layer-level opacity

[ansis]

Let's say I want to draw a semi-transparent street layer on top of satellite imagery.

If we just change the line opacity it looks something like this:

Ideally, we can get rid of the overlapping. I'm seeing two possible approaches.

One approach would be to draw opaque lines to a framebuffer, and then copy them over with a different opacity, but this is probably fairly slow (completely untested).

The second approach would be to use the depth buffer to prevent double drawing. This works, but has some bugs:

There are hairline cracks caused by anti-aliasing where the roads connect. The fragment is not opaque, but the depth buffer still gets written preventing the other line from being drawn there. To avoid this the anti-aliasing would need to be drawn later, separately. During the second draw, most of the fragments would be depth culled, so this shouldn't add a fragment shading cost. This change would also need to be made for #77 to be implemented.

The second problem is round linecaps. See the intersection on the right. The only way I see getting around this is adding something to the datatiles to specify whether a road ends in a junction or dead end. Or drawing linecaps/joins separately might fix this problem.

[kkaefer]

Good points. I think we should test drawing a layer to a framebuffer and copying over the result; it might actually be relatively fast. This would also allow us to switch to using GL_POINTS for drawing round line joins, then drawing disconnected line segments on top. This is the way Google does it.

[ansis]

I'll try with framebuffers its probably cleaner and just as fast.

Last week I worked on the other approach. Overall it looks good, but has noticeable defects in some edge cases. There are more draw calls (two for each semi-transparent) but I haven't notice a performance difference.

On a side note, what are your thoughts on using framebuffers to cache rendered tiles for panning?

[ansis]

I implemented this with the framebuffer approach. Seems to be just as fast, and simpler (both conceptually, code, and style-wise), so this seems to be the way to go.

[cyrilchapon]

As of 2018, is that still possible ?

[ansis]

@cyrilchapon it is not possible anymore. The framebuffer approach was removed because it turned out to be a performance trap. It was too easy to create slow maps by adding more than a few layers with layer-level opacity.

[cyrilchapon]

If I hypothetically have a mvt server and a "dissolved coverage" to show, semi transparently. What would be the best approach then ?

I don't need and don't want the cumulative approach (multiplying features opacity), I literally want to have polygon visually dissolved runtime

[vin-ni]

Still no solution to this?

[waissbluth]

also wondering about the status of this -- it would be very useful to manage the transparency of a layer that has overlapping features

EDIT: answering myself #4090