Initial mainline video capture and camera support for Rockchip RK3588 (collabora.com)

by mfilion 28 comments 91 points
Read article View on HN

28 comments

[−] jauntywundrkind 32d ago
Worth noting that, well, alas, camera support is incredibly incredibly incredibly cursed, period. It feels like, broadly, with all the image blocks, everyone makes really neat really good hardware thats chock-a-block full of capabilities that are un- or poorly documented or really hard to support for reasons, etc. Its a pretty bespoke high throughput pipeline with a lot of special domain knowledge very unlike anything else on computing.

Intel's IPU6 has been a ~4 year travail to get going (thankfully IPU7 landed fairly quickly however!) https://www.phoronix.com/news/Intel-IPU6-Camera-Challenge-25 https://www.phoronix.com/news/Intel-IPU7-Linux-6.17

AMD similarly has only just gotten the Strix Halo ISP near working: https://www.phoronix.com/news/AMD-ISP4-For-Linux-7.2

The whole video world seems like a nightmare. Difficult world of hardware. Just the worst Intellectual Property hostage taking banditry from awful awful valent legally predatory people everywhere, a dark forest ready to leap out of the dark and attack you if you dare use a computing to deal in bits that represent moving images.

[−] ironman1478 32d ago
I work in this field and this is 100 percent true. It's really hard to learn about too. A lot of textbooks go over the algorithms in the chips in an idealized form. The actual versions are so messy and different that the textbooks aren't even useful sometimes, especially if you work on custom ISPs. It's cursed, but it's fun.
[−] skoocda 31d ago
Are there any resources you would recommend for learning about real implementations?
[−] ironman1478 30d ago
Unfortunately no. It's all proprietary and locked up. However, I can say that the optimizations are not always about IQ (though that is a major factor). It's also about making things run fast enough in a low latency environment. Those two requirements lead to strange hardware designs that lead to strange register interfaces.
[−] XorNot 31d ago
Wait is IPU6 working now? My work laptop has it and the only thing I know is it might work provided I use some user space relay driver and a bunch of other things.

Has this situation improved?

[−] jauntywundrkind 31d ago
I looked and looked and looked. I don't know why but IPU7 is reported to be pretty good. But IPU6 might still be in perpetual limbo; I don't know. What kernel are you on btw? Sorry for your pain.
[−] Aurornis 32d ago
Great to see progress on mainlining more support for common and powerful chips.

The work required to get this one piece into mainline over 5-6 years reveals why most chip vendors aren’t aiming for mainline by default:

> A few iterations of the rkcif driver later, the basic driver providing support for the PX30 VIP and the RK3568 VICAP was accepted (October 2025). After more than five years of development, including 25 iterations and three renamings, this was a major milestone. On the other hand, there was still a lot to do, of course. For instance, the Rockchip MIPI CSI-2 receiver unit that is coupled closely to the VICAP required a mainline driver as well.

It’s never as simple as submitting existing work upstream and making a few changes. It takes a lot of development and a willingness to rewrite everything, possibly multiple times, to track the goals of upstream.

[−] Palomides 32d ago
I really feel like that should be table stakes if your entire business is making chips to run Linux, though

after working professionally with their stuff I'm really not a fan of Rockchip

[−] yjftsjthsd-h 32d ago

> why most chip vendors aren’t aiming for mainline by default:

> It’s never as simple as submitting existing work upstream and making a few changes.

If they had started by working with upstream, then they wouldn't have to go through unnecessary revisions trying to adapt the thing they already wrote.

[−] chucklenorris 31d ago
wow, i have a few of these laying around. i also bought some imx678 sensors i wanted to use with them. i tried pretty hard to make a driver work with these but it was impossible to get the isp working without modifying the kernel itself so i gave up. That convinced me to never buy hw that doesn't have drivers in the mainline kernel.
[−] tov_objorkin 31d ago
The product has a typical lifespan of 3–5 years, they just don't need LTS. RKISP(ImageSignalProcessor) is piece of code glued to the kernel, fast and cheap. The mainstream version provides proper integration with Linux multimedia subsystems.
[−] 4fterd4rk 32d ago
I guess I don't understand... why would the SOC manufacturer spend the money on integrating this stuff if they don't intend on also spending the money to enable it on the software side?
[−] bullen 32d ago
Some 3588 CMs are sold out.

So it might be too late as 3688 will be too hot...

Just like routers get dd-wrt when sold out!

[−] rasz 31d ago

>RK3588 ISP, as the vendor kernel driver is not upstreamable for various reasons

Its always IP. Afair even Raspberry Pi didnt open ISP, and support for MIPI took yeaaaaars.

[−] ACCount37 32d ago
Good. Video capture on second grade Linux SoCs is hell - lots of blobs and weird custom vendor SDKs that work with the vendor's own happy path use case demos and nothing else.

I hope that the more SoCs get mainline V4L2, the more likely the future SoCs are going to be to use it instead of doing something non-standard and awful.

[−] Muromec 31d ago
I still have not figured whether hdmi rx works on my orange pi or not