Libav 0.7.1
|
00001 /* 00002 * H.26L/H.264/AVC/JVT/14496-10/... direct mb/block decoding 00003 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at> 00004 * 00005 * This file is part of Libav. 00006 * 00007 * Libav is free software; you can redistribute it and/or 00008 * modify it under the terms of the GNU Lesser General Public 00009 * License as published by the Free Software Foundation; either 00010 * version 2.1 of the License, or (at your option) any later version. 00011 * 00012 * Libav is distributed in the hope that it will be useful, 00013 * but WITHOUT ANY WARRANTY; without even the implied warranty of 00014 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 00015 * Lesser General Public License for more details. 00016 * 00017 * You should have received a copy of the GNU Lesser General Public 00018 * License along with Libav; if not, write to the Free Software 00019 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 00020 */ 00021 00028 #include "internal.h" 00029 #include "dsputil.h" 00030 #include "avcodec.h" 00031 #include "mpegvideo.h" 00032 #include "h264.h" 00033 #include "rectangle.h" 00034 #include "thread.h" 00035 00036 //#undef NDEBUG 00037 #include <assert.h> 00038 00039 00040 static int get_scale_factor(H264Context * const h, int poc, int poc1, int i){ 00041 int poc0 = h->ref_list[0][i].poc; 00042 int td = av_clip(poc1 - poc0, -128, 127); 00043 if(td == 0 || h->ref_list[0][i].long_ref){ 00044 return 256; 00045 }else{ 00046 int tb = av_clip(poc - poc0, -128, 127); 00047 int tx = (16384 + (FFABS(td) >> 1)) / td; 00048 return av_clip((tb*tx + 32) >> 6, -1024, 1023); 00049 } 00050 } 00051 00052 void ff_h264_direct_dist_scale_factor(H264Context * const h){ 00053 MpegEncContext * const s = &h->s; 00054 const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ]; 00055 const int poc1 = h->ref_list[1][0].poc; 00056 int i, field; 00057 for(field=0; field<2; field++){ 00058 const int poc = h->s.current_picture_ptr->field_poc[field]; 00059 const int poc1 = h->ref_list[1][0].field_poc[field]; 00060 for(i=0; i < 2*h->ref_count[0]; i++) 00061 h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16); 00062 } 00063 00064 for(i=0; i<h->ref_count[0]; i++){ 00065 h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i); 00066 } 00067 } 00068 00069 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){ 00070 MpegEncContext * const s = &h->s; 00071 Picture * const ref1 = &h->ref_list[1][0]; 00072 int j, old_ref, rfield; 00073 int start= mbafi ? 16 : 0; 00074 int end = mbafi ? 16+2*h->ref_count[0] : h->ref_count[0]; 00075 int interl= mbafi || s->picture_structure != PICT_FRAME; 00076 00077 /* bogus; fills in for missing frames */ 00078 memset(map[list], 0, sizeof(map[list])); 00079 00080 for(rfield=0; rfield<2; rfield++){ 00081 for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){ 00082 int poc = ref1->ref_poc[colfield][list][old_ref]; 00083 00084 if (!interl) 00085 poc |= 3; 00086 else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed 00087 poc= (poc&~3) + rfield + 1; 00088 00089 for(j=start; j<end; j++){ 00090 if(4*h->ref_list[0][j].frame_num + (h->ref_list[0][j].reference&3) == poc){ 00091 int cur_ref= mbafi ? (j-16)^field : j; 00092 map[list][2*old_ref + (rfield^field) + 16] = cur_ref; 00093 if(rfield == field || !interl) 00094 map[list][old_ref] = cur_ref; 00095 break; 00096 } 00097 } 00098 } 00099 } 00100 } 00101 00102 void ff_h264_direct_ref_list_init(H264Context * const h){ 00103 MpegEncContext * const s = &h->s; 00104 Picture * const ref1 = &h->ref_list[1][0]; 00105 Picture * const cur = s->current_picture_ptr; 00106 int list, j, field; 00107 int sidx= (s->picture_structure&1)^1; 00108 int ref1sidx= (ref1->reference&1)^1; 00109 00110 for(list=0; list<2; list++){ 00111 cur->ref_count[sidx][list] = h->ref_count[list]; 00112 for(j=0; j<h->ref_count[list]; j++) 00113 cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3); 00114 } 00115 00116 if(s->picture_structure == PICT_FRAME){ 00117 memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0])); 00118 memcpy(cur->ref_poc [1], cur->ref_poc [0], sizeof(cur->ref_poc [0])); 00119 } 00120 00121 cur->mbaff= FRAME_MBAFF; 00122 00123 h->col_fieldoff= 0; 00124 if(s->picture_structure == PICT_FRAME){ 00125 int cur_poc = s->current_picture_ptr->poc; 00126 int *col_poc = h->ref_list[1]->field_poc; 00127 h->col_parity= (FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc)); 00128 ref1sidx=sidx= h->col_parity; 00129 }else if(!(s->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff){ // FL -> FL & differ parity 00130 h->col_fieldoff= 2*(h->ref_list[1][0].reference) - 3; 00131 } 00132 00133 if(cur->pict_type != AV_PICTURE_TYPE_B || h->direct_spatial_mv_pred) 00134 return; 00135 00136 for(list=0; list<2; list++){ 00137 fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0); 00138 if(FRAME_MBAFF) 00139 for(field=0; field<2; field++) 00140 fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1); 00141 } 00142 } 00143 00144 static void await_reference_mb_row(H264Context * const h, Picture *ref, int mb_y) 00145 { 00146 int ref_field = ref->reference - 1; 00147 int ref_field_picture = ref->field_picture; 00148 int ref_height = 16*h->s.mb_height >> ref_field_picture; 00149 00150 if(!HAVE_PTHREADS || !(h->s.avctx->active_thread_type&FF_THREAD_FRAME)) 00151 return; 00152 00153 //FIXME it can be safe to access mb stuff 00154 //even if pixels aren't deblocked yet 00155 00156 ff_thread_await_progress((AVFrame*)ref, FFMIN(16*mb_y >> ref_field_picture, ref_height-1), 00157 ref_field_picture && ref_field); 00158 } 00159 00160 static void pred_spatial_direct_motion(H264Context * const h, int *mb_type){ 00161 MpegEncContext * const s = &h->s; 00162 int b8_stride = 2; 00163 int b4_stride = h->b_stride; 00164 int mb_xy = h->mb_xy, mb_y = s->mb_y; 00165 int mb_type_col[2]; 00166 const int16_t (*l1mv0)[2], (*l1mv1)[2]; 00167 const int8_t *l1ref0, *l1ref1; 00168 const int is_b8x8 = IS_8X8(*mb_type); 00169 unsigned int sub_mb_type= MB_TYPE_L0L1; 00170 int i8, i4; 00171 int ref[2]; 00172 int mv[2]; 00173 int list; 00174 00175 assert(h->ref_list[1][0].reference&3); 00176 00177 await_reference_mb_row(h, &h->ref_list[1][0], s->mb_y + !!IS_INTERLACED(*mb_type)); 00178 00179 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM) 00180 00181 00182 /* ref = min(neighbors) */ 00183 for(list=0; list<2; list++){ 00184 int left_ref = h->ref_cache[list][scan8[0] - 1]; 00185 int top_ref = h->ref_cache[list][scan8[0] - 8]; 00186 int refc = h->ref_cache[list][scan8[0] - 8 + 4]; 00187 const int16_t *C= h->mv_cache[list][ scan8[0] - 8 + 4]; 00188 if(refc == PART_NOT_AVAILABLE){ 00189 refc = h->ref_cache[list][scan8[0] - 8 - 1]; 00190 C = h-> mv_cache[list][scan8[0] - 8 - 1]; 00191 } 00192 ref[list] = FFMIN3((unsigned)left_ref, (unsigned)top_ref, (unsigned)refc); 00193 if(ref[list] >= 0){ 00194 //this is just pred_motion() but with the cases removed that cannot happen for direct blocks 00195 const int16_t * const A= h->mv_cache[list][ scan8[0] - 1 ]; 00196 const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ]; 00197 00198 int match_count= (left_ref==ref[list]) + (top_ref==ref[list]) + (refc==ref[list]); 00199 if(match_count > 1){ //most common 00200 mv[list]= pack16to32(mid_pred(A[0], B[0], C[0]), 00201 mid_pred(A[1], B[1], C[1]) ); 00202 }else { 00203 assert(match_count==1); 00204 if(left_ref==ref[list]){ 00205 mv[list]= AV_RN32A(A); 00206 }else if(top_ref==ref[list]){ 00207 mv[list]= AV_RN32A(B); 00208 }else{ 00209 mv[list]= AV_RN32A(C); 00210 } 00211 } 00212 }else{ 00213 int mask= ~(MB_TYPE_L0 << (2*list)); 00214 mv[list] = 0; 00215 ref[list] = -1; 00216 if(!is_b8x8) 00217 *mb_type &= mask; 00218 sub_mb_type &= mask; 00219 } 00220 } 00221 if(ref[0] < 0 && ref[1] < 0){ 00222 ref[0] = ref[1] = 0; 00223 if(!is_b8x8) 00224 *mb_type |= MB_TYPE_L0L1; 00225 sub_mb_type |= MB_TYPE_L0L1; 00226 } 00227 00228 if(!(is_b8x8|mv[0]|mv[1])){ 00229 fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1); 00230 fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1); 00231 fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4); 00232 fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4); 00233 *mb_type= (*mb_type & ~(MB_TYPE_8x8|MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_P1L0|MB_TYPE_P1L1))|MB_TYPE_16x16|MB_TYPE_DIRECT2; 00234 return; 00235 } 00236 00237 if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL 00238 if(!IS_INTERLACED(*mb_type)){ // AFR/FR -> AFL/FL 00239 mb_y = (s->mb_y&~1) + h->col_parity; 00240 mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride; 00241 b8_stride = 0; 00242 }else{ 00243 mb_y += h->col_fieldoff; 00244 mb_xy += s->mb_stride*h->col_fieldoff; // non zero for FL -> FL & differ parity 00245 } 00246 goto single_col; 00247 }else{ // AFL/AFR/FR/FL -> AFR/FR 00248 if(IS_INTERLACED(*mb_type)){ // AFL /FL -> AFR/FR 00249 mb_y = s->mb_y&~1; 00250 mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride; 00251 mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy]; 00252 mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride]; 00253 b8_stride = 2+4*s->mb_stride; 00254 b4_stride *= 6; 00255 00256 sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */ 00257 if( (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) 00258 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) 00259 && !is_b8x8){ 00260 *mb_type |= MB_TYPE_16x8 |MB_TYPE_DIRECT2; /* B_16x8 */ 00261 }else{ 00262 *mb_type |= MB_TYPE_8x8; 00263 } 00264 }else{ // AFR/FR -> AFR/FR 00265 single_col: 00266 mb_type_col[0] = 00267 mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy]; 00268 00269 sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */ 00270 if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){ 00271 *mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_16x16 */ 00272 }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){ 00273 *mb_type |= MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16)); 00274 }else{ 00275 if(!h->sps.direct_8x8_inference_flag){ 00276 /* FIXME save sub mb types from previous frames (or derive from MVs) 00277 * so we know exactly what block size to use */ 00278 sub_mb_type += (MB_TYPE_8x8-MB_TYPE_16x16); /* B_SUB_4x4 */ 00279 } 00280 *mb_type |= MB_TYPE_8x8; 00281 } 00282 } 00283 } 00284 00285 await_reference_mb_row(h, &h->ref_list[1][0], mb_y); 00286 00287 l1mv0 = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]]; 00288 l1mv1 = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]]; 00289 l1ref0 = &h->ref_list[1][0].ref_index [0][4*mb_xy]; 00290 l1ref1 = &h->ref_list[1][0].ref_index [1][4*mb_xy]; 00291 if(!b8_stride){ 00292 if(s->mb_y&1){ 00293 l1ref0 += 2; 00294 l1ref1 += 2; 00295 l1mv0 += 2*b4_stride; 00296 l1mv1 += 2*b4_stride; 00297 } 00298 } 00299 00300 00301 if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){ 00302 int n=0; 00303 for(i8=0; i8<4; i8++){ 00304 int x8 = i8&1; 00305 int y8 = i8>>1; 00306 int xy8 = x8+y8*b8_stride; 00307 int xy4 = 3*x8+y8*b4_stride; 00308 int a,b; 00309 00310 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8])) 00311 continue; 00312 h->sub_mb_type[i8] = sub_mb_type; 00313 00314 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1); 00315 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1); 00316 if(!IS_INTRA(mb_type_col[y8]) && !h->ref_list[1][0].long_ref 00317 && ( (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1) 00318 || (l1ref0[xy8] < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){ 00319 a=b=0; 00320 if(ref[0] > 0) 00321 a= mv[0]; 00322 if(ref[1] > 0) 00323 b= mv[1]; 00324 n++; 00325 }else{ 00326 a= mv[0]; 00327 b= mv[1]; 00328 } 00329 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4); 00330 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4); 00331 } 00332 if(!is_b8x8 && !(n&3)) 00333 *mb_type= (*mb_type & ~(MB_TYPE_8x8|MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_P1L0|MB_TYPE_P1L1))|MB_TYPE_16x16|MB_TYPE_DIRECT2; 00334 }else if(IS_16X16(*mb_type)){ 00335 int a,b; 00336 00337 fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1); 00338 fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1); 00339 if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref 00340 && ( (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1) 00341 || (l1ref0[0] < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1 00342 && h->x264_build>33U))){ 00343 a=b=0; 00344 if(ref[0] > 0) 00345 a= mv[0]; 00346 if(ref[1] > 0) 00347 b= mv[1]; 00348 }else{ 00349 a= mv[0]; 00350 b= mv[1]; 00351 } 00352 fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4); 00353 fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4); 00354 }else{ 00355 int n=0; 00356 for(i8=0; i8<4; i8++){ 00357 const int x8 = i8&1; 00358 const int y8 = i8>>1; 00359 00360 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8])) 00361 continue; 00362 h->sub_mb_type[i8] = sub_mb_type; 00363 00364 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, mv[0], 4); 00365 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, mv[1], 4); 00366 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1); 00367 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1); 00368 00369 assert(b8_stride==2); 00370 /* col_zero_flag */ 00371 if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref && ( l1ref0[i8] == 0 00372 || (l1ref0[i8] < 0 && l1ref1[i8] == 0 00373 && h->x264_build>33U))){ 00374 const int16_t (*l1mv)[2]= l1ref0[i8] == 0 ? l1mv0 : l1mv1; 00375 if(IS_SUB_8X8(sub_mb_type)){ 00376 const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride]; 00377 if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){ 00378 if(ref[0] == 0) 00379 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4); 00380 if(ref[1] == 0) 00381 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4); 00382 n+=4; 00383 } 00384 }else{ 00385 int m=0; 00386 for(i4=0; i4<4; i4++){ 00387 const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride]; 00388 if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){ 00389 if(ref[0] == 0) 00390 AV_ZERO32(h->mv_cache[0][scan8[i8*4+i4]]); 00391 if(ref[1] == 0) 00392 AV_ZERO32(h->mv_cache[1][scan8[i8*4+i4]]); 00393 m++; 00394 } 00395 } 00396 if(!(m&3)) 00397 h->sub_mb_type[i8]+= MB_TYPE_16x16 - MB_TYPE_8x8; 00398 n+=m; 00399 } 00400 } 00401 } 00402 if(!is_b8x8 && !(n&15)) 00403 *mb_type= (*mb_type & ~(MB_TYPE_8x8|MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_P1L0|MB_TYPE_P1L1))|MB_TYPE_16x16|MB_TYPE_DIRECT2; 00404 } 00405 } 00406 00407 static void pred_temp_direct_motion(H264Context * const h, int *mb_type){ 00408 MpegEncContext * const s = &h->s; 00409 int b8_stride = 2; 00410 int b4_stride = h->b_stride; 00411 int mb_xy = h->mb_xy, mb_y = s->mb_y; 00412 int mb_type_col[2]; 00413 const int16_t (*l1mv0)[2], (*l1mv1)[2]; 00414 const int8_t *l1ref0, *l1ref1; 00415 const int is_b8x8 = IS_8X8(*mb_type); 00416 unsigned int sub_mb_type; 00417 int i8, i4; 00418 00419 assert(h->ref_list[1][0].reference&3); 00420 00421 await_reference_mb_row(h, &h->ref_list[1][0], s->mb_y + !!IS_INTERLACED(*mb_type)); 00422 00423 if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL 00424 if(!IS_INTERLACED(*mb_type)){ // AFR/FR -> AFL/FL 00425 mb_y = (s->mb_y&~1) + h->col_parity; 00426 mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride; 00427 b8_stride = 0; 00428 }else{ 00429 mb_y += h->col_fieldoff; 00430 mb_xy += s->mb_stride*h->col_fieldoff; // non zero for FL -> FL & differ parity 00431 } 00432 goto single_col; 00433 }else{ // AFL/AFR/FR/FL -> AFR/FR 00434 if(IS_INTERLACED(*mb_type)){ // AFL /FL -> AFR/FR 00435 mb_y = s->mb_y&~1; 00436 mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride; 00437 mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy]; 00438 mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride]; 00439 b8_stride = 2+4*s->mb_stride; 00440 b4_stride *= 6; 00441 00442 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */ 00443 00444 if( (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) 00445 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) 00446 && !is_b8x8){ 00447 *mb_type |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */ 00448 }else{ 00449 *mb_type |= MB_TYPE_8x8|MB_TYPE_L0L1; 00450 } 00451 }else{ // AFR/FR -> AFR/FR 00452 single_col: 00453 mb_type_col[0] = 00454 mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy]; 00455 00456 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */ 00457 if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){ 00458 *mb_type |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */ 00459 }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){ 00460 *mb_type |= MB_TYPE_L0L1|MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16)); 00461 }else{ 00462 if(!h->sps.direct_8x8_inference_flag){ 00463 /* FIXME save sub mb types from previous frames (or derive from MVs) 00464 * so we know exactly what block size to use */ 00465 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */ 00466 } 00467 *mb_type |= MB_TYPE_8x8|MB_TYPE_L0L1; 00468 } 00469 } 00470 } 00471 00472 await_reference_mb_row(h, &h->ref_list[1][0], mb_y); 00473 00474 l1mv0 = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]]; 00475 l1mv1 = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]]; 00476 l1ref0 = &h->ref_list[1][0].ref_index [0][4*mb_xy]; 00477 l1ref1 = &h->ref_list[1][0].ref_index [1][4*mb_xy]; 00478 if(!b8_stride){ 00479 if(s->mb_y&1){ 00480 l1ref0 += 2; 00481 l1ref1 += 2; 00482 l1mv0 += 2*b4_stride; 00483 l1mv1 += 2*b4_stride; 00484 } 00485 } 00486 00487 { 00488 const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]}; 00489 const int *dist_scale_factor = h->dist_scale_factor; 00490 int ref_offset; 00491 00492 if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){ 00493 map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0]; 00494 map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1]; 00495 dist_scale_factor =h->dist_scale_factor_field[s->mb_y&1]; 00496 } 00497 ref_offset = (h->ref_list[1][0].mbaff<<4) & (mb_type_col[0]>>3); //if(h->ref_list[1][0].mbaff && IS_INTERLACED(mb_type_col[0])) ref_offset=16 else 0 00498 00499 if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){ 00500 int y_shift = 2*!IS_INTERLACED(*mb_type); 00501 assert(h->sps.direct_8x8_inference_flag); 00502 00503 for(i8=0; i8<4; i8++){ 00504 const int x8 = i8&1; 00505 const int y8 = i8>>1; 00506 int ref0, scale; 00507 const int16_t (*l1mv)[2]= l1mv0; 00508 00509 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8])) 00510 continue; 00511 h->sub_mb_type[i8] = sub_mb_type; 00512 00513 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1); 00514 if(IS_INTRA(mb_type_col[y8])){ 00515 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1); 00516 fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4); 00517 fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4); 00518 continue; 00519 } 00520 00521 ref0 = l1ref0[x8 + y8*b8_stride]; 00522 if(ref0 >= 0) 00523 ref0 = map_col_to_list0[0][ref0 + ref_offset]; 00524 else{ 00525 ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset]; 00526 l1mv= l1mv1; 00527 } 00528 scale = dist_scale_factor[ref0]; 00529 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1); 00530 00531 { 00532 const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride]; 00533 int my_col = (mv_col[1]<<y_shift)/2; 00534 int mx = (scale * mv_col[0] + 128) >> 8; 00535 int my = (scale * my_col + 128) >> 8; 00536 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4); 00537 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4); 00538 } 00539 } 00540 return; 00541 } 00542 00543 /* one-to-one mv scaling */ 00544 00545 if(IS_16X16(*mb_type)){ 00546 int ref, mv0, mv1; 00547 00548 fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1); 00549 if(IS_INTRA(mb_type_col[0])){ 00550 ref=mv0=mv1=0; 00551 }else{ 00552 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset] 00553 : map_col_to_list0[1][l1ref1[0] + ref_offset]; 00554 const int scale = dist_scale_factor[ref0]; 00555 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0]; 00556 int mv_l0[2]; 00557 mv_l0[0] = (scale * mv_col[0] + 128) >> 8; 00558 mv_l0[1] = (scale * mv_col[1] + 128) >> 8; 00559 ref= ref0; 00560 mv0= pack16to32(mv_l0[0],mv_l0[1]); 00561 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]); 00562 } 00563 fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1); 00564 fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4); 00565 fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4); 00566 }else{ 00567 for(i8=0; i8<4; i8++){ 00568 const int x8 = i8&1; 00569 const int y8 = i8>>1; 00570 int ref0, scale; 00571 const int16_t (*l1mv)[2]= l1mv0; 00572 00573 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8])) 00574 continue; 00575 h->sub_mb_type[i8] = sub_mb_type; 00576 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1); 00577 if(IS_INTRA(mb_type_col[0])){ 00578 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1); 00579 fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4); 00580 fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4); 00581 continue; 00582 } 00583 00584 assert(b8_stride == 2); 00585 ref0 = l1ref0[i8]; 00586 if(ref0 >= 0) 00587 ref0 = map_col_to_list0[0][ref0 + ref_offset]; 00588 else{ 00589 ref0 = map_col_to_list0[1][l1ref1[i8] + ref_offset]; 00590 l1mv= l1mv1; 00591 } 00592 scale = dist_scale_factor[ref0]; 00593 00594 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1); 00595 if(IS_SUB_8X8(sub_mb_type)){ 00596 const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride]; 00597 int mx = (scale * mv_col[0] + 128) >> 8; 00598 int my = (scale * mv_col[1] + 128) >> 8; 00599 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4); 00600 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4); 00601 }else 00602 for(i4=0; i4<4; i4++){ 00603 const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride]; 00604 int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]]; 00605 mv_l0[0] = (scale * mv_col[0] + 128) >> 8; 00606 mv_l0[1] = (scale * mv_col[1] + 128) >> 8; 00607 AV_WN32A(h->mv_cache[1][scan8[i8*4+i4]], 00608 pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1])); 00609 } 00610 } 00611 } 00612 } 00613 } 00614 00615 void ff_h264_pred_direct_motion(H264Context * const h, int *mb_type){ 00616 if(h->direct_spatial_mv_pred){ 00617 pred_spatial_direct_motion(h, mb_type); 00618 }else{ 00619 pred_temp_direct_motion(h, mb_type); 00620 } 00621 }