Chessboard pattern partition search

This commit enables a chessboard pattern constrained partition
search for 720p and above resolutions. The scheme applies stricter
partition search to alternative blocks based on its above/left
neighboring blocks' partition range, as well as that of the
collocated blocks in the previous frame. It is currently turned
on at 16x16 block size level. The chessboard pattern is flipped
per coding frame.

The speed 3 runtime is reduced:
park_joy_1080p, 652832 ms -> 607738 ms (7% speed-up)
pedestrian_area_1080p, 215998 ms -> 200589 ms (8% speed-up)

The compression performance is changed:
hd     -0.223%
stdhd  -0.295%

Change-Id: I2d4d123ae89f7171562f618febb4d81789575b19

diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 0d9002bacd2ffce67679bdcb9e0a83b72ece3583..3b478746e9dd15bbade0b1c289ffa647f065ed37 100644 (file)
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -1880,6 +1880,60 @@ static void auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
   *max_block_size = max_size;
 }
 
+// TODO(jingning) refactor functions setting partition search range
+static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                int mi_row, int mi_col, BLOCK_SIZE bsize,
+                                BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
+  int mi_width  = num_8x8_blocks_wide_lookup[bsize];
+  int mi_height = num_8x8_blocks_high_lookup[bsize];
+  int idx, idy;
+
+  MODE_INFO *mi;
+  MODE_INFO **prev_mi =
+      &cm->prev_mi_grid_visible[mi_row * cm->mi_stride + mi_col];
+  BLOCK_SIZE bs, min_size, max_size;
+
+  min_size = BLOCK_64X64;
+  max_size = BLOCK_4X4;
+
+  if (prev_mi) {
+    for (idy = 0; idy < mi_height; ++idy) {
+      for (idx = 0; idx < mi_width; ++idx) {
+        mi = prev_mi[idy * cm->mi_stride + idx];
+        bs = mi ? mi->mbmi.sb_type : bsize;
+        min_size = MIN(min_size, bs);
+        max_size = MAX(max_size, bs);
+      }
+    }
+  }
+
+  if (xd->left_available) {
+    for (idy = 0; idy < mi_height; ++idy) {
+      mi = xd->mi[idy * cm->mi_stride - 1];
+      bs = mi ? mi->mbmi.sb_type : bsize;
+      min_size = MIN(min_size, bs);
+      max_size = MAX(max_size, bs);
+    }
+  }
+
+  if (xd->up_available) {
+    for (idx = 0; idx < mi_width; ++idx) {
+      mi = xd->mi[idx - cm->mi_stride];
+      bs = mi ? mi->mbmi.sb_type : bsize;
+      min_size = MIN(min_size, bs);
+      max_size = MAX(max_size, bs);
+    }
+  }
+
+  if (min_size == max_size) {
+    min_size = min_partition_size[min_size];
+    max_size = max_partition_size[max_size];
+  }
+
+  *min_bs = min_size;
+  *max_bs = max_size;
+}
+
 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
   vpx_memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
 }
@@ -1934,6 +1988,15 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
     set_offsets(cpi, tile, mi_row, mi_col, bsize);
     x->mb_energy = vp9_block_energy(cpi, x, bsize);
   }
+
+  if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
+    int cb_partition_search_ctrl = ((pc_tree->index == 0 || pc_tree->index == 3)
+        + get_chessboard_index(cm->current_video_frame)) & 0x1;
+
+    if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
+      set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
+  }
+
   // Determine partition types in search according to the speed features.
   // The threshold set here has to be of square block size.
   if (cpi->sf.auto_min_max_partition_size) {
@@ -2030,6 +2093,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
         if (cpi->sf.adaptive_motion_search)
           load_pred_mv(x, ctx);
 
+        pc_tree->split[i]->index = i;
         rd_pick_partition(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx,
                           subsize, &this_rate, &this_dist, i != 3,
                           best_rd - sum_rd, pc_tree->split[i]);
@@ -2225,6 +2289,7 @@ static void encode_rd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
     }
 
     vp9_zero(cpi->mb.pred_mv);
+    cpi->pc_root->index = 0;
 
     if ((sf->partition_search_type == SEARCH_PARTITION &&
          sf->use_lastframe_partitioning) ||

diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c
index 7315dd45446772f0d700c566dc8afcb1520914a9..6ca1bc525dfd9aadd7b37cf44f74885db828063d 100644 (file)
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -110,10 +110,12 @@ static void set_good_speed_feature(VP9_COMP *cpi, VP9_COMMON *cm,
   if (speed >= 3) {
     sf->tx_size_search_method = frame_is_intra_only(cm) ? USE_FULL_RD
                                                         : USE_LARGESTALL;
-    if (MIN(cm->width, cm->height) >= 720)
+    if (MIN(cm->width, cm->height) >= 720) {
       sf->disable_split_mask = DISABLE_ALL_SPLIT;
-    else
+      sf->cb_partition_search = frame_is_boosted(cpi) ? 0 : 1;
+    } else {
       sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
+    }
 
     sf->adaptive_pred_interp_filter = 0;
     sf->cb_pred_filter_search = 1;
@@ -334,6 +336,7 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
   sf->adaptive_motion_search = 0;
   sf->adaptive_pred_interp_filter = 0;
   sf->cb_pred_filter_search = 0;
+  sf->cb_partition_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 929acaf3e10845ead414b8bea0245e9550636882..de731cee1a99032e2836e431a14eca7e7bc35d8a 100644 (file)
--- a/vp9/encoder/vp9_speed_features.h
+++ b/vp9/encoder/vp9_speed_features.h
@@ -286,6 +286,8 @@ typedef struct SPEED_FEATURES {
   // Chessboard pattern prediction filter type search
   int cb_pred_filter_search;
 
+  int cb_partition_search;
+
   // Fast quantization process path
   int use_quant_fp;