1 files changed, 77 insertions, 0 deletions
diff --git a/core/mixer/mixer_neon.cpp b/core/mixer/mixer_neon.cpp
index a3468926..f3d54fec 100644
--- a/core/mixer/mixer_neon.cpp
+++ b/core/mixer/mixer_neon.cpp
@@ -305,3 +305,80 @@ void Mix_<NEONTag>(const al::span<const float> InSamples, const al::span<FloatBu
             dst[pos] += InSamples[pos] * gain;
     }
 }
+
+template<>
+void Mix_<NEONTag>(const al::span<const float> InSamples, float *OutBuffer, float &CurrentGain,
+    const float TargetGain, const size_t Counter)
+{
+    const float delta{(Counter > 0) ? 1.0f / static_cast<float>(Counter) : 0.0f};
+    const auto min_len = minz(Counter, InSamples.size());
+    const auto aligned_len = minz((min_len+3) & ~size_t{3}, InSamples.size()) - min_len;
+
+    float *RESTRICT dst{al::assume_aligned<16>(OutBuffer)};
+    float gain{CurrentGain};
+    const float step{(TargetGain-gain) * delta};
+
+    size_t pos{0};
+    if(!(std::abs(step) > std::numeric_limits<float>::epsilon()))
+        gain = TargetGain;
+    else
+    {
+        float step_count{0.0f};
+        /* Mix with applying gain steps in aligned multiples of 4. */
+        if(size_t todo{min_len >> 2})
+        {
+            const float32x4_t four4{vdupq_n_f32(4.0f)};
+            const float32x4_t step4{vdupq_n_f32(step)};
+            const float32x4_t gain4{vdupq_n_f32(gain)};
+            float32x4_t step_count4{vdupq_n_f32(0.0f)};
+            step_count4 = vsetq_lane_f32(1.0f, step_count4, 1);
+            step_count4 = vsetq_lane_f32(2.0f, step_count4, 2);
+            step_count4 = vsetq_lane_f32(3.0f, step_count4, 3);
+
+            do {
+                const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
+                float32x4_t dry4 = vld1q_f32(&dst[pos]);
+                dry4 = vmlaq_f32(dry4, val4, vmlaq_f32(gain4, step4, step_count4));
+                step_count4 = vaddq_f32(step_count4, four4);
+                vst1q_f32(&dst[pos], dry4);
+                pos += 4;
+            } while(--todo);
+            /* NOTE: step_count4 now represents the next four counts after the
+             * last four mixed samples, so the lowest element represents the
+             * next step count to apply.
+             */
+            step_count = vgetq_lane_f32(step_count4, 0);
+        }
+        /* Mix with applying left over gain steps that aren't aligned multiples of 4. */
+        for(size_t leftover{min_len&3};leftover;++pos,--leftover)
+        {
+            dst[pos] += InSamples[pos] * (gain + step*step_count);
+            step_count += 1.0f;
+        }
+        if(pos == Counter)
+            gain = TargetGain;
+        else
+            gain += step*step_count;
+
+        /* Mix until pos is aligned with 4 or the mix is done. */
+        for(size_t leftover{aligned_len&3};leftover;++pos,--leftover)
+            dst[pos] += InSamples[pos] * gain;
+    }
+    CurrentGain = gain;
+
+    if(!(std::abs(gain) > GainSilenceThreshold))
+        return;
+    if(size_t todo{(InSamples.size()-pos) >> 2})
+    {
+        const float32x4_t gain4 = vdupq_n_f32(gain);
+        do {
+            const float32x4_t val4 = vld1q_f32(&InSamples[pos]);
+            float32x4_t dry4 = vld1q_f32(&dst[pos]);
+            dry4 = vmlaq_f32(dry4, val4, gain4);
+            vst1q_f32(&dst[pos], dry4);
+            pos += 4;
+        } while(--todo);
+    }
+    for(size_t leftover{(InSamples.size()-pos)&3};leftover;++pos,--leftover)
+        dst[pos] += InSamples[pos] * gain;
+}