Label and Billboard depth test regression (1.135)

[ppoulainpro]

What happened?

Hello;
Release 1.135 has changed the billboard depth test logic from automatic to manual (Switch from automatic to manual depth testing for billboards #12821).
Unfortunately, it introduces some regression when using label or billboard within entities.

Best regards.
Pascal

Cesium 1.135:

Cesium 1.134

Reproduction steps

See sandcastle in attachment.

Sandcastle example

https://sandcastle.cesium.com/#c=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

Environment

Browser: Chrome,Edge
CesiumJS Version: 1.135
Operating System:W11

[jjspace]

Thanks for the report @ppoulainpro, I can confirm I see the issue.

@mzschwartz5 any ideas on this? I think you're the most familiar with the picking changes recently

[mzschwartz5]

(edited)

It looks like this has to do with translucency, which is definitely an area I was wary about messing up, and clearly I have some work to do there.

Debugging in Spectorjs, I see that at one point in a given frame, the labels render correctly:

And even with the plane rendering on top, afterwards:

But then these outlines are rendered in a separate, translucent pass, and because the manual globe depth texture wasn't updated after the opaque pass, these translucent fragments just render on top of anything from that pass.

[mzschwartz5]

The fix here may be as simple as updating the globe depth texture after the opaque pass so that the translucent pass has that info in the billboard fragment shader.... of course, we'll have to consider if that has any unintended consequences.

[mzschwartz5]

(skip to next comment)

After thinking about this more, I think it's a little more complicated. Since the billboard pass's manual depth testing reads from one depth texture (globe depth) and writes to another (standard depth target), that also means anything drawn in that pass will not depth test against anything else drawn in that pass.

I think (a posssible) solution is a bit more complicated:

1. Turn back on automatic depth testing for billboards
2. Split off the opaque labels-and-billboards pass from the opaque pass into its own pass.
3. Instead of the standard depth target, create a new depth target for that pass
4. Then after, merge that new depth target into the frame's standard depth target.

(The billboard's fragment shader will stay the same - doing manual testing against the globe depth / depth plane as needed).

This is... more complicated than I like, and adds some performance overhead with copying the depth targets. Also, the same treatment may not work for the transparency pass... will need to give this some more thought. It's not good news though :/ may require going back to the drawing board (once again) on billboard depth issues.

[mzschwartz5]

Still thinking, here's another idea that's simpler and (I think) more likely to work:

1. Turn back on automatic depth testing for billboards.
2. Keep the manual depth testing as well.
3. The depth that gets written out should be artificially set to at least the globe depth (plus some tiny increment), so that the billboard doesn't get clipped by the globe (the original issue).

So we get to keep our manual depth testing logic, but also get the benefits of the automatic testing, while tricking the automatic test into not clipping our billboard badly.

[mzschwartz5]

I have a proof of concept of the above working. It's a pretty simple fix in the end. Will raise a PR at some point this week and aim to have it patched in the next release.

[mzschwartz5]

Another instance of the issue in the community forums here. Let's try to get the PR in the next release.

[ppoulainpro]

I have done some test of the PR with use cases which require labels and billboards (dynamic entities, vctr with styling). It seems ok.
Thanks for your action.