Rearchitecturing the stiched up goose part 1

Run loop filter per row instead of per MB, this also should make it
much easier to switch to per frame filtering and also doing so in a
seperate thread in the future if some volunteer wants to try.
Overall decoding speedup of 1.7% (single thread on pentium dual / cathedral sample)
This change also allows some optimizations to be tried that would not have
been possible before.

Originally committed as revision 21270 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/h264.c

@@ -666,6 +666,7 @@ static void free_tables(H264Context *h){
     av_freep(&h->non_zero_count);
     av_freep(&h->slice_table_base);
     h->slice_table= NULL;
+    av_freep(&h->list_counts);
 
     av_freep(&h->mb2b_xy);
     av_freep(&h->mb2b8_xy);
@@ -756,7 +757,7 @@ int ff_h264_alloc_tables(H264Context *h){
 
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t), fail)
 
-    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t), fail)
+    FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 32 * sizeof(uint8_t), fail)
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 
@@ -764,6 +765,7 @@ int ff_h264_alloc_tables(H264Context *h){
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t), fail);
     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 32*big_mb_num * sizeof(uint16_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)
 
     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
@@ -945,12 +947,7 @@ int ff_h264_frame_start(H264Context *h){
 
 static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
     MpegEncContext * const s = &h->s;
-    int i;
-    int step    = 1;
-    int offset  = 1;
-    int uvoffset= 1;
     int top_idx = 1;
-    int skiplast= 0;
 
     src_y  -=   linesize;
     src_cb -= uvlinesize;
@@ -958,8 +955,6 @@ static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src
 
     if(!simple && FRAME_MBAFF){
         if(s->mb_y&1){
-            offset  = MB_MBAFF ? 1 : 17;
-            uvoffset= MB_MBAFF ? 1 : 9;
             if(!MB_MBAFF){
                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 0)= *(uint64_t*)(src_y +  15*linesize);
                 *(uint64_t*)(h->top_borders[0][s->mb_x]+ 8)= *(uint64_t*)(src_y +8+15*linesize);
@@ -968,39 +963,19 @@ static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src
                     *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+7*uvlinesize);
                 }
             }
-        }else{
-            if(!MB_MBAFF){
-                h->left_border[0]= h->top_borders[0][s->mb_x][15];
-                if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
-                    h->left_border[34   ]= h->top_borders[0][s->mb_x][16+7  ];
-                    h->left_border[34+18]= h->top_borders[0][s->mb_x][16+8+7];
-                }
-                skiplast= 1;
-            }
-            offset  =
-            uvoffset=
-            top_idx = MB_MBAFF ? 0 : 1;
-        }
-        step= MB_MBAFF ? 2 : 1;
+        }else if(MB_MBAFF){
+            top_idx = 0;
+        }else
+            return;
     }
 
     // There are two lines saved, the line above the the top macroblock of a pair,
     // and the line above the bottom macroblock
-    h->left_border[offset]= h->top_borders[top_idx][s->mb_x][15];
-    for(i=1; i<17 - skiplast; i++){
-        h->left_border[offset+i*step]= src_y[15+i*  linesize];
-    }
 
     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
     *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
 
     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
-        h->left_border[uvoffset+34   ]= h->top_borders[top_idx][s->mb_x][16+7];
-        h->left_border[uvoffset+34+18]= h->top_borders[top_idx][s->mb_x][24+7];
-        for(i=1; i<9 - skiplast; i++){
-            h->left_border[uvoffset+34   +i*step]= src_cb[7+i*uvlinesize];
-            h->left_border[uvoffset+34+18+i*step]= src_cr[7+i*uvlinesize];
-        }
         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
         *(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
     }
@@ -1013,21 +988,15 @@ static inline void xchg_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_c
     int deblock_left;
     int deblock_top;
     int mb_xy;
-    int step    = 1;
-    int offset  = 1;
-    int uvoffset= 1;
     int top_idx = 1;
 
     if(!simple && FRAME_MBAFF){
         if(s->mb_y&1){
-            offset  = MB_MBAFF ? 1 : 17;
-            uvoffset= MB_MBAFF ? 1 : 9;
+            if(!MB_MBAFF)
+                return;
         }else{
-            offset  =
-            uvoffset=
             top_idx = MB_MBAFF ? 0 : 1;
         }
-        step= MB_MBAFF ? 2 : 1;
     }
 
     if(h->deblocking_filter == 2) {
@@ -1049,14 +1018,10 @@ if(xchg)\
     a= b;\
 b= t;
 
-    if(deblock_left){
-        for(i = !deblock_top; i<16; i++){
-            XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, xchg);
-        }
-        XCHG(h->left_border[offset+i*step], src_y [i*  linesize], temp8, 1);
-    }
-
     if(deblock_top){
+        if(deblock_left){
+            XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x-1]+8), *(uint64_t*)(src_y -7), temp64, 1);
+        }
         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
         XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
         if(s->mb_x+1 < s->mb_width){
@@ -1065,15 +1030,11 @@ b= t;
     }
 
     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
-        if(deblock_left){
-            for(i = !deblock_top; i<8; i++){
-                XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, xchg);
-                XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, xchg);
-            }
-            XCHG(h->left_border[uvoffset+34   +i*step], src_cb[i*uvlinesize], temp8, 1);
-            XCHG(h->left_border[uvoffset+34+18+i*step], src_cr[i*uvlinesize], temp8, 1);
-        }
         if(deblock_top){
+            if(deblock_left){
+                XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x-1]+16), *(uint64_t*)(src_cb -7), temp64, 1);
+                XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x-1]+24), *(uint64_t*)(src_cr -7), temp64, 1);
+            }
             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
             XCHG(*(uint64_t*)(h->top_borders[top_idx][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
         }
@@ -1103,6 +1064,8 @@ static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
 
+    h->list_counts[mb_xy]= h->list_count;
+
     if (!simple && MB_FIELD) {
         linesize   = h->mb_linesize   = s->linesize * 2;
         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
@@ -1322,7 +1285,7 @@ static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
     if(h->cbp || IS_INTRA(mb_type))
         s->dsp.clear_blocks(h->mb);
 
-    if(h->deblocking_filter) {
+    if(h->deblocking_filter && 0) {
         backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
         fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
         h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
@@ -2174,6 +2137,70 @@ int ff_h264_get_slice_type(H264Context *h)
     }
 }
 
+static void loop_filter(H264Context *h){
+    MpegEncContext * const s = &h->s;
+    uint8_t  *dest_y, *dest_cb, *dest_cr;
+    int linesize, uvlinesize, mb_x, mb_y;
+    const int end_mb_y= s->mb_y + FRAME_MBAFF;
+    const int old_slice_type= h->slice_type;
+
+    if(h->deblocking_filter) {
+        for(mb_x= 0; mb_x<s->mb_width; mb_x++){
+            for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
+                int list, mb_xy, mb_type, is_complex;
+                mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
+                h->slice_num= h->slice_table[mb_xy];
+                mb_type= s->current_picture.mb_type[mb_xy];
+                h->list_count= h->list_counts[mb_xy];
+                if(h->list_count==2){
+                    h->slice_type= h->slice_type_nos= FF_B_TYPE;
+                }else if(h->list_count==1){
+                    h->slice_type= h->slice_type_nos= FF_P_TYPE;
+                }else
+                    h->slice_type= h->slice_type_nos= FF_I_TYPE;
+
+                if(FRAME_MBAFF)
+                    h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
+
+                is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0; //FIXME qscale might be wrong
+
+                s->mb_x= mb_x;
+                s->mb_y= mb_y;
+                dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
+                dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
+                dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
+                    //FIXME simplify above
+
+                if (MB_FIELD) {
+                    linesize   = h->mb_linesize   = s->linesize * 2;
+                    uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
+                    if(mb_y&1){ //FIXME move out of this function?
+                        dest_y -= s->linesize*15;
+                        dest_cb-= s->uvlinesize*7;
+                        dest_cr-= s->uvlinesize*7;
+                    }
+                } else {
+                    linesize   = h->mb_linesize   = s->linesize;
+                    uvlinesize = h->mb_uvlinesize = s->uvlinesize;
+                }
+                backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, !is_complex);
+                fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
+                h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
+                h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
+
+                if (is_complex && FRAME_MBAFF) {
+                    ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
+                } else {
+                    ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
+                }
+            }
+        }
+    }
+    h->slice_type= old_slice_type;
+    s->mb_x= 0;
+    s->mb_y= end_mb_y - FRAME_MBAFF;
+}
+
 static int decode_slice(struct AVCodecContext *avctx, void *arg){
     H264Context *h = *(void**)arg;
     MpegEncContext * const s = &h->s;
@@ -2222,6 +2249,7 @@ static int decode_slice(struct AVCodecContext *avctx, void *arg){
 
             if( ++s->mb_x >= s->mb_width ) {
                 s->mb_x = 0;
+                loop_filter(h);
                 ff_draw_horiz_band(s, 16*s->mb_y, 16);
                 ++s->mb_y;
                 if(FIELD_OR_MBAFF_PICTURE) {
@@ -2259,6 +2287,7 @@ static int decode_slice(struct AVCodecContext *avctx, void *arg){
 
             if(++s->mb_x >= s->mb_width){
                 s->mb_x=0;
+                loop_filter(h);
                 ff_draw_horiz_band(s, 16*s->mb_y, 16);
                 ++s->mb_y;
                 if(FIELD_OR_MBAFF_PICTURE) {

libavcodec/h264.h

@@ -300,7 +300,7 @@ typedef struct H264Context{
      * is 64 if not available.
      */
     DECLARE_ALIGNED_8(uint8_t, non_zero_count_cache[6*8]);
-    uint8_t (*non_zero_count)[16];
+    uint8_t (*non_zero_count)[32];
 
     /**
      * Motion vector cache.
@@ -423,6 +423,7 @@ typedef struct H264Context{
      */
     unsigned int ref_count[2];   ///< counts frames or fields, depending on current mb mode
     unsigned int list_count;
+    uint8_t *list_counts;            ///< Array of list_count per MB specifying the slice type
     Picture *short_ref[32];
     Picture *long_ref[32];
     Picture default_ref_list[2][32]; ///< base reference list for all slices of a coded picture
@@ -736,8 +737,8 @@ static void fill_caches(H264Context *h, int mb_type, int for_deblock){
     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
 
     //FIXME deblocking could skip the intra and nnz parts.
-    if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
-        return;
+//     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
+//         return;
 
     /* Wow, what a mess, why didn't they simplify the interlacing & intra
      * stuff, I can't imagine that these complex rules are worth it. */
@@ -793,20 +794,33 @@ static void fill_caches(H264Context *h, int mb_type, int for_deblock){
         left_type[0] = h->slice_table[left_xy[0] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[0]] : 0;
         left_type[1] = h->slice_table[left_xy[1] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[1]] : 0;
 
-        if(MB_MBAFF && !IS_INTRA(mb_type)){
+        if(!IS_INTRA(mb_type)){
             int list;
             for(list=0; list<h->list_count; list++){
-                //These values where changed for ease of performing MC, we need to change them back
-                //FIXME maybe we can make MC and loop filter use the same values or prevent
-                //the MC code from changing ref_cache and rather use a temporary array.
-                if(USES_LIST(mb_type,list)){
-                    int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
+                int8_t *ref;
+                int y, b_xy;
+                if(!USES_LIST(mb_type, list)){
+                    fill_rectangle(  h->mv_cache[list][scan8[0]], 4, 4, 8, pack16to32(0,0), 4);
                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
-                    *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
-                    ref += h->b8_stride;
+                    *(uint32_t*)&h->ref_cache[list][scan8[ 2]] =
                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
-                    *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
+                    *(uint32_t*)&h->ref_cache[list][scan8[10]] = ((LIST_NOT_USED)&0xFF)*0x01010101;
+                    continue;
                 }
+
+                ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
+                *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
+                *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
+                ref += h->b8_stride;
+                *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
+                *(uint32_t*)&h->ref_cache[list][scan8[10]] = (pack16to32(ref[0],ref[1])&0x00FF00FF)*0x0101;
+
+                b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
+                for(y=0; y<4; y++){
+                    *(uint64_t*)h->mv_cache[list][scan8[0]+0 + 8*y]= *(uint64_t*)s->current_picture.motion_val[list][b_xy + 0 + y*h->b_stride];
+                    *(uint64_t*)h->mv_cache[list][scan8[0]+2 + 8*y]= *(uint64_t*)s->current_picture.motion_val[list][b_xy + 2 + y*h->b_stride];
+                }
+
             }
         }
     }else{
@@ -1196,6 +1210,23 @@ static inline void write_back_non_zero_count(H264Context *h){
     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
+
+    //FIXME sort better how things are stored in non_zero_count
+
+
+    h->non_zero_count[mb_xy][13]= h->non_zero_count_cache[6+8*1];
+    h->non_zero_count[mb_xy][14]= h->non_zero_count_cache[6+8*2];
+    h->non_zero_count[mb_xy][15]= h->non_zero_count_cache[6+8*3];
+    h->non_zero_count[mb_xy][16]= h->non_zero_count_cache[5+8*1];
+    h->non_zero_count[mb_xy][17]= h->non_zero_count_cache[5+8*2];
+    h->non_zero_count[mb_xy][18]= h->non_zero_count_cache[5+8*3];
+    h->non_zero_count[mb_xy][19]= h->non_zero_count_cache[4+8*1];
+    h->non_zero_count[mb_xy][20]= h->non_zero_count_cache[4+8*2];
+    h->non_zero_count[mb_xy][21]= h->non_zero_count_cache[4+8*3];
+
+    h->non_zero_count[mb_xy][22]= h->non_zero_count_cache[1+8*1];
+    h->non_zero_count[mb_xy][23]= h->non_zero_count_cache[1+8*4];
+
 }
 
 static inline void write_back_motion(H264Context *h, int mb_type){
@@ -1271,7 +1302,7 @@ static void decode_mb_skip(H264Context *h){
     const int mb_xy= h->mb_xy;
     int mb_type=0;
 
-    memset(h->non_zero_count[mb_xy], 0, 16);
+    memset(h->non_zero_count[mb_xy], 0, 32);
     memset(h->non_zero_count_cache + 8, 0, 8*5); //FIXME ugly, remove pfui
 
     if(MB_FIELD)

libavcodec/h264_cabac.c

@@ -1392,7 +1392,7 @@ decode_intra_mb:
         // In deblocking, the quantizer is 0
         s->current_picture.qscale_table[mb_xy]= 0;
         // All coeffs are present
-        memset(h->non_zero_count[mb_xy], 16, 16);
+        memset(h->non_zero_count[mb_xy], 16, 32);
         s->current_picture.mb_type[mb_xy]= mb_type;
         h->last_qscale_diff = 0;
         return 0;

libavcodec/h264_cavlc.c

@@ -620,7 +620,7 @@ decode_intra_mb:
         // In deblocking, the quantizer is 0
         s->current_picture.qscale_table[mb_xy]= 0;
         // All coeffs are present
-        memset(h->non_zero_count[mb_xy], 16, 16);
+        memset(h->non_zero_count[mb_xy], 16, 32);
 
         s->current_picture.mb_type[mb_xy]= mb_type;
         return 0;

libavcodec/h264_loopfilter.c

@@ -620,7 +620,7 @@ static av_always_inline void filter_mb_dir(H264Context *h, int mb_x, int mb_y, u
         // Do not use s->qscale as luma quantizer because it has not the same
         // value in IPCM macroblocks.
         qp = ( s->current_picture.qscale_table[mb_xy] + s->current_picture.qscale_table[mbn_xy] + 1 ) >> 1;
-        //tprintf(s->avctx, "filter mb:%d/%d dir:%d edge:%d, QPy:%d, QPc:%d, QPcn:%d\n", mb_x, mb_y, dir, edge, qp, h->chroma_qp, s->current_picture.qscale_table[mbn_xy]);
+        //tprintf(s->avctx, "filter mb:%d/%d dir:%d edge:%d, QPy:%d, QPc:%d, QPcn:%d\n", mb_x, mb_y, dir, edge, qp, h->chroma_qp[0], s->current_picture.qscale_table[mbn_xy]);
         tprintf(s->avctx, "filter mb:%d/%d dir:%d edge:%d, QPy:%d ls:%d uvls:%d", mb_x, mb_y, dir, edge, qp, linesize, uvlinesize);
         //{ int i; for (i = 0; i < 4; i++) tprintf(s->avctx, " bS[%d]:%d", i, bS[i]); tprintf(s->avctx, "\n"); }
         if( dir == 0 ) {
@@ -650,6 +650,7 @@ void ff_h264_filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint
     const int mvy_limit = IS_INTERLACED(mb_type) ? 2 : 4;
     int first_vertical_edge_done = 0;
     av_unused int dir;
+    int list;
 
     //for sufficiently low qp, filtering wouldn't do anything
     //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
@@ -663,6 +664,35 @@ void ff_h264_filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint
         }
     }
 
+    h->non_zero_count_cache[7+8*1]=h->non_zero_count[mb_xy][0];
+    h->non_zero_count_cache[7+8*2]=h->non_zero_count[mb_xy][1];
+    h->non_zero_count_cache[7+8*3]=h->non_zero_count[mb_xy][2];
+    h->non_zero_count_cache[7+8*4]=h->non_zero_count[mb_xy][3];
+    h->non_zero_count_cache[4+8*4]=h->non_zero_count[mb_xy][4];
+    h->non_zero_count_cache[5+8*4]=h->non_zero_count[mb_xy][5];
+    h->non_zero_count_cache[6+8*4]=h->non_zero_count[mb_xy][6];
+
+    h->non_zero_count_cache[1+8*2]=h->non_zero_count[mb_xy][9];
+    h->non_zero_count_cache[2+8*2]=h->non_zero_count[mb_xy][8];
+    h->non_zero_count_cache[2+8*1]=h->non_zero_count[mb_xy][7];
+
+    h->non_zero_count_cache[1+8*5]=h->non_zero_count[mb_xy][12];
+    h->non_zero_count_cache[2+8*5]=h->non_zero_count[mb_xy][11];
+    h->non_zero_count_cache[2+8*4]=h->non_zero_count[mb_xy][10];
+
+    h->non_zero_count_cache[6+8*1]=h->non_zero_count[mb_xy][13];
+    h->non_zero_count_cache[6+8*2]=h->non_zero_count[mb_xy][14];
+    h->non_zero_count_cache[6+8*3]=h->non_zero_count[mb_xy][15];
+    h->non_zero_count_cache[5+8*1]=h->non_zero_count[mb_xy][16];
+    h->non_zero_count_cache[5+8*2]=h->non_zero_count[mb_xy][17];
+    h->non_zero_count_cache[5+8*3]=h->non_zero_count[mb_xy][18];
+    h->non_zero_count_cache[4+8*1]=h->non_zero_count[mb_xy][19];
+    h->non_zero_count_cache[4+8*2]=h->non_zero_count[mb_xy][20];
+    h->non_zero_count_cache[4+8*3]=h->non_zero_count[mb_xy][21];
+
+    h->non_zero_count_cache[1+8*1]=h->non_zero_count[mb_xy][22];
+    h->non_zero_count_cache[1+8*4]=h->non_zero_count[mb_xy][23];
+
     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
     if(!h->pps.cabac && h->pps.transform_8x8_mode){
         int top_type, left_type[2];
@@ -687,16 +717,16 @@ void ff_h264_filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint
 
         if(IS_8x8DCT(mb_type)){
             h->non_zero_count_cache[scan8[0   ]]= h->non_zero_count_cache[scan8[1   ]]=
-            h->non_zero_count_cache[scan8[2   ]]= h->non_zero_count_cache[scan8[3   ]]= h->cbp & 1;
+            h->non_zero_count_cache[scan8[2   ]]= h->non_zero_count_cache[scan8[3   ]]= h->cbp_table[mb_xy] & 1;
 
             h->non_zero_count_cache[scan8[0+ 4]]= h->non_zero_count_cache[scan8[1+ 4]]=
-            h->non_zero_count_cache[scan8[2+ 4]]= h->non_zero_count_cache[scan8[3+ 4]]= h->cbp & 2;
+            h->non_zero_count_cache[scan8[2+ 4]]= h->non_zero_count_cache[scan8[3+ 4]]= h->cbp_table[mb_xy] & 2;
 
             h->non_zero_count_cache[scan8[0+ 8]]= h->non_zero_count_cache[scan8[1+ 8]]=
-            h->non_zero_count_cache[scan8[2+ 8]]= h->non_zero_count_cache[scan8[3+ 8]]= h->cbp & 4;
+            h->non_zero_count_cache[scan8[2+ 8]]= h->non_zero_count_cache[scan8[3+ 8]]= h->cbp_table[mb_xy] & 4;
 
             h->non_zero_count_cache[scan8[0+12]]= h->non_zero_count_cache[scan8[1+12]]=
-            h->non_zero_count_cache[scan8[2+12]]= h->non_zero_count_cache[scan8[3+12]]= h->cbp & 8;
+            h->non_zero_count_cache[scan8[2+12]]= h->non_zero_count_cache[scan8[3+12]]= h->cbp_table[mb_xy] & 8;
         }
     }