Cut the size of mvd_table by yet another factor of 2.

The code read/write code itself was 1 cycle faster, overall its
likely more due to cache effects


git-svn-id: file:///var/local/repositories/ffmpeg/trunk@22048 9553f0bf-9b14-0410-a0b8-cfaf0461ba5b

libavcodec/h264.c

@@ -757,8 +757,8 @@ int ff_h264_alloc_tables(H264Context *h){
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
-    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 32*big_mb_num * sizeof(uint8_t), fail);
-    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 32*big_mb_num * sizeof(uint8_t), fail);
+    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*big_mb_num * sizeof(uint8_t), fail);
+    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*big_mb_num * sizeof(uint8_t), fail);
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 32*big_mb_num * sizeof(uint8_t) , fail);
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 
@@ -775,7 +775,7 @@ int ff_h264_alloc_tables(H264Context *h){
             const int b8_xy= 2*x + 2*y*h->b8_stride;
 
             h->mb2b_xy [mb_xy]= b_xy;
-            h->mb2br_xy[mb_xy]= FMO ? b_xy : (b_xy % (8*h->b_stride));
+            h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
             h->mb2b8_xy[mb_xy]= b8_xy;
         }
     }

libavcodec/h264.h

@@ -1070,23 +1070,23 @@ static void fill_decode_caches(H264Context *h, int mb_type){
             if( CABAC ) {
                 /* XXX beurk, Load mvd */
                 if(USES_LIST(top_type, list)){
-                    const int b_xy= h->mb2br_xy[top_xy] + 3*h->b_stride;
+                    const int b_xy= h->mb2br_xy[top_xy];
                     AV_COPY64(h->mvd_cache[list][scan8[0] + 0 - 1*8], h->mvd_table[list][b_xy + 0]);
                 }else{
                     AV_ZERO64(h->mvd_cache[list][scan8[0] + 0 - 1*8]);
                 }
                 if(USES_LIST(left_type[0], list)){
-                    const int b_xy= h->mb2br_xy[left_xy[0]] + 3;
-                    AV_COPY16(h->mvd_cache[list][scan8[0] - 1 + 0*8], h->mvd_table[list][b_xy + h->b_stride*left_block[0]]);
-                    AV_COPY16(h->mvd_cache[list][scan8[0] - 1 + 1*8], h->mvd_table[list][b_xy + h->b_stride*left_block[1]]);
+                    const int b_xy= h->mb2br_xy[left_xy[0]] + 6;
+                    AV_COPY16(h->mvd_cache[list][scan8[0] - 1 + 0*8], h->mvd_table[list][b_xy - left_block[0]]);
+                    AV_COPY16(h->mvd_cache[list][scan8[0] - 1 + 1*8], h->mvd_table[list][b_xy - left_block[1]]);
                 }else{
                     AV_ZERO16(h->mvd_cache [list][scan8[0] - 1 + 0*8]);
                     AV_ZERO16(h->mvd_cache [list][scan8[0] - 1 + 1*8]);
                 }
                 if(USES_LIST(left_type[1], list)){
-                    const int b_xy= h->mb2br_xy[left_xy[1]] + 3;
-                    AV_COPY16(h->mvd_cache[list][scan8[0] - 1 + 2*8], h->mvd_table[list][b_xy + h->b_stride*left_block[2]]);
-                    AV_COPY16(h->mvd_cache[list][scan8[0] - 1 + 3*8], h->mvd_table[list][b_xy + h->b_stride*left_block[3]]);
+                    const int b_xy= h->mb2br_xy[left_xy[1]] + 6;
+                    AV_COPY16(h->mvd_cache[list][scan8[0] - 1 + 2*8], h->mvd_table[list][b_xy - left_block[2]]);
+                    AV_COPY16(h->mvd_cache[list][scan8[0] - 1 + 3*8], h->mvd_table[list][b_xy - left_block[3]]);
                 }else{
                     AV_ZERO16(h->mvd_cache [list][scan8[0] - 1 + 2*8]);
                     AV_ZERO16(h->mvd_cache [list][scan8[0] - 1 + 3*8]);
@@ -1424,13 +1424,15 @@ static inline void write_back_motion(H264Context *h, int mb_type){
             AV_COPY128(mv_dst + y*b_stride, mv_src + 8*y);
         }
         if( CABAC ) {
-            uint8_t (*mvd_dst)[2] = &h->mvd_table[list][FMO ? b_xy : h->mb2br_xy[h->mb_xy]];
+            uint8_t (*mvd_dst)[2] = &h->mvd_table[list][FMO ? 8*h->mb_xy : h->mb2br_xy[h->mb_xy]];
             uint8_t (*mvd_src)[2] = &h->mvd_cache[list][scan8[0]];
             if(IS_SKIP(mb_type))
-                fill_rectangle(mvd_dst, 4, 4, h->b_stride, 0, 2);
-            else
-            for(y=0; y<4; y++){
-                AV_COPY64(mvd_dst + y*b_stride, mvd_src + 8*y);
+                AV_ZERO128(mvd_dst);
+            else{
+            AV_COPY64(mvd_dst, mvd_src + 8*3);
+            for(y=0; y<3; y++){
+                AV_COPY16(mvd_dst + 3 + 3 - y, mvd_src + 3 + 8*y);
+            }
             }
         }