Le jeudi 11 avril 2019 à 14:00 +0200, Philipp Zabel a écrit :
> On Thu, 2019-04-11 at 12:18 +0200, Hans Verkuil wrote:
> > On 4/11/19 10:22 AM, Philipp Zabel wrote:
> > > On Wed, 2019-04-10 at 18:24 +0200, Hans Verkuil wrote:
> > > > On 4/10/19 6:11 PM, Nicolas Dufresne wrote:
> > > > > Le mercredi 10 avril 2019 à 16:22 +0200, Philipp Zabel a écrit :
> > > > > > On Wed, 2019-04-10 at 15:43 +0200, Hans Verkuil wrote:
> > > > > > [...]
> > > > > > > > @@ -1126,12 +1127,32 @@ static int coda_enum_frameintervals(struct file *file, void *fh,
> > > > > >
> > > > > > [...]
> > > > > > > Why support VIDIOC_ENUM_FRAMEINTERVALS at all? It makes no sense for a codec.
> > > > > > > I'd remove it altogether.
> > > > >
> > > > > It does make sense, since framerate is the only information that can be
> > > > > used to produce a specific bitrate. If you don't enumerate the rates,
> > > > > then you may endup with a miss-match of what userspace wants, which
> > > > > will result in a different rate then what the user-space anticipated.
> > > > > That being said, I expect these intervals to be really wide. Venus HW
> > > > > uses a Q16 internally, which is precise enough that we could just
> > > > > ignore the interval.
> > > >
> > > > So the problem is that for an encoder where you desires a specific
> > > > constant bitrate, the encoder also needs to know the framerate.
> > > >
> > > > And the driver then would have to support ENUM_FRAMEINTERVALS and the
> > > > S_PARM ioctl so userspace can set the framerate.
> > >
> > > Ah right, that was it. Encoding works fine without ENUM_FRAMEINTERVALS,
> > > as coda has no meaningful frame rate limitations. I had implemented
> > > S_PARM since it is required for CBR encoding, and v4l2-compliance then
> > > complained about ENUM_FRAMEINTERVALS missing:
> > >
> > > cde29ef313de ("[media] coda: Use S_PARM to set nominal framerate for h.264 encoder")
> > > 07b6080d4e6d ("media: coda: implement ENUM_FRAMEINTERVALS")
> >
> > And I think that's right. If you advertise framerate support, then
> > userspace needs to know about the capabilities.
> >
> > I do think ENUM_FRAMEINTERVALS should return some sane values. Right now
> > coda advertises 1/65535 fps to 65535 fps (or thereabout).
>
> Those are the limits imposed by the frame rate registers.
>
> > I would probably return 1 fps to 200 fps (200 Hz is the highest framerate
> > defined by CTA-861-G).
>
> Why? We can certainly encode more than 200 fps in real time if only the
> frame size is small enough, and for offline transcoding the nominal
> frame rate can be anything the coded format supports.
> Smaller than 1 fps might be useful for time lapse videos as well.
>
> > This can also eventually be a helper function in v4l2-mem2mem.c since I expect
> > this to be the same for all (?) encoders.
>
> For example h.264 has an optional 32-bit numerator and denominator
> (num_units_in_tick and time_scale in the Video Usability Information).
> I don't see any reason to limit this artificially in the driver.
>
> > > Hmm, not necesarily. I hadn't thought about that before, but if the
> > > decoder supports frame skipping this could be used to advertise
> > > available reductions in frame rate.
> >
> > That assumes that you want to use S_PARM for that as well. And does the decoder
> > provide the framerate encoded in the bitstream to the driver?
>
> That is a difficult question. Yes, depending on CODA hardware variant,
> codec, and actual bitstream (where this information is optional), the
> firmware may or may not report the frame rate. Maybe it is fundamentally
> not a good idea to try to control frame skipping via the nominal frame
> rate instead of directly.
The Xilinx VCU decoder will allocate a different amount of cores per
session depending on the framerate we give it. This is how they manage
to handle more stream for lower rate/resolution etc. And it makes the
S_PARM call mandatory, since faking one will always lead to some
streams not to work properly, or less streams being handled in
parallel.
I personally don't see why we should in anyway be restrictive to the
information we are allowed to pass to a driver. Specially that is does
not matter if the driver uses it or not. And because there is no way we
will be able to think about every possible use cases popping in the
future, and it does not make backward compatibility to be a little more
open here.
>
> > Using S_PARM for this only works if the driver can know the encoded framerate.
> >
> > Anyway, this is probably something to discuss once a driver supports frame
> > skipping for the decoder.
>
> Right. Since the coda driver does not support frame skipping yet, I'll
> return -ENOTTY for decoder ENUM_FRAMEINTERVALS.
>
> regards
> Philipp
Attachment:
signature.asc
Description: This is a digitally signed message part
