Hi, Il giorno Fri, 29 Apr 2022 18:49:31 +0200 Luca Ceresoli <luca.ceresoli@xxxxxxxxxxx> ha scritto: > Hello, > > I am working to implement a driver for the Tegra20 camera > interface (VI) and have it merged in mainline Linux. I have already > done part of the work, but now I'm facing some difficulties, > especially with respect to interaction with the syncpt feature of > host1x. I have a few questions that hopefully someone here can reply > to, but first let me lay some background. > > First, the good news. > > A good thing is I have an old reference image using a vendor kernel > based on 3.1 which works and I can capture video from a parallel smart > sensor in YUV. This is on a custom hardware. > > Second, there is a (staging) video capture driver in mainline [0], but > the huge delta between 3.1 and mainline makes the two drivers not > trivial to compare. > > Now to the bad news. > > First, that the mainline driver only implements CSI input for > tegra210 and I only have a parallel camera and a tegra20. The > tegra20 documentation I have does not document parallel parallel video > at all. > > Second, the VI is heavily based on the host1x peripheral, and > especially the Synchronization Point (syncpt) feature. Syncpts are > also barely documented and I don't have any previous experience with > this kind of device. And the syncpt (and perhaps host1x too) > implementation in the vendor kernel vs mainline is very different. > > And here's my status. > > The mainline staging driver is already structured to support different > SoCs, so I added tegra20 there, initially as a copy of tegra210. I did > the same to add VIP (parallel input) support in addition to CSI. Then > I filled the gaps as far as I could and now I have register access in > the mainline driver that produces pretty much the same writes as the > vendor kernel does, but it does so using the same driver structure > that is in mainline, quite cleanly. > > Now I am stuck at trying to understand how to make syncpt code work. > As said the syncpt APIs are very different from what I see in the old > vendor kernel. Thus I cannot try to do make the mainline driver behave > like the old one. Moreover the mainline driver does many more calls to > syncpt code than the old one does. > > And finally here are some questions. > > 1. Is there any good documentation on host1x, especially syncpt? > > 2. Is the VIP (parallel input) section of VI (video capture) for > Tegra20 documented anywhere? > > 3. The old driver accesses the syncpt using fixed IDs from #defines > (15 for VI, 16 for CSI), while the mainline driver uses IDs obtained > at runtime from host1x_syncpt_request(). Is it correct that one can > use any of the 32 available syncpts, and the old driver using fixed > values is just an implementation choice? > > 4. I am currently working on a version of > tegra_channel_capture_frame() [1] for tegra20 VIP and other code > involving syncpt. The mainline driver calls many different > host1x_syncpt_*() functions, while the old driver does call very few > of them. Can anybody clarify the meaning of the various syncpt calls? > > 5. In tegra_channel_capture_frame() the call to host1x_syncpt_wait() > always returns -EAGAIN. Where would you start to investigate the > reason of this timeout? > > I think it is enough for today, thank you in advance for any feedback! > > [0] > https://elixir.bootlin.com/linux/latest/source/drivers/staging/media/tegra-video > > [1] > https://elixir.bootlin.com/linux/v5.17.5/source/drivers/staging/media/tegra-video/tegra210.c#L325 Should it be any helpful, here is a reference to the old driver in the vendor 3.1 branch that I am using as a reference: https://github.com/SKIDATA/linux/blob/l4t-3.1.y/drivers/media/video/tegra_v4l2_camera.c -- Luca Ceresoli, Bootlin Embedded Linux and Kernel engineering https://bootlin.com
