1 files changed, 50 insertions, 0 deletions
diff --git a/Alc/mixer_neon.c b/Alc/mixer_neon.c
new file mode 100644
index 00000000..10385e69
--- /dev/null
+++ b/Alc/mixer_neon.c
@@ -0,0 +1,50 @@
+#include "config.h"
+
+#ifdef HAVE_ARM_NEON_H
+#include <arm_neon.h>
+#endif
+
+#include "AL/al.h"
+#include "AL/alc.h"
+#include "alMain.h"
+#include "alu.h"
+
+
+static __inline void ApplyCoeffs(ALuint Offset, ALfloat (*RESTRICT Values)[2],
+                                 ALfloat (*RESTRICT Coeffs)[2],
+                                 ALfloat left, ALfloat right)
+{
+    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 < HRIR_LENGTH;c += 2)
+    {
+        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]);
+
+        vals = vmlaq_f32(vals, coefs, leftright4);
+
+        vst1_f32((float32_t*)&Values[o0][0], vget_low_f32(vals));
+        vst1_f32((float32_t*)&Values[o1][0], vget_high_f32(vals));
+    }
+}
+
+#define SUFFIX Neon
+#define SAMPLER point32
+#include "mixer_inc.c"
+#undef SAMPLER
+#define SAMPLER lerp32
+#include "mixer_inc.c"
+#undef SAMPLER
+#define SAMPLER cubic32
+#include "mixer_inc.c"
+#undef SAMPLER
+#undef SUFFIX