Add hit-detection support for WebGL

[tsauerwein]

This PR implements forEachFeatureAtPixel for WebGL. From the idea it uses the same technique as the canvas renderer: We have a 1x1px framebuffer in which we draw feature-by-feature (in reverse order) and check if there is some color.

Commit e01e556 uses an optimization, that should also be used for the canvas hit-detection: Using the new renderBuffer we build a box around the given position. And then we only check those features that intersect this box. This drastically improves the performance for WebGL.

Speaking about performance, canvas gives better results. I made some tests showing 1000, 5000, ... features on a map (zoomed-out so that really all features are shown). Then I measured the time to do a forEachFeatureAtPixel call.

# of features   Canvas    WebGL
1000            10 ms     20 - 80 ms
5000            35 ms     70 - 350 ms
10000           60 ms     140 - 650 ms
20000           120 ms    340 - 1400 ms
50000           320 ms    800 - 3000 ms 

For WebGL I have given a range, because it depends on where you click with the optimization made in e01e556. If you click somewhere in the map where are no features, it's faster.

But still, for WebGL, it's too slow to run forEachFeatureAtPixel on mousemove. We came up with a hasFeatureAtPixel method, for which I will open a separate PR (here is already an example to give you an idea).

[elemoine]

The above, texture-related, code should probably shared. It is not specific to hit detection. If we change the filtering parameters for the rendering to the screen I don't want to have to change some other code for the hit detection.

[tsauerwein]

Agree.

[elemoine]

I am looking forward to reviewing this in more detail, but it looks great already! Thanks a lot.

[elemoine]

The groupEnd variable is not used. And it looks like the linter (or compiler) is not able to catch this when the variables are all defined on the same line.

[tsauerwein]

Thanks, fixed with tsauerwein@abb5fad

[elemoine]

This looks very good to me. I just added two comments, one about an unused variable (groupEnd), and one suggesting to factor out some WebGL texture-related code.

[elemoine]

One thing: we need to implement "skip feature", but I think this can be done as a separate PR.

[fredj]

The ? is not needed (untested)

[tsauerwein]

In fact it is required, because ol.layer.LayerState is only a typedef.

[elemoine]

This looks good to merge to me.

Any objections?

[fredj]

LGTM

[gberaudo]

+1

[tsauerwein]

Thanks for the reviews!