- const int di = IS_8x8DCT(mb_type) ? 4 : 1;
- for(i=0; i<16; i+=di){
- if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
- uint8_t * const ptr= dest_y + block_offset[i];
- idct_add(ptr, h->mb + i*16, linesize);
+ if(IS_INTRA16x16(mb_type)){
+ for(i=0; i<16; i++){
+ if(h->non_zero_count_cache[ scan8[i] ])
+ idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
+ else if(h->mb[i*16])
+ idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
+ }
+ }else{
+ const int di = IS_8x8DCT(mb_type) ? 4 : 1;
+ for(i=0; i<16; i+=di){
+ int nnz = h->non_zero_count_cache[ scan8[i] ];
+ if(nnz){
+ if(nnz==1 && h->mb[i*16])
+ idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
+ else
+ idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
+ }