Implement MixDirect_SSE separately from the C and Neon versions

4 files changed, 58 insertions, 23 deletions

diff --git a/Alc/mixer_c.c b/Alc/mixer_c.c
index 5cb5acb4..aecf03a5 100644
--- a/Alc/mixer_c.c
+++ b/Alc/mixer_c.c
@@ -36,14 +36,6 @@ static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
 }
 
 
-static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value, const ALfloat *DrySend)
-{
-    ALuint c;
-    for(c = 0;c < MaxChannels;c++)
-        Output[c] += value*DrySend[c];
-}
-
-
 #define SUFFIX C
 #include "mixer_inc.c"
 #undef SUFFIX
diff --git a/Alc/mixer_inc.c b/Alc/mixer_inc.c
index ff12026b..359662cc 100644
--- a/Alc/mixer_inc.c
+++ b/Alc/mixer_inc.c
@@ -30,10 +30,9 @@ static __inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*RESTRICT Values)[2
 static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
                                  ALfloat (*RESTRICT Coeffs)[2],
                                  ALfloat left, ALfloat right);
-static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value,
-                                const ALfloat *DrySend);
 
 
+#ifndef NO_MIXDIRECT_HRTF
 void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
   const ALfloat *RESTRICT data, ALuint srcchan,
   ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
@@ -149,8 +148,9 @@ void MixDirect_Hrtf(ALsource *Source, ALCdevice *Device, DirectParams *params,
                                      Coeffs[0][1] * right;
     }
 }
+#endif
 
-
+#ifndef NO_MIXDIRECT
 void MixDirect(ALsource *Source, ALCdevice *Device, DirectParams *params,
   const ALfloat *RESTRICT data, ALuint srcchan,
   ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
@@ -176,22 +176,26 @@ void MixDirect(ALsource *Source, ALCdevice *Device, DirectParams *params,
     if(OutPos == 0)
     {
         value = lpFilter2PC(DryFilter, srcchan, data[pos]);
-        ApplyValue(ClickRemoval, -value, DrySend);
+        for(c = 0;c < MaxChannels;c++)
+            ClickRemoval[c] -= value*DrySend[c];
     }
     for(pos = 0;pos < BufferSize;pos++)
     {
         value = lpFilter2P(DryFilter, srcchan, data[pos]);
-        ApplyValue(DryBuffer[OutPos], value, DrySend);
+        for(c = 0;c < MaxChannels;c++)
+            DryBuffer[OutPos][c] += value*DrySend[c];
         OutPos++;
     }
     if(OutPos == SamplesToDo)
     {
         value = lpFilter2PC(DryFilter, srcchan, data[pos]);
-        ApplyValue(PendingClicks, value, DrySend);
+        for(c = 0;c < MaxChannels;c++)
+            PendingClicks[c] += value*DrySend[c];
     }
 }
+#endif
 
-
+#ifndef NO_MIXSEND
 void MixSend(SendParams *params, const ALfloat *RESTRICT data, ALuint srcchan,
   ALuint OutPos, ALuint SamplesToDo, ALuint BufferSize)
 {
@@ -229,6 +233,8 @@ void MixSend(SendParams *params, const ALfloat *RESTRICT data, ALuint srcchan,
         WetPendingClicks[0] += value * WetSend;
     }
 }
+#endif
+
 
 #undef MixSend
 #undef MixDirect
diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 4824c597..f7a9fed5 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -54,14 +54,6 @@ static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
 }
 
 
-static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value, const ALfloat *DrySend)
-{
-    ALuint c;
-    for(c = 0;c < MaxChannels;c++)
-        Output[c] += value*DrySend[c];
-}
-
-
 #define SUFFIX Neon
 #include "mixer_inc.c"
 #undef SUFFIX
diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c
index 2824beb4..101b296e 100644
--- a/Alc/mixer_sse.c
+++ b/Alc/mixer_sse.c
@@ -9,6 +9,9 @@
 #include "alMain.h"
 #include "alu.h"
 
+#include "alSource.h"
+#include "mixer_defs.h"
+
 
 static __inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*RESTRICT Values)[2],
                                      ALfloat (*RESTRICT Coeffs)[2],
@@ -76,6 +79,48 @@ static __inline void ApplyValue(ALfloat *RESTRICT Output, ALfloat value, const A
 }
 
 
+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 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];
+
+    pos = 0;
+    if(OutPos == 0)
+    {
+        value = lpFilter2PC(DryFilter, srcchan, data[pos]);
+        ApplyValue(ClickRemoval, -value, DrySend);
+    }
+    for(pos = 0;pos < BufferSize;pos++)
+    {
+        value = lpFilter2P(DryFilter, srcchan, data[pos]);
+        ApplyValue(DryBuffer[OutPos], value, DrySend);
+        OutPos++;
+    }
+    if(OutPos == SamplesToDo)
+    {
+        value = lpFilter2PC(DryFilter, srcchan, data[pos]);
+        ApplyValue(PendingClicks, value, DrySend);
+    }
+}
+#define NO_MIXDIRECT
+
+
 #define SUFFIX SSE
 #include "mixer_inc.c"
 #undef SUFFIX