vp8: Add temporal denoising for UV-channel.

C version and sse2 version, and off by default.
For the test clip used, the sse2 performance improved by ~5.6%

Change-Id: Ic2d815968849db51b9d62085d7a490d0e01574f6

4 files changed, 416 insertions, 27 deletions

diff --git a/vp8/common/rtcd_defs.pl b/vp8/common/rtcd_defs.pl
index cbfd76a8d..ba2693968 100644
--- a/vp8/common/rtcd_defs.pl
+++ b/vp8/common/rtcd_defs.pl
@@ -554,6 +554,9 @@ $vp8_yv12_copy_partial_frame_neon_asm=vp8_yv12_copy_partial_frame_neon;
 if (vpx_config("CONFIG_TEMPORAL_DENOISING") eq "yes") {
     add_proto qw/int vp8_denoiser_filter/, "unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
     specialize qw/vp8_denoiser_filter sse2 neon/;
+    add_proto qw/int vp8_denoiser_filter_uv/, "unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising";
+    specialize qw/vp8_denoiser_filter_uv sse2/;
+
 }
 
 # End of encoder only functions
diff --git a/vp8/encoder/denoising.c b/vp8/encoder/denoising.c
index 0f2e5f17b..9ad411366 100644
--- a/vp8/encoder/denoising.c
+++ b/vp8/encoder/denoising.c
@@ -191,6 +191,148 @@ int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride,
     return FILTER_BLOCK;
 }
 
+int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg_uv,
+                             int mc_avg_uv_stride,
+                             unsigned char *running_avg_uv,
+                             int avg_uv_stride,
+                             unsigned char *sig,
+                             int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    unsigned char *running_avg_uv_start = running_avg_uv;
+    unsigned char *sig_start = sig;
+    int sum_diff_thresh;
+    int r, c;
+    int sum_diff = 0;
+    int sum_block = 0;
+    int adj_val[3] = {3, 4, 6};
+    int shift_inc1 = 0;
+    int shift_inc2 = 1;
+    /* If motion_magnitude is small, making the denoiser more aggressive by
+     * increasing the adjustment for each level. Add another increment for
+     * blocks that are labeled for increase denoising. */
+    if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) {
+      if (increase_denoising) {
+        shift_inc1 = 1;
+        shift_inc2 = 2;
+      }
+      adj_val[0] += shift_inc2;
+      adj_val[1] += shift_inc2;
+      adj_val[2] += shift_inc2;
+    }
+
+    // Avoid denoising color signal if its close to average level.
+    for (r = 0; r < 8; ++r) {
+      for (c = 0; c < 8; ++c) {
+        sum_block += sig[c];
+      }
+      sig += sig_stride;
+    }
+    if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+      return COPY_BLOCK;
+    }
+
+    sig -= sig_stride * 8;
+    for (r = 0; r < 8; ++r) {
+      for (c = 0; c < 8; ++c) {
+        int diff = 0;
+        int adjustment = 0;
+        int absdiff = 0;
+
+        diff = mc_running_avg_uv[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_uv[c] = mc_running_avg_uv[c];
+          sum_diff += diff;
+        } else {
+          if (absdiff >= 4 && 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_uv[c] = 255;
+            else
+              running_avg_uv[c] = sig[c] + adjustment;
+            sum_diff += adjustment;
+          } else {
+            if ((sig[c] - adjustment) < 0)
+              running_avg_uv[c] = 0;
+            else
+              running_avg_uv[c] = sig[c] - adjustment;
+            sum_diff -= adjustment;
+          }
+        }
+      }
+      /* Update pointers for next iteration. */
+      sig += sig_stride;
+      mc_running_avg_uv += mc_avg_uv_stride;
+      running_avg_uv += avg_uv_stride;
+    }
+
+    sum_diff_thresh= SUM_DIFF_THRESHOLD_UV;
+    if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+    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 threshold.
+      int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        sig -= sig_stride * 8;
+        mc_running_avg_uv -= mc_avg_uv_stride * 8;
+        running_avg_uv -= avg_uv_stride * 8;
+        for (r = 0; r < 8; ++r) {
+          for (c = 0; c < 8; ++c) {
+            int diff = mc_running_avg_uv[c] - sig[c];
+            int adjustment = abs(diff);
+            if (adjustment > delta)
+              adjustment = delta;
+            if (diff > 0) {
+              // Bring denoised signal down.
+              if (running_avg_uv[c] - adjustment < 0)
+                running_avg_uv[c] = 0;
+              else
+                running_avg_uv[c] = running_avg_uv[c] - adjustment;
+              sum_diff -= adjustment;
+            } else if (diff < 0) {
+              // Bring denoised signal up.
+              if (running_avg_uv[c] + adjustment > 255)
+                running_avg_uv[c] = 255;
+              else
+                running_avg_uv[c] = running_avg_uv[c] + adjustment;
+              sum_diff += adjustment;
+            }
+          }
+          // TODO(marpan): Check here if abs(sum_diff) has gone below the
+          // threshold sum_diff_thresh, and if so, we can exit the row loop.
+          sig += sig_stride;
+          mc_running_avg_uv += mc_avg_uv_stride;
+          running_avg_uv += avg_uv_stride;
+        }
+        if (abs(sum_diff) > sum_diff_thresh)
+          return COPY_BLOCK;
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+
+    vp8_copy_mem8x8(running_avg_uv_start, avg_uv_stride, sig_start,
+                    sig_stride);
+    return FILTER_BLOCK;
+}
+
 int vp8_denoiser_allocate(VP8_DENOISER *denoiser, int width, int height,
                           int num_mb_rows, int num_mb_cols)
 {
@@ -260,6 +402,8 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
     unsigned int motion_magnitude2;
     unsigned int sse_thresh;
     int sse_diff_thresh = 0;
+    // Denoise the UV channel.
+    int apply_color_denoise = 0;
     // Spatial loop filter: only applied selectively based on
     // temporal filter state of block relative to top/left neighbors.
     int apply_spatial_loop_filter = 1;
@@ -267,6 +411,8 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
     MV_REFERENCE_FRAME zero_frame = x->best_zeromv_reference_frame;
 
     enum vp8_denoiser_decision decision = FILTER_BLOCK;
+    enum vp8_denoiser_decision decision_u = FILTER_BLOCK;
+    enum vp8_denoiser_decision decision_v = FILTER_BLOCK;
 
     if (zero_frame)
     {
@@ -376,11 +522,37 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
 
         /* Filter. */
         decision = vp8_denoiser_filter(mc_running_avg_y, mc_avg_y_stride,
-                                         running_avg_y, avg_y_stride,
-                                         x->thismb, 16, motion_magnitude2,
-                                         x->increase_denoising);
+                                       running_avg_y, avg_y_stride,
+                                       x->thismb, 16, motion_magnitude2,
+                                       x->increase_denoising);
         denoiser->denoise_state[block_index] = motion_magnitude2 > 0 ?
             kFilterNonZeroMV : kFilterZeroMV;
+        // Only denoise UV for zero motion, and if y channel was denoised.
+        if (apply_color_denoise &&
+            motion_magnitude2 == 0 &&
+            decision == FILTER_BLOCK) {
+          unsigned char *mc_running_avg_u =
+              denoiser->yv12_mc_running_avg.u_buffer + recon_uvoffset;
+          unsigned char *running_avg_u =
+              denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset;
+          unsigned char *mc_running_avg_v =
+              denoiser->yv12_mc_running_avg.v_buffer + recon_uvoffset;
+          unsigned char *running_avg_v =
+              denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset;
+          int mc_avg_uv_stride = denoiser->yv12_mc_running_avg.uv_stride;
+          int avg_uv_stride = denoiser->yv12_running_avg[INTRA_FRAME].uv_stride;
+          int signal_stride = x->block[16].src_stride;
+          decision_u =
+              vp8_denoiser_filter_uv(mc_running_avg_u, mc_avg_uv_stride,
+                                      running_avg_u, avg_uv_stride,
+                                      x->block[16].src + *x->block[16].base_src,
+                                      signal_stride, motion_magnitude2, 0);
+          decision_v =
+              vp8_denoiser_filter_uv(mc_running_avg_v, mc_avg_uv_stride,
+                                      running_avg_v, avg_uv_stride,
+                                      x->block[20].src + *x->block[20].base_src,
+                                      signal_stride, motion_magnitude2, 0);
+        }
     }
     if (decision == COPY_BLOCK)
     {
@@ -393,7 +565,21 @@ void vp8_denoiser_denoise_mb(VP8_DENOISER *denoiser,
                 denoiser->yv12_running_avg[INTRA_FRAME].y_stride);
         denoiser->denoise_state[block_index] = kNoFilter;
     }
-    // Option to selectively deblock the denoised signal.
+    if (apply_color_denoise) {
+      if (decision_u == COPY_BLOCK) {
+        vp8_copy_mem8x8(
+            x->block[16].src + *x->block[16].base_src, x->block[16].src_stride,
+            denoiser->yv12_running_avg[INTRA_FRAME].u_buffer + recon_uvoffset,
+            denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
+      }
+      if (decision_v == COPY_BLOCK) {
+        vp8_copy_mem8x8(
+            x->block[20].src + *x->block[20].base_src, x->block[16].src_stride,
+            denoiser->yv12_running_avg[INTRA_FRAME].v_buffer + recon_uvoffset,
+            denoiser->yv12_running_avg[INTRA_FRAME].uv_stride);
+      }
+    }
+    // Option to selectively deblock the denoised signal, for y channel only.
     if (apply_spatial_loop_filter) {
       loop_filter_info lfi;
       int apply_filter_col = 0;
diff --git a/vp8/encoder/denoising.h b/vp8/encoder/denoising.h
index 6db0785a0..8f1bfa51d 100644
--- a/vp8/encoder/denoising.h
+++ b/vp8/encoder/denoising.h
@@ -22,6 +22,11 @@ extern "C" {
 #define SUM_DIFF_THRESHOLD_HIGH (16 * 16 * 3)
 #define MOTION_MAGNITUDE_THRESHOLD (8*3)
 
+#define SUM_DIFF_THRESHOLD_UV (96)   // (8 * 8 * 1.5)
+#define SUM_DIFF_THRESHOLD_HIGH_UV (8 * 8 * 2)
+#define SUM_DIFF_FROM_AVG_THRESH_UV (8 * 8 * 4)
+#define MOTION_MAGNITUDE_THRESHOLD_UV (8*3)
+
 enum vp8_denoiser_decision
 {
   COPY_BLOCK,
diff --git a/vp8/encoder/x86/denoising_sse2.c b/vp8/encoder/x86/denoising_sse2.c
index ff439dd64..4d1452cc1 100644
--- a/vp8/encoder/x86/denoising_sse2.c
+++ b/vp8/encoder/x86/denoising_sse2.c
@@ -17,10 +17,23 @@
 #include <emmintrin.h>
 #include "vpx_ports/emmintrin_compat.h"
 
-union sum_union {
-    __m128i v;
-    signed char e[16];
-};
+/* Compute the sum of all pixel differences of this MB. */
+static inline unsigned int abs_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));
+  unsigned int sum_diff = _mm_cvtsi128_si32(hgfedcba);
+
+  return abs(sum_diff);
+}
 
 int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
                              int mc_avg_y_stride,
@@ -103,16 +116,10 @@ int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
 
     {
         /* Compute the sum of all pixel differences of this MB. */
-        union sum_union s;
-        int sum_diff = 0;
-        s.v = acc_diff;
-        sum_diff = s.e[0] + s.e[1] + s.e[2] + s.e[3] + s.e[4] + s.e[5]
-                 + s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11]
-                 + s.e[12] + s.e[13] + s.e[14] + s.e[15];
-
+        unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
         sum_diff_thresh = SUM_DIFF_THRESHOLD;
         if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
-        if (abs(sum_diff) > sum_diff_thresh) {
+        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
@@ -123,7 +130,7 @@ int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
 
           // The delta is set by the excess of absolute pixel diff over the
           // threshold.
-          int delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
+          int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
           // Only apply the adjustment for max delta up to 3.
           if (delta < 4) {
             const __m128i k_delta = _mm_set1_epi8(delta);
@@ -162,16 +169,9 @@ int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
              mc_running_avg_y += mc_avg_y_stride;
              running_avg_y += avg_y_stride;
             }
-            {
-              // Update the sum of all pixel differences of this MB.
-              union sum_union s;
-              s.v = acc_diff;
-              sum_diff = s.e[0] + s.e[1] + s.e[2] + s.e[3] + s.e[4] + s.e[5]
-                       + s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11]
-                       + s.e[12] + s.e[13] + s.e[14] + s.e[15];
-              if (abs(sum_diff) > sum_diff_thresh) {
-                return COPY_BLOCK;
-              }
+            abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+            if (abs_sum_diff > sum_diff_thresh) {
+              return COPY_BLOCK;
             }
           } else {
             return COPY_BLOCK;
@@ -182,3 +182,198 @@ int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
     vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
     return FILTER_BLOCK;
 }
+
+int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg,
+                             int mc_avg_stride,
+                             unsigned char *running_avg, int avg_stride,
+                             unsigned char *sig, int sig_stride,
+                             unsigned int motion_magnitude,
+                             int increase_denoising) {
+    unsigned char *running_avg_start = running_avg;
+    unsigned char *sig_start = sig;
+    int sum_diff_thresh;
+    int r;
+    int shift_inc  = (increase_denoising &&
+        motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 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 <= MOTION_MAGNITUDE_THRESHOLD_UV) ?
+                        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);
+
+    {
+      const __m128i k_1 = _mm_set1_epi16(1);
+      __m128i vec_sum_block = _mm_setzero_si128();
+
+      // Avoid denoising color signal if its close to average level.
+      for (r = 0; r < 8; ++r) {
+        const __m128i v_sig = _mm_loadl_epi64((__m128i *)(&sig[0]));
+        const __m128i v_sig_unpack = _mm_unpacklo_epi8(v_sig, k_0);
+        vec_sum_block = _mm_add_epi16(vec_sum_block, v_sig_unpack);
+        sig += sig_stride;
+      }
+      sig -= sig_stride * 8;
+      {
+        const __m128i hg_fe_dc_ba = _mm_madd_epi16(vec_sum_block, 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));
+        const int sum_block = _mm_cvtsi128_si32(hgfedcba);
+        if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
+          return COPY_BLOCK;
+        }
+      }
+    }
+
+    for (r = 0; r < 4; ++r) {
+        /* Calculate differences */
+        const __m128i v_sig_low = _mm_castpd_si128(
+            _mm_load_sd((double *)(&sig[0])));
+        const __m128i v_sig = _mm_castpd_si128(
+            _mm_loadh_pd(_mm_castsi128_pd(v_sig_low),
+                         (double *)(&sig[sig_stride])));
+        const __m128i v_mc_running_avg_low = _mm_castpd_si128(
+            _mm_load_sd((double *)(&mc_running_avg[0])));
+        const __m128i v_mc_running_avg = _mm_castpd_si128(
+            _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
+                         (double *)(&mc_running_avg[mc_avg_stride])));
+        const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
+        const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
+        /* 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;
+        __m128i v_running_avg;
+
+        /* 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 = _mm_adds_epu8(v_sig, padj);
+        v_running_avg = _mm_subs_epu8(v_running_avg, nadj);
+
+        _mm_storel_pd((double *)&running_avg[0],
+                      _mm_castsi128_pd(v_running_avg));
+        _mm_storeh_pd((double *)&running_avg[avg_stride],
+                      _mm_castsi128_pd(v_running_avg));
+
+        /* 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 * 2;
+        mc_running_avg += mc_avg_stride * 2;
+        running_avg += avg_stride * 2;
+    }
+
+    {
+        unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+        sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
+        if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
+        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) >> 8) + 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 * 8;
+            mc_running_avg -= mc_avg_stride * 8;
+            running_avg -= avg_stride * 8;
+            for (r = 0; r < 4; ++r) {
+              // Calculate differences.
+              const __m128i v_sig_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&sig[0])));
+              const __m128i v_sig = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_sig_low),
+                               (double *)(&sig[sig_stride])));
+              const __m128i v_mc_running_avg_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&mc_running_avg[0])));
+              const __m128i v_mc_running_avg = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
+                               (double *)(&mc_running_avg[mc_avg_stride])));
+              const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
+              const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
+              // 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;
+              const __m128i v_running_avg_low = _mm_castpd_si128(
+                  _mm_load_sd((double *)(&running_avg[0])));
+              __m128i v_running_avg = _mm_castpd_si128(
+                  _mm_loadh_pd(_mm_castsi128_pd(v_running_avg_low),
+                               (double *)(&running_avg[avg_stride])));
+              padj = _mm_andnot_si128(diff_sign, adj);
+              nadj = _mm_and_si128(diff_sign, adj);
+              // Calculate filtered value.
+              v_running_avg = _mm_subs_epu8(v_running_avg, padj);
+              v_running_avg = _mm_adds_epu8(v_running_avg, nadj);
+
+              _mm_storel_pd((double *)&running_avg[0],
+                            _mm_castsi128_pd(v_running_avg));
+              _mm_storeh_pd((double *)&running_avg[avg_stride],
+                            _mm_castsi128_pd(v_running_avg));
+
+             // Accumulate the adjustments.
+             acc_diff = _mm_subs_epi8(acc_diff, padj);
+             acc_diff = _mm_adds_epi8(acc_diff, nadj);
+
+             // Update pointers for next iteration.
+             sig += sig_stride * 2;
+             mc_running_avg += mc_avg_stride * 2;
+             running_avg += avg_stride * 2;
+            }
+            abs_sum_diff = abs_sum_diff_16x1(acc_diff);
+            if (abs_sum_diff > sum_diff_thresh) {
+              return COPY_BLOCK;
+            }
+          } else {
+            return COPY_BLOCK;
+          }
+        }
+    }
+
+    vp8_copy_mem8x8(running_avg_start, avg_stride, sig_start, sig_stride);
+    return FILTER_BLOCK;
+}