Re: [PATCH v7 2/4] media: videodev2: add V4L2_FMT_FLAG_FIXED_RESOLUTION

[Nicolas Dufresne]

Le mardi 11 juin 2019 Ã 20:58 -0400, Nicolas Dufresne a Ãcrit :
> Le mardi 11 juin 2019 Ã 10:52 +0200, Philipp Zabel a Ãcrit :
> > On Wed, 2019-06-05 at 15:39 +0200, Hans Verkuil wrote:
> > > Hi Maxime,
> > > 
> > > I am wondering if this flag shouldn't be inverted: you set
> > > V4L2_FMT_FLAG_DYN_RESOLUTION if dynamic resolution is supported,
> > > otherwise it isn't.
> > > 
> > > Can all the existing mainlined codec drivers handle midstream
> > > resolution changes?
> > > 
> > > s5p-mfc, venus and mediatek can, but I see no SOURCE_CHANGE event in
> > > the coda drivers, so I suspect that that can't handle this.
> > > 
> > > Philipp, what is the status of the coda driver for dynamic resolution
> > > changes?
> > 
> > FTR, to my knowledge there is no dynamic resolution change support in
> > the firmware, as there is no signal (interrupt nor picture run return
> > value) to indicate that different headers were parsed.
> > 
> > I am planning to add the initial source change event required by the
> > current decoder API documentation, but I am afraid there will be no
> > support for source changes due to mid-stream resolution changes due to
> > firmware limitations.
> 
> I'm far from familiar with this IP, but at least on CODA988, I can read
> from the manual that the workflow is to first guess the allocation, and
> if you guess it wrong, an error is returned. What seems to match the
> SOURCE_CHANGE event in that version would be in the picture status
> register, the bit 20, which is documented as triggered if the stream
> requires bigger buffers sizes, or more buffers. After fixing that, you
> should, if I read correctly, retry.
> 
> It does not notify if the buffers are too large, but you can detect,
> since there is register with the output stream information. This
> basically means that for V4L2 restriction, you'd have to bounce the
> buffers on frame size boundary or something like thisÃ
> 
> This workflow is very similar to how OMX works, but V4L2 is even less
> flexible on allocation vs format, forcing more re-allocation.

Oh, actually, you already bounce the buffers through the color
converter, so I guess this can all be hidden internally.

> 
> > regards
> > Philipp