shithub: libvpx

Download patch

ref: a9f479682ac36288d63f2851f5d13a51be20a7ac
parent: a75551585b0fb1ea8705e7e5c53433cfffabbab6
parent: 80465dae8811f8d62742d592b92e641a0ca0c9d9
author: JackyChen <[email protected]>
date: Tue Oct 7 06:51:55 EDT 2014

Merge "Add SSE2 code and unit test for VP9 denoiser."

--- a/test/test.mk
+++ b/test/test.mk
@@ -135,12 +135,15 @@
 LIBVPX_TEST_SRCS-$(CONFIG_SPATIAL_SVC) += svc_test.cc
 endif
 
+ifeq ($(CONFIG_VP9_TEMPORAL_DENOISING),yes)
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_TEMPORAL_DENOISING) += vp9_denoiser_sse2_test.cc
+endif
+
 endif # VP9
 
 LIBVPX_TEST_SRCS-$(CONFIG_ENCODERS)    += sad_test.cc
 
 endif # CONFIG_SHARED
-
 
 ##
 ## TEST DATA
--- /dev/null
+++ b/test/vp9_denoiser_sse2_test.cc
@@ -1,0 +1,102 @@
+/*
+ *  Copyright (c) 2014 The WebM 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 <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+#include "vpx_scale/yv12config.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+
+using libvpx_test::ACMRandom;
+
+namespace {
+
+const int kNumPixels = 64 * 64;
+class VP9DenoiserTest
+    : public ::testing::TestWithParam<int> {
+ public:
+  virtual ~VP9DenoiserTest() {}
+
+  virtual void SetUp() {
+    bs = (BLOCK_SIZE)GetParam();
+  }
+
+  virtual void TearDown() { libvpx_test::ClearSystemState(); }
+
+ protected:
+  BLOCK_SIZE bs;
+};
+
+TEST_P(VP9DenoiserTest, BitexactCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 4000;
+
+  // Allocate the space for input and output,
+  // where sig_block is the block to be denoised,
+  // mc_avg_block is the denoised reference block,
+  // avg_block_c is the denoised result from C code,
+  // avg_block_sse2 is the denoised result from SSE2 code.
+  DECLARE_ALIGNED_ARRAY(16, uint8_t, sig_block, kNumPixels);
+  DECLARE_ALIGNED_ARRAY(16, uint8_t, mc_avg_block, kNumPixels);
+  DECLARE_ALIGNED_ARRAY(16, uint8_t, avg_block_c, kNumPixels);
+  DECLARE_ALIGNED_ARRAY(16, uint8_t, avg_block_sse2, kNumPixels);
+
+  for (int i = 0; i < count_test_block; ++i) {
+    // Generate random motion magnitude, 20% of which exceed the threshold.
+    uint8_t motion_magnitude_random
+              = rnd.Rand8() % (uint8_t)(MOTION_MAGNITUDE_THRESHOLD * 1.2);
+
+    // Initialize a test block with random number in range [0, 255].
+    for (int j = 0; j < kNumPixels; ++j) {
+      int temp = 0;
+      sig_block[j] = rnd.Rand8();
+      // The pixels in mc_avg_block are generated by adding a random
+      // number in range [-19, 19] to corresponding pixels in sig_block.
+      temp = sig_block[j] + (rnd.Rand8() % 2 == 0? -1 : 1) *
+             (rnd.Rand8()%20);
+      // Clip.
+      mc_avg_block[j] = (temp < 0? 0 : (temp > 255? 255 : temp));
+    }
+
+    ASM_REGISTER_STATE_CHECK(vp9_denoiser_filter_c(sig_block, 64,
+                             mc_avg_block, 64, avg_block_c, 64,
+                             0, bs, motion_magnitude_random));
+
+    ASM_REGISTER_STATE_CHECK(vp9_denoiser_filter_sse2(sig_block, 64,
+                             mc_avg_block, 64, avg_block_sse2, 64,
+                             0, bs, motion_magnitude_random));
+
+    // Test bitexactness.
+    for (int h = 0; h < (4 << b_height_log2_lookup[bs]); ++h) {
+      for (int w = 0; w < (4 << b_width_log2_lookup[bs]); ++w) {
+        EXPECT_EQ(avg_block_c[h * 64 + w], avg_block_sse2[h * 64 + w]);
+      }
+    }
+  }
+}
+
+// Test for all block size.
+INSTANTIATE_TEST_CASE_P(
+    SSE2, VP9DenoiserTest,
+    ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4, BLOCK_8X8,
+                      BLOCK_8X16, BLOCK_16X8, BLOCK_16X16, BLOCK_16X32,
+                      BLOCK_32X16, BLOCK_32X32, BLOCK_32X64, BLOCK_64X32,
+                      BLOCK_64X64));
+}  // namespace
--- a/vp9/common/vp9_common_data.c
+++ b/vp9/common/vp9_common_data.c
@@ -36,7 +36,6 @@
 const int num_pels_log2_lookup[BLOCK_SIZES] =
   {4, 5, 5, 6, 7, 7, 8, 9, 9, 10, 11, 11, 12};
 
-
 const PARTITION_TYPE partition_lookup[][BLOCK_SIZES] = {
   {  // 4X4
     // 4X4, 4X8,8X4,8X8,8X16,16X8,16X16,16X32,32X16,32X32,32X64,64X32,64X64
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -1115,6 +1115,14 @@
 add_proto qw/void vp9_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
 specialize qw/vp9_subtract_block neon/, "$sse2_x86inc";
 
+#
+# Denoiser
+#
+if (vpx_config("CONFIG_VP9_TEMPORAL_DENOISING") eq "yes") {
+  add_proto qw/int vp9_denoiser_filter/, "const uint8_t *sig, int sig_stride, const uint8_t *mc_avg, int mc_avg_stride, uint8_t *avg, int avg_stride, int increase_denoising, BLOCK_SIZE bs, int motion_magnitude";
+  specialize qw/vp9_denoiser_filter sse2/;
+}
+
 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 # the transform coefficients are held in 32-bit
 # values, so the assembler code for  vp9_block_error can no longer be used.
--- a/vp9/encoder/vp9_denoiser.c
+++ b/vp9/encoder/vp9_denoiser.c
@@ -31,9 +31,6 @@
 static void make_grayscale(YV12_BUFFER_CONFIG *yuv);
 #endif
 
-static const int widths[]  = {4, 4, 8, 8,  8, 16, 16, 16, 32, 32, 32, 64, 64};
-static const int heights[] = {4, 8, 4, 8, 16,  8, 16, 32, 16, 32, 64, 32, 64};
-
 static int absdiff_thresh(BLOCK_SIZE bs, int increase_denoising) {
   (void)bs;
   return 3 + (increase_denoising ? 1 : 0);
@@ -52,7 +49,9 @@
 }
 
 static unsigned int sse_thresh(BLOCK_SIZE bs, int increase_denoising) {
-  return widths[bs] * heights[bs] * (increase_denoising ? 60 : 40);
+  return (4 << b_width_log2_lookup[bs]) *
+         (4 << b_height_log2_lookup[bs]) *
+         (increase_denoising ? 60 : 40);
 }
 
 static int sse_diff_thresh(BLOCK_SIZE bs, int increase_denoising,
@@ -61,25 +60,31 @@
       noise_motion_thresh(bs, increase_denoising)) {
     return 0;
   } else {
-    return widths[bs] * heights[bs] * 20;
+    return (4 << b_width_log2_lookup[bs]) *
+           (4 << b_height_log2_lookup[bs]) * 20;
   }
 }
 
-static int total_adj_strong_thresh(BLOCK_SIZE bs, int increase_denoising) {
-  return widths[bs] * heights[bs] * (increase_denoising ? 3 : 2);
+int total_adj_strong_thresh(BLOCK_SIZE bs, int increase_denoising) {
+  return (4 << b_width_log2_lookup[bs]) *
+         (4 << b_height_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
 }
 
 static int total_adj_weak_thresh(BLOCK_SIZE bs, int increase_denoising) {
-  return widths[bs] * heights[bs] * (increase_denoising ? 3 : 2);
+  return (4 << b_width_log2_lookup[bs]) *
+         (4 << b_height_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
 }
 
-static VP9_DENOISER_DECISION denoiser_filter(const uint8_t *sig, int sig_stride,
-                                             const uint8_t *mc_avg,
-                                             int mc_avg_stride,
-                                             uint8_t *avg, int avg_stride,
-                                             int increase_denoising,
-                                             BLOCK_SIZE bs,
-                                             int motion_magnitude) {
+// TODO(jackychen): If increase_denoising is enabled in the future,
+// we might need to update the code for calculating 'total_adj' in
+// case the C code is not bit-exact with corresponding sse2 code.
+int vp9_denoiser_filter_c(const uint8_t *sig, int sig_stride,
+                          const uint8_t *mc_avg,
+                          int mc_avg_stride,
+                          uint8_t *avg, int avg_stride,
+                          int increase_denoising,
+                          BLOCK_SIZE bs,
+                          int motion_magnitude) {
   int r, c;
   const uint8_t *sig_start = sig;
   const uint8_t *mc_avg_start = mc_avg;
@@ -102,8 +107,8 @@
   }
 
   // First attempt to apply a strong temporal denoising filter.
-  for (r = 0; r < heights[bs]; ++r) {
-    for (c = 0; c < widths[bs]; ++c) {
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+    for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) {
       diff = mc_avg[c] - sig[c];
       absdiff = abs(diff);
 
@@ -152,8 +157,8 @@
   mc_avg =  mc_avg_start;
   avg = avg_start;
   sig = sig_start;
-  for (r = 0; r < heights[bs]; ++r) {
-    for (c = 0; c < widths[bs]; ++c) {
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+    for (c = 0; c < (4 << b_width_log2_lookup[bs]); ++c) {
       diff = mc_avg[c] - sig[c];
       adj = abs(diff);
       if (adj > delta) {
@@ -193,8 +198,8 @@
 static void copy_block(uint8_t *dest, int dest_stride,
                        const uint8_t *src, int src_stride, BLOCK_SIZE bs) {
   int r;
-  for (r = 0; r < heights[bs]; ++r) {
-    vpx_memcpy(dest, src, widths[bs]);
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+    vpx_memcpy(dest, src, (4 << b_width_log2_lookup[bs]));
     dest += dest_stride;
     src += src_stride;
   }
@@ -336,10 +341,10 @@
                                          &motion_magnitude);
 
   if (decision == FILTER_BLOCK) {
-    decision = denoiser_filter(src.buf, src.stride,
-                               mc_avg_start, mc_avg.y_stride,
-                               avg_start, avg.y_stride,
-                               0, bs, motion_magnitude);
+    decision = vp9_denoiser_filter(src.buf, src.stride,
+                                 mc_avg_start, mc_avg.y_stride,
+                                 avg_start, avg.y_stride,
+                                 0, bs, motion_magnitude);
   }
 
   if (decision == FILTER_BLOCK) {
--- a/vp9/encoder/vp9_denoiser.h
+++ b/vp9/encoder/vp9_denoiser.h
@@ -55,6 +55,10 @@
 #endif
                        int border);
 
+#if CONFIG_VP9_TEMPORAL_DENOISING
+int total_adj_strong_thresh(BLOCK_SIZE bs, int increase_denoising);
+#endif
+
 void vp9_denoiser_free(VP9_DENOISER *denoiser);
 
 #ifdef __cplusplus
--- /dev/null
+++ b/vp9/encoder/x86/vp9_denoiser_sse2.c
@@ -1,0 +1,474 @@
+/*
+ *  Copyright (c) 2014 The WebM 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 <emmintrin.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+
+#include "vpx_ports/emmintrin_compat.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+#include "vpx_mem/vpx_mem.h"
+
+// Compute the sum of all pixel differences of this MB.
+static INLINE int sum_diff_16x1(__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));
+  int sum_diff = _mm_cvtsi128_si32(hgfedcba);
+  return sum_diff;
+}
+
+// Denoise a 16x1 vector.
+static INLINE __m128i vp9_denoiser_16x1_sse2(const uint8_t *sig,
+                                             const uint8_t *mc_running_avg_y,
+                                             uint8_t *running_avg_y,
+                                             const __m128i k_0,
+                                             const __m128i k_4,
+                                             const __m128i k_8,
+                                             const __m128i k_16,
+                                             const __m128i l3,
+                                             const __m128i l32,
+                                             const __m128i l21,
+                                             __m128i acc_diff) {
+  // Calculate differences
+  const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
+  const __m128i v_mc_running_avg_y = _mm_loadu_si128(
+                                     (__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((__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);
+  return acc_diff;
+}
+
+// Denoise a 16x1 vector with a weaker filter.
+static INLINE __m128i vp9_denoiser_adj_16x1_sse2(const uint8_t *sig,
+                                              const uint8_t *mc_running_avg_y,
+                                              uint8_t *running_avg_y,
+                                              const __m128i k_0,
+                                              const __m128i k_delta,
+                                              __m128i acc_diff) {
+  __m128i v_running_avg_y = _mm_loadu_si128((__m128i *)(&running_avg_y[0]));
+  // Calculate differences.
+  const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
+  const __m128i v_mc_running_avg_y =
+                _mm_loadu_si128((__m128i *)(&mc_running_avg_y[0]));
+  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 delta to get the adjustment.
+  const __m128i adj =
+                _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
+  // Restore the sign and get positive and negative adjustments.
+  __m128i padj, nadj;
+  padj = _mm_andnot_si128(diff_sign, adj);
+  nadj = _mm_and_si128(diff_sign, adj);
+  // Calculate filtered value.
+  v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
+  v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
+  _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+  // Accumulate the adjustments.
+  acc_diff = _mm_subs_epi8(acc_diff, padj);
+  acc_diff = _mm_adds_epi8(acc_diff, nadj);
+  return acc_diff;
+}
+
+static int vp9_denoiser_4xM_sse2(const uint8_t *sig, int sig_stride,
+                                 const uint8_t *mc_running_avg_y,
+                                 int mc_avg_y_stride,
+                                 uint8_t *running_avg_y, int avg_y_stride,
+                                 int increase_denoising,
+                                 BLOCK_SIZE bs,
+                                 int motion_magnitude) {
+  int sum_diff_thresh;
+  int r;
+  int shift_inc  = (increase_denoising &&
+                   motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
+  unsigned char sig_buffer[2][16], mc_running_buffer[2][16],
+                running_buffer[2][16];
+  __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 <= MOTION_MAGNITUDE_THRESHOLD) ?
+                     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);
+  int sum_diff = 0;
+
+  for (r = 0; r < ((4 << b_height_log2_lookup[bs]) >> 2); ++r) {
+    vpx_memcpy(sig_buffer[r], sig, 4);
+    vpx_memcpy(sig_buffer[r] + 4, sig + sig_stride, 4);
+    vpx_memcpy(sig_buffer[r] + 8, sig + sig_stride * 2, 4);
+    vpx_memcpy(sig_buffer[r] + 12, sig + sig_stride * 3, 4);
+    vpx_memcpy(mc_running_buffer[r], mc_running_avg_y, 4);
+    vpx_memcpy(mc_running_buffer[r] + 4, mc_running_avg_y +
+               mc_avg_y_stride, 4);
+    vpx_memcpy(mc_running_buffer[r] + 8, mc_running_avg_y +
+               mc_avg_y_stride * 2, 4);
+    vpx_memcpy(mc_running_buffer[r] + 12, mc_running_avg_y +
+               mc_avg_y_stride * 3, 4);
+    vpx_memcpy(running_buffer[r], running_avg_y, 4);
+    vpx_memcpy(running_buffer[r] + 4, running_avg_y +
+               avg_y_stride, 4);
+    vpx_memcpy(running_buffer[r] + 8, running_avg_y +
+               avg_y_stride * 2, 4);
+    vpx_memcpy(running_buffer[r] + 12, running_avg_y +
+               avg_y_stride * 3, 4);
+    acc_diff = vp9_denoiser_16x1_sse2(sig_buffer[r],
+                                      mc_running_buffer[r],
+                                      running_buffer[r],
+                                      k_0, k_4, k_8, k_16,
+                                      l3, l32, l21, acc_diff);
+    vpx_memcpy(running_avg_y, running_buffer[r], 4);
+    vpx_memcpy(running_avg_y + avg_y_stride, running_buffer[r] + 4, 4);
+    vpx_memcpy(running_avg_y + avg_y_stride * 2,
+               running_buffer[r] + 8, 4);
+    vpx_memcpy(running_avg_y + avg_y_stride * 3,
+               running_buffer[r] + 12, 4);
+    // Update pointers for next iteration.
+    sig += (sig_stride << 2);
+    mc_running_avg_y += (mc_avg_y_stride << 2);
+    running_avg_y += (avg_y_stride << 2);
+  }
+
+  {
+    sum_diff = sum_diff_16x1(acc_diff);
+    sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising),
+      // checK if we can still apply some (weaker) temporal filtering to
+      // this block, that would otherwise not be denoised at all. Simplest
+      // is to apply an additional adjustment to running_avg_y to bring it
+      // closer to sig. The adjustment is capped by a maximum delta, and
+      // chosen such that in most cases the resulting sum_diff will be
+      // within the accceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over the
+      // threshold.
+      int delta = ((abs(sum_diff) - sum_diff_thresh)
+                  >> num_pels_log2_lookup[bs]) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        const __m128i k_delta = _mm_set1_epi8(delta);
+        running_avg_y -= avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        sum_diff = 0;
+        for (r = 0; r < ((4 << b_height_log2_lookup[bs]) >> 2); ++r) {
+          acc_diff = vp9_denoiser_adj_16x1_sse2(
+                             sig_buffer[r], mc_running_buffer[r],
+                             running_buffer[r], k_0, k_delta,
+                             acc_diff);
+          vpx_memcpy(running_avg_y, running_buffer[r], 4);
+          vpx_memcpy(running_avg_y + avg_y_stride, running_buffer[r] + 4, 4);
+          vpx_memcpy(running_avg_y + avg_y_stride * 2,
+                     running_buffer[r] + 8, 4);
+          vpx_memcpy(running_avg_y + avg_y_stride * 3,
+                     running_buffer[r] + 12, 4);
+          // Update pointers for next iteration.
+          running_avg_y += (avg_y_stride << 2);
+        }
+        sum_diff = sum_diff_16x1(acc_diff);
+        if (abs(sum_diff) > sum_diff_thresh) {
+          return COPY_BLOCK;
+        }
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+  }
+  return FILTER_BLOCK;
+}
+
+static int vp9_denoiser_8xM_sse2(const uint8_t *sig, int sig_stride,
+                                 const uint8_t *mc_running_avg_y,
+                                 int mc_avg_y_stride,
+                                 uint8_t *running_avg_y, int avg_y_stride,
+                                 int increase_denoising,
+                                 BLOCK_SIZE bs,
+                                 int motion_magnitude) {
+  int sum_diff_thresh;
+  int r;
+  int shift_inc  = (increase_denoising &&
+                  motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
+  unsigned char sig_buffer[8][16], mc_running_buffer[8][16],
+                running_buffer[8][16];
+  __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 <= MOTION_MAGNITUDE_THRESHOLD) ?
+                      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);
+  int sum_diff = 0;
+
+  for (r = 0; r < ((4 << b_height_log2_lookup[bs]) >> 1); ++r) {
+    vpx_memcpy(sig_buffer[r], sig, 8);
+    vpx_memcpy(sig_buffer[r] + 8, sig + sig_stride, 8);
+    vpx_memcpy(mc_running_buffer[r], mc_running_avg_y, 8);
+    vpx_memcpy(mc_running_buffer[r] + 8, mc_running_avg_y +
+               mc_avg_y_stride, 8);
+    vpx_memcpy(running_buffer[r], running_avg_y, 8);
+    vpx_memcpy(running_buffer[r] + 8, running_avg_y +
+               avg_y_stride, 8);
+    acc_diff = vp9_denoiser_16x1_sse2(sig_buffer[r],
+                                      mc_running_buffer[r],
+                                      running_buffer[r],
+                                      k_0, k_4, k_8, k_16,
+                                      l3, l32, l21, acc_diff);
+    vpx_memcpy(running_avg_y, running_buffer[r], 8);
+    vpx_memcpy(running_avg_y + avg_y_stride, running_buffer[r] + 8, 8);
+    // Update pointers for next iteration.
+    sig += (sig_stride << 1);
+    mc_running_avg_y += (mc_avg_y_stride << 1);
+    running_avg_y += (avg_y_stride << 1);
+  }
+
+  {
+    sum_diff = sum_diff_16x1(acc_diff);
+    sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising),
+      // checK if we can still apply some (weaker) temporal filtering to
+      // this block, that would otherwise not be denoised at all. Simplest
+      // is to apply an additional adjustment to running_avg_y to bring it
+      // closer to sig. The adjustment is capped by a maximum delta, and
+      // chosen such that in most cases the resulting sum_diff will be
+      // within the accceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over the
+      // threshold.
+      int delta = ((abs(sum_diff) - sum_diff_thresh)
+                  >> num_pels_log2_lookup[bs]) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        const __m128i k_delta = _mm_set1_epi8(delta);
+        running_avg_y -= avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        for (r = 0; r < ((4 << b_height_log2_lookup[bs]) >> 1); ++r) {
+          acc_diff = vp9_denoiser_adj_16x1_sse2(
+                           sig_buffer[r], mc_running_buffer[r],
+                           running_buffer[r], k_0, k_delta,
+                           acc_diff);
+          vpx_memcpy(running_avg_y, running_buffer[r], 8);
+          vpx_memcpy(running_avg_y + avg_y_stride, running_buffer[r] + 8, 8);
+          // Update pointers for next iteration.
+          running_avg_y += (avg_y_stride << 1);
+        }
+        sum_diff = sum_diff_16x1(acc_diff);
+        if (abs(sum_diff) > sum_diff_thresh) {
+          return COPY_BLOCK;
+        }
+      } else {
+          return COPY_BLOCK;
+      }
+    }
+  }
+  return FILTER_BLOCK;
+}
+
+static int vp9_denoiser_64_32_16xM_sse2(const uint8_t *sig, int sig_stride,
+                                        const uint8_t *mc_running_avg_y,
+                                        int mc_avg_y_stride,
+                                        uint8_t *running_avg_y,
+                                        int avg_y_stride,
+                                        int increase_denoising, BLOCK_SIZE bs,
+                                        int motion_magnitude) {
+  int sum_diff_thresh;
+  int r, c;
+  int shift_inc  = (increase_denoising &&
+                   motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 1 : 0;
+  __m128i acc_diff[4][4];
+  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 <= MOTION_MAGNITUDE_THRESHOLD) ?
+                     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);
+  int sum_diff = 0;
+
+  for (int i = 0; i < 4; i++) {
+    for (int j = 0; j < 4; j++) {
+      acc_diff[i][j] = _mm_setzero_si128();
+    }
+  }
+
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); r++) {
+    for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+      acc_diff[c>>4][r>>4] = vp9_denoiser_16x1_sse2(
+                               sig, mc_running_avg_y,
+                               running_avg_y,
+                               k_0, k_4, k_8, k_16,
+                               l3, l32, l21, acc_diff[c>>4][r>>4]);
+      // Update pointers for next iteration.
+      sig += 16;
+      mc_running_avg_y += 16;
+      running_avg_y += 16;
+    }
+
+    if ((r + 1) % 16 == 0 || (bs == BLOCK_16X8 && r == 7)) {
+      for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+        sum_diff += sum_diff_16x1(acc_diff[c>>4][r>>4]);
+      }
+    }
+
+    // Update pointers for next iteration.
+    sig = sig - 16 * ((4 << b_width_log2_lookup[bs]) >> 4) + sig_stride;
+    mc_running_avg_y = mc_running_avg_y -
+                      16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                      mc_avg_y_stride;
+    running_avg_y = running_avg_y -
+                    16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                    avg_y_stride;
+  }
+
+  {
+    sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+    if (abs(sum_diff) > sum_diff_thresh) {
+      int delta = ((abs(sum_diff) - sum_diff_thresh)
+                  >> num_pels_log2_lookup[bs]) + 1;
+
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        const __m128i k_delta = _mm_set1_epi8(delta);
+        sig -= sig_stride * (4 << b_height_log2_lookup[bs]);
+        mc_running_avg_y -= mc_avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        running_avg_y -= avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        sum_diff = 0;
+        for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+          for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+            acc_diff[c>>4][r>>4] = vp9_denoiser_adj_16x1_sse2(
+                                             sig, mc_running_avg_y,
+                                             running_avg_y, k_0,
+                                             k_delta, acc_diff[c>>4][r>>4]);
+            // Update pointers for next iteration.
+            sig += 16;
+            mc_running_avg_y += 16;
+            running_avg_y += 16;
+          }
+
+          if ((r + 1) % 16 == 0 || (bs == BLOCK_16X8 && r == 7)) {
+            for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+              sum_diff += sum_diff_16x1(acc_diff[c>>4][r>>4]);
+            }
+          }
+          sig = sig - 16 * ((4 << b_width_log2_lookup[bs]) >> 4) + sig_stride;
+          mc_running_avg_y = mc_running_avg_y -
+                             16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                             mc_avg_y_stride;
+          running_avg_y = running_avg_y -
+                          16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                          avg_y_stride;
+        }
+        if (abs(sum_diff) > sum_diff_thresh) {
+          return COPY_BLOCK;
+        }
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+  }
+  return FILTER_BLOCK;
+}
+
+int vp9_denoiser_filter_sse2(const uint8_t *sig, int sig_stride,
+                             const uint8_t *mc_avg,
+                             int mc_avg_stride,
+                             uint8_t *avg, int avg_stride,
+                             int increase_denoising,
+                             BLOCK_SIZE bs,
+                             int motion_magnitude) {
+  if (bs == BLOCK_4X4 || bs == BLOCK_4X8) {
+    return vp9_denoiser_4xM_sse2(sig, sig_stride,
+                                 mc_avg, mc_avg_stride,
+                                 avg, avg_stride,
+                                 increase_denoising,
+                                 bs, motion_magnitude);
+  } else if (bs == BLOCK_8X4 || bs == BLOCK_8X8 || bs == BLOCK_8X16) {
+    return vp9_denoiser_8xM_sse2(sig, sig_stride,
+                                 mc_avg, mc_avg_stride,
+                                 avg, avg_stride,
+                                 increase_denoising,
+                                 bs, motion_magnitude);
+  } else if (bs == BLOCK_16X8 || bs == BLOCK_16X16 || bs == BLOCK_16X32 ||
+             bs == BLOCK_32X16|| bs == BLOCK_32X32 || bs == BLOCK_32X64 ||
+             bs == BLOCK_64X32 || bs == BLOCK_64X64) {
+    return vp9_denoiser_64_32_16xM_sse2(sig, sig_stride,
+                                        mc_avg, mc_avg_stride,
+                                        avg, avg_stride,
+                                        increase_denoising,
+                                        bs, motion_magnitude);
+  } else {
+    return COPY_BLOCK;
+  }
+}
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -120,6 +120,10 @@
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_dct_sse2.c
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_dct32x32_sse2.c
 
+ifeq ($(CONFIG_VP9_TEMPORAL_DENOISING),yes)
+VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_denoiser_sse2.c
+endif
+
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_dct32x32_avx2.c
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_dct_avx2.c
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_error_intrin_avx2.c