Re-add optional OIDN denoise as an external executable.

[bruvzg]

Implements godotengine/godot-proposals#7640

Alternative variant #82831 (load dynamic library directly).

Pros: can use executables for the different architecture (there's no ARM64 prebuild version for macOS in official releases).
Cons: oidnDenoise is a sample implementation and command line arguments might change in feature releases, temporary images are written/read to disk.

[jcostello]

How can I test this?

[akien-mga]

Suggested change

	The path to the directory containing the OIDN denoise executable.
	The path to the directory containing the Open Image Denoise (OIDN) executable, used optionally for denoising lightmaps. It can be downloaded from [url=https://www.openimagedenoise.org/downloads.html]openimagedenoise.org[/url].

[bruvzg]

I feel like we might need some sort of special interface for all tools like this, like a dedicated "External tool manager" next to template manager. Which present the categorized list of all external tools, checks the current status of the tool (using both user configured location and normal PATH lookup) and maybe provide a button to download and automatically configure each tool.

[akien-mga]

That would be great. The UX currently is really subpar, especially with Blender that throws a warning in new projects because it's enabled by default but the path isn't configured by default.

[akien-mga]

Tested successfully on Linux with the Global Illumination demo.

[Calinou]

Tested locally, it works as expected. GPU acceleration is working correctly on NVIDIA with CUDA libraries installed on the system (confirmed by using watch -n1 nvidia-smi while baking lightmaps).

Testing project: test_lightmap_preview_bake_4.x.zip

No denoiser

JNLM

OIDN

There are more seams visible with OIDN, but the normal information in this scene is kind of broken due to regressions in the OBJ importer I need to bisect (these aren't new in 4.2).

[Calinou]

I forgot to mention: LightmapGI's denoiser_strength documentation should note that it only has an effect when using the JNLM denoiser, not OIDN.

[bruvzg]

Moved the check to the beginning of the bake function and changed the path to oidn_denoise_path.

[akien-mga]

Thanks!

[atafra]

I wouldn't recommend using oidnDenoise for this. It was never meant to be used in production, it's just an example app. Also, you would need to keep writing and loading images from disk, which often would be much slower than actually denoising the image on the GPU. I think dynamically loading the library would make a lot more sense, both in terms of robustness and performance.

[Calinou]

I think dynamically loading the library would make a lot more sense, both in terms of robustness and performance.

See #82831. This is indeed a better approach, but we need precompiled releases with ARM64 macOS libraries to be available for this approach to be viable.

[atafra]

ARM64 macOS builds are planned to be added starting with the next OIDN release.