Re: [PATCH 10/10] venus: dec: make decoder compliant with stateful codec API

From: Stanimir Varbanov
Date: Fri May 31 2019 - 04:04:58 EST


Hi,

On 5/27/19 11:18 AM, Tomasz Figa wrote:
> On Mon, May 27, 2019 at 4:39 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
>>
>> On 5/27/19 5:51 AM, Tomasz Figa wrote:
>>> On Tue, May 21, 2019 at 9:27 PM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
>>>>
>>>> On 5/21/19 11:09 AM, Tomasz Figa wrote:
>>>>> Hi Stan,
>>>>>
>>>>> On Mon, May 20, 2019 at 11:47 PM Stanimir Varbanov
>>>>> <stanimir.varbanov@xxxxxxxxxx> wrote:
>>>>>>
>>>>>> Hi Tomasz,
>>>>>>
>>>>>> On 4/24/19 3:39 PM, Tomasz Figa wrote:
>>>>>>> On Wed, Apr 24, 2019 at 9:15 PM Stanimir Varbanov
>>>>>>> <stanimir.varbanov@xxxxxxxxxx> wrote:
>>>>>>>>
>>>>>>>> Hi Hans,
>>>>>>>>
>>>>>>>> On 2/15/19 3:44 PM, Hans Verkuil wrote:
>>>>>>>>> Hi Stanimir,
>>>>>>>>>
>>>>>>>>> I never paid much attention to this patch series since others were busy
>>>>>>>>> discussing it and I had a lot of other things on my plate, but then I heard
>>>>>>>>> that this patch made G_FMT blocking.
>>>>>>>>
>>>>>>>> OK, another option could be to block REQBUF(CAPTURE) until event from hw
>>>>>>>> is received that the stream is parsed and the resolution is correctly
>>>>>>>> set by application. Just to note that I'd think to this like a temporal
>>>>>>>> solution until gstreamer implements v4l events.
>>>>>>>>
>>>>>>>> Is that looks good to you?
>>>>>>>
>>>>>>> Hmm, I thought we concluded that gstreamer sets the width and height
>>>>>>> in OUTPUT queue before querying the CAPTURE queue and so making the
>>>>>>> driver calculate the CAPTURE format based on what's set on OUTPUT
>>>>>>> would work fine. Did I miss something?
>>>>>>
>>>>>> Nobody is miss something.
>>>>>>
>>>>>> First some background about how Venus implements stateful codec API.
>>>>>>
>>>>>> The Venus firmware can generate two events "sufficient" and
>>>>>> "insufficient" buffer requirements (this includes decoder output buffer
>>>>>> size and internal/scratch buffer sizes). Presently I always set minimum
>>>>>> possible decoder resolution no matter what the user said, and by that
>>>>>> way I'm sure that "insufficient" event will always be triggered by the
>>>>>> firmware (the other reason to take this path is because this is the
>>>>>> least-common-divider for all supported Venus hw/fw versions thus common
>>>>>> code in the driver). The reconfiguration (during codec Initialization
>>>>>> sequence) is made from STREAMON(CAPTURE) context. Now, to make that
>>>>>> re-configuration happen I need to wait for "insufficient" event from
>>>>>> firmware in order to know the real coded resolution.
>>>>>>
>>>>>> In the case of gstreamer where v4l2_events support is missing I have to
>>>>>> block (wait for firmware event) REQBUF(CAPTURE) (vb2::queue_setup) or
>>>>>> STREAMON(CAPTURE) (vb2::start_streaming).
>>>>>>
>>>>>> I tried to set the coded resolution to the firmware as-is it set by
>>>>>> gstreamer but then I cannot receive the "sufficient" event for VP8 and
>>>>>> VP9 codecs. So I return back to the solution with minimum resolution above.
>>>>>>
>>>>>> I'm open for suggestions.
>>>>>
>>>>> I think you could still keep setting the minimum size and wait for the
>>>>> "insufficient" event. At the same time, you could speculatively
>>>>> advertise the expected "sufficient" size on the CAPTURE queue before
>>>>> the hardware signals those. Even if you mispredict them, you'll get
>>>>> the event, update the CAPTURE resolution and send the source change
>>>>> event to the application, which would then give you the correct
>>>>> buffers. Would that work for you?
>>>>
>>>> As I understand it this still would require event support, which gstreamer
>>>> doesn't have.
>>>
>>> I don't think it matches what I remember from the earlier discussion.
>>> As long as Gstreamer sets the visible resolution (from the container
>>> AFAIR) on OUTPUT, the driver would adjust it to something that is
>>> expected to be the right framebuffer resolution and so Gstreamer would
>>> be able to continue. Of course if the expected value doesn't match, it
>>> wouldn't work, but it's the same as currently for Coda AFAICT.
>>>
>>>>
>>>> I think it is OK to have REQBUFS sleep in this case. However, I would only
>>>
>>> Why REQBUFS? While that could possibly allow us to allocate the right
>>> buffers, Gstreamer wouldn't be able to know the right format, because
>>> it would query it before REQBUFS, wouldn't it?
>>
>> Oops, you are right. It's got to be in G_FMT(CAPTURE), but *only* if
>> nobody subscribed to the SOURCE_CHANGE event.
>>
>>>
>>> For this reason, s5p-mfc makes G_FMT(CAPTURE) blocking and if we
>>> decide to forcefully keep the compatibility, even with in drivers, we
>>> should probably do the same here.
>>>
>>>> enable this behavior if the application didn't subscribe to the SOURCE_CHANGE
>>>> event. That's easy enough to check in the driver. And that means that if the
>>>> application is well written, then the driver will behave in a completely
>>>> standard way that the compliance test can check.
>>>
>>> I guess one could have some helpers for this. They would listen to the
>>> source change events internally and block / wake-up appropriate ioctls
>>> whenever necessary.
>>
>> I really do not want this for new drivers. gstreamer should be fixed.
>> A blocking G_FMT is just plain bad. Only those drivers that do this, can
>> still block if nobody subscribed to EVENT_SOURCE_CHANGE.
>
> Yeah and that's why I just suggested to mimic coda, which doesn't
> block, but apparently gstreamer still works with it.

Unfortunately in Venus case that is not an easy task (as I tried to
explain why above).

To have an unified and common code for all different SoCs and
firmware/hardware versions I decided to set the minimum supported
resolution for the decoder (no matter what the user said) and trigger
the reconfiguration event always. Something more, I need the event also
to retrieve the minimum capture buffers
(V4L2_CID_MIN_BUFFERS_FOR_CAPTURE) and sizes for capture and
internal/scratch buffers as well, thus I really need to wait for that
event.

So, just to confirm - you are fine with blocking G_FMT (not REQBUF) when
the user doesn't subscribe for v4l2 events?

>
>>
>>> Another question: If we intend this to be implemented in new drivers
>>> too, should it be documented in the spec?
>>
>> We most certainly do NOT want to implement this in new drivers.
>>
>
> Makes sense.
>
> When venus was merged initially, did it already have a blocking G_FMT?

No it isn't.

--
regards,
Stan