Feedback VFX Graph Instancing

Hi Qriva,

Thanks for your feedback!

I'll try to answer your questions:

1. Yes, kind of, let me explain. When we instance a VFX, we put it together with others, sharing their memory buffers. This means that we will run compute shaders (inits, updates, outputupdates) on all those instances at the same time. For draw calls (outputs) we want to do the same, but it is not always possible. However, even when we need multiple drawcalls, it should be faster when instanced, as we can set the material once and do many drawcalls in a row.

2. I think the main criteria for placing blocks should be clarity. There is probably not a big difference, and there are pros and cons. Updates can run more than once depending on time step vs outputs can be rendered in more than one pass, etc. If you have some blocks that could be in either place, the best way to be sure is to test in both places.

3. Yes, anything derived from Object (textures, meshes, graphicsbuffers) that is exposed will disable instancing. This is because we would need to bind different data in the shader, so they can't be run together. In future versions we would like to improve this and group those that belong to the same batch and have the same value for all their exposed objects.

4. This is because the value in exposed textures can change dynamically, so we don't know if they are going to be used or not when we decide to place it together with other instances (at creation). With the future improvement mentioned in 3, empty exposed objects should not break batching anymore.

5. That is a good question. The number there should be 900, all the VisualEffects get updated individually. VisualEfect Update function is multi-threaded, and maybe the profiler counters are only counting the ones running on main thread. If you have 8 cores, 900 / 8 = 112.5 ~ 113. Multi-threaded update was introduced in 2021.2, I think, so it should be available for both versions you are comparing. There are some reasons that can prevent multi-threaded update, like output events or mesh sampling, I don't know which one is causing that for your effect.

6. I think it is really sorting, not alpha blending itself. In those cases, instancing should still be enabled for computes, as explained in 1. Sorting (used by default with alpha blending) requires an indirection that currently makes it impossible to render more than one VFX in the same drawcall. They will still be rendered in a row, when possible.

7. Even if you disable sorting for the particles, transparent objects need to be sorted by the engine and rendered back to front. This means that unity instancing batcher (different from our VFX instancing) will have the last say in which instances can be rendered together. If you arrange your VFX in a different way or rotate the camera around, you should see that number change dynamically.

8. I'm afraid I can't help with that one, maybe someone else in the team will be able to give you more info.

Regarding your suggestion, that should be already there in the version you are testing (but not in the first version). Can you double check?


Finally, all the information here is subject to change in the future. This feature required big changes internally and we will need to tweak and improve some things. That is why your feedback is so valuable, thanks!

Also, I would like to use this opportunity to thank you for all the help you provide in the forums, it is really appreciated

 

Well, in that case it is true that we could ignore it, but at the moment we just look at the exposed variables list, connected or not. If it is not being used, there should be no problem for the user to remove it.

Then I'm not sure why you see 900 updates in the previous version. Maybe the counters worked properly then, maybe the multi-threaded update is disabled in that version.

Ok, I understand now. I would not consider the feature to be partly enabled in that case, though. Compute shaders will work the same as fully batched, and rendering will still be faster than if it was not instanced. There is little action for the user in that case, other than explaining what is going on when looking at a GPU capture.
We will consider your suggestion for future versions, and investigate the best way to show this kind of information.

Once an effect gets compiled, the subgraph becomes part of the graph. We can't batch subgraphs because they don't exist in runtime.

 

Hi!
You need to be in 2022.2 or later to use instancing.
You can find the instancing mode in the Visual Effect Asset inspector.