That one time when something is actually unique to 3D, though:D

Should be good, yeah :D

🧐 wasn't aware it was changed to completed, but I'm afraid that still a bit early even if it's getting closer and closer. Will investigate

This new one is actually heavier than the heavy one I had, and it feels like it's put to good use. I'll do some more tests and post about it later

What's the GPU in it?

Sadly after my previous high end Asus laptop I'd be hesitent to go for them again right away 🫠

That looks nice, too. You have one?

So lovely!

from a WIP thing

Have you tried simply brute forcing a lot of line segments?

At some point it's distinguishible from any actual curves, and with some simple culling/binning you can restrict the actual evaluation count significantly.

And if you place the points carefully you can control the tolerance as well

Me every day, but for physics-adjacent topics in computer graphics by and for computer scientists

🤔

Tested various version in my shader, interesting how it becomes fully pathological when simply adding a `const`: gist.github.com/vassvik/621c...

For the binary classification dot(n0, n1) < cos(t0 + t1) = cos(t0)cos(t1) - sqrt((1 - cos^2(t0))(1-cos^2(t1))) = cos(t0)cos(t1) - sqrt(1 - cos^2(t0) - cos^2(t1) + cos^2(t0)cos^2(t1)) should do, no?

What happened to the image? 😅

TFW it works, and the gains are massive

Colorado by number of splines that pass the culling.

Colors are (R, G, B, Y, M, C, W) = (1, 2, 3, 4 ,5, 6, >=7), respectively.

First image uses the outer circle, second image uses the purple region.

I measured in the middle of a chonky thing that does *a lot more*. The runtime of the entire thing changed from 2400 microseconds to 500.

Lots of registers indeed

Never actually tested the switch version, but the difference between the array version and any of the computed ones was night and day