Message-ID: <11665334.4LxpHo8hza@jernej-laptop>
Date:   Tue, 01 Oct 2019 00:58:35 +0200
From:   Jernej Škrabec <jernej.skrabec@...l.net>
To:     Hans Verkuil <hverkuil-cisco@...all.nl>
Cc:     mchehab@...nel.org, paul.kocialkowski@...tlin.com,
        mripard@...nel.org, pawel@...iak.com, m.szyprowski@...sung.com,
        kyungmin.park@...sung.com, tfiga@...omium.org, wens@...e.org,
        gregkh@...uxfoundation.org, boris.brezillon@...labora.com,
        linux-media@...r.kernel.org, linux-kernel@...r.kernel.org,
        devel@...verdev.osuosl.org, linux-arm-kernel@...ts.infradead.org,
        ezequiel@...labora.com, jonas@...boo.se
Subject: Re: [PATCH v2 0/6] media: cedrus: h264: Support multi-slice frames

Dne torek, 01. oktober 2019 ob 00:43:34 CEST je Hans Verkuil napisal(a):
> On 10/1/19 12:27 AM, Jernej Škrabec wrote:
> > Dne ponedeljek, 30. september 2019 ob 10:10:48 CEST je Hans Verkuil
> > 
> > napisal(a):
> >> On 9/29/19 10:00 PM, Jernej Skrabec wrote:
> >>> This series adds support for decoding multi-slice H264 frames along with
> >>> support for V4L2_DEC_CMD_FLUSH and V4L2_BUF_FLAG_M2M_HOLD_CAPTURE_BUF.
> >>> 
> >>> Code was tested by modified ffmpeg, which can be found here:
> >>> https://github.com/jernejsk/FFmpeg, branch mainline-test
> >>> It has to be configured with at least following options:
> >>> --enable-v4l2-request --enable-libudev --enable-libdrm
> >>> 
> >>> Samples used for testing:
> >>> http://jernej.libreelec.tv/videos/h264/BA1_FT_C.mp4
> >>> http://jernej.libreelec.tv/videos/h264/h264.mp4
> >>> 
> >>> Command line used for testing:
> >>> ffmpeg -hwaccel drm -hwaccel_device /dev/dri/card0 -i h264.mp4 -pix_fmt
> >>> bgra -f fbdev /dev/fb0
> >>> 
> >>> Please note that V4L2_DEC_CMD_FLUSH was not tested because I'm
> >>> not sure how. ffmpeg follows exactly which slice is last in frame
> >>> and sets hold flag accordingly. Improper usage of hold flag would
> >>> corrupt ffmpeg assumptions and it would probably crash. Any ideas
> >>> how to test this are welcome!
> >> 
> >> It can be tested partially at least: if ffmpeg tells you it is the last
> >> slice, then queue the slice with the HOLD flag set, then call FLUSH
> >> afterwards. This should clear the HOLD flag again. In this case the
> >> queued
> >> buffer is probably not yet processed, so the flag is cleared before the
> >> decode job starts.
> >> 
> >> You can also try to add a sleep before calling FLUSH to see what happens
> >> if the last queued buffer is already decoded.
> > 
> > Ok, I tried following code:
> > https://github.com/jernejsk/FFmpeg/blob/flush_test/libavcodec/
> > v4l2_request.c#L220-L233
> > 
> > But results are not good. It seems that out_vb in flush command is always
> > NULL and so it always marks capture buffer as done, which leads to kernel
> > warnings.
> > 
> > dmesg output with debugging messages is here: http://ix.io/1Ks8
> > 
> > Currently I'm not sure what is going on. Shouldn't be output buffer queued
> > and waiting to MEDIA_REQUEST_IOC_QUEUE which is executed after flush
> > command as it can be seen from ffmpeg code linked above?
> 
> Argh, I forgot about the fact that this uses requests.
> 
> The FLUSH should happen *after* the MEDIA_REQUEST_IOC_QUEUE ioctl. Otherwise
> it has no effect. As long as the request hasn't been queued, the buffer is
> also not queued to the driver, so out_vb will indeed be NULL.

Well, flush cmd has effect if it is called before MEDIA_REQUEST_IOC_QUEUE ioctl 
as it can be seen from linked dmesg output. I guess there is nothing that we 
can do to prevent wrong usage?

BTW, if capture buffer is marked as done, shouldn't be also removed from the 
queue?

Best regards,
Jernej

> 
> Sorry for the confusion.
> 
> Regards,
> 
> 	Hans
> 
> > Best regards,
> > Jernej
> > 
> >> Regards,
> >> 
> >> 	Hans
> >> 	
> >>> Thanks to Jonas for adjusting ffmpeg.
> >>> 
> >>> Please let me know what you think.
> >>> 
> >>> Best regards,
> >>> Jernej
> >>> 
> >>> Changes from v1:
> >>> - added Rb tags
> >>> - updated V4L2_DEC_CMD_FLUSH documentation
> >>> - updated first slice detection in Cedrus
> >>> - hold capture buffer flag is set according to source format
> >>> - added v4l m2m stateless_(try_)decoder_cmd ioctl helpers
> >>> 
> >>> Hans Verkuil (2):
> >>>   vb2: add V4L2_BUF_FLAG_M2M_HOLD_CAPTURE_BUF
> >>>   videodev2.h: add V4L2_DEC_CMD_FLUSH
> >>> 
> >>> Jernej Skrabec (4):
> >>>   media: v4l2-mem2mem: add stateless_(try_)decoder_cmd ioctl helpers
> >>>   media: cedrus: Detect first slice of a frame
> >>>   media: cedrus: h264: Support multiple slices per frame
> >>>   media: cedrus: Add support for holding capture buffer
> >>>  
> >>>  Documentation/media/uapi/v4l/buffer.rst       | 13 ++++++
> >>>  .../media/uapi/v4l/vidioc-decoder-cmd.rst     | 10 +++-
> >>>  .../media/uapi/v4l/vidioc-reqbufs.rst         |  6 +++
> >>>  .../media/videodev2.h.rst.exceptions          |  1 +
> >>>  .../media/common/videobuf2/videobuf2-v4l2.c   |  8 +++-
> >>>  drivers/media/v4l2-core/v4l2-mem2mem.c        | 35 ++++++++++++++
> >>>  drivers/staging/media/sunxi/cedrus/cedrus.h   |  1 +
> >>>  .../staging/media/sunxi/cedrus/cedrus_dec.c   | 11 +++++
> >>>  .../staging/media/sunxi/cedrus/cedrus_h264.c  | 11 ++++-
> >>>  .../staging/media/sunxi/cedrus/cedrus_hw.c    |  8 ++--
> >>>  .../staging/media/sunxi/cedrus/cedrus_video.c | 14 ++++++
> >>>  include/media/v4l2-mem2mem.h                  | 46 +++++++++++++++++++
> >>>  include/media/videobuf2-core.h                |  3 ++
> >>>  include/media/videobuf2-v4l2.h                |  5 ++
> >>>  include/uapi/linux/videodev2.h                | 14 ++++--
> >>>  15 files changed, 175 insertions(+), 11 deletions(-)

Powered by blists - more mailing lists