Use AMD Compressionator for faster texture compression

[julhe]

I'm working with large textures and found the compression times to be extrem long, usually up to multiple minutes. Basis Universal is slightly faster, but is inferiero in quality and loading times.*

Using https://github.com/GPUOpen-Tools/compressonator and the encoder only takes around two seconds for the same texture in BC7, which is obviously a massive improvement!**

So, from my perspective, it would be usefull to integrate the Compressonator SDK into Godot, since that means super super fast encoding for pratically all used formats. Windows, Mac and Linux are also supported and the license so far doesn't look limiting.

I could implement this myself, but before starting, I would like some feedback. 🙂

---

*By the way: That's because Basis in Godot is using UASTC. UASTC still needs to be translated into an actual GPU format, thus adding loading time, see https://github.com/BinomialLLC/basis_universal/wiki/How-to-benchmark-or-directly-call-the-UASTC-functions-without-using-.basis

**That speed up is due to thier encoder being GPU accelerated via OpenCL. A fallback to MT-CPU exists. So, since Godot texture encoding is single-threaded, compressonator would still be faster, even if there is no compute support.

[Calinou]

I understand the sentiment, but I'm not sure if switching to Compressonator is the best thing to do.

In 4.0.alpha, Godot already uses etcpak (ETC2) and libsquish (S3TC) for texture compression, both of which are very fast. (To my knowledge, there is no CPU encoder beating etcpak right now.)
Resource importing is also multi-threaded, unlike in Godot 3.x. This means that even if each texture is imported with a single thread, multiple textures will be imported at the same time, which spreads the load over multiple CPU threads.

The issue with BC7 compression is that not only it only works on desktop platforms, but it doesn't work on macOS as Apple has never supported OpenGL past 4.1. Godot 3.x does have BPTC support when using GLES3, but it's rarely ever used due to this limitation. I don't remember if this support is still present in 4.0.alpha.

Also, my concern about the Compressonator SDK is that it looks bloated at a glance – much more so than what we currently have. Remember that the Godot editor can run on a lot of platforms, including Android, ARM Linux/macOS and HTML5. Does Compressonator build on those platforms? We also want to keep build times and binary sizes reasonable.

In the far future, using a GPU-based compressor like Betsy is planned, but not for 4.0. Betsy uses Vulkan compute instead of OpenCL or CUDA, so it doesn't suffer from poor driver support.

PS: To speed up VRAM-compressed texture importing, disable ETC2 import in the Project Settings if you don't target mobile devices in GLES3, and disable ETC import if you don't target mobile devices in GLES2. This works in both 4.0 and 3.x.

[julhe]

Yes, after some digging I share your concerns about bloat and incompatability. So I guess integrating Compressonator directly is off then.

I think however there is still room for improvement, at least for the BC7 encoder, which it's currently single-thread. (Other encoders as well, but they are fast). I guess I'll look into making it multi-threaded then, hope that doesn't interfere with the parallel asset import?

Alternatetively, just an idea, the transcoded output from Basis Universal could be cached to bring the loading times on pair with VRAM Compressed. But I don't know enough about Godots loading mechanism if that makes any sense or is difficult to implement.

[Calinou]

I guess I'll look into making it multi-threaded then, hope that doesn't interfere with the parallel asset import?

This might spawn a lot of threads if importing a lot of textures, therefore reducing performance. It'd have to be made parallel only when there is one texture to import at once (to speed up reimporting of a single large texture while iterating).

[fire]

Compression times are not extreme long anymore due to betsy support. This discussion can be closed?

[Calinou]

Indeed, this has largely been addressed now thanks to Betsy integration.