Port "cost_coeff speed improvements" to vp9.

About a 5% faster overall encode (perf cycles) at speed zero!

Change-Id: Iaf013ba75884415cd824e98349f654ffb1c3ef33

vp9/encoder/vp9_rdopt.c

@@ -387,47 +387,70 @@ static int cost_coeffs(MACROBLOCK *x,
     cost = token_costs[0][0][pt][EOB_TOKEN];
     c = 0;
   } else {
-    int band_left = *band_count++;
+    if (use_fast_coef_costing) {
+      int band_left = *band_count++;
 
-    // dc token
+      // dc token
-    int v = qcoeff[0];
+      int v = qcoeff[0];
-    int16_t prev_t;
+      int16_t prev_t;
-    EXTRABIT e;
+      cost = vp9_get_token_cost(v, &prev_t, cat6_high_cost);
-    vp9_get_token_extra(v, &prev_t, &e);
+      cost += (*token_costs)[0][pt][prev_t];
-    cost = (*token_costs)[0][pt][prev_t] +
-        vp9_get_cost(prev_t, e, cat6_high_cost);
 
-    token_cache[0] = vp9_pt_energy_class[prev_t];
+      token_cache[0] = vp9_pt_energy_class[prev_t];
-    ++token_costs;
+      ++token_costs;
 
-    // ac tokens
+      // ac tokens
-    for (c = 1; c < eob; c++) {
+      for (c = 1; c < eob; c++) {
-      const int rc = scan[c];
+        const int rc = scan[c];
-      int16_t t;
+        int16_t t;
 
-      v = qcoeff[rc];
+        v = qcoeff[rc];
-      vp9_get_token_extra(v, &t, &e);
+        cost += vp9_get_token_cost(v, &t, cat6_high_cost);
-      if (use_fast_coef_costing) {
+        cost += (*token_costs)[!prev_t][!prev_t][t];
-        cost += (*token_costs)[!prev_t][!prev_t][t] +
+        prev_t = t;
-            vp9_get_cost(t, e, cat6_high_cost);
+        if (!--band_left) {
-      } else {
+          band_left = *band_count++;
-        pt = get_coef_context(nb, token_cache, c);
+          ++token_costs;
-        cost += (*token_costs)[!prev_t][pt][t] +
+        }
-            vp9_get_cost(t, e, cat6_high_cost);
-        token_cache[rc] = vp9_pt_energy_class[t];
       }
-      prev_t = t;
-      if (!--band_left) {
-        band_left = *band_count++;
-        ++token_costs;
-      }
-    }
 
-    // eob token
+      // eob token
-    if (band_left) {
+      if (band_left)
-      if (use_fast_coef_costing) {
         cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
-      } else {
+
+    } else {  // !use_fast_coef_costing
+      int band_left = *band_count++;
+
+      // dc token
+      int v = qcoeff[0];
+      int16_t tok;
+      unsigned int (*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
+      cost = vp9_get_token_cost(v, &tok, cat6_high_cost);
+      cost += (*token_costs)[0][pt][tok];
+
+      token_cache[0] = vp9_pt_energy_class[tok];
+      ++token_costs;
+
+      tok_cost_ptr = &((*token_costs)[!tok]);
+
+      // ac tokens
+      for (c = 1; c < eob; c++) {
+        const int rc = scan[c];
+
+        v = qcoeff[rc];
+        cost += vp9_get_token_cost(v, &tok, cat6_high_cost);
+        pt = get_coef_context(nb, token_cache, c);
+        cost += (*tok_cost_ptr)[pt][tok];
+        token_cache[rc] = vp9_pt_energy_class[tok];
+        if (!--band_left) {
+          band_left = *band_count++;
+          ++token_costs;
+        }
+        tok_cost_ptr = &((*token_costs)[!tok]);
+      }
+
+      // eob token
+      if (band_left) {
         pt = get_coef_context(nb, token_cache, c);
         cost += (*token_costs)[0][pt][EOB_TOKEN];
       }

vp9/encoder/vp9_tokenize.c

@@ -50,6 +50,35 @@ static const TOKENVALUE dct_cat_lt_10_value_tokens[] = {
 const TOKENVALUE *vp9_dct_cat_lt_10_value_tokens = dct_cat_lt_10_value_tokens +
     (sizeof(dct_cat_lt_10_value_tokens) / sizeof(*dct_cat_lt_10_value_tokens))
     / 2;
+// The corresponding costs of the extrabits for the tokens in the above table
+// are stored in the table below. The values are obtained from looking up the
+// entry for the specified extrabits in the table corresponding to the token
+// (as defined in cost element vp9_extra_bits)
+// e.g. {9, 63} maps to cat5_cost[63 >> 1], {1, 1} maps to sign_cost[1 >> 1]
+static const int dct_cat_lt_10_value_cost[] = {
+  3773, 3750, 3704, 3681, 3623, 3600, 3554, 3531,
+  3432, 3409, 3363, 3340, 3282, 3259, 3213, 3190,
+  3136, 3113, 3067, 3044, 2986, 2963, 2917, 2894,
+  2795, 2772, 2726, 2703, 2645, 2622, 2576, 2553,
+  3197, 3116, 3058, 2977, 2881, 2800,
+  2742, 2661, 2615, 2534, 2476, 2395,
+  2299, 2218, 2160, 2079,
+  2566, 2427, 2334, 2195, 2023, 1884, 1791, 1652,
+  1893, 1696, 1453, 1256, 1229, 864,
+  512, 512, 512, 512, 0,
+  512, 512, 512, 512,
+  864, 1229, 1256, 1453, 1696, 1893,
+  1652, 1791, 1884, 2023, 2195, 2334, 2427, 2566,
+  2079, 2160, 2218, 2299, 2395, 2476, 2534, 2615,
+  2661, 2742, 2800, 2881, 2977, 3058, 3116, 3197,
+  2553, 2576, 2622, 2645, 2703, 2726, 2772, 2795,
+  2894, 2917, 2963, 2986, 3044, 3067, 3113, 3136,
+  3190, 3213, 3259, 3282, 3340, 3363, 3409, 3432,
+  3531, 3554, 3600, 3623, 3681, 3704, 3750, 3773,
+};
+const int *vp9_dct_cat_lt_10_value_cost = dct_cat_lt_10_value_cost +
+    (sizeof(dct_cat_lt_10_value_cost) / sizeof(*dct_cat_lt_10_value_cost))
+    / 2;
 
 // Array indices are identical to previously-existing CONTEXT_NODE indices
 const vpx_tree_index vp9_coef_tree[TREE_SIZE(ENTROPY_TOKENS)] = {

vp9/encoder/vp9_tokenize.h

@@ -74,6 +74,7 @@ extern const int16_t *vp9_dct_value_cost_ptr;
  */
 extern const TOKENVALUE *vp9_dct_value_tokens_ptr;
 extern const TOKENVALUE *vp9_dct_cat_lt_10_value_tokens;
+extern const int *vp9_dct_cat_lt_10_value_cost;
 extern const int16_t vp9_cat6_low_cost[256];
 extern const int vp9_cat6_high_cost[64];
 extern const int vp9_cat6_high10_high_cost[256];
@@ -117,6 +118,18 @@ static INLINE int16_t vp9_get_token(int v) {
   return vp9_dct_cat_lt_10_value_tokens[v].token;
 }
 
+static INLINE int vp9_get_token_cost(int v, int16_t *token,
+                                     const int *cat6_high_table) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) {
+    EXTRABIT extrabits;
+    *token = CATEGORY6_TOKEN;
+    extrabits = abs(v) - CAT6_MIN_VAL;
+    return vp9_cat6_low_cost[extrabits & 0xff] +
+           cat6_high_table[extrabits >> 8];
+  }
+  *token = vp9_dct_cat_lt_10_value_tokens[v].token;
+  return vp9_dct_cat_lt_10_value_cost[v];
+}
 
 #ifdef __cplusplus
 }  // extern "C"