[PATCH 0/3] Enable JPEG Encoder on RK3566/RK3568

[Nicolas Dufresne]

Le lundi 20 juin 2022 à 15:40 +0200, Michael Grzeschik a écrit :
> Hi Nicolas,
> Hi Andrzej,
> 
> On Mon, May 02, 2022 at 08:20:50AM -0400, Nicolas Dufresne wrote:
> > Le samedi 30 avril 2022 à 01:20 +0200, Michael Grzeschik a écrit :
> > > Since the RK3568 seems to get more and more of attention at the moment
> > > I would like to ask if somebody is planning to write support the RKVENC
> > > in mainline. That is the VEPU540 core refered to in the RK3568 TRM.
> > > 
> > > I would start with that in the next weeks, taking the bootlin H1 support
> > > as an reference to handle the request_api for the encoder part. Which is
> > > currently completely untouched in mainline AFAIK.
> > 
> > I'm very happy to ear you'd be interest in helping with this. I'm adding Andrzej
> > in CC, as he started some base work using VP8 encoder on RK3399 (same HW found
> > on other RK SoC) to make this possible. Note that these are a new type of
> > encoders and a specification is needed to ensure they all have a coherent work-
> > flow. The first step was obviously to have working prototype, Bootlin started
> > with H264 (but never posted anything on the mailing list). I'm under the
> > impression they have abandoned it. Andrzej have decided to use VP8 as it is
> > simpler. The prototype should help understand the basic flow of an encoder and
> > produce a specification for this.
> 
> I just have tested the bootlin v4l2-h264-encoder stack on the rk3399
> with v5.19-rc2. It is working as expected.
> 
> Since there have been some rumours about using the VEPU121 core on the
> rk3568, which could have h264 support as well. I have tested this
> stack on that core as well.
> 
> The findings are; It is producing some data, but the stream created
> is not possible to be decoded. Tested with vlc and gstreamer.

Oh well, I'd say unless vendor document it, we'll not going to just enable it
there. There is another (and faster) encoder there with an vendor but fully Open
Source stack anyway. Thanks for the extra step here.

> 
> Be it due to misconfiguration or the core is really not h264 capable.
> 
> Find the latest code here:
> 
> https://git.pengutronix.de/cgit/mgr/linux/log/?h=v5.19/topic/rk3568-vepu-h264-stateless-bootlin
> 
> > From there, the "easy" part is to make per codec controls, to configure the
> > encoder. Demonstrating fixed QP, this is about were I believe Bootlin stopped.
> 
> I have seen that the userspace tool is filling in some initial values
> in the PPS and SPS headers, which are not adjustable. I bet you refer to that.
> 
> > And I personally believe some OSS userland (not just tests) that can handle
> > multiple reference, perhaps more advance GOP pattern and some basic rate control
> > would help build confidence in the uAPI.
> 
> I found the following CTRL Types being defined in the bootlin stack
> for the request API.
> 
> V4L2_CID_STATELESS_H264_ENCODE_PARAMS
> V4L2_CID_STATELESS_H264_ENCODE_RC
> V4L2_CID_STATELESS_H264_ENCODE_FEEDBACK
> 
> The FEEDBACK CID is the information we gain after one frame was
> encoded. With this data the RC (Rate Control) parameter is prepared for
> the next frame.
> 
> You mean to have some user interface to make the rate control user
> configurable?

RC feedback controls are going to all be vendor specific. This one cannot be
called with a generic name like "ENCODE_FEEDBACK" since it is not generic. In
fact, news Hantro chip don't have this feedback, they have something else but it
is not used by the vendor software (not open source, but you can find leaks I
suppose). They only use the size of the compressed frames, and do RC in software
there.

In addition to that, with the current trend, "feedback" controls would have to
use RO Request, so that one can still queue multiple encode and be able to match
the feedback to a specific request/frame.

> 
> > Let us know how we can help, we should >void doing the same thing, as
> > this is already quite a large project that can easily take over a year
> > to become mainline ready.
> 
> It would probably help to share some codebase. If you already have
> something for that rkvenc codec, be it VP8 support. It would probably
> make sense for me to build up on that. Or to discuss common code
> patterns on structures.

We are doing the exercise with VP8 using the original Hantro G1 design, as fond
on IMX8MQ and RK3288. Though we only test on NXP boards at the moment.

Andrzej is on vacation this week, I'll check with him why hist draft isn't
accessible to the public. Meanwhile, the userland / GStreamer part is drafted
(careful its a draft) here:

https://gitlab.freedesktop.org/benjamin.gaignard1/gstreamer/-/tree/WIP_V4L2_VP8_STATELESS_ENCODER

Currently we are still working on very basic bit of the puzzle, understanding
the problem in order to produce a proper Stateless Encoder specification. As an
example, the Hantro VP8 encoder brought interesting aspect, as an example it
produces 2 buffers. Shall we offset these into 1 buffer, shall we use planes to
zero copy that to user, or shall we memcpy in the driver .. Shall the VP8 frame
header be produced by the driver (even though the HW don't produce it) or shall
we have more stateless VP8 format for both cases. Its a big series of tiny
things like this that we need to study and weight, and compare against other
Hardware designs.

regards,
Nicolas