Added ARM Neon SIMD optimizations for AEC3

webrtc/modules/audio_processing/aec3/matched_filter.cc

Index: webrtc/modules/audio_processing/aec3/matched_filter.cc
diff --git a/webrtc/modules/audio_processing/aec3/matched_filter.cc b/webrtc/modules/audio_processing/aec3/matched_filter.cc
index 7bb5778999879f91012641f82f72c491b1df11c4..4c6e0d70525ad9eba85cb52043d49b82751af999 100644
--- a/webrtc/modules/audio_processing/aec3/matched_filter.cc
+++ b/webrtc/modules/audio_processing/aec3/matched_filter.cc
@@ -9,6 +9,9 @@
  */
 #include "webrtc/modules/audio_processing/aec3/matched_filter.h"
 
+#if defined(WEBRTC_HAS_NEON)
+#include <arm_neon.h>
+#endif
 #include "webrtc/typedefs.h"
 #if defined(WEBRTC_ARCH_X86_FAMILY)
 #include <emmintrin.h>
@@ -22,6 +25,114 @@
 namespace webrtc {
 namespace aec3 {
 
+#if defined(WEBRTC_HAS_NEON)
+
+void MatchedFilterCore_NEON(size_t x_start_index,
+                            float x2_sum_threshold,
+                            rtc::ArrayView<const float> x,
+                            rtc::ArrayView<const float> y,
+                            rtc::ArrayView<float> h,
+                            bool* filters_updated,
+                            float* error_sum) {
+  const int h_size = static_cast<int>(h.size());
+  const int x_size = static_cast<int>(x.size());
+  RTC_DCHECK_EQ(0, h_size % 4);
+
+  // Process for all samples in the sub-block.
+  for (size_t i = 0; i < kSubBlockSize; ++i) {
+    // Apply the matched filter as filter * x, and compute x * x.
+
+    RTC_DCHECK_GT(x_size, x_start_index);
+    const float* x_p = &x[x_start_index];
+    const float* h_p = &h[0];
+
+    // Initialize values for the accumulation.
+    float32x4_t s_128 = vdupq_n_f32(0);
+    float32x4_t x2_sum_128 = vdupq_n_f32(0);
+    float x2_sum = 0.f;
+    float s = 0;
+
+    // Compute loop chunk sizes until, and after, the wraparound of the circular
+    // buffer for x.
+    const int chunk1 =
+        std::min(h_size, static_cast<int>(x_size - x_start_index));
+
+    // Perform the loop in two chunks.
+    const int chunk2 = h_size - chunk1;
+    for (int limit : {chunk1, chunk2}) {
+      // Perform 128 bit vector operations.
+      const int limit_by_4 = limit >> 2;
+      for (int k = limit_by_4; k > 0; --k, h_p += 4, x_p += 4) {
+        // Load the data into 128 bit vectors.
+        const float32x4_t x_k = vld1q_f32(x_p);
+        const float32x4_t h_k = vld1q_f32(h_p);
+        // Compute and accumulate x * x and h * x.
+        x2_sum_128 = vmlaq_f32(x2_sum_128, x_k, x_k);
+        s_128 = vmlaq_f32(s_128, h_k, x_k);
+      }
+
+      // Perform non-vector operations for any remaining items.
+      for (int k = limit - limit_by_4 * 4; k > 0; --k, ++h_p, ++x_p) {
+        const float x_k = *x_p;
+        x2_sum += x_k * x_k;
+        s += *h_p * x_k;
+      }
+
+      x_p = &x[0];
+    }
+
+    // Combine the accumulated vector and scalar values.
+    float* v = reinterpret_cast<float*>(&x2_sum_128);
+    x2_sum += v[0] + v[1] + v[2] + v[3];
+    v = reinterpret_cast<float*>(&s_128);
+    s += v[0] + v[1] + v[2] + v[3];
+
+    // Compute the matched filter error.
+    const float e = std::min(32767.f, std::max(-32768.f, y[i] - s));
+    *error_sum += e * e;
+
+    // Update the matched filter estimate in an NLMS manner.
+    if (x2_sum > x2_sum_threshold) {
+      RTC_DCHECK_LT(0.f, x2_sum);
+      const float alpha = 0.7f * e / x2_sum;
+      const float32x4_t alpha_128 = vmovq_n_f32(alpha);
+
+      // filter = filter + 0.7 * (y - filter * x) / x * x.
+      float* h_p = &h[0];
+      x_p = &x[x_start_index];
+
+      // Perform the loop in two chunks.
+      for (int limit : {chunk1, chunk2}) {
+        // Perform 128 bit vector operations.
+        const int limit_by_4 = limit >> 2;
+        for (int k = limit_by_4; k > 0; --k, h_p += 4, x_p += 4) {
+          // Load the data into 128 bit vectors.
+          float32x4_t h_k = vld1q_f32(h_p);
+          const float32x4_t x_k = vld1q_f32(x_p);
+          // Compute h = h + alpha * x.
+          h_k = vmlaq_f32(h_k, alpha_128, x_k);
+
+          // Store the result.
+          vst1q_f32(h_p, h_k);
+        }
+
+        // Perform non-vector operations for any remaining items.
+        for (int k = limit - limit_by_4 * 4; k > 0; --k, ++h_p, ++x_p) {
+          *h_p += alpha * *x_p;
+        }
+
+        x_p = &x[0];
+      }
+
+      *filters_updated = true;
+    }
+
+    x_start_index = x_start_index > 0 ? x_start_index - 1 : x_size - 1;
+  }
+}
+
+#endif
+
 #if defined(WEBRTC_ARCH_X86_FAMILY)
 
 void MatchedFilterCore_SSE2(size_t x_start_index,
@@ -227,6 +338,13 @@ void MatchedFilter::Update(const DownsampledRenderBuffer& render_buffer,
                                      &filters_updated, &error_sum);
         break;
 #endif
+#if defined(WEBRTC_HAS_NEON)
+      case Aec3Optimization::kNeon:
+        aec3::MatchedFilterCore_NEON(x_start_index, x2_sum_threshold,
+                                     render_buffer.buffer, y, filters_[n],
+                                     &filters_updated, &error_sum);
+        break;
+#endif
       default:
         aec3::MatchedFilterCore(x_start_index, x2_sum_threshold,
                                 render_buffer.buffer, y, filters_[n],