Michael Niedermayer wrote:
On Thu, Feb 07, 2008 at 01:58:32PM -0000, Måns Rullgård wrote:
Michael Niedermayer wrote:
[...]
That way the buffer_fullness stored in the syncpoint will always match exactly the amount the decoder has in its buffer when reading the syncpoint. If it has more in its buffer it just would change its clock to one running at 100.1% and when it has less in its buffer it would choose a 99.9% clock as reference. (Or any approximetaly equivalent process)
That the buffer fullness is off by N bits doesn't tell you how much too fast or too slow your clock is, only the sign of the error.
yes
Knowing also the magnitude of the error allows much more rapid convergence.
I am not so sure about this. I mean i dont dispute that more information should improve it, but i think its good enough with the too much/too little.
A simple example, lets assume we have a decoder with a clock that drifts by up to D between syncpoints. That is, in the most ideal case we would have to accept that we are D off when we reach a syncpoint, assuming we synced to the previous perfectly.
Now lets assume that we are within -2D .. +2D at syncpoint x, and we apply a +D correction if we are <0 and -D if we are >0. This correction could be applied slowly until the next syncpoint. What matters is that after the correction we are within -D .. +D and with the drift thats again -2D .. +2D at syncpoint x+1. Thus above is a proof by induction that just knowing the sign and the worst case clock drift is sufficient to be within a factor of 2 of the best achiveable clock sync. (comparing worst cases, not average)
This is not how clock sync is usually done. A typical implementation involves a PLL-type construct to make the local clock accurately track the sender clock. Once locked, there is very little drift. To correctly compensate for what little drift inevitably remains, the size of the error must be known. The time difference can of course be computed from the difference in buffer fullness and the received bitrate, it merely takes a little more work on the receiver side.
Providing the timestamp in the stream makes this trivial and independent of the buffering mechanism actually used. Only specifying expected buffer fullness (according to a reference model) requires that the receiver at the very least simulate the reference model,
I think most receivers will use something quite similar to the reference model thus making this unneeded. Though yes a receiver using a different buffer model might need to simulate the reference one.
I think it very unlikely that any real implementation will use whatever precise buffer model we choose. Just about any implementation is likely to immediately extract the elementary streams of interest, and discard everything else, such as container headers and unwanted elementary streams.
But i have difficulty imageing a sufficiently different buffer model. I mean a receiver with split buffers could just be taking the sum of their buffer_fullness. A reciver which removes packets later or not instantaneously would just traverse the last few packets to find out how much the refernce buffer would contain.
I'm not saying it would very difficult to simulate the reference buffer, but something is always more than nothing.
Anyway if you say a timestamp would be better, i suspect rich would be harder to convince than me.
Of that I am certain. -- Måns Rullgård mans@mansr.com