[Bug 30191] New: Stuttering mmdevapi audio (Mac OS X)

http://bugs.winehq.org/show_bug.cgi?id=30191

Bug #: 30191
           Summary: Stuttering mmdevapi audio (Mac OS X)
           Product: Wine
           Version: 1.4
          Platform: x86
        OS/Version: Mac OS X
            Status: NEW
          Severity: normal
          Priority: P2
         Component: mmdevapi
        AssignedTo: [e-mail]
        ReportedBy: [e-mail]
                CC: [e-mail]
    Classification: Unclassified

The interactive mmdevapi render tests stutters quasi continuously on my Mac OS
X "early 2009" nVidia Mac mini.

Actually, stuttering can be interpreted as an indicator of low-latency, which
is good :-)  Consider winecoreaudio's 20ms period.  If audio would only start
at e.g. t=9ms, then at period time t=20, underrun would be very far when
writing the next frames.

The reasons for stuttering are known from winealsa, yet the details are
different: lead-in and regularity.

Lead-in:
Note that mmdevapi/render.c:test_worst_case does not behave like XAudio2 with
winecoreaudio and wineoss, due to their use of a 20ms period.  Alexey's
analysis in bug #28723, comment #19 is that XAudio then writes a lead-in of
3/4*4 = 3 periods, unlike the 1 period written in the 10ms period case. 
test_worst_case uses 0 or 1 period.
So XAudio2 apps may not stutter, whereas the mmdevapi test does.  That does not
imply that our winecoreaudio code is great, rather than that we're forbidden to
reduce the period to the 10ms that native uses everywhere.

Unlike ALSA, I see no notion of underrun in CoreAudio.  So I don't see when
writing a lead-in would be appropriate, except at start.  I think we need
something else.

Regularity/stability:
XAudio2 and test_worst_case write at most one period worth of frames per event.
 Writing nothing is a guarantee of underrun (unless the event rate is
artificially high to counter this).  Indeed, adding a trace("MISS") to
test_worse_case every time GetCurrentPadding is too high to add more samples
correlates with some, but not all underruns.

Again, my ntdll CreateTimerQueue stabilisator patch from bug #30071 helps a
little, as otherwise callbacks are mostly 21ms apart rather than 20ms. That
clock skew also accounts for a few underruns.

However the major point is: XAudio2 needs event signalling to match decrease of
padding, such that it adds the next period of data.  My ALSA work-in progress
code does exactly that, but CoreAudio needs it even more due to the absence of
the lead-in safety-belt.