In-browser polygon tesselation

[mourner]

Closes #682. Not ready yet. To do:

• implement holed polygon -> single ring polygon conversion algorithm as described in FIST paper
• implement outer/inner rings classification
• ditch libtess code completely
• fix artifacts on some complex water polygons
• make triangulation code a separate repo https://github.com/mapbox/earcut
• fix fill antialiasing @ansis
• fix fill patterns @ansis
• try out indexed vertices approach and see if it improves performance @ansis
• wait for Polygon winding order & holes vs outer rings mapnik-vector-tile#59 and ditch ring classification code to fix low-zoom water bleeding artifacts
• make sure there are no other glaring artifacts browsing through the map
• fix tests

[mourner]

@ansis you can try out the indexed approach using earcut indexed branch. mapbox/earcut#5 Lets do it in a separate earcut-indexed branch.

Earcut returns {vertices: [...], indices: [...]} object. Vertices is a flat JS array. You can turn it into a typed one with e.g. new Int16Array(vertices).

[jfirebaugh]

@mourner Modulo fixing the tests, is this good to merge as-is, leaving investigation of the indexed approach as a follow-up ticket?

[mourner]

@jfirebaugh no, mapbox/mapnik-vector-tile#59 is a blocker — we need to ditch unreliable ring classification hack, otherwise low zoom levels (<7) are completely broken.

[jfirebaugh]

Is that something we need to worry about for user supplied GeoJSON as well?

[mourner]

@jfirebaugh no, because GeoJSON spec explicitly defines the first ring as outer and other rings as holes for each polygon, and differentiates between polygons in a multipolygon.

[ansis]

@mourner It looks like -native already assumes winding order is clockwise for outer rings and counterclockwise for holes: https://github.com/mapbox/mapbox-gl-native/blob/47966a143176a337733fe63b4121d2d00c48359a/src/mbgl/renderer/fill_bucket.cpp#L98

Is fixing the order of rings all that we need? mapbox/mapnik-vector-tile#59 (comment)

Do we also need better clipping/unioning in the vector tiles?

[mourner]

@ansis we need to try this. It may produce incorrect ring classification sometimes due to degeneracies — points can belong to several rings at once. This problem is solved by doing a union in native, but we can't afford operations like this in JS because they're expensive (probably more expensive than triangulation).

Once we can classify rings reliably, cleaner geometries (union, etc. mapbox/mapnik-vector-tile#53) will be nice to solve minor artifacts on bad data, but not a blocker.

[mourner]

I'll probably squash the PR into 2 commits, the first one and the rest. Such a mess currently after all the conflicting rebases...

[mourner]

-> #1606