I've said elsewhere that my pbn algorithm is really a merge sort in reverse, with random swaps. So that might be one direction to explore. Another option is to see if we can improve on NuSan's shader they came up with after reading my notebook: https://twitter.com/NuSan_fx/status/1464273180059750401 https://www.shadertoy.com/view/7lV3Ry

*he (hadn't checked the twitter bio for pronouns so defaulted to "they")

Also, this is way above my math level but might be relevant: http://abdallagafar.com/publications/dyadic-nets/

(I've promoted http://abdallagafar.com/ before right? His work is so good)

The ACM link has an 18min video: https://dl.acm.org/doi/abs/10.1145/3450626.3459880

Oh, they actually wrote a follow-up with actual blue noise! https://dl.acm.org/doi/abs/10.1145/3550454.3555519

I had missed all this activity! Amazing :)

I have now reduced NuSan’s shader to a few lines of code.

Neat!

You have to force the width/height into a power of two if you want the closed form to render properly, it seems

I think this works now

Bayer still does funky things on my screen - and the pb one possibly too, but if so it's masked by the swapping

🤔 maybe it's an artifact of using Plot.raster? My main contribution to the shader code is that I eliminated any notion of width and height—so the function itself is completely independent of the target geometry (provided you stay in the [0-65535] range for x and y).

Well, if we figure out what breaks the Bayer code we also know what breaks the pb code

Am I reading this right and are we making a 2¹⁶ by 2¹⁶ Bayer matrix here?

Oh right, 32 bit numbers, duh

yes, except we're not producing them all :)

Right, right, closed form :)

My guess is that the cause of t he the issue is that we're not "tiling" x,y correctly. If so I was wondering if the problem started here. Clearly not.

So presuming `bayer` is working right, the problem might be the x and y values sent to it. Is Plot.raster guaranteed to always send integer values over, and increase those integer values one step per pixel?

yes (and double checked by logging the values)

is your browser set to 100%?

Yep

Also seems to happen in both Chromium and Firefox browsers. Maybe it's an issue with my graphics card?

… I don't see any problem on my screen so it's hard to guess. Maybe we can make a pure canvas version which would help eliminate any issue with Plot

I got some ready-made functions for that I think :)

Why do I suddenly hear a carillon playing? https://www.youtube.com/watch?v=8qvdAWLcPyU

Sorry, was just figuring out what the `b = y` was for. Mind if I try if Interleaved Gradient Noise works as a mix function?

try all you want! (use the tinker mode if you don't want to fork or to add to history.) I've tried several things that are much faster but less nicely distributed (e.g. sqrt instead of sin). There's potential to be at least 5x faster.

I suspect that if we manage to avoid the casting between floating point and integers (for bitmasking) and inline that we'd get close to the speed of unmodified bayer

Ah, forgot about the tinker mode :)

When I profile it's the 16 calls to Math.sin that take about 90% of the CPU time.