shithub: libvpx

Download patch

ref: 4d98b50be5eb79fb0bd8ac4d30239de1f22daa0f
parent: af2a7369537cc842c9d4e484b1cb979db68dfe06
parent: ba70f1601108c537ffb48b4f0e63ccd5c7af25ef
author: Yunqing Wang <yunqingwang@google.com>
date: Mon Aug 18 15:00:24 EDT 2014

Merge "Add early termination in transform size search"

--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -490,11 +490,11 @@
                              {INT64_MAX, INT64_MAX},
                              {INT64_MAX, INT64_MAX},
                              {INT64_MAX, INT64_MAX}};
-  TX_SIZE n, m;
+  int n, m;
   int s0, s1;
   const TX_SIZE max_mode_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
   int64_t best_rd = INT64_MAX;
-  TX_SIZE best_tx = TX_4X4;
+  TX_SIZE best_tx = max_tx_size;
 
   const vp9_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc.tx_probs);
   assert(skip_prob > 0);
@@ -501,7 +501,7 @@
   s0 = vp9_cost_bit(skip_prob, 0);
   s1 = vp9_cost_bit(skip_prob, 1);
 
-  for (n = TX_4X4; n <= max_tx_size; n++) {
+  for (n = max_tx_size; n >= 0;  n--) {
     txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
                      &sse[n], ref_best_rd, 0, bs, n,
                      cpi->sf.use_fast_coef_costing);
@@ -522,6 +522,13 @@
       rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
       rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
     }
+
+    // Early termination in transform size search.
+    if (cpi->sf.tx_size_search_breakout &&
+        (rd[n][1] == INT64_MAX ||
+        (n < max_tx_size && rd[n][1] > rd[n + 1][1]) ||
+        s[n] == 1))
+      break;
 
     if (rd[n][1] < best_rd) {
       best_tx = n;
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -59,8 +59,6 @@
   if (speed >= 1) {
     sf->use_square_partition_only = !frame_is_intra_only(cm);
     sf->less_rectangular_check  = 1;
-    sf->tx_size_search_method = frame_is_boosted(cpi) ? USE_FULL_RD
-                                                      : USE_LARGESTALL;
 
     if (MIN(cm->width, cm->height) >= 720)
       sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
@@ -80,9 +78,14 @@
     sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
     sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
     sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
+
+    sf->tx_size_search_breakout = 1;
   }
 
   if (speed >= 2) {
+    sf->tx_size_search_method = frame_is_boosted(cpi) ? USE_FULL_RD
+                                                      : USE_LARGESTALL;
+
     if (MIN(cm->width, cm->height) >= 720) {
       sf->lf_motion_threshold = LOW_MOTION_THRESHOLD;
       sf->last_partitioning_redo_frequency = 3;
@@ -391,6 +394,7 @@
   // Recode loop tolerence %.
   sf->recode_tolerance = 25;
   sf->default_interp_filter = SWITCHABLE;
+  sf->tx_size_search_breakout = 0;
 
   if (oxcf->mode == REALTIME) {
     set_rt_speed_feature(cpi, sf, oxcf->speed, oxcf->content);
--- a/vp9/encoder/vp9_speed_features.h
+++ b/vp9/encoder/vp9_speed_features.h
@@ -376,6 +376,10 @@
 
   // default interp filter choice
   INTERP_FILTER default_interp_filter;
+
+  // Early termination in transform size search, which only applies while
+  // tx_size_search_method is USE_FULL_RD.
+  int tx_size_search_breakout;
 } SPEED_FEATURES;
 
 struct VP9_COMP;