Use a separate method to set initial HRTF coefficients

4 files changed, 52 insertions, 6 deletions

diff --git a/Alc/mixer_c.c b/Alc/mixer_c.c
index 0fdcc087..fa87c3d4 100644
--- a/Alc/mixer_c.c
+++ b/Alc/mixer_c.c
@@ -59,6 +59,18 @@ void ALfilterState_processC(ALfilterState *filter, ALfloat *restrict dst, const
 }
 
 
+static inline void SetupCoeffs(ALfloat (*restrict OutCoeffs)[2],
+                               const HrtfParams *hrtfparams,
+                               ALuint IrSize, ALuint Counter)
+{
+    ALuint c;
+    for(c = 0;c < IrSize;c++)
+    {
+        OutCoeffs[c][0] = hrtfparams->Coeffs[c][0] - (hrtfparams->CoeffStep[c][0]*Counter);
+        OutCoeffs[c][1] = hrtfparams->Coeffs[c][1] - (hrtfparams->CoeffStep[c][1]*Counter);
+    }
+}
+
 static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
                                    const ALuint IrSize,
                                    ALfloat (*restrict Coeffs)[2],
diff --git a/Alc/mixer_inc.c b/Alc/mixer_inc.c
index b4635b43..63003f03 100644
--- a/Alc/mixer_inc.c
+++ b/Alc/mixer_inc.c
@@ -14,6 +14,9 @@
 #define MixHrtf MERGE(MixHrtf_,SUFFIX)
 
 
+static inline void SetupCoeffs(ALfloat (*restrict OutCoeffs)[2],
+                               const HrtfParams *hrtfparams,
+                               ALuint IrSize, ALuint Counter);
 static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
                                    const ALuint irSize,
                                    ALfloat (*restrict Coeffs)[2],
@@ -33,13 +36,8 @@ void MixHrtf(ALfloat (*restrict OutBuffer)[BUFFERSIZE], const ALfloat *data,
     ALuint Delay[2];
     ALfloat left, right;
     ALuint pos;
-    ALuint c;
 
-    for(c = 0;c < IrSize;c++)
-    {
-        Coeffs[c][0] = hrtfparams->Coeffs[c][0] - (hrtfparams->CoeffStep[c][0]*Counter);
-        Coeffs[c][1] = hrtfparams->Coeffs[c][1] - (hrtfparams->CoeffStep[c][1]*Counter);
-    }
+    SetupCoeffs(Coeffs, hrtfparams, IrSize, Counter);
     Delay[0] = hrtfparams->Delay[0] - (hrtfparams->DelayStep[0]*Counter);
     Delay[1] = hrtfparams->Delay[1] - (hrtfparams->DelayStep[1]*Counter);
 
diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 8ce17644..a068d02b 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -9,6 +9,25 @@
 #include "hrtf.h"
 
 
+static inline void SetupCoeffs(ALfloat (*restrict OutCoeffs)[2],
+                               const HrtfParams *hrtfparams,
+                               ALuint IrSize, ALuint Counter)
+{
+    ALuint c;
+    float32x4_t counter4;
+    {
+        float32x2_t counter2 = vdup_n_f32(-(float)Counter);
+        counter4 = vcombine_f32(counter2, counter2);
+    }
+    for(c = 0;c < IrSize;c += 2)
+    {
+        float32x4_t step4 = vld1q_f32((float32_t*)hrtfparams->CoeffStep[c]);
+        float32x4_t coeffs = vld1q_f32((float32_t*)hrtfparams->Coeffs[c]);
+        coeffs = vmlaq_f32(coeffs, step4, counter4);
+        vst1q_f32((float32_t*)OutCoeffs[c], coeffs);
+    }
+}
+
 static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
                                    const ALuint IrSize,
                                    ALfloat (*restrict Coeffs)[2],
diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c
index d86cf749..75f50f88 100644
--- a/Alc/mixer_sse.c
+++ b/Alc/mixer_sse.c
@@ -19,6 +19,23 @@
 #include "mixer_defs.h"
 
 
+static inline void SetupCoeffs(ALfloat (*restrict OutCoeffs)[2],
+                               const HrtfParams *hrtfparams,
+                               ALuint IrSize, ALuint Counter)
+{
+    const __m128 counter4 = _mm_set1_ps((float)Counter);
+    __m128 coeffs, step4;
+    ALuint i;
+
+    for(i = 0;i < IrSize;i += 2)
+    {
+        step4  = _mm_load_ps(&hrtfparams->CoeffStep[i][0]);
+        coeffs = _mm_load_ps(&hrtfparams->Coeffs[i][0]);
+        coeffs = _mm_sub_ps(coeffs, _mm_mul_ps(step4, counter4));
+        _mm_store_ps(&OutCoeffs[i][0], coeffs);
+    }
+}
+
 static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
                                    const ALuint IrSize,
                                    ALfloat (*restrict Coeffs)[2],