Hi Tomasz,

On Thu, Jun 14, 2018 at 10:00:43PM +0900, Tomasz Figa wrote:

Hi Maxime,

On Wed, Jun 13, 2018 at 11:07 PM Maxime Ripard
[off-list ref] wrote:

quoted

Hi,

Here is a preliminary version of the H264 decoding support in the
cedrus driver.

Thanks for the series! Let me reply inline to some of the points raised here.

quoted

As you might already know, the cedrus driver relies on the Request
API, and is a reverse engineered driver for the video decoding engine
found on the Allwinner SoCs.

This work has been possible thanks to the work done by the people
behind libvdpau-sunxi found here:
https://github.com/linux-sunxi/libvdpau-sunxi/

This driver is based on the last version of the cedrus driver sent by
Paul, based on Request API v13 sent by Hans:
https://lkml.org/lkml/2018/5/7/316

Just FYI, there is v15 already. :)

Yeah, we know, Paul is currently working on rebasing to that version :)

quoted

This driver has been tested only with baseline profile videos, and is
missing a few key features to decode videos with higher profiles.
This has been tested using our cedrus-frame-test tool, which should be
a quite generic v4l2-to-drm decoder using the request API to
demonstrate the video decoding:
https://github.com/free-electrons/cedrus-frame-test/, branch h264

However, sending this preliminary version, I'd really like to start a
discussion and get some feedback on the user-space API for the H264
controls exposed through the request API.

I've been using the controls currently integrated into ChromeOS that
have a working version of this particular setup. However, these
controls have a number of shortcomings and inconsistencies with other
decoding API. I've worked with libva so far, but I've noticed already
that:

Note that these controls are supposed to be defined exactly like the
bitstream headers deserialized into C structs in memory. I believe
Pawel (on CC) defined them based on the actual H264 specification.

quoted

  - The kernel UAPI expects to have the nal_ref_idc variable, while
    libva only exposes whether that frame is a reference frame or
    not. I've looked at the rockchip driver in the ChromeOS tree, and
    our own driver, and they both need only the information about
    whether the frame is a reference one or not, so maybe we should
    change this?

The fact that 2 drivers only need partial information doesn't mean
that we should ignore the data being already in the bitstream. IMHO
this API should to provide all the metadata available in the stream to
the kernel driver, as a replacement for bitstream parsing in firmware
(or in kernel... yuck).

The point is more that libva will only pass the result of (nal_ref_idc
!= 0). So in the libva plugin, you won't be able to fill the proper
value to the kernel, since you don't have access to it.

quoted

  - The H264 bitstream exposes the picture default reference list (for
    both list 0 and list 1), the slice reference list and an override
    flag. The libva will only pass the reference list to be used (so
    either the picture default's or the slice's) depending on the
    override flag. The kernel UAPI wants the picture default reference
    list and the slice reference list, but doesn't expose the override
    flag, which prevents us from configuring properly the
    hardware. Our video decoding engine needs the three information,
    but we can easily adapt to having only one. However, having two
    doesn't really work for us.

Where does the override flag come from? If it's in the bitstream, then
I guess it was just missed when creating the structures.

It's in the bitstream yeah. I'll add it then.

Maxime

-- 
Maxime Ripard, Bootlin (formerly Free Electrons)
Embedded Linux and Kernel engineering
https://bootlin.com