Fix voxel GI issues

[RandomShaper]

UPDATE: To avoid multiple 'update' notices throughout, just know that the full text has been revised after enhancing the fixes.

---

Changes kept in two commits because, even if related, they address different issues:

• Commit 1: Fix breakages of volumetric fog on voxel GI changes

  Fixes the book-keeping of related uniforms. Without this change, some of them could fail to be created when needed and some others could be pointing to outdated voxel GI buffers. For the uniform sets related to GI instances, a minor sets group idiom is introduced, that helps tracking validity and saves some checks, while also assisting developers in detecting bugs.

  Fixes Vulkan: Volumetric Fog breaks when number of visible VoxelGI nodes changes #69565.
  Fixes Vulkan: Volumetric fog glitches out in builds that use optimize=none, but not other optimization levels #76243.

• Commit 2: Fix unsupported sampler filter used for voxel GI

  Fixes a validation error* about an R8G8 texture being sampled linearly when the driver reports no support for that.

*:

drivers\vulkan\vulkan_context.cpp:267 - VALIDATION - Message Id Number: 188609398 | Message Id Name: VUID-vkCmdDispatch-magFilter-04553
Validation Error: [ VUID-vkCmdDispatch-magFilter-04553 ] Object 0: handle = 0x8f7057000000562c, name = RID:505478996034037, type = VK_OBJECT_TYPE_DESCRIPTOR_SET; Object 1: handle = 0xa182620000000079, name = RID:163208757251, type = VK_OBJECT_TYPE_SAMPLER; Object 2: handle = 0xf325e500000055e3, name = RenderBuffer forward_clustered/voxel_gi View, type = VK_OBJECT_TYPE_IMAGE_VIEW; | MessageID = 0xb3df376 | Descriptor set VkDescriptorSet 0x8f7057000000562c[RID:505478996034037] encountered the following validation error at vkCmdDispatch time: Sampler (VkSampler 0xa182620000000079[RID:163208757251]) is set to use VK_FILTER_LINEAR with compareEnable is set to VK_FALSE, but image view's (VkImageView 0xf325e500000055e3[RenderBuffer forward_clustered/voxel_gi View]) format (VK_FORMAT_R8G8_UINT) does not contain VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT in its format features. The Vulkan spec states: If a VkSampler created with magFilter or minFilter equal to VK_FILTER_LINEAR and compareEnable equal to VK_FALSE is used to sample a VkImageView as a result of this command, then the image view's format features must contain VK_FORMAT_FEATURE_SAMPLED_IMAGE_FILTER_LINEAR_BIT (https://vulkan.lunarg.com/doc/view/1.3.216.0/windows/1.3-extensions/vkspec.html#VUID-vkCmdDispatch-magFilter-04553)
	Objects - 3
		Object[0] - VK_OBJECT_TYPE_DESCRIPTOR_SET, Handle -8110887271382624724, Name "RID:505478996034037"
		Object[1] - VK_OBJECT_TYPE_SAMPLER, Handle -6808771934491246471, Name "RID:163208757251"
		Object[2] - VK_OBJECT_TYPE_IMAGE_VIEW, Handle -926082360191986205, Name "RenderBuffer forward_clustered/voxel_gi View"

[RandomShaper]

Specifying the format here feels too hard-coded to my taste, but I couldn't find a way of querying the format of the actual texture. This is not the end of the world, but has a risk of getting out of sync with the code that creates the texture.

[Calinou]

Fixes a validation error about an R8G8 texture being sampled linearly when the drivers reports no support for that. Besides the validation error, without the appropriate fallback to nearest, when looking at the scene from certain angles, the volumetric fog starts flooding the whole viewport with white in a few frames and there's no way back other than resizing the viewport.

Does this fix #76243? I thought it was due to compiler optimization levels but that may have been a fluke.

[RandomShaper]

Does this fix #76243? I thought it was due to compiler optimization levels but that may have been a fluke.

Oh, that's exactly what one of the commits addresses! Not completely sure if its the same root cause or it's instead a different bug that looks like the same. I'm being optimistic and choosing to believe it's the same, and so updating the PR description.

[RandomShaper]

Hmm... I've been able to reproduce the fog glitch once. I'll investigate a bit more before concluding this PR doesn't fix it.

[RandomShaper]

I think I've found the glitched fog issue! Will finish updating the PR tomorrow. Tagging as Needs work.

[RandomShaper]

OK. I finally think this is good for reviewing. And fixes the flood-of-white-or-black fog issue! I've overhauled the PR description.

[RandomShaper]

I think this block of checks would belong better to just before using the uniform sets. They will be invalid by that time because of deletion of dependencies, so doing the check there is to me more correct. The idea would be being reactive to what actually happened instead of being proactive in guessing what may have happened. That, of course, without detriment of being proactive as an optimization in cases where that can save some checks (a bit like this PR does by assuming the whole group of sets is either valid or invalid).

In any case, this is a suggestion to the rendering maintainers. I didn't want to attempt something like that in this PR, which would broaden its scope needlessly.

[clayjohn]

That makes sense. There isn't much of a reason to be caching this so early

[RandomShaper]

For the records, I've still seen the floody fog glitch once even with this changes, in the D3D12 renderer. But I believe this PR fixes it like 99% since without it I've been able to reproduce the issue consistently if I force voxel_instances_changed = true always.

[clayjohn]

This looks good as a minimalistic approach to fixing the issues. Indeed clearing those uniform sets should be moved out of the gi_setup code as it creates a weird dependency between the two classes.

[akien-mga]

Thanks!

[akien-mga]

Cherry-picked for 4.0.3.