Attempt to restore the Neon-enhanced ApplyCoeffsStep method

Unable to test, but it hopefully works.

1 files changed, 21 insertions, 7 deletions

diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 65e702eb..dd7c1c80 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -8,6 +8,7 @@
 #include "AL/alc.h"
 #include "alMain.h"
 #include "alu.h"
+#include "hrtf.h"
 
 
 static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
@@ -16,16 +17,29 @@ static inline void ApplyCoeffsStep(ALuint Offset, ALfloat (*restrict Values)[2],
                                    const ALfloat (*restrict CoeffStep)[2],
                                    ALfloat left, ALfloat right)
 {
-    float32x4_t coeffs, deltas;
     ALuint c;
-
+    float32x4_t leftright4;
+    {
+        float32x2_t leftright2 = vdup_n_f32(0.0);
+        leftright2 = vset_lane_f32(left, leftright2, 0);
+        leftright2 = vset_lane_f32(right, leftright2, 1);
+        leftright4 = vcombine_f32(leftright2, leftright2);
+    }
     for(c = 0;c < IrSize;c += 2)
     {
-        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];
+        const ALuint o0 = (Offset+c)&HRIR_MASK;
+        const ALuint o1 = (o0+1)&HRIR_MASK;
+        float32x4_t vals = vcombine_f32(vld1_f32((float32_t*)&Values[o0][0]),
+                                        vld1_f32((float32_t*)&Values[o1][0]));
+        float32x4_t coefs = vld1q_f32((float32_t*)&Coeffs[c][0]);
+        float32x4_t deltas = vld1q_f32(&CoeffStep[c][0]);
+
+        vals = vmlaq_f32(vals, coefs, leftright4);
+        coefs = vaddq_f32(coefs, deltas);
+
+        vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
+        vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
+        vst1q_f32(&Coeffs[c][0], coefs);
     }
 }