Vulkan: Random meshes vanish in large scenes with lights that have a projector texture

[DAShoe1]

Godot version

4.0.beta3

System information

Windows 11, AMD Vega64 Driver 22.8.1, Vulkan

Issue description

When a scene has a large number of MeshInstance3D nodes (more than 980 in the MRP) and a Spotlight3D or Omnilight3D with a projector texture is added to the scene and moved around, Godot will randomly stop drawing meshes. Reloading the scene will restore them. This occurs within the editor as well as at runtime.

light_projector_bug.mp4

Steps to reproduce

• Add at least 980 MeshInstance3D nodes with a BoxMesh.
• Add a Spotlight3D or Omnilight3D, enable shadows, and assign a texture to the projector.
• Move the light around the scene.

Minimal reproduction project

Light_Projector_Test.zip

[Calinou]

I can confirm this on 4.0.beta 282e50a (Fedora 36, Radeon RX 6900 XT with RADV). This only occurs on Vulkan Clustered, not Vulkan Mobile. Increasing Max Clustered Elements to 2048 in the advanced Project Setting doesn't improve the situation.

After moving the light for a while, cubes start to disappear more slowly. It eventually stabilizes to something that looks like this:

[73v3]

I'm seeing almost identical behaviour in my project on Beta 4 (Windows 10. NVIDIA GTX 2070 super) Vulkan mobile forward_plus.

Moving the player causes random tiled meshinstance3Ds to vanish.

Behaviour stops when I:

• move the tiles further away from player
• turn off the omni light on the player
• drop the number of 3d tiles in the scene.

I'm not using a projector texture. Just a vanilla omnilight3d.

Sorry to barge into your thread but it looks like the same issue and may help clarify the underlying bug.

Project (20Mb, sorry)

VanishingMeshes.zip

[Calinou]

@73v3 Your issue occurs when using Vulkan Mobile, so it is unrelated. Please open a separate issue with the minimal reproduction project attached.

[73v3]

Apologies. Renderer was actually set to forward_plus. Switching to mobile fixes the issue.

[dzil123]

I'm getting this error when closing the editor after observing the behavior:

ERROR: Pages in use exist at exit in PagedAllocator: N33RendererSceneRenderImplementation22RenderForwardClustered32GeometryInstanceSurfaceDataCacheE
   at: ~PagedAllocator (./core/templates/paged_allocator.h:140)

[clayjohn]

It looks like the issue is coming from here:

godot/servers/rendering/renderer_rd/forward_clustered/render_forward_clustered.cpp

Lines 3732 to 3736 in e25d928

	void RenderForwardClustered::GeometryInstanceForwardClustered::set_softshadow_projector_pairing(bool p_softshadow, bool p_projector) {
	using_projectors = p_projector;
	using_softshadows = p_softshadow;
	_mark_dirty();
	}

_mark_dirty() is called which frees the surface data, but then for some reason, the surface isn't being reallocated. This will take a bit of investigation to figure out.

We need to check why the surface isn't being reconstructed in some cases (it appears totally random, so maybe some undefined behaviour has snuck in).

[clayjohn]

Looks like this is caused by the multithreaded culler which is only activated if the number of instances is a scene is greater than 1000. The threshold is configurable, so it can be increased to avoid this issue in the meantime

rendering/limits/spatial_indexer/threaded_cull_minimum_instances can be increased, as long as it is larger than the number of elements in the scene this issue should go away