Do the filtering separately from the mixing

4 files changed, 43 insertions, 59 deletions

diff --git a/Alc/mixer.c b/Alc/mixer.c
index a7882c23..41eba4ae 100644
--- a/Alc/mixer.c
+++ b/Alc/mixer.c
@@ -191,6 +191,16 @@ static void Resample(enum Resampler Resampler, const ALfloat *data, ALuint frac,
 }
 
 
+static void Filter2P(FILTER *filter, ALuint chan, ALfloat *RESTRICT dst,
+                     const ALfloat *RESTRICT src, ALuint numsamples)
+{
+    ALuint i;
+    for(i = 0;i < numsamples;i++)
+        dst[i] = lpFilter2P(filter, chan, src[i]);
+    dst[i] = lpFilter2PC(filter, chan, src[i]);
+}
+
+
 ALvoid MixSource(ALsource *Source, ALCdevice *Device, ALuint SamplesToDo)
 {
     ALbufferlistitem *BufferListItem;
@@ -432,20 +442,29 @@ ALvoid MixSource(ALsource *Source, ALCdevice *Device, ALuint SamplesToDo)
         SrcData += BufferPrePadding*NumChannels;
         for(i = 0;i < NumChannels;i++)
         {
+            DirectParams *directparms = &Source->Params.Direct;
+            ALIGN(16) ALfloat FilteredData[BUFFERSIZE];
             ALfloat ResampledData[BUFFERSIZE];
 
             Resample(Resampler, SrcData+i, DataPosFrac, increment,
                      NumChannels, ResampledData, BufferSize);
 
-            Source->Params.DryMix(Source, Device, &Source->Params.Direct,
-                                  ResampledData, i, OutPos, SamplesToDo,
+            Filter2P(&directparms->iirFilter, i, FilteredData, ResampledData,
+                     BufferSize);
+            Source->Params.DryMix(Source, Device, directparms,
+                                  FilteredData, i, OutPos, SamplesToDo,
                                   BufferSize);
+
             for(j = 0;j < Device->NumAuxSends;j++)
             {
-                if(!Source->Params.Send[j].Slot)
+                SendParams *sendparms = &Source->Params.Send[j];
+                if(!sendparms->Slot)
                     continue;
-                Source->Params.WetMix(&Source->Params.Send[j], ResampledData, i,
-                                      OutPos, SamplesToDo, BufferSize);
+
+                Filter2P(&sendparms->iirFilter, i, FilteredData, ResampledData,
+                         BufferSize);
+                Source->Params.WetMix(sendparms, FilteredData, OutPos,
+                                      SamplesToDo, BufferSize);
             }
         }
         for(i = 0;i < BufferSize;i++)
diff --git a/Alc/mixer_defs.h b/Alc/mixer_defs.h
index fb3d3d8f..7e930a39 100644
--- a/Alc/mixer_defs.h
+++ b/Alc/mixer_defs.h
@@ -13,16 +13,16 @@ struct SendParams;
 /* C mixers */
 void MixDirect_Hrtf_C(struct ALsource*,ALCdevice*,struct DirectParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint,ALuint);
 void MixDirect_C(struct ALsource*,ALCdevice*,struct DirectParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint,ALuint);
-void MixSend_C(struct SendParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint,ALuint);
+void MixSend_C(struct SendParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint);
 
 /* SSE mixers */
 void MixDirect_Hrtf_SSE(struct ALsource*,ALCdevice*,struct DirectParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint,ALuint);
 void MixDirect_SSE(struct ALsource*,ALCdevice*,struct DirectParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint,ALuint);
-void MixSend_SSE(struct SendParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint,ALuint);
+void MixSend_SSE(struct SendParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint);
 
 /* Neon mixers */
 void MixDirect_Hrtf_Neon(struct ALsource*,ALCdevice*,struct DirectParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint,ALuint);
 void MixDirect_Neon(struct ALsource*,ALCdevice*,struct DirectParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint,ALuint);
-void MixSend_Neon(struct SendParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint,ALuint);
+void MixSend_Neon(struct SendParams*,const ALfloat*RESTRICT,ALuint,ALuint,ALuint);
 
 #endif /* MIXER_DEFS_H */
diff --git a/Alc/mixer_inc.c b/Alc/mixer_inc.c
index c3df342e..f60ade65 100644
--- a/Alc/mixer_inc.c
+++ b/Alc/mixer_inc.c
@@ -53,15 +53,12 @@ void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
     ALIGN(16) ALfloat Coeffs[HRIR_LENGTH][2];
     ALuint Delay[2];
     ALfloat left, right;
-    FILTER *DryFilter;
-    ALfloat value;
     ALuint pos;
     ALuint c;
 
     DryBuffer = Device->DryBuffer;
     ClickRemoval = Device->ClickRemoval;
     PendingClicks = Device->PendingClicks;
-    DryFilter = &params->iirFilter;
 
     pos = 0;
     for(c = 0;c < IrSize;c++)
@@ -75,9 +72,7 @@ void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
 
     if(LIKELY(OutPos == 0))
     {
-        value = lpFilter2PC(DryFilter, srcchan, data[pos]);
-
-        History[Offset&SRC_HISTORY_MASK] = value;
+        History[Offset&SRC_HISTORY_MASK] = data[pos];
         left  = lerp(History[(Offset-(Delay[0]>>HRTFDELAY_BITS))&SRC_HISTORY_MASK],
                      History[(Offset-(Delay[0]>>HRTFDELAY_BITS)-1)&SRC_HISTORY_MASK],
                      (Delay[0]&HRTFDELAY_MASK)*(1.0f/HRTFDELAY_FRACONE));
@@ -92,9 +87,7 @@ void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
     }
     for(pos = 0;pos < BufferSize && Counter > 0;pos++)
     {
-        value = lpFilter2P(DryFilter, srcchan, data[pos]);
-
-        History[Offset&SRC_HISTORY_MASK] = value;
+        History[Offset&SRC_HISTORY_MASK] = data[pos];
         left  = lerp(History[(Offset-(Delay[0]>>HRTFDELAY_BITS))&SRC_HISTORY_MASK],
                      History[(Offset-(Delay[0]>>HRTFDELAY_BITS)-1)&SRC_HISTORY_MASK],
                      (Delay[0]&HRTFDELAY_MASK)*(1.0f/HRTFDELAY_FRACONE));
@@ -121,9 +114,7 @@ void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
     Delay[1] >>= HRTFDELAY_BITS;
     for(;pos < BufferSize;pos++)
     {
-        value = lpFilter2P(DryFilter, srcchan, data[pos]);
-
-        History[Offset&SRC_HISTORY_MASK] = value;
+        History[Offset&SRC_HISTORY_MASK] = data[pos];
         left = History[(Offset-Delay[0])&SRC_HISTORY_MASK];
         right = History[(Offset-Delay[1])&SRC_HISTORY_MASK];
 
@@ -139,9 +130,7 @@ void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
     }
     if(LIKELY(OutPos == SamplesToDo))
     {
-        value = lpFilter2PC(DryFilter, srcchan, data[pos]);
-
-        History[Offset&SRC_HISTORY_MASK] = value;
+        History[Offset&SRC_HISTORY_MASK] = data[pos];
         left = History[(Offset-Delay[0])&SRC_HISTORY_MASK];
         right = History[(Offset-Delay[1])&SRC_HISTORY_MASK];
 
@@ -161,16 +150,13 @@ void MixDirect(ALsource *Source, ALCdevice *Device, DirectParams *params,
     ALfloat (*RESTRICT DryBuffer)[MaxChannels];
     ALfloat *RESTRICT ClickRemoval, *RESTRICT PendingClicks;
     ALIGN(16) ALfloat DrySend[MaxChannels];
-    FILTER *DryFilter;
     ALuint pos;
-    ALfloat value;
     ALuint c;
     (void)Source;
 
     DryBuffer = Device->DryBuffer;
     ClickRemoval = Device->ClickRemoval;
     PendingClicks = Device->PendingClicks;
-    DryFilter = &params->iirFilter;
 
     for(c = 0;c < MaxChannels;c++)
         DrySend[c] = params->Gains[srcchan][c];
@@ -178,28 +164,25 @@ void MixDirect(ALsource *Source, ALCdevice *Device, DirectParams *params,
     pos = 0;
     if(OutPos == 0)
     {
-        value = lpFilter2PC(DryFilter, srcchan, data[pos]);
         for(c = 0;c < MaxChannels;c++)
-            ClickRemoval[c] -= value*DrySend[c];
+            ClickRemoval[c] -= data[pos]*DrySend[c];
     }
     for(pos = 0;pos < BufferSize;pos++)
     {
-        value = lpFilter2P(DryFilter, srcchan, data[pos]);
         for(c = 0;c < MaxChannels;c++)
-            DryBuffer[OutPos][c] += value*DrySend[c];
+            DryBuffer[OutPos][c] += data[pos]*DrySend[c];
         OutPos++;
     }
     if(OutPos == SamplesToDo)
     {
-        value = lpFilter2PC(DryFilter, srcchan, data[pos]);
         for(c = 0;c < MaxChannels;c++)
-            PendingClicks[c] += value*DrySend[c];
+            PendingClicks[c] += data[pos]*DrySend[c];
     }
 }
 #endif
 
 #ifndef NO_MIXSEND
-void MixSend(SendParams *params, const ALfloat *RESTRICT data, ALuint srcchan,
+void MixSend(SendParams *params, const ALfloat *RESTRICT data,
   ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
 {
     ALeffectslot *Slot;
@@ -207,33 +190,27 @@ void MixSend(SendParams *params, const ALfloat *RESTRICT data, ALuint srcchan,
     ALfloat *WetBuffer;
     ALfloat *WetClickRemoval;
     ALfloat *WetPendingClicks;
-    FILTER  *WetFilter;
     ALuint pos;
-    ALfloat value;
 
     Slot = params->Slot;
     WetBuffer = Slot->WetBuffer;
     WetClickRemoval = Slot->ClickRemoval;
     WetPendingClicks = Slot->PendingClicks;
-    WetFilter = &params->iirFilter;
     WetSend = params->Gain;
 
     pos = 0;
     if(OutPos == 0)
     {
-        value = lpFilter2PC(WetFilter, srcchan, data[pos]);
-        WetClickRemoval[0] -= value * WetSend;
+        WetClickRemoval[0] -= data[pos] * WetSend;
     }
     for(pos = 0;pos < BufferSize;pos++)
     {
-        value = lpFilter2P(WetFilter, srcchan, data[pos]);
-        WetBuffer[OutPos] += value * WetSend;
+        WetBuffer[OutPos] += data[pos] * WetSend;
         OutPos++;
     }
     if(OutPos == SamplesToDo)
     {
-        value = lpFilter2PC(WetFilter, srcchan, data[pos]);
-        WetPendingClicks[0] += value * WetSend;
+        WetPendingClicks[0] += data[pos] * WetSend;
     }
 }
 #endif
diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c
index 27b550e0..434857c1 100644
--- a/Alc/mixer_sse.c
+++ b/Alc/mixer_sse.c
@@ -134,9 +134,8 @@ void MixDirect_SSE(ALsource *Source, ALCdevice *Device, DirectParams *params,
 {
     ALfloat (*RESTRICT DryBuffer)[MaxChannels];
     ALfloat *RESTRICT ClickRemoval, *RESTRICT PendingClicks;
-    ALIGN(16) ALfloat DrySend[MaxChannels];
+    ALfloat DrySend[MaxChannels];
     ALIGN(16) ALfloat value[4];
-    FILTER *DryFilter;
     ALuint pos;
     ALuint c;
     (void)Source;
@@ -144,7 +143,6 @@ void MixDirect_SSE(ALsource *Source, ALCdevice *Device, DirectParams *params,
     DryBuffer = Device->DryBuffer;
     ClickRemoval = Device->ClickRemoval;
     PendingClicks = Device->PendingClicks;
-    DryFilter = &params->iirFilter;
 
     for(c = 0;c < MaxChannels;c++)
         DrySend[c] = params->Gains[srcchan][c];
@@ -152,20 +150,12 @@ void MixDirect_SSE(ALsource *Source, ALCdevice *Device, DirectParams *params,
     pos = 0;
     if(OutPos == 0)
     {
-        value[0] = lpFilter2PC(DryFilter, srcchan, data[pos]);
         for(c = 0;c < MaxChannels;c++)
-            ClickRemoval[c] -= value[0]*DrySend[c];
+            ClickRemoval[c] -= data[pos]*DrySend[c];
     }
     for(pos = 0;pos < BufferSize-3;pos += 4)
     {
-        __m128 val4;
-
-        value[0] = lpFilter2P(DryFilter, srcchan, data[pos  ]);
-        value[1] = lpFilter2P(DryFilter, srcchan, data[pos+1]);
-        value[2] = lpFilter2P(DryFilter, srcchan, data[pos+2]);
-        value[3] = lpFilter2P(DryFilter, srcchan, data[pos+3]);
-        val4 = _mm_load_ps(value);
-
+        const __m128 val4 = _mm_load_ps(&data[pos]);
         for(c = 0;c < MaxChannels;c++)
         {
             const __m128 gain = _mm_set1_ps(DrySend[c]);
@@ -190,16 +180,14 @@ void MixDirect_SSE(ALsource *Source, ALCdevice *Device, DirectParams *params,
     }
     for(;pos < BufferSize;pos++)
     {
-        value[0] = lpFilter2P(DryFilter, srcchan, data[pos]);
         for(c = 0;c < MaxChannels;c++)
-            DryBuffer[OutPos][c] += value[0]*DrySend[c];
+            DryBuffer[OutPos][c] += data[pos]*DrySend[c];
         OutPos++;
     }
     if(OutPos == SamplesToDo)
     {
-        value[0] = lpFilter2PC(DryFilter, srcchan, data[pos]);
         for(c = 0;c < MaxChannels;c++)
-            PendingClicks[c] += value[0]*DrySend[c];
+            PendingClicks[c] += data[pos]*DrySend[c];
     }
 }
 #define NO_MIXDIRECT