faster cabac rdo. up to 10% faster at q0, but negligible at normal bitrates.

git-svn-id: svn://svn.videolan.org/x264/trunk@647 df754926-b1dd-0310-bc7b-ec298dee348c

diff --git a/encoder/cabac.c b/encoder/cabac.c
index 7b0223896e67bd25f8dcc12e25e7d1f878478a34..908a4600997b8661edcad20f62b242399232992d 100644 (file)
--- a/encoder/cabac.c
+++ b/encoder/cabac.c
@@ -783,12 +783,17 @@ static void block_residual_write_cabac( x264_t *h, x264_cabac_t *cb, int i_ctxBl
         if( i_prefix != 0 )
         {
             const int i_ctxIdxInc = 5 + X264_MIN( 4, i_abslevelgt1 ) + i_ctx_level;
+#ifdef RDO_SKIP_BS
+            cb->f8_bits_encoded += cabac_prefix_size[i_prefix][cb->state[i_ctxIdxInc]];
+            cb->state[i_ctxIdxInc] = cabac_prefix_transition[i_prefix][cb->state[i_ctxIdxInc]];
+#else
             int j;
             for( j = 0; j < i_prefix - 1; j++ )
                 x264_cabac_encode_decision( cb, i_ctxIdxInc, 1 );
             if( i_prefix < 14 )
                 x264_cabac_encode_decision( cb, i_ctxIdxInc, 0 );
-            else /* suffix */
+#endif
+            if( i_prefix >= 14 )
                 x264_cabac_encode_ue_bypass( cb, 0, i_coeff_abs_m1[i] - 14 );
 
             i_abslevelgt1++;
@@ -797,6 +802,9 @@ static void block_residual_write_cabac( x264_t *h, x264_cabac_t *cb, int i_ctxBl
             i_abslevel1++;
 
         /* write sign */
+#ifdef RDO_SKIP_BS
+        if( i_prefix == 0 )
+#endif
         x264_cabac_encode_bypass( cb, i_coeff_sign[i] );
     }
 }

diff --git a/encoder/rdo.c b/encoder/rdo.c
index e10702f51a86ba5ba76c45e3dabd0bf514f28ff7..b204639f5d3cbee6cdb1ef07b73317fdad3d2bfe 100644 (file)
--- a/encoder/rdo.c
+++ b/encoder/rdo.c
@@ -26,6 +26,9 @@
 
 #define RDO_SKIP_BS
 
+static int cabac_prefix_transition[15][128];
+static int cabac_prefix_size[15][128];
+
 /* CAVLC: produces exactly the same bit count as a normal encode */
 /* this probably still leaves some unnecessary computations */
 #define bs_write1(s,v)     ((s)->i_bits_encoded += 1)
@@ -213,8 +216,6 @@ int x264_rd_cost_i8x8_chroma( x264_t *h, int i_lambda2, int i_mode, int b_dct )
 #define LAMBDA_BITS 4
 
 /* precalculate the cost of coding abs_level_m1 */
-static int cabac_prefix_transition[15][128];
-static int cabac_prefix_size[15][128];
 void x264_rdo_init( )
 {
     int i_prefix;
@@ -318,7 +319,7 @@ static void quant_trellis_cabac( x264_t *h, int16_t *dct,
     uint8_t cabac_state_last[64];
     const int b_interlaced = h->mb.b_interlaced;
     const int f = 1 << 15; // no deadzone
-    int i_last_nnz = -1;
+    int i_last_nnz;
     int i, j;
 
     // (# of coefs) * (# of ctx) * (# of levels tried) = 1024
@@ -331,21 +332,25 @@ static void quant_trellis_cabac( x264_t *h, int16_t *dct,
     int i_levels_used = 1;
 
     /* init coefs */
-    for( i = b_ac; i < i_coefs; i++ )
-    {
-        int coef = dct[zigzag[i]];
-        abs_coefs[i] = abs(coef);
-        signs[i] = coef < 0 ? -1 : 1;
-        if( f <= abs_coefs[i] * quant_mf[zigzag[i]] )
-            i_last_nnz = i;
-    }
+    for( i = i_coefs-1; i >= b_ac; i-- )
+        if( (unsigned)(dct[zigzag[i]] * quant_mf[zigzag[i]] + f-1) >= 2*f )
+            break;
 
-    if( i_last_nnz == -1 )
+    if( i < b_ac )
     {
         memset( dct, 0, i_coefs * sizeof(*dct) );
         return;
     }
 
+    i_last_nnz = i;
+
+    for( ; i >= b_ac; i-- )
+    {
+        int coef = dct[zigzag[i]];
+        abs_coefs[i] = abs(coef);
+        signs[i] = coef < 0 ? -1 : 1;
+    }
+
     /* init trellis */
     for( i = 1; i < 8; i++ )
         nodes_cur[i].score = TRELLIS_SCORE_MAX;