Inline vp9_get_coef_context() (and remove vp9_ prefix).

Makes cost_coeffs() a lot faster:
4x4: 236 -> 181 cycles
8x8: 888 -> 588 cycles
16x16: 3550 -> 2483 cycles
32x32: 17392 -> 12010 cycles

Total encode time of first 50 frames of bus (speed 0) @ 1500kbps goes
from 2min51.6 to 2min43.9, i.e. 4.7% overall speedup.

Change-Id: I16b8d595946393c8dc661599550b3f37f5718896

vp9/common/vp9_entropy.c

@@ -445,32 +445,6 @@ vp9_extra_bit vp9_extra_bits[12] = {
 
 #include "vp9/common/vp9_default_coef_probs.h"
 
-// This function updates and then returns n AC coefficient context
-// This is currently a placeholder function to allow experimentation
-// using various context models based on the energy earlier tokens
-// within the current block.
-//
-// For now it just returns the previously used context.
-#define MAX_NEIGHBORS 2
-int vp9_get_coef_context(const int *scan, const int *neighbors,
-                         int nb_pad, uint8_t *token_cache, int c, int l) {
-  int eob = l;
-  assert(nb_pad == MAX_NEIGHBORS);
-  if (c == eob) {
-    return 0;
-  } else {
-    int ctx;
-    assert(neighbors[MAX_NEIGHBORS * c + 0] >= 0);
-    if (neighbors[MAX_NEIGHBORS * c + 1] >= 0) {
-      ctx = (1 + token_cache[scan[neighbors[MAX_NEIGHBORS * c + 0]]] +
-             token_cache[scan[neighbors[MAX_NEIGHBORS * c + 1]]]) >> 1;
-    } else {
-      ctx = token_cache[scan[neighbors[MAX_NEIGHBORS * c + 0]]];
-    }
-    return ctx;
-  }
-};
-
 void vp9_default_coef_probs(VP9_COMMON *pc) {
   vpx_memcpy(pc->fc.coef_probs[TX_4X4], default_coef_probs_4x4,
              sizeof(pc->fc.coef_probs[TX_4X4]));

vp9/common/vp9_entropy.h

@@ -148,8 +148,27 @@ static int get_coef_band(const uint8_t * band_translate, int coef_index) {
     ? (COEF_BANDS-1) : band_translate[coef_index];
 }
 
-extern int vp9_get_coef_context(const int *scan, const int *neighbors,
-                                int nb_pad, uint8_t *token_cache, int c, int l);
+#define MAX_NEIGHBORS 2
+static INLINE int get_coef_context(const int *scan, const int *neighbors,
+                                   int nb_pad, uint8_t *token_cache,
+                                   int c, int l) {
+  int eob = l;
+  assert(nb_pad == MAX_NEIGHBORS);
+  if (c == eob) {
+    return 0;
+  } else {
+    int ctx;
+    assert(neighbors[MAX_NEIGHBORS * c + 0] >= 0);
+    if (neighbors[MAX_NEIGHBORS * c + 1] >= 0) {
+      ctx = (1 + token_cache[scan[neighbors[MAX_NEIGHBORS * c + 0]]] +
+             token_cache[scan[neighbors[MAX_NEIGHBORS * c + 1]]]) >> 1;
+    } else {
+      ctx = token_cache[scan[neighbors[MAX_NEIGHBORS * c + 0]]];
+    }
+    return ctx;
+  }
+}
+
 const int *vp9_get_coef_neighbors_handle(const int *scan, int *pad);
 
 

vp9/decoder/vp9_detokenize.c

@@ -172,8 +172,8 @@ static int decode_coefs(FRAME_CONTEXT *fc, const MACROBLOCKD *xd,
     if (c >= seg_eob)
       break;
     if (c)
-      pt = vp9_get_coef_context(scan, nb, pad, token_cache,
-                                c, default_eob);
+      pt = get_coef_context(scan, nb, pad, token_cache,
+                            c, default_eob);
     band = get_coef_band(band_translate, c);
     prob = coef_probs[band][pt];
 #if !CONFIG_BALANCED_COEFTREE
@@ -186,8 +186,8 @@ SKIP_START:
     if (c >= seg_eob)
       break;
     if (c)
-      pt = vp9_get_coef_context(scan, nb, pad, token_cache,
-                                c, default_eob);
+      pt = get_coef_context(scan, nb, pad, token_cache,
+                            c, default_eob);
     band = get_coef_band(band_translate, c);
     prob = coef_probs[band][pt];
 

vp9/encoder/vp9_encodemb.c

@@ -116,7 +116,7 @@ static int trellis_get_coeff_context(const int *scan,
                                      int pad, int l) {
   int bak = token_cache[scan[idx]], pt;
   token_cache[scan[idx]] = vp9_pt_energy_class[token];
-  pt = vp9_get_coef_context(scan, nb, pad, token_cache, idx + 1, l);
+  pt = get_coef_context(scan, nb, pad, token_cache, idx + 1, l);
   token_cache[scan[idx]] = bak;
   return pt;
 }

vp9/encoder/vp9_rdopt.c

@@ -402,7 +402,7 @@ static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb,
 
       v = qcoeff_ptr[rc];
       t = vp9_dct_value_tokens_ptr[v].token;
-      pt = vp9_get_coef_context(scan, nb, pad, token_cache, c, default_eob);
+      pt = get_coef_context(scan, nb, pad, token_cache, c, default_eob);
       cost += token_costs[!prev_t][band][pt][t] + vp9_dct_value_cost_ptr[v];
       token_cache[rc] = vp9_pt_energy_class[t];
       prev_t = t;
@@ -410,7 +410,7 @@ static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb,
 
     // eob token
     if (c < seg_eob) {
-      pt = vp9_get_coef_context(scan, nb, pad, token_cache, c, default_eob);
+      pt = get_coef_context(scan, nb, pad, token_cache, c, default_eob);
       cost += token_costs[0][get_coef_band(band_translate, c)][pt]
                          [DCT_EOB_TOKEN];
     }

vp9/encoder/vp9_tokenize.c

@@ -193,7 +193,7 @@ static void tokenize_b(int plane, int block, BLOCK_SIZE_TYPE bsize,
     int v = 0;
     rc = scan[c];
     if (c)
-      pt = vp9_get_coef_context(scan, nb, pad, token_cache, c, default_eob);
+      pt = get_coef_context(scan, nb, pad, token_cache, c, default_eob);
     if (c < eob) {
       v = qcoeff_ptr[rc];
       assert(-DCT_MAX_VALUE <= v  &&  v < DCT_MAX_VALUE);