Apply HRTF mixer coefficients with stepping using SSE

4 files changed, 65 insertions, 11 deletions

diff --git a/Alc/mixer_c.c b/Alc/mixer_c.c
index f59b3190..eeb75bf4 100644
--- a/Alc/mixer_c.c
+++ b/Alc/mixer_c.c
@@ -6,6 +6,22 @@
 #include "alu.h"
 
 
+static __inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*RESTRICT Values)[2],
+                                     ALfloat (*RESTRICT Coeffs)[2],
+                                     ALfloat (*RESTRICT CoeffStep)[2],
+                                     ALfloat left, ALfloat right)
+{
+    ALuint c;
+    for(c = 0;c < HRIR_LENGTH;c++)
+    {
+        const ALuint off = (Offset+c)&HRIR_MASK;
+        Values[off][0] += Coeffs[c][0] * left;
+        Values[off][1] += Coeffs[c][1] * right;
+        Coeffs[c][0] += CoeffStep[c][0];
+        Coeffs[c][1] += CoeffStep[c][1];
+    }
+}
+
 static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
                                  ALfloat (*RESTRICT Coeffs)[2],
                                  ALfloat left, ALfloat right)
diff --git a/Alc/mixer_inc.c b/Alc/mixer_inc.c
index b7051f64..3e9d7f46 100644
--- a/Alc/mixer_inc.c
+++ b/Alc/mixer_inc.c
@@ -107,15 +107,7 @@ void MERGE4(MixDirect_Hrtf_,SAMPLER,_,SUFFIX)(
             Values[Offset&HRIR_MASK][1] = 0.0f;
             Offset++;
 
-            for(c = 0;c < HRIR_LENGTH;c++)
-            {
-                const ALuint off = (Offset+c)&HRIR_MASK;
-                Values[off][0] += Coeffs[c][0] * left;
-                Values[off][1] += Coeffs[c][1] * right;
-                Coeffs[c][0] += CoeffStep[c][0];
-                Coeffs[c][1] += CoeffStep[c][1];
-            }
-
+            ApplyCoeffsStep(Offset, Values, Coeffs, CoeffStep, left, right);
             DryBuffer[OutPos][FrontLeft]  += Values[Offset&HRIR_MASK][0];
             DryBuffer[OutPos][FrontRight] += Values[Offset&HRIR_MASK][1];
 
diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 10385e69..b550b837 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -10,6 +10,22 @@
 #include "alu.h"
 
 
+static __inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*RESTRICT Values)[2],
+                                     ALfloat (*RESTRICT Coeffs)[2],
+                                     ALfloat (*RESTRICT CoeffStep)[2],
+                                     ALfloat left, ALfloat right)
+{
+    ALuint c;
+    for(c = 0;c < HRIR_LENGTH;c++)
+    {
+        const ALuint off = (Offset+c)&HRIR_MASK;
+        Values[off][0] += Coeffs[c][0] * left;
+        Values[off][1] += Coeffs[c][1] * right;
+        Coeffs[c][0] += CoeffStep[c][0];
+        Coeffs[c][1] += CoeffStep[c][1];
+    }
+}
+
 static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
                                  ALfloat (*RESTRICT Coeffs)[2],
                                  ALfloat left, ALfloat right)
diff --git a/Alc/mixer_sse.c b/Alc/mixer_sse.c
index 2da2b6b8..bbca54ec 100644
--- a/Alc/mixer_sse.c
+++ b/Alc/mixer_sse.c
@@ -10,18 +10,48 @@
 #include "alu.h"
 
 
+static __inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*RESTRICT Values)[2],
+                                     ALfloat (*RESTRICT Coeffs)[2],
+                                     ALfloat (*RESTRICT CoeffStep)[2],
+                                     ALfloat left, ALfloat right)
+{
+    const __m128 lrlr = { left, right, left, right };
+    __m128 vals = { 0.0f, 0.0f, 0.0f, 0.0f };
+    __m128 coeffs, coeffstep;
+    ALuint c;
+    for(c = 0;c < HRIR_LENGTH;c += 2)
+    {
+        const ALuint o0 = (Offset++)&HRIR_MASK;
+        const ALuint o1 = (Offset++)&HRIR_MASK;
+
+        coeffs = _mm_load_ps(&Coeffs[c][0]);
+        vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
+        vals = _mm_loadh_pi(vals, (__m64*)&Values[o1][0]);
+
+        vals = _mm_add_ps(vals, _mm_mul_ps(coeffs, lrlr));
+        _mm_storel_pi((__m64*)&Values[o0][0], vals);
+        _mm_storeh_pi((__m64*)&Values[o1][0], vals);
+
+        coeffstep = _mm_load_ps(&CoeffStep[c][0]);
+        coeffs = _mm_add_ps(coeffs, coeffstep);
+        _mm_store_ps(&Coeffs[c][0], coeffs);
+    }
+}
+
 static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
                                  ALfloat (*RESTRICT Coeffs)[2],
                                  ALfloat left, ALfloat right)
 {
     const __m128 lrlr = { left, right, left, right };
+    __m128 vals = { 0.0f, 0.0f, 0.0f, 0.0f };
+    __m128 coeffs;
     ALuint c;
     for(c = 0;c < HRIR_LENGTH;c += 2)
     {
         const ALuint o0 = (Offset++)&HRIR_MASK;
         const ALuint o1 = (Offset++)&HRIR_MASK;
-        __m128 coeffs = _mm_load_ps(&Coeffs[c][0]);
-        __m128 vals = { 0.0f, 0.0f, 0.0f, 0.0f };
+
+        coeffs = _mm_load_ps(&Coeffs[c][0]);
         vals = _mm_loadl_pi(vals, (__m64*)&Values[o0][0]);
         vals = _mm_loadh_pi(vals, (__m64*)&Values[o1][0]);