Switch to tesselated fill rendering with earcut

[mourner]

A fresh recut of #948. Had to squish into one commit because all previous branches were a big mess. cc @jfirebaugh @ansis @kkaefer

fixes #2269
fixes #682

[mourner]

All good except one fill pattern render test! Rendering is super-blazing-fast.

At first I thought some sliver lines I saw were earcut artifacts, but they turn out to be present on master as well:

[mourner]

The fill-pattern/zoomed test looks very good actually, and there is definitely something fishy about the diff algorithm. Should we add it to ignore list @jfirebaugh? Otherwise the PR looks great.

[jfirebaugh]

We should update the expected rendering for all of the fill tests.

[lucaswoj]

@jfirebaugh @mourner Great work! Do you have an estimate on how long it will be until this can land in a release? It resolves #930, which is blocking a client of ours.

[jfirebaugh]

The blocker for adopting earcut is getting confidence that the tessellation is good enough. This is a function of both the quality of our vector tile generation pipeline, and the robustness of earcut itself, and myself, @mourner, @lbud, @flippmoke, and @springmeyer are all pushing on this at the moment. We should have a better estimation next week.

[lucaswoj]

Rebased this on master (with #1645) and pushed to earcut-final-final.

[jfirebaugh]

@lucaswoj Looks like earcut-final-final is still pointing at the same commit as earcut-final. Feel free to just force push earcut-final.

[lucaswoj]

Ok. Force pushed here and deleted earcut-final-final. (Just in case, previous SHA was 44ce255)

[mourner]

@lucaswoj let's rebase on master to make sure it's in mergeable state

[lucaswoj]

Done @mourner 🚀 (previous SHA was 494bb19)

[lucaswoj]

We can close #930 when this lands

[averas]

Do you have a fresh estimate on when this could be merged? I'm making heavy use of fill-layers and this branch is much more performant than master.

[mourner]

Did you commit this unintentionally?

[lucaswoj]

@mourner Intentionally (though perhaps not not in good practice). I was debugging some serverside problems with @springmeyer.

[lucaswoj]

This is ready to ship pending

• a final code review of this PR
• a code review of Fix query tests for v2 features mapbox-gl-test-suite#91

cc @mourner @jfirebaugh @jakepruitt @ansis @springmeyer

[jfirebaugh]

Let's eliminate convertCoords, rewriting classifyRings and the code that generates the flattened array for earcut in terms of the {x: x, y: y} point-geometry objects. This will both simplify the code and avoid GC churning a large number of 2-element arrays.

[mourner]

@jfirebaugh yep, done.

[mourner]

The test failed because of some really weird Git/NPM issue — tried rebuilding without cache and it still failed with some "git bad object" errors. Not sure whether this is something temporary or broken in one of the subdeps.

[mourner]

Ah, it failed because the test suite ref it pointed to no longer exists for some reason.

[mourner]

OK, rebased again with updated test-suite and it passes.

[jfirebaugh]

👍 🚢 🎉

[springmeyer]

\o/ Thank you everyone!!