Re: [PATCH v3 3/9] DocBook/v4l: Add compressed video formats used on MT8173 codec driver

[Nicolas Dufresne]

Le mercredi 13 juillet 2016 Ã 10:00 +0800, tiffany lin a ÃcritÂ:
> Hi Nicolas,
> 
> On Tue, 2016-07-12 at 15:14 -0400, Nicolas Dufresne wrote:
> > Le mardi 12 juillet 2016 Ã 15:08 -0400, Nicolas Dufresne a Ãcrit :
> > > Le mardi 12 juillet 2016 Ã 16:16 +0800, Wu-Cheng Li (æåè) a Ãcrit
> :
> > > > Decoder hardware produces MT21 (compressed). Image processor
> can
> > > > convert it to a format that can be input of display driver.
> > > > Tiffany.
> > > > When do you plan to upstream image processor (mtk-mdp)?
> > > > >Â
> > > > > It can be as input format for encoder, MDP and display
> drivers in
> > > > our
> > > > > platform.
> > > > I remember display driver can only accept uncompressed MT21.
> Right?
> > > > Basically V4L2_PIX_FMT_MT21 is compressed and is like an opaque
> > > > format. It's not usable until it's decompressed and converted
> by
> > > > image
> > > > processor.
> > >Â
> > > Previously it was described as MediaTek block mode, and now as a
> > > MediaTek compressed format. It makes me think you have no idea
> what
> > > this pixel format really is. Is that right ?
> > >Â
> > > The main reason why I keep asking, is that we often find
> similarities
> > > between what vendor like to call their proprietary formats. Doing
> the
> > > proper research helps not creating a mess like in Android where
> you
> > > have a lot of formats that all point to the same format. I
> believe
> > > there was the same concern when Samsung wanted to introduce their
> Z-
> > > flip-Z NV12 tile format. In the end they simply provided
> sufficient
> > > documentation so we could document it and implement software
> > > converters
> > > for test and validation purpose.
> >Â
> > Here's the kind of information we want in the documentation.
> >Â
> > https://chromium.googlesource.com/chromium/src/media/+/master/base/
> vide
> > o_types.h#40
> >Â
> >ÂÂ // MediaTek proprietary format. MT21 is similar to NV21 except
> the memory
> >ÂÂ // layout and pixel layout (swizzles). 12bpp with Y plane
> followed by a 2x2
> >ÂÂ // interleaved VU plane. Each image contains two buffers -- Y
> plane and VU
> >ÂÂ // plane. Two planes can be non-contiguous in memory. The
> starting addresses
> >ÂÂ // of Y plane and VU plane are 4KB alignment.
> >ÂÂ // Suppose image dimension is (width, height). For both Y plane
> and VU plane:
> >ÂÂ // Row pitch = ((width+15)/16) * 16.
> >ÂÂ // Plane size = Row pitch * (((height+31)/32)*32)
> >Â
> > Now obviously this is incomplete, as the swizzling need to be
> documented of course.
> >Â
> Because it's finally a compressed format from our codec hw, we cannot
> describe its swizzling.

Can you clarify why it cannot be described ? I don't buy any argument
that it's impossible to convert without hardware. Intel did document
their compressed formats.

Debugging an integration with such format that we have literally
decided to no make public is a pain. A lot of decoder got abandoned or
never used because of that.

It would be nice to explain the architecture you are suggesting to make
this decoder useful. How will userspace application be able to use your
decoder. Will the image processor usable outside the display path ?
People might want to do transcoding, or other relatively common task
that does not imply a display.

regards,
Nicolas

p.s.
A small note to Wu-Cheng, the definition of the DRM format in your
Chromium branch is not correct. DRM uses format modifier, such that the
format is the family, in your case it's most likely NV21, and the
modifier is the compression or tiling algorithm in place (or both as
needed).