Be smarter about bind group usage

[cart]

Currently bind groups are re-created each frame to prevent "garbage" from stacking up from removed resources. @kvark has advised me that this approach isn't optimal :smile:

As a short term fix, he suggested just retaining them for X frames before dropping them. We could also make WgpuResources aware of BindGroup resource dependencies. The information is all there, it would just be a matter of storing a RenderResource->Vec<BindGroupId> mapping, then cleaning up bind groups for removed resources.

[cart]

We are now much smarter about this. There are a few very scoped / cheap cases that still create bind groups every frame, but for pretty much all of the important stuff we now cache them.

[superdump]

Hmm. Is that true? This made me want to look through to see what we're doing. I'm going to call out ones that are not being cached:

• queue_shadow_view_bind_group (1 per frame)
• queue_mesh_bind_group (2 per frame)
• queue_mesh2d_bind_group (1 per frame)
• queue_mesh2d_view_bind_groups (1 per 2d view per frame)
• queue_sprites (1 per frame)
• queue_uinodes (1 per frame)

Which is relatively few, but still. :)

[HugoPeters1024]

We also teach our users to recreate bind groups each frame in our examples. e.g.:

https://github.com/HugoPeters1024/bevy/blob/badfa2015e22d93bc40dce0928673601191e580a/examples/shader/compute_shader_game_of_life.rs#L77-L80

It's also not entirely clear to me if there is a canonical solution to that available right now.

[cart]

Hmm. Is that true? This made me want to look through to see what we're doing. I'm going to call out ones that are not being cached:

From the perspective I was speaking from I think it is true. Iirc this issue was written in the context of per-entity / per-asset bind groups being created each frame. We are being way smarter about bind group usage now. Mission accomplished.

I don't think we have any reason to believe that creating a small number of bind groups each frame is a cause for concern. And if it is, I believe that deserves to be a separate discussion.

[davids91]

It is a big concern in case the bindgroups have significant data stored in them, e.g. here: https://github.com/davids91/shocovox/blob/2f5e583ba1465d33fae4be9bb11fad3cded6e383/src/octree/raytracing/bevy/types.rs#L79

The view I implemented stores the Viewport and the renderable data in one bind group. While the renderable data ( octree ) does not change very often, Viewport does. This essentially means that the whole structure is copied to the GPU at every frame which is unnecessary, and is a big performance hit.

I1ve been suggested to set the RenderAssetUsage to RENDER_WORLD only, but that would mean I am not able to change the Viewport .

Is there any best practice here I am missing? Should I open up another issue for this? This seemed relevant to this one..