[mythtv] Re: mythtv-dev digest, Vol 1 #138 - 17 msgs

[Dr. J. S. Pezaris]

   Date: Mon, 04 Nov 2002 06:01:14 -0500
   From: John Coiner <jcoiner@stanfordalumni.org>
   To: mythtv-dev@snowman.net
   Subject: [mythtv] Re: Audio problem when changing channels
   
   Our A/V sync works by varying the rate at which video frames are 
   displayed, to keep sync with the audio. We can't vary how fast the audio 
   is played; once we request a sample rate from the sound card, we're stuck 
   with that (or, as close as the sound card gets to it, some cards are off 
   by a few percent.)
   
   On each frame, we calculate the error, how far off we are from perfect 
   sync. The first (stupid) implementation I did corrected the entire error 
   on each frame, but this produces jittery (i.e. jerky) video output. The 
   reason is that our error calculation is based on querying the sound 
   driver to find out where it is -- but the value returned by the sound 
   driver is inexact, and the calculated error has a lot of noise as a result.
   
   The solution is to correct only a small fraction of the measured error 
   on each frame (currently we correct 1/50th of the error.) This is 
   sufficient to maintain sync, and it also smooths out the noise, so that 
   the video output has the lowest possible jitter.
   
   Everything I've described so far exists to maintain A/V sync at steady 
   state, when mythtv has been running, unpaused, on the same channel, for 
   at least a few seconds. But we don't always have steady state-- a 
   channel change is a disruptive event, which empties all the buffers, and 
   usually puts the A/V out of sync. With '1/50th' feedback, it took too 
   long to resync after a channel change. To deal with this, if we detect a 
   very large A/V sync error, we use stiffer feedback to get back into sync 
   quickly.
   
Uhm, you've heard of PID controllers, yes?  In general, in a feedback
system, one controls the driving term with a constant k1 times the error
(thus P, "proportional"), plus a constant k2 times the derivative of the
error (D, "derivative"), plus a constant k3 times the integral of the error
(I, "integral").  The proportional term does most of the work (and was the
first you implemented, initially with k1 = 1, then later with k1 = 1/50).
The differential term is vital to insure good response to large transients
and depending on the system k2 can often be near 1 (as you suggest, you're
using stiffer feedback when there's a large error; it might be interesting
to experiment with using the derivative instead).  The integral term k3 is
often overlooked but is very important for maintaining long term stability
as errors below what is nominally corrected by the P term can slowly
accumulate.  If the servo signal is noisy (as you suggest the sound
driver's answers to position queries can be), then the integral term
becomes doubly important since it will act to smooth the servo noise.

Cheers,

	- pz.

-- 
John Pezaris, Ph.D.
pz@hms.harvard.edu