Implement a Neon-enhanced MixRow

1 files changed, 25 insertions, 0 deletions

diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
index 96936ef5..073f62c8 100644
--- a/Alc/mixer_neon.c
+++ b/Alc/mixer_neon.c
@@ -118,3 +118,28 @@ void Mix_Neon(const ALfloat *data, ALuint OutChans, ALfloat (*restrict OutBuffer
             OutBuffer[c][OutPos+pos] += data[pos]*gain;
     }
 }
+
+void MixRow_Neon(ALfloat *OutBuffer, const ALfloat *Mtx, ALfloat (*restrict data)[BUFFERSIZE], ALuint InChans, ALuint BufferSize)
+{
+    float32x4_t gain4;
+    ALuint c;
+
+    for(c = 0;c < InChans;c++)
+    {
+        ALuint pos = 0;
+        ALfloat gain = Mtx[c];
+        if(!(fabsf(gain) > GAIN_SILENCE_THRESHOLD))
+            continue;
+
+        gain4 = vdupq_n_f32(gain);
+        for(;BufferSize-pos > 3;pos += 4)
+        {
+            const float32x4_t val4 = vld1q_f32(&data[c][pos]);
+            float32x4_t dry4 = vld1q_f32(&OutBuffer[pos]);
+            dry4 = vmlaq_f32(dry4, val4, gain4);
+            vst1q_f32(&OutBuffer[pos], dry4);
+        }
+        for(;pos < BufferSize;pos++)
+            OutBuffer[pos] += data[c][pos]*gain;
+    }
+}