Reland of Added ARM Neon SIMD optimizations for AEC3

webrtc/modules/audio_processing/aec3/adaptive_fir_filter_unittest.cc

Index: webrtc/modules/audio_processing/aec3/adaptive_fir_filter_unittest.cc
diff --git a/webrtc/modules/audio_processing/aec3/adaptive_fir_filter_unittest.cc b/webrtc/modules/audio_processing/aec3/adaptive_fir_filter_unittest.cc
index 6d1a5820fe7f0e6c752d9e84cc212d3406a87c2a..4560958bfd0ef350e46d97a5a3ddcb34484627d0 100644
--- a/webrtc/modules/audio_processing/aec3/adaptive_fir_filter_unittest.cc
+++ b/webrtc/modules/audio_processing/aec3/adaptive_fir_filter_unittest.cc
@@ -10,6 +10,7 @@
 
 #include "webrtc/modules/audio_processing/aec3/adaptive_fir_filter.h"
 
+#include <math.h>
 #include <algorithm>
 #include <numeric>
 #include <string>
@@ -41,10 +42,114 @@
 
 }  // namespace
 
+#if defined(WEBRTC_HAS_NEON)
+// Verifies that the optimized methods for filter adaptation are similar to
+// their reference counterparts.
+TEST(AdaptiveFirFilter, FilterAdaptationNeonOptimizations) {
+  RenderBuffer render_buffer(Aec3Optimization::kNone, 3, 12,
+                             std::vector<size_t>(1, 12));
+  Random random_generator(42U);
+  std::vector<std::vector<float>> x(3, std::vector<float>(kBlockSize, 0.f));
+  FftData S_C;
+  FftData S_NEON;
+  FftData G;
+  Aec3Fft fft;
+  std::vector<FftData> H_C(10);
+  std::vector<FftData> H_NEON(10);
+  for (auto& H_j : H_C) {
+    H_j.Clear();
+  }
+  for (auto& H_j : H_NEON) {
+    H_j.Clear();
+  }
+
+  for (size_t k = 0; k < 30; ++k) {
+    RandomizeSampleVector(&random_generator, x[0]);
+    render_buffer.Insert(x);
+  }
+
+  for (size_t j = 0; j < G.re.size(); ++j) {
+    G.re[j] = j / 10001.f;
+  }
+  for (size_t j = 1; j < G.im.size() - 1; ++j) {
+    G.im[j] = j / 20001.f;
+  }
+  G.im[0] = 0.f;
+  G.im[G.im.size() - 1] = 0.f;
+
+  AdaptPartitions_NEON(render_buffer, G, H_NEON);
+  AdaptPartitions(render_buffer, G, H_C);
+  AdaptPartitions_NEON(render_buffer, G, H_NEON);
+  AdaptPartitions(render_buffer, G, H_C);
+
+  for (size_t l = 0; l < H_C.size(); ++l) {
+    for (size_t j = 0; j < H_C[l].im.size(); ++j) {
+      EXPECT_NEAR(H_C[l].re[j], H_NEON[l].re[j], fabs(H_C[l].re[j] * 0.00001f));
+      EXPECT_NEAR(H_C[l].im[j], H_NEON[l].im[j], fabs(H_C[l].im[j] * 0.00001f));
+    }
+  }
+
+  ApplyFilter_NEON(render_buffer, H_NEON, &S_NEON);
+  ApplyFilter(render_buffer, H_C, &S_C);
+  for (size_t j = 0; j < S_C.re.size(); ++j) {
+    EXPECT_NEAR(S_C.re[j], S_NEON.re[j], fabs(S_C.re[j] * 0.00001f));
+    EXPECT_NEAR(S_C.im[j], S_NEON.im[j], fabs(S_C.re[j] * 0.00001f));
+  }
+}
+
+// Verifies that the optimized method for frequency response computation is
+// bitexact to the reference counterpart.
+TEST(AdaptiveFirFilter, UpdateFrequencyResponseNeonOptimization) {
+  const size_t kNumPartitions = 12;
+  std::vector<FftData> H(kNumPartitions);
+  std::vector<std::array<float, kFftLengthBy2Plus1>> H2(kNumPartitions);
+  std::vector<std::array<float, kFftLengthBy2Plus1>> H2_NEON(kNumPartitions);
+
+  for (size_t j = 0; j < H.size(); ++j) {
+    for (size_t k = 0; k < H[j].re.size(); ++k) {
+      H[j].re[k] = k + j / 3.f;
+      H[j].im[k] = j + k / 7.f;
+    }
+  }
+
+  UpdateFrequencyResponse(H, &H2);
+  UpdateFrequencyResponse_NEON(H, &H2_NEON);
+
+  for (size_t j = 0; j < H2.size(); ++j) {
+    for (size_t k = 0; k < H[j].re.size(); ++k) {
+      EXPECT_FLOAT_EQ(H2[j][k], H2_NEON[j][k]);
+    }
+  }
+}
+
+// Verifies that the optimized method for echo return loss computation is
+// bitexact to the reference counterpart.
+TEST(AdaptiveFirFilter, UpdateErlNeonOptimization) {
+  const size_t kNumPartitions = 12;
+  std::vector<std::array<float, kFftLengthBy2Plus1>> H2(kNumPartitions);
+  std::array<float, kFftLengthBy2Plus1> erl;
+  std::array<float, kFftLengthBy2Plus1> erl_NEON;
+
+  for (size_t j = 0; j < H2.size(); ++j) {
+    for (size_t k = 0; k < H2[j].size(); ++k) {
+      H2[j][k] = k + j / 3.f;
+    }
+  }
+
+  UpdateErlEstimator(H2, &erl);
+  UpdateErlEstimator_NEON(H2, &erl_NEON);
+
+  for (size_t j = 0; j < erl.size(); ++j) {
+    EXPECT_FLOAT_EQ(erl[j], erl_NEON[j]);
+  }
+}
+
+#endif
+
 #if defined(WEBRTC_ARCH_X86_FAMILY)
 // Verifies that the optimized methods for filter adaptation are bitexact to
 // their reference counterparts.
-TEST(AdaptiveFirFilter, FilterAdaptationOptimizations) {
+TEST(AdaptiveFirFilter, FilterAdaptationSse2Optimizations) {
   bool use_sse2 = (WebRtc_GetCPUInfo(kSSE2) != 0);
   if (use_sse2) {
     RenderBuffer render_buffer(Aec3Optimization::kNone, 3, 12,
@@ -95,7 +200,7 @@
 
 // Verifies that the optimized method for frequency response computation is
 // bitexact to the reference counterpart.
-TEST(AdaptiveFirFilter, UpdateFrequencyResponseOptimization) {
+TEST(AdaptiveFirFilter, UpdateFrequencyResponseSse2Optimization) {
   bool use_sse2 = (WebRtc_GetCPUInfo(kSSE2) != 0);
   if (use_sse2) {
     const size_t kNumPartitions = 12;
@@ -123,7 +228,7 @@
 
 // Verifies that the optimized method for echo return loss computation is
 // bitexact to the reference counterpart.
-TEST(AdaptiveFirFilter, UpdateErlOptimization) {
+TEST(AdaptiveFirFilter, UpdateErlSse2Optimization) {
   bool use_sse2 = (WebRtc_GetCPUInfo(kSSE2) != 0);
   if (use_sse2) {
     const size_t kNumPartitions = 12;