Enable chessboard inter prediction filter type search

This commit enables a chessboard pattern prediction filter type
search scheme for rate-distortion optimization speed-up. For the
inferred motion vector modes, the encoder can re-use its above/left
neighbor blocks' prediction filter type and skip a full test on
all possible filter types. Such operation is turned on/off
alternatively in a chessboard manner.

It is turned on in speed 3. For test clip pedestrian 1080p, the
runtime is reduced from 231500 ms -> 221700 ms. The compression
performance is changed:
derf:  -0.147%
yt:    -0.134%
hd:    -0.079%
stdhd: -0.220%

Change-Id: I1912f278e7576c2dc632688e3ad7a257410c605a

3 files changed, 18 insertions, 3 deletions

diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index cdc79345b..03a3c36f4 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -2040,7 +2040,18 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   uint8_t *orig_dst[MAX_MB_PLANE];
   int orig_dst_stride[MAX_MB_PLANE];
   int rs = 0;
-  INTERP_FILTER best_filter = cm->interp_filter;
+  INTERP_FILTER best_filter = SWITCHABLE;
+
+  int bsl = mi_width_log2_lookup[bsize];
+  int pred_filter_search = cpi->sf.cb_pred_filter_search ?
+      (((mi_row + mi_col) >> bsl)) & 0x01 : 0;
+
+  if (pred_filter_search && this_mode != NEWMV) {
+    if (xd->up_available)
+      best_filter = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
+    else if (xd->left_available)
+      best_filter = xd->mi[-1]->mbmi.interp_filter;
+  }
 
   if (is_comp_pred) {
     if (frame_mv[refs[0]].as_int == INVALID_MV ||
@@ -2124,10 +2135,9 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
     rd_opt->filter_cache[i] = INT64_MAX;
 
   if (cm->interp_filter != BILINEAR) {
-    best_filter = EIGHTTAP;
     if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
       best_filter = EIGHTTAP;
-    } else {
+    } else if (best_filter == SWITCHABLE) {
       int newbest;
       int tmp_rate_sum = 0;
       int64_t tmp_dist_sum = 0;
diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c
index 8fb4aa566..7315dd454 100644
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -116,6 +116,7 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
       sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
 
     sf->adaptive_pred_interp_filter = 0;
+    sf->cb_pred_filter_search = 1;
 
     sf->lf_motion_threshold = LOW_MOTION_THRESHOLD;
     sf->last_partitioning_redo_frequency = 3;
@@ -332,6 +333,7 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
   sf->use_lp32x32fdct = 0;
   sf->adaptive_motion_search = 0;
   sf->adaptive_pred_interp_filter = 0;
+  sf->cb_pred_filter_search = 0;
   sf->use_quant_fp = 0;
   sf->reference_masking = 0;
   sf->partition_search_type = SEARCH_PARTITION;
diff --git a/vp9/encoder/vp9_speed_features.h b/vp9/encoder/vp9_speed_features.h
index bdbbe5888..929acaf3e 100644
--- a/vp9/encoder/vp9_speed_features.h
+++ b/vp9/encoder/vp9_speed_features.h
@@ -283,6 +283,9 @@ typedef struct SPEED_FEATURES {
   // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
   int adaptive_pred_interp_filter;
 
+  // Chessboard pattern prediction filter type search
+  int cb_pred_filter_search;
+
   // Fast quantization process path
   int use_quant_fp;