[PATCH v2 0/4] media: HEVC: RPS clean up

Alex Bee knaerzche at gmail.com
Fri Oct 15 07:14:28 PDT 2021


Am 15.10.21 um 16:06 schrieb Benjamin Gaignard:
>
> Le 15/10/2021 à 12:33, Alex Bee a écrit :
>> Hi Benjamin, Jernej
>> Am 12.10.21 um 18:08 schrieb Jernej Škrabec:
>>> CC: Alex Bee
>>>
>>> Alex, please take a look to these patches too.
>> These patches don't remove anything that would be need for rkvdec 
>> hevc - but indeed - we need some more:
>> https://github.com/LibreELEC/LibreELEC.tv/blob/master/projects/Rockchip/patches/linux/default/linux-2001-v4l-wip-rkvdec-hevc.patch#L242-L305 
>>
>>
>> v4l2_ctrl_hevc_sps:
>> __u8    video_parameter_set_id
>> __u8    seq_parameter_set_id
>>
>> v4l2_ctrl_hevc_pps:
>> __u8    pic_parameter_set_id
>> __u16    short_term_ref_pic_set_size
>> __u16    long_term_ref_pic_set_size
>>
>> As far as I can see, they are all part of the spec and should be 
>> therefore good to go in the uapi.
>
> Do you have any plan to upstream these fields in HEVC uAPI ?

I might be upstreaming them at some point, yes.

With this I just wanted to underline Jernej said: HEVC uapi is NOT ready 
to get unstaged yet.

>
> Regards,
> Benjamin
>
>>
>> As you might now, even rkvdec is a frame-based decoder, it doesn't 
>> fully parse slice headers in HW for HEVC and we need to set 
>> references in SW which requires looping over the slices. Downstream 
>> we have a hack to give num_slices in v4l2_ctrl_hevc_sps for doing that.
>> That could fully go away, if V4L2_CID_MPEG_VIDEO_HEVC_SLICE_PARAMS 
>> could get dynamic array control support and would make upstreaming 
>> this a lot easier - as far as I'm concered this would be required for 
>> RPi HEVC decoder as well.
>> As a last resort we could also implement a HW specifc control à la
>> V4L2_CID_HANTRO_HEVC_SLICE_HEADER_SKIP - but I'd like to avoid that, 
>> knowing it would certainly be better from performance pov.
>>
>> Alex.
>>>
>>> Dne torek, 12. oktober 2021 ob 17:57:50 CEST je Benjamin Gaignard 
>>> napisal(a):
>>>>
>>>> Le 12/10/2021 à 17:34, Jernej Škrabec a écrit :
>>>>> Hi Benjamin!
>>>>>
>>>>> Dne torek, 12. oktober 2021 ob 16:35:48 CEST je Benjamin Gaignard
>>> napisal(a):
>>>>>> This series aims to clean up Reference Picture Set usage and flags.
>>>>>>
>>>>>> Long term flag was named with RPS prefix while it is not used for 
>>>>>> RPS
>>>>>> but for mark long term references in DBP. Remane it and remove 
>>>>>> the two
>>>>>> other useless RPS flags.
>>>>>>
>>>>>> Clarify documentation about RPS lists content and make sure that 
>>>>>> Hantro
>>>>>> driver use them correctly (i.e without look up in DBP).
>>>>>>
>>>>>> These patches are the last in my backlog impacting HEVC uAPI.
>>>>>>   From my point of view, once they get merged, you could start 
>>>>>> talking
>>>>>> about how move HEVC uAPI to stable.
>>>>> With your changes, HEVC uAPI controls still won't be complete. Cedrus
>>> needs
>>>>> entry point control, which in turn needs dynamic array support. 
>>>>> I'm a bit
>>> lazy
>>>>> implementing that control, but I guess I can take a look in a 
>>>>> month or so.
>>>>> rkvdec also needs more fields for HEVC. With patches collected here:
>>>>> https://github.com/LibreELEC/LibreELEC.tv/blob/master/projects/Rockchip/ 
>>>>>
>>>>> patches/linux/default/linux-2001-v4l-wip-rkvdec-hevc.patch
>>>>> fluster HEVC test score is reportedly 121/135 (8-bit tests only).
>>>>
>>>> Hi Jernej,
>>>>
>>>> Thanks for your feedback, getting a list of missing items in HEVC uAPI
>>>> will definitively help to fill the hope.
>>>> The patch you mention for rkvdec are already merged in mainline 
>>>> kernel (at
>>>> least for uAPI part).
>>>
>>> Are they? What about:
>>> video_parameter_set_id
>>> seq_parameter_set_id
>>> pic_parameter_set_id
>>> short_term_ref_pic_set_size
>>> long_term_ref_pic_set_size
>>>
>>> At least I don't see them in linux-next. Maybe that information can be
>>> obtained in some other way?
>>>
>>>> Cedrus needs are about num_entry_point_offsets, offset_len_minus1 and
>>> entry_point_offset_minus1[ i ]
>>>> in HEVC specifications ?
>>>
>>> Yes, Cedrus needs to know whole list of entry points. I don't think 
>>> we need to
>>> worry about offset_len_minus1, list could be pre-processed - just 
>>> number of
>>> entry points and their values.
>>>
>>> Best regards,
>>> Jernej
>>>
>>>>
>>>> Regards,
>>>> Benjamin
>>>>
>>>>>
>>>>> I would certainly wait with moving HEVC uAPI to stable.
>>>>>
>>>>> Best regards,
>>>>> Jernej
>>>>>
>>>>>> version 2:
>>>>>> - change DPB field name from rps to flags
>>>>>>
>>>>>> Please note that the only purpose of commits 3 and 4 is to allow 
>>>>>> to test
>>>>>> G2 hardware block for IMX8MQ until a proper solution isuing power 
>>>>>> domain
>>>>>> can be found. Do not merge them.
>>>>>>
>>>>>> GStreamer HEVC plugin merge request can be found here:
>>>>>> https://gitlab.freedesktop.org/gstreamer/gstreamer/-/merge_requests/1079 
>>>>>>
>>>>>>
>>>>>> With those piece of code fluster score is 77/147.
>>>>>>
>>>>>> Benjamin
>>>>>>
>>>>>> Benjamin Gaignard (4):
>>>>>>     media: hevc: Remove RPS named flags
>>>>>>     media: hevc: Embedded indexes in RPS
>>>>>>     media: hantro: Use syscon instead of 'ctrl' register
>>>>>>     arm64: dts: imx8mq: Add node to G2 hardware
>>>>>>
>>>>>>    .../media/v4l/ext-ctrls-codec.rst             | 14 +++---
>>>>>>    arch/arm64/boot/dts/freescale/imx8mq.dtsi     | 43 
>>>>>> +++++++++++++----
>>>>>>    drivers/staging/media/hantro/hantro.h         |  5 +-
>>>>>>    .../staging/media/hantro/hantro_g2_hevc_dec.c | 27 +++--------
>>>>>>    drivers/staging/media/hantro/imx8m_vpu_hw.c   | 48 
>>>>>> ++++++++++++-------
>>>>>>    .../staging/media/sunxi/cedrus/cedrus_h265.c  |  2 +-
>>>>>>    include/media/hevc-ctrls.h                    |  6 +--
>>>>>>    7 files changed, 84 insertions(+), 61 deletions(-)
>>>>>>
>>>>>> -- 
>>>>>> 2.30.2
>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>>
>>



More information about the linux-arm-kernel mailing list