Change blending options in attempt to fix overlapping billboard fringes. #5066
Conversation
|
Here's an example of what I'm talking about with labels sticking into mountains. I know the text looks crowded and bad, but there are not supposed to be any holes in the glyphs. Cesium 1.27 is too old to load this format of gist. Cesium 1.28 needs to be manually set to "Natural Earth II" base map, but then it works, and it looks perfect: Cesium 1.30 exhibits the label depth strangeness, which means it is not introduced by this branch:
So, it looks like this is a separate, but likely related, issue. |
Agreed. Can you submit a separate issue for this? |
Source/Scene/BillboardCollection.js
Outdated
| @@ -1465,7 +1465,8 @@ define([ | |||
| this._rsTranslucent = RenderState.fromCache({ | |||
| depthTest : { | |||
| enabled : true, | |||
| func : WebGLConstants.LEQUAL // Allows label glyphs and billboards to overlap. | |||
| // Allows label glyphs and billboards to overlap. | |||
| func : ((this._blendOption === BlendOption.TRANSLUCENT) ? WebGLConstants.LEQUAL : WebGLConstants.LESS) | |||
There was a problem hiding this comment.
What does this._blendOption === BlendOption.TRANSLUCENT really mean? Does it mean "this billboard collection is for a label collection with backgrounds?" If so, can we just asked the label collection or have the label collection pass in a private flag so this code is more obvious?
There was a problem hiding this comment.
Good eye, but no. this._blendOption === BlendOption.TRANSLUCENT means that the entire billboard is being written in the translucent pass, without an opaque pass. You'll notice on #L1471 that we already use the same test to determine whether or not to write to the depth buffer. This is because, when the opaque pass is NOT in use, we rely on the depth buffer to save us from OIT making everything blend with everything, even if it means "punching holes" in things. When the 2-pass system is in use, we're avoiding the "punching holes" bug by using the depth buffer only in the opaque pass and not the translucent pass.
So, the difference between single-pass and 2-pass controls a difference in strategy of how we use the depth buffer, and thus should also control the depth test function. In single-pass mode we use LEQUAL because the depth buffer is in use and we need to overlap fragments at the same depth. In 2-pass mode we use LESS to avoid the fringe overlap problem.
Basically, with this change, single-pass mode has returned to the behavior from before the "punching holes" fix, making it fully backwards-compatible.
There was a problem hiding this comment.
OK, can you improve the Allows label glyphs and billboards to overlap comment to better explain this?
|
Also, update CHANGES.md. |
|
New issue #5083 submitted for the label depth problem mentioned above. |
|
Does anyone else want to look at this? @bagnell? |
|
Updated. |
|
OK with me. Does anyone else want to test this out? I labeled this |
|
Now that #5099 was merged to master, and I merged master in here, this fix is looking better than any of the screenshots posted originally.
|
|
This looks good to me. |
|
This is ready. |
If this works, it should fix #5007. Hopefully it whacks this mole without other ones popping up.
To see old behavior, compare GeoJSON in 1.27 to GeoJSON in 1.30. The newer version is on the right:
This branch offers this:
If you look closely at the 1.27 shot, you can see in the "rounded" corners of the billboard where the old bug was "punching holes" in the ones behind, making the corners appear more squared off. In the 1.30 shot, everything is messed up, but the "punching holes" bug at least is gone.
On this branch, this demo appears more orderly.
There's a slight issue remaining, where low-alpha fringes of the inset icons punch holes in the background color of the same icon. This may actually be a bug with the alpha values coming out of the PinBuilder, since the output billboard should be calling for opaque fragments there in the interior, regardless of the inset graphic.Edit: Issue #5097, fixed in #5099.Anyway, I tested this with label backgrounds, and with non-backgrounded labels. It seems to mostly work. There's some strangeness where a non-background label that sticks into terrain will have its opaque fragments occluded at a slightly different depth from the translucent fragments. Edit: Issue #5083. But mostly I think this is a big improvement.