Residual coding to cache energy class of tokens.

Proposal for tuning the residual coding by changing how the context
from previous tokens is calculated. Storing the energy class of previous
tokens instead of the token itself eases the critical path of
HW implementations.

Change-Id: I6d71d856b84518f6c88de771ddd818436f794bab

3 files changed, 14 insertions, 4 deletions

diff --git a/vp9/encoder/vp9_encodemb.c b/vp9/encoder/vp9_encodemb.c
index b7f60b127..62ef9a04f 100644
--- a/vp9/encoder/vp9_encodemb.c
+++ b/vp9/encoder/vp9_encodemb.c
@@ -20,6 +20,9 @@
 #include "vp9/common/vp9_systemdependent.h"
 #include "vp9_rtcd.h"
 
+DECLARE_ALIGNED(16, extern const uint8_t,
+                vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);
+
 void vp9_subtract_block(int rows, int cols,
                         int16_t *diff_ptr, int diff_stride,
                         const uint8_t *src_ptr, int src_stride,
@@ -105,7 +108,7 @@ static int trellis_get_coeff_context(const int *scan,
                                      uint8_t *token_cache,
                                      int pad, int l) {
   int bak = token_cache[scan[idx]], pt;
-  token_cache[scan[idx]] = token;
+  token_cache[scan[idx]] = vp9_pt_energy_class[token];
   pt = vp9_get_coef_context(scan, nb, pad, token_cache, idx + 1, l);
   token_cache[scan[idx]] = bak;
   return pt;
@@ -189,7 +192,8 @@ static void optimize_b(VP9_COMMON *const cm, MACROBLOCK *mb,
   *(tokens[eob] + 1) = *(tokens[eob] + 0);
   next = eob;
   for (i = 0; i < eob; i++)
-    token_cache[scan[i]] = vp9_dct_value_tokens_ptr[qcoeff_ptr[scan[i]]].token;
+    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
+        qcoeff_ptr[scan[i]]].token];
   nb = vp9_get_coef_neighbors_handle(scan, &pad);
 
   for (i = eob; i-- > i0;) {
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 2eb3f9b29..aaba2bb11 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -46,6 +46,9 @@
 /* Factor to weigh the rate for switchable interp filters */
 #define SWITCHABLE_INTERP_RATE_FACTOR 1
 
+DECLARE_ALIGNED(16, extern const uint8_t,
+                vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);
+
 const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
   {ZEROMV,    LAST_FRAME,   NONE},
   {DC_PRED,   INTRA_FRAME,  NONE},
@@ -366,7 +369,7 @@ static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb,
 
       if (!c || token_cache[scan[c - 1]])
         cost += vp9_cost_bit(coef_probs[band][pt][0], 1);
-      token_cache[scan[c]] = t;
+      token_cache[scan[c]] = vp9_pt_energy_class[t];
     }
     if (c < seg_eob) {
       if (c)
diff --git a/vp9/encoder/vp9_tokenize.c b/vp9/encoder/vp9_tokenize.c
index 08efc84d4..eb79de1d9 100644
--- a/vp9/encoder/vp9_tokenize.c
+++ b/vp9/encoder/vp9_tokenize.c
@@ -36,6 +36,9 @@ extern vp9_coeff_stats tree_update_hist_16x16[BLOCK_TYPES];
 extern vp9_coeff_stats tree_update_hist_32x32[BLOCK_TYPES];
 #endif  /* ENTROPY_STATS */
 
+DECLARE_ALIGNED(16, extern const uint8_t,
+                vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);
+
 static TOKENVALUE dct_value_tokens[DCT_MAX_VALUE * 2];
 const TOKENVALUE *vp9_dct_value_tokens_ptr;
 static int dct_value_cost[DCT_MAX_VALUE * 2];
@@ -228,7 +231,7 @@ static void tokenize_b(int plane, int block, BLOCK_SIZE_TYPE bsize,
       if (!t->skip_eob_node)
         ++cpi->common.fc.eob_branch_counts[tx_size][type][ref][band][pt];
     }
-    token_cache[scan[c]] = token;
+    token_cache[scan[c]] = vp9_pt_energy_class[token];
     ++t;
   } while (c < eob && ++c < seg_eob);