Request key frame in vpuenc (Freescale)

[Carlos Rafael Giani]

On 2013-10-02 00:56, Javier Jardón wrote:
> On 30 September 2013 17:11, Chuck Crisler <ccrisler at mutualink.net> wrote:
>> I can't remember exactly, but there is an option in the encoder structure to
>> instruct the VPU encoder to generate an I-Frame. Just make sure that you
>> clear the flag after the I-Frame is generated. Of course, the structure item
>> name may vary depending on which VPU you are using. Your encoder element
>> probably already controls when to generate an I-Frame. When I worked with a
>> Freescale VPU we had to determine the I-Frame rate ourselves, the encoder
>> didn't count the frames.
> Hi Chuck,
>
> thanks for the hint. As you seggested, nForceIPicture=1 is the one
> needed to generate the keyframe.
> The problem I have now is that I think I have to send the SPS/PPS for
> any new segment I created, but currently I get this from the parser
> after the encoder:
>
> 0:00:20.366602668 23704    0x20120 WARN               h264parse
> gsth264parse.c:926:gst_h264_parse_handle_frame:<h264parse0> no SPS/PPS
> yet, nal Type: 1 Slice, Size: 28105 will be dropped
>
> I checked and seems that the headers are generated only one time in
> the beggining of the stream, do you know if there is anything I can do
> to tell the vpu to generate new headers again?
>
> Thanks
>
>

To give additional details:

The VPU encoder can produce h.264 baseline content, in the Annex.B (= 
"byte-stream") format. Access units do not seem to be generated, at 
least I do not see typical au content in the NAL units.
The VPU generates the SPS/PPS headers during startup, and whenever it 
deems it necessary later on. There is no apparent mechanism to trigger a 
new SPS/PPS header generation.

The problem seems to be when a new segment occurs. Since the headers 
were created only once, only the previous segment ever saw headers - the 
new one didn't. h264parse apparently considers this an error, and 
expects SPS/PPS data for every segment. But since the VPU output is 
using the byte-stream format, the headers are part of the stream, and 
cannot simply be set as codec-data caps (which would otherwise be an 
easy and fully working solution).

I see two options:
1) Extract the SPS/PPS headers when they are generated, store them, and 
install a set_event handler that looks for new_segment events. If such 
an event occurs, set a flag. The next time a buffer is about to be sent 
downstream, add the headers to this buffer's data (which contains a NAL 
unit).
2) Remove the byte-stream wrapping, and sent the data downstream. Set 
the codec-data caps field to whatever is the SPS/PPS content.

Perhaps it would also be possible to set the codec-data caps even though 
Annex.B is sent downstream? Would this break something? If not, this 
could be an option 3, perhaps the best of all.

Opinions?