Use a more specialized mixer function for B-Format to HRTF

9 files changed, 66 insertions, 27 deletions

diff --git a/Alc/ALu.c b/Alc/ALu.c
index 41377eec..a9f89de0 100644
--- a/Alc/ALu.c
+++ b/Alc/ALu.c
@@ -94,18 +94,18 @@ extern inline void aluMatrixfSet(aluMatrixf *matrix,
                                  ALfloat m30, ALfloat m31, ALfloat m32, ALfloat m33);
 
 
-static inline HrtfMixerFunc SelectHrtfMixer(void)
+static inline HrtfDirectMixerFunc SelectHrtfMixer(void)
 {
 #ifdef HAVE_SSE
     if((CPUCapFlags&CPU_CAP_SSE))
-        return MixHrtf_SSE;
+        return MixDirectHrtf_SSE;
 #endif
 #ifdef HAVE_NEON
     if((CPUCapFlags&CPU_CAP_NEON))
-        return MixHrtf_Neon;
+        return MixDirectHrtf_Neon;
 #endif
 
-    return MixHrtf_C;
+    return MixDirectHrtf_C;
 }
 
 
@@ -1504,17 +1504,14 @@ ALvoid aluMixData(ALCdevice *device, ALvoid *buffer, ALsizei size)
             int ridx = GetChannelIdxByName(device->RealOut, FrontRight);
             if(lidx != -1 && ridx != -1)
             {
-                HrtfMixerFunc HrtfMix = SelectHrtfMixer();
+                HrtfDirectMixerFunc HrtfMix = SelectHrtfMixer();
                 ALuint irsize = device->Hrtf_IrSize;
-                MixHrtfParams hrtfparams;
-                memset(&hrtfparams, 0, sizeof(hrtfparams));
                 for(c = 0;c < device->Dry.NumChannels;c++)
                 {
-                    hrtfparams.Current = &device->Hrtf_Params[c];
-                    hrtfparams.Target = &device->Hrtf_Params[c];
                     HrtfMix(device->RealOut.Buffer, lidx, ridx,
-                        device->Dry.Buffer[c], 0, device->Hrtf_Offset, 0,
-                        irsize, &hrtfparams, &device->Hrtf_State[c], SamplesToDo
+                        device->Dry.Buffer[c], device->Hrtf_Offset, irsize,
+                        device->Hrtf_Coeffs[c], device->Hrtf_Values[c],
+                        SamplesToDo
                     );
                 }
                 device->Hrtf_Offset += SamplesToDo;
diff --git a/Alc/mixer_c.c b/Alc/mixer_c.c
index 5b3e7b74..c79d6061 100644
--- a/Alc/mixer_c.c
+++ b/Alc/mixer_c.c
@@ -166,6 +166,7 @@ static inline void ApplyCoeffs(ALuint Offset, ALfloat (*restrict Values)[2],
 }
 
 #define MixHrtf MixHrtf_C
+#define MixDirectHrtf MixDirectHrtf_C
 #include "mixer_inc.c"
 #undef MixHrtf
 
diff --git a/Alc/mixer_defs.h b/Alc/mixer_defs.h
index 8b934c58..be88379b 100644
--- a/Alc/mixer_defs.h
+++ b/Alc/mixer_defs.h
@@ -25,6 +25,10 @@ void MixHrtf_C(ALfloat (*restrict OutBuffer)[BUFFERSIZE], ALuint lidx, ALuint ri
                const ALfloat *data, ALuint Counter, ALuint Offset, ALuint OutPos,
                const ALuint IrSize, const struct MixHrtfParams *hrtfparams,
                struct HrtfState *hrtfstate, ALuint BufferSize);
+void MixDirectHrtf_C(ALfloat (*restrict OutBuffer)[BUFFERSIZE], ALuint lidx, ALuint ridx,
+                     const ALfloat *data, ALuint Offset, const ALuint IrSize,
+                     ALfloat (*restrict Coeffs)[2], ALfloat (*restrict Values)[2],
+                     ALuint BufferSize);
 void Mix_C(const ALfloat *data, ALuint OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
            struct MixGains *Gains, ALuint Counter, ALuint OutPos, ALuint BufferSize);
 void MixRow_C(ALfloat *OutBuffer, const ALfloat *Mtx, ALfloat (*restrict data)[BUFFERSIZE],
@@ -35,6 +39,10 @@ void MixHrtf_SSE(ALfloat (*restrict OutBuffer)[BUFFERSIZE], ALuint lidx, ALuint
                  const ALfloat *data, ALuint Counter, ALuint Offset, ALuint OutPos,
                  const ALuint IrSize, const struct MixHrtfParams *hrtfparams,
                  struct HrtfState *hrtfstate, ALuint BufferSize);
+void MixDirectHrtf_SSE(ALfloat (*restrict OutBuffer)[BUFFERSIZE], ALuint lidx, ALuint ridx,
+                       const ALfloat *data, ALuint Offset, const ALuint IrSize,
+                       ALfloat (*restrict Coeffs)[2], ALfloat (*restrict Values)[2],
+                       ALuint BufferSize);
 void Mix_SSE(const ALfloat *data, ALuint OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
              struct MixGains *Gains, ALuint Counter, ALuint OutPos, ALuint BufferSize);
 void MixRow_SSE(ALfloat *OutBuffer, const ALfloat *Mtx, ALfloat (*restrict data)[BUFFERSIZE],
@@ -78,6 +86,10 @@ void MixHrtf_Neon(ALfloat (*restrict OutBuffer)[BUFFERSIZE], ALuint lidx, ALuint
                   const ALfloat *data, ALuint Counter, ALuint Offset, ALuint OutPos,
                   const ALuint IrSize, const struct MixHrtfParams *hrtfparams,
                   struct HrtfState *hrtfstate, ALuint BufferSize);
+void MixDirectHrtf_Neon(ALfloat (*restrict OutBuffer)[BUFFERSIZE], ALuint lidx, ALuint ridx,
+                        const ALfloat *data, ALuint Offset, const ALuint IrSize,
+                        ALfloat (*restrict Coeffs)[2], ALfloat (*restrict Values)[2],
+                        ALuint BufferSize);
 void Mix_Neon(const ALfloat *data, ALuint OutChans, ALfloat (*restrict OutBuffer)[BUFFERSIZE],
               struct MixGains *Gains, ALuint Counter, ALuint OutPos, ALuint BufferSize);
 void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Mtx, ALfloat (*restrict data)[BUFFERSIZE],
diff --git a/Alc/mixer_inc.c b/Alc/mixer_inc.c
index e9d81c2a..25dc2b58 100644
--- a/Alc/mixer_inc.c
+++ b/Alc/mixer_inc.c
@@ -114,3 +114,36 @@ skip_stepping:
         OutPos += todo;
     }
 }
+
+void MixDirectHrtf(ALfloat (*restrict OutBuffer)[BUFFERSIZE], ALuint lidx, ALuint ridx,
+                   const ALfloat *data, ALuint Offset, const ALuint IrSize,
+                   ALfloat (*restrict Coeffs)[2], ALfloat (*restrict Values)[2],
+                   ALuint BufferSize)
+{
+    ALfloat out[MAX_UPDATE_SAMPLES][2];
+    ALfloat insample;
+    ALuint pos, i;
+
+    for(pos = 0;pos < BufferSize;)
+    {
+        ALuint todo = minu(BufferSize-pos, MAX_UPDATE_SAMPLES);
+
+        for(i = 0;i < todo;i++)
+        {
+            Values[(Offset+IrSize)&HRIR_MASK][0] = 0.0f;
+            Values[(Offset+IrSize)&HRIR_MASK][1] = 0.0f;
+            Offset++;
+
+            insample = *(data++);
+            ApplyCoeffs(Offset, Values, IrSize, Coeffs, insample, insample);
+            out[i][0] = Values[Offset&HRIR_MASK][0];
+            out[i][1] = Values[Offset&HRIR_MASK][1];
+        }
+
+        for(i = 0;i < todo;i++)
+            OutBuffer[lidx][pos+i] += out[i][0];
+        for(i = 0;i < todo;i++)
+            OutBuffer[ridx][pos+i] += out[i][1];
+        pos += todo;
+    }
+}
diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 48e41703..fcd0387a 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -70,6 +70,7 @@ static inline void ApplyCoeffs(ALuint Offset, ALfloat (*restrict Values)[2],
 }
 
 #define MixHrtf MixHrtf_Neon
+#define MixDirectHrtf MixDirectHrtf_Neon
 #include "mixer_inc.c"
 #undef MixHrtf
 
diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c
index 120ac4a0..6914e500 100644
--- a/Alc/mixer_sse.c
+++ b/Alc/mixer_sse.c
@@ -186,6 +186,7 @@ static inline void ApplyCoeffs(ALuint Offset, ALfloat (*restrict Values)[2],
 }
 
 #define MixHrtf MixHrtf_SSE
+#define MixDirectHrtf MixDirectHrtf_SSE
 #include "mixer_inc.c"
 #undef MixHrtf
 
diff --git a/Alc/panning.c b/Alc/panning.c
index 65506303..70494547 100644
--- a/Alc/panning.c
+++ b/Alc/panning.c
@@ -758,9 +758,8 @@ static void InitHQPanning(ALCdevice *device, const AmbDecConf *conf, const ALuin
 
 static void InitHrtfPanning(ALCdevice *device)
 {
-    ALfloat hrtf_coeffs[4][HRIR_LENGTH][2];
     size_t count = 4;
-    ALuint i, j;
+    ALuint i;
 
     for(i = 0;i < count;i++)
     {
@@ -773,21 +772,12 @@ static void InitHrtfPanning(ALCdevice *device)
     device->FOAOut.Ambi = device->Dry.Ambi;
     device->FOAOut.CoeffCount = device->Dry.CoeffCount;
 
-    memset(hrtf_coeffs, 0, sizeof(hrtf_coeffs));
-    device->Hrtf_IrSize = BuildBFormatHrtf(device->Hrtf, hrtf_coeffs, device->Dry.NumChannels);
+    memset(device->Hrtf_Coeffs, 0, sizeof(device->Hrtf_Coeffs));
+    device->Hrtf_IrSize = BuildBFormatHrtf(device->Hrtf, device->Hrtf_Coeffs,
+                                           device->Dry.NumChannels);
 
     /* Round up to the nearest multiple of 8 */
     device->Hrtf_IrSize = (device->Hrtf_IrSize+7)&~7;
-    for(i = 0;i < device->Dry.NumChannels;i++)
-    {
-        for(j = 0;j < HRIR_LENGTH;j++)
-        {
-            device->Hrtf_Params[i].Coeffs[j][0] = hrtf_coeffs[i][j][0];
-            device->Hrtf_Params[i].Coeffs[j][1] = hrtf_coeffs[i][j][1];
-        }
-        device->Hrtf_Params[i].Delay[0] = 0;
-        device->Hrtf_Params[i].Delay[1] = 0;
-    }
 }
 
 static void InitUhjPanning(ALCdevice *device)
diff --git a/OpenAL32/Include/alMain.h b/OpenAL32/Include/alMain.h
index 3d099ca2..b173c515 100644
--- a/OpenAL32/Include/alMain.h
+++ b/OpenAL32/Include/alMain.h
@@ -599,8 +599,8 @@ struct ALCdevice_struct
     ALCenum Hrtf_Status;
 
     /* HRTF filter state for dry buffer content */
-    HrtfState Hrtf_State[4];
-    HrtfParams Hrtf_Params[4];
+    alignas(16) ALfloat Hrtf_Values[4][HRIR_LENGTH][2];
+    alignas(16) ALfloat Hrtf_Coeffs[4][HRIR_LENGTH][2];
     ALuint Hrtf_Offset;
     ALuint Hrtf_IrSize;
 
diff --git a/OpenAL32/Include/alu.h b/OpenAL32/Include/alu.h
index 9913a117..1d922881 100644
--- a/OpenAL32/Include/alu.h
+++ b/OpenAL32/Include/alu.h
@@ -167,6 +167,10 @@ typedef void (*HrtfMixerFunc)(ALfloat (*restrict OutBuffer)[BUFFERSIZE], ALuint
                               const ALfloat *data, ALuint Counter, ALuint Offset, ALuint OutPos,
                               const ALuint IrSize, const MixHrtfParams *hrtfparams,
                               HrtfState *hrtfstate, ALuint BufferSize);
+typedef void (*HrtfDirectMixerFunc)(ALfloat (*restrict OutBuffer)[BUFFERSIZE],
+                                    ALuint lidx, ALuint ridx, const ALfloat *data, ALuint Offset,
+                                    const ALuint IrSize, ALfloat (*restrict Coeffs)[2],
+                                    ALfloat (*restrict Values)[2], ALuint BufferSize);
 
 
 #define GAIN_SILENCE_THRESHOLD  (0.00001f) /* -100dB */