Posts: 8
Joined: Fri Aug 26, 2016 9:08 pm

H264 Keyframe Question

Wed Aug 14, 2019 1:06 pm

Hi guys I dabble with OpenHD which is an offshoot of the wifibroadcast project.

We have been investigating the new raspivid h264 slices parameter. In the course of testing one guy discovered that he gets improved results with our h264 byte stream when he sets keyframes to “-1” . That’s right, a negative number. What is happening here- how is it evaluating the negative?

While I have your attention. Slices seems to work fine on our pi display program and ok on platforms running gstreamer (maybe with slightly more artifacts than non slice stream). Some android h264 decoders do not work well at all with slices. Any tips on how best to encode and decode when using slices?


Posts: 3
Joined: Sat May 11, 2019 1:17 am

Re: H264 Keyframe Question

Wed Aug 14, 2019 2:19 pm

Hi, I was also doing some tests with us on our openHD (Wifibroadcast based system) .

It does appear as if latency is reduced by 15-20ms but it's really hard to determine the exact reduction without proper tools.

Here is a short video I put together, the first part has Keyframes set to "-1" and the 2nd half has keyframes set to "5" I think.

Appreciate any info on keyframes -1.....

I started doing some tests with those settings some time ago because I can't remember where but I read that someone was using those settings.

Raspberry Pi Engineer & Forum Moderator
Raspberry Pi Engineer & Forum Moderator
Posts: 7149
Joined: Wed Dec 04, 2013 11:27 am
Location: ZZ9 Plural Z Alpha, aka just outside Cambridge.

Re: H264 Keyframe Question

Mon Aug 19, 2019 12:30 pm

Setting the intraperiod is how many there are between every I frame. You can only start decoding a stream from an I-frame (technically an IDR frame, but all of our I-frames are).
Using "raspivid -g -1" is a little bit of a mess over signed vs unsigned values in the code. The underlying value used by the firmware is unsigned. It is read off the command line with "sscanf("%u") as unsigned, but into a signed value. Adding -v prints it out as signed.
Even if that were fixed so that they are all unsigned, there is a niggle in sscanf("%u", "-1"). The C99 spec says
"Matches an optionally signed decimal integer, whose format is the same as expected for the subject sequence of the strtoul function with the value 10 for the base argument. The corresponding argument shall be a pointer to unsigned integer."
So that actually returns 4294967295, and that is the number of frames that I would expect the encoder to produce between I frames (ignoring potential truncation elsewhere into 16bit values). At 30fps that would be around 4.5 years, so effectively never.
If you really want never, then specify 0 as that is defined as producing one I-frame, and then all P-frames. In doing so be aware that you will never be able to seek or error recover the stream. Lost data will almost certainly result in an unplayable stream.

If you're gaining significantly on latency then I'd expect it mainly to be down to the reduced size of the data being sent across the link. The actual encode latency is likely to be a smidge higher compared to an I-frame due to the need for motion estimation over the frame, although you'll then gain some of that back as there is less data to be CABAC encoded.

Not much I can say over compatibility with slice decoding on other platforms. Where it may be getting upset is on timestamps, as all slices in a frame should have the same timestamp. I know that when I was last working on Android it didn't like having fragmented encoded frames, therefore it may well be the same still.
Fundamentally using slices is just chopping the frame up into smaller independent encodes, and each encode can be thrown down the pipe as soon as it is complete rather than once the full frame is encoded. This compromises compression efficiency slightly as it can't reference across the slice boundary.
On a 1080p encode it takes around 40ms to encode (pipelined, hence it can be longer than a frame time), therefore if slice support worked above 720p you would expect to see that time roughly halve.
Software Engineer at Raspberry Pi Trading. Views expressed are still personal views.
I'm not interested in doing contracts for bespoke functionality - please don't ask.

Return to “OpenMAX”