Neon match to vp8 temporal denoiser fix

Now match the "C" version of "Fix to reduce block
artifacts from vp8 temporal denoiser."
(see change id Id9b56e59e33f3c22e79d2f89f763bdde246fdf3f)

Change-Id: I99e569bb6af4ae3532621127e12bf917a48ba08e

1 files changed, 82 insertions, 5 deletions

diff --git a/vp8/encoder/arm/neon/denoising_neon.c b/vp8/encoder/arm/neon/denoising_neon.c
index b8e403419..78cc6fa37 100644
--- a/vp8/encoder/arm/neon/denoising_neon.c
+++ b/vp8/encoder/arm/neon/denoising_neon.c
@@ -68,8 +68,8 @@ int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y,
     int64x2_t v_sum_diff_total = vdupq_n_s64(0);
 
     /* Go over lines. */
-    int i;
-    for (i = 0; i < 16; ++i) {
+    int r;
+    for (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);
@@ -145,14 +145,91 @@ int vp8_denoiser_filter_neon(unsigned char *mc_running_avg_y,
 
     /* Too much adjustments => copy block. */
     {
-        const int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+        int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
                                       vget_low_s64(v_sum_diff_total));
-        const int s0 = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+        int sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
         int sum_diff_thresh = SUM_DIFF_THRESHOLD;
 
         if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
-        if (s0 > sum_diff_thresh)
+        if (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 = ((sum_diff - sum_diff_thresh) >> 8) + 1;
+          // Only apply the adjustment for max delta up to 3.
+          if (delta < 4) {
+            const uint8x16_t k_delta = vmovq_n_u8(delta);
+            sig -= sig_stride * 16;
+            mc_running_avg_y -= mc_running_avg_y_stride * 16;
+            running_avg_y -= running_avg_y_stride * 16;
+            for (r = 0; r < 16; ++r) {
+              uint8x16_t v_running_avg_y = vld1q_u8(running_avg_y);
+              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);
+              // Clamp absolute difference to delta to get the adjustment.
+              const uint8x16_t v_abs_adjustment =
+                  vminq_u8(v_abs_diff, (k_delta));
+
+              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);
+
+              v_running_avg_y = vqsubq_u8(v_running_avg_y, v_pos_adjustment);
+              v_running_avg_y = vqaddq_u8(v_running_avg_y, v_neg_adjustment);
+
+              /* Store results. */
+              vst1q_u8(running_avg_y, v_running_avg_y);
+
+              {
+                  const int8x16_t v_sum_diff =
+                      vqsubq_s8(vreinterpretq_s8_u8(v_neg_adjustment),
+                                vreinterpretq_s8_u8(v_pos_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;
+            }
+            {
+              // Update the sum of all pixel differences of this MB.
+              x = vqadd_s64(vget_high_s64(v_sum_diff_total),
+                            vget_low_s64(v_sum_diff_total));
+              sum_diff = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
+
+              if (sum_diff > sum_diff_thresh) {
+                return COPY_BLOCK;
+              }
+            }
+          } else {
             return COPY_BLOCK;
+          }
+        }
     }
 
     /* Tell above level that block was filtered. */