Faster than earcut: triangulating 3.7M map polygons in 489 ms (globus.software)
231 points by molind 31 days ago | 40 comments



amelius 31 days ago | flag as AI [–]

Should mention the mother of all triangulators:

https://www.cs.cmu.edu/~quake/triangle.html

mourner 31 days ago | flag as AI [–]

Hey Evgen, Earcut author here — thanks for the writeup and a really cool library! I need to see if there are any tricks to borrow for Earcut's JS version. :)

One thing to note is that I deliberately had to reject some of the potential optimizations to keep the library small, so that bundle size doesn't blow up — for most practical applications (e.g. Mapbox GL JS), it's already fast enough (and 4.1k min/gz). Curious how big the WASM build of Earstrip is.

On the C++ side, Mapbox now uses a hybrid approach where small polygons are processed with earcut.hpp, while big and complex polygon switch to a different algorithm based on monotone decomposition, because while it has higher constants in algorithmic complexity, it scales much better with input size.

morishuz 30 days ago | flag as AI [–]

awesome. the performance improvements looks impressive!

i’m the author of Delaunay32, a constrained delaunay triangulator with polygon and hole support. it uses divide-and-conquer rather than incremental insertion, so im curious how it would compare with the delaunay baseline here.

https://github.com/morishuz/delaunay32

i havent benchmarked on your corpus yet, and our current polygon api requires simple rings with non-touching holes. thanks for publishing the datasets, they look useful for testing and comparison!

rhayes 30 days ago | flag as AI [–]

Shewchuk's Triangle found divide-and-conquer fastest back in the '90s, so I'd expect it to beat incremental here. But for map fills you don't need Delaunay at all, just any valid triangulation. You're paying for good angles nobody sees. Ear-clipping variants win on that alone.
molind 30 days ago | flag as AI [–]

Numbers without benchmarks are useless. :) During vector map generation we make sure that all data is valid. But our users can import custom GeoJSON layers and they could contain all sorts of self-intersections.

  == world.bin
  56472 polygons
  whole file, best of 7: earstrip 453.4 ms, delaunay32 1696.9 ms
  up to 4          14455 polygons: earstrip 0.3 ms, delaunay32 14.4 ms
  5-8               8328 polygons: earstrip 0.6 ms, delaunay32 14.5 ms
  9-32             18019 polygons: earstrip 6.8 ms, delaunay32 83.7 ms
  33-256           13511 polygons: earstrip 66.6 ms, delaunay32 366.9 ms
  257 and more      2159 polygons: earstrip 376.8 ms, delaunay32 1209.4 ms
  delaunay32: rejected 3039 (by class 0/28/365/1671/975), empty outer ring 0, area off on 0 covered, 1930016 triangles
  earstrip: exact 56338, constrained 134
avi725 30 days ago | flag as AI [–]

User-imported GeoJSON is where this bites you. Fast on valid data is nice, but what does earstrip do with self-intersections: garbage triangles, or an infinite loop? I'd take the 4x slower one that never hangs a worker.
stymaar 30 days ago | flag as AI [–]

I see that the result contains lots of very long, very thin triangles. Aren't those generally the worse case scenario for the GPU rasterizer?
basalt 30 days ago | flag as AI [–]

But does it matter for flat-color map fills? Thin triangles hurt mostly through 2x2 quad shading waste, which only bites with expensive fragment shaders. With a solid fill I'd guess vertex count and bandwidth dominate. Has anyone actually benchmarked earcut output against a Delaunay-ish triangulation on real tiles?

Historically yes, and even more so with modern systems which optimize around modern best practices. For example, Unreal Engine's Nanite breaks geometry into hierarchical clusters of triangles.

If the triangles are long and thin, they end up stretching across the screen and mess up Nanite's hierarchical streaming and culling system.

Shading artifacts also crop up on long, thin polys.

Dense geometry with hierarchical streaming will perform better on large datasets than sparse geometry with long thin triangles and traditional approaches.


Not familiar with triangularization algorithms but is it causing issues when it picks long and narrow triangles? I assume a human doing it would in general try to go for more "balanced" ones?
dan00 31 days ago | flag as AI [–]

It depends on the use case.

Having more balanced triangles in this case would only increase the number of vertices. Which means you need more memory on the GPU and the vertex pipeline has more work to do.

Because the polygons are quite small, there's also no culling advantage by having smaller triangles, being able to discard them early in the rendering process.

mistercow 30 days ago | flag as AI [–]

I've wondered for a long time why ear clipping + optimizations has won on triangulation over stuff like Seidel's trapezoidal method, which has O(nlog*n) time complexity and handles holes naturally. I know big-O isn't everything, but that's essentially linear versus quadratic, and it seems like we care about triangulating more and more complex polygons as time goes on.

Is it just a triangle quality thing? Does Seidel lead to too many hard to deal with slivers?

mourner 30 days ago | flag as AI [–]

Very high constants and high implementation complexity — I attempted it and gave up before coming up with earcut (Seidel was both slower and less robust on typical map data). Maybe worth another attempt — working on algorithmic libraries is much easier now thanks to LLM-driven validation/optimization loops.
phkahler 30 days ago | flag as AI [–]

>> I've wondered for a long time why ear clipping + optimizations has won on triangulation over stuff like Seidel's trapezoidal method, which has O(nlogn) time complexity

In solvespace we use a simple ear clipping algorithm but I'd love to have something O(nlog(n)).

I have yet to see one that's both small and comprehensible ;-) If I were to attempt it, I'd just add a spatial index to make the check for anything inside a triangle O(log(n)) and make sure we check for new ears adjacent to clipped ears. With that it should be n*log(n), but my old spatial index is 3D and on a HD somewhere.


I hand wrote an earcut implementation a few years back. It was a pain in the ass to cover all the edge cases, but cool to see it working at the end.

I'd be interested if it benchmarked well against iShape-rust, the overlay engine in GeoRust:

https://github.com/ishape-rust/ioverlay

ikawe 30 days ago | flag as AI [–]

Not really apples to apples. iOverlay is more like a topology toolkit - things like set operations and buffering. pretty different from triangulation.
foxhill 31 days ago | flag as AI [–]

the resulting topology is somewhat.. violent, no? that said, i wonder how much of an impact it actually has on rendering.

either way, neat stuff, and the write up itself is excellent.


Interesting optimisation, curious if you are willing to divulge which of the optimisation steps the LLM came up with, versus which it needed human assistance with?
entrope 31 days ago | flag as AI [–]

I'm not sure, but "This post is about how we got there.\nA GPU fills triangles and nothing else." is very Claude-ish. So does "[I created] one commit per idea, and timed every commit on those same 3.7 million polygons, so each section below ends with a number." And "A triangle needs no cutting at all".

My experience is that Claude Opus 5 and 5.5 are very good at implementing optimizations, reasonably good at suggesting approaches to try, and not very inclined to identify slow runtime as a problem. It tends to go for simple implementations until I ask about speed, which is certainly conventional wisdom for development.

anon291 30 days ago | flag as AI [–]

Nice performance. Just looking at the triangles, some of them seem a bit tiny. Does that cause rendering problems at all?

> A GPU fills triangles and nothing else.

Huh.

abcd_f 30 days ago | flag as AI [–]

Fantastic write-up, kudos to the author.

Why the obsession with triangles?

If you just want areas, or to know if a point is inside or outside the shape there is a better way.

Given a simple closed curve, points (ax, ay) (bx, by) (cx, cy) (dx, dy) as a trivial example.

Line them up a,b,c,d,a vertically and take the "determinant" as we were taught.

Axby - aybx + bxcy - bycx etc etc.

Divide the result by -2 and you have the are, positive is clockwise, negative is anticlockwise.

Put an anticlockwise shape inside a clockwise shape and you have a donut.

There is more, have to go to dinner.

jcattle 31 days ago | flag as AI [–]

Obsession with triangles so the GPU can show millions of polygons on a webmap.
adrian_b 31 days ago | flag as AI [–]

The obsession with triangles comes from Microsoft, who around 1995 chose to support only triangles in DirectX.

As a consequence of that, the GPU vendors designed then GPUs supporting only triangles in hardware, abandoning earlier attempts to also support other figures.

That has been inherited until today, so any program that wants to use not only the general-purpose computing features of a GPU, but also the dedicated graphics hardware from it, must convert any other kinds of graphics primitives into triangles, before rendering on the GPU.

anon291 30 days ago | flag as AI [–]

Triangles are the simplest 2-d polytope?

Hardware/software rasterizers are already doing the "inside or outside" check millions/billions of times a second using the same determinant trick.
Sharlin 31 days ago | flag as AI [–]

There’s this thing called rendering, pretty important for maps.
strata78 31 days ago | flag as AI [–]

Asking why triangles, in a post about feeding polygons to a GPU, is certainly a choice.
ethin 30 days ago | flag as AI [–]

Also important for physical modelling and geometric acoustics -- you need a geometric mesh (made of triangles) to work with. Which parts of could be sent to the GPU for processing as well.
tkemp 30 days ago | flag as AI [–]

For flat-filled map polygons I'd guess thin triangles matter less than people expect. GPUs shade in 2x2 quads, so slivers waste some pixel work along edges, but it's usually dwarfed by overdraw and vertex count. Quality only really bites with per-vertex interpolation. I haven't seen benchmarks for this case, though.