Track the virtual and real output buffers ecplicitly

3 files changed, 79 insertions, 46 deletions

diff --git a/Alc/ALc.c b/Alc/ALc.c
index fd44ed28..fb9f2caa 100644
--- a/Alc/ALc.c
+++ b/Alc/ALc.c
@@ -1855,6 +1855,10 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
 
     al_free(device->DryBuffer);
     device->DryBuffer = NULL;
+    device->VirtOut.Buffer = NULL;
+    device->VirtOut.NumChannels = 0;
+    device->RealOut.Buffer = NULL;
+    device->RealOut.NumChannels = 0;
 
     UpdateClockBase(device);
 
@@ -2124,6 +2128,21 @@ static ALCenum UpdateDeviceParams(ALCdevice *device, const ALCint *attrList)
         return ALC_INVALID_DEVICE;
     }
 
+    if(device->Hrtf || device->Uhj_Encoder)
+    {
+        device->VirtOut.Buffer = device->DryBuffer;
+        device->VirtOut.NumChannels = device->NumChannels;
+        device->RealOut.Buffer = device->DryBuffer + device->NumChannels;
+        device->RealOut.NumChannels = 2;
+    }
+    else
+    {
+        device->VirtOut.Buffer = NULL;
+        device->VirtOut.NumChannels = 0;
+        device->RealOut.Buffer = device->DryBuffer;
+        device->RealOut.NumChannels = device->NumChannels;
+    }
+
     SetMixerFPUMode(&oldMode);
     V0(device->Backend,lock)();
     context = ATOMIC_LOAD(&device->ContextList);
@@ -2261,6 +2280,10 @@ static ALCvoid FreeDevice(ALCdevice *device)
 
     al_free(device->DryBuffer);
     device->DryBuffer = NULL;
+    device->VirtOut.Buffer = NULL;
+    device->VirtOut.NumChannels = 0;
+    device->RealOut.Buffer = NULL;
+    device->RealOut.NumChannels = 0;
 
     al_free(device);
 }
@@ -3361,6 +3384,10 @@ ALC_API ALCdevice* ALC_APIENTRY alcOpenDevice(const ALCchar *deviceName)
     device->Render_Mode = NormalRender;
     AL_STRING_INIT(device->DeviceName);
     device->DryBuffer = NULL;
+    device->VirtOut.Buffer = NULL;
+    device->VirtOut.NumChannels = 0;
+    device->RealOut.Buffer = NULL;
+    device->RealOut.NumChannels = 0;
 
     ATOMIC_INIT(&device->ContextList, NULL);
 
@@ -3616,6 +3643,10 @@ ALC_API ALCdevice* ALC_APIENTRY alcCaptureOpenDevice(const ALCchar *deviceName,
 
     AL_STRING_INIT(device->DeviceName);
     device->DryBuffer = NULL;
+    device->VirtOut.Buffer = NULL;
+    device->VirtOut.NumChannels = 0;
+    device->RealOut.Buffer = NULL;
+    device->RealOut.NumChannels = 0;
 
     InitUIntMap(&device->BufferMap, ~0);
     InitUIntMap(&device->EffectMap, ~0);
@@ -3806,6 +3837,10 @@ ALC_API ALCdevice* ALC_APIENTRY alcLoopbackOpenDeviceSOFT(const ALCchar *deviceN
     device->Render_Mode = NormalRender;
     AL_STRING_INIT(device->DeviceName);
     device->DryBuffer = NULL;
+    device->VirtOut.Buffer = NULL;
+    device->VirtOut.NumChannels = 0;
+    device->RealOut.Buffer = NULL;
+    device->RealOut.NumChannels = 0;
 
     ATOMIC_INIT(&device->ContextList, NULL);
 
diff --git a/Alc/ALu.c b/Alc/ALu.c
index 38a4ef71..599ba3c2 100644
--- a/Alc/ALu.c
+++ b/Alc/ALu.c
@@ -581,8 +581,8 @@ ALvoid CalcNonAttnSourceParams(ALvoice *voice, const ALsource *ALSource, const A
                  * channels and write FrontLeft and FrontRight inputs to the
                  * first and second outputs.
                  */
-                voice->Direct.OutBuffer += voice->Direct.OutChannels;
-                voice->Direct.OutChannels = 2;
+                voice->Direct.OutBuffer = Device->RealOut.Buffer;
+                voice->Direct.OutChannels = Device->RealOut.NumChannels;
                 for(c = 0;c < num_channels;c++)
                 {
                     for(j = 0;j < MAX_OUTPUT_CHANNELS;j++)
@@ -632,8 +632,8 @@ ALvoid CalcNonAttnSourceParams(ALvoice *voice, const ALsource *ALSource, const A
             /* Full HRTF rendering. Skip the virtual channels and render each
              * input channel to the real outputs.
              */
-            voice->Direct.OutBuffer += voice->Direct.OutChannels;
-            voice->Direct.OutChannels = 2;
+            voice->Direct.OutBuffer = Device->RealOut.Buffer;
+            voice->Direct.OutChannels = Device->RealOut.NumChannels;
             for(c = 0;c < num_channels;c++)
             {
                 if(chans[c].channel == LFE)
@@ -1131,8 +1131,8 @@ ALvoid CalcSourceParams(ALvoice *voice, const ALsource *ALSource, const ALCconte
         ALfloat dirfact = 1.0f;
         ALfloat coeffs[MAX_AMBI_COEFFS];
 
-        voice->Direct.OutBuffer += voice->Direct.OutChannels;
-        voice->Direct.OutChannels = 2;
+        voice->Direct.OutBuffer = Device->RealOut.Buffer;
+        voice->Direct.OutChannels = Device->RealOut.NumChannels;
 
         if(Distance > FLT_EPSILON)
         {
@@ -1393,26 +1393,13 @@ ALvoid aluMixData(ALCdevice *device, ALvoid *buffer, ALsizei size)
 
     while(size > 0)
     {
-        ALfloat (*OutBuffer)[BUFFERSIZE];
-        ALuint OutChannels;
-
         IncrementRef(&device->MixCount);
 
-        OutBuffer = device->DryBuffer;
-        OutChannels = device->NumChannels;
-
         SamplesToDo = minu(size, BUFFERSIZE);
-        for(c = 0;c < OutChannels;c++)
-            memset(OutBuffer[c], 0, SamplesToDo*sizeof(ALfloat));
-        if(device->Hrtf || device->Uhj_Encoder)
-        {
-            /* Set OutBuffer/OutChannels to correspond to the actual output
-             * with HRTF. Make sure to clear them too. */
-            OutBuffer += OutChannels;
-            OutChannels = 2;
-            for(c = 0;c < OutChannels;c++)
-                memset(OutBuffer[c], 0, SamplesToDo*sizeof(ALfloat));
-        }
+        for(c = 0;c < device->VirtOut.NumChannels;c++)
+            memset(device->VirtOut.Buffer[c], 0, SamplesToDo*sizeof(ALfloat));
+        for(c = 0;c < device->RealOut.NumChannels;c++)
+            memset(device->RealOut.Buffer[c], 0, SamplesToDo*sizeof(ALfloat));
 
         V0(device->Backend,lock)();
 
@@ -1495,12 +1482,13 @@ ALvoid aluMixData(ALCdevice *device, ALvoid *buffer, ALsizei size)
             ALuint irsize = GetHrtfIrSize(device->Hrtf);
             MixHrtfParams hrtfparams;
             memset(&hrtfparams, 0, sizeof(hrtfparams));
-            for(c = 0;c < device->NumChannels;c++)
+            for(c = 0;c < device->VirtOut.NumChannels;c++)
             {
                 hrtfparams.Current = &device->Hrtf_Params[c];
                 hrtfparams.Target = &device->Hrtf_Params[c];
-                HrtfMix(OutBuffer, device->DryBuffer[c], 0, device->Hrtf_Offset,
-                    0, irsize, &hrtfparams, &device->Hrtf_State[c], SamplesToDo
+                HrtfMix(device->RealOut.Buffer, device->VirtOut.Buffer[c], 0,
+                    device->Hrtf_Offset, 0, irsize, &hrtfparams,
+                    &device->Hrtf_State[c], SamplesToDo
                 );
             }
             device->Hrtf_Offset += SamplesToDo;
@@ -1510,7 +1498,8 @@ ALvoid aluMixData(ALCdevice *device, ALvoid *buffer, ALsizei size)
             if(device->Uhj_Encoder)
             {
                 /* Encode to stereo-compatible 2-channel UHJ output. */
-                EncodeUhj2(device->Uhj_Encoder, OutBuffer, device->DryBuffer, SamplesToDo);
+                EncodeUhj2(device->Uhj_Encoder, device->RealOut.Buffer,
+                           device->VirtOut.Buffer, SamplesToDo);
             }
             if(device->Bs2b)
             {
@@ -1518,17 +1507,20 @@ ALvoid aluMixData(ALCdevice *device, ALvoid *buffer, ALsizei size)
                 for(i = 0;i < SamplesToDo;i++)
                 {
                     float samples[2];
-                    samples[0] = OutBuffer[0][i];
-                    samples[1] = OutBuffer[1][i];
+                    samples[0] = device->RealOut.Buffer[0][i];
+                    samples[1] = device->RealOut.Buffer[1][i];
                     bs2b_cross_feed(device->Bs2b, samples);
-                    OutBuffer[0][i] = samples[0];
-                    OutBuffer[1][i] = samples[1];
+                    device->RealOut.Buffer[0][i] = samples[0];
+                    device->RealOut.Buffer[1][i] = samples[1];
                 }
             }
         }
 
         if(buffer)
         {
+            ALfloat (*OutBuffer)[BUFFERSIZE] = device->RealOut.Buffer;
+            ALuint OutChannels = device->RealOut.NumChannels;;
+
 #define WRITE(T, a, b, c, d) do {               \
     Write_##T((a), (b), (c), (d));              \
     buffer = (T*)buffer + (c)*(d);              \
diff --git a/Alc/effects/reverb.c b/Alc/effects/reverb.c
index 995ad6a5..a92fbd5b 100644
--- a/Alc/effects/reverb.c
+++ b/Alc/effects/reverb.c
@@ -364,6 +364,9 @@ static ALboolean ALreverbState_deviceUpdate(ALreverbState *State, ALCdevice *Dev
     if(!AllocLines(frequency, State))
         return AL_FALSE;
 
+    /* WARNING: This assumes the real output follows the virtual output in the
+     * device's DryBuffer.
+     */
     State->ExtraChannels = (Device->Hrtf || Device->Uhj_Encoder) ? 2 : 0;
 
     // Calculate the modulation filter coefficient.  Notice that the exponent
@@ -666,16 +669,19 @@ static ALvoid UpdateMixedPanning(const ALCdevice *Device, const ALfloat *Reflect
     ALfloat length;
     ALuint i;
 
-    /* With HRTF, the normal output provides a panned reverb channel when a
-     * non-0-length vector is specified, while the real stereo output provides
-     * two other "direct" non-panned reverb channels.
+    /* With HRTF or UHJ, the normal output provides a panned reverb channel
+     * when a non-0-length vector is specified, while the real stereo output
+     * provides two other "direct" non-panned reverb channels.
+     *
+     * WARNING: This assumes the real output follows the virtual output in the
+     * device's DryBuffer.
      */
     memset(State->Early.PanGain, 0, sizeof(State->Early.PanGain));
     length = sqrtf(ReflectionsPan[0]*ReflectionsPan[0] + ReflectionsPan[1]*ReflectionsPan[1] + ReflectionsPan[2]*ReflectionsPan[2]);
     if(!(length > FLT_EPSILON))
     {
-        for(i = 0;i < 2;i++)
-            State->Early.PanGain[i&3][Device->NumChannels+i] = Gain * EarlyGain;
+        for(i = 0;i < Device->RealOut.NumChannels;i++)
+            State->Early.PanGain[i&3][Device->VirtOut.NumChannels+i] = Gain * EarlyGain;
     }
     else
     {
@@ -691,19 +697,19 @@ static ALvoid UpdateMixedPanning(const ALCdevice *Device, const ALfloat *Reflect
         length = minf(length, 1.0f);
 
         CalcDirectionCoeffs(pan, coeffs);
-        ComputePanningGains(Device->AmbiCoeffs, Device->NumChannels, coeffs, Gain, DirGains);
-        for(i = 0;i < Device->NumChannels;i++)
+        ComputePanningGains(Device->AmbiCoeffs, Device->VirtOut.NumChannels, coeffs, Gain, DirGains);
+        for(i = 0;i < Device->VirtOut.NumChannels;i++)
             State->Early.PanGain[3][i] = DirGains[i] * EarlyGain * length;
-        for(i = 0;i < 2;i++)
-            State->Early.PanGain[i&3][Device->NumChannels+i] = Gain * EarlyGain * (1.0f-length);
+        for(i = 0;i < Device->RealOut.NumChannels;i++)
+            State->Early.PanGain[i&3][Device->VirtOut.NumChannels+i] = Gain * EarlyGain * (1.0f-length);
     }
 
     memset(State->Late.PanGain, 0, sizeof(State->Late.PanGain));
     length = sqrtf(LateReverbPan[0]*LateReverbPan[0] + LateReverbPan[1]*LateReverbPan[1] + LateReverbPan[2]*LateReverbPan[2]);
     if(!(length > FLT_EPSILON))
     {
-        for(i = 0;i < 2;i++)
-            State->Late.PanGain[i&3][Device->NumChannels+i] = Gain * LateGain;
+        for(i = 0;i < Device->RealOut.NumChannels;i++)
+            State->Late.PanGain[i&3][Device->VirtOut.NumChannels+i] = Gain * LateGain;
     }
     else
     {
@@ -715,11 +721,11 @@ static ALvoid UpdateMixedPanning(const ALCdevice *Device, const ALfloat *Reflect
         length = minf(length, 1.0f);
 
         CalcDirectionCoeffs(pan, coeffs);
-        ComputePanningGains(Device->AmbiCoeffs, Device->NumChannels, coeffs, Gain, DirGains);
-        for(i = 0;i < Device->NumChannels;i++)
+        ComputePanningGains(Device->AmbiCoeffs, Device->VirtOut.NumChannels, coeffs, Gain, DirGains);
+        for(i = 0;i < Device->VirtOut.NumChannels;i++)
             State->Late.PanGain[3][i] = DirGains[i] * LateGain * length;
-        for(i = 0;i < 2;i++)
-            State->Late.PanGain[i&3][Device->NumChannels+i] = Gain * LateGain * (1.0f-length);
+        for(i = 0;i < Device->RealOut.NumChannels;i++)
+            State->Late.PanGain[i&3][Device->VirtOut.NumChannels+i] = Gain * LateGain * (1.0f-length);
     }
 }