--HG-- extra : rebase_source : cfee1d5353fa3cacf4df8712fde15cd94e2ee3d4
The xine libmad adaptor seemed not to forward the pts to the metronome: It buffers the MPEG audio packets until a threshold is reached (MAD_MIN_SIZE: 2889 bytes) and then has libmad decode the packets which is send to audio out. The pts of the last audio packet is forwarded on to metronome which can then sync video with audio. For the channel4 channels MPEG audio packets have a size of 576 bytes which means it takes five packets to fill the buffer enough for processing. In the stream every fifth audio packet contains a pts. The result of this is: If after a seek, the last audio packet is the one with the pts, video and audio are in sync. If the pts is in any of the four previous ones no pts will reach metronome and video and audio will never be synced before a new seek and even then there's a one in five chance that video and audio are not synced. Other channels did not show this behaviour because e.g. BBC One has an audio packet size of about 750 bytes and send a pts every fifth packet as well. This means that not every pts from the stream gets through to metronome but some do. This also means that syncing after a seek is probably not as quick as it could be but it will sync. My workaround to this problem is to start decoding not only when a the buffer has reached a threshold but also when a pts != 0 arrives. This does mean however that the buffer isn't always filled to the theshold and decoding might not perform as well as it could. --HG-- extra : transplant_source : %EC%90%EB%AA%8A%C7%BD%A4%B7%EE%F5%E9%E8SY%89S%9D0s
This is optional, and some systems don't support it. POSIX defines the _POSIX_THREAD_PRIORITY_SCHEDULING to tell that support is present.
In dvd_input.c:file_read(), if the read is not done in a single access, the buffer pointer doesn't get advanced so that the second read overwrites the data of the first. I actually don't remember under which circumstances this could happen, but the flaw in the code is obvious. The patch is from NetBSD's pkgsrc, and this patch is attributed to an "unnamed contributor" in the CVS log.
In demux_loop(), a time value is calculated by adding to the fractional part. In case a second barrier is crossed, the value is not in its canonical form anymore - the fractional part is larger than 10^9-1. It should be normalized for portability. While I haven't found a formal requirement for this in POSIX, NetBSD's libpthread checks for it and complains.