Adding upstream version 110.0.1.upstream/110.0.1upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

12 files changed, 1043 insertions, 0 deletions

diff --git a/third_party/libwebrtc/modules/video_processing/util/denoiser_filter.cc b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter.cc
new file mode 100644
index 0000000000..0e1570114a
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter.cc
@@ -0,0 +1,65 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "modules/video_processing/util/denoiser_filter.h"
+
+#include "modules/video_processing/util/denoiser_filter_c.h"
+#include "rtc_base/checks.h"
+#include "rtc_base/system/arch.h"
+#include "system_wrappers/include/cpu_features_wrapper.h"
+
+#if defined(WEBRTC_ARCH_X86_FAMILY)
+#include "modules/video_processing/util/denoiser_filter_sse2.h"
+#elif defined(WEBRTC_HAS_NEON)
+#include "modules/video_processing/util/denoiser_filter_neon.h"
+#endif
+
+namespace webrtc {
+
+const int kMotionMagnitudeThreshold = 8 * 3;
+const int kSumDiffThreshold = 96;
+const int kSumDiffThresholdHigh = 448;
+
+std::unique_ptr<DenoiserFilter> DenoiserFilter::Create(
+    bool runtime_cpu_detection,
+    CpuType* cpu_type) {
+  std::unique_ptr<DenoiserFilter> filter;
+
+  if (cpu_type != nullptr)
+    *cpu_type = CPU_NOT_NEON;
+  if (runtime_cpu_detection) {
+// If we know the minimum architecture at compile time, avoid CPU detection.
+#if defined(WEBRTC_ARCH_X86_FAMILY)
+#if defined(__SSE2__)
+    filter.reset(new DenoiserFilterSSE2());
+#else
+    // x86 CPU detection required.
+    if (GetCPUInfo(kSSE2)) {
+      filter.reset(new DenoiserFilterSSE2());
+    } else {
+      filter.reset(new DenoiserFilterC());
+    }
+#endif
+#elif defined(WEBRTC_HAS_NEON)
+    filter.reset(new DenoiserFilterNEON());
+    if (cpu_type != nullptr)
+      *cpu_type = CPU_NEON;
+#else
+    filter.reset(new DenoiserFilterC());
+#endif
+  } else {
+    filter.reset(new DenoiserFilterC());
+  }
+
+  RTC_DCHECK(filter.get() != nullptr);
+  return filter;
+}
+
+}  // namespace webrtc
diff --git a/third_party/libwebrtc/modules/video_processing/util/denoiser_filter.h b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter.h
new file mode 100644
index 0000000000..1d574f4a4f
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter.h
@@ -0,0 +1,50 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_H_
+#define MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_H_
+
+#include <stdint.h>
+
+#include <memory>
+
+namespace webrtc {
+
+extern const int kMotionMagnitudeThreshold;
+extern const int kSumDiffThreshold;
+extern const int kSumDiffThresholdHigh;
+
+enum DenoiserDecision { COPY_BLOCK, FILTER_BLOCK };
+enum CpuType { CPU_NEON, CPU_NOT_NEON };
+
+class DenoiserFilter {
+ public:
+  static std::unique_ptr<DenoiserFilter> Create(bool runtime_cpu_detection,
+                                                CpuType* cpu_type);
+
+  virtual ~DenoiserFilter() {}
+  virtual uint32_t Variance16x8(const uint8_t* a,
+                                int a_stride,
+                                const uint8_t* b,
+                                int b_stride,
+                                unsigned int* sse) = 0;
+  virtual DenoiserDecision MbDenoise(const uint8_t* mc_running_avg_y,
+                                     int mc_avg_y_stride,
+                                     uint8_t* running_avg_y,
+                                     int avg_y_stride,
+                                     const uint8_t* sig,
+                                     int sig_stride,
+                                     uint8_t motion_magnitude,
+                                     int increase_denoising) = 0;
+};
+
+}  // namespace webrtc
+
+#endif  // MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_H_
diff --git a/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_c.cc b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_c.cc
new file mode 100644
index 0000000000..55c0ea7b35
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_c.cc
@@ -0,0 +1,126 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "modules/video_processing/util/denoiser_filter_c.h"
+
+#include <stdlib.h>
+#include <string.h>
+
+namespace webrtc {
+
+uint32_t DenoiserFilterC::Variance16x8(const uint8_t* a,
+                                       int a_stride,
+                                       const uint8_t* b,
+                                       int b_stride,
+                                       uint32_t* sse) {
+  int sum = 0;
+  *sse = 0;
+  a_stride <<= 1;
+  b_stride <<= 1;
+
+  for (int i = 0; i < 8; i++) {
+    for (int j = 0; j < 16; j++) {
+      const int diff = a[j] - b[j];
+      sum += diff;
+      *sse += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+  return *sse - ((static_cast<int64_t>(sum) * sum) >> 7);
+}
+
+DenoiserDecision DenoiserFilterC::MbDenoise(const uint8_t* mc_running_avg_y,
+                                            int mc_avg_y_stride,
+                                            uint8_t* running_avg_y,
+                                            int avg_y_stride,
+                                            const uint8_t* sig,
+                                            int sig_stride,
+                                            uint8_t motion_magnitude,
+                                            int increase_denoising) {
+  int sum_diff_thresh = 0;
+  int sum_diff = 0;
+  int adj_val[3] = {3, 4, 6};
+  int shift_inc1 = 0;
+  int shift_inc2 = 1;
+  int col_sum[16] = {0};
+  if (motion_magnitude <= kMotionMagnitudeThreshold) {
+    if (increase_denoising) {
+      shift_inc1 = 1;
+      shift_inc2 = 2;
+    }
+    adj_val[0] += shift_inc2;
+    adj_val[1] += shift_inc2;
+    adj_val[2] += shift_inc2;
+  }
+
+  for (int r = 0; r < 16; ++r) {
+    for (int c = 0; c < 16; ++c) {
+      int diff = 0;
+      int adjustment = 0;
+      int absdiff = 0;
+
+      diff = mc_running_avg_y[c] - sig[c];
+      absdiff = abs(diff);
+
+      // When `diff` <= |3 + shift_inc1|, use pixel value from
+      // last denoised raw.
+      if (absdiff <= 3 + shift_inc1) {
+        running_avg_y[c] = mc_running_avg_y[c];
+        col_sum[c] += diff;
+      } else {
+        if (absdiff >= 4 + shift_inc1 && absdiff <= 7)
+          adjustment = adj_val[0];
+        else if (absdiff >= 8 && absdiff <= 15)
+          adjustment = adj_val[1];
+        else
+          adjustment = adj_val[2];
+
+        if (diff > 0) {
+          if ((sig[c] + adjustment) > 255)
+            running_avg_y[c] = 255;
+          else
+            running_avg_y[c] = sig[c] + adjustment;
+
+          col_sum[c] += adjustment;
+        } else {
+          if ((sig[c] - adjustment) < 0)
+            running_avg_y[c] = 0;
+          else
+            running_avg_y[c] = sig[c] - adjustment;
+
+          col_sum[c] -= adjustment;
+        }
+      }
+    }
+
+    // Update pointers for next iteration.
+    sig += sig_stride;
+    mc_running_avg_y += mc_avg_y_stride;
+    running_avg_y += avg_y_stride;
+  }
+
+  for (int c = 0; c < 16; ++c) {
+    if (col_sum[c] >= 128) {
+      col_sum[c] = 127;
+    }
+    sum_diff += col_sum[c];
+  }
+
+  sum_diff_thresh =
+      increase_denoising ? kSumDiffThresholdHigh : kSumDiffThreshold;
+  if (abs(sum_diff) > sum_diff_thresh)
+    return COPY_BLOCK;
+
+  return FILTER_BLOCK;
+}
+
+}  // namespace webrtc
diff --git a/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_c.h b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_c.h
new file mode 100644
index 0000000000..5633c171f0
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_c.h
@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_C_H_
+#define MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_C_H_
+
+#include <stdint.h>
+
+#include "modules/video_processing/util/denoiser_filter.h"
+
+namespace webrtc {
+
+class DenoiserFilterC : public DenoiserFilter {
+ public:
+  DenoiserFilterC() {}
+  uint32_t Variance16x8(const uint8_t* a,
+                        int a_stride,
+                        const uint8_t* b,
+                        int b_stride,
+                        unsigned int* sse) override;
+  DenoiserDecision MbDenoise(const uint8_t* mc_running_avg_y,
+                             int mc_avg_y_stride,
+                             uint8_t* running_avg_y,
+                             int avg_y_stride,
+                             const uint8_t* sig,
+                             int sig_stride,
+                             uint8_t motion_magnitude,
+                             int increase_denoising) override;
+};
+
+}  // namespace webrtc
+
+#endif  // MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_C_H_
diff --git a/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_neon.cc b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_neon.cc
new file mode 100644
index 0000000000..e1e6ed4f18
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_neon.cc
@@ -0,0 +1,182 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "modules/video_processing/util/denoiser_filter_neon.h"
+
+#include <arm_neon.h>
+
+namespace webrtc {
+
+const int kSumDiffThresholdHighNeon = 600;
+
+static int HorizontalAddS16x8(const int16x8_t v_16x8) {
+  const int32x4_t a = vpaddlq_s16(v_16x8);
+  const int64x2_t b = vpaddlq_s32(a);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+static int HorizontalAddS32x4(const int32x4_t v_32x4) {
+  const int64x2_t b = vpaddlq_s32(v_32x4);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+static void VarianceNeonW8(const uint8_t* a,
+                           int a_stride,
+                           const uint8_t* b,
+                           int b_stride,
+                           int w,
+                           int h,
+                           uint32_t* sse,
+                           int64_t* sum) {
+  int16x8_t v_sum = vdupq_n_s16(0);
+  int32x4_t v_sse_lo = vdupq_n_s32(0);
+  int32x4_t v_sse_hi = vdupq_n_s32(0);
+
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 8) {
+      const uint8x8_t v_a = vld1_u8(&a[j]);
+      const uint8x8_t v_b = vld1_u8(&b[j]);
+      const uint16x8_t v_diff = vsubl_u8(v_a, v_b);
+      const int16x8_t sv_diff = vreinterpretq_s16_u16(v_diff);
+      v_sum = vaddq_s16(v_sum, sv_diff);
+      v_sse_lo =
+          vmlal_s16(v_sse_lo, vget_low_s16(sv_diff), vget_low_s16(sv_diff));
+      v_sse_hi =
+          vmlal_s16(v_sse_hi, vget_high_s16(sv_diff), vget_high_s16(sv_diff));
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+
+  *sum = HorizontalAddS16x8(v_sum);
+  *sse =
+      static_cast<uint32_t>(HorizontalAddS32x4(vaddq_s32(v_sse_lo, v_sse_hi)));
+}
+
+uint32_t DenoiserFilterNEON::Variance16x8(const uint8_t* a,
+                                          int a_stride,
+                                          const uint8_t* b,
+                                          int b_stride,
+                                          uint32_t* sse) {
+  int64_t sum = 0;
+  VarianceNeonW8(a, a_stride << 1, b, b_stride << 1, 16, 8, sse, &sum);
+  return *sse - ((sum * sum) >> 7);
+}
+
+DenoiserDecision DenoiserFilterNEON::MbDenoise(const uint8_t* mc_running_avg_y,
+                                               int mc_running_avg_y_stride,
+                                               uint8_t* running_avg_y,
+                                               int running_avg_y_stride,
+                                               const uint8_t* sig,
+                                               int sig_stride,
+                                               uint8_t motion_magnitude,
+                                               int increase_denoising) {
+  // If motion_magnitude is small, making the denoiser more aggressive by
+  // increasing the adjustment for each level, level1 adjustment is
+  // increased, the deltas stay the same.
+  int shift_inc =
+      (increase_denoising && motion_magnitude <= kMotionMagnitudeThreshold) ? 1
+                                                                            : 0;
+  int sum_diff_thresh = 0;
+  const uint8x16_t v_level1_adjustment = vmovq_n_u8(
+      (motion_magnitude <= kMotionMagnitudeThreshold) ? 4 + shift_inc : 3);
+  const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
+  const uint8x16_t v_delta_level_2_and_3 = vdupq_n_u8(2);
+  const uint8x16_t v_level1_threshold = vmovq_n_u8(4 + shift_inc);
+  const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
+  const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
+  int64x2_t v_sum_diff_total = vdupq_n_s64(0);
+
+  // Go over lines.
+  for (int r = 0; r < 16; ++r) {
+    // Load inputs.
+    const uint8x16_t v_sig = vld1q_u8(sig);
+    const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
+
+    // Calculate absolute difference and sign masks.
+    const uint8x16_t v_abs_diff = vabdq_u8(v_sig, v_mc_running_avg_y);
+    const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig, v_mc_running_avg_y);
+    const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig, v_mc_running_avg_y);
+
+    // Figure out which level that put us in.
+    const uint8x16_t v_level1_mask = vcleq_u8(v_level1_threshold, v_abs_diff);
+    const uint8x16_t v_level2_mask = vcleq_u8(v_level2_threshold, v_abs_diff);
+    const uint8x16_t v_level3_mask = vcleq_u8(v_level3_threshold, v_abs_diff);
+
+    // Calculate absolute adjustments for level 1, 2 and 3.
+    const uint8x16_t v_level2_adjustment =
+        vandq_u8(v_level2_mask, v_delta_level_1_and_2);
+    const uint8x16_t v_level3_adjustment =
+        vandq_u8(v_level3_mask, v_delta_level_2_and_3);
+    const uint8x16_t v_level1and2_adjustment =
+        vaddq_u8(v_level1_adjustment, v_level2_adjustment);
+    const uint8x16_t v_level1and2and3_adjustment =
+        vaddq_u8(v_level1and2_adjustment, v_level3_adjustment);
+
+    // Figure adjustment absolute value by selecting between the absolute
+    // difference if in level0 or the value for level 1, 2 and 3.
+    const uint8x16_t v_abs_adjustment =
+        vbslq_u8(v_level1_mask, v_level1and2and3_adjustment, v_abs_diff);
+
+    // Calculate positive and negative adjustments. Apply them to the signal
+    // and accumulate them. Adjustments are less than eight and the maximum
+    // sum of them (7 * 16) can fit in a signed char.
+    const uint8x16_t v_pos_adjustment =
+        vandq_u8(v_diff_pos_mask, v_abs_adjustment);
+    const uint8x16_t v_neg_adjustment =
+        vandq_u8(v_diff_neg_mask, v_abs_adjustment);
+
+    uint8x16_t v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
+    v_running_avg_y = vqsubq_u8(v_running_avg_y, v_neg_adjustment);
+
+    // Store results.
+    vst1q_u8(running_avg_y, v_running_avg_y);
+
+    // Sum all the accumulators to have the sum of all pixel differences
+    // for this macroblock.
+    {
+      const int8x16_t v_sum_diff =
+          vqsubq_s8(vreinterpretq_s8_u8(v_pos_adjustment),
+                    vreinterpretq_s8_u8(v_neg_adjustment));
+      const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff);
+      const int32x4_t fedc_ba98_7654_3210 =
+          vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+      const int64x2_t fedcba98_76543210 = vpaddlq_s32(fedc_ba98_7654_3210);
+
+      v_sum_diff_total = vqaddq_s64(v_sum_diff_total, fedcba98_76543210);
+    }
+
+    // Update pointers for next iteration.
+    sig += sig_stride;
+    mc_running_avg_y += mc_running_avg_y_stride;
+    running_avg_y += running_avg_y_stride;
+  }
+
+  // Too much adjustments => copy block.
+  int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                          vget_low_s64(v_sum_diff_total));
+  int sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+  sum_diff_thresh =
+      increase_denoising ? kSumDiffThresholdHighNeon : kSumDiffThreshold;
+  if (sum_diff > sum_diff_thresh)
+    return COPY_BLOCK;
+
+  // Tell above level that block was filtered.
+  running_avg_y -= running_avg_y_stride * 16;
+  sig -= sig_stride * 16;
+
+  return FILTER_BLOCK;
+}
+
+}  // namespace webrtc
diff --git a/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_neon.h b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_neon.h
new file mode 100644
index 0000000000..4d9f271e5a
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_neon.h
@@ -0,0 +1,38 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_NEON_H_
+#define MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_NEON_H_
+
+#include "modules/video_processing/util/denoiser_filter.h"
+
+namespace webrtc {
+
+class DenoiserFilterNEON : public DenoiserFilter {
+ public:
+  DenoiserFilterNEON() {}
+  uint32_t Variance16x8(const uint8_t* a,
+                        int a_stride,
+                        const uint8_t* b,
+                        int b_stride,
+                        unsigned int* sse) override;
+  DenoiserDecision MbDenoise(const uint8_t* mc_running_avg_y,
+                             int mc_avg_y_stride,
+                             uint8_t* running_avg_y,
+                             int avg_y_stride,
+                             const uint8_t* sig,
+                             int sig_stride,
+                             uint8_t motion_magnitude,
+                             int increase_denoising) override;
+};
+
+}  // namespace webrtc
+
+#endif  // MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_NEON_H_
diff --git a/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_sse2.cc b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_sse2.cc
new file mode 100644
index 0000000000..5ca5f0cf34
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_sse2.cc
@@ -0,0 +1,200 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "modules/video_processing/util/denoiser_filter_sse2.h"
+
+#include <emmintrin.h>
+#include <stdlib.h>
+#include <string.h>
+
+namespace webrtc {
+
+static void Get8x8varSse2(const uint8_t* src,
+                          int src_stride,
+                          const uint8_t* ref,
+                          int ref_stride,
+                          unsigned int* sse,
+                          int* sum) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i vsum = _mm_setzero_si128();
+  __m128i vsse = _mm_setzero_si128();
+
+  for (int i = 0; i < 8; i += 2) {
+    const __m128i src0 = _mm_unpacklo_epi8(
+        _mm_loadl_epi64((const __m128i*)(src + i * src_stride)), zero);
+    const __m128i ref0 = _mm_unpacklo_epi8(
+        _mm_loadl_epi64((const __m128i*)(ref + i * ref_stride)), zero);
+    const __m128i diff0 = _mm_sub_epi16(src0, ref0);
+
+    const __m128i src1 = _mm_unpacklo_epi8(
+        _mm_loadl_epi64((const __m128i*)(src + (i + 1) * src_stride)), zero);
+    const __m128i ref1 = _mm_unpacklo_epi8(
+        _mm_loadl_epi64((const __m128i*)(ref + (i + 1) * ref_stride)), zero);
+    const __m128i diff1 = _mm_sub_epi16(src1, ref1);
+
+    vsum = _mm_add_epi16(vsum, diff0);
+    vsum = _mm_add_epi16(vsum, diff1);
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
+  }
+
+  // sum
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
+  *sum = static_cast<int16_t>(_mm_extract_epi16(vsum, 0));
+
+  // sse
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
+  *sse = _mm_cvtsi128_si32(vsse);
+}
+
+static void VarianceSSE2(const unsigned char* src,
+                         int src_stride,
+                         const unsigned char* ref,
+                         int ref_stride,
+                         int w,
+                         int h,
+                         uint32_t* sse,
+                         int64_t* sum,
+                         int block_size) {
+  *sse = 0;
+  *sum = 0;
+
+  for (int i = 0; i < h; i += block_size) {
+    for (int j = 0; j < w; j += block_size) {
+      uint32_t sse0 = 0;
+      int32_t sum0 = 0;
+
+      Get8x8varSse2(src + src_stride * i + j, src_stride,
+                    ref + ref_stride * i + j, ref_stride, &sse0, &sum0);
+      *sse += sse0;
+      *sum += sum0;
+    }
+  }
+}
+
+// Compute the sum of all pixel differences of this MB.
+static uint32_t AbsSumDiff16x1(__m128i acc_diff) {
+  const __m128i k_1 = _mm_set1_epi16(1);
+  const __m128i acc_diff_lo =
+      _mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_hi =
+      _mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
+  const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
+  const __m128i hgfe_dcba =
+      _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
+  const __m128i hgfedcba =
+      _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
+  unsigned int sum_diff = abs(_mm_cvtsi128_si32(hgfedcba));
+
+  return sum_diff;
+}
+
+uint32_t DenoiserFilterSSE2::Variance16x8(const uint8_t* src,
+                                          int src_stride,
+                                          const uint8_t* ref,
+                                          int ref_stride,
+                                          uint32_t* sse) {
+  int64_t sum = 0;
+  VarianceSSE2(src, src_stride << 1, ref, ref_stride << 1, 16, 8, sse, &sum, 8);
+  return *sse - ((sum * sum) >> 7);
+}
+
+DenoiserDecision DenoiserFilterSSE2::MbDenoise(const uint8_t* mc_running_avg_y,
+                                               int mc_avg_y_stride,
+                                               uint8_t* running_avg_y,
+                                               int avg_y_stride,
+                                               const uint8_t* sig,
+                                               int sig_stride,
+                                               uint8_t motion_magnitude,
+                                               int increase_denoising) {
+  DenoiserDecision decision = FILTER_BLOCK;
+  unsigned int sum_diff_thresh = 0;
+  int shift_inc =
+      (increase_denoising && motion_magnitude <= kMotionMagnitudeThreshold) ? 1
+                                                                            : 0;
+  __m128i acc_diff = _mm_setzero_si128();
+  const __m128i k_0 = _mm_setzero_si128();
+  const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+  const __m128i k_8 = _mm_set1_epi8(8);
+  const __m128i k_16 = _mm_set1_epi8(16);
+  // Modify each level's adjustment according to motion_magnitude.
+  const __m128i l3 = _mm_set1_epi8(
+      (motion_magnitude <= kMotionMagnitudeThreshold) ? 7 + shift_inc : 6);
+  // Difference between level 3 and level 2 is 2.
+  const __m128i l32 = _mm_set1_epi8(2);
+  // Difference between level 2 and level 1 is 1.
+  const __m128i l21 = _mm_set1_epi8(1);
+
+  for (int r = 0; r < 16; ++r) {
+    // Calculate differences.
+    const __m128i v_sig =
+        _mm_loadu_si128(reinterpret_cast<const __m128i*>(&sig[0]));
+    const __m128i v_mc_running_avg_y =
+        _mm_loadu_si128(reinterpret_cast<const __m128i*>(&mc_running_avg_y[0]));
+    __m128i v_running_avg_y;
+    const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+    const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+    // Obtain the sign. FF if diff is negative.
+    const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+    // Clamp absolute difference to 16 to be used to get mask. Doing this
+    // allows us to use _mm_cmpgt_epi8, which operates on signed byte.
+    const __m128i clamped_absdiff =
+        _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_16);
+    // Get masks for l2 l1 and l0 adjustments.
+    const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
+    const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
+    const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
+    // Get adjustments for l2, l1, and l0.
+    __m128i adj2 = _mm_and_si128(mask2, l32);
+    const __m128i adj1 = _mm_and_si128(mask1, l21);
+    const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
+    __m128i adj, padj, nadj;
+
+    // Combine the adjustments and get absolute adjustments.
+    adj2 = _mm_add_epi8(adj2, adj1);
+    adj = _mm_sub_epi8(l3, adj2);
+    adj = _mm_andnot_si128(mask0, adj);
+    adj = _mm_or_si128(adj, adj0);
+
+    // Restore the sign and get positive and negative adjustments.
+    padj = _mm_andnot_si128(diff_sign, adj);
+    nadj = _mm_and_si128(diff_sign, adj);
+
+    // Calculate filtered value.
+    v_running_avg_y = _mm_adds_epu8(v_sig, padj);
+    v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
+    _mm_storeu_si128(reinterpret_cast<__m128i*>(running_avg_y),
+                     v_running_avg_y);
+
+    // Adjustments <=7, and each element in acc_diff can fit in signed
+    // char.
+    acc_diff = _mm_adds_epi8(acc_diff, padj);
+    acc_diff = _mm_subs_epi8(acc_diff, nadj);
+
+    // Update pointers for next iteration.
+    sig += sig_stride;
+    mc_running_avg_y += mc_avg_y_stride;
+    running_avg_y += avg_y_stride;
+  }
+
+  // Compute the sum of all pixel differences of this MB.
+  unsigned int abs_sum_diff = AbsSumDiff16x1(acc_diff);
+  sum_diff_thresh =
+      increase_denoising ? kSumDiffThresholdHigh : kSumDiffThreshold;
+  if (abs_sum_diff > sum_diff_thresh)
+    decision = COPY_BLOCK;
+  return decision;
+}
+
+}  // namespace webrtc
diff --git a/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_sse2.h b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_sse2.h
new file mode 100644
index 0000000000..8fe4b905ae
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/denoiser_filter_sse2.h
@@ -0,0 +1,40 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_SSE2_H_
+#define MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_SSE2_H_
+
+#include <stdint.h>
+
+#include "modules/video_processing/util/denoiser_filter.h"
+
+namespace webrtc {
+
+class DenoiserFilterSSE2 : public DenoiserFilter {
+ public:
+  DenoiserFilterSSE2() {}
+  uint32_t Variance16x8(const uint8_t* a,
+                        int a_stride,
+                        const uint8_t* b,
+                        int b_stride,
+                        unsigned int* sse) override;
+  DenoiserDecision MbDenoise(const uint8_t* mc_running_avg_y,
+                             int mc_avg_y_stride,
+                             uint8_t* running_avg_y,
+                             int avg_y_stride,
+                             const uint8_t* sig,
+                             int sig_stride,
+                             uint8_t motion_magnitude,
+                             int increase_denoising) override;
+};
+
+}  // namespace webrtc
+
+#endif  // MODULES_VIDEO_PROCESSING_UTIL_DENOISER_FILTER_SSE2_H_
diff --git a/third_party/libwebrtc/modules/video_processing/util/noise_estimation.cc b/third_party/libwebrtc/modules/video_processing/util/noise_estimation.cc
new file mode 100644
index 0000000000..c72f764901
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/noise_estimation.cc
@@ -0,0 +1,113 @@
+/*
+ *  Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "modules/video_processing/util/noise_estimation.h"
+#if DISPLAYNEON
+#include <android/log.h>
+#endif
+
+namespace webrtc {
+
+void NoiseEstimation::Init(int width, int height, CpuType cpu_type) {
+  int mb_cols = width >> 4;
+  int mb_rows = height >> 4;
+  consec_low_var_.reset(new uint32_t[mb_cols * mb_rows]());
+  width_ = width;
+  height_ = height;
+  mb_cols_ = width_ >> 4;
+  mb_rows_ = height_ >> 4;
+  cpu_type_ = cpu_type;
+}
+
+void NoiseEstimation::GetNoise(int mb_index, uint32_t var, uint32_t luma) {
+  consec_low_var_[mb_index]++;
+  num_static_block_++;
+  if (consec_low_var_[mb_index] >= kConsecLowVarFrame &&
+      (luma >> 6) < kAverageLumaMax && (luma >> 6) > kAverageLumaMin) {
+    // Normalized var by the average luma value, this gives more weight to
+    // darker blocks.
+    int nor_var = var / (luma >> 10);
+    noise_var_ +=
+        nor_var > kBlockSelectionVarMax ? kBlockSelectionVarMax : nor_var;
+    num_noisy_block_++;
+  }
+}
+
+void NoiseEstimation::ResetConsecLowVar(int mb_index) {
+  consec_low_var_[mb_index] = 0;
+}
+
+void NoiseEstimation::UpdateNoiseLevel() {
+  // TODO(jackychen): Tune a threshold for numb_noisy_block > T to make the
+  // condition more reasonable.
+  // No enough samples implies the motion of the camera or too many moving
+  // objects in the frame.
+  if (num_static_block_ <
+          (0.65 * mb_cols_ * mb_rows_ / NOISE_SUBSAMPLE_INTERVAL) ||
+      !num_noisy_block_) {
+#if DISPLAY
+    printf("Not enough samples. %d \n", num_static_block_);
+#elif DISPLAYNEON
+    __android_log_print(ANDROID_LOG_DEBUG, "DISPLAY",
+                        "Not enough samples. %d \n", num_static_block_);
+#endif
+    noise_var_ = 0;
+    noise_var_accum_ = 0;
+    num_noisy_block_ = 0;
+    num_static_block_ = 0;
+    return;
+  } else {
+#if DISPLAY
+    printf("%d %d fraction = %.3f\n", num_static_block_,
+           mb_cols_ * mb_rows_ / NOISE_SUBSAMPLE_INTERVAL,
+           percent_static_block_);
+#elif DISPLAYNEON
+    __android_log_print(ANDROID_LOG_DEBUG, "DISPLAY", "%d %d fraction = %.3f\n",
+                        num_static_block_,
+                        mb_cols_ * mb_rows_ / NOISE_SUBSAMPLE_INTERVAL,
+                        percent_static_block_);
+#endif
+    // Normalized by the number of noisy blocks.
+    noise_var_ /= num_noisy_block_;
+    // Get the percentage of static blocks.
+    percent_static_block_ = static_cast<double>(num_static_block_) /
+                            (mb_cols_ * mb_rows_ / NOISE_SUBSAMPLE_INTERVAL);
+    num_noisy_block_ = 0;
+    num_static_block_ = 0;
+  }
+  // For the first frame just update the value with current noise_var_,
+  // otherwise, use the averaging window.
+  if (noise_var_accum_ == 0) {
+    noise_var_accum_ = noise_var_;
+  } else {
+    noise_var_accum_ = (noise_var_accum_ * 15 + noise_var_) / 16;
+  }
+#if DISPLAY
+  printf("noise_var_accum_ = %.1f, noise_var_ = %d.\n", noise_var_accum_,
+         noise_var_);
+#elif DISPLAYNEON
+  __android_log_print(ANDROID_LOG_DEBUG, "DISPLAY",
+                      "noise_var_accum_ = %.1f, noise_var_ = %d.\n",
+                      noise_var_accum_, noise_var_);
+#endif
+  // Reset noise_var_ for the next frame.
+  noise_var_ = 0;
+}
+
+uint8_t NoiseEstimation::GetNoiseLevel() {
+  int noise_thr = cpu_type_ ? kNoiseThreshold : kNoiseThresholdNeon;
+  UpdateNoiseLevel();
+  if (noise_var_accum_ > noise_thr) {
+    return 1;
+  }
+  return 0;
+}
+
+}  // namespace webrtc
diff --git a/third_party/libwebrtc/modules/video_processing/util/noise_estimation.h b/third_party/libwebrtc/modules/video_processing/util/noise_estimation.h
new file mode 100644
index 0000000000..4c5f10f1d8
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/noise_estimation.h
@@ -0,0 +1,63 @@
+/*
+ *  Copyright (c) 2016 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef MODULES_VIDEO_PROCESSING_UTIL_NOISE_ESTIMATION_H_
+#define MODULES_VIDEO_PROCESSING_UTIL_NOISE_ESTIMATION_H_
+
+#include <cstdint>
+#include <memory>
+
+#include "modules/video_processing/util/denoiser_filter.h"
+
+namespace webrtc {
+
+#define DISPLAY 0      // Rectangle diagnostics
+#define DISPLAYNEON 0  // Rectangle diagnostics on NEON
+
+const int kNoiseThreshold = 150;
+const int kNoiseThresholdNeon = 70;
+const int kConsecLowVarFrame = 6;
+const int kAverageLumaMin = 20;
+const int kAverageLumaMax = 220;
+const int kBlockSelectionVarMax = kNoiseThreshold << 1;
+
+// TODO(jackychen): To test different sampling strategy.
+// Collect noise data every NOISE_SUBSAMPLE_INTERVAL blocks.
+#define NOISE_SUBSAMPLE_INTERVAL 41
+
+class NoiseEstimation {
+ public:
+  void Init(int width, int height, CpuType cpu_type);
+  // Collect noise data from one qualified block.
+  void GetNoise(int mb_index, uint32_t var, uint32_t luma);
+  // Reset the counter for consecutive low-var blocks.
+  void ResetConsecLowVar(int mb_index);
+  // Update noise level for current frame.
+  void UpdateNoiseLevel();
+  // 0: low noise, 1: high noise
+  uint8_t GetNoiseLevel();
+
+ private:
+  int width_;
+  int height_;
+  int mb_rows_;
+  int mb_cols_;
+  int num_noisy_block_;
+  int num_static_block_;
+  CpuType cpu_type_;
+  uint32_t noise_var_;
+  double noise_var_accum_;
+  double percent_static_block_;
+  std::unique_ptr<uint32_t[]> consec_low_var_;
+};
+
+}  // namespace webrtc
+
+#endif  // MODULES_VIDEO_PROCESSING_UTIL_NOISE_ESTIMATION_H_
diff --git a/third_party/libwebrtc/modules/video_processing/util/skin_detection.cc b/third_party/libwebrtc/modules/video_processing/util/skin_detection.cc
new file mode 100644
index 0000000000..76399d6e73
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/skin_detection.cc
@@ -0,0 +1,96 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "modules/video_processing/util/skin_detection.h"
+
+namespace webrtc {
+
+// Fixed-point skin color model parameters.
+static const int skin_mean[5][2] = {{7463, 9614},
+                                    {6400, 10240},
+                                    {7040, 10240},
+                                    {8320, 9280},
+                                    {6800, 9614}};
+static const int skin_inv_cov[4] = {4107, 1663, 1663, 2157};  // q16
+static const int skin_threshold[6] = {1570636, 1400000, 800000,
+                                      800000,  800000,  800000};  // q18
+
+// Thresholds on luminance.
+static const int y_low = 40;
+static const int y_high = 220;
+
+// Evaluates the Mahalanobis distance measure for the input CbCr values.
+static int EvaluateSkinColorDifference(int cb, int cr, int idx) {
+  const int cb_q6 = cb << 6;
+  const int cr_q6 = cr << 6;
+  const int cb_diff_q12 =
+      (cb_q6 - skin_mean[idx][0]) * (cb_q6 - skin_mean[idx][0]);
+  const int cbcr_diff_q12 =
+      (cb_q6 - skin_mean[idx][0]) * (cr_q6 - skin_mean[idx][1]);
+  const int cr_diff_q12 =
+      (cr_q6 - skin_mean[idx][1]) * (cr_q6 - skin_mean[idx][1]);
+  const int cb_diff_q2 = (cb_diff_q12 + (1 << 9)) >> 10;
+  const int cbcr_diff_q2 = (cbcr_diff_q12 + (1 << 9)) >> 10;
+  const int cr_diff_q2 = (cr_diff_q12 + (1 << 9)) >> 10;
+  const int skin_diff =
+      skin_inv_cov[0] * cb_diff_q2 + skin_inv_cov[1] * cbcr_diff_q2 +
+      skin_inv_cov[2] * cbcr_diff_q2 + skin_inv_cov[3] * cr_diff_q2;
+  return skin_diff;
+}
+
+static int SkinPixel(const uint8_t y, const uint8_t cb, const uint8_t cr) {
+  if (y < y_low || y > y_high) {
+    return 0;
+  } else {
+    if (MODEL_MODE == 0) {
+      return (EvaluateSkinColorDifference(cb, cr, 0) < skin_threshold[0]);
+    } else {
+      // Exit on grey.
+      if (cb == 128 && cr == 128)
+        return 0;
+      // Exit on very strong cb.
+      if (cb > 150 && cr < 110)
+        return 0;
+      // Exit on (another) low luminance threshold if either color is high.
+      if (y < 50 && (cb > 140 || cr > 140))
+        return 0;
+      for (int i = 0; i < 5; i++) {
+        int diff = EvaluateSkinColorDifference(cb, cr, i);
+        if (diff < skin_threshold[i + 1]) {
+          return 1;
+        } else if (diff > (skin_threshold[i + 1] << 3)) {
+          // Exit if difference is much large than the threshold.
+          return 0;
+        }
+      }
+      return 0;
+    }
+  }
+}
+
+bool MbHasSkinColor(const uint8_t* y_src,
+                    const uint8_t* u_src,
+                    const uint8_t* v_src,
+                    const int stride_y,
+                    const int stride_u,
+                    const int stride_v,
+                    const int mb_row,
+                    const int mb_col) {
+  const uint8_t* y = y_src + ((mb_row << 4) + 8) * stride_y + (mb_col << 4) + 8;
+  const uint8_t* u = u_src + ((mb_row << 3) + 4) * stride_u + (mb_col << 3) + 4;
+  const uint8_t* v = v_src + ((mb_row << 3) + 4) * stride_v + (mb_col << 3) + 4;
+  // Use 2x2 average of center pixel to compute skin area.
+  uint8_t y_avg = (*y + *(y + 1) + *(y + stride_y) + *(y + stride_y + 1)) >> 2;
+  uint8_t u_avg = (*u + *(u + 1) + *(u + stride_u) + *(u + stride_u + 1)) >> 2;
+  uint8_t v_avg = (*v + *(v + 1) + *(v + stride_v) + *(v + stride_v + 1)) >> 2;
+  return SkinPixel(y_avg, u_avg, v_avg) == 1;
+}
+
+}  // namespace webrtc
diff --git a/third_party/libwebrtc/modules/video_processing/util/skin_detection.h b/third_party/libwebrtc/modules/video_processing/util/skin_detection.h
new file mode 100644
index 0000000000..7f2e17aa87
--- /dev/null
+++ b/third_party/libwebrtc/modules/video_processing/util/skin_detection.h
@@ -0,0 +1,30 @@
+/*
+ *  Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef MODULES_VIDEO_PROCESSING_UTIL_SKIN_DETECTION_H_
+#define MODULES_VIDEO_PROCESSING_UTIL_SKIN_DETECTION_H_
+
+namespace webrtc {
+
+#define MODEL_MODE 0
+
+typedef unsigned char uint8_t;
+bool MbHasSkinColor(const uint8_t* y_src,
+                    const uint8_t* u_src,
+                    const uint8_t* v_src,
+                    int stride_y,
+                    int stride_u,
+                    int stride_v,
+                    int mb_row,
+                    int mb_col);
+
+}  // namespace webrtc
+
+#endif  // MODULES_VIDEO_PROCESSING_UTIL_SKIN_DETECTION_H_