Restructure spatial direct MV generation so that the zero predictior shortcut
authorMichael Niedermayer <michaelni@gmx.at>
Sat, 13 Feb 2010 13:02:25 +0000 (13:02 +0000)
committerMichael Niedermayer <michaelni@gmx.at>
Sat, 13 Feb 2010 13:02:25 +0000 (13:02 +0000)
is closer at the top.
50-130 cpu cycles faster depending on which kind of direct MBs are meassured.

Originally committed as revision 21788 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/h264_direct.c

index 96445cb..a8fb6b3 100644 (file)
@@ -151,11 +151,72 @@ static void pred_spatial_direct_motion(H264Context * const h, int *mb_type){
     const int is_b8x8 = IS_8X8(*mb_type);
     unsigned int sub_mb_type;
     int i8, i4;
+    int ref[2];
+    int mv[2];
+    int list;
 
     assert(h->ref_list[1][0].reference&3);
 
 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
 
+    *mb_type |= MB_TYPE_L0L1;
+    sub_mb_type |= MB_TYPE_L0L1;
+
+    /* ref = min(neighbors) */
+    for(list=0; list<2; list++){
+        int left_ref = h->ref_cache[list][scan8[0] - 1];
+        int top_ref  = h->ref_cache[list][scan8[0] - 8];
+        int refc = h->ref_cache[list][scan8[0] - 8 + 4];
+        const int16_t *C= h->mv_cache[list][ scan8[0] - 8 + 4];
+        if(refc == PART_NOT_AVAILABLE){
+            refc = h->ref_cache[list][scan8[0] - 8 - 1];
+            C    = h-> mv_cache[list][scan8[0] - 8 - 1];
+        }
+        ref[list] = FFMIN3((unsigned)left_ref, (unsigned)top_ref, (unsigned)refc);
+        if(ref[list] >= 0){
+            //this is just pred_motion() but with the cases removed that cannot happen for direct blocks
+            const int16_t * const A= h->mv_cache[list][ scan8[0] - 1 ];
+            const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
+
+            int match_count= (left_ref==ref[list]) + (top_ref==ref[list]) + (refc==ref[list]);
+            if(match_count > 1){ //most common
+                mv[list]= (mid_pred(A[0], B[0], C[0])&0xFFFF)
+                            +(mid_pred(A[1], B[1], C[1])<<16);
+            }else {
+                assert(match_count==1);
+                if(left_ref==ref[list]){
+                    mv[list]= *(uint32_t*)A;
+                }else if(top_ref==ref[list]){
+                    mv[list]= *(uint32_t*)B;
+                }else{
+                    mv[list]= *(uint32_t*)C;
+                }
+            }
+        }else{
+            int mask= ~(MB_TYPE_L0 << (2*list));
+            mv[list] = 0;
+            ref[list] = -1;
+            if(!is_b8x8)
+                *mb_type &= mask;
+            sub_mb_type &= mask;
+        }
+    }
+    if(ref[0] < 0 && ref[1] < 0){
+        ref[0] = ref[1] = 0;
+        if(!is_b8x8)
+            *mb_type |= MB_TYPE_L0L1;
+        sub_mb_type |= MB_TYPE_L0L1;
+    }
+
+    if(!is_b8x8 && (mv[0]|mv[1]) == 0){
+        fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
+        fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
+        fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
+        fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
+        *mb_type= (*mb_type & ~(MB_TYPE_8x8|MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_P1L0|MB_TYPE_P1L1))|MB_TYPE_16x16|MB_TYPE_DIRECT2;
+        return;
+    }
+
     if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
         if(!IS_INTERLACED(*mb_type)){                    //     AFR/FR    -> AFL/FL
             mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride;
@@ -172,31 +233,31 @@ static void pred_spatial_direct_motion(H264Context * const h, int *mb_type){
             b8_stride *= 3;
             b4_stride *= 6;
 
-            sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
+            sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
                 && !is_b8x8){
-                *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
+                *mb_type   |= MB_TYPE_16x8 |MB_TYPE_DIRECT2; /* B_16x8 */
             }else{
-                *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
+                *mb_type   |= MB_TYPE_8x8;
             }
         }else{                                           //     AFR/FR    -> AFR/FR
 single_col:
             mb_type_col[0] =
             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
 
-            sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
+            sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
             if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
-                *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
+                *mb_type   |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_16x16 */
             }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
-                *mb_type   |= MB_TYPE_L0L1|MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
+                *mb_type   |= MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
             }else{
                 if(!h->sps.direct_8x8_inference_flag){
                     /* FIXME save sub mb types from previous frames (or derive from MVs)
                     * so we know exactly what block size to use */
-                    sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
+                    sub_mb_type += (MB_TYPE_8x8-MB_TYPE_16x16); /* B_SUB_4x4 */
                 }
-                *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
+                *mb_type   |= MB_TYPE_8x8;
             }
         }
     }
@@ -214,64 +275,8 @@ single_col:
         }
     }
 
-    {
-        int ref[2];
-        int mv[2];
-        int list;
-
-        /* ref = min(neighbors) */
-        for(list=0; list<2; list++){
-            int left_ref = h->ref_cache[list][scan8[0] - 1];
-            int top_ref  = h->ref_cache[list][scan8[0] - 8];
-            int refc = h->ref_cache[list][scan8[0] - 8 + 4];
-            const int16_t *C= h->mv_cache[list][ scan8[0] - 8 + 4];
-            if(refc == PART_NOT_AVAILABLE){
-                refc = h->ref_cache[list][scan8[0] - 8 - 1];
-                C    = h-> mv_cache[list][scan8[0] - 8 - 1];
-            }
-            ref[list] = FFMIN3((unsigned)left_ref, (unsigned)top_ref, (unsigned)refc);
-            if(ref[list] >= 0){
-                //this is just pred_motion() but with the cases removed that cannot happen for direct blocks
-                const int16_t * const A= h->mv_cache[list][ scan8[0] - 1 ];
-                const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
-
-                int match_count= (left_ref==ref[list]) + (top_ref==ref[list]) + (refc==ref[list]);
-                if(match_count > 1){ //most common
-                    mv[list]= (mid_pred(A[0], B[0], C[0])&0xFFFF)
-                             +(mid_pred(A[1], B[1], C[1])<<16);
-                }else {
-                    assert(match_count==1);
-                    if(left_ref==ref[list]){
-                        mv[list]= *(uint32_t*)A;
-                    }else if(top_ref==ref[list]){
-                        mv[list]= *(uint32_t*)B;
-                    }else{
-                        mv[list]= *(uint32_t*)C;
-                    }
-                }
-            }else{
-                int mask= ~(MB_TYPE_L0 << (2*list));
-                mv[list] = 0;
-                ref[list] = -1;
-                if(!is_b8x8)
-                    *mb_type &= mask;
-                sub_mb_type &= mask;
-            }
-        }
-        if(ref[0] < 0 && ref[1] < 0){
-            ref[0] = ref[1] = 0;
-            if(!is_b8x8)
-                *mb_type |= MB_TYPE_L0L1;
-            sub_mb_type |= MB_TYPE_L0L1;
-        }
 
-        if(!is_b8x8 && (mv[0]|mv[1]) == 0){
-            fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
-            fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
-            fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
-            fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
-            *mb_type= (*mb_type & ~(MB_TYPE_8x8|MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_P1L0|MB_TYPE_P1L1))|MB_TYPE_16x16|MB_TYPE_DIRECT2;
-        }else if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
+        if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
             int n=0;
             for(i8=0; i8<4; i8++){
                 int x8 = i8&1;