Em 29-07-2011 05:36, Laurent Pinchart escreveu: > Hi Mauro, > > On Friday 29 July 2011 06:02:54 Mauro Carvalho Chehab wrote: >> Em 28-07-2011 19:57, Sylwester Nawrocki escreveu: >>> On 07/28/2011 03:20 PM, Mauro Carvalho Chehab wrote: >>>> Accumulating sub-dev controls at the video node is the right thing to >>>> do. >>>> >>>> An MC-aware application will need to handle with that, but that doesn't >>>> sound to be hard. All such application would need to do is to first >>>> probe the subdev controls, and, when parsing the videodev controls, not >>>> register controls with duplicated ID's, or to mark them with some >>>> special attribute. >>> >>> IMHO it's not a big issue in general. Still, both subdev and the host >>> device may support same control id. And then even though the control ids >>> are same on the subdev and the host they could mean physically different >>> controls (since when registering a subdev at the host driver the host's >>> controls take precedence and doubling subdev controls are skipped). >> >> True, but, except for specific usecases, the host control is enough. > > Not for embedded devices. In many case the control won't even be implemented > by the host. If your system has two sensors connected to the same host, they > will both expose an exposure time control. Which one do you configure through > the video node ? The two sensors will likely have different bounds for the > same control, how do you report that ? If the device has two sensors that are mutually exclusive, they should be mapped as two different inputs. The showed control should be the one used by the currently active input. If the sensors aren't mutually exclusive, then two different video nodes will be shown in userspace. > Those controls are also quite useless for generic V4L2 applications, which > will want auto-exposure anyway. This needs to be implemented in userspace in > libv4l. Several webcams export exposure controls. Why shouldn't those controls be exposed to userspace anymore? Ok, if the hardware won't support 3A algorithm, libv4l will implement it, eventually using an extra hardware-aware code to get the best performance for that specific device, but this doesn't mean that the user should always use it. Btw, the 3A algorithm is one of the things I don't like on my cell phone: while it works most of the time, sometimes I want to disable it and manually adjust, as it produces dark images, when there's a very bright light somewhere on the image background. Manually adjusting the exposure time and aperture is something relevant for some users. >>> Also there might be some preference at user space, at which stage of the >>> pipeline to apply some controls. This is where the subdev API helps, and >>> plain video node API does not. >> >> Again, this is for specific usecases. On such cases, what is expected is >> that the more generic control will be exported via V4L2 API. >> >>>>> Thus it's a bit hard to imagine that we could do something like >>>>> "optionally not to inherit controls" as the subdev/MC API is optional. >>>>> :) >>>> >>>> This was actually implemented. There are some cases at ivtv/cx18 driver >>>> where both the bridge and a subdev provides the same control (audio >>>> volume, for example). The idea is to allow the bridge driver to touch >>>> at the subdev control without exposing it to userspace, since the >>>> desire was that the bridge driver itself would expose such control, >>>> using a logic that combines changing the subdev and the bridge >>>> registers for volume. >>> >>> This seem like hard coding a policy in the driver;) Then there is no way >>> (it might not be worth the effort though) to play with volume level at >>> both devices, e.g. to obtain optimal S/N ratio. >> >> In general, playing with just one control is enough. Andy had a different >> opinion when this issue were discussed, and he thinks that playing with >> both is better. At the end, this is a developers decision, depending on >> how much information (and bug reports) he had. > > ivtv/cx18 is a completely different use case, where hardware configurations > are known, and experiments possible to find out which control(s) to use and > how. In this case you can't know in advance what the sensor and host drivers > will be used for. Why not? I never saw an embedded hardware that allows physically changing the sensor. > Even if you did, fine image quality tuning requires > accessing pretty much all controls individually anyway. The same is also true for non-embedded hardware. The only situation where V4L2 API is not enough is when there are two controls of the same type active. For example, 2 active volume controls, one at the audio demod, and another at the bridge. There may have some cases where you can do the same thing at the sensor or at a DSP block. This is where MC API gives an improvement, by allowing changing both, instead of just one of the controls. >>> This is a hack...sorry, just joking ;-) Seriously, I think the >>> situation with the userspace subdevs is a bit different. Because with one >>> API we directly expose some functionality for applications, with other >>> we code it in the kernel, to make the devices appear uniform at user >>> space. >> >> Not sure if I understood you. V4L2 export drivers functionality to >> userspace in an uniform way. MC api is for special applications that might >> need to access some internal functions on embedded devices. >> >> Of course, there are some cases where it doesn't make sense to export a >> subdev control via V4L2 API. >> >>>>> Also, the sensor subdev can be configured in the video node driver as >>>>> well as through the subdev device node. Both APIs can do the same >>>>> thing but in order to let the subdev API work as expected the video >>>>> node driver must be forbidden to configure the subdev. >>>> >>>> Why? For the sensor, a V4L2 API call will look just like a bridge driver >>>> call. The subdev will need a mutex anyway, as two MC applications may >>>> be opening it simultaneously. I can't see why it should forbid changing >>>> the control from the bridge driver call. >>> >>> Please do not forget there might be more than one subdev to configure and >>> that the bridge itself is also a subdev (which exposes a scaler >>> interface, for instance). A situation pretty much like in Figure 4.4 [1] >>> (after the scaler there is also a video node to configure, but we may >>> assume that pixel resolution at the scaler pad 1 is same as at the video >>> node). Assuming the format and crop configuration flow is from sensor to >>> host scaler direction, if we have tried to configure _all_ subdevs when >>> the last stage of the pipeline is configured (i.e. video node) the whole >>> scaler and crop/composition configuration we have been destroyed at that >>> time. And there is more to configure than VIDIOC_S_FMT can do. >> >> Think from users perspective: all user wants is to see a video of a given >> resolution. S_FMT (and a few other VIDIOC_* calls) have everything that >> the user wants: the desired resolution, framerate and format. >> >> Specialized applications indeed need more, in order to get the best images >> for certain types of usages. So, MC is there. >> >> Such applications will probably need to know exactly what's the sensor, >> what are their bugs, how it is connected, what are the DSP blocks in the >> patch, how the DSP algorithms are implemented, etc, in order to obtain the >> the perfect image. >> >> Even on embedded devices like smartphones and tablets, I predict that both >> types of applications will be developed and used: people may use a generic >> application like flash player, and an specialized application provided by >> the manufacturer. Users can even develop their own applications generic >> apps using V4L2 directly, at the devices that allow that. >> >> As I said before: both application types are welcome. We just need to >> warrant that a pure V4L application will work reasonably well. > > That's why we have libv4l. The driver simply doesn't receive enough > information to configure the hardware correctly from the VIDIOC_* calls. And > as mentioned above, 3A algorithms, required by "simple" V4L2 applications, > need to be implemented in userspace anyway. It is OK to improve users experience via libv4l. What I'm saying is that it is NOT OK to remove V4L2 API support from the driver, forcing users to use some hardware plugin at libv4l. >>> Allowing the bridge driver to configure subdevs at all times would >>> prevent the subdev/MC API to work. >> >> Well, then we need to think on an alternative for that. It seems an >> interesting theme for the media workshop at the Kernel Summit/2011. >> >>>>> There is a conflict there that in order to use >>>>> 'optional' API the 'main' API behaviour must be affected.... >>>> >>>> It is optional from userspace perspective. A V4L2-only application >>>> should be able to work with all drivers. However, a MC-aware >>>> application will likely be specific for some hardware, as it will need >>>> to know some device-specific stuff. >>>> >>>> Both kinds of applications are welcome, but dropping support for >>>> V4L2-only applications is the wrong thing to do. >>>> >>>>> And I really cant use V4L2 API only as is because it's too limited. >>>> >>>> Why? >>> >>> For instance there is really yet no support for scaler and composition >>> onto a target buffer in the Video Capture Interface (we also use sensors >>> with built in scalers). It's difficult to efficiently manage >>> capture/preview pipelines. It is impossible to discover the system >>> topology. >> >> Scaler were always supported by V4L2: if the resolution specified by S_FMT >> is not what the sensor provides, then scale. All non-embedded drivers with >> sensor or bridge supports scale does that. >> >> Composition is not properly supported yet. It could make sense to add it to >> V4L. How do you think MC API would help with composite? >> >> Managing capture/preview pipelines will require some support at V4L2 level. >> This is a problem that needs to be addressed there anyway, as buffers for >> preview/capture need to be allocated. There's an RFC about that, but I >> don't think it covers the pipelines for it. > > Managing pipelines is a policy decision, and needs to be implemented in > userspace. Once again, the solution here is libv4l. If V4L2 API is not enough, implementing it on libv4l won't solve, as userspace apps will use V4L2 API for requresting it. >> Discovering the system topology indeed is not part of V4L2 API and will >> never be. This is MC API business. There's no overlap with V4L2. >> >>>>> Might be that's why I see more and more often migration to OpenMAX >>>>> recently. >>>> >>>> I don't think so. People may be adopting OpenMAX just because of some >>>> marketing strategy from the OpenMAX forum. We don't spend money to >>>> announce V4L2 ;) >>>> >>> :) >>> : >>>> I think that writing a pure OpenMAX driver is the wrong thing to do, as, >>>> at the long term, it will cost _a_lot_ for the vendors to maintain >>>> something that will never be merged upstream. >>> >>> In general it depends on priorities. For some chip vendors it might be >>> more important to provide a solution in short time frame. Getting >>> something in mainline isn't effortless and spending lot's of time on >>> this for some parties is unacceptable. >> >> Adding something at a forum like OpenMAX is probably not an easy task ;) It >> generally takes a long time to change something on open forum >> specifications. Also, participating on those forums require lots of money, >> with membership and travel expenses. > > Correct me if I'm wrong (I'm not an OpenMAX specialist), but I think it > doesn't. OpenMAX isn't really open, and vendors implement proprietary > extensions as needed. My understanding is that tt's easier for them to use > OpenMAX for that, because they can more or less claim OpenMAX support even > with the proprietary extensions, something they couldn't do with V4L2. > > If we want to compete there (and I certainly do), we need to provide vendors > with clear, easy and well documented APIs that fulfill their needs out of the > box. That's of course an evermoving target. Agreed. > >> Adding a new driver that follows the API's is not a long-term effort. The >> delay is basically one kernel cycle, e. g. about 3-4 months. >> >> Most of the delays on merging drivers for embedded systems, however, take a >> way longer than that, unfortunately. From what I see, what is delaying >> driver submission is due to: >> - the lack of how to work with the Linux community. Some developers take a >> long time to get the 'modus operandi'; >> - the need of API changes. It is still generally faster to add a new API >> addition at the Kernel than on an open forum; >> - the discussions inside the various teams (generally from the same >> company, or the company and their customers) about the better way to >> implement some feature. >> >> All the above also happens when developing an OpenMAX driver: companies >> need to learn how to work with the OpenMax forums, API changes may be >> required, so they need to participate at the forums, the several internal >> teams and customers will be discussing the requirements. >> >> I bet that there's also one additional step: to submit the driver to some >> company that will check the driver compliance with the official API. Such >> certification process is generally expensive and takes a long time. > > I could be wrong, but I very much doubt this ever happens. > >> At the end of the day, they'll also spend a lot of time to have the driver >> done, or they'll end by releasing a "not-quite-openmax" driver, and then >> needing to rework on that, due to customers complains, or to >> certification-compliance, loosing time and money. > -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
