iiinova — where your digital self keeps evolving

Yea makes sense! I've been doing some weird shenanigans to make procedural shader text effects on my end so I may jot them down some time

Still though, if the SDF solution works and the memory isn't a problem, that may indeed be the way to go.

100

each curve for each pixel, which probably isn't much slower than the scanline sweeper, although I haven't worked out the distance to a quadratic Bezier on paper yet (you'd want to leverage the monotonicity so you don't have to evaluate all cubic roots, compared to the general solution).

100

Mathias Westerdahl 🇺🇦 ✈️ Gamescom @jcash.bsky.social · 03/08/2026

Fast is definitely a bit harder since you'd have to query the closest distance to a curve from each pixel (at which point any analytical line anti-aliasing technique would work for arbitrary thicknesses). I guess depending on the font, you could use the same curve packing layout and brute-force...

100

Reposted by Scuff3D Rook

I think the vacation shenanigans are nearly done. I've done a due diligence of using #Slug or #SweepingScanline and it seems to comeout in favor of the sweep algo. Faster, simpler code, single curve texture. Thx @scuff3drook.bsky.social ! @defold.com #gamedev

1102

Scuff3D Rook @scuff3drook.bsky.social · 30/07/2026

I concede that a talented programmer can extract value from LLMs as a tool, just as they can a static analyzer, theorem prover, or any tool really. My point is that in a few generations, there will be far *fewer* talented programmers than before, and there isn't a pipeline to replace what's lost.

100

Scuff3D Rook @scuff3drook.bsky.social · 30/07/2026

Also, please do use the "thread" feature when replying ;). It's a lot easier to follow your train of thought when the replies are sequenced chronologically.

000

Scuff3D Rook @scuff3drook.bsky.social · 30/07/2026

I think the actual throughput gain is debateable but I wouldn't use it in the current state for strictly ethical reasons regardless of its efficacy 🤷‍♂️. I can't/don't impose that on others but it's a personal decision.

200

Scuff3D Rook @scuff3drook.bsky.social · 28/07/2026

While writing this, I was reminded of alextardif.com/AI.html from @longbool.bsky.social which is a good (albeit sobering) read as well

alextardif.com

Alex Tardif: Graphics Programmer

Scuff3D Rook @scuff3drook.bsky.social · 28/07/2026

rookandpossum.com/posts/anti-s... This is an off-my-chest sort of post about what I perceive to be the incredible madness in the industry writ-large with respect to AI.

591

Scuff3D Rook @scuff3drook.bsky.social · 26/07/2026

Very nice work! Some of those torture test fonts are terrifying

000

Scuff3D Rook @scuff3drook.bsky.social · 23/07/2026

GLSL : HLSL :: AT&T assembly : Intel assembly

000

Scuff3D Rook @scuff3drook.bsky.social · 23/07/2026

... partitioning scheme) works well if you end up needing it. As for other differences, are you seeing the differences with the same gamma applied to your exponent? Asking mainly out of curiosity, since this "antialias falloff" is technically a stylistic parameter you can control

100

Scuff3D Rook @scuff3drook.bsky.social · 23/07/2026

Hey glad the experiments with the technique are working out for you! One tip is to clamp very small coverage values to 0. I just saw your other note about performance also -- for fonts like Roboto it shouldn't be a big deal, but some sort of horizontal partitioning scheme (with an optional vertical

300

Scuff3D Rook @scuff3drook.bsky.social · 21/07/2026

Complex log gets you a linearly varying angle. Quaternion log gets you the 3D version. This is why interpolating the log is the constant angular velocity operation (slerp) just as blending the logarithms of complex numbers and exponentiating is a constant angular velocity rotation

000

Scuff3D Rook @scuff3drook.bsky.social · 21/07/2026

bsky.app/profile/scuf... Yea, taking the logarithm is analogous to "linearizing" the quaternion to get the 3d analogue of an angle, just like taking the log of a complex number in the complex plane gets you the angle

100

More to the point though, even though the unit sphere is 2D in a degrees of freedom sense, it's not as helpful to think of it as such (the embedding is useful) -- it's the same with quaternions is similar IMO (it's more useful to think of them as topological spheres embedded in 4D)

100

Anyways, the main point of this is that IMO, topology is often best understood by considering the local neighborhood (in terms of metric, local curvature, connectivity, etc.), as opposed to trying to understand things at a global level. That's how it clicks better for me at least!

000

The fact that the neighborhood of a quaternion is locally linear suggests you can use a logarithms and exponential operators to go from a quaternion to the linear space and back. Blending in the linear space and exponentiating is how you derive slerp.

100

..in the plane perpendicular to the axis. Every quaternion has this connection to other "nearby" quaternions on the manifold. Since normalizing a quaternion gets you back on the manifold and the manifold has constant curvature you can blend quats as though they were points on the sphere

100

bsky.app/profile/scuf... The way I think about it is in terms of the space local to a quaternion. For every quaternion, there is a local tangent space which is a 3-dimensional tangent plane. Perturbations of the quaternion work geometrically by perturbing the angle itself, or the axis of rotation..

100

The dimensionality honestly isn't as important a property as the fact that there is an algebraic structure though. So you can take quaternion derivatives, blend/average quaternions, and so on without "leaving the manifold" so to speak.

It's vaild to think of quaternions as a 3-manifold embedded in R4 if you want, just as a sphere is a 2-manifold in R3. Topologically, every quaternion has a locally linear "tangent space" analogous to a Lie algebra, which is why so many operations on quaternions are well-defined

100

R&P -- Tell me about a bug you fixed...

rookandpossum.com/posts/tell-m... Quick post, expanding on last week's post on interview questions. This post expands on a few responses I've given over the years when prompted to talk about a "tricky bug I fixed."

rookandpossum.com

Continuing with the theme of past interview questions, let's look at some responses to the common "tell me about a bug you fixed" prompt.

000

Scuff3D Rook @scuff3drook.bsky.social · 16/07/2026

All that said though, nothing beats the guesswork out of a hypothesis like a profiler :)

Scuff3D Rook @scuff3drook.bsky.social · 16/07/2026

Memory loads can only be hidden insofar as you have waves that are able to be resident. The other thing is to only compare shader execution times with stable shader clocks enabled, since there can be quite a bit of deviation depending on other system activity.

100

Scuff3D Rook @scuff3drook.bsky.social · 16/07/2026

I would use Nvidia Nsight to be sure (it's an amazing tool), but this strikes me as an occupancy problem in one of two ways. Code motion can have a dramatic impact on register allocation, so you may be crossing a register block allocation boundary.

ceruleite.bsky.social @ceruleite.bsky.social · 15/07/2026

Reposted by Scuff3D Rook

2026 Advances in Real-Time Rendering in Games course line up is here - check it out! advances.realtimerendering.com

16127

Scuff3D Rook @scuff3drook.bsky.social · 15/07/2026

I've asked AMD dev tech for an Aftermath equivalent over the years... hoping the LLM nonsense dies so we can get back to principled engineering again.

Scuff3D Rook @scuff3drook.bsky.social · 15/07/2026

Yea that would be nice for sure, but I'll concede this one since unlike the CPU world where the ISAs for x86/arm are known, gpus still are in a world where ASICs remain unfortunately proprietary for the most part

Scuff3D Rook @scuff3drook.bsky.social · 14/07/2026

Debugging TDRs on AMD GPUs is admittedly difficult, but the *reasons* for the difficulty is because you have to somehow manage to repro the TDR on-demand with an attached capture tool, which is quite hard to do in a large-scale project

100

Scuff3D Rook @scuff3drook.bsky.social · 14/07/2026

AMD decides that what graphics programmers were missing from TDR-analysis on AMD GPUs was the ability to use LLMs to analyze crash dumps... I need to LLM bubble to burst already

140

Scuff3D Rook @scuff3drook.bsky.social · 12/07/2026

This first installment has a couple systems interview questions and a couple graphics programming questions. Coming posts will have a similar format (until I move onto the next thing).

020

Scuff3D Rook @scuff3drook.bsky.social · 12/07/2026

In this week's post, I'm kicking off a mini-series to present interview questions I've received in the past, how I responded, and some thoughts on how I would respond today (with over a decade of hindsight and additional experience). rookandpossum.com/posts/interv...

rookandpossum.com

R&P -- Interview Questions Received, Reviewed

Let's look at some interview questions I've seen over the years, how I responded, and how I might respond today.

120

Scuff3D Rook @scuff3drook.bsky.social · 09/07/2026

I'm familiar with this in the context of Lp spaces en.wikipedia.org/wiki/Lp_space, but yes, absolutely relevant for 2D/3D encodings. Intuitively, dividing by the norm always gets you back to the unit circle, and the shape of the unit circle is dictated by the metric.

en.wikipedia.org

Lp space - Wikipedia

Scuff3D Rook @scuff3drook.bsky.social · 08/07/2026

Youtube has this "feature" where for every comment posted, they generate a slop response, and they even give you an option to auto-reply to everything with slop. The responses are often aggressively patronizing or just flat out wrong (have seen bad math, bad suggestions, etc.). It sucks 🙄

Scuff3D Rook @scuff3drook.bsky.social · 07/07/2026

For simpler fonts though (like Roboto, Geist, etc.), no partitioning is needed, and the shader will run just as quickly in my experience

111

Scuff3D Rook @scuff3drook.bsky.social · 07/07/2026

For performance, what worked best for me was using slightly overlapping horizontal partitions of curves such that each pixel never needs to look up more than one partition at a time. This is "technically" incorrect for some magnification levels, but not actually noticeable

100

Scuff3D Rook @scuff3drook.bsky.social · 07/07/2026

Let me know if you need any assistance! FYI there was a one line shader code bug which was pushed recently in case you're using the code in the paper directly (unfortunate bug crept in during paper code formatting which wasn't caught until today).

Unlike stronger copyleft licenses, MPL 2 doesn't "propagate" to other files in a larger work, so it should still be easier to use in the context of a bigger open source project also, and is compatible with other licenses. See www.mozilla.org/en-US/MPL/2.... Hope this helps!

120

The license also offers some degree of protection from patent trolls, similar to Apache. Rationale for this license is that I don't want to restrict usage for companies/people trying to ship something to users, but want to avoid the *code* from being redistributed in a less permissive way.

Thanks for asking, this has been on my TODO list for a while so I've taken care of it now. The code in the paper (and on the site generally) is all under MPL 2.0, which allows you to redistribute the code in binary form as you like (e.g. in games), but propagates for source distributions.

120

I saw a few comments that were curious if I was going to patent anything or publish the paper in an actual journal ("no" and "no" are the answers), so I wanted to also mention what my actual goals/motivations are. With that, I'm onto the next thing!