| 1 | /* |
| 2 | * Motion estimation |
| 3 | * Copyright (c) 2000,2001 Fabrice Bellard |
| 4 | * Copyright (c) 2002-2004 Michael Niedermayer |
| 5 | * |
| 6 | * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at> |
| 7 | * |
| 8 | * This file is part of FFmpeg. |
| 9 | * |
| 10 | * FFmpeg is free software; you can redistribute it and/or |
| 11 | * modify it under the terms of the GNU Lesser General Public |
| 12 | * License as published by the Free Software Foundation; either |
| 13 | * version 2.1 of the License, or (at your option) any later version. |
| 14 | * |
| 15 | * FFmpeg is distributed in the hope that it will be useful, |
| 16 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 17 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 18 | * Lesser General Public License for more details. |
| 19 | * |
| 20 | * You should have received a copy of the GNU Lesser General Public |
| 21 | * License along with FFmpeg; if not, write to the Free Software |
| 22 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 23 | */ |
| 24 | |
| 25 | /** |
| 26 | * @file |
| 27 | * Motion estimation. |
| 28 | */ |
| 29 | |
| 30 | #include <stdlib.h> |
| 31 | #include <stdio.h> |
| 32 | #include <limits.h> |
| 33 | |
| 34 | #include "avcodec.h" |
| 35 | #include "mathops.h" |
| 36 | #include "mpegutils.h" |
| 37 | #include "mpegvideo.h" |
| 38 | |
| 39 | #undef NDEBUG |
| 40 | #include <assert.h> |
| 41 | |
| 42 | #define P_LEFT P[1] |
| 43 | #define P_TOP P[2] |
| 44 | #define P_TOPRIGHT P[3] |
| 45 | #define P_MEDIAN P[4] |
| 46 | #define P_MV1 P[9] |
| 47 | |
| 48 | #define ME_MAP_SHIFT 3 |
| 49 | #define ME_MAP_MV_BITS 11 |
| 50 | |
| 51 | static int sad_hpel_motion_search(MpegEncContext * s, |
| 52 | int *mx_ptr, int *my_ptr, int dmin, |
| 53 | int src_index, int ref_index, |
| 54 | int size, int h); |
| 55 | |
| 56 | static inline unsigned update_map_generation(MotionEstContext *c) |
| 57 | { |
| 58 | c->map_generation+= 1<<(ME_MAP_MV_BITS*2); |
| 59 | if(c->map_generation==0){ |
| 60 | c->map_generation= 1<<(ME_MAP_MV_BITS*2); |
| 61 | memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE); |
| 62 | } |
| 63 | return c->map_generation; |
| 64 | } |
| 65 | |
| 66 | /* shape adaptive search stuff */ |
| 67 | typedef struct Minima{ |
| 68 | int height; |
| 69 | int x, y; |
| 70 | int checked; |
| 71 | }Minima; |
| 72 | |
| 73 | static int minima_cmp(const void *a, const void *b){ |
| 74 | const Minima *da = (const Minima *) a; |
| 75 | const Minima *db = (const Minima *) b; |
| 76 | |
| 77 | return da->height - db->height; |
| 78 | } |
| 79 | |
| 80 | #define FLAG_QPEL 1 //must be 1 |
| 81 | #define FLAG_CHROMA 2 |
| 82 | #define FLAG_DIRECT 4 |
| 83 | |
| 84 | static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){ |
| 85 | const int offset[3]= { |
| 86 | y*c-> stride + x, |
| 87 | ((y*c->uvstride + x)>>1), |
| 88 | ((y*c->uvstride + x)>>1), |
| 89 | }; |
| 90 | int i; |
| 91 | for(i=0; i<3; i++){ |
| 92 | c->src[0][i]= src [i] + offset[i]; |
| 93 | c->ref[0][i]= ref [i] + offset[i]; |
| 94 | } |
| 95 | if(ref_index){ |
| 96 | for(i=0; i<3; i++){ |
| 97 | c->ref[ref_index][i]= ref2[i] + offset[i]; |
| 98 | } |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | static int get_flags(MotionEstContext *c, int direct, int chroma){ |
| 103 | return ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0) |
| 104 | + (direct ? FLAG_DIRECT : 0) |
| 105 | + (chroma ? FLAG_CHROMA : 0); |
| 106 | } |
| 107 | |
| 108 | static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby, |
| 109 | const int size, const int h, int ref_index, int src_index, |
| 110 | me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){ |
| 111 | MotionEstContext * const c= &s->me; |
| 112 | const int stride= c->stride; |
| 113 | const int hx= subx + (x<<(1+qpel)); |
| 114 | const int hy= suby + (y<<(1+qpel)); |
| 115 | uint8_t * const * const ref= c->ref[ref_index]; |
| 116 | uint8_t * const * const src= c->src[src_index]; |
| 117 | int d; |
| 118 | //FIXME check chroma 4mv, (no crashes ...) |
| 119 | av_assert2(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)); |
| 120 | if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){ |
| 121 | const int time_pp= s->pp_time; |
| 122 | const int time_pb= s->pb_time; |
| 123 | const int mask= 2*qpel+1; |
| 124 | if(s->mv_type==MV_TYPE_8X8){ |
| 125 | int i; |
| 126 | for(i=0; i<4; i++){ |
| 127 | int fx = c->direct_basis_mv[i][0] + hx; |
| 128 | int fy = c->direct_basis_mv[i][1] + hy; |
| 129 | int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4)); |
| 130 | int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4)); |
| 131 | int fxy= (fx&mask) + ((fy&mask)<<(qpel+1)); |
| 132 | int bxy= (bx&mask) + ((by&mask)<<(qpel+1)); |
| 133 | |
| 134 | uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1); |
| 135 | if(qpel){ |
| 136 | c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride); |
| 137 | c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride); |
| 138 | }else{ |
| 139 | c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8); |
| 140 | c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8); |
| 141 | } |
| 142 | } |
| 143 | }else{ |
| 144 | int fx = c->direct_basis_mv[0][0] + hx; |
| 145 | int fy = c->direct_basis_mv[0][1] + hy; |
| 146 | int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp); |
| 147 | int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp); |
| 148 | int fxy= (fx&mask) + ((fy&mask)<<(qpel+1)); |
| 149 | int bxy= (bx&mask) + ((by&mask)<<(qpel+1)); |
| 150 | |
| 151 | if(qpel){ |
| 152 | c->qpel_put[1][fxy](c->temp , ref[0] + (fx>>2) + (fy>>2)*stride , stride); |
| 153 | c->qpel_put[1][fxy](c->temp + 8 , ref[0] + (fx>>2) + (fy>>2)*stride + 8 , stride); |
| 154 | c->qpel_put[1][fxy](c->temp + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8*stride, stride); |
| 155 | c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride); |
| 156 | c->qpel_avg[1][bxy](c->temp , ref[8] + (bx>>2) + (by>>2)*stride , stride); |
| 157 | c->qpel_avg[1][bxy](c->temp + 8 , ref[8] + (bx>>2) + (by>>2)*stride + 8 , stride); |
| 158 | c->qpel_avg[1][bxy](c->temp + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8*stride, stride); |
| 159 | c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride); |
| 160 | }else{ |
| 161 | av_assert2((fx>>1) + 16*s->mb_x >= -16); |
| 162 | av_assert2((fy>>1) + 16*s->mb_y >= -16); |
| 163 | av_assert2((fx>>1) + 16*s->mb_x <= s->width); |
| 164 | av_assert2((fy>>1) + 16*s->mb_y <= s->height); |
| 165 | av_assert2((bx>>1) + 16*s->mb_x >= -16); |
| 166 | av_assert2((by>>1) + 16*s->mb_y >= -16); |
| 167 | av_assert2((bx>>1) + 16*s->mb_x <= s->width); |
| 168 | av_assert2((by>>1) + 16*s->mb_y <= s->height); |
| 169 | |
| 170 | c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16); |
| 171 | c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16); |
| 172 | } |
| 173 | } |
| 174 | d = cmp_func(s, c->temp, src[0], stride, 16); |
| 175 | }else |
| 176 | d= 256*256*256*32; |
| 177 | return d; |
| 178 | } |
| 179 | |
| 180 | static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby, |
| 181 | const int size, const int h, int ref_index, int src_index, |
| 182 | me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){ |
| 183 | MotionEstContext * const c= &s->me; |
| 184 | const int stride= c->stride; |
| 185 | const int uvstride= c->uvstride; |
| 186 | const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel? |
| 187 | const int hx= subx + (x<<(1+qpel)); |
| 188 | const int hy= suby + (y<<(1+qpel)); |
| 189 | uint8_t * const * const ref= c->ref[ref_index]; |
| 190 | uint8_t * const * const src= c->src[src_index]; |
| 191 | int d; |
| 192 | //FIXME check chroma 4mv, (no crashes ...) |
| 193 | int uvdxy; /* no, it might not be used uninitialized */ |
| 194 | if(dxy){ |
| 195 | if(qpel){ |
| 196 | c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h) |
| 197 | if(chroma){ |
| 198 | int cx= hx/2; |
| 199 | int cy= hy/2; |
| 200 | cx= (cx>>1)|(cx&1); |
| 201 | cy= (cy>>1)|(cy&1); |
| 202 | uvdxy= (cx&1) + 2*(cy&1); |
| 203 | //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264 |
| 204 | } |
| 205 | }else{ |
| 206 | c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h); |
| 207 | if(chroma) |
| 208 | uvdxy= dxy | (x&1) | (2*(y&1)); |
| 209 | } |
| 210 | d = cmp_func(s, c->temp, src[0], stride, h); |
| 211 | }else{ |
| 212 | d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h); |
| 213 | if(chroma) |
| 214 | uvdxy= (x&1) + 2*(y&1); |
| 215 | } |
| 216 | if(chroma){ |
| 217 | uint8_t * const uvtemp= c->temp + 16*stride; |
| 218 | c->hpel_put[size+1][uvdxy](uvtemp , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1); |
| 219 | c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1); |
| 220 | d += chroma_cmp_func(s, uvtemp , src[1], uvstride, h>>1); |
| 221 | d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1); |
| 222 | } |
| 223 | return d; |
| 224 | } |
| 225 | |
| 226 | static int cmp_simple(MpegEncContext *s, const int x, const int y, |
| 227 | int ref_index, int src_index, |
| 228 | me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){ |
| 229 | return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0); |
| 230 | } |
| 231 | |
| 232 | static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y, |
| 233 | const int size, const int h, int ref_index, int src_index, |
| 234 | me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){ |
| 235 | if(flags&FLAG_DIRECT){ |
| 236 | return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL); |
| 237 | }else{ |
| 238 | return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA); |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby, |
| 243 | const int size, const int h, int ref_index, int src_index, |
| 244 | me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){ |
| 245 | if(flags&FLAG_DIRECT){ |
| 246 | return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL); |
| 247 | }else{ |
| 248 | return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL, flags&FLAG_CHROMA); |
| 249 | } |
| 250 | } |
| 251 | |
| 252 | /** @brief compares a block (either a full macroblock or a partition thereof) |
| 253 | against a proposed motion-compensated prediction of that block |
| 254 | */ |
| 255 | static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby, |
| 256 | const int size, const int h, int ref_index, int src_index, |
| 257 | me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){ |
| 258 | if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size) |
| 259 | && av_builtin_constant_p(subx) && av_builtin_constant_p(suby) |
| 260 | && flags==0 && h==16 && size==0 && subx==0 && suby==0){ |
| 261 | return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func); |
| 262 | }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby) |
| 263 | && subx==0 && suby==0){ |
| 264 | return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags); |
| 265 | }else{ |
| 266 | return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags); |
| 267 | } |
| 268 | } |
| 269 | |
| 270 | static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby, |
| 271 | const int size, const int h, int ref_index, int src_index, |
| 272 | me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){ |
| 273 | if(flags&FLAG_DIRECT){ |
| 274 | return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0); |
| 275 | }else{ |
| 276 | return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA); |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby, |
| 281 | const int size, const int h, int ref_index, int src_index, |
| 282 | me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){ |
| 283 | if(flags&FLAG_DIRECT){ |
| 284 | return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1); |
| 285 | }else{ |
| 286 | return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA); |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | #include "motion_est_template.c" |
| 291 | |
| 292 | static int zero_cmp(MpegEncContext *s, uint8_t *a, uint8_t *b, |
| 293 | int stride, int h) |
| 294 | { |
| 295 | return 0; |
| 296 | } |
| 297 | |
| 298 | static void zero_hpel(uint8_t *a, const uint8_t *b, ptrdiff_t stride, int h){ |
| 299 | } |
| 300 | |
| 301 | int ff_init_me(MpegEncContext *s){ |
| 302 | MotionEstContext * const c= &s->me; |
| 303 | int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT); |
| 304 | int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255); |
| 305 | |
| 306 | if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -FFMIN(ME_MAP_SIZE, MAX_SAB_SIZE)){ |
| 307 | av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n"); |
| 308 | return -1; |
| 309 | } |
| 310 | //special case of snow is needed because snow uses its own iterative ME code |
| 311 | if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != AV_CODEC_ID_SNOW){ |
| 312 | av_log(s->avctx, AV_LOG_ERROR, "me_method is only allowed to be set to zero and epzs; for hex,umh,full and others see dia_size\n"); |
| 313 | return -1; |
| 314 | } |
| 315 | |
| 316 | c->avctx= s->avctx; |
| 317 | |
| 318 | if(cache_size < 2*dia_size && !c->stride){ |
| 319 | av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n"); |
| 320 | } |
| 321 | |
| 322 | ff_set_cmp(&s->mecc, s->mecc.me_pre_cmp, c->avctx->me_pre_cmp); |
| 323 | ff_set_cmp(&s->mecc, s->mecc.me_cmp, c->avctx->me_cmp); |
| 324 | ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, c->avctx->me_sub_cmp); |
| 325 | ff_set_cmp(&s->mecc, s->mecc.mb_cmp, c->avctx->mb_cmp); |
| 326 | |
| 327 | c->flags = get_flags(c, 0, c->avctx->me_cmp &FF_CMP_CHROMA); |
| 328 | c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA); |
| 329 | c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp &FF_CMP_CHROMA); |
| 330 | |
| 331 | /*FIXME s->no_rounding b_type*/ |
| 332 | if(s->flags&CODEC_FLAG_QPEL){ |
| 333 | c->sub_motion_search= qpel_motion_search; |
| 334 | c->qpel_avg = s->qdsp.avg_qpel_pixels_tab; |
| 335 | if (s->no_rounding) |
| 336 | c->qpel_put = s->qdsp.put_no_rnd_qpel_pixels_tab; |
| 337 | else |
| 338 | c->qpel_put = s->qdsp.put_qpel_pixels_tab; |
| 339 | }else{ |
| 340 | if(c->avctx->me_sub_cmp&FF_CMP_CHROMA) |
| 341 | c->sub_motion_search= hpel_motion_search; |
| 342 | else if( c->avctx->me_sub_cmp == FF_CMP_SAD |
| 343 | && c->avctx-> me_cmp == FF_CMP_SAD |
| 344 | && c->avctx-> mb_cmp == FF_CMP_SAD) |
| 345 | c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles |
| 346 | else |
| 347 | c->sub_motion_search= hpel_motion_search; |
| 348 | } |
| 349 | c->hpel_avg = s->hdsp.avg_pixels_tab; |
| 350 | if (s->no_rounding) |
| 351 | c->hpel_put = s->hdsp.put_no_rnd_pixels_tab; |
| 352 | else |
| 353 | c->hpel_put = s->hdsp.put_pixels_tab; |
| 354 | |
| 355 | if(s->linesize){ |
| 356 | c->stride = s->linesize; |
| 357 | c->uvstride= s->uvlinesize; |
| 358 | }else{ |
| 359 | c->stride = 16*s->mb_width + 32; |
| 360 | c->uvstride= 8*s->mb_width + 16; |
| 361 | } |
| 362 | |
| 363 | /* 8x8 fullpel search would need a 4x4 chroma compare, which we do |
| 364 | * not have yet, and even if we had, the motion estimation code |
| 365 | * does not expect it. */ |
| 366 | if (s->codec_id != AV_CODEC_ID_SNOW) { |
| 367 | if ((c->avctx->me_cmp & FF_CMP_CHROMA) /* && !s->mecc.me_cmp[2] */) |
| 368 | s->mecc.me_cmp[2] = zero_cmp; |
| 369 | if ((c->avctx->me_sub_cmp & FF_CMP_CHROMA) && !s->mecc.me_sub_cmp[2]) |
| 370 | s->mecc.me_sub_cmp[2] = zero_cmp; |
| 371 | c->hpel_put[2][0]= c->hpel_put[2][1]= |
| 372 | c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel; |
| 373 | } |
| 374 | |
| 375 | if(s->codec_id == AV_CODEC_ID_H261){ |
| 376 | c->sub_motion_search= no_sub_motion_search; |
| 377 | } |
| 378 | |
| 379 | return 0; |
| 380 | } |
| 381 | |
| 382 | #define CHECK_SAD_HALF_MV(suffix, x, y) \ |
| 383 | {\ |
| 384 | d = s->mecc.pix_abs[size][(x ? 1 : 0) + (y ? 2 : 0)](NULL, pix, ptr + ((x) >> 1), stride, h); \ |
| 385 | d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\ |
| 386 | COPY3_IF_LT(dminh, d, dx, x, dy, y)\ |
| 387 | } |
| 388 | |
| 389 | static int sad_hpel_motion_search(MpegEncContext * s, |
| 390 | int *mx_ptr, int *my_ptr, int dmin, |
| 391 | int src_index, int ref_index, |
| 392 | int size, int h) |
| 393 | { |
| 394 | MotionEstContext * const c= &s->me; |
| 395 | const int penalty_factor= c->sub_penalty_factor; |
| 396 | int mx, my, dminh; |
| 397 | uint8_t *pix, *ptr; |
| 398 | int stride= c->stride; |
| 399 | LOAD_COMMON |
| 400 | |
| 401 | av_assert2(c->sub_flags == 0); |
| 402 | |
| 403 | if(c->skip){ |
| 404 | *mx_ptr = 0; |
| 405 | *my_ptr = 0; |
| 406 | return dmin; |
| 407 | } |
| 408 | |
| 409 | pix = c->src[src_index][0]; |
| 410 | |
| 411 | mx = *mx_ptr; |
| 412 | my = *my_ptr; |
| 413 | ptr = c->ref[ref_index][0] + (my * stride) + mx; |
| 414 | |
| 415 | dminh = dmin; |
| 416 | |
| 417 | if (mx > xmin && mx < xmax && |
| 418 | my > ymin && my < ymax) { |
| 419 | int dx=0, dy=0; |
| 420 | int d, pen_x, pen_y; |
| 421 | const int index= (my<<ME_MAP_SHIFT) + mx; |
| 422 | const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; |
| 423 | const int l= score_map[(index- 1 )&(ME_MAP_SIZE-1)]; |
| 424 | const int r= score_map[(index+ 1 )&(ME_MAP_SIZE-1)]; |
| 425 | const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)]; |
| 426 | mx<<=1; |
| 427 | my<<=1; |
| 428 | |
| 429 | |
| 430 | pen_x= pred_x + mx; |
| 431 | pen_y= pred_y + my; |
| 432 | |
| 433 | ptr-= stride; |
| 434 | if(t<=b){ |
| 435 | CHECK_SAD_HALF_MV(y2 , 0, -1) |
| 436 | if(l<=r){ |
| 437 | CHECK_SAD_HALF_MV(xy2, -1, -1) |
| 438 | if(t+r<=b+l){ |
| 439 | CHECK_SAD_HALF_MV(xy2, +1, -1) |
| 440 | ptr+= stride; |
| 441 | }else{ |
| 442 | ptr+= stride; |
| 443 | CHECK_SAD_HALF_MV(xy2, -1, +1) |
| 444 | } |
| 445 | CHECK_SAD_HALF_MV(x2 , -1, 0) |
| 446 | }else{ |
| 447 | CHECK_SAD_HALF_MV(xy2, +1, -1) |
| 448 | if(t+l<=b+r){ |
| 449 | CHECK_SAD_HALF_MV(xy2, -1, -1) |
| 450 | ptr+= stride; |
| 451 | }else{ |
| 452 | ptr+= stride; |
| 453 | CHECK_SAD_HALF_MV(xy2, +1, +1) |
| 454 | } |
| 455 | CHECK_SAD_HALF_MV(x2 , +1, 0) |
| 456 | } |
| 457 | }else{ |
| 458 | if(l<=r){ |
| 459 | if(t+l<=b+r){ |
| 460 | CHECK_SAD_HALF_MV(xy2, -1, -1) |
| 461 | ptr+= stride; |
| 462 | }else{ |
| 463 | ptr+= stride; |
| 464 | CHECK_SAD_HALF_MV(xy2, +1, +1) |
| 465 | } |
| 466 | CHECK_SAD_HALF_MV(x2 , -1, 0) |
| 467 | CHECK_SAD_HALF_MV(xy2, -1, +1) |
| 468 | }else{ |
| 469 | if(t+r<=b+l){ |
| 470 | CHECK_SAD_HALF_MV(xy2, +1, -1) |
| 471 | ptr+= stride; |
| 472 | }else{ |
| 473 | ptr+= stride; |
| 474 | CHECK_SAD_HALF_MV(xy2, -1, +1) |
| 475 | } |
| 476 | CHECK_SAD_HALF_MV(x2 , +1, 0) |
| 477 | CHECK_SAD_HALF_MV(xy2, +1, +1) |
| 478 | } |
| 479 | CHECK_SAD_HALF_MV(y2 , 0, +1) |
| 480 | } |
| 481 | mx+=dx; |
| 482 | my+=dy; |
| 483 | |
| 484 | }else{ |
| 485 | mx<<=1; |
| 486 | my<<=1; |
| 487 | } |
| 488 | |
| 489 | *mx_ptr = mx; |
| 490 | *my_ptr = my; |
| 491 | return dminh; |
| 492 | } |
| 493 | |
| 494 | static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4) |
| 495 | { |
| 496 | const int xy= s->mb_x + s->mb_y*s->mb_stride; |
| 497 | |
| 498 | s->p_mv_table[xy][0] = mx; |
| 499 | s->p_mv_table[xy][1] = my; |
| 500 | |
| 501 | /* has already been set to the 4 MV if 4MV is done */ |
| 502 | if(mv4){ |
| 503 | int mot_xy= s->block_index[0]; |
| 504 | |
| 505 | s->current_picture.motion_val[0][mot_xy ][0] = mx; |
| 506 | s->current_picture.motion_val[0][mot_xy ][1] = my; |
| 507 | s->current_picture.motion_val[0][mot_xy + 1][0] = mx; |
| 508 | s->current_picture.motion_val[0][mot_xy + 1][1] = my; |
| 509 | |
| 510 | mot_xy += s->b8_stride; |
| 511 | s->current_picture.motion_val[0][mot_xy ][0] = mx; |
| 512 | s->current_picture.motion_val[0][mot_xy ][1] = my; |
| 513 | s->current_picture.motion_val[0][mot_xy + 1][0] = mx; |
| 514 | s->current_picture.motion_val[0][mot_xy + 1][1] = my; |
| 515 | } |
| 516 | } |
| 517 | |
| 518 | /** |
| 519 | * get fullpel ME search limits. |
| 520 | */ |
| 521 | static inline void get_limits(MpegEncContext *s, int x, int y) |
| 522 | { |
| 523 | MotionEstContext * const c= &s->me; |
| 524 | int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL)); |
| 525 | int max_range = MAX_MV >> (1 + !!(c->flags&FLAG_QPEL)); |
| 526 | /* |
| 527 | if(c->avctx->me_range) c->range= c->avctx->me_range >> 1; |
| 528 | else c->range= 16; |
| 529 | */ |
| 530 | if (s->unrestricted_mv) { |
| 531 | c->xmin = - x - 16; |
| 532 | c->ymin = - y - 16; |
| 533 | c->xmax = - x + s->width; |
| 534 | c->ymax = - y + s->height; |
| 535 | } else if (s->out_format == FMT_H261){ |
| 536 | // Search range of H261 is different from other codec standards |
| 537 | c->xmin = (x > 15) ? - 15 : 0; |
| 538 | c->ymin = (y > 15) ? - 15 : 0; |
| 539 | c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0; |
| 540 | c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0; |
| 541 | } else { |
| 542 | c->xmin = - x; |
| 543 | c->ymin = - y; |
| 544 | c->xmax = - x + s->mb_width *16 - 16; |
| 545 | c->ymax = - y + s->mb_height*16 - 16; |
| 546 | } |
| 547 | if(!range || range > max_range) |
| 548 | range = max_range; |
| 549 | if(range){ |
| 550 | c->xmin = FFMAX(c->xmin,-range); |
| 551 | c->xmax = FFMIN(c->xmax, range); |
| 552 | c->ymin = FFMAX(c->ymin,-range); |
| 553 | c->ymax = FFMIN(c->ymax, range); |
| 554 | } |
| 555 | } |
| 556 | |
| 557 | static inline void init_mv4_ref(MotionEstContext *c){ |
| 558 | const int stride= c->stride; |
| 559 | |
| 560 | c->ref[1][0] = c->ref[0][0] + 8; |
| 561 | c->ref[2][0] = c->ref[0][0] + 8*stride; |
| 562 | c->ref[3][0] = c->ref[2][0] + 8; |
| 563 | c->src[1][0] = c->src[0][0] + 8; |
| 564 | c->src[2][0] = c->src[0][0] + 8*stride; |
| 565 | c->src[3][0] = c->src[2][0] + 8; |
| 566 | } |
| 567 | |
| 568 | static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift) |
| 569 | { |
| 570 | MotionEstContext * const c= &s->me; |
| 571 | const int size= 1; |
| 572 | const int h=8; |
| 573 | int block; |
| 574 | int P[10][2]; |
| 575 | int dmin_sum=0, mx4_sum=0, my4_sum=0, i; |
| 576 | int same=1; |
| 577 | const int stride= c->stride; |
| 578 | uint8_t *mv_penalty= c->current_mv_penalty; |
| 579 | int saftey_cliping= s->unrestricted_mv && (s->width&15) && (s->height&15); |
| 580 | |
| 581 | init_mv4_ref(c); |
| 582 | |
| 583 | for(block=0; block<4; block++){ |
| 584 | int mx4, my4; |
| 585 | int pred_x4, pred_y4; |
| 586 | int dmin4; |
| 587 | static const int off[4]= {2, 1, 1, -1}; |
| 588 | const int mot_stride = s->b8_stride; |
| 589 | const int mot_xy = s->block_index[block]; |
| 590 | |
| 591 | if(saftey_cliping){ |
| 592 | c->xmax = - 16*s->mb_x + s->width - 8*(block &1); |
| 593 | c->ymax = - 16*s->mb_y + s->height - 8*(block>>1); |
| 594 | } |
| 595 | |
| 596 | P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0]; |
| 597 | P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1]; |
| 598 | |
| 599 | if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift); |
| 600 | |
| 601 | /* special case for first line */ |
| 602 | if (s->first_slice_line && block<2) { |
| 603 | c->pred_x= pred_x4= P_LEFT[0]; |
| 604 | c->pred_y= pred_y4= P_LEFT[1]; |
| 605 | } else { |
| 606 | P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0]; |
| 607 | P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1]; |
| 608 | P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0]; |
| 609 | P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1]; |
| 610 | if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift); |
| 611 | if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift); |
| 612 | if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); |
| 613 | if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift); |
| 614 | |
| 615 | P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
| 616 | P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); |
| 617 | |
| 618 | c->pred_x= pred_x4 = P_MEDIAN[0]; |
| 619 | c->pred_y= pred_y4 = P_MEDIAN[1]; |
| 620 | } |
| 621 | P_MV1[0]= mx; |
| 622 | P_MV1[1]= my; |
| 623 | if(saftey_cliping) |
| 624 | for(i=1; i<10; i++){ |
| 625 | if (s->first_slice_line && block<2 && i>1 && i<9) |
| 626 | continue; |
| 627 | if (i>4 && i<9) |
| 628 | continue; |
| 629 | if(P[i][0] > (c->xmax<<shift)) P[i][0]= (c->xmax<<shift); |
| 630 | if(P[i][1] > (c->ymax<<shift)) P[i][1]= (c->ymax<<shift); |
| 631 | } |
| 632 | |
| 633 | dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift); |
| 634 | |
| 635 | dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h); |
| 636 | |
| 637 | if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) { |
| 638 | int dxy; |
| 639 | const int offset= ((block&1) + (block>>1)*stride)*8; |
| 640 | uint8_t *dest_y = c->scratchpad + offset; |
| 641 | if(s->quarter_sample){ |
| 642 | uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride; |
| 643 | dxy = ((my4 & 3) << 2) | (mx4 & 3); |
| 644 | |
| 645 | if(s->no_rounding) |
| 646 | s->qdsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y, ref, stride); |
| 647 | else |
| 648 | s->qdsp.put_qpel_pixels_tab[1][dxy](dest_y, ref, stride); |
| 649 | }else{ |
| 650 | uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride; |
| 651 | dxy = ((my4 & 1) << 1) | (mx4 & 1); |
| 652 | |
| 653 | if(s->no_rounding) |
| 654 | s->hdsp.put_no_rnd_pixels_tab[1][dxy](dest_y , ref , stride, h); |
| 655 | else |
| 656 | s->hdsp.put_pixels_tab [1][dxy](dest_y , ref , stride, h); |
| 657 | } |
| 658 | dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor; |
| 659 | }else |
| 660 | dmin_sum+= dmin4; |
| 661 | |
| 662 | if(s->quarter_sample){ |
| 663 | mx4_sum+= mx4/2; |
| 664 | my4_sum+= my4/2; |
| 665 | }else{ |
| 666 | mx4_sum+= mx4; |
| 667 | my4_sum+= my4; |
| 668 | } |
| 669 | |
| 670 | s->current_picture.motion_val[0][s->block_index[block]][0] = mx4; |
| 671 | s->current_picture.motion_val[0][s->block_index[block]][1] = my4; |
| 672 | |
| 673 | if(mx4 != mx || my4 != my) same=0; |
| 674 | } |
| 675 | |
| 676 | if(same) |
| 677 | return INT_MAX; |
| 678 | |
| 679 | if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) { |
| 680 | dmin_sum += s->mecc.mb_cmp[0](s, |
| 681 | s->new_picture.f->data[0] + |
| 682 | s->mb_x * 16 + s->mb_y * 16 * stride, |
| 683 | c->scratchpad, stride, 16); |
| 684 | } |
| 685 | |
| 686 | if(c->avctx->mb_cmp&FF_CMP_CHROMA){ |
| 687 | int dxy; |
| 688 | int mx, my; |
| 689 | int offset; |
| 690 | |
| 691 | mx= ff_h263_round_chroma(mx4_sum); |
| 692 | my= ff_h263_round_chroma(my4_sum); |
| 693 | dxy = ((my & 1) << 1) | (mx & 1); |
| 694 | |
| 695 | offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize; |
| 696 | |
| 697 | if(s->no_rounding){ |
| 698 | s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8); |
| 699 | s->hdsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8); |
| 700 | }else{ |
| 701 | s->hdsp.put_pixels_tab [1][dxy](c->scratchpad , s->last_picture.f->data[1] + offset, s->uvlinesize, 8); |
| 702 | s->hdsp.put_pixels_tab [1][dxy](c->scratchpad + 8, s->last_picture.f->data[2] + offset, s->uvlinesize, 8); |
| 703 | } |
| 704 | |
| 705 | dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture.f->data[1] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad, s->uvlinesize, 8); |
| 706 | dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture.f->data[2] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad + 8, s->uvlinesize, 8); |
| 707 | } |
| 708 | |
| 709 | c->pred_x= mx; |
| 710 | c->pred_y= my; |
| 711 | |
| 712 | switch(c->avctx->mb_cmp&0xFF){ |
| 713 | /*case FF_CMP_SSE: |
| 714 | return dmin_sum+ 32*s->qscale*s->qscale;*/ |
| 715 | case FF_CMP_RD: |
| 716 | return dmin_sum; |
| 717 | default: |
| 718 | return dmin_sum+ 11*c->mb_penalty_factor; |
| 719 | } |
| 720 | } |
| 721 | |
| 722 | static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){ |
| 723 | MotionEstContext * const c= &s->me; |
| 724 | |
| 725 | c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize; |
| 726 | c->src[1][0] = c->src[0][0] + s->linesize; |
| 727 | if(c->flags & FLAG_CHROMA){ |
| 728 | c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize; |
| 729 | c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize; |
| 730 | c->src[1][1] = c->src[0][1] + s->uvlinesize; |
| 731 | c->src[1][2] = c->src[0][2] + s->uvlinesize; |
| 732 | } |
| 733 | } |
| 734 | |
| 735 | static int interlaced_search(MpegEncContext *s, int ref_index, |
| 736 | int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select) |
| 737 | { |
| 738 | MotionEstContext * const c= &s->me; |
| 739 | const int size=0; |
| 740 | const int h=8; |
| 741 | int block; |
| 742 | int P[10][2]; |
| 743 | uint8_t * const mv_penalty= c->current_mv_penalty; |
| 744 | int same=1; |
| 745 | const int stride= 2*s->linesize; |
| 746 | int dmin_sum= 0; |
| 747 | const int mot_stride= s->mb_stride; |
| 748 | const int xy= s->mb_x + s->mb_y*mot_stride; |
| 749 | |
| 750 | c->ymin>>=1; |
| 751 | c->ymax>>=1; |
| 752 | c->stride<<=1; |
| 753 | c->uvstride<<=1; |
| 754 | init_interlaced_ref(s, ref_index); |
| 755 | |
| 756 | for(block=0; block<2; block++){ |
| 757 | int field_select; |
| 758 | int best_dmin= INT_MAX; |
| 759 | int best_field= -1; |
| 760 | |
| 761 | for(field_select=0; field_select<2; field_select++){ |
| 762 | int dmin, mx_i, my_i; |
| 763 | int16_t (*mv_table)[2]= mv_tables[block][field_select]; |
| 764 | |
| 765 | if(user_field_select){ |
| 766 | av_assert1(field_select==0 || field_select==1); |
| 767 | av_assert1(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1); |
| 768 | if(field_select_tables[block][xy] != field_select) |
| 769 | continue; |
| 770 | } |
| 771 | |
| 772 | P_LEFT[0] = mv_table[xy - 1][0]; |
| 773 | P_LEFT[1] = mv_table[xy - 1][1]; |
| 774 | if(P_LEFT[0] > (c->xmax<<1)) P_LEFT[0] = (c->xmax<<1); |
| 775 | |
| 776 | c->pred_x= P_LEFT[0]; |
| 777 | c->pred_y= P_LEFT[1]; |
| 778 | |
| 779 | if(!s->first_slice_line){ |
| 780 | P_TOP[0] = mv_table[xy - mot_stride][0]; |
| 781 | P_TOP[1] = mv_table[xy - mot_stride][1]; |
| 782 | P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0]; |
| 783 | P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1]; |
| 784 | if(P_TOP[1] > (c->ymax<<1)) P_TOP[1] = (c->ymax<<1); |
| 785 | if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1); |
| 786 | if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1); |
| 787 | if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1); |
| 788 | |
| 789 | P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
| 790 | P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); |
| 791 | } |
| 792 | P_MV1[0]= mx; //FIXME not correct if block != field_select |
| 793 | P_MV1[1]= my / 2; |
| 794 | |
| 795 | dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1); |
| 796 | |
| 797 | dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h); |
| 798 | |
| 799 | mv_table[xy][0]= mx_i; |
| 800 | mv_table[xy][1]= my_i; |
| 801 | |
| 802 | if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) { |
| 803 | int dxy; |
| 804 | |
| 805 | //FIXME chroma ME |
| 806 | uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride; |
| 807 | dxy = ((my_i & 1) << 1) | (mx_i & 1); |
| 808 | |
| 809 | if(s->no_rounding){ |
| 810 | s->hdsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref , stride, h); |
| 811 | }else{ |
| 812 | s->hdsp.put_pixels_tab [size][dxy](c->scratchpad, ref , stride, h); |
| 813 | } |
| 814 | dmin = s->mecc.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h); |
| 815 | dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor; |
| 816 | }else |
| 817 | dmin+= c->mb_penalty_factor; //field_select bits |
| 818 | |
| 819 | dmin += field_select != block; //slightly prefer same field |
| 820 | |
| 821 | if(dmin < best_dmin){ |
| 822 | best_dmin= dmin; |
| 823 | best_field= field_select; |
| 824 | } |
| 825 | } |
| 826 | { |
| 827 | int16_t (*mv_table)[2]= mv_tables[block][best_field]; |
| 828 | |
| 829 | if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all |
| 830 | if(mv_table[xy][1]&1) same=0; |
| 831 | if(mv_table[xy][1]*2 != my) same=0; |
| 832 | if(best_field != block) same=0; |
| 833 | } |
| 834 | |
| 835 | field_select_tables[block][xy]= best_field; |
| 836 | dmin_sum += best_dmin; |
| 837 | } |
| 838 | |
| 839 | c->ymin<<=1; |
| 840 | c->ymax<<=1; |
| 841 | c->stride>>=1; |
| 842 | c->uvstride>>=1; |
| 843 | |
| 844 | if(same) |
| 845 | return INT_MAX; |
| 846 | |
| 847 | switch(c->avctx->mb_cmp&0xFF){ |
| 848 | /*case FF_CMP_SSE: |
| 849 | return dmin_sum+ 32*s->qscale*s->qscale;*/ |
| 850 | case FF_CMP_RD: |
| 851 | return dmin_sum; |
| 852 | default: |
| 853 | return dmin_sum+ 11*c->mb_penalty_factor; |
| 854 | } |
| 855 | } |
| 856 | |
| 857 | static inline int get_penalty_factor(int lambda, int lambda2, int type){ |
| 858 | switch(type&0xFF){ |
| 859 | default: |
| 860 | case FF_CMP_SAD: |
| 861 | return lambda>>FF_LAMBDA_SHIFT; |
| 862 | case FF_CMP_DCT: |
| 863 | return (3*lambda)>>(FF_LAMBDA_SHIFT+1); |
| 864 | case FF_CMP_W53: |
| 865 | return (4*lambda)>>(FF_LAMBDA_SHIFT); |
| 866 | case FF_CMP_W97: |
| 867 | return (2*lambda)>>(FF_LAMBDA_SHIFT); |
| 868 | case FF_CMP_SATD: |
| 869 | case FF_CMP_DCT264: |
| 870 | return (2*lambda)>>FF_LAMBDA_SHIFT; |
| 871 | case FF_CMP_RD: |
| 872 | case FF_CMP_PSNR: |
| 873 | case FF_CMP_SSE: |
| 874 | case FF_CMP_NSSE: |
| 875 | return lambda2>>FF_LAMBDA_SHIFT; |
| 876 | case FF_CMP_BIT: |
| 877 | return 1; |
| 878 | } |
| 879 | } |
| 880 | |
| 881 | void ff_estimate_p_frame_motion(MpegEncContext * s, |
| 882 | int mb_x, int mb_y) |
| 883 | { |
| 884 | MotionEstContext * const c= &s->me; |
| 885 | uint8_t *pix, *ppix; |
| 886 | int sum, mx, my, dmin; |
| 887 | int varc; ///< the variance of the block (sum of squared (p[y][x]-average)) |
| 888 | int vard; ///< sum of squared differences with the estimated motion vector |
| 889 | int P[10][2]; |
| 890 | const int shift= 1+s->quarter_sample; |
| 891 | int mb_type=0; |
| 892 | Picture * const pic= &s->current_picture; |
| 893 | |
| 894 | init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0); |
| 895 | |
| 896 | av_assert0(s->quarter_sample==0 || s->quarter_sample==1); |
| 897 | av_assert0(s->linesize == c->stride); |
| 898 | av_assert0(s->uvlinesize == c->uvstride); |
| 899 | |
| 900 | c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp); |
| 901 | c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp); |
| 902 | c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp); |
| 903 | c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV; |
| 904 | |
| 905 | get_limits(s, 16*mb_x, 16*mb_y); |
| 906 | c->skip=0; |
| 907 | |
| 908 | /* intra / predictive decision */ |
| 909 | pix = c->src[0][0]; |
| 910 | sum = s->mpvencdsp.pix_sum(pix, s->linesize); |
| 911 | varc = s->mpvencdsp.pix_norm1(pix, s->linesize) - |
| 912 | (((unsigned) sum * sum) >> 8) + 500; |
| 913 | |
| 914 | pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8; |
| 915 | pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8; |
| 916 | c->mb_var_sum_temp += (varc+128)>>8; |
| 917 | |
| 918 | switch(s->me_method) { |
| 919 | case ME_ZERO: |
| 920 | default: |
| 921 | mx = 0; |
| 922 | my = 0; |
| 923 | dmin = 0; |
| 924 | break; |
| 925 | case ME_X1: |
| 926 | case ME_EPZS: |
| 927 | { |
| 928 | const int mot_stride = s->b8_stride; |
| 929 | const int mot_xy = s->block_index[0]; |
| 930 | |
| 931 | P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0]; |
| 932 | P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1]; |
| 933 | |
| 934 | if(P_LEFT[0] > (c->xmax<<shift)) P_LEFT[0] = (c->xmax<<shift); |
| 935 | |
| 936 | if(!s->first_slice_line) { |
| 937 | P_TOP[0] = s->current_picture.motion_val[0][mot_xy - mot_stride ][0]; |
| 938 | P_TOP[1] = s->current_picture.motion_val[0][mot_xy - mot_stride ][1]; |
| 939 | P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0]; |
| 940 | P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1]; |
| 941 | if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1] = (c->ymax<<shift); |
| 942 | if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift); |
| 943 | if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift); |
| 944 | |
| 945 | P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
| 946 | P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); |
| 947 | |
| 948 | if(s->out_format == FMT_H263){ |
| 949 | c->pred_x = P_MEDIAN[0]; |
| 950 | c->pred_y = P_MEDIAN[1]; |
| 951 | }else { /* mpeg1 at least */ |
| 952 | c->pred_x= P_LEFT[0]; |
| 953 | c->pred_y= P_LEFT[1]; |
| 954 | } |
| 955 | }else{ |
| 956 | c->pred_x= P_LEFT[0]; |
| 957 | c->pred_y= P_LEFT[1]; |
| 958 | } |
| 959 | |
| 960 | } |
| 961 | dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16); |
| 962 | |
| 963 | break; |
| 964 | } |
| 965 | |
| 966 | /* At this point (mx,my) are full-pell and the relative displacement */ |
| 967 | ppix = c->ref[0][0] + (my * s->linesize) + mx; |
| 968 | |
| 969 | vard = s->mecc.sse[0](NULL, pix, ppix, s->linesize, 16); |
| 970 | |
| 971 | pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8; |
| 972 | c->mc_mb_var_sum_temp += (vard+128)>>8; |
| 973 | |
| 974 | if(mb_type){ |
| 975 | int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100); |
| 976 | int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20; |
| 977 | c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score); |
| 978 | |
| 979 | if(mb_type == CANDIDATE_MB_TYPE_INTER){ |
| 980 | c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16); |
| 981 | set_p_mv_tables(s, mx, my, 1); |
| 982 | }else{ |
| 983 | mx <<=shift; |
| 984 | my <<=shift; |
| 985 | } |
| 986 | if(mb_type == CANDIDATE_MB_TYPE_INTER4V){ |
| 987 | h263_mv4_search(s, mx, my, shift); |
| 988 | |
| 989 | set_p_mv_tables(s, mx, my, 0); |
| 990 | } |
| 991 | if(mb_type == CANDIDATE_MB_TYPE_INTER_I){ |
| 992 | interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1); |
| 993 | } |
| 994 | }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){ |
| 995 | int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100); |
| 996 | int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20; |
| 997 | c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score); |
| 998 | |
| 999 | if (vard*2 + 200*256 > varc) |
| 1000 | mb_type|= CANDIDATE_MB_TYPE_INTRA; |
| 1001 | if (varc*2 + 200*256 > vard || s->qscale > 24){ |
| 1002 | // if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){ |
| 1003 | mb_type|= CANDIDATE_MB_TYPE_INTER; |
| 1004 | c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16); |
| 1005 | if (s->mpv_flags & FF_MPV_FLAG_MV0) |
| 1006 | if(mx || my) |
| 1007 | mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference |
| 1008 | }else{ |
| 1009 | mx <<=shift; |
| 1010 | my <<=shift; |
| 1011 | } |
| 1012 | if((s->flags&CODEC_FLAG_4MV) |
| 1013 | && !c->skip && varc>50<<8 && vard>10<<8){ |
| 1014 | if(h263_mv4_search(s, mx, my, shift) < INT_MAX) |
| 1015 | mb_type|=CANDIDATE_MB_TYPE_INTER4V; |
| 1016 | |
| 1017 | set_p_mv_tables(s, mx, my, 0); |
| 1018 | }else |
| 1019 | set_p_mv_tables(s, mx, my, 1); |
| 1020 | if((s->flags&CODEC_FLAG_INTERLACED_ME) |
| 1021 | && !c->skip){ //FIXME varc/d checks |
| 1022 | if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX) |
| 1023 | mb_type |= CANDIDATE_MB_TYPE_INTER_I; |
| 1024 | } |
| 1025 | }else{ |
| 1026 | int intra_score, i; |
| 1027 | mb_type= CANDIDATE_MB_TYPE_INTER; |
| 1028 | |
| 1029 | dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16); |
| 1030 | if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip) |
| 1031 | dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1); |
| 1032 | |
| 1033 | if((s->flags&CODEC_FLAG_4MV) |
| 1034 | && !c->skip && varc>50<<8 && vard>10<<8){ |
| 1035 | int dmin4= h263_mv4_search(s, mx, my, shift); |
| 1036 | if(dmin4 < dmin){ |
| 1037 | mb_type= CANDIDATE_MB_TYPE_INTER4V; |
| 1038 | dmin=dmin4; |
| 1039 | } |
| 1040 | } |
| 1041 | if((s->flags&CODEC_FLAG_INTERLACED_ME) |
| 1042 | && !c->skip){ //FIXME varc/d checks |
| 1043 | int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0); |
| 1044 | if(dmin_i < dmin){ |
| 1045 | mb_type = CANDIDATE_MB_TYPE_INTER_I; |
| 1046 | dmin= dmin_i; |
| 1047 | } |
| 1048 | } |
| 1049 | |
| 1050 | set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V); |
| 1051 | |
| 1052 | /* get intra luma score */ |
| 1053 | if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){ |
| 1054 | intra_score= varc - 500; |
| 1055 | }else{ |
| 1056 | unsigned mean = (sum+128)>>8; |
| 1057 | mean*= 0x01010101; |
| 1058 | |
| 1059 | for(i=0; i<16; i++){ |
| 1060 | *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean; |
| 1061 | *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean; |
| 1062 | *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean; |
| 1063 | *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean; |
| 1064 | } |
| 1065 | |
| 1066 | intra_score= s->mecc.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16); |
| 1067 | } |
| 1068 | intra_score += c->mb_penalty_factor*16; |
| 1069 | |
| 1070 | if(intra_score < dmin){ |
| 1071 | mb_type= CANDIDATE_MB_TYPE_INTRA; |
| 1072 | s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup |
| 1073 | }else |
| 1074 | s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0; |
| 1075 | |
| 1076 | { |
| 1077 | int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100); |
| 1078 | int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20; |
| 1079 | c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score); |
| 1080 | } |
| 1081 | } |
| 1082 | |
| 1083 | s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type; |
| 1084 | } |
| 1085 | |
| 1086 | int ff_pre_estimate_p_frame_motion(MpegEncContext * s, |
| 1087 | int mb_x, int mb_y) |
| 1088 | { |
| 1089 | MotionEstContext * const c= &s->me; |
| 1090 | int mx, my, dmin; |
| 1091 | int P[10][2]; |
| 1092 | const int shift= 1+s->quarter_sample; |
| 1093 | const int xy= mb_x + mb_y*s->mb_stride; |
| 1094 | init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0); |
| 1095 | |
| 1096 | av_assert0(s->quarter_sample==0 || s->quarter_sample==1); |
| 1097 | |
| 1098 | c->pre_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp); |
| 1099 | c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV; |
| 1100 | |
| 1101 | get_limits(s, 16*mb_x, 16*mb_y); |
| 1102 | c->skip=0; |
| 1103 | |
| 1104 | P_LEFT[0] = s->p_mv_table[xy + 1][0]; |
| 1105 | P_LEFT[1] = s->p_mv_table[xy + 1][1]; |
| 1106 | |
| 1107 | if(P_LEFT[0] < (c->xmin<<shift)) P_LEFT[0] = (c->xmin<<shift); |
| 1108 | |
| 1109 | /* special case for first line */ |
| 1110 | if (s->first_slice_line) { |
| 1111 | c->pred_x= P_LEFT[0]; |
| 1112 | c->pred_y= P_LEFT[1]; |
| 1113 | P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]= |
| 1114 | P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME |
| 1115 | } else { |
| 1116 | P_TOP[0] = s->p_mv_table[xy + s->mb_stride ][0]; |
| 1117 | P_TOP[1] = s->p_mv_table[xy + s->mb_stride ][1]; |
| 1118 | P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0]; |
| 1119 | P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1]; |
| 1120 | if(P_TOP[1] < (c->ymin<<shift)) P_TOP[1] = (c->ymin<<shift); |
| 1121 | if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift); |
| 1122 | if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift); |
| 1123 | |
| 1124 | P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
| 1125 | P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); |
| 1126 | |
| 1127 | c->pred_x = P_MEDIAN[0]; |
| 1128 | c->pred_y = P_MEDIAN[1]; |
| 1129 | } |
| 1130 | |
| 1131 | dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16); |
| 1132 | |
| 1133 | s->p_mv_table[xy][0] = mx<<shift; |
| 1134 | s->p_mv_table[xy][1] = my<<shift; |
| 1135 | |
| 1136 | return dmin; |
| 1137 | } |
| 1138 | |
| 1139 | static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y, |
| 1140 | int16_t (*mv_table)[2], int ref_index, int f_code) |
| 1141 | { |
| 1142 | MotionEstContext * const c= &s->me; |
| 1143 | int mx, my, dmin; |
| 1144 | int P[10][2]; |
| 1145 | const int shift= 1+s->quarter_sample; |
| 1146 | const int mot_stride = s->mb_stride; |
| 1147 | const int mot_xy = mb_y*mot_stride + mb_x; |
| 1148 | uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV; |
| 1149 | int mv_scale; |
| 1150 | |
| 1151 | c->penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp); |
| 1152 | c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp); |
| 1153 | c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp); |
| 1154 | c->current_mv_penalty= mv_penalty; |
| 1155 | |
| 1156 | get_limits(s, 16*mb_x, 16*mb_y); |
| 1157 | |
| 1158 | switch(s->me_method) { |
| 1159 | case ME_ZERO: |
| 1160 | default: |
| 1161 | mx = 0; |
| 1162 | my = 0; |
| 1163 | dmin = 0; |
| 1164 | break; |
| 1165 | case ME_X1: |
| 1166 | case ME_EPZS: |
| 1167 | P_LEFT[0] = mv_table[mot_xy - 1][0]; |
| 1168 | P_LEFT[1] = mv_table[mot_xy - 1][1]; |
| 1169 | |
| 1170 | if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift); |
| 1171 | |
| 1172 | /* special case for first line */ |
| 1173 | if (!s->first_slice_line) { |
| 1174 | P_TOP[0] = mv_table[mot_xy - mot_stride ][0]; |
| 1175 | P_TOP[1] = mv_table[mot_xy - mot_stride ][1]; |
| 1176 | P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0]; |
| 1177 | P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1]; |
| 1178 | if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift); |
| 1179 | if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift); |
| 1180 | if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift); |
| 1181 | |
| 1182 | P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
| 1183 | P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); |
| 1184 | } |
| 1185 | c->pred_x = P_LEFT[0]; |
| 1186 | c->pred_y = P_LEFT[1]; |
| 1187 | |
| 1188 | if(mv_table == s->b_forw_mv_table){ |
| 1189 | mv_scale= (s->pb_time<<16) / (s->pp_time<<shift); |
| 1190 | }else{ |
| 1191 | mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift); |
| 1192 | } |
| 1193 | |
| 1194 | dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16); |
| 1195 | |
| 1196 | break; |
| 1197 | } |
| 1198 | |
| 1199 | dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16); |
| 1200 | |
| 1201 | if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip) |
| 1202 | dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1); |
| 1203 | |
| 1204 | // s->mb_type[mb_y*s->mb_width + mb_x]= mb_type; |
| 1205 | mv_table[mot_xy][0]= mx; |
| 1206 | mv_table[mot_xy][1]= my; |
| 1207 | |
| 1208 | return dmin; |
| 1209 | } |
| 1210 | |
| 1211 | static inline int check_bidir_mv(MpegEncContext * s, |
| 1212 | int motion_fx, int motion_fy, |
| 1213 | int motion_bx, int motion_by, |
| 1214 | int pred_fx, int pred_fy, |
| 1215 | int pred_bx, int pred_by, |
| 1216 | int size, int h) |
| 1217 | { |
| 1218 | //FIXME optimize? |
| 1219 | //FIXME better f_code prediction (max mv & distance) |
| 1220 | //FIXME pointers |
| 1221 | MotionEstContext * const c= &s->me; |
| 1222 | uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame |
| 1223 | uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame |
| 1224 | int stride= c->stride; |
| 1225 | uint8_t *dest_y = c->scratchpad; |
| 1226 | uint8_t *ptr; |
| 1227 | int dxy; |
| 1228 | int src_x, src_y; |
| 1229 | int fbmin; |
| 1230 | uint8_t **src_data= c->src[0]; |
| 1231 | uint8_t **ref_data= c->ref[0]; |
| 1232 | uint8_t **ref2_data= c->ref[2]; |
| 1233 | |
| 1234 | if(s->quarter_sample){ |
| 1235 | dxy = ((motion_fy & 3) << 2) | (motion_fx & 3); |
| 1236 | src_x = motion_fx >> 2; |
| 1237 | src_y = motion_fy >> 2; |
| 1238 | |
| 1239 | ptr = ref_data[0] + (src_y * stride) + src_x; |
| 1240 | s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride); |
| 1241 | |
| 1242 | dxy = ((motion_by & 3) << 2) | (motion_bx & 3); |
| 1243 | src_x = motion_bx >> 2; |
| 1244 | src_y = motion_by >> 2; |
| 1245 | |
| 1246 | ptr = ref2_data[0] + (src_y * stride) + src_x; |
| 1247 | s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride); |
| 1248 | }else{ |
| 1249 | dxy = ((motion_fy & 1) << 1) | (motion_fx & 1); |
| 1250 | src_x = motion_fx >> 1; |
| 1251 | src_y = motion_fy >> 1; |
| 1252 | |
| 1253 | ptr = ref_data[0] + (src_y * stride) + src_x; |
| 1254 | s->hdsp.put_pixels_tab[size][dxy](dest_y , ptr , stride, h); |
| 1255 | |
| 1256 | dxy = ((motion_by & 1) << 1) | (motion_bx & 1); |
| 1257 | src_x = motion_bx >> 1; |
| 1258 | src_y = motion_by >> 1; |
| 1259 | |
| 1260 | ptr = ref2_data[0] + (src_y * stride) + src_x; |
| 1261 | s->hdsp.avg_pixels_tab[size][dxy](dest_y , ptr , stride, h); |
| 1262 | } |
| 1263 | |
| 1264 | fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor |
| 1265 | +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor |
| 1266 | + s->mecc.mb_cmp[size](s, src_data[0], dest_y, stride, h); // FIXME new_pic |
| 1267 | |
| 1268 | if(c->avctx->mb_cmp&FF_CMP_CHROMA){ |
| 1269 | } |
| 1270 | //FIXME CHROMA !!! |
| 1271 | |
| 1272 | return fbmin; |
| 1273 | } |
| 1274 | |
| 1275 | /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/ |
| 1276 | static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y) |
| 1277 | { |
| 1278 | MotionEstContext * const c= &s->me; |
| 1279 | const int mot_stride = s->mb_stride; |
| 1280 | const int xy = mb_y *mot_stride + mb_x; |
| 1281 | int fbmin; |
| 1282 | int pred_fx= s->b_bidir_forw_mv_table[xy-1][0]; |
| 1283 | int pred_fy= s->b_bidir_forw_mv_table[xy-1][1]; |
| 1284 | int pred_bx= s->b_bidir_back_mv_table[xy-1][0]; |
| 1285 | int pred_by= s->b_bidir_back_mv_table[xy-1][1]; |
| 1286 | int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0]; |
| 1287 | int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1]; |
| 1288 | int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0]; |
| 1289 | int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1]; |
| 1290 | const int flags= c->sub_flags; |
| 1291 | const int qpel= flags&FLAG_QPEL; |
| 1292 | const int shift= 1+qpel; |
| 1293 | const int xmin= c->xmin<<shift; |
| 1294 | const int ymin= c->ymin<<shift; |
| 1295 | const int xmax= c->xmax<<shift; |
| 1296 | const int ymax= c->ymax<<shift; |
| 1297 | #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by)) |
| 1298 | #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by)) |
| 1299 | int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by); |
| 1300 | uint8_t map[256] = { 0 }; |
| 1301 | |
| 1302 | map[hashidx&255] = 1; |
| 1303 | |
| 1304 | fbmin= check_bidir_mv(s, motion_fx, motion_fy, |
| 1305 | motion_bx, motion_by, |
| 1306 | pred_fx, pred_fy, |
| 1307 | pred_bx, pred_by, |
| 1308 | 0, 16); |
| 1309 | |
| 1310 | if(s->avctx->bidir_refine){ |
| 1311 | int end; |
| 1312 | static const uint8_t limittab[5]={0,8,32,64,80}; |
| 1313 | const int limit= limittab[s->avctx->bidir_refine]; |
| 1314 | static const int8_t vect[][4]={ |
| 1315 | { 0, 0, 0, 1}, { 0, 0, 0,-1}, { 0, 0, 1, 0}, { 0, 0,-1, 0}, { 0, 1, 0, 0}, { 0,-1, 0, 0}, { 1, 0, 0, 0}, {-1, 0, 0, 0}, |
| 1316 | |
| 1317 | { 0, 0, 1, 1}, { 0, 0,-1,-1}, { 0, 1, 1, 0}, { 0,-1,-1, 0}, { 1, 1, 0, 0}, {-1,-1, 0, 0}, { 1, 0, 0, 1}, {-1, 0, 0,-1}, |
| 1318 | { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0}, |
| 1319 | { 0, 0,-1, 1}, { 0, 0, 1,-1}, { 0,-1, 1, 0}, { 0, 1,-1, 0}, {-1, 1, 0, 0}, { 1,-1, 0, 0}, { 1, 0, 0,-1}, {-1, 0, 0, 1}, |
| 1320 | { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0}, |
| 1321 | |
| 1322 | { 0, 1, 1, 1}, { 0,-1,-1,-1}, { 1, 1, 1, 0}, {-1,-1,-1, 0}, { 1, 1, 0, 1}, {-1,-1, 0,-1}, { 1, 0, 1, 1}, {-1, 0,-1,-1}, |
| 1323 | { 0,-1, 1, 1}, { 0, 1,-1,-1}, {-1, 1, 1, 0}, { 1,-1,-1, 0}, { 1, 1, 0,-1}, {-1,-1, 0, 1}, { 1, 0,-1, 1}, {-1, 0, 1,-1}, |
| 1324 | { 0, 1,-1, 1}, { 0,-1, 1,-1}, { 1,-1, 1, 0}, {-1, 1,-1, 0}, {-1, 1, 0, 1}, { 1,-1, 0,-1}, { 1, 0, 1,-1}, {-1, 0,-1, 1}, |
| 1325 | { 0, 1, 1,-1}, { 0,-1,-1, 1}, { 1, 1,-1, 0}, {-1,-1, 1, 0}, { 1,-1, 0, 1}, {-1, 1, 0,-1}, {-1, 0, 1, 1}, { 1, 0,-1,-1}, |
| 1326 | |
| 1327 | { 1, 1, 1, 1}, {-1,-1,-1,-1}, |
| 1328 | { 1, 1, 1,-1}, {-1,-1,-1, 1}, { 1, 1,-1, 1}, {-1,-1, 1,-1}, { 1,-1, 1, 1}, {-1, 1,-1,-1}, {-1, 1, 1, 1}, { 1,-1,-1,-1}, |
| 1329 | { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1}, |
| 1330 | }; |
| 1331 | static const uint8_t hash[]={ |
| 1332 | HASH8( 0, 0, 0, 1), HASH8( 0, 0, 0,-1), HASH8( 0, 0, 1, 0), HASH8( 0, 0,-1, 0), HASH8( 0, 1, 0, 0), HASH8( 0,-1, 0, 0), HASH8( 1, 0, 0, 0), HASH8(-1, 0, 0, 0), |
| 1333 | |
| 1334 | HASH8( 0, 0, 1, 1), HASH8( 0, 0,-1,-1), HASH8( 0, 1, 1, 0), HASH8( 0,-1,-1, 0), HASH8( 1, 1, 0, 0), HASH8(-1,-1, 0, 0), HASH8( 1, 0, 0, 1), HASH8(-1, 0, 0,-1), |
| 1335 | HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0), |
| 1336 | HASH8( 0, 0,-1, 1), HASH8( 0, 0, 1,-1), HASH8( 0,-1, 1, 0), HASH8( 0, 1,-1, 0), HASH8(-1, 1, 0, 0), HASH8( 1,-1, 0, 0), HASH8( 1, 0, 0,-1), HASH8(-1, 0, 0, 1), |
| 1337 | HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0), |
| 1338 | |
| 1339 | HASH8( 0, 1, 1, 1), HASH8( 0,-1,-1,-1), HASH8( 1, 1, 1, 0), HASH8(-1,-1,-1, 0), HASH8( 1, 1, 0, 1), HASH8(-1,-1, 0,-1), HASH8( 1, 0, 1, 1), HASH8(-1, 0,-1,-1), |
| 1340 | HASH8( 0,-1, 1, 1), HASH8( 0, 1,-1,-1), HASH8(-1, 1, 1, 0), HASH8( 1,-1,-1, 0), HASH8( 1, 1, 0,-1), HASH8(-1,-1, 0, 1), HASH8( 1, 0,-1, 1), HASH8(-1, 0, 1,-1), |
| 1341 | HASH8( 0, 1,-1, 1), HASH8( 0,-1, 1,-1), HASH8( 1,-1, 1, 0), HASH8(-1, 1,-1, 0), HASH8(-1, 1, 0, 1), HASH8( 1,-1, 0,-1), HASH8( 1, 0, 1,-1), HASH8(-1, 0,-1, 1), |
| 1342 | HASH8( 0, 1, 1,-1), HASH8( 0,-1,-1, 1), HASH8( 1, 1,-1, 0), HASH8(-1,-1, 1, 0), HASH8( 1,-1, 0, 1), HASH8(-1, 1, 0,-1), HASH8(-1, 0, 1, 1), HASH8( 1, 0,-1,-1), |
| 1343 | |
| 1344 | HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1), |
| 1345 | HASH8( 1, 1, 1,-1), HASH8(-1,-1,-1, 1), HASH8( 1, 1,-1, 1), HASH8(-1,-1, 1,-1), HASH8( 1,-1, 1, 1), HASH8(-1, 1,-1,-1), HASH8(-1, 1, 1, 1), HASH8( 1,-1,-1,-1), |
| 1346 | HASH8( 1, 1,-1,-1), HASH8(-1,-1, 1, 1), HASH8( 1,-1,-1, 1), HASH8(-1, 1, 1,-1), HASH8( 1,-1, 1,-1), HASH8(-1, 1,-1, 1), |
| 1347 | }; |
| 1348 | |
| 1349 | #define CHECK_BIDIR(fx,fy,bx,by)\ |
| 1350 | if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\ |
| 1351 | &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\ |
| 1352 | &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\ |
| 1353 | int score;\ |
| 1354 | map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\ |
| 1355 | score= check_bidir_mv(s, motion_fx+fx, motion_fy+fy, motion_bx+bx, motion_by+by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);\ |
| 1356 | if(score < fbmin){\ |
| 1357 | hashidx += HASH(fx,fy,bx,by);\ |
| 1358 | fbmin= score;\ |
| 1359 | motion_fx+=fx;\ |
| 1360 | motion_fy+=fy;\ |
| 1361 | motion_bx+=bx;\ |
| 1362 | motion_by+=by;\ |
| 1363 | end=0;\ |
| 1364 | }\ |
| 1365 | } |
| 1366 | #define CHECK_BIDIR2(a,b,c,d)\ |
| 1367 | CHECK_BIDIR(a,b,c,d)\ |
| 1368 | CHECK_BIDIR(-(a),-(b),-(c),-(d)) |
| 1369 | |
| 1370 | do{ |
| 1371 | int i; |
| 1372 | int borderdist=0; |
| 1373 | end=1; |
| 1374 | |
| 1375 | CHECK_BIDIR2(0,0,0,1) |
| 1376 | CHECK_BIDIR2(0,0,1,0) |
| 1377 | CHECK_BIDIR2(0,1,0,0) |
| 1378 | CHECK_BIDIR2(1,0,0,0) |
| 1379 | |
| 1380 | for(i=8; i<limit; i++){ |
| 1381 | int fx= motion_fx+vect[i][0]; |
| 1382 | int fy= motion_fy+vect[i][1]; |
| 1383 | int bx= motion_bx+vect[i][2]; |
| 1384 | int by= motion_by+vect[i][3]; |
| 1385 | if(borderdist<=0){ |
| 1386 | int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin); |
| 1387 | int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin); |
| 1388 | if((a|b) < 0) |
| 1389 | map[(hashidx+hash[i])&255] = 1; |
| 1390 | } |
| 1391 | if(!map[(hashidx+hash[i])&255]){ |
| 1392 | int score; |
| 1393 | map[(hashidx+hash[i])&255] = 1; |
| 1394 | score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16); |
| 1395 | if(score < fbmin){ |
| 1396 | hashidx += hash[i]; |
| 1397 | fbmin= score; |
| 1398 | motion_fx=fx; |
| 1399 | motion_fy=fy; |
| 1400 | motion_bx=bx; |
| 1401 | motion_by=by; |
| 1402 | end=0; |
| 1403 | borderdist--; |
| 1404 | if(borderdist<=0){ |
| 1405 | int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin); |
| 1406 | int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin); |
| 1407 | borderdist= FFMIN(a,b); |
| 1408 | } |
| 1409 | } |
| 1410 | } |
| 1411 | } |
| 1412 | }while(!end); |
| 1413 | } |
| 1414 | |
| 1415 | s->b_bidir_forw_mv_table[xy][0]= motion_fx; |
| 1416 | s->b_bidir_forw_mv_table[xy][1]= motion_fy; |
| 1417 | s->b_bidir_back_mv_table[xy][0]= motion_bx; |
| 1418 | s->b_bidir_back_mv_table[xy][1]= motion_by; |
| 1419 | |
| 1420 | return fbmin; |
| 1421 | } |
| 1422 | |
| 1423 | static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y) |
| 1424 | { |
| 1425 | MotionEstContext * const c= &s->me; |
| 1426 | int P[10][2]; |
| 1427 | const int mot_stride = s->mb_stride; |
| 1428 | const int mot_xy = mb_y*mot_stride + mb_x; |
| 1429 | const int shift= 1+s->quarter_sample; |
| 1430 | int dmin, i; |
| 1431 | const int time_pp= s->pp_time; |
| 1432 | const int time_pb= s->pb_time; |
| 1433 | int mx, my, xmin, xmax, ymin, ymax; |
| 1434 | int16_t (*mv_table)[2]= s->b_direct_mv_table; |
| 1435 | |
| 1436 | c->current_mv_penalty= c->mv_penalty[1] + MAX_MV; |
| 1437 | ymin= xmin=(-32)>>shift; |
| 1438 | ymax= xmax= 31>>shift; |
| 1439 | |
| 1440 | if (IS_8X8(s->next_picture.mb_type[mot_xy])) { |
| 1441 | s->mv_type= MV_TYPE_8X8; |
| 1442 | }else{ |
| 1443 | s->mv_type= MV_TYPE_16X16; |
| 1444 | } |
| 1445 | |
| 1446 | for(i=0; i<4; i++){ |
| 1447 | int index= s->block_index[i]; |
| 1448 | int min, max; |
| 1449 | |
| 1450 | c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0]; |
| 1451 | c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1]; |
| 1452 | c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3)); |
| 1453 | c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3)); |
| 1454 | // c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3); |
| 1455 | // c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3); |
| 1456 | |
| 1457 | max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift; |
| 1458 | min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift; |
| 1459 | max+= 16*mb_x + 1; // +-1 is for the simpler rounding |
| 1460 | min+= 16*mb_x - 1; |
| 1461 | xmax= FFMIN(xmax, s->width - max); |
| 1462 | xmin= FFMAX(xmin, - 16 - min); |
| 1463 | |
| 1464 | max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift; |
| 1465 | min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift; |
| 1466 | max+= 16*mb_y + 1; // +-1 is for the simpler rounding |
| 1467 | min+= 16*mb_y - 1; |
| 1468 | ymax= FFMIN(ymax, s->height - max); |
| 1469 | ymin= FFMAX(ymin, - 16 - min); |
| 1470 | |
| 1471 | if(s->mv_type == MV_TYPE_16X16) break; |
| 1472 | } |
| 1473 | |
| 1474 | av_assert2(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16); |
| 1475 | |
| 1476 | if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){ |
| 1477 | s->b_direct_mv_table[mot_xy][0]= 0; |
| 1478 | s->b_direct_mv_table[mot_xy][1]= 0; |
| 1479 | |
| 1480 | return 256*256*256*64; |
| 1481 | } |
| 1482 | |
| 1483 | c->xmin= xmin; |
| 1484 | c->ymin= ymin; |
| 1485 | c->xmax= xmax; |
| 1486 | c->ymax= ymax; |
| 1487 | c->flags |= FLAG_DIRECT; |
| 1488 | c->sub_flags |= FLAG_DIRECT; |
| 1489 | c->pred_x=0; |
| 1490 | c->pred_y=0; |
| 1491 | |
| 1492 | P_LEFT[0] = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift); |
| 1493 | P_LEFT[1] = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift); |
| 1494 | |
| 1495 | /* special case for first line */ |
| 1496 | if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped |
| 1497 | P_TOP[0] = av_clip(mv_table[mot_xy - mot_stride ][0], xmin<<shift, xmax<<shift); |
| 1498 | P_TOP[1] = av_clip(mv_table[mot_xy - mot_stride ][1], ymin<<shift, ymax<<shift); |
| 1499 | P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1 ][0], xmin<<shift, xmax<<shift); |
| 1500 | P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1 ][1], ymin<<shift, ymax<<shift); |
| 1501 | |
| 1502 | P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]); |
| 1503 | P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]); |
| 1504 | } |
| 1505 | |
| 1506 | dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16); |
| 1507 | if(c->sub_flags&FLAG_QPEL) |
| 1508 | dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16); |
| 1509 | else |
| 1510 | dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16); |
| 1511 | |
| 1512 | if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip) |
| 1513 | dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1); |
| 1514 | |
| 1515 | get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed |
| 1516 | |
| 1517 | mv_table[mot_xy][0]= mx; |
| 1518 | mv_table[mot_xy][1]= my; |
| 1519 | c->flags &= ~FLAG_DIRECT; |
| 1520 | c->sub_flags &= ~FLAG_DIRECT; |
| 1521 | |
| 1522 | return dmin; |
| 1523 | } |
| 1524 | |
| 1525 | void ff_estimate_b_frame_motion(MpegEncContext * s, |
| 1526 | int mb_x, int mb_y) |
| 1527 | { |
| 1528 | MotionEstContext * const c= &s->me; |
| 1529 | const int penalty_factor= c->mb_penalty_factor; |
| 1530 | int fmin, bmin, dmin, fbmin, bimin, fimin; |
| 1531 | int type=0; |
| 1532 | const int xy = mb_y*s->mb_stride + mb_x; |
| 1533 | init_ref(c, s->new_picture.f->data, s->last_picture.f->data, |
| 1534 | s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2); |
| 1535 | |
| 1536 | get_limits(s, 16*mb_x, 16*mb_y); |
| 1537 | |
| 1538 | c->skip=0; |
| 1539 | |
| 1540 | if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) { |
| 1541 | int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0 |
| 1542 | |
| 1543 | score= ((unsigned)(score*score + 128*256))>>16; |
| 1544 | c->mc_mb_var_sum_temp += score; |
| 1545 | s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE |
| 1546 | s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0; |
| 1547 | |
| 1548 | return; |
| 1549 | } |
| 1550 | |
| 1551 | if (s->codec_id == AV_CODEC_ID_MPEG4) |
| 1552 | dmin= direct_search(s, mb_x, mb_y); |
| 1553 | else |
| 1554 | dmin= INT_MAX; |
| 1555 | //FIXME penalty stuff for non mpeg4 |
| 1556 | c->skip=0; |
| 1557 | fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + |
| 1558 | 3 * penalty_factor; |
| 1559 | |
| 1560 | c->skip=0; |
| 1561 | bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + |
| 1562 | 2 * penalty_factor; |
| 1563 | av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]); |
| 1564 | |
| 1565 | c->skip=0; |
| 1566 | fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor; |
| 1567 | av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin); |
| 1568 | |
| 1569 | if(s->flags & CODEC_FLAG_INTERLACED_ME){ |
| 1570 | //FIXME mb type penalty |
| 1571 | c->skip=0; |
| 1572 | c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV; |
| 1573 | fimin= interlaced_search(s, 0, |
| 1574 | s->b_field_mv_table[0], s->b_field_select_table[0], |
| 1575 | s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0); |
| 1576 | c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV; |
| 1577 | bimin= interlaced_search(s, 2, |
| 1578 | s->b_field_mv_table[1], s->b_field_select_table[1], |
| 1579 | s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0); |
| 1580 | }else |
| 1581 | fimin= bimin= INT_MAX; |
| 1582 | |
| 1583 | { |
| 1584 | int score= fmin; |
| 1585 | type = CANDIDATE_MB_TYPE_FORWARD; |
| 1586 | |
| 1587 | if (dmin <= score){ |
| 1588 | score = dmin; |
| 1589 | type = CANDIDATE_MB_TYPE_DIRECT; |
| 1590 | } |
| 1591 | if(bmin<score){ |
| 1592 | score=bmin; |
| 1593 | type= CANDIDATE_MB_TYPE_BACKWARD; |
| 1594 | } |
| 1595 | if(fbmin<score){ |
| 1596 | score=fbmin; |
| 1597 | type= CANDIDATE_MB_TYPE_BIDIR; |
| 1598 | } |
| 1599 | if(fimin<score){ |
| 1600 | score=fimin; |
| 1601 | type= CANDIDATE_MB_TYPE_FORWARD_I; |
| 1602 | } |
| 1603 | if(bimin<score){ |
| 1604 | score=bimin; |
| 1605 | type= CANDIDATE_MB_TYPE_BACKWARD_I; |
| 1606 | } |
| 1607 | |
| 1608 | score= ((unsigned)(score*score + 128*256))>>16; |
| 1609 | c->mc_mb_var_sum_temp += score; |
| 1610 | s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE |
| 1611 | } |
| 1612 | |
| 1613 | if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){ |
| 1614 | type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT; |
| 1615 | if(fimin < INT_MAX) |
| 1616 | type |= CANDIDATE_MB_TYPE_FORWARD_I; |
| 1617 | if(bimin < INT_MAX) |
| 1618 | type |= CANDIDATE_MB_TYPE_BACKWARD_I; |
| 1619 | if(fimin < INT_MAX && bimin < INT_MAX){ |
| 1620 | type |= CANDIDATE_MB_TYPE_BIDIR_I; |
| 1621 | } |
| 1622 | //FIXME something smarter |
| 1623 | if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB |
| 1624 | if (s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && |
| 1625 | s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy]) |
| 1626 | type |= CANDIDATE_MB_TYPE_DIRECT0; |
| 1627 | } |
| 1628 | |
| 1629 | s->mb_type[mb_y*s->mb_stride + mb_x]= type; |
| 1630 | } |
| 1631 | |
| 1632 | /* find best f_code for ME which do unlimited searches */ |
| 1633 | int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type) |
| 1634 | { |
| 1635 | if(s->me_method>=ME_EPZS){ |
| 1636 | int score[8]; |
| 1637 | int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2); |
| 1638 | uint8_t * fcode_tab= s->fcode_tab; |
| 1639 | int best_fcode=-1; |
| 1640 | int best_score=-10000000; |
| 1641 | |
| 1642 | if(s->msmpeg4_version) |
| 1643 | range= FFMIN(range, 16); |
| 1644 | else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL) |
| 1645 | range= FFMIN(range, 256); |
| 1646 | |
| 1647 | for(i=0; i<8; i++) score[i]= s->mb_num*(8-i); |
| 1648 | |
| 1649 | for(y=0; y<s->mb_height; y++){ |
| 1650 | int x; |
| 1651 | int xy= y*s->mb_stride; |
| 1652 | for(x=0; x<s->mb_width; x++){ |
| 1653 | if(s->mb_type[xy] & type){ |
| 1654 | int mx= mv_table[xy][0]; |
| 1655 | int my= mv_table[xy][1]; |
| 1656 | int fcode= FFMAX(fcode_tab[mx + MAX_MV], |
| 1657 | fcode_tab[my + MAX_MV]); |
| 1658 | int j; |
| 1659 | |
| 1660 | if(mx >= range || mx < -range || |
| 1661 | my >= range || my < -range) |
| 1662 | continue; |
| 1663 | |
| 1664 | for(j=0; j<fcode && j<8; j++){ |
| 1665 | if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy]) |
| 1666 | score[j]-= 170; |
| 1667 | } |
| 1668 | } |
| 1669 | xy++; |
| 1670 | } |
| 1671 | } |
| 1672 | |
| 1673 | for(i=1; i<8; i++){ |
| 1674 | if(score[i] > best_score){ |
| 1675 | best_score= score[i]; |
| 1676 | best_fcode= i; |
| 1677 | } |
| 1678 | } |
| 1679 | |
| 1680 | return best_fcode; |
| 1681 | }else{ |
| 1682 | return 1; |
| 1683 | } |
| 1684 | } |
| 1685 | |
| 1686 | void ff_fix_long_p_mvs(MpegEncContext * s) |
| 1687 | { |
| 1688 | MotionEstContext * const c= &s->me; |
| 1689 | const int f_code= s->f_code; |
| 1690 | int y, range; |
| 1691 | av_assert0(s->pict_type==AV_PICTURE_TYPE_P); |
| 1692 | |
| 1693 | range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code); |
| 1694 | |
| 1695 | av_assert0(range <= 16 || !s->msmpeg4_version); |
| 1696 | av_assert0(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)); |
| 1697 | |
| 1698 | if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range; |
| 1699 | |
| 1700 | if(s->flags&CODEC_FLAG_4MV){ |
| 1701 | const int wrap= s->b8_stride; |
| 1702 | |
| 1703 | /* clip / convert to intra 8x8 type MVs */ |
| 1704 | for(y=0; y<s->mb_height; y++){ |
| 1705 | int xy= y*2*wrap; |
| 1706 | int i= y*s->mb_stride; |
| 1707 | int x; |
| 1708 | |
| 1709 | for(x=0; x<s->mb_width; x++){ |
| 1710 | if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){ |
| 1711 | int block; |
| 1712 | for(block=0; block<4; block++){ |
| 1713 | int off= (block& 1) + (block>>1)*wrap; |
| 1714 | int mx = s->current_picture.motion_val[0][ xy + off ][0]; |
| 1715 | int my = s->current_picture.motion_val[0][ xy + off ][1]; |
| 1716 | |
| 1717 | if( mx >=range || mx <-range |
| 1718 | || my >=range || my <-range){ |
| 1719 | s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V; |
| 1720 | s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA; |
| 1721 | s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA; |
| 1722 | } |
| 1723 | } |
| 1724 | } |
| 1725 | xy+=2; |
| 1726 | i++; |
| 1727 | } |
| 1728 | } |
| 1729 | } |
| 1730 | } |
| 1731 | |
| 1732 | /** |
| 1733 | * |
| 1734 | * @param truncate 1 for truncation, 0 for using intra |
| 1735 | */ |
| 1736 | void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select, |
| 1737 | int16_t (*mv_table)[2], int f_code, int type, int truncate) |
| 1738 | { |
| 1739 | MotionEstContext * const c= &s->me; |
| 1740 | int y, h_range, v_range; |
| 1741 | |
| 1742 | // RAL: 8 in MPEG-1, 16 in MPEG-4 |
| 1743 | int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code); |
| 1744 | |
| 1745 | if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range; |
| 1746 | |
| 1747 | h_range= range; |
| 1748 | v_range= field_select_table ? range>>1 : range; |
| 1749 | |
| 1750 | /* clip / convert to intra 16x16 type MVs */ |
| 1751 | for(y=0; y<s->mb_height; y++){ |
| 1752 | int x; |
| 1753 | int xy= y*s->mb_stride; |
| 1754 | for(x=0; x<s->mb_width; x++){ |
| 1755 | if (s->mb_type[xy] & type){ // RAL: "type" test added... |
| 1756 | if (!field_select_table || field_select_table[xy] == field_select) { |
| 1757 | if( mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range |
| 1758 | || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){ |
| 1759 | |
| 1760 | if(truncate){ |
| 1761 | if (mv_table[xy][0] > h_range-1) mv_table[xy][0]= h_range-1; |
| 1762 | else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range; |
| 1763 | if (mv_table[xy][1] > v_range-1) mv_table[xy][1]= v_range-1; |
| 1764 | else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range; |
| 1765 | }else{ |
| 1766 | s->mb_type[xy] &= ~type; |
| 1767 | s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA; |
| 1768 | mv_table[xy][0]= |
| 1769 | mv_table[xy][1]= 0; |
| 1770 | } |
| 1771 | } |
| 1772 | } |
| 1773 | } |
| 1774 | xy++; |
| 1775 | } |
| 1776 | } |
| 1777 | } |