[Bug 14717] resampled sound is horrible

http://bugs.winehq.org/show_bug.cgi?id=14717

Andrew Eikum  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #38186|0                           |1
        is obsolete|                            |

--- Comment #261 from Andrew Eikum  2012-01-03 15:36:31 CST ---
Created attachment 38246
  --> http://bugs.winehq.org/attachment.cgi?id=38246
Speex resampler again

Another patchset, possibly final. A couple of changes:

I removed the "convert to float" patch, and instead just removed freqAdjust
entirely in the patch that introduces Speex. It wasn't really used anyway.

I made one change to each of the first two patches. The first patch has this
new diff hunk:

diff --git a/dlls/dsound/dsound_convert.c b/dlls/dsound/dsound_convert.c
index 63e3dbe..dd982c6 100644
--- a/dlls/dsound/dsound_convert.c
+++ b/dlls/dsound/dsound_convert.c
@@ -66,8 +66,8 @@ static inline void src_advance(const void **src, UINT stride,
INT *count, UINT *
         ULONG adv = (*freqAcc >> DSOUND_FREQSHIFT);
         *freqAcc &= (1 << DSOUND_FREQSHIFT) - 1;
         *(const char **)src += adv * stride;
-        *count -= adv;
     }
+    *count -= 1;
 }

The original patch changed the "count" (aka "size") variable to be in terms of
the number of output frames, but the conversion functions still decremeted it
in step with the src parameter. That caused data corruption and crashes, which
this diff fixes.

The second patch was changed as follows (diff of a diff):

+static inline void cp_fields(const IDirectSoundBufferImpl *dsb, const BYTE
*ibuf,
          UINT istride, UINT ostride, UINT count, UINT freqAcc, UINT adj)
  {
      DirectSoundDevice *device = dsb->device;
 -    INT istep = dsb->pwfx->wBitsPerSample / 8, ostep =
device->pwfx->wBitsPerSample / 8;
 +    float value;
 +    ULONG adv;
-+    DWORD ipos = 0, opos = 0;
++    DWORD ipos = dsb->sec_mixpos, opos = 0;

since the get() functions work on (dsb->buff+ipos), not on the passed ibuf
value (which is actually unused, and removed in the following patch. perhaps
the following patch should be squashed? it's quite small).

Concerning channel conversions, I believe the patchset here more flexible than
the old channel conversion. The old code supported (secondary -> primary):
n -> n
6 -> 2
2 -> 8 (hack, first 2 only)
2 -> 6 (hack, first 2 only)
2 -> 1
1 -> 2
silence (catchall)

While the new code supports:
n -> n
1 -> 2
n -> 1
n -> 2 (catchall)

I tested each patch against the dsound tests and four applications
(pikachugame.exe, Dracula Origin Demo, Plants Vs Zombies Demo, and
HighwayPursuit). The entire patch sequence together has been tested on my
computer against many more applications in addition to those.