Added skip switches for SB32 and SB64

Added switches and code to skip/breakout from
doing SB32 and SB64 tests based on whether
the 16x16 MB tests used split modes. Also to
optionally skip 64x64 if 16x16 was chosen over
32x32.

Impact varies depending on clip from a few %
up to almost 50% on encode speed. Only the
split mode breakout is currently enabled.

Change-Id: Ib5836140b064b350ffa3057778ed2cadcc495cf8

vp9/encoder/vp9_encodeframe.c

@@ -741,17 +741,18 @@ static void set_offsets(VP9_COMP *cpi,
   }
 }
 
-static void pick_mb_modes(VP9_COMP *cpi,
-                          int mb_row,
-                          int mb_col,
-                          TOKENEXTRA **tp,
-                          int *totalrate,
-                          int *totaldist) {
+static int pick_mb_modes(VP9_COMP *cpi,
+                         int mb_row,
+                         int mb_col,
+                         TOKENEXTRA **tp,
+                         int *totalrate,
+                         int *totaldist) {
   VP9_COMMON *const cm = &cpi->common;
   MACROBLOCK *const x = &cpi->mb;
   MACROBLOCKD *const xd = &x->e_mbd;
   int i;
   int recon_yoffset, recon_uvoffset;
+  int splitmodes_used = 0;
   ENTROPY_CONTEXT_PLANES left_context[2];
   ENTROPY_CONTEXT_PLANES above_context[2];
   ENTROPY_CONTEXT_PLANES *initial_above_context_ptr = cm->above_context
@@ -818,6 +819,8 @@ static void pick_mb_modes(VP9_COMP *cpi,
       *totalrate += r;
       *totaldist += d;
 
+      splitmodes_used += (mbmi->mode == SPLITMV);
+
       // Dummy encode, do not do the tokenization
       encode_macroblock(cpi, tp, recon_yoffset, recon_uvoffset, 0,
                         mb_row + y_idx, mb_col + x_idx);
@@ -849,6 +852,8 @@ static void pick_mb_modes(VP9_COMP *cpi,
   vpx_memcpy(initial_above_context_ptr,
              above_context,
              sizeof(above_context));
+
+  return splitmodes_used;
 }
 
 static void pick_sb_modes(VP9_COMP *cpi,
@@ -1106,6 +1111,7 @@ static void encode_sb_row(VP9_COMP *cpi,
     int sb32_rate = 0, sb32_dist = 0;
     int is_sb[4];
     int sb64_rate = INT_MAX, sb64_dist;
+    int sb64_skip = 0;
     ENTROPY_CONTEXT_PLANES l[4], a[4];
     TOKENEXTRA *tp_orig = *tp;
 
@@ -1115,18 +1121,27 @@ static void encode_sb_row(VP9_COMP *cpi,
       const int x_idx = (i & 1) << 1, y_idx = i & 2;
       int mb_rate = 0, mb_dist = 0;
       int sb_rate = INT_MAX, sb_dist;
+      int splitmodes_used = 0;
+      int sb32_skip = 0;
 
       if (mb_row + y_idx >= cm->mb_rows || mb_col + x_idx >= cm->mb_cols)
         continue;
 
       xd->sb_index = i;
 
-      pick_mb_modes(cpi, mb_row + y_idx, mb_col + x_idx,
-                    tp, &mb_rate, &mb_dist);
+      splitmodes_used = pick_mb_modes(cpi, mb_row + y_idx, mb_col + x_idx,
+                                      tp, &mb_rate, &mb_dist);
+
       mb_rate += vp9_cost_bit(cm->sb32_coded, 0);
 
-      if (!(((cm->mb_cols & 1) && mb_col + x_idx == cm->mb_cols - 1) ||
-            ((cm->mb_rows & 1) && mb_row + y_idx == cm->mb_rows - 1))) {
+      if (cpi->sf.splitmode_breakout) {
+        sb32_skip = splitmodes_used;
+        sb64_skip += splitmodes_used;
+      }
+
+      if ( !sb32_skip &&
+           !(((cm->mb_cols & 1) && mb_col + x_idx == cm->mb_cols - 1) ||
+             ((cm->mb_rows & 1) && mb_row + y_idx == cm->mb_rows - 1))) {
         /* Pick a mode assuming that it applies to all 4 of the MBs in the SB */
         pick_sb_modes(cpi, mb_row + y_idx, mb_col + x_idx,
                       tp, &sb_rate, &sb_dist);
@@ -1144,6 +1159,11 @@ static void encode_sb_row(VP9_COMP *cpi,
         is_sb[i] = 0;
         sb32_rate += mb_rate;
         sb32_dist += mb_dist;
+
+        // If we used 16x16 instead of 32x32 then skip 64x64 (if enabled).
+        if (cpi->sf.mb16_breakout) {
+          ++sb64_skip;
+        }
       }
 
       /* Encode SB using best computed mode(s) */
@@ -1159,7 +1179,8 @@ static void encode_sb_row(VP9_COMP *cpi,
     memcpy(cm->left_context, &l, sizeof(l));
     sb32_rate += vp9_cost_bit(cm->sb64_coded, 0);
 
-    if (!(((cm->mb_cols & 3) && mb_col + 3 >= cm->mb_cols) ||
+    if (!sb64_skip &&
+        !(((cm->mb_cols & 3) && mb_col + 3 >= cm->mb_cols) ||
           ((cm->mb_rows & 3) && mb_row + 3 >= cm->mb_rows))) {
       pick_sb64_modes(cpi, mb_row, mb_col, tp, &sb64_rate, &sb64_dist);
       sb64_rate += vp9_cost_bit(cm->sb64_coded, 1);

vp9/encoder/vp9_onyx_if.c

@@ -691,9 +691,10 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
   sf->optimize_coefficients = 1;
 #endif
   sf->no_skip_block4x4_search = 1;
-
   sf->first_step = 0;
   sf->max_step_search_steps = MAX_MVSEARCH_STEPS;
+  sf->splitmode_breakout = 0;
+  sf->mb16_breakout = 0;
 
   // Set rd thresholds based on mode and speed setting
   set_rd_speed_thresholds(cpi, mode, speed);
@@ -704,6 +705,9 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
       break;
 
     case 1:
+      sf->splitmode_breakout = 1;
+      sf->mb16_breakout = 0;
+
       if (speed > 0) {
         /* Disable coefficient optimization above speed 0 */
         sf->optimize_coefficients = 0;

vp9/encoder/vp9_onyx_int.h

@@ -259,7 +259,8 @@ typedef struct {
   int optimize_coefficients;
   int no_skip_block4x4_search;
   int search_best_filter;
-
+  int splitmode_breakout;
+  int mb16_breakout;
 } SPEED_FEATURES;
 
 typedef struct {