On Sun, Jan 10, 2021 at 11:01:29AM +0100, Volker Rümelin wrote: > A mix of bug fixes and improvements. > > Patches 01/23 - 11/23 have a few SDL fixes and add audio recording > functions to the SDL audio backend. > > Patch 12/23 removes unnecessary code from audio/audio.c. > > Patches 13/23 - 16/23 fix a few PulseAudio backend bugs. > > Patch 17/23 shows a PulseAudio backend bug. So far I don't think I > broke any fine tuned user settings and I don't want to do it here. > > Patches 18/23 - 19/23 are the first steps towards glitch free and > lower latency PulseAudio playback and recording. Currently qemu > uses incredibly large buffers in the PulseAudio backend. For > playback this just increases the playback latency but doesn't > improve dropout safety, because PulseAudio can't access this buffer > directly. With these patches it's possible to move the large qemu > buffer to the PulseAudio server side and just keep a small buffer > on the qemu side. On the PulseAudio server side PulseAudio tries > to place a part of these buffers directly on the hardware and > PulseAudio runs with a higher priority than qemu, so it has a > better chance to deliver audio data in time. > > Here is an example to show how this works: > -device intel-hda -device hda-duplex,audiodev=audio0 > -machine pcspk-audiodev=audio0 -audiodev pa,id=audio0, > out.buffer-length=14000,out.latency=46440,in.latency=46440 > > Due to a bug in the PulseAudio backend, these command line options > actually decrease the playback latency compared to current defaults. > For playback with defaults (16 bits, stereo, 44100 samples/s) > we have a 15ms server side buffer + 2 * 46.44ms qemu audio buffer > + 23.22ms hda codec buffer = 131.1ms latency. With my example it's > 46.44ms + 2 * 14ms + 23.22ms = 97.66ms latency and I guess you > won't hear any drop outs. Btw.: 14ms = 10ms timer-period + 4ms > additional playback data the hda codec can produce in timer-period > time. > > Patches 20/23 - 23/23 fix small issues with DirectSound.
Nice overall improvements. Some minor nits (see replies to some patches). Can be fixed incrementally though, so I'll do some more testing and if nothing blows up queue the series for merge. thanks, Gerd
