Please always include units and system specs when sharing benchmarks. But seems like 4x speedup for oneliner? Impressive! Can we fix 2 more lines and have 64x speedup :)

I think this is the proper thing to do (limit leaf size to something more adequate). I'm wondering, will lowering limit to 100 improve the performance further or will it start to degrade the performance?

Lowering this value has a few potential performance issues:

• Slower GPU rendering due to rendering more batches.
• Increased CPU and GPU memory usage due to more shared/duplicated vertices between leaf nodes.
• Higher CPU threading overhead.

So, for measuring performance you not only need to look at this existing failure case, but also check that on models with millions of polygons the GPU rendering and memory usage are still ok. For both multires and regular meshes.

I think 100 will definitely be too low. It means if you have 10 million polygons, you will have 100k GPU batches, which is too many. If we need that fine-grained scheduling for brush operations, we could look at storing GPU batches at inner nodes to cover multiple leaf nodes.

The current limit is set to 10k, but there are still 49398 vertices in a leaf node? How does that happen?

The current limit is set to 10k, but there are still 49398 vertices in a leaf node? How does that happen?

Because in Multires it is using the grids face count instead of vertex count to limit the leaf nodes, so in level 1 there will be always 4 vertices per face. I can change that, but @Julien Kaspar (JulienKaspar) told me that even with this patch the performance in his computer is still the same, so maybe the issue is not entirely related to this.
Also, dyntopo is already using a leaf limit of 100, not sure if there is a specific reason for that.

In D8442#206027, @Pablo Dobarro (pablodp606) wrote:

Because in Multires it is using the grids face count instead of vertex count to limit the leaf nodes, so in level 1 there will be always 4 vertices per face. I can change that, but @Julien Kaspar (JulienKaspar) told me that even with this patch the performance in his computer is still the same, so maybe the issue is not entirely related to this.

I think it is good to use vertex count instead, so that leaf nodes have approximately the same number of vertices regardless of the multires level. Otherwise performance will probably change too much depending on the multires level.

And then on top of that you can tweak the limit for best performance.

Also, dyntopo is already using a leaf limit of 100, not sure if there is a specific reason for that.

I'm not sure either, it seems too low to me regarding GPU and threading overhead.

I update the leaf limit to a more reasonable number after D8454. I still get a consistent performance improvement after this patch. Maybe we can try to commit this to 2.91 and see if it causes other performance problems.