This week, I reviewed the last available version of the Linux KMS Color API. Specifically, I explored the proposed API by Harry Wentland and Alex Hung (AMD), their implementation for the AMD display driver and tracked the parallel efforts of Uma Shankar and Chaitanya Kumar Borah (Intel) in bringing this plane color management to life. With this API in place, compositors will be able to provide better HDR support and advanced color management for Linux users.

To get a hands-on feel for the API's potential, I developed a fork of drm_info compatible with the new color properties. This allowed me to visualize the display hardware color management capabilities being exposed. If you're curious and want to peek behind the curtain, you can find my exploratory work on the drm_info/kms_color branch. The README there will guide you through the simple compilation and installation process.

Note: You will need to update libdrm to match the proposed API. You can find an updated version in my personal repository here. To avoid potential conflicts with your official libdrm installation, you can compile and install it in a local directory. Then, use the following command: export LD_LIBRARY_PATH="/usr/local/lib/"

In this post, I invite you to familiarize yourself with the new API that is about to be released. You can start doing as I did below: just deploy a custom kernel with the necessary patches and visualize the interface with the help of drm_info. Or, better yet, if you are a userspace developer, you can start developing user cases by experimenting with it.

The more eyes the better.

KMS Color API on AMD

The great news is that AMD's driver implementation for plane color operations is being developed right alongside their Linux KMS Color API proposal, so it's easy to apply to your kernel branch and check it out. You can find details of their progress in the AMD's series.

I just needed to compile a custom kernel with this series applied, intentionally leaving out the AMD_PRIVATE_COLOR flag. The AMD_PRIVATE_COLOR flag guards driver-specific color plane properties, which experimentally expose hardware capabilities while we don't have the generic KMS plane color management interface available.

If you don't know or don't remember the details of AMD driver specific color properties, you can learn more about this work in my blog posts [1] [2] [3]. As driver-specific color properties and KMS colorops are redundant, the driver only advertises one of them, as you can see in AMD workaround patch 24.

So, with the custom kernel image ready, I installed it on a system powered by AMD DCN3 hardware (i.e. my Steam Deck). Using my custom drm_info, I could clearly see the Plane Color Pipeline with eight color operations as below:

└───"COLOR_PIPELINE" (atomic): enum {Bypass, Color Pipeline 258} = Bypass
    ├───Bypass
    └───Color Pipeline 258
        ├───Color Operation 258
        │   ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, Multiplier, 3D LUT} = 1D Curve
        │   ├───"BYPASS" (atomic): range [0, 1] = 1
        │   └───"CURVE_1D_TYPE" (atomic): enum {sRGB EOTF, PQ 125 EOTF, BT.2020 Inverse OETF} = sRGB EOTF
        ├───Color Operation 263
        │   ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, Multiplier, 3D LUT} = Multiplier
        │   ├───"BYPASS" (atomic): range [0, 1] = 1
        │   └───"MULTIPLIER" (atomic): range [0, UINT64_MAX] = 0
        ├───Color Operation 268
        │   ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, Multiplier, 3D LUT} = 3x4 Matrix
        │   ├───"BYPASS" (atomic): range [0, 1] = 1
        │   └───"DATA" (atomic): blob = 0
        ├───Color Operation 273
        │   ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, Multiplier, 3D LUT} = 1D Curve
        │   ├───"BYPASS" (atomic): range [0, 1] = 1
        │   └───"CURVE_1D_TYPE" (atomic): enum {sRGB Inverse EOTF, PQ 125 Inverse EOTF, BT.2020 OETF} = sRGB Inverse EOTF
        ├───Color Operation 278
        │   ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, Multiplier, 3D LUT} = 1D LUT
        │   ├───"BYPASS" (atomic): range [0, 1] = 1
        │   ├───"SIZE" (atomic, immutable): range [0, UINT32_MAX] = 4096
        │   ├───"LUT1D_INTERPOLATION" (immutable): enum {Linear} = Linear
        │   └───"DATA" (atomic): blob = 0
        ├───Color Operation 285
        │   ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, Multiplier, 3D LUT} = 3D LUT
        │   ├───"BYPASS" (atomic): range [0, 1] = 1
        │   ├───"SIZE" (atomic, immutable): range [0, UINT32_MAX] = 17
        │   ├───"LUT3D_INTERPOLATION" (immutable): enum {Tetrahedral} = Tetrahedral
        │   └───"DATA" (atomic): blob = 0
        ├───Color Operation 292
        │   ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, Multiplier, 3D LUT} = 1D Curve
        │   ├───"BYPASS" (atomic): range [0, 1] = 1
        │   └───"CURVE_1D_TYPE" (atomic): enum {sRGB EOTF, PQ 125 EOTF, BT.2020 Inverse OETF} = sRGB EOTF
        └───Color Operation 297
            ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, Multiplier, 3D LUT} = 1D LUT
            ├───"BYPASS" (atomic): range [0, 1] = 1
            ├───"SIZE" (atomic, immutable): range [0, UINT32_MAX] = 4096
            ├───"LUT1D_INTERPOLATION" (immutable): enum {Linear} = Linear
            └───"DATA" (atomic): blob = 0

Note that Gamescope is currently using AMD driver-specific color properties implemented by me, Autumn Ashton and Harry Wentland. It doesn't use this KMS Color API, and therefore COLOR_PIPELINE is set to Bypass. Once the API is accepted upstream, all users of the driver-specific API (including Gamescope) should switch to the KMS generic API, as this will be the official plane color management interface of the Linux kernel.

KMS Color API on Intel

On the Intel side, the driver implementation available upstream was built upon an earlier iteration of the API. This meant I had to apply a few tweaks to bring it in line with the latest specifications. You can explore their latest work here. For a more simplified handling, combining the V9 of the Linux Color API, Intel's contributions, and my necessary adjustments, check out my dedicated branch.

I then compiled a kernel from this integrated branch and deployed it on a system featuring Intel TigerLake GT2 graphics. Running my custom drm_info revealed a Plane Color Pipeline with three color operations as follows:

├───"COLOR_PIPELINE" (atomic): enum {Bypass, Color Pipeline 480} = Bypass
│   ├───Bypass
│   └───Color Pipeline 480
│       ├───Color Operation 480
│       │   ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, 1D LUT Mult Seg, 3x3 Matrix, Multiplier, 3D LUT} = 1D LUT Mult Seg
│       │   ├───"BYPASS" (atomic): range [0, 1] = 1
│       │   ├───"HW_CAPS" (atomic, immutable): blob = 484
│       │   └───"DATA" (atomic): blob = 0
│       ├───Color Operation 487
│       │   ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, 1D LUT Mult Seg, 3x3 Matrix, Multiplier, 3D LUT} = 3x3 Matrix
│       │   ├───"BYPASS" (atomic): range [0, 1] = 1
│       │   └───"DATA" (atomic): blob = 0
│       └───Color Operation 492
│           ├───"TYPE" (immutable): enum {1D Curve, 1D LUT, 3x4 Matrix, 1D LUT Mult Seg, 3x3 Matrix, Multiplier, 3D LUT} = 1D LUT Mult Seg
│           ├───"BYPASS" (atomic): range [0, 1] = 1
│           ├───"HW_CAPS" (atomic, immutable): blob = 496
│           └───"DATA" (atomic): blob = 0

Observe that Intel's approach introduces additional properties like "HW_CAPS" at the color operation level, along with two new color operation types: 1D LUT with Multiple Segments and 3x3 Matrix. It's important to remember that this implementation is based on an earlier stage of the KMS Color API and is awaiting review.

A Shout-Out to Those Who Made This Happen

I'm impressed by the solid implementation and clear direction of the V9 of the KMS Color API. It aligns with the many insightful discussions we've had over the past years. A huge thank you to Harry Wentland and Alex Hung for their dedication in bringing this to fruition!

Beyond their efforts, I deeply appreciate Uma and Chaitanya's commitment to updating Intel's driver implementation to align with the freshest version of the KMS Color API. The collaborative spirit of the AMD and Intel developers in sharing their color pipeline work upstream is invaluable. We're now gaining a much clearer picture of the color capabilities embedded in modern display hardware, all thanks to their hard work, comprehensive documentation, and engaging discussions.

Finally, thanks all the userspace developers, color science experts, and kernel developers from various vendors who actively participate in the upstream discussions, meetings, workshops, each iteration of this API and the crucial code review process. I'm happy to be part of the final stages of this long kernel journey, but I know that when it comes to colors, one step is completed for new challenges to be unlocked.

Looking forward to meeting you in this year Linux Display Next hackfest, organized by AMD in Toronto, to further discuss HDR, advanced color management, and other display trends.

19 May 2025 9:05pm GMT

Earlier in this blog we looked at how to generate a justified block of text, which is nowadays usually done with the Knuth-Plass algorithm. Sadly this is not, by itself, enough to create a finished product. Processing all input text with the algorithm produces one very long column of text. This works fine if you end result is a scroll (of the papyrus variety), but people tend to prefer their text it book format. Thus you need a way to split that text column into pages.

The simplest solution is to decide that a page has some N number of lines and page breaks happen at exact these places. This works (and is commonly used) but it has certain drawbacks. From a typographical point of view, there are at least three things that should be avoided:

1. Orphan lines, a paragraph at the bottom of a page that has only one line of text.
2. Widow lines, a paragraph that starts a new page and has only one line of text.
3. Spread imbalance, where the two pages on a spread have different number of lines on them (when both are "full pages")

Determining "globally optimal" page splits for a given text is not a simple problem, because every pagination choice affects every pagination choice that comes after it. If you stare at the problem long and hard enough, eventually you realize something.

The two problems are exactly the same and can be solved in the same way. I have implemented this in the Chapterizer book generator program. The dynamic programming algorithm for both is so similar that they could, in theory, use shared code. I chose not to do it because sometimes it is just so much simpler to not be DRY. They are so similar, in fact, that a search space optimization that I had to do in the paragraph justification case was also needed for pagination even though the data set size is typically much smaller in pagination. The optimization implementation turned out to be exactly the same for both cases.

The program now creates optimal page splits and in addition prints a text log of all pages with widows, orphans and imbalances it could not optimize away.

Has this been done before?

Probably. Everything has already been invented multiple times. I don't know for sure, so instead I'm going to post my potentially incorrect answer here on the Internet for other people to fact check.

I am not aware of any book generation program doing global page optimization. LibreOffice and Word definitely do not do it. As far as I know LaTeX processes pages one by one and once a page is generated it never returns to it, probably because computers in the early 80s did not have enough memory to hold the entire document in RAM at the same time. I have never used Indesign so I don't know what it does.

Books created earlier than about the mid eighties were typeset by hand and they seem to try to avoid orphans and widows at the cost of some page spreads having more lines than other spreads. Modern books seem to optimize for filling the page fully rather than avoiding widows and orphans. Since most books nowadays are created with Indesign (I think), it would seem to imply that Indesign does not do optimal page splitting or at least it is not enabled by default.

When implementing the page splitter the reasons for not doing global page splitting became clear. Computing both paragraph and page optimization is too slow for a "real time" GUI app. This sort of an optimization really only makes sense for a classical "LaTeX-style" book where there is one main text flow. Layout programs like Scribus and Indesign support richer magazine style layouts. This sort of a page splitter does not work in the general case, so in order to use it they would need to have a separate simpler document format.

But perhaps the biggest issue is that if different pages may have different number of lines on each page, it leads to problems in the page's visual appearance. Anything written at the bottom of the page (like page numbers) need to be re-placed based on how many lines of text actually ended up on the page. Doing this in code in a batch system is easy, doing the same with GUI widgets in "real time" is hard.

19 May 2025 8:10pm GMT

It's everyone's favorite time of year, election season! …Okay, maybe not the most exciting thing-but an extremely important one nonetheless.

For anyone who doesn't know, GNOME is comprised of many parts: individual contributors and maintainers, adhoc teams of volunteers, a bunch of open source software in the form of apps and libraries, a whole bunch of infrastructure, and-importantly-a nonprofit foundation. The GNOME Foundation exists to help manage and support the organizational side of GNOME, act as the official face of the project to third parties, and delegate authority when/where it makes the most sense. The GNOME Foundation itself is governed by its elected Board of Directors.

If you contribute to GNOME, you're eligible to become a member of the GNOME Foundation, which gets you some perks (like an @gnome.org email address and Matrix account, blog hosting and syndication, and access to Nextcloud and video conferencing tools)-but most importantly, GNOME Foundation members vote to elect the Board of Directors. If you contribute to GNOME, I highly recommend you become a member: it looks good for you, but it also helps ensure the GNOME Foundation is directly influenced and governed by contributors themselves.

I'm Running for the Board!

I realized today I never actually announced this on my blog (just via social media), but this past March I was appointed to the GNOME Foundation Board of Directors to fill a vacancy.

However, the seat I filled was up for re-election in this very next cycle, so I'm happy to officially announce: I'm running for the GNOME Foundation Board of Directors! As part of announcing my candidacy, I was asked to share why I would like to serve on the board. I posted this on the GNOME Discourse, but for convenience, I've copied it below:

Hey everyone,

I'm Cassidy (cassidyjames pretty much everywhere)! I have been involved in GNOME design since 2015, and was a contributor to the wider FreeDesktop ecosystem before that via elementary OS since around 2010. I am employed by Endless, where I am the community architect/experience lead.

I am particularly proud of my work in early design, communication, and advocacy around both the FreeDesktop color scheme (i.e. dark style) and accent color preferences, both of which are now widely supported across FreeDesktop OSes and the app ecosystem. At elementary I coordinated volunteer projects, lead the user experience design, launched and managed OEM partnerships, and largely maintained our communication by writing and editing regular update announcements and other blog posts. Over the past year I helped organize GUADEC 2024 in Denver, and continue to contribute to the GNOME design team and Flathub documentation and curation.

I was appointed to the GNOME Foundation board in March to fill a vacancy, and I am excited to earn your vote to continue my work on the board. If elected, I will continue my focus on:

• Clearer and more frequent communication from the GNOME Foundation, including by helping write and edit blog posts and announcements

• Exploring and supporting fundraising opportunities including with users, OEMs, and downstream projects

• Ensuring Flathub continues to be recognized as the premier Linux app store, especially as it moves to enable financially supporting the developers of FOSS apps

• More widely communicating the impact, influence, and importance of GNOME and Flathub to raise awareness beyond the existing contributor community

• Helping ensure that the Foundation reflects the interests of the contributor community

I feel like the GNOME Foundation is at an important transformation point, and I look forward to helping steer things in the right direction for an effective, sustainable organization in support of the GNOME community. Regardless of whether I am elected, I will continue to contribute to design and communication as much as I'm able.

Thank you for your consideration!

Become a Member, and Vote!

Voting will be open for two weeks beginning June 5, 2025. If you contribute to GNOME, now is a great time to ensure you're a member so you can vote in time; check the GNOME Discourse announcement for all of the specific dates and details. And don't forget to actually vote once it begins. :)

19 May 2025 12:00am GMT

It's strange being alive while so much bad shit is going on in the world, right? With our big brains that invented smartphones, quantum computers and Haskell, surely we could figure out how to stop Benjamin Netenyahu from starving hundreds of thousands of children? (Or, causing "high levels of acute food insecurity" as the UN refer to it).

Nothing in the world is simple, though, is it.

Back in 1914 when European leaders kicked off the First World War, the collective imagination of a war dated back to an era where the soldiers wore colourful jackets and the most sophisticated weapon was a gun with a knife attached. The reality of WWI was machine guns, tanks and poison gas. All that new technology took people by surprise, and made for one of the deadliest wars in history.

If you're reading this, then however old or young you are, your life has been marked by rapid technological changes. Things are still changing rapidly. And therein lies the problem.

In amongst the bad news I am seeing some reasons to be optimistic. The best defense against exploitation is education. As a society it feels like we're starting to get a grip on why living standards for everyone except the rich are nosediving.

Lets go back to an older technology which changed the world centuries ago: books. I am going to recommend a few books.

Technofeudalism (by Yannis Varoufakis)

The book's preface outlines the theory: capital has mutated into a much more powerful and dangerous form of itself. Two things caused it: the "privatization of the internet", and the manner in which Western governments and central banks responded to the financial crisis of 2008. The strongest part of the book is the detailed telling of this story, from the beginnings of capitalism and its metamorphoses during the 20th century, to the panicked central banks keeping interest rates near zero for over a decade, effectively printing money and giving it to the wealthy, who in turn decided it was best to hang onto all of it. Out of this he declares capitalism itself is dead, replaced by a more powerful force: technofuedalism.

Yanis' concept of technofuedalism is this:

Markets, the medium of capitalism, have been replaced by digital trading platforms which look like, but as not, markets and are better understood as fiefdoms. And profit, the engine of capitalism, has been replaced with its feudal predecessor: rent. Specifically, it is a form of rent that must be paid for access to those platforms and to the cloud more broadly.

Many people depend on cloud platforms for basic needs. For example, access to work. Think about how many people earn a living through apps: Uber drivers, food delivery couriers, freelancers advertising via Google or Meta, and so on. But it's not just individuals. Many capitalist businesses now rely on sites like Amazon.com for most of their sales. Everyone has to pay cloud rent to the overlord.

Yanis likens this to a colonial town where all the shops are owned by the same guy, who happens to be named Jeff. This isn't your traditional monopoly, though - because cloud platforms also "personalize" your experience of the site. You get recommendations perfectly tailed to your needs. For consumers, the platform owners control what you see. For participants, they control who is paying attention. This he calls cloud capital.

The concept of cloud capital needs better definition in the book, but I think the attention economy is the most interesting aspect, and it is what leads to the other counterintuitive side effect: many people creating value for cloud platforms do it for little or no money. Apple doesn't pay you to make app store apps. Tiktok don't pay you to upload videos. The book claims that capitalist companies like General Motors pay about 80% of their income to workers as salary payments, while Big Tech companies tend to spend less than 1% of their income paying workers.

In my last status update I mentioned some disillusionment with open source projects in the age of AI. Here's another perspective: contributing to some open source projects now feels like giving free labour to cloud platform owners.

The Free Software movement dates from the 1980s, when operating systems were a source of power. Microsoft created an illegal monopoly on operating systems in the 90s and became the richest and most powerful company in the world; but today, operating systems are a commodity, and Microsoft makes more money from its cloud platform Azure.

It's great that we maintain ethical, open computing platforms like GNOME, but the power struggle has moved on. I don't expect to see huge innovations in desktop or mobile operating systems in the next decade.

Meanwhile, maintaining ethical cloud platforms is still a minority pursuit. Writing software doesn't feel like the difficult part, here. The work needed if we have the will to climb out of this technofuedal hole is community organization and communication. The most valuable thing the major cloud platforms have is our attention. (And perhaps the most valuable thing we have in the open source world, is our existing communities and events, such as the Linux App Summit).

Why does this book give me hope? It gives a structure to the last 18 years of fucked up goings on in the world of power and politics. And it helps us analyze exactly what makes the big tech companies of the USA and China so powerful.

If the cloudalists got to you already and you don't have the attention span to buy and read a book any more, don't worry! There's also a video.

The Trading Game (by Gary Stevenson)

I'm late to the party with this one. Gary started a blog ten years ago (wealtheconomics.org), and now runs a online video channel (GarysEconomics).

He knows a lot about money and the super-rich. He knows that people are addicted to accumulating wealth and power. He knows that living standards for working people are getting worse. He knows that politicians won't admit that the two things are linked. And he has over a million subscribers to his channel who know that too.

Why does it give me hope? First, he's focused on helping us understand the problem. He does have a clickbait solution - "Tax wealth, not work" - but he also acknowledges that it's slow, difficult and messy to affect national politics. He's realistic about how difficult it is to tax the super-rich in a world of tax havens. And he's been going at it for 10 years already.

Careless People (by Sarah Wynn-Williams)

I listened to long discussion of this book on a podcast called Chisme Corporativo, run by two chicas from Mexico working to demystify the world of big business and USA power that controls much of the world.

The fact that Chisme Corporativo exists makes me very happy. If we're going to live in a world where US corporations have more power than your own government - particularly the case in Latin America - then it makes a lot of sense to learn about US corporations, the people who run them, and why they make the decisions they do.

The book review quotes a part where Mark Zuckerberg finally realized that Facebook was instrumental in the first Tromp election campaign, and just how much power that endows the company with.

And he digested this bombshell for three hours, and his thought process led him to this: "Maybe I should run for president!"

That's the type of person we are dealing with.

What's next

Inequality keeps rising and living standards are getting worse for everyone except the super rich. But we are learning more and more about the people and the processes responsible. Information is a seed for bringing society into better balance again.

I'm going to leave you with this quote I stole blatantly from Gary Stevenson's website:

"If we can really understand the problem, the answer will come out of it, because the answer is not separate from the problem."
― J. Krishnamurti

Have you read any of these books? What else would you add to this list?

18 May 2025 10:17pm GMT

Hi!

Today wlroots 0.19.0 has finally been released! Among the newly supported protocols, color-management-v1 lays the first stone of HDR support (backend and renderer bits are still being reviewed) and ext-image-copy-capture-v1 enhances the previous screen capture protocol with better performance. Explicit synchronization is now fully supported, and display-only devices such as gud or DisplayLink can now be used with wlroots. See the release notes for more details! I hope I'll be able to go back to some feature work and reviews now that the release is out of the way.

In other graphics news, I've finished my review of the core DRM patches for the new KMS color pipeline. Other kernel folks have reviewed the patches, we're just waiting on a user-space implementation now (which various compositor folks are working on). I've started a discussion about libliftoff support.

In addition to wlroots, this month I've also released a new version of my mobile IRC client, Goguma 0.8.0. delthas has sent a patch to synchronize pinned and muted conversations across devices via soju. Thanks to pounce, Goguma now supports message reactions (not included in the release):

My extended-isupport IRCv3 specification has been accepted. It allows servers to advertise metadata such as the maximum nickname length or IRC network name early (before the user provides a nickname and authentication details), which is useful for building nice connection UIs. I've posted another proposal for IRC network icons.

go-smtp 0.22.0 has been released with an improved DATA command API, RRVS support (Require Recipient Valid Since), and custom hello after reset or STARTTLS. I've also spent quite a bit of time reaching out to companies for XDC 2025 sponsorships.

See you next month!

14 May 2025 10:00pm GMT

It has been almost a year since my last update on the Rockchip NPU, and though I'm a bit sad that I haven't had more time to work on it, I'm happy that I found some time earlier this year for this.

Quoting from my last update on the Rockchip NPU driver:

The kernel driver is able to fully use the three cores in the NPU, giving us the possibility of running 4 simultaneous object detection inferences such as the one below on a stream, at almost 30 frames per second.

All feedback has been incorporated in a new revision of the kernel driver and it was submitted to the Linux kernel mailing list.

Though I'm very happy with the direction the kernel driver is taking, I would have liked to make faster progress on it. I have spent the time since the first revision on making the Etnaviv NPU driver ready to be deployed in production (will be blogging about this soon), and also had to take some non-upstream work to pay my bills.

Next I plan to cleanup the userspace driver so it's ready for review, and then I will go for a third revision of the kernel driver.

12 May 2025 5:30am GMT