commit a0943783526e9b2c429c02ad1597fa2d6bfbd9fb
author Debargha Mukherjee <debargha@google.com> Sat Nov 10 16:17:55 2018 -0800
committer Debargha Mukherjee <debargha@google.com> Mon Nov 12 17:27:26 2018 +0000
tree d629bdcd17dbd5d7b27a6cb34c9efcd5838aa56e
parent d9cdf8f768b248fdb2378f252c73c441fd8def0b

Adds speed feature to prune wedge search

Adds a speed feature to disable wedge search if the difference
between the two best single predictors so far is small. The
threshold is tighter for modes with newmv than without.

About 3-4% speed-up with 0.03 coding efficiency loss for lowres
and midres with 30 frames in end-usage=q mode.

STATS_CHANGED

Change-Id: I038750a5a1b9ad1e11ed6e40dde1ea9bee882f68

av1/encoder/rdopt.c

diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index 5b7747e..cd49118 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -7894,6 +7894,19 @@
                                          preds1, residual1, diff10, strides);
     *calc_pred_masked_compound = 0;
   }
+  if (cpi->sf.prune_wedge_pred_diff_based && compound_type == COMPOUND_WEDGE) {
+    unsigned int sse;
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+      (void)cpi->fn_ptr[bsize].vf(CONVERT_TO_BYTEPTR(*preds0), *strides,
+                                  CONVERT_TO_BYTEPTR(*preds1), *strides, &sse);
+    else
+      (void)cpi->fn_ptr[bsize].vf(*preds0, *strides, *preds1, *strides, &sse);
+    const unsigned int mse =
+        ROUND_POWER_OF_TWO(sse, num_pels_log2_lookup[bsize]);
+    // If two predictors are very similar, skip wedge compound mode search
+    if (mse < 8 || (!have_newmv_in_inter_mode(this_mode) && mse < 64))
+      return INT64_MAX;
+  }
 
   best_rd_cur =
       pick_interinter_mask(cpi, x, bsize, *preds0, *preds1, residual1, diff10);

av1/encoder/speed_features.c

diff --git a/av1/encoder/speed_features.c b/av1/encoder/speed_features.c
index f3c5104..48abe7c 100644
--- a/av1/encoder/speed_features.c
+++ b/av1/encoder/speed_features.c
@@ -199,6 +199,7 @@
   sf->intra_tx_size_search_init_depth_sqr = 1;
   sf->intra_angle_estimation = 1;
   sf->selective_ref_frame = 1;
+  sf->prune_wedge_pred_diff_based = 1;
 
   if (speed >= 1) {
     sf->gm_erroradv_type = GM_ERRORADV_TR_1;
@@ -443,6 +444,7 @@
   sf->use_accurate_subpel_search = USE_8_TAPS;
   sf->disable_wedge_search_var_thresh = 0;
   sf->fast_wedge_sign_estimate = 0;
+  sf->prune_wedge_pred_diff_based = 0;
   sf->drop_ref = 0;
   sf->skip_intra_in_interframe = 1;
   sf->txb_split_cap = 1;

av1/encoder/speed_features.h

diff --git a/av1/encoder/speed_features.h b/av1/encoder/speed_features.h
index 8f0f9a6..7041eab 100644
--- a/av1/encoder/speed_features.h
+++ b/av1/encoder/speed_features.h
@@ -446,6 +446,9 @@
   // Whether fast wedge sign estimate is used
   int fast_wedge_sign_estimate;
 
+  // Whether to prune wedge search based on predictor difference
+  int prune_wedge_pred_diff_based;
+
   // These bit masks allow you to enable or disable intra modes for each
   // transform size separately.
   int intra_y_mode_mask[TX_SIZES];