bevy_solari: RIS for Direct Lighting #19620
Conversation
alice-i-cecile
added
C-Feature
|
It looks like your PR has been selected for a highlight in the next release blog post, but you didn't provide a release note. Please review the instructions for writing release notes, then expand or revise the content in the release notes directory to showcase your changes. |
JMS55
changed the title
JMS55
added
the
S-Needs-Review
| }, | ||
| }; | ||
| use camera_controller::{CameraController, CameraControllerPlugin}; | ||
| use std::f32::consts::PI; | ||
|
|
||
| /// `bevy_solari` demo. |
There was a problem hiding this comment.
I would really appreciate comments on exactly what users need to call to test this.
There was a problem hiding this comment.
It's intended that you just run cargo run --example solari --features bevy_solari, like any other example (it will warn you if you're missing the feature).
The extra options are meant more for me to use for development, and less for users to test.
| @@ -0,0 +1,117 @@ | |||
| #import bevy_core_pipeline::tonemapping::tonemapping_luminance as luminance | |||
There was a problem hiding this comment.
If you want a full approval from me, I'm going to need a bit of a comment-based walkthrough of each of the steps that we're taking inside of this shader. I understand the conceptual approach of restir, but I'm having trouble mapping it to this code because of my shader noobiness.
There was a problem hiding this comment.
initial_and_temporal:
- Load a bunch of data about the current pixel (depth, normal, albedo, etc)
- Call generate_initial_reservoir() to get a reservoir (see the ReSTIR paper/course)
- Create an empty reservoir
- 32 times:
- Generate a random point on a light source (light sample)
- Calculate a resampling weight for that sample, based on the brightness of the light (without tracing any rays to test visibility)
- Add it to the reservoir
- After keeping the best of the 32 samples, trace a ray to test visibility
- Write the reservoir to a buffer
spatial_and_shade:
- Load the same pixel data again
- Grab the reservoir from the buffer
- Calculate the light contribution again
- Shade the current pixel (emissive + (radiance * reservoir weight * brdf * view exposure)).
There was a problem hiding this comment.
Eventually these steps will be expanded to reuse reservoirs temporally and spatially for better quality, but I'll leave that to the next PR.
| let gpixel = textureLoad(gbuffer, global_id.xy, 0); | ||
| let world_position = reconstruct_world_position(global_id.xy, depth); | ||
| let world_normal = octahedral_decode(unpack_24bit_normal(gpixel.a)); | ||
| let base_color = pow(unpack4x8unorm(gpixel.r).rgb, vec3(2.2)); |
There was a problem hiding this comment.
To me, the 2.2 sounds oddly specific and is used twice in this shader, maybe it justifies a comment?
Or is it an "obvious" value if you know the technique?
There was a problem hiding this comment.
I'm not too sure what it's doing, but this code is copied from pbr_input_from_deferred_gbuffer. It's unpacking the gbuffer data, which is in a compressed format.
There was a problem hiding this comment.
pow 2.2 is fake gamma correction
There was a problem hiding this comment.
(this should ideally be an implementation detail of the gbuffer, and access done through a function)
There was a problem hiding this comment.
I am content enough with the code and docs quality here, and have a basic understanding of the rendering algorithm in play :) I also feel strongly that this is the right direction.
That said, please don't merge this without at least one rendering expert review.
Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
JMS55
added
S-Ready-For-Final-Review
|
|
||
| let size = (view_size.x * view_size.y) as u64 * RESERVOIR_STRUCT_SIZE; | ||
|
|
||
| let reservoirs_a = render_device.create_buffer(&BufferDescriptor { |
There was a problem hiding this comment.
Do you really need new buffers every frame? Or could you cache the descriptor.
There was a problem hiding this comment.
It already is cached :)
There was a problem hiding this comment.
Ah, yup, glazed right past that!
| solari_lighting.as_deref().unwrap().clone(), | ||
| SkipDeferredLighting, | ||
| )); | ||
| solari_lighting.as_mut().unwrap().reset = false; |
There was a problem hiding this comment.
Don't feel super strongly about this since we all know these patterns are hard to express but might be cleaner with a First system that sets it to false each frame.
There was a problem hiding this comment.
Ehh this is already what I did for taa. I'd rather keep it consistent.
# Objective - Start the realtime direct lighting work for bevy solari ## Solution - Setup all the CPU-side code for the realtime lighting path (minus some parts for the temporal reuse I haven't written yet) - Implement RIS with 32 samples to choose a good random light - Don't sample a disk for the directional light, just treat it as a single point. This is faster and not much worse quality. ## Future - Spatiotemporal reuse (ReSTIR DI) - Denoiser (DLSS-RR) - Light tile optimization for faster light selection - Indirect lighting (ReSTIR GI) ## Testing - Run the solari example to see realtime - Run the solari example with `-- --pathtracer` to see the existing pathtracer --- ## Showcase 1 frame direct lighting:  Accumulated pathtracer output:  --------- Co-authored-by: Alice Cecile <alice.i.cecile@gmail.com>
Objective
Solution
Future
Testing
-- --pathtracerto see the existing pathtracerShowcase
1 frame direct lighting:
Accumulated pathtracer output:
