Support for AMD FSR 2.0

[MarioPL98]

What problem does this solve or what need does it fill?

Great way to increase FPS, especially in 4k res or above. Open source and supports all gpus.

What solution would you like?

Implementing FSR 2.0 in a way that is simple to use and configure.

What alternative(s) have you considered?

Simple upscaling, however it is much worse in terms of quality.

Additional context

I suspect everyone is familiar with what FSR 2.0 is and how good it is.

[alice-i-cecile]

I suspect everyone is familiar with what FSR 2.0 is and how good it is.

I'm not! Do you have a good link to read?

[nicopap]

I don't think everyone is aware of what FSR is. FSR is an upscalling technology developed by AMD. It's very lightweight yet it results in really good quality image.

There are two versions. From what I hear FSR1 is very easy to implement, but FSR2 requires some more involvement I think. It might require motion vectors.

Here is some discussion on it:

• AMD's slide deck: https://raw.githubusercontent.com/GPUOpen-Effects/FidelityFX-FSR/master/docs/FidelityFX-FSR-Overview-Integration.pdf and https://github.com/GPUOpen-Effects/FidelityFX-FSR2
• Godot discussions: Implement AMD FidelityFX Super Resolution (FSR) in the Vulkan renderer godotengine/godot-proposals#2809 and Integrate AMD FSR 2 in the Vulkan renderer (currently, FSR 1.0 is supported) godotengine/godot-proposals#4280

[bjorn3]

Looks like it indeed requires motion vectors. The source repo seems to be https://github.com/GPUOpen-Effects/FidelityFX-FSR2 which is under the MIT license.

[james7132]

The shading language requirements are potentially possible to use with wgpu/naga. Not sure how much deeper the integration needs to go, or if we need to have speciifc native hardware access to support this.

[superdump]

FSR 2.0 is apparently similar to what they helped Unreal to do for their temporal super resolution TAA/upscaling. It would be good as part of an effort to get some form of TAA working. Even if it would be fun to get a TAA implementation that we would make ourselves to work, using FSR 2.0 working would likely be more effective and efficient.

[Andrewp2]

Shader Model 6 (cs__6_0) has some wave-related functions, which was a sticking point when I was investigating building a LBVH on compute shaders in wgpu (LBVH requires a radix sort, which requires a prefix sum, which uses a decoupled lookback algorithm which AFAIK wgpu currently can't support: https://raphlinus.github.io/gpu/2021/11/17/prefix-sum-portable.html (although the prefix sum doesn't have to use decoupled lookback, it's just faster)). There could be a similar issue here.

After a brief lookover of FSR 2.0, I don't immediately see anything like that. FSR 2.0 is super cool and I would love to get something like that in-engine.

[FraserLee]

I'm also super intrigued by FSR2. Quality wise- it seems to be in the same rough order as DLSS, while actually open-source and available to people like us.

@nicopap From what I hear FSR1 is very easy to implement

Would it be at all worth someone (possibly me) trying to implement this in bevy, just to see how it integrates into the pipeline? Even if we don't have all the requirements (motion vectors, etc) yet, it could still be useful in the interim.

[nicopap]

@FraserLee No idea how difficult it is in practice. I meant by "easy to implement" that everything FRS1 requires is already available in bevy. A first step would be evaluating https://github.com/GPUOpen-Effects/FidelityFX-FSR and the godot implementation, and see if it's in the ballpark of what you can do.

I personally wouldn't feel confident implementing it.

[FraserLee]

Thanks for the leads, I think I’ll give it a shot.

…

On Tue, Aug 2, 2022 at 01:20 Nicola Papale ***@***.***> wrote: @FraserLee <https://github.com/FraserLee> No idea how difficult it is in practice. I meant by "easy to implement" that everything FRS1 requires is already available in bevy. A first step would be evaluating https://github.com/GPUOpen-Effects/FidelityFX-FSR and the godot implementation, and see if it's in the ballpark of what you can do. I personally wouldn't feel confident implementing it. — Reply to this email directly, view it on GitHub <#5232 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHIIGGPISDLI5BKGFO5H25LVXCV3VANCNFSM522QBNZA> . You are receiving this because you were mentioned.Message ID: ***@***.***>