[Bug 56702] New: Resampler cutoff frequency for upsampling seems too low

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=56702

Bug ID: 56702
           Summary: Resampler cutoff frequency for upsampling seems too
                    low
           Product: Wine
           Version: 9.8
          Hardware: x86-64
                OS: Linux
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: dsound
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

When you use wine directsound and mix in a secondary buffer whose sample rate
is below your device sample rate (for example, upsampling 8000/11025 to
44100/48000), the resampled output has a frequency cutoff at around 1/4 of the
secondary buffer's sample rate (1/2 nyquist frequency) instead of where it's
supposed to be - around 1/2 the sample rate (or ~85% of nyquist freq, going by
make_fir).  You can also visually verify this if you view the spectrum of the
resampled output.

This can be very noticeable especially if your input sounds are of a lower
sample rate but still have higher frequencies present.

Other related aspects seem to be working correctly (mmdev opening the device at
requested sample rate, pulseaudio negotiating sample format correctly according
to logs, no other resampling in the chain).

It seems that the fir resampler logic (mixer.c, fir.h, make_fir) hasn't
substantially changed since it was first added over a decade ago, so this seems
less likely to be a regression.

Additionally, I confirmed that the linear resampler from DSound-Fast_Mixer
patch (from wine-staging) does not have this issue - you can force it to always
use the linear resampler by setting HQBuffersMax reg key to 0.

I don't have a test case to upload, but it should be fairly simple to
reproduce.
Tested on wine-devel 9.8 and 9.9 (linux mint 21.3 / ubuntu 22.04)

Thanks!

[Bug 56702] Resampler cutoff frequency for upsampling seems too low

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=56702

Ken Sharp  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
    Staged patchset|                            |https://gitlab.winehq.org/w
                   |                            |ine/wine-staging/-/tree/mas
                   |                            |ter/patches/dsound-Fast_Mix
                   |                            |er
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |STAGED

--- Comment #1 from Ken Sharp  ---
Are you aware of any applications affected by this?

[Bug 56702] Resampler cutoff frequency for upsampling seems too low

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=56702

--- Comment #2 from [e-mail] ---
Yes, it's an old now-defunct mmo - Asheron's Call (retail servers are no longer
up, but there are emulators and the client can still be downloaded).  It's also
likely that other older apps/games might use lower sampling rate sounds with
directsound.

The game client itself uses a pretty wide variety of sampling rates for its
sounds/secondary buffers (8000/11025 mono for the older sounds all the way up
to 44100 stereo for the newer), so it is pretty obvious when lower samplerate
sounds get incorrectly resampled.

Also, I noticed if you simply double both $approx_bandwidth and $fir_step
values (from 0.85/120 to 1.70/240, respectively) and regenerate fir.h with it,
that the resulting spectrographs appear more in line with expectations and the
resampled audio sounds better as well.

Don't take that as a real fix, however, as I'm not an expert in dsp.  Not sure
if that is actually the correct thing to do or if there's an actual calculation
bug in make_fir (or even mixer.c), so I'm hoping someone who knows this stuff
well might chime in :)

[Bug 56702] Resampler cutoff frequency for upsampling seems too low

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=56702

--- Comment #3 from [e-mail] ---
Might have misspoke in my last comment - those changes to make_fir actually
introduce aliasing.  Please ignore that, my apologies.

Still messing with it, will update if I have something more to share.

[Bug 56702] Resampler cutoff frequency for upsampling seems too low

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=56702

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |WORKSFORME
             Status|STAGED                      |RESOLVED

--- Comment #4 from [e-mail] ---
Please ignore this bug - I made a mistake and the dsound resampler works fine. 
I ended up comparing the spectrograms of the original extracted sounds and
recorded sounds (with both the staging patch and without), and the resampler
bandpass/bandwidth is correct as per comments in make_fir.

[Bug 56702] Resampler cutoff frequency for upsampling seems too low

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=56702

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |CLOSED

--- Comment #5 from [e-mail] ---
Closing the bug

[Bug 56702] Resampler cutoff frequency for upsampling seems too low

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=56702

Zeb Figura  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|WORKSFORME                  |INVALID