Mercurial > libavcodec.hg
comparison h264.h @ 11288:d5dd13f345fc libavcodec
Store data in direct_table interleaved.
seems 20cpu cycles faster
| author | michael |
|---|---|
| date | Thu, 25 Feb 2010 15:27:55 +0000 |
| parents | c8fea332d5d9 |
| children | 411ab09ada91 |
comparison
equal
deleted
inserted
replaced
| 11287:c8fea332d5d9 | 11288:d5dd13f345fc |
|---|---|
| 1086 fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, MB_TYPE_16x16>>1, 1); | 1086 fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, MB_TYPE_16x16>>1, 1); |
| 1087 | 1087 |
| 1088 if(IS_DIRECT(top_type)){ | 1088 if(IS_DIRECT(top_type)){ |
| 1089 AV_WN32A(&h->direct_cache[scan8[0] - 1*8], 0x01010101*(MB_TYPE_DIRECT2>>1)); | 1089 AV_WN32A(&h->direct_cache[scan8[0] - 1*8], 0x01010101*(MB_TYPE_DIRECT2>>1)); |
| 1090 }else if(IS_8X8(top_type)){ | 1090 }else if(IS_8X8(top_type)){ |
| 1091 int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride; | 1091 int b8_xy = 4*top_xy; |
| 1092 h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy]; | 1092 h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy + 2]; |
| 1093 h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1]; | 1093 h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 3]; |
| 1094 }else{ | 1094 }else{ |
| 1095 AV_WN32A(&h->direct_cache[scan8[0] - 1*8], 0x01010101*(MB_TYPE_16x16>>1)); | 1095 AV_WN32A(&h->direct_cache[scan8[0] - 1*8], 0x01010101*(MB_TYPE_16x16>>1)); |
| 1096 } | 1096 } |
| 1097 | 1097 |
| 1098 if(IS_DIRECT(left_type[0])) | 1098 if(IS_DIRECT(left_type[0])) |
| 1099 h->direct_cache[scan8[0] - 1 + 0*8]= MB_TYPE_DIRECT2>>1; | 1099 h->direct_cache[scan8[0] - 1 + 0*8]= MB_TYPE_DIRECT2>>1; |
| 1100 else if(IS_8X8(left_type[0])) | 1100 else if(IS_8X8(left_type[0])) |
| 1101 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)]; | 1101 h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[4*left_xy[0] + 1 + (left_block[0]&~1)]; |
| 1102 else | 1102 else |
| 1103 h->direct_cache[scan8[0] - 1 + 0*8]= MB_TYPE_16x16>>1; | 1103 h->direct_cache[scan8[0] - 1 + 0*8]= MB_TYPE_16x16>>1; |
| 1104 | 1104 |
| 1105 if(IS_DIRECT(left_type[1])) | 1105 if(IS_DIRECT(left_type[1])) |
| 1106 h->direct_cache[scan8[0] - 1 + 2*8]= MB_TYPE_DIRECT2>>1; | 1106 h->direct_cache[scan8[0] - 1 + 2*8]= MB_TYPE_DIRECT2>>1; |
| 1107 else if(IS_8X8(left_type[1])) | 1107 else if(IS_8X8(left_type[1])) |
| 1108 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)]; | 1108 h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[4*left_xy[1] + 1 + (left_block[2]&~1)]; |
| 1109 else | 1109 else |
| 1110 h->direct_cache[scan8[0] - 1 + 2*8]= MB_TYPE_16x16>>1; | 1110 h->direct_cache[scan8[0] - 1 + 2*8]= MB_TYPE_16x16>>1; |
| 1111 } | 1111 } |
| 1112 } | 1112 } |
| 1113 } | 1113 } |
| 1431 } | 1431 } |
| 1432 } | 1432 } |
| 1433 | 1433 |
| 1434 if(h->slice_type_nos == FF_B_TYPE && CABAC){ | 1434 if(h->slice_type_nos == FF_B_TYPE && CABAC){ |
| 1435 if(IS_8X8(mb_type)){ | 1435 if(IS_8X8(mb_type)){ |
| 1436 uint8_t *direct_table = &h->direct_table[b8_xy]; | 1436 uint8_t *direct_table = &h->direct_table[4*h->mb_xy]; |
| 1437 direct_table[1+0*h->b8_stride] = h->sub_mb_type[1]>>1; | 1437 direct_table[1] = h->sub_mb_type[1]>>1; |
| 1438 direct_table[0+1*h->b8_stride] = h->sub_mb_type[2]>>1; | 1438 direct_table[2] = h->sub_mb_type[2]>>1; |
| 1439 direct_table[1+1*h->b8_stride] = h->sub_mb_type[3]>>1; | 1439 direct_table[3] = h->sub_mb_type[3]>>1; |
| 1440 } | 1440 } |
| 1441 } | 1441 } |
| 1442 } | 1442 } |
| 1443 | 1443 |
| 1444 static inline int get_dct8x8_allowed(H264Context *h){ | 1444 static inline int get_dct8x8_allowed(H264Context *h){ |