Use a non-interleaved DryBuffer

10 files changed, 68 insertions, 79 deletions

diff --git a/Alc/ALu.c b/Alc/ALu.c
index 22f347f5..b83e7df1 100644
--- a/Alc/ALu.c
+++ b/Alc/ALu.c
@@ -797,7 +797,7 @@ static __inline ALubyte aluF2UB(ALfloat val)
 static void Write_##T##_##N(ALCdevice *device, T *RESTRICT buffer,            \
                             ALuint SamplesToDo)                               \
 {                                                                             \
-    ALfloat (*RESTRICT DryBuffer)[MaxChannels] = device->DryBuffer;           \
+    ALfloat (*RESTRICT DryBuffer)[BUFFERSIZE] = device->DryBuffer;            \
     const enum Channel *ChanMap = device->DevChannels;                        \
     ALuint i, j;                                                              \
                                                                               \
@@ -807,7 +807,7 @@ static void Write_##T##_##N(ALCdevice *device, T *RESTRICT buffer,            \
         enum Channel chan = ChanMap[j];                                       \
                                                                               \
         for(i = 0;i < SamplesToDo;i++)                                        \
-            out[i*N] = func(DryBuffer[i][chan]);                              \
+            out[i*N] = func(DryBuffer[chan][i]);                              \
     }                                                                         \
 }
 
@@ -913,7 +913,8 @@ ALvoid aluMixData(ALCdevice *device, ALvoid *buffer, ALsizei size)
     while(size > 0)
     {
         SamplesToDo = minu(size, BUFFERSIZE);
-        memset(device->DryBuffer, 0, SamplesToDo*MaxChannels*sizeof(ALfloat));
+        for(c = 0;c < MaxChannels;c++)
+            memset(device->DryBuffer[c], 0, SamplesToDo*sizeof(ALfloat));
 
         ALCdevice_Lock(device);
         ctx = device->ContextList;
@@ -1002,7 +1003,7 @@ ALvoid aluMixData(ALCdevice *device, ALvoid *buffer, ALsizei size)
         {
             for(i = 0;i < SamplesToDo;i++)
             {
-                device->DryBuffer[i][FrontCenter] += device->ClickRemoval[FrontCenter];
+                device->DryBuffer[FrontCenter][i] += device->ClickRemoval[FrontCenter];
                 device->ClickRemoval[FrontCenter] -= device->ClickRemoval[FrontCenter] * (1.0f/256.0f);
             }
             device->ClickRemoval[FrontCenter] += device->PendingClicks[FrontCenter];
@@ -1011,38 +1012,41 @@ ALvoid aluMixData(ALCdevice *device, ALvoid *buffer, ALsizei size)
         else if(device->FmtChans == DevFmtStereo)
         {
             /* Assumes the first two channels are FrontLeft and FrontRight */
-            for(i = 0;i < SamplesToDo;i++)
+            for(c = 0;c < 2;c++)
             {
-                for(c = 0;c < 2;c++)
+                ALfloat offset = device->ClickRemoval[c];
+                for(i = 0;i < SamplesToDo;i++)
                 {
-                    device->DryBuffer[i][c] += device->ClickRemoval[c];
-                    device->ClickRemoval[c] -= device->ClickRemoval[c] * (1.0f/256.0f);
+                    device->DryBuffer[c][i] += offset;
+                    offset -= offset * (1.0f/256.0f);
                 }
-            }
-            for(c = 0;c < 2;c++)
-            {
-                device->ClickRemoval[c] += device->PendingClicks[c];
+                device->ClickRemoval[c] = offset + device->PendingClicks[c];
                 device->PendingClicks[c] = 0.0f;
             }
             if(device->Bs2b)
             {
+                float samples[2];
                 for(i = 0;i < SamplesToDo;i++)
-                    bs2b_cross_feed(device->Bs2b, &device->DryBuffer[i][0]);
+                {
+                    samples[0] = device->DryBuffer[FrontLeft][i];
+                    samples[1] = device->DryBuffer[FrontRight][i];
+                    bs2b_cross_feed(device->Bs2b, samples);
+                    device->DryBuffer[FrontLeft][i] = samples[0];
+                    device->DryBuffer[FrontRight][i] = samples[1];
+                }
             }
         }
         else
         {
-            for(i = 0;i < SamplesToDo;i++)
+            for(c = 0;c < MaxChannels;c++)
             {
-                for(c = 0;c < MaxChannels;c++)
+                ALfloat offset = device->ClickRemoval[c];
+                for(i = 0;i < SamplesToDo;i++)
                 {
-                    device->DryBuffer[i][c] += device->ClickRemoval[c];
-                    device->ClickRemoval[c] -= device->ClickRemoval[c] * (1.0f/256.0f);
+                    device->DryBuffer[c][i] += offset;
+                    offset -= offset * (1.0f/256.0f);
                 }
-            }
-            for(c = 0;c < MaxChannels;c++)
-            {
-                device->ClickRemoval[c] += device->PendingClicks[c];
+                device->ClickRemoval[c] = offset + device->PendingClicks[c];
                 device->PendingClicks[c] = 0.0f;
             }
         }
diff --git a/Alc/alcDedicated.c b/Alc/alcDedicated.c
index 64c2910b..2fb3dbbe 100644
--- a/Alc/alcDedicated.c
+++ b/Alc/alcDedicated.c
@@ -66,19 +66,16 @@ static ALvoid DedicatedUpdate(ALeffectState *effect, ALCdevice *device, const AL
         state->gains[LFE] = Gain;
 }
 
-static ALvoid DedicatedProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[MaxChannels])
+static ALvoid DedicatedProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[BUFFERSIZE])
 {
     ALdedicatedState *state = (ALdedicatedState*)effect;
     const ALfloat *gains = state->gains;
-    ALuint i, s;
+    ALuint i, c;
 
-    for(i = 0;i < SamplesToDo;i++)
+    for(c = 0;c < MaxChannels;c++)
     {
-        ALfloat sample;
-
-        sample = SamplesIn[i];
-        for(s = 0;s < MaxChannels;s++)
-            SamplesOut[i][s] = sample * gains[s];
+        for(i = 0;i < SamplesToDo;i++)
+            SamplesOut[c][i] = SamplesIn[i] * gains[c];
     }
 }
 
diff --git a/Alc/alcEcho.c b/Alc/alcEcho.c
index e3c3df96..1a586488 100644
--- a/Alc/alcEcho.c
+++ b/Alc/alcEcho.c
@@ -126,7 +126,7 @@ static ALvoid EchoUpdate(ALeffectState *effect, ALCdevice *Device, const ALeffec
     ComputeAngleGains(Device, atan2f(+lrpan, 0.0f), (1.0f-dirGain)*F_PI, gain, state->Gain[1]);
 }
 
-static ALvoid EchoProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[MaxChannels])
+static ALvoid EchoProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[BUFFERSIZE])
 {
     ALechoState *state = (ALechoState*)effect;
     const ALuint mask = state->BufferLength-1;
@@ -141,12 +141,12 @@ static ALvoid EchoProcess(ALeffectState *effect, ALuint SamplesToDo, const ALflo
         /* First tap */
         smp = state->SampleBuffer[(offset-tap1) & mask];
         for(k = 0;k < MaxChannels;k++)
-            SamplesOut[i][k] += smp * state->Gain[0][k];
+            SamplesOut[k][i] += smp * state->Gain[0][k];
 
         /* Second tap */
         smp = state->SampleBuffer[(offset-tap2) & mask];
         for(k = 0;k < MaxChannels;k++)
-            SamplesOut[i][k] += smp * state->Gain[1][k];
+            SamplesOut[k][i] += smp * state->Gain[1][k];
 
         // Apply damping and feedback gain to the second tap, and mix in the
         // new sample
diff --git a/Alc/alcModulator.c b/Alc/alcModulator.c
index ce91e95a..46bf1717 100644
--- a/Alc/alcModulator.c
+++ b/Alc/alcModulator.c
@@ -84,7 +84,7 @@ static __inline ALfloat hpFilter1P(FILTER *iir, ALuint offset, ALfloat input)
 
 #define DECL_TEMPLATE(func)                                                   \
 static void Process##func(ALmodulatorState *state, ALuint SamplesToDo,        \
-  const ALfloat *SamplesIn, ALfloat (*SamplesOut)[MaxChannels])               \
+  const ALfloat *SamplesIn, ALfloat (*SamplesOut)[BUFFERSIZE])                \
 {                                                                             \
     const ALuint step = state->step;                                          \
     ALuint index = state->index;                                              \
@@ -102,7 +102,7 @@ static void Process##func(ALmodulatorState *state, ALuint SamplesToDo,        \
         samp = hpFilter1P(&state->iirFilter, 0, samp);                        \
                                                                               \
         for(k = 0;k < MaxChannels;k++)                                        \
-            SamplesOut[i][k] += state->Gain[k] * samp;                        \
+            SamplesOut[k][i] += state->Gain[k] * samp;                        \
     }                                                                         \
     state->index = index;                                                     \
 }
@@ -160,7 +160,7 @@ static ALvoid ModulatorUpdate(ALeffectState *effect, ALCdevice *Device, const AL
     }
 }
 
-static ALvoid ModulatorProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[MaxChannels])
+static ALvoid ModulatorProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[BUFFERSIZE])
 {
     ALmodulatorState *state = (ALmodulatorState*)effect;
 
diff --git a/Alc/alcReverb.c b/Alc/alcReverb.c
index 13f90511..7fe28fa9 100644
--- a/Alc/alcReverb.c
+++ b/Alc/alcReverb.c
@@ -546,7 +546,7 @@ static __inline ALvoid EAXVerbPass(ALverbState *State, ALfloat in, ALfloat *earl
 
 // This processes the reverb state, given the input samples and an output
 // buffer.
-static ALvoid VerbProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[MaxChannels])
+static ALvoid VerbProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[BUFFERSIZE])
 {
     ALverbState *State = (ALverbState*)effect;
     ALuint index, c;
@@ -566,13 +566,13 @@ static ALvoid VerbProcess(ALeffectState *effect, ALuint SamplesToDo, const ALflo
 
         // Output the results.
         for(c = 0;c < MaxChannels;c++)
-            SamplesOut[index][c] += panGain[c] * out[c&3];
+            SamplesOut[c][index] += panGain[c] * out[c&3];
     }
 }
 
 // This processes the EAX reverb state, given the input samples and an output
 // buffer.
-static ALvoid EAXVerbProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[MaxChannels])
+static ALvoid EAXVerbProcess(ALeffectState *effect, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[BUFFERSIZE])
 {
     ALverbState *State = (ALverbState*)effect;
     ALuint index, c;
@@ -584,7 +584,7 @@ static ALvoid EAXVerbProcess(ALeffectState *effect, ALuint SamplesToDo, const AL
         EAXVerbPass(State, SamplesIn[index], early, late);
 
         for(c = 0;c < MaxChannels;c++)
-            SamplesOut[index][c] += State->Early.PanGain[c]*early[c&3] +
+            SamplesOut[c][index] += State->Early.PanGain[c]*early[c&3] +
                                     State->Late.PanGain[c]*late[c&3];
     }
 }
diff --git a/Alc/mixer_inc.c b/Alc/mixer_inc.c
index f60ade65..db0941f0 100644
--- a/Alc/mixer_inc.c
+++ b/Alc/mixer_inc.c
@@ -41,7 +41,7 @@ void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
 {
     const ALint *RESTRICT DelayStep = params->Hrtf.DelayStep;
     const ALuint IrSize = GetHrtfIrSize(Device->Hrtf);
-    ALfloat (*RESTRICT DryBuffer)[MaxChannels];
+    ALfloat (*RESTRICT DryBuffer)[BUFFERSIZE];
     ALfloat *RESTRICT ClickRemoval, *RESTRICT PendingClicks;
     ALfloat (*RESTRICT CoeffStep)[2] = params->Hrtf.CoeffStep;
     ALfloat (*RESTRICT TargetCoeffs)[2] = params->Hrtf.Coeffs[srcchan];
@@ -103,8 +103,8 @@ void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
         Offset++;
 
         ApplyCoeffsStep(Offset, Values, IrSize, Coeffs, CoeffStep, left, right);
-        DryBuffer[OutPos][FrontLeft]  += Values[Offset&HRIR_MASK][0];
-        DryBuffer[OutPos][FrontRight] += Values[Offset&HRIR_MASK][1];
+        DryBuffer[FrontLeft][OutPos]  += Values[Offset&HRIR_MASK][0];
+        DryBuffer[FrontRight][OutPos] += Values[Offset&HRIR_MASK][1];
 
         OutPos++;
         Counter--;
@@ -123,8 +123,8 @@ void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
         Offset++;
 
         ApplyCoeffs(Offset, Values, IrSize, Coeffs, left, right);
-        DryBuffer[OutPos][FrontLeft]  += Values[Offset&HRIR_MASK][0];
-        DryBuffer[OutPos][FrontRight] += Values[Offset&HRIR_MASK][1];
+        DryBuffer[FrontLeft][OutPos]  += Values[Offset&HRIR_MASK][0];
+        DryBuffer[FrontRight][OutPos] += Values[Offset&HRIR_MASK][1];
 
         OutPos++;
     }
@@ -147,7 +147,7 @@ void MixDirect(ALsource *Source, ALCdevice *Device, DirectParams *params,
   const ALfloat *RESTRICT data, ALuint srcchan,
   ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
 {
-    ALfloat (*RESTRICT DryBuffer)[MaxChannels];
+    ALfloat (*RESTRICT DryBuffer)[BUFFERSIZE];
     ALfloat *RESTRICT ClickRemoval, *RESTRICT PendingClicks;
     ALIGN(16) ALfloat DrySend[MaxChannels];
     ALuint pos;
@@ -167,13 +167,12 @@ void MixDirect(ALsource *Source, ALCdevice *Device, DirectParams *params,
         for(c = 0;c < MaxChannels;c++)
             ClickRemoval[c] -= data[pos]*DrySend[c];
     }
-    for(pos = 0;pos < BufferSize;pos++)
+    for(c = 0;c < MaxChannels;c++)
     {
-        for(c = 0;c < MaxChannels;c++)
-            DryBuffer[OutPos][c] += data[pos]*DrySend[c];
-        OutPos++;
+        for(pos = 0;pos < BufferSize;pos++)
+            DryBuffer[c][OutPos+pos] += data[pos]*DrySend[c];
     }
-    if(OutPos == SamplesToDo)
+    if(OutPos+pos == SamplesToDo)
     {
         for(c = 0;c < MaxChannels;c++)
             PendingClicks[c] += data[pos]*DrySend[c];
diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c
index 434857c1..88d2c1ee 100644
--- a/Alc/mixer_sse.c
+++ b/Alc/mixer_sse.c
@@ -132,10 +132,9 @@ void MixDirect_SSE(ALsource *Source, ALCdevice *Device, DirectParams *params,
   const ALfloat *RESTRICT data, ALuint srcchan,
   ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
 {
-    ALfloat (*RESTRICT DryBuffer)[MaxChannels];
+    ALfloat (*RESTRICT DryBuffer)[BUFFERSIZE];
     ALfloat *RESTRICT ClickRemoval, *RESTRICT PendingClicks;
     ALfloat DrySend[MaxChannels];
-    ALIGN(16) ALfloat value[4];
     ALuint pos;
     ALuint c;
     (void)Source;
@@ -153,38 +152,28 @@ void MixDirect_SSE(ALsource *Source, ALCdevice *Device, DirectParams *params,
         for(c = 0;c < MaxChannels;c++)
             ClickRemoval[c] -= data[pos]*DrySend[c];
     }
-    for(pos = 0;pos < BufferSize-3;pos += 4)
+    for(c = 0;c < MaxChannels;c++)
     {
-        const __m128 val4 = _mm_load_ps(&data[pos]);
-        for(c = 0;c < MaxChannels;c++)
+        const __m128 gain = _mm_set1_ps(DrySend[c]);
+        for(pos = 0;pos < BufferSize-3;pos += 4)
         {
-            const __m128 gain = _mm_set1_ps(DrySend[c]);
-            __m128 dry4;
-
-            value[0] = DryBuffer[OutPos  ][c];
-            value[1] = DryBuffer[OutPos+1][c];
-            value[2] = DryBuffer[OutPos+2][c];
-            value[3] = DryBuffer[OutPos+3][c];
-            dry4 = _mm_load_ps(value);
-
+            const __m128 val4 = _mm_load_ps(&data[pos]);
+            __m128 dry4 = _mm_load_ps(&DryBuffer[c][OutPos+pos]);
             dry4 = _mm_add_ps(dry4, _mm_mul_ps(val4, gain));
-
-            _mm_store_ps(value, dry4);
-            DryBuffer[OutPos  ][c] = value[0];
-            DryBuffer[OutPos+1][c] = value[1];
-            DryBuffer[OutPos+2][c] = value[2];
-            DryBuffer[OutPos+3][c] = value[3];
+            _mm_store_ps(&DryBuffer[c][OutPos+pos], dry4);
         }
-
-        OutPos += 4;
     }
-    for(;pos < BufferSize;pos++)
+    if(pos < BufferSize)
     {
+        ALuint oldpos = pos;
         for(c = 0;c < MaxChannels;c++)
-            DryBuffer[OutPos][c] += data[pos]*DrySend[c];
-        OutPos++;
+        {
+            pos = oldpos;
+            for(;pos < BufferSize;pos++)
+                DryBuffer[c][OutPos+pos] += data[pos]*DrySend[c];
+        }
     }
-    if(OutPos == SamplesToDo)
+    if(OutPos+pos == SamplesToDo)
     {
         for(c = 0;c < MaxChannels;c++)
             PendingClicks[c] += data[pos]*DrySend[c];
diff --git a/OpenAL32/Include/alAuxEffectSlot.h b/OpenAL32/Include/alAuxEffectSlot.h
index dd8d3d76..d1b00e37 100644
--- a/OpenAL32/Include/alAuxEffectSlot.h
+++ b/OpenAL32/Include/alAuxEffectSlot.h
@@ -40,7 +40,7 @@ struct ALeffectState {
     ALvoid (*Destroy)(ALeffectState *State);
     ALboolean (*DeviceUpdate)(ALeffectState *State, ALCdevice *Device);
     ALvoid (*Update)(ALeffectState *State, ALCdevice *Device, const ALeffectslot *Slot);
-    ALvoid (*Process)(ALeffectState *State, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[MaxChannels]);
+    ALvoid (*Process)(ALeffectState *State, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[BUFFERSIZE]);
 };
 
 ALeffectState *NoneCreate(void);
diff --git a/OpenAL32/Include/alMain.h b/OpenAL32/Include/alMain.h
index e32e2313..6b8d5fab 100644
--- a/OpenAL32/Include/alMain.h
+++ b/OpenAL32/Include/alMain.h
@@ -618,7 +618,7 @@ struct ALCdevice_struct
     ALuint  NumChan;
 
     // Dry path buffer mix
-    ALIGN(16) ALfloat DryBuffer[BUFFERSIZE][MaxChannels];
+    ALIGN(16) ALfloat DryBuffer[MaxChannels][BUFFERSIZE];
 
     ALIGN(16) ALfloat ClickRemoval[MaxChannels];
     ALIGN(16) ALfloat PendingClicks[MaxChannels];
diff --git a/OpenAL32/alAuxEffectSlot.c b/OpenAL32/alAuxEffectSlot.c
index ab884cf7..048248aa 100644
--- a/OpenAL32/alAuxEffectSlot.c
+++ b/OpenAL32/alAuxEffectSlot.c
@@ -405,7 +405,7 @@ static ALvoid NoneUpdate(ALeffectState *State, ALCdevice *Device, const ALeffect
     (void)Device;
     (void)Slot;
 }
-static ALvoid NoneProcess(ALeffectState *State, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[MaxChannels])
+static ALvoid NoneProcess(ALeffectState *State, ALuint SamplesToDo, const ALfloat *SamplesIn, ALfloat (*SamplesOut)[BUFFERSIZE])
 {
     (void)State;
     (void)SamplesToDo;