bevyengine
Mix of HDR and non-HDR cameras breaks rendering
Bevy version
0.9 and main
Relevant system information
AdapterInfo { name: "NVIDIA GeForce RTX 2070", vendor: 4318, device: 7938, device_type: DiscreteGpu, driver: "NVIDIA", driver_info: "517.48", backend: Vulkan }
What you did
Change the split_screen.rs example to add hdr: true on one camera but not the other (doesn't matter which one).
What went wrong
Background becomes black, default clear color (grey) not applied anymore. I both cases (whether hdr is set on left or right), only the right camera shows something.
Testing this on another 4-camera 🎆 Hanabi example, where things animate, I can confirm clearing doesn't work (all particles render on the previous frame) and the HDR camera stops rendering:
Although the latter could be a bug in 🎆 Hanabi, the similarity of artifacts might help diagnosing the issue.
Additional information
I'm guessing the cameras have incompatible render target, and there's no "smart" resolution applying tonemapping only to HDR ones.
It's unclear if this is supposed to work in the first place but logged as bug because @superdump said it should "maybe" be supported.
There are two problems:
- The upscaling render pass unconditionally clears the
RenderPassColorAttachment - The
UpscalingPipelineRenderPipelineDescriptorconfiguresNonefor the colour attachment's blend state
If I instead unconditionally load the colour attachment texture, and alpha blend it, then the desired result is produced. I think the following should be done as a proper solution:
- Clear the out texture only for the initial upscale into it, otherwise load it (or make this configurable)
- Use alpha blending for composing one camera on top of another - not sure whether other options make more sense?
Think I found the same issue.
I'll copy paste my question from Discord:
https://github.com/bevyengine/bevy/pull/10325 should let us implement a proper fix for this now
We don't need clear_color: ClearColorConfig::None in the split screen example anymore.
Also happening on 0.13.1. Here's a minimal example that replicates it.
Tested on M1: AdapterInfo { name: "Apple M1 Max", vendor: 0, device: 0, device_type: IntegratedGpu, driver: "", driver_info: "", backend: Metal }
