[pulseaudio-discuss] regression with tsched=0 and small fragments?

[pl bossart]

Hi all,
I have lost the afternoon chasing an audio quality issue with PA
configured with tsched=0 and small fragments (<30ms). I am doing this
on purpose to profile the behavior of PulseAudio with lots of wakeups
and samples provided in chunks at once.
With the latest 0.9.19 version I have cracking noise but no reports of
underflows, it works fine maybe once in 20 trials. With 0.9.15, no
issue at all, I can even go down to fragments of 2ms without any
issue. This is on a Atom board with classic HDAudio. I have no problem
at all in both cases with aplay -Dhw. What I noticed though is that
aplay seems to select a period size and buffer size that are multiples
of 128 bytes. For example a period of 5ms at 48kHz is actually
configured as 5.33ms and 256 frames; I suspect an alignement issue
that would cause these noises.

I have an intuition that the regression was introduced by the ALSA
configuration rewrite
SHA:557c4295107dc7374c850b0bd5331dd35e8fdd0f ,
but before I go on with a dichotomy and try to fix the issue I wanted
to check if others faced the same problem.
Thanks for your feedback
- Pierre