicion is that Unity (and the GPU) prefers submeshes to be contiguous areas in the vertex buffer. Is that correct? Does anyone have real-world experience in how much that affects performance?

Also am I right assuming that the stats window reports “Verts” as the total size of vertex buffers and not the number of actually rendered buffers / references in the index buffers?

For the record: This is with shadows and depth buffer disabled, so there isn’t anything that would inflate the stats (and if it did it would also inflate the number of tris).

For anyone who’s wondering: After some further benchmarking it appears that it doesn’t affect performance in my case. So it might have some negative effect in other situations (I guess when submeshes are highly “scattered” in small fragments across the buffer) but in my case, where each portion of the buffer is relatively large, it did not noticeable affect the FPS.

So the number reported by the stats window did not really have any real-world consequences.

Let me know if someone has other experiences or thinks I’m missing something important here such some other negative consequences of “unordered” vertex buffers.

Does anyone know what term I should be googling to find some more information about this? I don’t mean in Unity but in DirectX or GPUs in general. Something like “vertex buffer fragmentation”?

Vertex cache optimization.

https://tomforsyth1000.github.io/papers/fast_vert_cache_opt.html