Re: [PATCH v3 1/9] media: hevc: Modify structures to follow H265 ITU spec

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Le jeudi 25 février 2021 à 18:34 +0000, John Cox a écrit :
> On Thu, 25 Feb 2021 19:05:55 +0100, you wrote:
> 
> > Dne ?etrtek, 25. februar 2021 ob 18:34:48 CET je Ezequiel Garcia napisal(a):
> > > Hey Jernej,
> > > 
> > > On Thu, 2021-02-25 at 18:01 +0100, Jernej Škrabec wrote:
> > > > Hi Ezequiel,
> > > > 
> > > > Dne ?etrtek, 25. februar 2021 ob 14:09:52 CET je Ezequiel Garcia 
> > napisal(a):
> > > > > Hi Benjamin,
> > > > > 
> > > > > Thanks for the good work.
> > > > > 
> > > > > On Mon, 2021-02-22 at 13:23 +0100, Benjamin Gaignard wrote:
> > > > > > The H.265 ITU specification (section 7.4) define the general
> > > > > > slice segment header semantics.
> > > > > > Modified/added fields are:
> > > > > > - video_parameter_set_id: (7.4.3.1) identifies the VPS for
> > > > > > reference by other syntax elements.
> > > > > > - seq_parameter_set_id: (7.4.3.2.1) specifies the value of
> > > > > > the vps_video_parameter_set_id of the active VPS.
> > > > > > - chroma_format_idc: (7.4.3.2.1) specifies the chroma sampling
> > > > > >  relative to the luma sampling
> > > > > > - pic_parameter_set_id: (7.4.3.3.1) identifies the PPS for
> > > > > > reference by other syntax elements
> > > > > > - num_ref_idx_l0_default_active_minus1: (7.4.3.3.1) specifies
> > > > > > the inferred value of num_ref_idx_l0_active_minus1
> > > > > > - num_ref_idx_l1_default_active_minus1: (7.4.3.3.1) specifies
> > > > > > the inferred value of num_ref_idx_l1_active_minus1
> > > > > > - slice_segment_addr: (7.4.7.1) specifies the address of
> > > > > > the first coding tree block in the slice segment
> > > > > > - num_entry_point_offsets: (7.4.7.1) specifies the number of
> > > > > > entry_point_offset_minus1[ i ] syntax elements in the slice header
> > > > > > 
> > > > > > Add HEVC decode params contains the information used in section
> > > > > > "8.3 Slice decoding process" of the specification to let the
> > > > > > hardware
> > > > > > perform decoding of a slices.
> > > > > > 
> > > > > > Adapt Cedrus driver according to these changes.
> > > > > > 
> > > > > > Signed-off-by: Benjamin Gaignard <benjamin.gaignard@xxxxxxxxxxxxx>
> > > > > > ---
> > > > > > version 3:
> > > > > > - Add documentation about the new structuers and fields.
> > > > > > 
> > > > > > version 2:
> > > > > > - remove all change related to scaling
> > > > > > - squash commits to a coherent split
> > > > > > - be more verbose about the added fields
> > > > > > 
> > > > > >  .../media/v4l/ext-ctrls-codec.rst             | 126 ++++++++++++++
> > +---
> > > > > >  .../media/v4l/vidioc-queryctrl.rst            |   6 +
> > > > > >  drivers/media/v4l2-core/v4l2-ctrls.c          |  26 +++-
> > > > > >  drivers/staging/media/sunxi/cedrus/cedrus.c   |   6 +
> > > > > >  drivers/staging/media/sunxi/cedrus/cedrus.h   |   1 +
> > > > > >  .../staging/media/sunxi/cedrus/cedrus_dec.c   |   2 +
> > > > > >  .../staging/media/sunxi/cedrus/cedrus_h265.c  |   6 +-
> > > > > >  include/media/hevc-ctrls.h                    |  45 +++++--
> > > > > >  8 files changed, 186 insertions(+), 32 deletions(-)
> > > > > > 
> > > > > > diff --git a/Documentation/userspace-api/media/v4l/ext-ctrls-
> > > > > > codec.rst 
> > b/
> > > > Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
> > > > > > index 00944e97d638..5e6d77e858c0 100644
> > > > > > --- a/Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
> > > > > > +++ b/Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
> > > > > > @@ -3109,6 +3109,15 @@ enum
> > > > > > v4l2_mpeg_video_hevc_size_of_length_field -
> > > > > >      :stub-columns: 0
> > > > > >      :widths:       1 1 2
> > > > > >  
> > > > > > +    * - __u8
> > > > > > +      - ``video_parameter_set_id``
> > > > > > +      - Identifies the VPS for reference by other syntax elements
> > > > > > +    * - __u8
> > > > > > +      - ``seq_parameter_set_id?``
> > > > > > +      - Specifies the value of the vps_video_parameter_set_id of
> > > > > > the 
> > > > active VPS
> > > > > > +    * - __u8
> > > > > > +      - ``chroma_format_idc``
> > > > > > +      - Specifies the chroma sampling relative to the luma sampling
> > > > > 
> > > > > None of these fields seem needed for the Hantro G2 driver,
> > > > > so I suggest you drop them for now.
> > > > > 
> > > > > >      * - __u16
> > > > > >        - ``pic_width_in_luma_samples``
> > > > > >        -
> > > > > > @@ -3172,6 +3181,9 @@ enum v4l2_mpeg_video_hevc_size_of_length_field
> > > > > > -
> > > > > >      * - __u8
> > > > > >        - ``chroma_format_idc``
> > > > > >        -
> > > > > > +    * - __u8
> > > > > > +      - ``num_slices``
> > > > > > +
> > > > > 
> > > > > Not used, but also doesn't seem part of the SPS syntax. If we have to
> > > > > pass the number of slices, we'll need another mechanism.
> > > > > 
> > > > > >       -
> > > > > >      * - __u64
> > > > > >        - ``flags``
> > > > > >        - See :ref:`Sequence Parameter Set Flags <hevc_sps_flags>`
> > > > > > @@ -3231,9 +3243,18 @@ enum
> > > > > > v4l2_mpeg_video_hevc_size_of_length_field -
> > > > > >      :stub-columns: 0
> > > > > >      :widths:       1 1 2
> > > > > >  
> > > > > > +    * - __u8
> > > > > > +      - ``pic_parameter_set_id``
> > > > > > +      - Identifies the PPS for reference by other syntax elements
> > > > > 
> > > > > Not used.
> > > > > 
> > > > > >      * - __u8
> > > > > >        - ``num_extra_slice_header_bits``
> > > > > >        -
> > > > > > +    * - __u8
> > > > > > +      - ``num_ref_idx_l0_default_active_minus1``
> > > > > > +      - Specifies the inferred value of
> > > > > > num_ref_idx_l0_active_minus1
> > > > > > +    * - __u8
> > > > > > +      - ``num_ref_idx_l1_default_active_minus1``
> > > > > > +      - Specifies the inferred value of
> > > > > > num_ref_idx_l1_active_minus1
> > > > > >      * - __s8
> > > > > >        - ``init_qp_minus26``
> > > > > >        -
> > > > > > @@ -3342,6 +3363,12 @@ enum
> > > > > > v4l2_mpeg_video_hevc_size_of_length_field -
> > > > > >      * -
> > > > > > ``V4L2_HEVC_PPS_FLAG_SLICE_SEGMENT_HEADER_EXTENSION_PRESENT``
> > > > > >        - 0x00040000
> > > > > >        -
> > > > > > +    * - ``V4L2_HEVC_PPS_FLAG_DEBLOCKING_FILTER_CONTROL_PRESENT``
> > > > > > +      - 0x00080000
> > > > > > +      -
> > > > > > +    * - ``V4L2_HEVC_PPS_FLAG_UNIFORM_SPACING``
> > > > > > +      - 0x00100000
> > > > > > +      -
> > > > > >  
> > > > > 
> > > > > I suggest to do all the PPS control changes in a separate patch,
> > > > > feels easier to review and cleaner as you can explain the
> > > > > changes with more detail in the commit description.
> > > > > 
> > > > > Looking at the PPS syntax for tiles, I'm wondering if these
> > > > > deserve their own control, which would be used if tiles are enabled,
> > > > > i.e. V4L2_HEVC_PPS_FLAG_TILES_ENABLED is set.
> > > > > 
> > > > >         __u8    
> > num_tile_columns_minus1;                                         
> > > > >         __u8    
> > num_tile_rows_minus1;                                            
> > > > >         __u8    
> > column_width_minus1[20];                                         
> > > > >         __u8    row_height_minus1[22];    
> > > > > 
> > > > > Not something we necessarily have to tackle now.
> > > > > 
> > > > > >  ``V4L2_CID_MPEG_VIDEO_HEVC_SLICE_PARAMS (struct)``
> > > > > >      Specifies various slice-specific parameters, especially from
> > > > > > the 
> > NAL 
> > > > unit
> > > > > > @@ -3366,6 +3393,12 @@ enum
> > > > > > v4l2_mpeg_video_hevc_size_of_length_field -
> > > > > >      * - __u32
> > > > > >        - ``data_bit_offset``
> > > > > >        - Offset (in bits) to the video data in the current slice
> > > > > > data.
> > > > > > +    * - __u32
> > > > > > +      - ``slice_segment_addr``
> > > > > > +      - Specifies the address of the first coding tree block in the
> > slice 
> > > > segment
> > > > > 
> > > > > Not used.
> > > > > 
> > > > > > +    * - __u32
> > > > > > +      - ``num_entry_point_offsets``
> > > > > > +      - Specifies the number of entry_point_offset_minus1[ i ]
> > > > > > syntax 
> > > > elements in the slice header
> > > > > 
> > > > > Not used.
> > > > 
> > > > While above two fields may not be used in Hantro, they are for sure
> > > > useful 
> > for 
> > > > Cedrus and RPi4. I would like to keep them, otherwise with such approach
> > HEVC 
> > > > will stay in staging for a long time. I'm still baffled why scaling
> > > > matrix 
> > > > control was dropped. It would fit well in Cedrus and RPi4 driver and
> > > > after 
> > a 
> > > > quick look, it seems that it was used in driver in later patch.
> > > > 
> > > 
> > > I'd like to make sure each modification we are making to the uAPI
> > > goes in the right direction, that is in the direction of moving
> > > the API out of staging.
> > > 
> > > Since reviewing each field is quite hard, and opens some discussions,
> > > I wanted to keep this patchset specific to what's needed for Hantro G2.
> > > 
> > > The Scaling matrix control is certainly a good one, as well as the ones
> > > needed for Cedrus and RPi4. However, I feel it's better to discuss
> > > them in their own "uAPI review" series so we can review all the changes
> > > with an API hat.
> > > 
> > > This way we decouple the Hantro G2 discussion and work from the API work.
> > > 
> > > Also please feel free to submit RFC patches fo Cedrus and RPi4
> > > (API and driver changes). We can certainly start the discussion around
> > > that,
> > > with driver changes in context.
> > 
> > I don't know much about RPi4 driver, only few implementation details, so 
> > you'll have to ping developer who wrote it. Regarding HEVC on Cedrus - it
> > has 
> > one pain point - it needs entry point table which in turn needs support for 
> > variable arrays in order to be feasable AFAIK. I don't plan to develop that.
> > Patches for scaling matrix and segment address were sent a bit more than a 
> > year ago but were turned down because they change control structures (among 
> > other things). Sorry to say, but I work on other things now, so Cedrus will 
> > have to wait. Alternatively, someone can take my patches from LibreELEC, 
> > update and submit them. They are in use for a long time.
> 
> It seems to me that H.265 should be the source for what fields are
> needed in the uapi - not whatever happens to be supported by current
> h/w. The standard defines the list of fields that can occur in the
> parameter sets and headers, and everything that is needed to decode 
> any slice_seqgment_data should be in the upai.  Eventually all those
> fields are going to be needed (they aren't in the standard just to look
> pretty) and given the reluctance I've observed to change the uapi once
> released they should be there from the start.
> 
> Now some hardware requires more fields that aren't in the standard -
> those can (and should) be added in h/w specific extensions.

While this is not wrong, there is obvious bitstream parameters that are not of
the HW accelerator domain. We did the work of deciding what to omit in the H264
uAPI, and we'll do that for HEVC for sure. But now is not the time yet.

Now the point that you seem to be missing is that Benjamin is trying to stage a
driver, not a uAPI. With this third driver, we can have more folks contributing
and testing toward our goal of unstaging the uAPI.

After this, we'll be able to iteratively add features (like scaling list) to
this driver along with the required uAPI and discuss each of these needed API
separatly. Note that previous iteration, the driver implementation for the
scaling list wasn't correct, and for this reason, the uAPI wasn't being
validated.

When this is getting stable enough, there will be a final stage were someone
will have to pick the uAPI and fill in the blank with remaining field that has a
strong probability to be used by later HW WITHIN the supported HEVC profiles et
have defined.

I emphasis on that because there is a lot of HEVC (and H264) extensions we
haven't covered or mentionned, and these will be added with dedicated controls
when we have a use case for it of course. Remember that adding controls is not a
problem, we need these basic control of the basic decoding, and for sure, over
the decade to come we will have to introduce new controls for other stuff we
haven't supported yet.

All this may look like taking forever, but as Jernej states, not everyone can
afford contributing upstream. So unless someone step in, this pace is about what
we can offer, and we strongly believe that it will converge. I do personnally
expect HEVC to take more time as the codec rationales are mostly kept secret,
and our answers burried into a mountain of reference code. It's also, if you
defocus from the film industry, just one CODEC. Notably, for the web, VP9 is
more important, and this will likely puts some shadow on this work.

Meawhile, you are strongly encouraged if you can afford it to open the
discussion around the RPi4 multi-stage HEVC decoder. While doing this, you could
send a patched for staging the RPi4 driver. The benefit would be that we get to
update your drivers when we update the uAPI, saving you times, and also giving
everyone more data on what is strictly needed.

> 
> Regards
> 
> JC
> 
> > Best regards,
> > Jernej
> > 
> > > 
> > > Hope I'm making sense here :)
> > > 
> > > Thanks,
> > > Ezequiel
> > > 
> > > > Best regards,
> > > > Jernej
> > > > 
> > > > > 
> > > > > >      * - __u8
> > > > > >        - ``nal_unit_type``
> > > > > >        -
> > > > > > @@ -3422,28 +3455,20 @@ enum
> > > > > > v4l2_mpeg_video_hevc_size_of_length_field 
> > -
> > > > > >      * - __u8
> > > > > >        - ``pic_struct``
> > > > > >        -
> > > > > > -    * - __u8
> > > > > > -      - ``num_active_dpb_entries``
> > > > > > -      - The number of entries in ``dpb``.
> > > > > 
> > > > > Need to explain in the commit description why this field is moved.
> > > > > 
> > > > > >      * - __u8
> > > > > >        - ``ref_idx_l0[V4L2_HEVC_DPB_ENTRIES_NUM_MAX]``
> > > > > >        - The list of L0 reference elements as indices in the DPB.
> > > > > >      * - __u8
> > > > > >        - ``ref_idx_l1[V4L2_HEVC_DPB_ENTRIES_NUM_MAX]``
> > > > > >        - The list of L1 reference elements as indices in the DPB.
> > > > > > +    * - __u16
> > > > > > +      - ``short_term_ref_pic_set_size``
> > > > > > +
> > > > > 
> > > > > Not used.
> > > > > 
> > > > > >       -
> > > > > > +    * - __u16
> > > > > > +      - ``long_term_ref_pic_set_size``
> > > > > > +      -
> > > > > 
> > > > > Not used.
> > > > > 
> > > > > >      * - __u8
> > > > > > -      - ``num_rps_poc_st_curr_before``
> > > > > > -      - The number of reference pictures in the short-term set that
> > come 
> > > > before
> > > > > > -        the current frame.
> > > > > 
> > > > > If this matches NumPocStCurrBefore from section 8.3.2 "Decoding
> > > > > process 
> > for 
> > > > reference picture set"
> > > > > then I would document that. And perhaps rename it to 
> > num_poc_st_curr_before.
> > > > > 
> > > > > > -    * - __u8
> > > > > > -      - ``num_rps_poc_st_curr_after``
> > > > > > -      - The number of reference pictures in the short-term set that
> > come 
> > > > after
> > > > > > -        the current frame.
> > > > > 
> > > > > Ditto.
> > > > > 
> > > > > > -    * - __u8
> > > > > > -      - ``num_rps_poc_lt_curr``
> > > > > > -      - The number of reference pictures in the long-term set.
> > > > > 
> > > > > Ditto.
> > > > > 
> > > > > Also, I'd like the changes that move fields from 
> > > > V4L2_CID_MPEG_VIDEO_HEVC_SLICE_PARAMS
> > > > > to the new V4L2_CID_MPEG_VIDEO_HEVC_DECODE_PARAMS control, to be in 
> > their
> > > > > patch.
> > > > > 
> > > > > That will allow us to put in the commit description a proper
> > > > > explanation of why are fields being moved. Nothing fancy, simply
> > > > > explaining that these variables come from section 8.3.2
> > > > > "Decoding process for reference picture set", which describes
> > > > > a process invoked once per picture, so they are not per-slice.
> > > > > 
> > > > > > -    * - __u8
> > > > > > -      - ``padding[7]``
> > > > > > +      - ``padding``
> > > > > >        - Applications and drivers must set this to zero.
> > > > > >      * - struct :c:type:`v4l2_hevc_dpb_entry`
> > > > > >        - ``dpb[V4L2_HEVC_DPB_ENTRIES_NUM_MAX]``
> > > > > > @@ -3646,3 +3671,74 @@ enum
> > > > > > v4l2_mpeg_video_hevc_size_of_length_field -
> > > > > >      so this has to come from client.
> > > > > >      This is applicable to H264 and valid Range is from 0 to 63.
> > > > > >      Source Rec. ITU-T H.264 (06/2019); G.7.4.1.1, G.8.8.1.
> > > > > > +
> > > > > > +``V4L2_CID_MPEG_VIDEO_HEVC_DECODE_PARAMS (struct)``
> > > > > > +    Specifies various decode parameters, especially the references 
> > picture 
> > > > order
> > > > > > +    count (POC) for all the lists (short, long, before, current, 
> > after) 
> > > > and the
> > > > > > +    number of entries for each of them.
> > > > > > +    These parameters are defined according to :ref:`hevc`.
> > > > > > +    They are described in section 8.3 "Slice decoding process" of
> > > > > > the
> > > > > > +    specification.
> > > > > > +
> > > > > > +.. c:type:: v4l2_ctrl_hevc_decode_params
> > > > > > +
> > > > > > +.. cssclass:: longtable
> > > > > > +
> > > > > > +.. flat-table:: struct v4l2_ctrl_hevc_decode_params
> > > > > > +    :header-rows:  0
> > > > > > +    :stub-columns: 0
> > > > > > +    :widths:       1 1 2
> > > > > > +
> > > > > > +    * - __s32
> > > > > > +      - ``pic_order_cnt_val``
> > > > > > +      -
> > > > > 
> > > > > Can be documented as:
> > > > > 
> > > > > """
> > > > > PicOrderCntVal as described in section 8.3.1 "Decoding process
> > > > > for picture order count" of the specification.
> > > > > """
> > > > > 
> > > > > Note that snake case is used to match the kernel style,
> > > > > but other than that we try to keep the HEVC spec variable
> > > > > names.
> > > > > 
> > > > > > +    * - __u8
> > > > > > +      - ``num_active_dpb_entries``
> > > > > > +      - The number of entries in ``dpb``.
> > > > > > +    * - struct :c:type:`v4l2_hevc_dpb_entry`
> > > > > > +      - ``dpb[V4L2_HEVC_DPB_ENTRIES_NUM_MAX]``
> > > > > > +      - The decoded picture buffer, for meta-data about reference 
> > frames.
> > > > > 
> > > > > The DPB is here, but it seems it's also in the slice control?
> > > > > 
> > > > > > +    * - __u8
> > > > > > +      - ``num_rps_poc_st_curr_before``
> > > > > > +      - The number of reference pictures in the short-term set that
> > come 
> > > > before
> > > > > > +        the current frame.
> > > > > > +    * - __u8
> > > > > > +      - ``num_rps_poc_st_curr_after``
> > > > > > +      - The number of reference pictures in the short-term set that
> > come 
> > > > after
> > > > > > +        the current frame.
> > > > > > +    * - __u8
> > > > > > +      - ``num_rps_poc_lt_curr``
> > > > > > +      - The number of reference pictures in the long-term set.
> > > > > > +    * - __u8
> > > > > > +      - ``rps_st_curr_before[V4L2_HEVC_DPB_ENTRIES_NUM_MAX]``
> > > > > > +      -
> > > > > > +    * - __u8
> > > > > > +      - ``rps_st_curr_after[V4L2_HEVC_DPB_ENTRIES_NUM_MAX]``
> > > > > > +      -
> > > > > > +    * - __u8
> > > > > > +      - ``rps_lt_curr[V4L2_HEVC_DPB_ENTRIES_NUM_MAX]``
> > > > > > +      -
> > > > > 
> > > > > Could you document these as well?
> > > > > 
> > > > > Thanks a lot,
> > > > > Ezequiel
> > > > > 
> > > > > 
> > > > 
> > > > 
> > > 
> > > 
> > > 
> > 
> 






[Index of Archives]     [Linux Input]     [Video for Linux]     [Gstreamer Embedded]     [Mplayer Users]     [Linux USB Devel]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [Yosemite Backpacking]

  Powered by Linux