[enhancement] GPU Acceleration

[Asten-Valentinus]

Currently, it goes quite slow when using on large node trees.
I think that to help with this perhaps GPU acceleration should be implemented.

[semagnum]

Do you have a specific implementation in mind? Blender's Python module is pretty limited in what it offers for hardware acceleration. There are separate Python libraries that can do this, but that would require bundling them with the add-on.

If there's specific operations you're thinking of that are slow, we could look into optimizing them. Might be able to do a lot without needing to add external Python libraries.

[Asten-Valentinus]

I noticed the problem with the 'Arrange Nodes' feature, although I see that it also applies to the 'Relax Brush'.

This is an example of a node tree I tried using this on (for reference of complexity):

[semagnum]

Thanks for that. That means it's more about the underlying code than the operators/modals themselves. The code requires nodes checking for collisions with other nodes, so a lot of the speed complexity is around number of (selected) nodes.

With geometry nodes, I've noticed the modal can trigger a viewport update, so if the mesh's geometry nodes and other modifiers are enabled, it can be drastically slower in my experience. Is it faster when these modifiers are turned off?

[Asten-Valentinus]

I tried what you suggested, and it does not speed up the process.

[semagnum]

Okay. I mean it's slow for me too; usually I workaround this by running only on selected (which usually works better with frames anyway).

Before we could even consider hardware acceleration, we'd have to refactor the code to support it. That would likely mean doing all the calculations with numpy. That may be fast enough to not need further acceleration.

[Asten-Valentinus]

Okay. I mean it's slow for me too; usually I workaround this by running only on selected (which usually works better with frames anyway).

Before we could even consider hardware acceleration, we'd have to refactor the code to support it. That would likely mean doing all the calculations with numpy. That may be fast enough to not need further acceleration.

Sounds good. As long as it's optimized I'm happy.