2010-09-02 16:06, pl bossart skrev: >> Agreed: You can pick those two patches, and then we add a third patch to >> both branches, which brings back the watermark for tsched devices and 20 >> ms for non-tsched. Assuming my suspicion is not disproved, of course. >> What does Pierre think of that? > > I don't want the watermark to be used for rewinds. The watermark is > there for timer-based scheduling, so that you have enough time to > wake-up from sleep and still refill the buffer. > The rewinds happens when the processor is already awake, pulseaudio up > and running and only the remix part needs to happen. Plus the > watermark varies and the logic could really be improved. > > Also I think 20ms for rewinds is way too much. This will kill your > actual latency. Imagine you have a low-latency app that starts, the > first sample would be heard after at best 20ms. Not acceptable for > speech or interactive sounds. > > But I agree that 256-bytes isn't fool-proof for heavy duty use cases > such 8ch 192kHz 32-bit float. > > So how about we keep 256 bytes (1.33 ms for 48kHz) but add a 1.33 ms > threshold to make sure we never rewind below. > > rewind_safeguard = max(256, pa_usec_to_bytes(1330)); > > This way you solve both the hardware issue (frequency independent) and > leave enough headroom for the system to avoid underflows. Whether 1,33 ms or 20 ms is best - I assume your guess is as good as mine. Colin, feel free to go ahead with Pierre's suggestion - it's likely to be good enough. As for the watermark usage, I admit to not knowing enough of CPU scheduling and wake-up times to either prove Pierre right or wrong. -- David Henningsson, Canonical Ltd. http://launchpad.net/~diwic
