comparison h264_mvpred.h @ 11292:411ab09ada91 libavcodec

Get rid of mb2b8_xy and b8_stride, change arrays organized based on b8_stride to ones based on mb_stride in h264. about 20 cpu cycles faster overall per MB
author michael
date Thu, 25 Feb 2010 23:44:42 +0000
parents 10c06a9bd3d9
children fd2f27fcb749
comparison
equal deleted inserted replaced
11291:1527e25ec9d4 11292:411ab09ada91
41 41
42 /* there is no consistent mapping of mvs to neighboring locations that will 42 /* there is no consistent mapping of mvs to neighboring locations that will
43 * make mbaff happy, so we can't move all this logic to fill_caches */ 43 * make mbaff happy, so we can't move all this logic to fill_caches */
44 if(FRAME_MBAFF){ 44 if(FRAME_MBAFF){
45 45
46 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\ 46 #define SET_DIAG_MV(MV_OP, REF_OP, XY, Y4)\
47 const int x4 = X4, y4 = Y4;\ 47 const int xy = XY, y4 = Y4;\
48 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\ 48 const int mb_type = mb_types[xy+(y4>>2)*s->mb_stride];\
49 if(!USES_LIST(mb_type,list))\ 49 if(!USES_LIST(mb_type,list))\
50 return LIST_NOT_USED;\ 50 return LIST_NOT_USED;\
51 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\ 51 mv = s->current_picture_ptr->motion_val[list][h->mb2b_xy[xy]+3 + y4*h->b_stride];\
52 h->mv_cache[list][scan8[0]-2][0] = mv[0];\ 52 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
53 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\ 53 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
54 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP; 54 return s->current_picture_ptr->ref_index[list][4*xy+1 + (y4&~1)] REF_OP;
55 55
56 if(topright_ref == PART_NOT_AVAILABLE 56 if(topright_ref == PART_NOT_AVAILABLE
57 && i >= scan8[0]+8 && (i&7)==4 57 && i >= scan8[0]+8 && (i&7)==4
58 && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){ 58 && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
59 const uint32_t *mb_types = s->current_picture_ptr->mb_type; 59 const uint32_t *mb_types = s->current_picture_ptr->mb_type;
61 AV_ZERO32(h->mv_cache[list][scan8[0]-2]); 61 AV_ZERO32(h->mv_cache[list][scan8[0]-2]);
62 *C = h->mv_cache[list][scan8[0]-2]; 62 *C = h->mv_cache[list][scan8[0]-2];
63 63
64 if(!MB_FIELD 64 if(!MB_FIELD
65 && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){ 65 && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
66 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1); 66 SET_DIAG_MV(*2, >>1, h->left_mb_xy[0]+s->mb_stride, (s->mb_y&1)*2+(i>>4)-1);
67 assert(h->left_mb_xy[0] == h->left_mb_xy[1]);
67 } 68 }
68 if(MB_FIELD 69 if(MB_FIELD
69 && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])){ 70 && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
70 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK. 71 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
71 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2); 72 SET_DIAG_MV(/2, <<1, h->left_mb_xy[i>=36], (- 1 + ((i-scan8[0])>>3)*2)&3);
72 } 73 }
73 } 74 }
74 #undef SET_DIAG_MV 75 #undef SET_DIAG_MV
75 } 76 }
76 77