On 02/07/2017 08:21 AM, Hugues FRUCHET wrote:
Hi,
Here is an update regarding MPEG-2 implementation based on ST video decoder:
* MPEG-2 API + DELTA kernel driver:
http://www.mail-archive.com/linux-media@vger.kernel.org/msg107405.html
* libv4l-codecparsers plugin including MPEG-2 back-end:
http://www.mail-archive.com/linux-media@vger.kernel.org/msg107812.html
Please note that this is implemented & functional using currently available
V4L2 control framework (no Request API), assuming that user side keeps unicity of
S_EXT_CTRL() / QBUF(OUTPUT) pair.
Request API will remove this constraint, but the point here is to define
control interface, as far as I have understood code, Request API will not
affect those control definitions.
Some updates inline thereafter regarding activities on this subject; me for
MPEG-2 on ST platform and Randy Li, Nicolas Dufresne, Ayaka for H264 on
Rockchip platform:
On 11/14/2016 10:55 AM, Hans Verkuil wrote:
On 10/27/2016 09:42 AM, Hugues FRUCHET wrote:
Hi,
This RFC aims to start discussions in order to define the codec
specific controls structures to fulfill the low-level decoder API
needed by non "Stream API" based decoders ("stateless" or "Frame API" based
decoders).
Let's refer to this as 'stateful' decoders and 'stateless' decoders.
This is the preferred terminology (and much more descriptive than
'Stream' vs 'Frame'). It's also not really a new API, although it does
rely on the Request API.
Several implementation exists now which runs on several SoC and
various software frameworks.
The idea is to find the communalities between all those
implementations and SoC to define a single unified interface in V4L2 includes.
Even if "Request API" is needed to pass those codec specific controls
from userspace down to kernel on a per-buffer basis, we can start
discussions and define the controls in parallel of its development.
We can even propose some implementations based on existing V4L2
control framework (which doesn't support "per-frame" basis) by
ensuring atomicity of sequence S_EXT_CTRL(header[i])/QBUF(stream[i]).
Constraint can then be relaxed when "Request API" is merged.
I would like to propose to work on a "per-codec" basis, having at
least
2 different SoC and 2 different frameworks to test and validate controls.
To do so, I have tried to identify some people that have worked on
this subject and have proposed some implementations, feel free to
correct me and enhance the list if needed:
* MPEG2/MPEG4
- Florent Revest for Allwinner A13 CedarX support [1] tested with
VLC -> libVA + sunxi-cedrus-drv-video -> V4L2
- Myself for STMicroelectronics Delta support [2] tested with
GStreamer V4L2 -> libv4l2 + libv4l-delta plugin -> V4L2
Available on ST platform with [2] & [2.1] patchset series.
* VP8
- Pawel Osciak for Rockchip RK3288, RK3399? VPU Support [3] tested
with Chromium -> V4L2
- Jung Zhao for Rockchip RK3288 VPU support [4] <cannot find the
framework used>
* H264
- Pawel Osciak for Rockchip RK3288, RK3399? VPU Support [5] tested
with Chromium -> V4L2
- Randy Li for Rockchip RK3288 VPU support [6] tested with VLC? ->
libVA + rockchip-va-driver -> V4L2 VLC? -> libVDPAU +
rockchip-va-driver -> V4L2
Tested with Gstreamer -> VA-API element -> Rockchip VA-API driver -> V4L2
https://github.com/rockchip-linux/libvdpau-rockchip
Study on-going for H264 userland/kernel partitioning in this thread:
Request API: stateless VPU: the buffer mechanism and DPB management:
http://www.mail-archive.com/linux-media@vger.kernel.org/msg107165.html
I can work to define MPEG2/MPEG4 controls and propose functional
implementations for those codecs, and will be glad to co-work with
you Florent.
I can help on H264 on a code review basis based on the functional
H264 setup I have in-house and codec knowledge, but I cannot provide
implementation in a reasonable timeframe, same for VP8.
Apart of very details of each codec, we have also to state about
generic concerns such as:
- new pixel format introduction (VP8 => VP8F, H264 => S264, MPG2 =>
MG2F, MPG4 => MG4F)
- new device caps to indicate that driver requires extra headers ?
maybe not needed because redundant with new pixel format
That's indeed typically signaled through the pixelformat.
[2] is implemented this way in [media] v4l: add parsed MPEG-2 support, two new pixel format MG1P
and MG2P ("P" for "Parsed") have been introduced:
+#define V4L2_PIX_FMT_MPEG1_PARSED v4l2_fourcc('M', 'G', '1', 'P') /*
MPEG1 with parsing metadata given through controls */
+#define V4L2_PIX_FMT_MPEG2_PARSED v4l2_fourcc('M', 'G', '2', 'P') /*
MPEG2 with parsing metadata given through controls */
libv4l plugin [2.1] intercepts the pixel format information negotiated
between user and driver (enum_fmt/try_fmt/get_fmt/s_fmt) and selects the
right parser to use to convert MPEG-2 compressed format to new "MPEG-2
parsed" compressed format required by kernel driver.
Plugin is designed to support several formats.
- continue to modify v4l2-controls.h ? or do we add some new specific
header files (H264 is huge!) ?
- how to manage sequence header & picture header, optional/extended
controls (MPEG2 sequence/picture extensions, H264 SEI, ...). Personally
I have added flags inside a single control structure, H264 is done in a
different way using several controls (SPS/PPS/SLICE/DECODE/...)
Thanks you to all of you for your attention and feel free to react on
this topic if you are interested to work on this subject.
As long as the V4L2 driver underpins the various solutions I am happy :-)
I do think that having a libv4l plugin will be useful since it will make
it easy for applications that support the stateful decoder to use the
same code for a stateless decoder by seamlessly using the plugin.
This does not prevent other approaches at the same time, of course.
Regards,
Hans
[2] implements new extended controls for MPEG-2 in [media] v4l: add
parsed MPEG-2 support
http://www.mail-archive.com/linux-media@vger.kernel.org/msg107406.html
This has been done in uapi/linux/v4l2-controls.h as extension of already
existing MPEG video controls, defining one control per "header" (so not
using a single control with selection flag):
#define V4L2_CID_MPEG_VIDEO_MPEG2_SEQ_HDR
#define V4L2_CID_MPEG_VIDEO_MPEG2_SEQ_EXT
#define V4L2_CID_MPEG_VIDEO_MPEG2_SEQ_DISPLAY_EXT
#define V4L2_CID_MPEG_VIDEO_MPEG2_SEQ_MATRIX_EXT
#define V4L2_CID_MPEG_VIDEO_MPEG2_PIC_HDR
#define V4L2_CID_MPEG_VIDEO_MPEG2_PIC_EXT
Those controls and their associated data structure have been defined
based on MPEG-2 standard ISO/IEC 13818-2.
Best regards,
Hugues.
[0] [ANN] Codec & Request API Brainstorm meeting Oct 10 &
11https://www.spinics.net/lists/linux-media/msg106699.html
[1] MPEG2 A13 CedarXhttp://www.spinics.net/lists/linux-media/msg104823.html
[1] MPEG4 A13 CedarXhttp://www.spinics.net/lists/linux-media/msg104817.html
[2] MPEG2 STi4xx
Deltahttp://www.spinics.net/lists/linux-media/msg106240.html
[2] MPEG2 DELTA kernel driver: http://www.mail-archive.com/linux-
me...@vger.kernel.org/msg107405.html
[2.1] MPEG2 libv4l plugin:
http://www.mail-archive.com/linux-media@vger.kernel.org/msg107812.html
[2] MPEG4 STi4xx Delta is also supported but not yet pushed
[3] VP8 Rockchip RK3288, RK3399?
VPUhttps://chromium.googlesource.com/chromiumos/overlays/chromiumos-overlay/+/refs/heads/master/sys-kernel/linux-headers/files/0002-CHROMIUM-v4l-Add-VP8-low-level-decoder-API-controls.patch
[4] VP8 Rockchip RK3288
VPUhttp://www.spinics.net/lists/linux-media/msg97997.html
[5] H264 Rockchip RK3288, RK3399?
VPUhttps://chromium.googlesource.com/chromiumos/overlays/chromiumos-overlay/+/refs/heads/master/sys-kernel/linux-headers/files/0001-CHROMIUM-media-headers-Import-V4L2-headers-from-Chro.patch
[6] H264 Rockchip RK3288
VPUhttp://www.spinics.net/lists/linux-media/msg105095.html
https://github.com/rockchip-linux/libvdpau-rockchip
https://github.com/rockchip-linux/gstreamer-rockchip
https://github.com/rockchip-linux/mpp
