Solari radiance cache

[JMS55]

Objective

• Make GI pass cheaper
• Improve GI quality after disocclusions
• Have multibounce GI instead of the current single bounce

Solution

• Add a persistent world cache based on spatial hashing
• Each hash key is a combination of position, normal, and cell size, where the cell size is based on the distance from the camera

1. Each entry/cell in the cache has a "life" counter. It starts at 30, and the first pass each frame decrements each entry by 1. When it reaches 0, the entry is deallocated.
2. Passes 2-4 compact and write a list of the active cells (life != 0) in the cache via parallel prefix sum
3. Pass 5 samples lighting for each cache cell (direct via light sampling, indirect via BRDF sampling and querying the cache at hit points)
4. Pass 6 blends the new radiance into the existing cell's radiance. Unlike ReSTIR/RIS we are not storing a single light sample/path. We are instead storing radiance.
5. The ReSTIR GI pass is basically the same, except at the sample point, instead of sampling direct lighting via another raytrace, it instead queries the world cache. Querying a cache cell resets its life to 30.

Issues

• Querying an empty cache cell initializes the position/normal used for sampling lighting for the cell to the values used in the query. This means that each cache entry has a fixed sample point to represent the entire volume, which can sometimes lead to problems if the chosen sample point poorly represents the volume.
• When transitioning LODs, cells start out as empty rather than initializing themselves via querying adjacent LOD levels. Not a huge problem in practice though, at least in the cornell box I was hard pressed to notice this as ReSTIR GI temporal reservoirs hide the issue. Yay for two level caching!
• I'm not entirely sure that cache entries will ever die in larger worlds. The cache can keep itself alive since cells can query each other (resetting their life to 30) when updating cells. Probably need to think about how to handle larger scenes.
• Worse stability in larger scenes like bistro. The jacobian leads to temporal bright spots with ReSTIR GI more often than not using the cache, and the cache itself can be very unstable. Might need some form of ReSTIR in the cache, like https://wangningbei.github.io/2023/ReSTIR_files/paper_ReSTIRGI.pdf. (EDIT: Added RIS which helps a ton. ReSTIR might be desirable for better responsiveness though)
• Lots more tweaking of hashing stuff possible.
• Like the rest of Solari, I haven't done any testing of dynamic content yet. Gotta get static content working first :)
• The pipeline setup code for Solari has gotten super messy. Ideas on how to fix it are welcome.

---

Showcase

[alice-i-cecile]

This is not a blocker for 0.17; moving milestones :)

[SparkyPotato]

Is the + 0.0001 meant to stop a div by 0?

[JMS55]

Nah, it goes after the division. It just kinda bugs out without it, not entirely sure why.

[SparkyPotato]

We should have some sort of util function for this...

Also, this should be moved to sampling.wgsl so we can share it with restir_di.wgsl

[JMS55]

Maybe? I haven't decided how I want to unify the RIS stuff yet. For instance this one dosen't use BRDF.

[atlv24]

I think Sparky was talking about the hash thing