Re: [PATCH v3 09/14] media: uapi: Add V4L2_CID_STATELESS_HEVC_ENTRY_POINT_OFFSET control

[Benjamin Gaignard <benjamin.gaignard@collabora.com>]

Le 28/02/2022 à 10:47, John Cox a écrit :
> On Sat, 26 Feb 2022 19:17:46 +0100, you wrote:
>
>> Dne sobota, 26. februar 2022 ob 18:59:18 CET je Jernej Škrabec napisal(a):
>>> Hi!
>>>
>>> Dne petek, 25. februar 2022 ob 20:30:20 CET je Jernej Škrabec napisal(a):
>>>> Hi!
>>>>
>>>> Dne petek, 25. februar 2022 ob 17:45:55 CET je Benjamin Gaignard
>> napisal(a):
>>>>> The number of 'entry point offset' could be very variable.
>>>>> Rather than use a large static array define a v4l2 dynamic array
>>>>> of integer control.
>>>> I suggest we should be more specific and say U32 (V4L2_CTRL_TYPE_U32).
>>>>
>>>>> The number of entry point offsets is reported by the elems field.
>>> I did few more tests and these are my findings:
>>> 1. dynamic array can't be set over size, specified in .dims array
>>> 2. entry point offsets are per slice, so if we make slices dynamic array,
>> then
>>> entry points would become two dimensional dynamic array
>>> 3. num_entry_point_offsets must be part of slice control, because it can be
>>> zero, but size of array can't be
>> Clarification: num_entry_point_offsets is needed in any case, even without entry
>> point list.
> I thought that at least one decoder wanted the entry points or have we
> decided that we don't? Whilst they aren't required for decode, what they
> give you, if your h/w supports it, is the ability to decode multiple
> lines of CTBs in parallel (with a 2 CTB offset between lines), so they
> aren't useless.
>
> If we do decided we want the entry point offset control, now or later,
> then num_entry_points in the slice_param allows you to determine which
> offsets apply to which slices without the complexity of a variable array
> of variable arrays.
>
> But whether or not we end up supporting entry point offsets, the Pi h/w
> requires me to generate 1 table entry for each CTB line and
> num_entry_points is the parameter I need for that, so I really want it
> in the slice_params.

I'm adding a control (dynamic array) and num_entry_points field in slice
parameters in the next version.
See here:
https://gitlab.collabora.com/benjamin.gaignard/for-upstream/-/commit/6b4efa4dc54de069893f5203d885fea2c1347916

Regards,
Benjamin

> Regards
>
> John Cox
>
>>> 4. fortunately, not setting entry points doesn't impact decoding
>> correctness.
>>> This is in line what John told me about them.
>>>
>>> Hans, can you comment points 1-3? I might misunderstand point 1.
>>>
>>> In short, it seems like we don't really need entry points, even if they are
>>> used in BSP library. In both cases, I got fluster score 119/138 (10-bit
>>> excluded), so we can just drop this patch, although I'm a bit uneasy not
>>> setting entry points...
>>>
>>> Best regards,
>>> Jernej
>>>
>>>>> Signed-off-by: Benjamin Gaignard <benjamin.gaignard@collabora.com>
>>>>> ---
>>>>>   .../userspace-api/media/v4l/ext-ctrls-codec.rst          | 9 +++++++++
>>>>>   include/media/hevc-ctrls.h                               | 1 +
>>>>>   2 files changed, 10 insertions(+)
>>>>>
>>>>> diff --git a/Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst b/
>>>> Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
>>>>> index 44a268a948c0..71f7dc1c1ccd 100644
>>>>> --- a/Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
>>>>> +++ b/Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
>>>>> @@ -3128,6 +3128,15 @@ enum v4l2_mpeg_video_hevc_size_of_length_field -
>>>>>   
>>>>>       \normalsize
>>>>>   
>>>>> +``V4L2_CID_STATELESS_HEVC_ENTRY_POINT_OFFSETS (integer)``
>>>> Here you have OFFSETS (plural) ...
>>>>
>>>>> +    Specifies the i-th entry point offset in bytes and is represented by
>>>>> +    offset_len_minus1 plus 1 bits.
>>>> You probably mean entry_point_offset_minus1? offset_len_minus1 just tells
>> how
>>>> much bits need to be read for each element and it's not important for
>> actual
>>>> decoding.
>>>>
>>>>> +    This control is a dynamically sized array. The number of entry
>> point
>>>>> +    offsets is reported by the ``elems`` field.
>>>>> +    This bitstream parameter is defined according to :ref:`hevc`.
>>>>> +    They are described in section 7.4.7.1 "General slice segment header
>>>>> +    semantics" of the specification.
>>>>> +
>>>>>   ``V4L2_CID_STATELESS_HEVC_SCALING_MATRIX (struct)``
>>>>>       Specifies the HEVC scaling matrix parameters used for the scaling
>>>> process
>>>>>       for transform coefficients.
>>>>> diff --git a/include/media/hevc-ctrls.h b/include/media/hevc-ctrls.h
>>>>> index 3016c1abb1d0..3f8a67924df3 100644
>>>>> --- a/include/media/hevc-ctrls.h
>>>>> +++ b/include/media/hevc-ctrls.h
>>>>> @@ -20,6 +20,7 @@
>>>>>   #define V4L2_CID_STATELESS_HEVC_DECODE_PARAMS	(V4L2_CID_CODEC_BASE +
>>>> 1012)
>>>>>   #define V4L2_CID_STATELESS_HEVC_DECODE_MODE	(V4L2_CID_CODEC_BASE +
>>>> 1015)
>>>>>   #define V4L2_CID_STATELESS_HEVC_START_CODE	(V4L2_CID_CODEC_BASE
>> +
>>> 1016)
>>>>> +#define V4L2_CID_STATELESS_HEVC_ENTRY_POINT_OFFSET (V4L2_CID_CODEC_BASE
>> +
>>>> 1017)
>>>>
>>>> ... and here you have OFFSET (singlular). I suggest plural form to be used
>>> in
>>>> all places, including subject line of this commit.
>>>>
>>>> Additionally, it would be nice if control is initialized, like so:
>>>> https://github.com/jernejsk/linux-1/commit/
>>>> f938e162cd8dd77c9f6f1b248d80144840a37bce
>>>>
>>>> Best regards,
>>>> Jernej
>>>>
>>>>>   
>>>>>   /* enum v4l2_ctrl_type type values */
>>>>>   #define V4L2_CTRL_TYPE_HEVC_SPS 0x0120
>>>>> -- 
>>>>> 2.32.0
>>>>>
>>>>>
>>>>
>>>>
>>>
>>>