Optimize extract_clusters and prepare_clusters systems

[rafalh]

Objective

When developing my game I realized extract_clusters and prepare_clusters systems are taking a lot of time despite me creating very little lights. Reducing number of clusters from the default 4096 to 2048 or less greatly improved performance and stabilized FPS (~300 -> 1000+). I debugged it and found out that the main reason for this is cloning VisiblePointLights in extract_clusters system. It contains light entities grouped by clusters that they affect. The problem is that we clone 4096 (assuming the default clusters configuration) vectors every frame. If many of them happen to be non-empty it starts to be a bottleneck because there is a lot of heap allocation. It wouldn't be a problem if we reused those vectors in following frames but we don't.

Solution

Avoid cloning multiple vectors and instead build a single vector containing data for all clusters.

I've recorded a trace in 3d_scene example with disabled v-sync before and after the change.
Mean FPS went from 424 to 990. Mean time for extract_clusters system was reduced from 210 us to 24 us and prepare_clusters from 189 us to 87 us.

---

Changelog

• Improved performance of extract_clusters and prepare_clusters systems for scenes where lights affect a big part of it.

[james7132]

Can you run this on some of our more intensive stress test examples? In this particular case, many_lights might be a better test than 3d_scene.

[rafalh]

Can you run this on some of our more intensive stress test examples? In this particular case, many_lights might be a better test than 3d_scene.

Here you are - tracy comparison for many_lights:

I showed data for 3d_scene first (which may be not something that we want to optimize) because it has small amount of lights (just 1) that has entire scene in range. As I understand it it is a pessimistic case where nearly all clusters are in range of a point light causing a lot of Vec allocations.
In many_lights example we have a lot of very small lights (radius = 0.3) so most of the clusters are empty and in such case cloning vectors is not a bottleneck.
I am not sure how much this optimization would affect real projects. I think it may affect some with indoor areas or if someone tries to model sun as a point light. But even if it was for simple examples like 3d_scene I think it is worth merging because of high impact on the FPS. Low FPS (when compared to other engines) in simple scenes may discourage people from using Bevy.

[kristoff3r]

I have hit this is as well multiple times, so it would be great to get it fixed.

I read this pretty thoroughly and it looks correct and like a good approach to me.

[atlv24]

good find!

[alice-i-cecile]

CI failure appears to be a network error; retrying the merge.