[FFmpeg-devel] [PATCH 3/3] avformat/dashenc: always attempt to enable prft on ldash mode

[James Almer]

On 2/19/2020 11:24 PM, Jeyapal, Karthick wrote:
> 
> On 2/20/20 7:19 AM, James Almer wrote:
>> On 2/19/2020 9:33 PM, Jeyapal, Karthick wrote:
>>>
>>> On 2/19/20 7:05 PM, James Almer wrote:
>>>> On 2/19/2020 8:50 AM, Jeyapal, Karthick wrote:
>>>>>
>>>>> On 2/19/20 4:21 PM, Thilo Borgmann wrote:
>>>>>> Am 19.02.20 um 06:18 schrieb Jeyapal, Karthick:
>>>>>>>
>>>>>>> On 2/18/20 9:43 PM, James Almer wrote:
>>>>>>>> Signed-off-by: James Almer <jamrial at gmail.com>
>>>>>>>> ---
>>>>>>>>  libavformat/dashenc.c | 5 +++++
>>>>>>>>  1 file changed, 5 insertions(+)
>>>>>>>>
>>>>>>>> diff --git a/libavformat/dashenc.c b/libavformat/dashenc.c
>>>>>>>> index b910cc22d0..045d2f4df6 100644
>>>>>>>> --- a/libavformat/dashenc.c
>>>>>>>> +++ b/libavformat/dashenc.c
>>>>>>>> @@ -1395,6 +1395,11 @@ static int dash_init(AVFormatContext *s)
>>>>>>>>          c->frag_type = FRAG_TYPE_EVERY_FRAME;
>>>>>>>>      }
>>>>>>>>  
>>>>>>>> +    if (c->ldash && !c->write_prft) {
>>>>>>>> +        av_log(s, AV_LOG_INFO, "Enabling Producer Reference Time element for Low Latency mode\n");
>>>>>>>> +        c->write_prft = 1;
>>>>>>>> +    }
>>>>>>>> +
>>>>>>> PRFT elements has a significant bitrate overhead, especially in streaming mode when each frame is a moof fragment.
>>>>>>> In terms of percentage of stream's bitrate this overhead will be a significant % for lower bitrate streams(such as audio streams).
>>>>>>> For any application which does not need PRFT this is an unnecessary wastage of bits. 
>>>>>>> Hence, I would advise against enabling PRFT without user control.
>>>>>>
>>>>>> Latest to-become spec for low latency mode declares it mandatory [1].
>>>>> I see. Now I understand the motive behind this change. 
>>>>>>
>>>>>> I see your point, though. What significance would this actually have, can you provide some numbers / examples?
>>>>> Sorry. I worked on this bitrate overhead optimizations around a year back. Hence, I don’t have the numbers with and without PRFT handy. 
>>>>> But I do have the final overhead numbers (without PRFT) for an audio stream. 
>>>>> CMAF Muxer overhead (for an AAC-LC codec) by Sampling frequency
>>>>> 16000 Hz - 14 Kbps
>>>>> 24000 Hz - 20 Kbps
>>>>> 32000 Hz - 28 Kbps
>>>>> 48000 Hz - 40 Kbps
>>>>>
>>>>> At 48KHz, the overhead due to CMAF was 40 Kbps which was significant by itself. 
>>>>> My random guess is that PRFT would add another 10Kbps - 20Kbps. But I could be wrong here, as I don’t remember exactly.
>>>>
>>>> prft is a 32 byte box per dash segment, and segment duration can be
>>>> configured. A 1 second long segment for a 96kbps 44kHz audio stream with
>>>> a single moof/mdat pair inside is about 12kb. 32 bytes aren't going to
>>>> affect it.
>>> Thanks for your clarification. If the prft is created only once per segment, then it is not a big overhead.
>>> I had encountered a case where pfrt was getting created once per fragment, and hence was worried.
>>
>> Actually, you're right, it's one per fragment. My mistake. But you can
>> control both fragment count per segment and frame count per fragment in
>> the dash muxer now, and for low latency dash one fragment per segment of
>> about 1 second each is recommended for audio streams.
> If that is the case, I would like to point out that not everybody might choose 1 second per fragment. 
> Anyone interested in reduced latency and stability would go for 1 frame per fragment. 
> In our tests(in production environments), 1 frame per fragment provides much smoother and stable behavior at low latency streaming than higher fragment sizes.
> And in such a case PRFT will add 12Kbps overhead for a 48Khz AAC-LC stream. Hence, I suggest we keep this behavior configurable.

One audio frame per fragment is a considerable overhead by itself just
by the excess of moof atoms. But in any case, i guess i can skip this
patch for the time being, while i look for a good way to for example
configure prft in a per adaptation set basis, or just enable it for
video streams.

>>
>>>>
>>>> The real overhead is in the CMAF fragmentation/segmentation. Each moof
>>>> box can be in the hundreds of bytes depending on frame count. The more
>>>> moof/mdat pairs (AKA CMAF Chunks) are used, the bigger the overhead.
>>>> Before my recent changes the dash muxer would always make one per frame
>>>> in streaming mode, which was excessive, especially for audio. But now
>>>> you can customize it in various ways.
>>>>
>>>>>
>>>>>> If that turns out to be actually significant, I don't know if we would prefer an override option to disable it and produce non-conformant manifests or live with the overhead.
>>>>> Yes. Having an option to control this behavior would be useful.
>>>>>>
>>>>>> -Thilo
>>>>>>
>>>>>> [1] https://dashif.org/docs/DASH-IF-IOP-CR-Low-Latency-Live-Community-Review-Dec-2019.pdf