Use default samplers in base uniform set when rendering to reflection probes

[clayjohn]

Fixes: #83962

The RenderBuffers used by ReflectionProbes doesn't create and maintain its own set of samplers. So basically what was happening was the Viewport RenderBuffers were invalidating the base_uniform_set (so it would get recreated with the new samplers). But then the ReflectionProbes would get rendered, and they don't have valid samplers, so they would give invalid samplers and the base_uniform_set would be null. Then rendering would fail for a frame. The Viewport would then get rendered and update the base_uniform_set to be valid. Then, in subsequent frames, the ReflectionProbes would just use the updated base_uniform_set.

This PR works, but I think it will bring us back to the state where the sampler gets ignored when changed (due to how the caching works). I need to do a bit more testing before this is ready.

[clayjohn]

It seems that we are going to need to implement a base_uniform_set cache. Right now the first Viewport to set a base_uniform set wins and its settings are used for everything else.

[clayjohn]

This iteration fixes #83962 and an unreported bug whereby the wrong bias is used.

However, there is still an existing unreported bug (#84083) that will require more invasive changes. Right now we maintain a render_base_uniform_set which should only hold data that doesn't change between draw calls. However, since we made the samplers change per viewport/render type, that assumption no longer holds true. render_base_uniform_set can change depending on the type of rendering going on (shadows, probes, scene draw) and per viewport.

This PR splits render_base_uniform_set into two, one for viewports (where bias is not default) and one for everything else (where bias is always default). This fixes the most obvious bug, but it will stop working if multiple viewports are used in a scene and the Viewports do not all have the same bias setting.

A perfect solution would be one where we either store the samplers in another location (one that is updated more frequently), or we store the render_base_uniform_set on a per-viewport basis with a default implementation for non-viewport rendering. That, however, would be a much more invasive change than this, so it might be best to hold that off for 4.3

[Calinou]

This iteration fixes #83962 and an unreported bug whereby the wrong bias is used.

Judging by your description, this may also fix #65690 🙂 I'll give it a try.
Edit: I tested the MRP linked in that issue, and this PR doesn't appear to fix that issue.

This fixes the most obvious bug, but it will stop working if multiple viewports are used in a scene and the Viewports do not all have the same bias setting.

I assume this is connected to #84083.

[Calinou]

Tested locally, it works as expected.

[clayjohn]

This iteration fixes #83962 and an unreported bug whereby the wrong bias is used.

Judging by your description, this may also fix #65690 🙂 I'll give it a try. Edit: This PR doesn't appear to fix that issue.

That's a different issue, ya.

This fixes the most obvious bug, but it will stop working if multiple viewports are used in a scene and the Viewports do not all have the same bias setting.

I assume this is connected to #84083.

That's right. I hadn't seen #84083 before. It is exactly what I was describing

[akien-mga]

Thanks!