Re: [PATCH v2 7/8] media: hevc: Add scaling matrix control

[Nicolas Dufresne]

Le lundi 14 juin 2021 à 09:50 -0300, Ezequiel Garcia a écrit :
> On Mon, 2021-06-14 at 09:43 +0200, Benjamin Gaignard wrote:
> > 
> > Le 14/06/2021 à 09:27, Hans Verkuil a écrit :
> > > On 10/06/2021 17:44, Benjamin Gaignard wrote:
> > > > HEVC scaling lists are used for the scaling process for transform
> > > > coefficients.
> > > > V4L2_HEVC_SPS_FLAG_SCALING_LIST_ENABLED has to set when they are
> > > > encoded in the bitstream.
> > > Comparing H264 with HEVC I noticed that the corresponding flag for H264 is
> > > called V4L2_H264_PPS_FLAG_SCALING_MATRIX_PRESENT.
> > > 
> > > Should those names be aligned? Also, it is part of PPS for H264 and SPS in HEVC,
> > > is that difference correct?
> > 
> > In ITU specifications ("7.4.3.2.1 General sequence parameter set RBSP semantics") this flag is define like that:
> > scaling_list_enabled_flag equal to 1 specifies that a scaling list is used for the scaling process for transform coefficients.
> > scaling_list_enabled_flag equal to 0 specifies that scaling list is not used for the scaling process for transform coefficients.
> > 
> > So for me the naming is correct.
> > 
> 
> The bitstream is really parsed in userspace (gstreamer, ffmpeg, chromium).
> Not all bitstream parameters need to be passed as-is, because the kernel
> is actually a sort of low-level layer in the decoder stack.
> 
> I think it's probably most appropriate to follow our H.264 interface
> semantics, see https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=54889c51b833d236228f983be16212fbe806bb89.
> 
> The H.264 story goes like this:
> 
> * Default scaling list is used for decoding, but can be modified
>   by a bitstream-provided scaling list.
>    
> * The scaling list modification can be in SPS or in PPS.
> 
> * For each frame, the SPS and PPS headers will tell if
>   a modified scaling list must be used for decoding,
>   and if it's provided in the SPS header or the PPS header.
> 
> The userspace parser must take care of this, and pass
> a scaling matrix to the kernel via V4L2_CID_MPEG_VIDEO_H264_SCALING_MATRIX,
> setting PPS_FLAG_SCALING_MATRIX_PRESENT.
> 
> This flag is at the PPS control, so the scaling modification
> can be applied on each frame.
> 
> In other words, the kernel interface is simpler, it just
> receives a scaling matrix and a flag telling if it must be used or not. 
> 
> We should probably clarify the documentation, so this process is more clear.
> 
> From the HEVC spec, it seems the process should be similar.
> The only difference is the SPS "scaling_list_enabled_flag" parameter,
> which doesn't exist in H.264.
> 
> Nicolas: what do you think?

I believe its a fine "driver convenience". In the sense that the flag might not
have been strictly needed, but may make the driver code simpler. Whenever
possible, and within the spec, I'd agree to keep things as consistant as
possible.

>From the doc, there seems to be a "default" or "flat" version of the scaling
matrix in H264. I think if I had notice this before, I would have pushed forward
the same semantic as the MPEG2 quantisation. In MPEG2, the quantisation control
is set to it's default value in the control framework. What I like of this
approach is that the control becomes always valid. Perhaps there is some
differences here and there I'm not noticing though.

Semantically, it would also be totally different if there is a HW fast path to
the "flat" scaling matrix, in which case you need that flag to enable it. I
think the fact one prepend SPS/PPS is simply to help locate the relevant part of
the specification.

> 
> Thanks,
> Ezequiel
>