Populate bmi for B_PRED only

Small decode performance gain (~1%) on keyframes.  No
noticeable gains on encode.  Also changed pick_intra4x4mby_modes()
to read the above and left block modes for keyframes only.

Change-Id: I1f4885252f5b3e9caf04d4e01e643960f910aba5

vp8/common/findnearmv.h

@@ -123,7 +123,21 @@ static B_PREDICTION_MODE left_block_mode(const MODE_INFO *cur_mb, int b)
     {
         /* On L edge, get from MB to left of us */
         --cur_mb;
-        b += 4;
+        switch (cur_mb->mbmi.mode)
+        {
+            case B_PRED:
+              return (cur_mb->bmi + b + 3)->as_mode;
+            case DC_PRED:
+                return B_DC_PRED;
+            case V_PRED:
+                return B_VE_PRED;
+            case H_PRED:
+                return B_HE_PRED;
+            case TM_PRED:
+                return B_TM_PRED;
+            default:
+                return B_DC_PRED;
+        }
     }
 
     return (cur_mb->bmi + b - 1)->as_mode;
@@ -135,7 +149,22 @@ static B_PREDICTION_MODE above_block_mode(const MODE_INFO *cur_mb, int b, int mi
     {
         /* On top edge, get from MB above us */
         cur_mb -= mi_stride;
-        b += 16;
+
+        switch (cur_mb->mbmi.mode)
+        {
+            case B_PRED:
+              return (cur_mb->bmi + b + 12)->as_mode;
+            case DC_PRED:
+                return B_DC_PRED;
+            case V_PRED:
+                return B_VE_PRED;
+            case H_PRED:
+                return B_HE_PRED;
+            case TM_PRED:
+                return B_TM_PRED;
+            default:
+                return B_DC_PRED;
+        }
     }
 
     return (cur_mb->bmi + b - 4)->as_mode;

vp8/decoder/decodemv.c

@@ -101,36 +101,6 @@ static void vp8_kfread_modes(VP8D_COMP *pbi, MODE_INFO *m, int mb_row, int mb_co
                 }
                 while (++i < 16);
             }
-            else
-            {
-                int BMode;
-                int i = 0;
-
-                switch (y_mode)
-                {
-                case DC_PRED:
-                    BMode = B_DC_PRED;
-                    break;
-                case V_PRED:
-                    BMode = B_VE_PRED;
-                    break;
-                case H_PRED:
-                    BMode = B_HE_PRED;
-                    break;
-                case TM_PRED:
-                    BMode = B_TM_PRED;
-                    break;
-                default:
-                    BMode = B_DC_PRED;
-                    break;
-                }
-
-                do
-                {
-                    m->bmi[i].as_mode = (B_PREDICTION_MODE)BMode;
-                }
-                while (++i < 16);
-            }
 
             m->mbmi.uv_mode = (MB_PREDICTION_MODE)vp8_read_uv_mode(bc, pbi->common.kf_uv_mode_prob);
         }

vp8/encoder/encodeintra.c

@@ -81,30 +81,6 @@ void vp8_encode_intra16x16mby(const VP8_ENCODER_RTCD *rtcd, MACROBLOCK *x)
     RECON_INVOKE(&rtcd->common->recon, recon_mby)
         (IF_RTCD(&rtcd->common->recon), &x->e_mbd);
 
-    // make sure block modes are set the way we want them for context updates
-    for (b = 0; b < 16; b++)
-    {
-        BLOCKD *d = &x->e_mbd.block[b];
-
-        switch (x->e_mbd.mode_info_context->mbmi.mode)
-        {
-        case DC_PRED:
-            d->bmi.as_mode = B_DC_PRED;
-            break;
-        case V_PRED:
-            d->bmi.as_mode = B_VE_PRED;
-            break;
-        case H_PRED:
-            d->bmi.as_mode = B_HE_PRED;
-            break;
-        case TM_PRED:
-            d->bmi.as_mode = B_TM_PRED;
-            break;
-        default:
-            d->bmi.as_mode = B_DC_PRED;
-            break;
-        }
-    }
 }
 
 void vp8_encode_intra16x16mbuv(const VP8_ENCODER_RTCD *rtcd, MACROBLOCK *x)

vp8/encoder/pickinter.c

@@ -140,8 +140,7 @@ static int pick_intra4x4block(
     MACROBLOCK *x,
     int ib,
     B_PREDICTION_MODE *best_mode,
-    B_PREDICTION_MODE above,
-    B_PREDICTION_MODE left,
+    unsigned int *mode_costs,
 
     int *bestrate,
     int *bestdistortion)
@@ -153,16 +152,6 @@ static int pick_intra4x4block(
     int best_rd = INT_MAX;       // 1<<30
     int rate;
     int distortion;
-    unsigned int *mode_costs;
-
-    if (x->e_mbd.frame_type == KEY_FRAME)
-    {
-        mode_costs = x->bmode_costs[above][left];
-    }
-    else
-    {
-        mode_costs = x->inter_bmode_costs;
-    }
 
     for (mode = B_DC_PRED; mode <= B_HE_PRED /*B_HU_PRED*/; mode++)
     {
@@ -202,20 +191,30 @@ static int pick_intra4x4mby_modes
     int cost = mb->mbmode_cost [xd->frame_type] [B_PRED];
     int error;
     int distortion = 0;
+    unsigned int *bmode_costs;
 
     vp8_intra_prediction_down_copy(xd);
 
+    bmode_costs = mb->inter_bmode_costs;
+
     for (i = 0; i < 16; i++)
     {
         MODE_INFO *const mic = xd->mode_info_context;
         const int mis = xd->mode_info_stride;
-        const B_PREDICTION_MODE A = above_block_mode(mic, i, mis);
-        const B_PREDICTION_MODE L = left_block_mode(mic, i);
 
         B_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode);
         int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(d);
 
-        pick_intra4x4block(rtcd, mb, i, &best_mode, A, L, &r, &d);
+        if (mb->e_mbd.frame_type == KEY_FRAME)
+        {
+            const B_PREDICTION_MODE A = above_block_mode(mic, i, mis);
+            const B_PREDICTION_MODE L = left_block_mode(mic, i);
+
+            bmode_costs  = mb->bmode_costs[A][L];
+        }
+
+
+        pick_intra4x4block(rtcd, mb, i, &best_mode, bmode_costs, &r, &d);
 
         cost += r;
         distortion += d;