Skip dynamic cluster resizing when GPU clustering is active

[holg]

Objective

Fixes #23438

The dynamic resizing code in assign_objects_to_clusters compares last frame's cluster index count against
ViewClusterBindings::MAX_INDICES (16384), which is the UBO limit. When GPU clustering (#23036) is active, storage
buffers are used and can hold far more indices. The resizing shrinks the grid unnecessarily, causing thrashing and
flickering.

#22874 added dynamic resizing before GPU clustering existed, so the UBO limit was correct. When #23036 landed, the
runtime index path was updated to skip the limit when storage buffers are available (line 500 in cluster/mod.rs), but
the dynamic resizing path was not.

Solution

Skip the dynamic resizing when global_cluster_settings.gpu_clustering is active.

// Skip when storage buffers are available
if config.dynamic_resizing()
    && !global_cluster_settings.supports_storage_buffers
    && let Some(last_frame_cluster_index_count) = ...

Testing

• Tested on Apple M2 Max (Metal), NVIDIA 5060 Ti (Vulkan), Intel i360P (Vulkan)
• Also tested via WASM/WebGPU in Chrome on macOS and Windows
• Flickering resolved in scenes with many clustered lights
• No change when GPU clustering is disabled — existing UBO path unaffected

[Copilot]

Pull request overview

This PR prevents clustered-light grid “dynamic resizing” from running when GPU clustering is enabled, avoiding unnecessary cluster-grid shrinking based on the uniform-buffer index limit and reducing flicker/thrashing in heavy clustered-light scenes.

Changes:

• Add a guard to skip dynamic resizing logic when global_cluster_settings.gpu_clustering is active.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The dynamic resizing guard is currently tied to gpu_clustering.is_none(), but the index-list limit it’s trying to avoid (ViewClusterBindings::MAX_INDICES) is only enforced when supports_storage_buffers is false (see crates/bevy_pbr/src/cluster/mod.rs around the n_indices() >= MAX_INDICES && !supports_storage_buffers check). On devices where compute shaders are unavailable (so gpu_clustering is None) but storage buffers are supported, this will still shrink clusters unnecessarily. Consider gating dynamic resizing on !global_cluster_settings.supports_storage_buffers (or otherwise using an “effective max indices” value that reflects whether storage buffers are in use).

Suggested change

	if config.dynamic_resizing()
	if config.dynamic_resizing()
	&& !global_cluster_settings.supports_storage_buffers

[holg]

changed to :
// Skip when storage buffers are available
if config.dynamic_resizing()
&& !global_cluster_settings.supports_storage_buffers

[pcwalton]

Going through my review queue. Is this still relevant/should I review it?

[holg]

Yes, still relevant:
the issue is still present on current main: the dynamic resizing block in assign_objects_to_clusters (crates/bevy_light/src/cluster/assign.rs) still shrinks the cluster grid against the 16384 UBO limit unconditionally,
while the runtime index path in the same file already skips that limit when storage buffers are available.
So with GPU clustering active the grid still gets resized down needlessly, causing the thrashing/flickering described in #23438.
as well it would help my bevy related crates to be pushed onto crates.io (eulumdat-bevy and gldf-bevy) , as i am relying on this... (right now they need my own fork, so i cannot publish)

[pcwalton]

I’m a little unsure about whether we shouldn’t have any limit, because clusters do have a cost, but I guess this at least makes it consistent.

[pcwalton]

Suggested change

	// Skip when storage buffers are available
	// We don’t need to do this if storage buffers are available, because we have plenty of space in that case.