Radxa Cubie A7Z – Radxa 4K Camera detected but no image output (need help + FFmpeg HW acceleration setup)

Looking at the Log i noticed you made the “Common Mistake” i described earlier.

Use /dev/video0 when the frame size is 1920x1080 and /dev/video1 when the frame size is 3840x2160, that will fix the errors.

Regarding libcamera, afaik no one ported it to A733, but i never bothered with that.

Furthermore, FFMpeg is not ported yet AFAIK.

You are making unfounded assumptions about what I did. First of all, the kernel crashes happen without even starting video at all. Simply using the query ioctls triggers it, probably on the control enumeration. Secondly, all the video I did try was 1280x720. I did not even attempt above that.

I will comment that the differences between /dev/video0 and /dev./video1 are not apparently documented. v4l2-ctl should reveal them, but it triggers the kernel crash.

libcamera does indeed compile – but (like the others in the thread) I cannot get basic video capture working reliably, so I have not assessed it. Don’t care about ffmpeg. As I said, I’d be happy with v4l2. My main issue is that (a) very basic video functions are flaky as heck whether they work at all, and (b) the logging noise suggests that core h/w support, specifically the ISP, is a very long way from being stable. The dmesg log is filled with “not for production use” messages from kernel modules. Developers have a right to be aware of that before they invest time in a platform.

Final follow-up, because I discovered some helpful insights for others.

The A733-based systems, including the A7A/S/Z, were really designed for Android. If you use Android, not really a problem. However, Allwinner’s non-Android software is effectively cut down from that Android setup, including the binaries, which is why it can involve non-standard, proprietary API usage, especially from V4L2. Some of that is in their gstreamer patches, but not all. Standard V4L2 usage will crash the kernel. There is no realistic prospect of this changing, as the architecture is all for Android.

Knowing this would have meant I could simply avoid the A733 boards, and saved myself a few weeks of effort. As it is, I did intend to use Android on A733 for a different part of the work, so it’s not all a waste. I’d just say that anyone who buys a smaller A733, like a 1Gb A7Z (as I did), thinking any CSI camera will be viable, should have a serious think about the challenges they will face, working with poorly-written and poorly-documented proprietary driver binaries from Allwinner that can possibly be overcome, but not without considerable work and frustration.

Of course, most ARM boards, especially with Allwinner SOCs, were designed for Android. The BSP kernel is an Android kernel. It’s not unique to the A733.
However, there’s a dedicated community (Armbian) which brings up features for various SOCs - and the Radxa distros. If you have some tokens to burn, Claude or other AIs are great at mainlining features. So it’s not like your out of options. And this would help everyone, not just you.

AI can help draft patches, but mainlining V4L2/ISP support still needs real hardware testing and review of Allwinner’s proprietary pieces; the media graph is the hard part, not the overlay. A more useful next step may be taking the crash dmesg and exact pipeline to the Armbian/linux-sunxi people, while using the vendor Android/BSP path for anything that must work soon. Standard V4L2 stability is unlikely to arrive from a generic AI-generated patch alone.