sound improvements - better queue handling, refactoring and cleanups

[Antoine Martin]

0.10 fixed a number of sound issues (ie: #362 and #379) but there is more we can and should do:

• refactor the code so both the client and server use the same logic for creating and destroying the pipelines (servers typiccally run on Linux so they tend to be more forgiving - but we still tend to play catch up with fixes already client side: #396)
• if the pipeline fails to start, we don't tell the client/server that requested it about the failure at the moment, which leaves a dead pipeline hanging. This is wasting cpu and showing the wrong state in the UI or "xpra info".
• the way we deal with overruns is to re-start the whole pipeline which is wasteful (and potentially problematic), we should instead use the fact that changing the min-threshold-time (and maybe other attributes?) flushes the whole queue, effectively clearing the overrun
• choose the sound quality: stereo vs mono, 48KHz, etc..
• use a python wrapper to access pulseaudio rather than execing commands, one of:
  • ​python-pulseaudio
  • ​pypactl

[Antoine Martin]

some minor improvements were made for 0.11, the rest should get done for 0.12

[Antoine Martin]

This is needed, but needs to be done early in the release cycle to find all the problems. Re-scheduling.

[Antoine Martin]

Some cleanups were done as part of #90: see r6315, r6308, etc

Too late for the large refactoring.. so re-scheduling it again.

[Antoine Martin]

See also: #663, #669 and #673.

[Antoine Martin]

Started in r7700.

[Antoine Martin]

r7757 fixes this sink warning found on OSX:

pipeline warning: ['gstbasesink.c(3638): gst_base_sink_chain_unlocked (): \
    /GstPipeline:pipeline1/GstOsxAudioSink:osxaudiosink1:\n\
    Received buffer without a new-segment. Assuming timestamps start from 0.']

afarr: please check that I haven't broken anything on win32 or osx.

[alas]

Sound still seems good, both OSX and Windows. SInk warning also gone.

Probably good to close this.

[Antoine Martin]

I'm not closing this ticket yet because there are still quite a few items that I haven't dealt with.

[Antoine Martin]

We can also deal with #669 and use multiprocessing.

[Antoine Martin]

We can't use ​multiprocessing with pygst because it relies on the gobject / glib main loop exclusively - and we can only have on of those.

[Antoine Martin]

deal with underruns better: try to use min-threshold-time

[Antoine Martin]

The "sound underrun" patch has been applied as part of r8786 for #669.

This may affect underrun and overrun events...

[Antoine Martin]

r8827 allows us to disable this new variable min queue code using:

XPRA_VARIABLE_MIN_QUEUE=0 xpra attach ...

[Antoine Martin]

Some improvements in r8870 (see commit message).

[Antoine Martin]

allows us to compress raw WAV using the standard packet compressors (only saves about 25% with lz4, probably not worth it)

[Antoine Martin]

I think this ticket is good to close, we can deal with testing in #669.

If someone has time, running benchmarks with the netcompress patch could be useful, to see what kind of trade-off we get: 25% compression is poor, but the CPU load is going to be minimal..