Use tx_size_cdf with CONFIG_EC_MULTISYMBOL.

This commit produces a bit-exact bitstream because prior calls to code
 the tx_size using aom_write_tree() would automatically convert the
 aom_tree_index on the fly to a CDF.
The tx_size_cdf[]'s are now computed once per frame.
Based on recent tests, this patch reduces decode time by 4%.

Change-Id: I920dbf036b8a9574ea6e65ae5f9c43eef1c3f864

av1/common/entropymode.c

@@ -1632,6 +1632,10 @@ static void init_mode_probs(FRAME_CONTEXT *fc) {
   av1_tree_to_cdf_2D(av1_intra_mode_tree, av1_kf_y_mode_prob, av1_kf_y_mode_cdf,
                      INTRA_MODES, INTRA_MODES);
   av1_tree_to_cdf(av1_segment_tree, fc->seg.tree_probs, fc->seg.tree_cdf);
+  for (int k = 0; k < MAX_TX_DEPTH; k++) {
+    av1_tree_to_cdf_1D(av1_tx_size_tree[k], fc->tx_size_probs[k],
+                       fc->tx_size_cdf[k], TX_SIZE_CONTEXTS);
+  }
 #endif
 #if CONFIG_DELTA_Q
   av1_copy(fc->delta_q_prob, default_delta_q_probs);
@@ -1684,6 +1688,12 @@ void av1_set_mode_cdfs(struct AV1Common *cm) {
     av1_tree_to_cdf(av1_ext_tx_tree, fc->inter_ext_tx_prob[i],
                     fc->inter_ext_tx_cdf[i]);
 #endif
+  for (i = 0; i < MAX_TX_DEPTH; i++) {
+    for (j = 0; j < TX_SIZE_CONTEXTS; j++) {
+      av1_tree_to_cdf(av1_tx_size_tree[i], fc->tx_size_probs[i][j],
+                      fc->tx_size_cdf[i][j]);
+    }
+  }
 }
 #endif
 

av1/common/entropymode.h

@@ -243,6 +243,8 @@ typedef struct frame_contexts {
      multiple copies for adaptation in tiles easier */
   aom_cdf_prob kf_y_cdf[INTRA_MODES][INTRA_MODES]
                        [INTRA_MODES + CONFIG_EC_ADAPT];
+  aom_cdf_prob tx_size_cdf[MAX_TX_DEPTH][TX_SIZE_CONTEXTS]
+                          [MAX_TX_DEPTH + 1 + CONFIG_EC_ADAPT];
 #if !CONFIG_EXT_TX
   aom_cdf_prob intra_ext_tx_cdf[EXT_TX_SIZES][TX_TYPES]
                                [TX_TYPES + CONFIG_EC_ADAPT];

av1/decoder/decodemv.c

@@ -404,8 +404,13 @@ static TX_SIZE read_selected_tx_size(AV1_COMMON *cm, MACROBLOCKD *xd,
   FRAME_COUNTS *counts = xd->counts;
   const int ctx = get_tx_size_context(xd);
   const int depth =
+#if CONFIG_EC_MULTISYMBOL
+      aom_read_symbol(r, cm->fc->tx_size_cdf[tx_size_cat][ctx], tx_size_cat + 2,
+                      ACCT_STR);
+#else
       aom_read_tree(r, av1_tx_size_tree[tx_size_cat],
                     cm->fc->tx_size_probs[tx_size_cat][ctx], ACCT_STR);
+#endif
   const TX_SIZE tx_size = depth_to_tx_size(depth);
 #if CONFIG_RECT_TX
   assert(!is_rect_tx(tx_size));

av1/encoder/bitstream.c

@@ -80,6 +80,7 @@ static struct av1_token palette_size_encodings[PALETTE_MAX_SIZE - 1];
 static struct av1_token palette_color_index_encodings[PALETTE_MAX_SIZE - 1]
                                                      [PALETTE_MAX_SIZE];
 #endif  // CONFIG_PALETTE
+#if !CONFIG_EC_MULTISYMBOL
 static const struct av1_token tx_size_encodings[MAX_TX_DEPTH][TX_SIZES] = {
   { { 0, 1 }, { 1, 1 } },                      // Max tx_size is 8X8
   { { 0, 1 }, { 2, 2 }, { 3, 2 } },            // Max tx_size is 16X16
@@ -88,6 +89,7 @@ static const struct av1_token tx_size_encodings[MAX_TX_DEPTH][TX_SIZES] = {
   { { 0, 1 }, { 2, 2 }, { 6, 3 }, { 14, 4 }, { 15, 4 } },  // Max tx_size 64X64
 #endif                                                     // CONFIG_TX64X64
 };
+#endif
 
 #if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
 static INLINE void write_uniform(aom_writer *w, int n, int v) {
@@ -456,9 +458,14 @@ static void write_selected_tx_size(const AV1_COMMON *cm, const MACROBLOCKD *xd,
         IMPLIES(is_rect_tx(tx_size), tx_size == max_txsize_rect_lookup[bsize]));
 #endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
 
+#if CONFIG_EC_MULTISYMBOL
+    aom_write_symbol(w, depth, cm->fc->tx_size_cdf[tx_size_cat][tx_size_ctx],
+                     tx_size_cat + 2);
+#else
     av1_write_token(w, av1_tx_size_tree[tx_size_cat],
                     cm->fc->tx_size_probs[tx_size_cat][tx_size_ctx],
                     &tx_size_encodings[tx_size_cat][depth]);
+#endif
   }
 }