]> git.sesse.net Git - ffmpeg/blobdiff - libavcodec/h264.h
Free encoder extradata in avcodec_close(). Should fix several small memory
[ffmpeg] / libavcodec / h264.h
index b49f6cc6f5acc80375c7b4edaf5bec9085309c1f..2c677c88b415efd92ca6906d212b5cc632c2bc88 100644 (file)
@@ -28,6 +28,7 @@
 #ifndef AVCODEC_H264_H
 #define AVCODEC_H264_H
 
+#include "libavutil/intreadwrite.h"
 #include "dsputil.h"
 #include "cabac.h"
 #include "mpegvideo.h"
@@ -282,9 +283,19 @@ typedef struct H264Context{
     int chroma_pred_mode;
     int intra16x16_pred_mode;
 
+    int topleft_mb_xy;
     int top_mb_xy;
+    int topright_mb_xy;
     int left_mb_xy[2];
 
+    int topleft_type;
+    int top_type;
+    int topright_type;
+    int left_type[2];
+
+    const uint8_t * left_block;
+    int topleft_partition;
+
     int8_t intra4x4_pred_mode_cache[5*8];
     int8_t (*intra4x4_pred_mode)[8];
     H264PredContext hpc;
@@ -420,6 +431,8 @@ typedef struct H264Context{
     int redundant_pic_count;
 
     int direct_spatial_mv_pred;
+    int col_parity;
+    int col_fieldoff;
     int dist_scale_factor[16];
     int dist_scale_factor_field[2][32];
     int map_col_to_list0[2][16+32];
@@ -578,6 +591,9 @@ typedef struct H264Context{
 
 extern const uint8_t ff_h264_chroma_qp[52];
 
+void ff_svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
+
+void ff_svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
 
 /**
  * Decode SEI
@@ -726,14 +742,10 @@ static inline int get_chroma_qp(H264Context *h, int t, int qscale){
 
 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my);
 
-static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deblock){
+static void fill_decode_neighbors(H264Context *h, int mb_type){
     MpegEncContext * const s = &h->s;
     const int mb_xy= h->mb_xy;
     int topleft_xy, top_xy, topright_xy, left_xy[2];
-    int topleft_type, top_type, topright_type, left_type[2];
-    const uint8_t * left_block;
-    int topleft_partition= -1;
-    int i;
     static const uint8_t left_block_options[4][16]={
         {0,1,2,3,7,10,8,11,7+0*8, 7+1*8, 7+2*8, 7+3*8, 2+0*8, 2+3*8, 2+1*8, 2+2*8},
         {2,2,3,3,8,11,8,11,7+2*8, 7+2*8, 7+3*8, 7+3*8, 2+1*8, 2+2*8, 2+1*8, 2+2*8},
@@ -741,11 +753,9 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
         {0,2,0,2,7,10,7,10,7+0*8, 7+2*8, 7+0*8, 7+2*8, 2+0*8, 2+3*8, 2+0*8, 2+3*8}
     };
 
-    top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
+    h->topleft_partition= -1;
 
-    //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;
+    top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
 
     /* Wow, what a mess, why didn't they simplify the interlacing & intra
      * stuff, I can't imagine that these complex rules are worth it. */
@@ -753,7 +763,7 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
     topleft_xy = top_xy - 1;
     topright_xy= top_xy + 1;
     left_xy[1] = left_xy[0] = mb_xy-1;
-    left_block = left_block_options[0];
+    h->left_block = left_block_options[0];
     if(FRAME_MBAFF){
         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[mb_xy-1]);
         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
@@ -762,12 +772,12 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
                 left_xy[1] = left_xy[0] = mb_xy - s->mb_stride - 1;
                 if (curr_mb_field_flag) {
                     left_xy[1] += s->mb_stride;
-                    left_block = left_block_options[3];
+                    h->left_block = left_block_options[3];
                 } else {
                     topleft_xy += s->mb_stride;
                     // take top left mv from the middle of the mb, as opposed to all other modes which use the bottom right partition
-                    topleft_partition = 0;
-                    left_block = left_block_options[1];
+                    h->topleft_partition = 0;
+                    h->left_block = left_block_options[1];
                 }
             }
         }else{
@@ -780,90 +790,47 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
                 left_xy[1] = left_xy[0] = mb_xy - 1;
                 if (curr_mb_field_flag) {
                     left_xy[1] += s->mb_stride;
-                    left_block = left_block_options[3];
+                    h->left_block = left_block_options[3];
                 } else {
-                    left_block = left_block_options[2];
+                    h->left_block = left_block_options[2];
                 }
             }
         }
     }
 
-    h->top_mb_xy = top_xy;
+    h->topleft_mb_xy = topleft_xy;
+    h->top_mb_xy     = top_xy;
+    h->topright_mb_xy= topright_xy;
     h->left_mb_xy[0] = left_xy[0];
     h->left_mb_xy[1] = left_xy[1];
-    if(for_deblock){
-
-        //for sufficiently low qp, filtering wouldn't do anything
-        //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
-        int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
-        int qp = s->current_picture.qscale_table[mb_xy];
-        if(qp <= qp_thresh
-           && (left_xy[0]<0 || ((qp + s->current_picture.qscale_table[left_xy[0]] + 1)>>1) <= qp_thresh)
-           && (top_xy   < 0 || ((qp + s->current_picture.qscale_table[top_xy    ] + 1)>>1) <= qp_thresh)){
-            if(!FRAME_MBAFF)
-                return 1;
-            if(   (left_xy[0]< 0            || ((qp + s->current_picture.qscale_table[left_xy[0]+s->mb_stride] + 1)>>1) <= qp_thresh)
-               && (top_xy    < s->mb_stride || ((qp + s->current_picture.qscale_table[top_xy    -s->mb_stride] + 1)>>1) <= qp_thresh))
-                return 1;
-        }
-        if(IS_INTRA(mb_type))
-            return 0;
-
-        AV_COPY64(&h->non_zero_count_cache[0+8*1], &h->non_zero_count[mb_xy][ 0]);
-        AV_COPY64(&h->non_zero_count_cache[0+8*2], &h->non_zero_count[mb_xy][ 8]);
-        *((uint32_t*)&h->non_zero_count_cache[0+8*5])= *((uint32_t*)&h->non_zero_count[mb_xy][16]);
-        *((uint32_t*)&h->non_zero_count_cache[4+8*3])= *((uint32_t*)&h->non_zero_count[mb_xy][20]);
-        AV_COPY64(&h->non_zero_count_cache[0+8*4], &h->non_zero_count[mb_xy][24]);
-
-        h->cbp= h->cbp_table[mb_xy];
-
-        top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
-        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;
-
-        {
-            int list;
-            for(list=0; list<h->list_count; list++){
-                int8_t *ref;
-                int y, b_stride;
-                int16_t (*mv_dst)[2];
-                int16_t (*mv_src)[2];
-
-                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]] =
-                    *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
-                    *(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]];
-                {
-                    int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
-                    *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
-                    *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][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(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
-                }
-
-                b_stride = h->b_stride;
-                mv_dst   = &h->mv_cache[list][scan8[0]];
-                mv_src   = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
-                for(y=0; y<4; y++){
-                    AV_COPY128(mv_dst + 8*y, mv_src + y*b_stride);
-                }
+    //FIXME do we need all in the context?
+    h->topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
+    h->top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
+    h->topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
+    h->left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
+    h->left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
+}
 
-            }
-        }
-    }else{
-        topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
-        top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
-        topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
-        left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
-        left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
+static void fill_decode_caches(H264Context *h, int mb_type){
+    MpegEncContext * const s = &h->s;
+    const int mb_xy= h->mb_xy;
+    int topleft_xy, top_xy, topright_xy, left_xy[2];
+    int topleft_type, top_type, topright_type, left_type[2];
+    const uint8_t * left_block= h->left_block;
+    int i;
 
+    topleft_xy   = h->topleft_mb_xy ;
+    top_xy       = h->top_mb_xy     ;
+    topright_xy  = h->topright_mb_xy;
+    left_xy[0]   = h->left_mb_xy[0] ;
+    left_xy[1]   = h->left_mb_xy[1] ;
+    topleft_type = h->topleft_type  ;
+    top_type     = h->top_type      ;
+    topright_type= h->topright_type ;
+    left_type[0] = h->left_type[0]  ;
+    left_type[1] = h->left_type[1]  ;
+
+    if(!IS_SKIP(mb_type)){
         if(IS_INTRA(mb_type)){
             int type_mask= h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
             h->topleft_samples_available=
@@ -943,7 +910,6 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
                 }
             }
         }
-    }
 
 
 /*
@@ -956,32 +922,28 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
 */
 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
     if(top_type){
-        *(uint32_t*)&h->non_zero_count_cache[4+8*0]= *(uint32_t*)&h->non_zero_count[top_xy][4+3*8];
-        if(!for_deblock){
+        AV_COPY32(&h->non_zero_count_cache[4+8*0], &h->non_zero_count[top_xy][4+3*8]);
             h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][1+1*8];
             h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][2+1*8];
 
             h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][1+2*8];
             h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][2+2*8];
-        }
-    }else if(!for_deblock){
+    }else {
             h->non_zero_count_cache[1+8*0]=
             h->non_zero_count_cache[2+8*0]=
 
             h->non_zero_count_cache[1+8*3]=
             h->non_zero_count_cache[2+8*3]=
-            *(uint32_t*)&h->non_zero_count_cache[4+8*0]= CABAC && !IS_INTRA(mb_type) ? 0 : 0x40404040;
+            AV_WN32A(&h->non_zero_count_cache[4+8*0], CABAC && !IS_INTRA(mb_type) ? 0 : 0x40404040);
     }
 
     for (i=0; i<2; i++) {
         if(left_type[i]){
             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[8+0+2*i]];
             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[8+1+2*i]];
-            if(!for_deblock){
                 h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[8+4+2*i]];
                 h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[8+5+2*i]];
-            }
-        }else if(!for_deblock){
+        }else{
                 h->non_zero_count_cache[3+8*1 + 2*8*i]=
                 h->non_zero_count_cache[3+8*2 + 2*8*i]=
                 h->non_zero_count_cache[0+8*1 +   8*i]=
@@ -989,54 +951,22 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
         }
     }
 
-    // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
-    if(for_deblock && !CABAC && h->pps.transform_8x8_mode){
-        if(IS_8x8DCT(top_type)){
-            h->non_zero_count_cache[4+8*0]=
-            h->non_zero_count_cache[5+8*0]= h->cbp_table[top_xy] & 4;
-            h->non_zero_count_cache[6+8*0]=
-            h->non_zero_count_cache[7+8*0]= h->cbp_table[top_xy] & 8;
-        }
-        if(IS_8x8DCT(left_type[0])){
-            h->non_zero_count_cache[3+8*1]=
-            h->non_zero_count_cache[3+8*2]= h->cbp_table[left_xy[0]]&2; //FIXME check MBAFF
-        }
-        if(IS_8x8DCT(left_type[1])){
-            h->non_zero_count_cache[3+8*3]=
-            h->non_zero_count_cache[3+8*4]= h->cbp_table[left_xy[1]]&8; //FIXME check MBAFF
-        }
-
-        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[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[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[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;
-        }
-    }
-
-    if( CABAC && !for_deblock) {
+    if( CABAC ) {
         // top_cbp
         if(top_type) {
             h->top_cbp = h->cbp_table[top_xy];
         } else if(IS_INTRA(mb_type)) {
-            h->top_cbp = 0x1C0;
+            h->top_cbp = 0x1CF;
         } else {
-            h->top_cbp = 0;
+            h->top_cbp = 0x00F;
         }
         // left_cbp
         if (left_type[0]) {
             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
         } else if(IS_INTRA(mb_type)) {
-            h->left_cbp = 0x1C0;
+            h->left_cbp = 0x1CF;
         } else {
-            h->left_cbp = 0;
+            h->left_cbp = 0x00F;
         }
         if (left_type[0]) {
             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
@@ -1045,12 +975,13 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
         }
     }
+    }
 
 #if 1
-    if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
+    if(IS_INTER(mb_type) || (IS_DIRECT(mb_type) && h->direct_spatial_mv_pred)){
         int list;
         for(list=0; list<h->list_count; list++){
-            if(!for_deblock && !USES_LIST(mb_type, list) && !IS_DIRECT(mb_type)){
+            if(!USES_LIST(mb_type, list)){
                 /*if(!h->mv_cache_clean[list]){
                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
@@ -1058,27 +989,21 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
                 }*/
                 continue;
             }
+            assert(!(IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred));
+
             h->mv_cache_clean[list]= 0;
 
             if(USES_LIST(top_type, list)){
                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
                 AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
-                if(for_deblock){
-                    int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
-                    h->ref_cache[list][scan8[0] + 0 - 1*8]=
-                    h->ref_cache[list][scan8[0] + 1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
-                    h->ref_cache[list][scan8[0] + 2 - 1*8]=
-                    h->ref_cache[list][scan8[0] + 3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
-                }else{
                     h->ref_cache[list][scan8[0] + 0 - 1*8]=
                     h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
                     h->ref_cache[list][scan8[0] + 2 - 1*8]=
                     h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
-                }
             }else{
                 AV_ZERO128(h->mv_cache[list][scan8[0] + 0 - 1*8]);
-                *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= (((for_deblock||top_type) ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
+                AV_WN32A(&h->ref_cache[list][scan8[0] + 0 - 1*8], ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101);
             }
 
             for(i=0; i<2; i++){
@@ -1086,60 +1011,52 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
                 if(USES_LIST(left_type[i], list)){
                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
-                    *(uint32_t*)h->mv_cache[list][cache_idx  ]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]];
-                    *(uint32_t*)h->mv_cache[list][cache_idx+8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]];
-                    if(for_deblock){
-                        int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[i]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
-                        h->ref_cache[list][cache_idx  ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)]];
-                        h->ref_cache[list][cache_idx+8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)]];
-                    }else{
+                    AV_COPY32(h->mv_cache[list][cache_idx  ], s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]]);
+                    AV_COPY32(h->mv_cache[list][cache_idx+8], s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]]);
                         h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
                         h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
-                    }
                 }else{
-                    *(uint32_t*)h->mv_cache [list][cache_idx  ]=
-                    *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
+                    AV_ZERO32(h->mv_cache [list][cache_idx  ]);
+                    AV_ZERO32(h->mv_cache [list][cache_idx+8]);
                     h->ref_cache[list][cache_idx  ]=
-                    h->ref_cache[list][cache_idx+8]= (for_deblock||left_type[i]) ? LIST_NOT_USED : PART_NOT_AVAILABLE;
+                    h->ref_cache[list][cache_idx+8]= (left_type[i]) ? LIST_NOT_USED : PART_NOT_AVAILABLE;
                 }
             }
 
-            if(for_deblock || ((IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred) && !FRAME_MBAFF))
-                continue;
-
             if(USES_LIST(topleft_type, list)){
-                const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
-                const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
-                *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
+                const int b_xy = h->mb2b_xy [topleft_xy] + 3 + h->b_stride + (h->topleft_partition & 2*h->b_stride);
+                const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (h->topleft_partition & h->b8_stride);
+                AV_COPY32(h->mv_cache[list][scan8[0] - 1 - 1*8], s->current_picture.motion_val[list][b_xy]);
                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
             }else{
-                *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
+                AV_ZERO32(h->mv_cache[list][scan8[0] - 1 - 1*8]);
                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
             }
 
             if(USES_LIST(topright_type, list)){
                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
-                *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
+                AV_COPY32(h->mv_cache[list][scan8[0] + 4 - 1*8], s->current_picture.motion_val[list][b_xy]);
                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
             }else{
-                *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
+                AV_ZERO32(h->mv_cache [list][scan8[0] + 4 - 1*8]);
                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
             }
 
-            if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
+            if((mb_type&(MB_TYPE_SKIP|MB_TYPE_DIRECT2)) && !FRAME_MBAFF)
                 continue;
 
+            if(!(mb_type&(MB_TYPE_SKIP|MB_TYPE_DIRECT2))) {
             h->ref_cache[list][scan8[5 ]+1] =
             h->ref_cache[list][scan8[7 ]+1] =
             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
             h->ref_cache[list][scan8[4 ]] =
             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
-            *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
-            *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
-            *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
-            *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
-            *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
+            AV_ZERO32(h->mv_cache [list][scan8[5 ]+1]);
+            AV_ZERO32(h->mv_cache [list][scan8[7 ]+1]);
+            AV_ZERO32(h->mv_cache [list][scan8[13]+1]); //FIXME remove past 3 (init somewhere else)
+            AV_ZERO32(h->mv_cache [list][scan8[4 ]]);
+            AV_ZERO32(h->mv_cache [list][scan8[12]]);
 
             if( CABAC ) {
                 /* XXX beurk, Load mvd */
@@ -1151,55 +1068,55 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
                 }
                 if(USES_LIST(left_type[0], list)){
                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
-                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 0*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[0]];
-                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[1]];
+                    AV_COPY32(h->mvd_cache[list][scan8[0] - 1 + 0*8], h->mvd_table[list][b_xy + h->b_stride*left_block[0]]);
+                    AV_COPY32(h->mvd_cache[list][scan8[0] - 1 + 1*8], h->mvd_table[list][b_xy + h->b_stride*left_block[1]]);
                 }else{
-                    *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
-                    *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
+                    AV_ZERO32(h->mvd_cache [list][scan8[0] - 1 + 0*8]);
+                    AV_ZERO32(h->mvd_cache [list][scan8[0] - 1 + 1*8]);
                 }
                 if(USES_LIST(left_type[1], list)){
                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
-                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 2*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[2]];
-                    *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 3*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[3]];
+                    AV_COPY32(h->mvd_cache[list][scan8[0] - 1 + 2*8], h->mvd_table[list][b_xy + h->b_stride*left_block[2]]);
+                    AV_COPY32(h->mvd_cache[list][scan8[0] - 1 + 3*8], h->mvd_table[list][b_xy + h->b_stride*left_block[3]]);
                 }else{
-                    *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
-                    *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
+                    AV_ZERO32(h->mvd_cache [list][scan8[0] - 1 + 2*8]);
+                    AV_ZERO32(h->mvd_cache [list][scan8[0] - 1 + 3*8]);
                 }
-                *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
-                *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
-                *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
-                *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
-                *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
+                AV_ZERO32(h->mvd_cache [list][scan8[5 ]+1]);
+                AV_ZERO32(h->mvd_cache [list][scan8[7 ]+1]);
+                AV_ZERO32(h->mvd_cache [list][scan8[13]+1]); //FIXME remove past 3 (init somewhere else)
+                AV_ZERO32(h->mvd_cache [list][scan8[4 ]]);
+                AV_ZERO32(h->mvd_cache [list][scan8[12]]);
 
                 if(h->slice_type_nos == FF_B_TYPE){
-                    fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
+                    fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, MB_TYPE_16x16>>1, 1);
 
                     if(IS_DIRECT(top_type)){
-                        *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
+                        AV_WN32A(&h->direct_cache[scan8[0] - 1*8], 0x01010101*(MB_TYPE_DIRECT2>>1));
                     }else if(IS_8X8(top_type)){
                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
                     }else{
-                        *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
+                        AV_WN32A(&h->direct_cache[scan8[0] - 1*8], 0x01010101*(MB_TYPE_16x16>>1));
                     }
 
                     if(IS_DIRECT(left_type[0]))
-                        h->direct_cache[scan8[0] - 1 + 0*8]= 1;
+                        h->direct_cache[scan8[0] - 1 + 0*8]= MB_TYPE_DIRECT2>>1;
                     else if(IS_8X8(left_type[0]))
                         h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[h->mb2b8_xy[left_xy[0]] + 1 + h->b8_stride*(left_block[0]>>1)];
                     else
-                        h->direct_cache[scan8[0] - 1 + 0*8]= 0;
+                        h->direct_cache[scan8[0] - 1 + 0*8]= MB_TYPE_16x16>>1;
 
                     if(IS_DIRECT(left_type[1]))
-                        h->direct_cache[scan8[0] - 1 + 2*8]= 1;
+                        h->direct_cache[scan8[0] - 1 + 2*8]= MB_TYPE_DIRECT2>>1;
                     else if(IS_8X8(left_type[1]))
                         h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[h->mb2b8_xy[left_xy[1]] + 1 + h->b8_stride*(left_block[2]>>1)];
                     else
-                        h->direct_cache[scan8[0] - 1 + 2*8]= 0;
+                        h->direct_cache[scan8[0] - 1 + 2*8]= MB_TYPE_16x16>>1;
                 }
             }
-
+            }
             if(FRAME_MBAFF){
 #define MAP_MVS\
                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
@@ -1236,13 +1153,7 @@ static av_always_inline int fill_caches(H264Context *h, int mb_type, int for_deb
     }
 #endif
 
-    if(!for_deblock)
         h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
-    return 0;
-}
-
-static void fill_decode_caches(H264Context *h, int mb_type){
-    fill_caches(h, mb_type, 0);
 }
 
 /**
@@ -1250,7 +1161,210 @@ static void fill_decode_caches(H264Context *h, int mb_type){
  * @returns non zero if the loop filter can be skiped
  */
 static int fill_filter_caches(H264Context *h, int mb_type){
-    return fill_caches(h, mb_type, 1);
+    MpegEncContext * const s = &h->s;
+    const int mb_xy= h->mb_xy;
+    int top_xy, left_xy[2];
+    int top_type, left_type[2];
+
+    top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
+
+    //FIXME deblocking could skip the intra and nnz parts.
+
+    /* Wow, what a mess, why didn't they simplify the interlacing & intra
+     * stuff, I can't imagine that these complex rules are worth it. */
+
+    left_xy[1] = left_xy[0] = mb_xy-1;
+    if(FRAME_MBAFF){
+        const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[mb_xy-1]);
+        const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
+        if(s->mb_y&1){
+            if (left_mb_field_flag != curr_mb_field_flag) {
+                left_xy[0] -= s->mb_stride;
+            }
+        }else{
+            if(curr_mb_field_flag){
+                top_xy      += s->mb_stride & (((s->current_picture.mb_type[top_xy    ]>>7)&1)-1);
+            }
+            if (left_mb_field_flag != curr_mb_field_flag) {
+                left_xy[1] += s->mb_stride;
+            }
+        }
+    }
+
+    h->top_mb_xy = top_xy;
+    h->left_mb_xy[0] = left_xy[0];
+    h->left_mb_xy[1] = left_xy[1];
+    {
+        //for sufficiently low qp, filtering wouldn't do anything
+        //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
+        int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
+        int qp = s->current_picture.qscale_table[mb_xy];
+        if(qp <= qp_thresh
+           && (left_xy[0]<0 || ((qp + s->current_picture.qscale_table[left_xy[0]] + 1)>>1) <= qp_thresh)
+           && (top_xy   < 0 || ((qp + s->current_picture.qscale_table[top_xy    ] + 1)>>1) <= qp_thresh)){
+            if(!FRAME_MBAFF)
+                return 1;
+            if(   (left_xy[0]< 0            || ((qp + s->current_picture.qscale_table[left_xy[1]             ] + 1)>>1) <= qp_thresh)
+               && (top_xy    < s->mb_stride || ((qp + s->current_picture.qscale_table[top_xy    -s->mb_stride] + 1)>>1) <= qp_thresh))
+                return 1;
+        }
+    }
+
+    if(h->deblocking_filter == 2){
+        h->top_type    = top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
+        h->left_type[0]= left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
+        h->left_type[1]= left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
+    }else{
+        h->top_type    = top_type     = h->slice_table[top_xy     ] < 0xFFFF ? s->current_picture.mb_type[top_xy]     : 0;
+        h->left_type[0]= left_type[0] = h->slice_table[left_xy[0] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[0]] : 0;
+        h->left_type[1]= left_type[1] = h->slice_table[left_xy[1] ] < 0xFFFF ? s->current_picture.mb_type[left_xy[1]] : 0;
+    }
+    if(IS_INTRA(mb_type))
+        return 0;
+
+    AV_COPY64(&h->non_zero_count_cache[0+8*1], &h->non_zero_count[mb_xy][ 0]);
+    AV_COPY64(&h->non_zero_count_cache[0+8*2], &h->non_zero_count[mb_xy][ 8]);
+    AV_COPY32(&h->non_zero_count_cache[0+8*5], &h->non_zero_count[mb_xy][16]);
+    AV_COPY32(&h->non_zero_count_cache[4+8*3], &h->non_zero_count[mb_xy][20]);
+    AV_COPY64(&h->non_zero_count_cache[0+8*4], &h->non_zero_count[mb_xy][24]);
+
+    h->cbp= h->cbp_table[mb_xy];
+
+    {
+        int list;
+        for(list=0; list<h->list_count; list++){
+            int8_t *ref;
+            int y, b_stride;
+            int16_t (*mv_dst)[2];
+            int16_t (*mv_src)[2];
+
+            if(!USES_LIST(mb_type, list)){
+                fill_rectangle(  h->mv_cache[list][scan8[0]], 4, 4, 8, pack16to32(0,0), 4);
+                AV_WN32A(&h->ref_cache[list][scan8[ 0]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
+                AV_WN32A(&h->ref_cache[list][scan8[ 2]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
+                AV_WN32A(&h->ref_cache[list][scan8[ 8]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
+                AV_WN32A(&h->ref_cache[list][scan8[10]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
+                continue;
+            }
+
+            ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
+            {
+                int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
+                AV_WN32A(&h->ref_cache[list][scan8[ 0]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
+                AV_WN32A(&h->ref_cache[list][scan8[ 2]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
+                ref += h->b8_stride;
+                AV_WN32A(&h->ref_cache[list][scan8[ 8]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
+                AV_WN32A(&h->ref_cache[list][scan8[10]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
+            }
+
+            b_stride = h->b_stride;
+            mv_dst   = &h->mv_cache[list][scan8[0]];
+            mv_src   = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
+            for(y=0; y<4; y++){
+                AV_COPY128(mv_dst + 8*y, mv_src + y*b_stride);
+            }
+
+        }
+    }
+
+
+/*
+0 . T T. T T T T
+1 L . .L . . . .
+2 L . .L . . . .
+3 . T TL . . . .
+4 L . .L . . . .
+5 L . .. . . . .
+*/
+//FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
+    if(top_type){
+        AV_COPY32(&h->non_zero_count_cache[4+8*0], &h->non_zero_count[top_xy][4+3*8]);
+    }
+
+    if(left_type[0]){
+        h->non_zero_count_cache[3+8*1]= h->non_zero_count[left_xy[0]][7+0*8];
+        h->non_zero_count_cache[3+8*2]= h->non_zero_count[left_xy[0]][7+1*8];
+        h->non_zero_count_cache[3+8*3]= h->non_zero_count[left_xy[0]][7+2*8];
+        h->non_zero_count_cache[3+8*4]= h->non_zero_count[left_xy[0]][7+3*8];
+    }
+
+    // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
+    if(!CABAC && h->pps.transform_8x8_mode){
+        if(IS_8x8DCT(top_type)){
+            h->non_zero_count_cache[4+8*0]=
+            h->non_zero_count_cache[5+8*0]= h->cbp_table[top_xy] & 4;
+            h->non_zero_count_cache[6+8*0]=
+            h->non_zero_count_cache[7+8*0]= h->cbp_table[top_xy] & 8;
+        }
+        if(IS_8x8DCT(left_type[0])){
+            h->non_zero_count_cache[3+8*1]=
+            h->non_zero_count_cache[3+8*2]= h->cbp_table[left_xy[0]]&2; //FIXME check MBAFF
+        }
+        if(IS_8x8DCT(left_type[1])){
+            h->non_zero_count_cache[3+8*3]=
+            h->non_zero_count_cache[3+8*4]= h->cbp_table[left_xy[1]]&8; //FIXME check MBAFF
+        }
+
+        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[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[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[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;
+        }
+    }
+
+    if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
+        int list;
+        for(list=0; list<h->list_count; list++){
+            if(USES_LIST(top_type, list)){
+                const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
+                const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
+                int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
+                AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
+                h->ref_cache[list][scan8[0] + 0 - 1*8]=
+                h->ref_cache[list][scan8[0] + 1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
+                h->ref_cache[list][scan8[0] + 2 - 1*8]=
+                h->ref_cache[list][scan8[0] + 3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
+            }else{
+                AV_ZERO128(h->mv_cache[list][scan8[0] + 0 - 1*8]);
+                AV_WN32A(&h->ref_cache[list][scan8[0] + 0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
+            }
+
+            if(!IS_INTERLACED(mb_type^left_type[0])){
+                if(USES_LIST(left_type[0], list)){
+                    const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
+                    const int b8_xy= h->mb2b8_xy[left_xy[0]] + 1;
+                    int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[0]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
+                    AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 0 ], s->current_picture.motion_val[list][b_xy + h->b_stride*0]);
+                    AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 8 ], s->current_picture.motion_val[list][b_xy + h->b_stride*1]);
+                    AV_COPY32(h->mv_cache[list][scan8[0] - 1 +16 ], s->current_picture.motion_val[list][b_xy + h->b_stride*2]);
+                    AV_COPY32(h->mv_cache[list][scan8[0] - 1 +24 ], s->current_picture.motion_val[list][b_xy + h->b_stride*3]);
+                    h->ref_cache[list][scan8[0] - 1 + 0 ]=
+                    h->ref_cache[list][scan8[0] - 1 + 8 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + h->b8_stride*0]];
+                    h->ref_cache[list][scan8[0] - 1 +16 ]=
+                    h->ref_cache[list][scan8[0] - 1 +24 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + h->b8_stride*1]];
+                }else{
+                    AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 0 ]);
+                    AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 8 ]);
+                    AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +16 ]);
+                    AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +24 ]);
+                    h->ref_cache[list][scan8[0] - 1 + 0  ]=
+                    h->ref_cache[list][scan8[0] - 1 + 8  ]=
+                    h->ref_cache[list][scan8[0] - 1 + 16 ]=
+                    h->ref_cache[list][scan8[0] - 1 + 24 ]= LIST_NOT_USED;
+                }
+            }
+        }
+    }
+
+    return 0;
 }
 
 /**
@@ -1273,8 +1387,8 @@ static inline void write_back_non_zero_count(H264Context *h){
 
     AV_COPY64(&h->non_zero_count[mb_xy][ 0], &h->non_zero_count_cache[0+8*1]);
     AV_COPY64(&h->non_zero_count[mb_xy][ 8], &h->non_zero_count_cache[0+8*2]);
-    *((uint32_t*)&h->non_zero_count[mb_xy][16]) = *((uint32_t*)&h->non_zero_count_cache[0+8*5]);
-    *((uint32_t*)&h->non_zero_count[mb_xy][20]) = *((uint32_t*)&h->non_zero_count_cache[4+8*3]);
+    AV_COPY32(&h->non_zero_count[mb_xy][16], &h->non_zero_count_cache[0+8*5]);
+    AV_COPY32(&h->non_zero_count[mb_xy][20], &h->non_zero_count_cache[4+8*3]);
     AV_COPY64(&h->non_zero_count[mb_xy][24], &h->non_zero_count_cache[0+8*4]);
 }
 
@@ -1324,29 +1438,18 @@ static inline void write_back_motion(H264Context *h, int mb_type){
     if(h->slice_type_nos == FF_B_TYPE && CABAC){
         if(IS_8X8(mb_type)){
             uint8_t *direct_table = &h->direct_table[b8_xy];
-            direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
-            direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
-            direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
+            direct_table[1+0*h->b8_stride] = h->sub_mb_type[1]>>1;
+            direct_table[0+1*h->b8_stride] = h->sub_mb_type[2]>>1;
+            direct_table[1+1*h->b8_stride] = h->sub_mb_type[3]>>1;
         }
     }
 }
 
 static inline int get_dct8x8_allowed(H264Context *h){
     if(h->sps.direct_8x8_inference_flag)
-        return !(*(uint64_t*)h->sub_mb_type & ((MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_8x8                )*0x0001000100010001ULL));
+        return !(AV_RN64A(h->sub_mb_type) & ((MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_8x8                )*0x0001000100010001ULL));
     else
-        return !(*(uint64_t*)h->sub_mb_type & ((MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_8x8|MB_TYPE_DIRECT2)*0x0001000100010001ULL));
-}
-
-static void predict_field_decoding_flag(H264Context *h){
-    MpegEncContext * const s = &h->s;
-    const int mb_xy= h->mb_xy;
-    int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
-                ? s->current_picture.mb_type[mb_xy-1]
-                : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
-                ? s->current_picture.mb_type[mb_xy-s->mb_stride]
-                : 0;
-    h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
+        return !(AV_RN64A(h->sub_mb_type) & ((MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_8x8|MB_TYPE_DIRECT2)*0x0001000100010001ULL));
 }
 
 /**
@@ -1366,9 +1469,11 @@ static void decode_mb_skip(H264Context *h){
     if( h->slice_type_nos == FF_B_TYPE )
     {
         // just for fill_caches. pred_direct_motion will set the real mb_type
-        mb_type|= MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2|MB_TYPE_SKIP;
-
+        mb_type|= MB_TYPE_L0L1|MB_TYPE_DIRECT2|MB_TYPE_SKIP;
+        if(h->direct_spatial_mv_pred){
+            fill_decode_neighbors(h, mb_type);
         fill_decode_caches(h, mb_type); //FIXME check what is needed and what not ...
+        }
         ff_h264_pred_direct_motion(h, &mb_type);
         mb_type|= MB_TYPE_SKIP;
     }
@@ -1377,6 +1482,7 @@ static void decode_mb_skip(H264Context *h){
         int mx, my;
         mb_type|= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P1L0|MB_TYPE_SKIP;
 
+        fill_decode_neighbors(h, mb_type);
         fill_decode_caches(h, mb_type); //FIXME check what is needed and what not ...
         pred_pskip_motion(h, &mx, &my);
         fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, 0, 1);