vp9_optimize_b: Combine extrabits cost with token lookup

About 0.6% fewer cycles spent in vp9_optimize_b.

Change-Id: I2ae62a78374c594ed81d4e3100a5848e2f6f2c4e

diff --git a/vp9/encoder/vp9_encodemb.c b/vp9/encoder/vp9_encodemb.c
index 1cf0b02..cf92dba 100644 (file)
--- a/vp9/encoder/vp9_encodemb.c
+++ b/vp9/encoder/vp9_encodemb.c
@@ -108,7 +108,6 @@ int vp9_optimize_b(MACROBLOCK *mb, int plane, int block, TX_SIZE tx_size,
   int rate0, rate1;
   int64_t error0, error1;
   int16_t t0, t1;
-  EXTRABIT e0;
   unsigned int(*const token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
       mb->token_costs[tx_size][type][ref];
   int best, band, pt, i, final_eob;
@@ -144,7 +143,7 @@ int vp9_optimize_b(MACROBLOCK *mb, int plane, int block, TX_SIZE tx_size,
       /* Evaluate the first possibility for this state. */
       rate0 = tokens[next][0].rate;
       rate1 = tokens[next][1].rate;
-      vp9_get_token_extra(x, &t0, &e0);
+      vp9_get_token_extracost(cat6_high_cost, x, &t0, &base_bits);
       /* Consider both possible successor states. */
       if (next < default_eob) {
         band = band_translate[i + 1];
@@ -155,7 +154,6 @@ int vp9_optimize_b(MACROBLOCK *mb, int plane, int block, TX_SIZE tx_size,
       UPDATE_RD_COST();
       /* And pick the best. */
       best = rd_cost1 < rd_cost0;
-      base_bits = vp9_get_cost(t0, e0, cat6_high_cost);
       dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
 #if CONFIG_VP9_HIGHBITDEPTH
       if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
@@ -193,9 +191,9 @@ int vp9_optimize_b(MACROBLOCK *mb, int plane, int block, TX_SIZE tx_size,
          */
         t0 = tokens[next][0].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
         t1 = tokens[next][1].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
-        e0 = 0;
+        base_bits = 0;
       } else {
-        vp9_get_token_extra(x, &t0, &e0);
+        vp9_get_token_extracost(cat6_high_cost, x, &t0, &base_bits);
         t1 = t0;
       }
       if (next < default_eob) {
@@ -213,7 +211,6 @@ int vp9_optimize_b(MACROBLOCK *mb, int plane, int block, TX_SIZE tx_size,
       UPDATE_RD_COST();
       /* And pick the best. */
       best = rd_cost1 < rd_cost0;
-      base_bits = vp9_get_cost(t0, e0, cat6_high_cost);
 
 #if CONFIG_VP9_HIGHBITDEPTH
       if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {

diff --git a/vp9/encoder/vp9_tokenize.h b/vp9/encoder/vp9_tokenize.h
index c7694af..4b8d7ad 100644 (file)
--- a/vp9/encoder/vp9_tokenize.h
+++ b/vp9/encoder/vp9_tokenize.h
@@ -79,12 +79,22 @@ extern const int16_t vp9_cat6_low_cost[256];
 extern const uint16_t vp9_cat6_high_cost[64];
 extern const uint16_t vp9_cat6_high10_high_cost[256];
 extern const uint16_t vp9_cat6_high12_high_cost[1024];
-static INLINE int vp9_get_cost(int16_t token, EXTRABIT extrabits,
-                               const uint16_t *cat6_high_table) {
-  if (token != CATEGORY6_TOKEN)
-    return vp9_extra_bits[token].cost[extrabits >> 1];
-  return vp9_cat6_low_cost[(extrabits >> 1) & 0xff] +
-         cat6_high_table[extrabits >> 9];
+
+static INLINE void vp9_get_token_extracost(const uint16_t *cat6_high_table,
+                                           int v, int16_t *token,
+                                           int *extracost) {
+  EXTRABIT extrabits;  // unsigned extrabits
+  v = abs(v);
+  if (v >= CAT6_MIN_VAL) {
+    *token = CATEGORY6_TOKEN;
+    extrabits = v - CAT6_MIN_VAL;
+    *extracost =
+        vp9_cat6_low_cost[extrabits & 0xff] + cat6_high_table[extrabits >> 8];
+  } else {
+    *token = vp9_dct_cat_lt_10_value_tokens[v].token;
+    extrabits = vp9_dct_cat_lt_10_value_tokens[v].extra >> 1;
+    *extracost = vp9_extra_bits[*token].cost[extrabits];
+  }
 }
 
 #if CONFIG_VP9_HIGHBITDEPTH