dsound: clamp input buffer size at 62.5 ms

This works around a DirectSound limitation where input host buffer sizes
smaller than 31.25 ms are basically unworkable and make PortAudio hang.
The workaround is to impose a minimal buffer size of 2*31.25 ms on
input-only and full-duplex streams. This is enough for the read cursor
to advance twice around the buffer, basically resulting in de facto
double buffering.

This change was tested with `paloopback` under a wide variety of
half/full-duplex, framesPerBuffer, and suggested latency parameters.
(Note the testing was done on top of #772 as otherwise paloopback is not
usable.)

Fixes #775

Author: dechamps

src/hostapi/dsound/pa_win_ds.c

@@ -326,7 +326,7 @@ typedef struct PaWinDsStream
 /* Set minimal latency based on the current OS version.
  * NT has higher latency.
  */
-static double PaWinDS_GetMinSystemLatencySeconds( void )
+static DWORD PaWinDS_GetWindowsMajorVersion( void )
 {
 /*
 NOTE: GetVersionEx() is deprecated as of Windows 8.1 and can not be used to reliably detect
@@ -338,23 +338,33 @@ See: http://www.codeproject.com/Articles/678606/Part-Overcoming-Windows-s-deprec
 */
 #pragma warning (disable : 4996) /* use of GetVersionEx */
 
+    OSVERSIONINFO osvi;
+    osvi.dwOSVersionInfoSize = sizeof(osvi);
+    GetVersionEx(&osvi);
+    return osvi.dwMajorVersion;
+
+#pragma warning (default : 4996)
+}
+
+
+/* Set minimal latency based on the current OS version.
+ * NT has higher latency.
+ */
+static double PaWinDS_GetMinSystemLatencySeconds( void )
+{
     double minLatencySeconds;
     /* Set minimal latency based on whether NT or other OS.
      * NT has higher latency.
      */
 
-    OSVERSIONINFO osvi;
-    osvi.dwOSVersionInfoSize = sizeof( osvi );
-    GetVersionEx( &osvi );
-    DBUG(("PA - PlatformId = 0x%x\n", osvi.dwPlatformId ));
-    DBUG(("PA - MajorVersion = 0x%x\n", osvi.dwMajorVersion ));
-    DBUG(("PA - MinorVersion = 0x%x\n", osvi.dwMinorVersion ));
+    const DWORD windowsMajorVersion = PaWinDS_GetWindowsMajorVersion();
+    DBUG(("PA - windowsMajorVersion = 0x%x\n", windowsMajorVersion));
     /* Check for NT */
-    if( (osvi.dwMajorVersion == 4) && (osvi.dwPlatformId == 2) )
+    if( (windowsMajorVersion == 4) && (windowsMajorVersion == 2) )
     {
         minLatencySeconds = PA_DS_WIN_NT_DEFAULT_LATENCY_;
     }
-    else if(osvi.dwMajorVersion >= 5)
+    else if(windowsMajorVersion >= 5)
     {
         minLatencySeconds = PA_DS_WIN_WDM_DEFAULT_LATENCY_;
     }
@@ -363,8 +373,6 @@ See: http://www.codeproject.com/Articles/678606/Part-Overcoming-Windows-s-deprec
         minLatencySeconds = PA_DS_WIN_9X_DEFAULT_LATENCY_;
     }
     return minLatencySeconds;
-
-#pragma warning (default : 4996)
 }
 
 
@@ -1741,7 +1749,7 @@ static HRESULT InitOutputBuffer( PaWinDsStream *stream, PaWinDsDeviceInfo *devic
 
 static void CalculateBufferSettings( unsigned long *hostBufferSizeFrames,
                                     unsigned long *pollingPeriodFrames,
-                                    int isFullDuplex,
+                                    int hasInput, int hasOutput,
                                     unsigned long suggestedInputLatencyFrames,
                                     unsigned long suggestedOutputLatencyFrames,
                                     double sampleRate, unsigned long userFramesPerBuffer )
@@ -1751,7 +1759,7 @@ static void CalculateBufferSettings( unsigned long *hostBufferSizeFrames,
     unsigned long pollingJitterFrames = (unsigned long)(sampleRate * PA_DS_POLLING_JITTER_SECONDS);
 
     unsigned long adjustedSuggestedOutputLatencyFrames = suggestedOutputLatencyFrames;
-    if( userFramesPerBuffer != paFramesPerBufferUnspecified && isFullDuplex )
+    if( userFramesPerBuffer != paFramesPerBufferUnspecified && hasInput && hasOutput )
     {
         /* In full duplex streams we know that the buffer adapter adds userFramesPerBuffer
            extra fixed latency. so we subtract it here as a fixed latency before computing
@@ -1773,6 +1781,21 @@ static void CalculateBufferSettings( unsigned long *hostBufferSizeFrames,
             userFramesPerBuffer;
     *hostBufferSizeFrames = intendedUserFramesPerBuffer
         + max( intendedUserFramesPerBuffer + pollingJitterFrames, targetBufferingLatencyFrames );
+
+    /* In some (most?) systems, DirectSound has an odd limitation where it always uses
+       a fixed 31.25 ms granularity for the read cursor, regardless of parameters.
+       This in turn means that if we allocate an input buffer that is less than 31.25 ms,
+       the read cursor will stay stuck at zero. See https://github.com/PortAudio/portaudio/issues/775
+       To work around this problem, ensure that the input host buffer is large enough
+       for at least two 31.25 ms buffer "halves".
+
+       On pre-Vista Windows, we don't do this because DirectSound is implemented very
+       differently, and is therefore unlikely to suffer from the same issue.
+    */
+    if( hasInput && PaWinDS_GetWindowsMajorVersion() >= 6 )
+    {
+        *hostBufferSizeFrames = max( *hostBufferSizeFrames, 2 * 0.03125 * sampleRate );
+    }
 }
 
 
@@ -2108,7 +2131,8 @@ static PaError OpenStream( struct PaUtilHostApiRepresentation *hostApi,
         else
         {
             CalculateBufferSettings( (unsigned long*)&stream->hostBufferSizeFrames, &pollingPeriodFrames,
-                    /* isFullDuplex = */ (inputParameters && outputParameters),
+                    /* hasInput = */ !!inputParameters,
+                    /* hasOutput = */ !!outputParameters,
                     suggestedInputLatencyFrames,
                     suggestedOutputLatencyFrames,
                     sampleRate, framesPerBuffer );