• Main Page
  • Related Pages
  • Modules
  • Data Structures
  • Files
  • Examples
  • File List
  • Globals

libavcodec/motion_est.c

Go to the documentation of this file.
00001 /*
00002  * Motion estimation
00003  * Copyright (c) 2000,2001 Fabrice Bellard
00004  * Copyright (c) 2002-2004 Michael Niedermayer
00005  *
00006  * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
00007  *
00008  * This file is part of Libav.
00009  *
00010  * Libav is free software; you can redistribute it and/or
00011  * modify it under the terms of the GNU Lesser General Public
00012  * License as published by the Free Software Foundation; either
00013  * version 2.1 of the License, or (at your option) any later version.
00014  *
00015  * Libav is distributed in the hope that it will be useful,
00016  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00017  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00018  * Lesser General Public License for more details.
00019  *
00020  * You should have received a copy of the GNU Lesser General Public
00021  * License along with Libav; if not, write to the Free Software
00022  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00023  */
00024 
00030 #include <stdlib.h>
00031 #include <stdio.h>
00032 #include <limits.h>
00033 #include "libavutil/intmath.h"
00034 #include "avcodec.h"
00035 #include "dsputil.h"
00036 #include "mathops.h"
00037 #include "mpegvideo.h"
00038 
00039 #undef NDEBUG
00040 #include <assert.h>
00041 
00042 #define SQ(a) ((a)*(a))
00043 
00044 #define P_LEFT P[1]
00045 #define P_TOP P[2]
00046 #define P_TOPRIGHT P[3]
00047 #define P_MEDIAN P[4]
00048 #define P_MV1 P[9]
00049 
00050 static inline int sad_hpel_motion_search(MpegEncContext * s,
00051                                   int *mx_ptr, int *my_ptr, int dmin,
00052                                   int src_index, int ref_index,
00053                                   int size, int h);
00054 
00055 static inline unsigned update_map_generation(MotionEstContext *c)
00056 {
00057     c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
00058     if(c->map_generation==0){
00059         c->map_generation= 1<<(ME_MAP_MV_BITS*2);
00060         memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
00061     }
00062     return c->map_generation;
00063 }
00064 
00065 /* shape adaptive search stuff */
00066 typedef struct Minima{
00067     int height;
00068     int x, y;
00069     int checked;
00070 }Minima;
00071 
00072 static int minima_cmp(const void *a, const void *b){
00073     const Minima *da = (const Minima *) a;
00074     const Minima *db = (const Minima *) b;
00075 
00076     return da->height - db->height;
00077 }
00078 
00079 #define FLAG_QPEL   1 //must be 1
00080 #define FLAG_CHROMA 2
00081 #define FLAG_DIRECT 4
00082 
00083 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){
00084     const int offset[3]= {
00085           y*c->  stride + x,
00086         ((y*c->uvstride + x)>>1),
00087         ((y*c->uvstride + x)>>1),
00088     };
00089     int i;
00090     for(i=0; i<3; i++){
00091         c->src[0][i]= src [i] + offset[i];
00092         c->ref[0][i]= ref [i] + offset[i];
00093     }
00094     if(ref_index){
00095         for(i=0; i<3; i++){
00096             c->ref[ref_index][i]= ref2[i] + offset[i];
00097         }
00098     }
00099 }
00100 
00101 static int get_flags(MotionEstContext *c, int direct, int chroma){
00102     return   ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
00103            + (direct ? FLAG_DIRECT : 0)
00104            + (chroma ? FLAG_CHROMA : 0);
00105 }
00106 
00107 static av_always_inline int cmp_direct_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
00108                       const int size, const int h, int ref_index, int src_index,
00109                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){
00110     MotionEstContext * const c= &s->me;
00111     const int stride= c->stride;
00112     const int hx= subx + (x<<(1+qpel));
00113     const int hy= suby + (y<<(1+qpel));
00114     uint8_t * const * const ref= c->ref[ref_index];
00115     uint8_t * const * const src= c->src[src_index];
00116     int d;
00117     //FIXME check chroma 4mv, (no crashes ...)
00118         assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
00119         if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
00120             const int time_pp= s->pp_time;
00121             const int time_pb= s->pb_time;
00122             const int mask= 2*qpel+1;
00123             if(s->mv_type==MV_TYPE_8X8){
00124                 int i;
00125                 for(i=0; i<4; i++){
00126                     int fx = c->direct_basis_mv[i][0] + hx;
00127                     int fy = c->direct_basis_mv[i][1] + hy;
00128                     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));
00129                     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));
00130                     int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
00131                     int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
00132 
00133                     uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
00134                     if(qpel){
00135                         c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
00136                         c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
00137                     }else{
00138                         c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
00139                         c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
00140                     }
00141                 }
00142             }else{
00143                 int fx = c->direct_basis_mv[0][0] + hx;
00144                 int fy = c->direct_basis_mv[0][1] + hy;
00145                 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
00146                 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
00147                 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
00148                 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
00149 
00150                 if(qpel){
00151                     c->qpel_put[1][fxy](c->temp               , ref[0] + (fx>>2) + (fy>>2)*stride               , stride);
00152                     c->qpel_put[1][fxy](c->temp + 8           , ref[0] + (fx>>2) + (fy>>2)*stride + 8           , stride);
00153                     c->qpel_put[1][fxy](c->temp     + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride     + 8*stride, stride);
00154                     c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
00155                     c->qpel_avg[1][bxy](c->temp               , ref[8] + (bx>>2) + (by>>2)*stride               , stride);
00156                     c->qpel_avg[1][bxy](c->temp + 8           , ref[8] + (bx>>2) + (by>>2)*stride + 8           , stride);
00157                     c->qpel_avg[1][bxy](c->temp     + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride     + 8*stride, stride);
00158                     c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
00159                 }else{
00160                     assert((fx>>1) + 16*s->mb_x >= -16);
00161                     assert((fy>>1) + 16*s->mb_y >= -16);
00162                     assert((fx>>1) + 16*s->mb_x <= s->width);
00163                     assert((fy>>1) + 16*s->mb_y <= s->height);
00164                     assert((bx>>1) + 16*s->mb_x >= -16);
00165                     assert((by>>1) + 16*s->mb_y >= -16);
00166                     assert((bx>>1) + 16*s->mb_x <= s->width);
00167                     assert((by>>1) + 16*s->mb_y <= s->height);
00168 
00169                     c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
00170                     c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
00171                 }
00172             }
00173             d = cmp_func(s, c->temp, src[0], stride, 16);
00174         }else
00175             d= 256*256*256*32;
00176     return d;
00177 }
00178 
00179 static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
00180                       const int size, const int h, int ref_index, int src_index,
00181                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){
00182     MotionEstContext * const c= &s->me;
00183     const int stride= c->stride;
00184     const int uvstride= c->uvstride;
00185     const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
00186     const int hx= subx + (x<<(1+qpel));
00187     const int hy= suby + (y<<(1+qpel));
00188     uint8_t * const * const ref= c->ref[ref_index];
00189     uint8_t * const * const src= c->src[src_index];
00190     int d;
00191     //FIXME check chroma 4mv, (no crashes ...)
00192         int uvdxy;              /* no, it might not be used uninitialized */
00193         if(dxy){
00194             if(qpel){
00195                 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
00196                 if(chroma){
00197                     int cx= hx/2;
00198                     int cy= hy/2;
00199                     cx= (cx>>1)|(cx&1);
00200                     cy= (cy>>1)|(cy&1);
00201                     uvdxy= (cx&1) + 2*(cy&1);
00202                     //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
00203                 }
00204             }else{
00205                 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
00206                 if(chroma)
00207                     uvdxy= dxy | (x&1) | (2*(y&1));
00208             }
00209             d = cmp_func(s, c->temp, src[0], stride, h);
00210         }else{
00211             d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
00212             if(chroma)
00213                 uvdxy= (x&1) + 2*(y&1);
00214         }
00215         if(chroma){
00216             uint8_t * const uvtemp= c->temp + 16*stride;
00217             c->hpel_put[size+1][uvdxy](uvtemp  , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
00218             c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
00219             d += chroma_cmp_func(s, uvtemp  , src[1], uvstride, h>>1);
00220             d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
00221         }
00222     return d;
00223 }
00224 
00225 static int cmp_simple(MpegEncContext *s, const int x, const int y,
00226                       int ref_index, int src_index,
00227                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){
00228     return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);
00229 }
00230 
00231 static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,
00232                       const int size, const int h, int ref_index, int src_index,
00233                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
00234     if(flags&FLAG_DIRECT){
00235         return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
00236     }else{
00237         return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
00238     }
00239 }
00240 
00241 static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
00242                       const int size, const int h, int ref_index, int src_index,
00243                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
00244     if(flags&FLAG_DIRECT){
00245         return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);
00246     }else{
00247         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);
00248     }
00249 }
00250 
00254 static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
00255                       const int size, const int h, int ref_index, int src_index,
00256                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
00257     if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)
00258        && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
00259        && flags==0 && h==16 && size==0 && subx==0 && suby==0){
00260         return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);
00261     }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)
00262        && subx==0 && suby==0){
00263         return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);
00264     }else{
00265         return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);
00266     }
00267 }
00268 
00269 static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
00270                       const int size, const int h, int ref_index, int src_index,
00271                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
00272     if(flags&FLAG_DIRECT){
00273         return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);
00274     }else{
00275         return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);
00276     }
00277 }
00278 
00279 static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
00280                       const int size, const int h, int ref_index, int src_index,
00281                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
00282     if(flags&FLAG_DIRECT){
00283         return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);
00284     }else{
00285         return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);
00286     }
00287 }
00288 
00289 #include "motion_est_template.c"
00290 
00291 static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
00292     return 0;
00293 }
00294 
00295 static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
00296 }
00297 
00298 int ff_init_me(MpegEncContext *s){
00299     MotionEstContext * const c= &s->me;
00300     int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
00301     int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
00302 
00303     if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){
00304         av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
00305         return -1;
00306     }
00307     //special case of snow is needed because snow uses its own iterative ME code
00308     if(s->me_method!=ME_ZERO && s->me_method!=ME_EPZS && s->me_method!=ME_X1 && s->avctx->codec_id != CODEC_ID_SNOW){
00309         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");
00310         return -1;
00311     }
00312 
00313     c->avctx= s->avctx;
00314 
00315     if(cache_size < 2*dia_size && !c->stride){
00316         av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
00317     }
00318 
00319     ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
00320     ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
00321     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
00322     ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
00323 
00324     c->flags    = get_flags(c, 0, c->avctx->me_cmp    &FF_CMP_CHROMA);
00325     c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
00326     c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp    &FF_CMP_CHROMA);
00327 
00328 /*FIXME s->no_rounding b_type*/
00329     if(s->flags&CODEC_FLAG_QPEL){
00330         c->sub_motion_search= qpel_motion_search;
00331         c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
00332         if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
00333         else               c->qpel_put= s->dsp.put_qpel_pixels_tab;
00334     }else{
00335         if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
00336             c->sub_motion_search= hpel_motion_search;
00337         else if(   c->avctx->me_sub_cmp == FF_CMP_SAD
00338                 && c->avctx->    me_cmp == FF_CMP_SAD
00339                 && c->avctx->    mb_cmp == FF_CMP_SAD)
00340             c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
00341         else
00342             c->sub_motion_search= hpel_motion_search;
00343     }
00344     c->hpel_avg= s->dsp.avg_pixels_tab;
00345     if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
00346     else               c->hpel_put= s->dsp.put_pixels_tab;
00347 
00348     if(s->linesize){
00349         c->stride  = s->linesize;
00350         c->uvstride= s->uvlinesize;
00351     }else{
00352         c->stride  = 16*s->mb_width + 32;
00353         c->uvstride=  8*s->mb_width + 16;
00354     }
00355 
00356     /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
00357      * not have yet, and even if we had, the motion estimation code
00358      * does not expect it. */
00359     if(s->codec_id != CODEC_ID_SNOW){
00360         if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
00361             s->dsp.me_cmp[2]= zero_cmp;
00362         }
00363         if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
00364             s->dsp.me_sub_cmp[2]= zero_cmp;
00365         }
00366         c->hpel_put[2][0]= c->hpel_put[2][1]=
00367         c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
00368     }
00369 
00370     if(s->codec_id == CODEC_ID_H261){
00371         c->sub_motion_search= no_sub_motion_search;
00372     }
00373 
00374     return 0;
00375 }
00376 
00377 static inline void no_motion_search(MpegEncContext * s,
00378                                     int *mx_ptr, int *my_ptr)
00379 {
00380     *mx_ptr = 16 * s->mb_x;
00381     *my_ptr = 16 * s->mb_y;
00382 }
00383 
00384 #define Z_THRESHOLD 256
00385 
00386 #define CHECK_SAD_HALF_MV(suffix, x, y) \
00387 {\
00388     d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
00389     d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
00390     COPY3_IF_LT(dminh, d, dx, x, dy, y)\
00391 }
00392 
00393 static inline int sad_hpel_motion_search(MpegEncContext * s,
00394                                   int *mx_ptr, int *my_ptr, int dmin,
00395                                   int src_index, int ref_index,
00396                                   int size, int h)
00397 {
00398     MotionEstContext * const c= &s->me;
00399     const int penalty_factor= c->sub_penalty_factor;
00400     int mx, my, dminh;
00401     uint8_t *pix, *ptr;
00402     int stride= c->stride;
00403     const int flags= c->sub_flags;
00404     LOAD_COMMON
00405 
00406     assert(flags == 0);
00407 
00408     if(c->skip){
00409 //    printf("S");
00410         *mx_ptr = 0;
00411         *my_ptr = 0;
00412         return dmin;
00413     }
00414 //    printf("N");
00415 
00416     pix = c->src[src_index][0];
00417 
00418     mx = *mx_ptr;
00419     my = *my_ptr;
00420     ptr = c->ref[ref_index][0] + (my * stride) + mx;
00421 
00422     dminh = dmin;
00423 
00424     if (mx > xmin && mx < xmax &&
00425         my > ymin && my < ymax) {
00426         int dx=0, dy=0;
00427         int d, pen_x, pen_y;
00428         const int index= (my<<ME_MAP_SHIFT) + mx;
00429         const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
00430         const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
00431         const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
00432         const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
00433         mx<<=1;
00434         my<<=1;
00435 
00436 
00437         pen_x= pred_x + mx;
00438         pen_y= pred_y + my;
00439 
00440         ptr-= stride;
00441         if(t<=b){
00442             CHECK_SAD_HALF_MV(y2 , 0, -1)
00443             if(l<=r){
00444                 CHECK_SAD_HALF_MV(xy2, -1, -1)
00445                 if(t+r<=b+l){
00446                     CHECK_SAD_HALF_MV(xy2, +1, -1)
00447                     ptr+= stride;
00448                 }else{
00449                     ptr+= stride;
00450                     CHECK_SAD_HALF_MV(xy2, -1, +1)
00451                 }
00452                 CHECK_SAD_HALF_MV(x2 , -1,  0)
00453             }else{
00454                 CHECK_SAD_HALF_MV(xy2, +1, -1)
00455                 if(t+l<=b+r){
00456                     CHECK_SAD_HALF_MV(xy2, -1, -1)
00457                     ptr+= stride;
00458                 }else{
00459                     ptr+= stride;
00460                     CHECK_SAD_HALF_MV(xy2, +1, +1)
00461                 }
00462                 CHECK_SAD_HALF_MV(x2 , +1,  0)
00463             }
00464         }else{
00465             if(l<=r){
00466                 if(t+l<=b+r){
00467                     CHECK_SAD_HALF_MV(xy2, -1, -1)
00468                     ptr+= stride;
00469                 }else{
00470                     ptr+= stride;
00471                     CHECK_SAD_HALF_MV(xy2, +1, +1)
00472                 }
00473                 CHECK_SAD_HALF_MV(x2 , -1,  0)
00474                 CHECK_SAD_HALF_MV(xy2, -1, +1)
00475             }else{
00476                 if(t+r<=b+l){
00477                     CHECK_SAD_HALF_MV(xy2, +1, -1)
00478                     ptr+= stride;
00479                 }else{
00480                     ptr+= stride;
00481                     CHECK_SAD_HALF_MV(xy2, -1, +1)
00482                 }
00483                 CHECK_SAD_HALF_MV(x2 , +1,  0)
00484                 CHECK_SAD_HALF_MV(xy2, +1, +1)
00485             }
00486             CHECK_SAD_HALF_MV(y2 ,  0, +1)
00487         }
00488         mx+=dx;
00489         my+=dy;
00490 
00491     }else{
00492         mx<<=1;
00493         my<<=1;
00494     }
00495 
00496     *mx_ptr = mx;
00497     *my_ptr = my;
00498     return dminh;
00499 }
00500 
00501 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
00502 {
00503     const int xy= s->mb_x + s->mb_y*s->mb_stride;
00504 
00505     s->p_mv_table[xy][0] = mx;
00506     s->p_mv_table[xy][1] = my;
00507 
00508     /* has already been set to the 4 MV if 4MV is done */
00509     if(mv4){
00510         int mot_xy= s->block_index[0];
00511 
00512         s->current_picture.f.motion_val[0][mot_xy    ][0] = mx;
00513         s->current_picture.f.motion_val[0][mot_xy    ][1] = my;
00514         s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
00515         s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
00516 
00517         mot_xy += s->b8_stride;
00518         s->current_picture.f.motion_val[0][mot_xy    ][0] = mx;
00519         s->current_picture.f.motion_val[0][mot_xy    ][1] = my;
00520         s->current_picture.f.motion_val[0][mot_xy + 1][0] = mx;
00521         s->current_picture.f.motion_val[0][mot_xy + 1][1] = my;
00522     }
00523 }
00524 
00528 static inline void get_limits(MpegEncContext *s, int x, int y)
00529 {
00530     MotionEstContext * const c= &s->me;
00531     int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
00532 /*
00533     if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
00534     else                   c->range= 16;
00535 */
00536     if (s->unrestricted_mv) {
00537         c->xmin = - x - 16;
00538         c->ymin = - y - 16;
00539         c->xmax = - x + s->mb_width *16;
00540         c->ymax = - y + s->mb_height*16;
00541     } else if (s->out_format == FMT_H261){
00542         // Search range of H261 is different from other codec standards
00543         c->xmin = (x > 15) ? - 15 : 0;
00544         c->ymin = (y > 15) ? - 15 : 0;
00545         c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
00546         c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
00547     } else {
00548         c->xmin = - x;
00549         c->ymin = - y;
00550         c->xmax = - x + s->mb_width *16 - 16;
00551         c->ymax = - y + s->mb_height*16 - 16;
00552     }
00553     if(range){
00554         c->xmin = FFMAX(c->xmin,-range);
00555         c->xmax = FFMIN(c->xmax, range);
00556         c->ymin = FFMAX(c->ymin,-range);
00557         c->ymax = FFMIN(c->ymax, range);
00558     }
00559 }
00560 
00561 static inline void init_mv4_ref(MotionEstContext *c){
00562     const int stride= c->stride;
00563 
00564     c->ref[1][0] = c->ref[0][0] + 8;
00565     c->ref[2][0] = c->ref[0][0] + 8*stride;
00566     c->ref[3][0] = c->ref[2][0] + 8;
00567     c->src[1][0] = c->src[0][0] + 8;
00568     c->src[2][0] = c->src[0][0] + 8*stride;
00569     c->src[3][0] = c->src[2][0] + 8;
00570 }
00571 
00572 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
00573 {
00574     MotionEstContext * const c= &s->me;
00575     const int size= 1;
00576     const int h=8;
00577     int block;
00578     int P[10][2];
00579     int dmin_sum=0, mx4_sum=0, my4_sum=0;
00580     int same=1;
00581     const int stride= c->stride;
00582     uint8_t *mv_penalty= c->current_mv_penalty;
00583 
00584     init_mv4_ref(c);
00585 
00586     for(block=0; block<4; block++){
00587         int mx4, my4;
00588         int pred_x4, pred_y4;
00589         int dmin4;
00590         static const int off[4]= {2, 1, 1, -1};
00591         const int mot_stride = s->b8_stride;
00592         const int mot_xy = s->block_index[block];
00593 
00594         P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
00595         P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
00596 
00597         if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
00598 
00599         /* special case for first line */
00600         if (s->first_slice_line && block<2) {
00601             c->pred_x= pred_x4= P_LEFT[0];
00602             c->pred_y= pred_y4= P_LEFT[1];
00603         } else {
00604             P_TOP[0]      = s->current_picture.f.motion_val[0][mot_xy - mot_stride             ][0];
00605             P_TOP[1]      = s->current_picture.f.motion_val[0][mot_xy - mot_stride             ][1];
00606             P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][0];
00607             P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + off[block]][1];
00608             if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
00609             if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
00610             if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
00611             if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
00612 
00613             P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
00614             P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
00615 
00616             c->pred_x= pred_x4 = P_MEDIAN[0];
00617             c->pred_y= pred_y4 = P_MEDIAN[1];
00618         }
00619         P_MV1[0]= mx;
00620         P_MV1[1]= my;
00621 
00622         dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
00623 
00624         dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
00625 
00626         if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
00627             int dxy;
00628             const int offset= ((block&1) + (block>>1)*stride)*8;
00629             uint8_t *dest_y = c->scratchpad + offset;
00630             if(s->quarter_sample){
00631                 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
00632                 dxy = ((my4 & 3) << 2) | (mx4 & 3);
00633 
00634                 if(s->no_rounding)
00635                     s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , stride);
00636                 else
00637                     s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , stride);
00638             }else{
00639                 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
00640                 dxy = ((my4 & 1) << 1) | (mx4 & 1);
00641 
00642                 if(s->no_rounding)
00643                     s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , stride, h);
00644                 else
00645                     s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , stride, h);
00646             }
00647             dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
00648         }else
00649             dmin_sum+= dmin4;
00650 
00651         if(s->quarter_sample){
00652             mx4_sum+= mx4/2;
00653             my4_sum+= my4/2;
00654         }else{
00655             mx4_sum+= mx4;
00656             my4_sum+= my4;
00657         }
00658 
00659         s->current_picture.f.motion_val[0][s->block_index[block]][0] = mx4;
00660         s->current_picture.f.motion_val[0][s->block_index[block]][1] = my4;
00661 
00662         if(mx4 != mx || my4 != my) same=0;
00663     }
00664 
00665     if(same)
00666         return INT_MAX;
00667 
00668     if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
00669         dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.f.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
00670     }
00671 
00672     if(c->avctx->mb_cmp&FF_CMP_CHROMA){
00673         int dxy;
00674         int mx, my;
00675         int offset;
00676 
00677         mx= ff_h263_round_chroma(mx4_sum);
00678         my= ff_h263_round_chroma(my4_sum);
00679         dxy = ((my & 1) << 1) | (mx & 1);
00680 
00681         offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
00682 
00683         if(s->no_rounding){
00684             s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad    , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
00685             s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
00686         }else{
00687             s->dsp.put_pixels_tab       [1][dxy](c->scratchpad    , s->last_picture.f.data[1] + offset, s->uvlinesize, 8);
00688             s->dsp.put_pixels_tab       [1][dxy](c->scratchpad + 8, s->last_picture.f.data[2] + offset, s->uvlinesize, 8);
00689         }
00690 
00691         dmin_sum += s->dsp.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);
00692         dmin_sum += s->dsp.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);
00693     }
00694 
00695     c->pred_x= mx;
00696     c->pred_y= my;
00697 
00698     switch(c->avctx->mb_cmp&0xFF){
00699     /*case FF_CMP_SSE:
00700         return dmin_sum+ 32*s->qscale*s->qscale;*/
00701     case FF_CMP_RD:
00702         return dmin_sum;
00703     default:
00704         return dmin_sum+ 11*c->mb_penalty_factor;
00705     }
00706 }
00707 
00708 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
00709     MotionEstContext * const c= &s->me;
00710 
00711     c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
00712     c->src[1][0] = c->src[0][0] + s->linesize;
00713     if(c->flags & FLAG_CHROMA){
00714         c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
00715         c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
00716         c->src[1][1] = c->src[0][1] + s->uvlinesize;
00717         c->src[1][2] = c->src[0][2] + s->uvlinesize;
00718     }
00719 }
00720 
00721 static int interlaced_search(MpegEncContext *s, int ref_index,
00722                              int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
00723 {
00724     MotionEstContext * const c= &s->me;
00725     const int size=0;
00726     const int h=8;
00727     int block;
00728     int P[10][2];
00729     uint8_t * const mv_penalty= c->current_mv_penalty;
00730     int same=1;
00731     const int stride= 2*s->linesize;
00732     int dmin_sum= 0;
00733     const int mot_stride= s->mb_stride;
00734     const int xy= s->mb_x + s->mb_y*mot_stride;
00735 
00736     c->ymin>>=1;
00737     c->ymax>>=1;
00738     c->stride<<=1;
00739     c->uvstride<<=1;
00740     init_interlaced_ref(s, ref_index);
00741 
00742     for(block=0; block<2; block++){
00743         int field_select;
00744         int best_dmin= INT_MAX;
00745         int best_field= -1;
00746 
00747         for(field_select=0; field_select<2; field_select++){
00748             int dmin, mx_i, my_i;
00749             int16_t (*mv_table)[2]= mv_tables[block][field_select];
00750 
00751             if(user_field_select){
00752                 assert(field_select==0 || field_select==1);
00753                 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
00754                 if(field_select_tables[block][xy] != field_select)
00755                     continue;
00756             }
00757 
00758             P_LEFT[0] = mv_table[xy - 1][0];
00759             P_LEFT[1] = mv_table[xy - 1][1];
00760             if(P_LEFT[0]       > (c->xmax<<1)) P_LEFT[0]       = (c->xmax<<1);
00761 
00762             c->pred_x= P_LEFT[0];
00763             c->pred_y= P_LEFT[1];
00764 
00765             if(!s->first_slice_line){
00766                 P_TOP[0]      = mv_table[xy - mot_stride][0];
00767                 P_TOP[1]      = mv_table[xy - mot_stride][1];
00768                 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
00769                 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
00770                 if(P_TOP[1]      > (c->ymax<<1)) P_TOP[1]     = (c->ymax<<1);
00771                 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
00772                 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
00773                 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
00774 
00775                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
00776                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
00777             }
00778             P_MV1[0]= mx; //FIXME not correct if block != field_select
00779             P_MV1[1]= my / 2;
00780 
00781             dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
00782 
00783             dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
00784 
00785             mv_table[xy][0]= mx_i;
00786             mv_table[xy][1]= my_i;
00787 
00788             if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
00789                 int dxy;
00790 
00791                 //FIXME chroma ME
00792                 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
00793                 dxy = ((my_i & 1) << 1) | (mx_i & 1);
00794 
00795                 if(s->no_rounding){
00796                     s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref    , stride, h);
00797                 }else{
00798                     s->dsp.put_pixels_tab       [size][dxy](c->scratchpad, ref    , stride, h);
00799                 }
00800                 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
00801                 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
00802             }else
00803                 dmin+= c->mb_penalty_factor; //field_select bits
00804 
00805             dmin += field_select != block; //slightly prefer same field
00806 
00807             if(dmin < best_dmin){
00808                 best_dmin= dmin;
00809                 best_field= field_select;
00810             }
00811         }
00812         {
00813             int16_t (*mv_table)[2]= mv_tables[block][best_field];
00814 
00815             if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
00816             if(mv_table[xy][1]&1) same=0;
00817             if(mv_table[xy][1]*2 != my) same=0;
00818             if(best_field != block) same=0;
00819         }
00820 
00821         field_select_tables[block][xy]= best_field;
00822         dmin_sum += best_dmin;
00823     }
00824 
00825     c->ymin<<=1;
00826     c->ymax<<=1;
00827     c->stride>>=1;
00828     c->uvstride>>=1;
00829 
00830     if(same)
00831         return INT_MAX;
00832 
00833     switch(c->avctx->mb_cmp&0xFF){
00834     /*case FF_CMP_SSE:
00835         return dmin_sum+ 32*s->qscale*s->qscale;*/
00836     case FF_CMP_RD:
00837         return dmin_sum;
00838     default:
00839         return dmin_sum+ 11*c->mb_penalty_factor;
00840     }
00841 }
00842 
00843 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
00844     int ymax= s->me.ymax>>interlaced;
00845     int ymin= s->me.ymin>>interlaced;
00846 
00847     if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
00848     if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
00849     if(mv[1] <       ymin) mv[1] =       ymin;
00850     if(mv[1] >       ymax) mv[1] =       ymax;
00851 }
00852 
00853 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
00854     MotionEstContext * const c= &s->me;
00855     Picture *p= s->current_picture_ptr;
00856     int mb_xy= mb_x + mb_y*s->mb_stride;
00857     int xy= 2*mb_x + 2*mb_y*s->b8_stride;
00858     int mb_type= s->current_picture.f.mb_type[mb_xy];
00859     int flags= c->flags;
00860     int shift= (flags&FLAG_QPEL) + 1;
00861     int mask= (1<<shift)-1;
00862     int x, y, i;
00863     int d=0;
00864     me_cmp_func cmpf= s->dsp.sse[0];
00865     me_cmp_func chroma_cmpf= s->dsp.sse[1];
00866 
00867     if(p_type && USES_LIST(mb_type, 1)){
00868         av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
00869         return INT_MAX/2;
00870     }
00871     assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
00872 
00873     for(i=0; i<4; i++){
00874         int xy= s->block_index[i];
00875         clip_input_mv(s, p->f.motion_val[0][xy], !!IS_INTERLACED(mb_type));
00876         clip_input_mv(s, p->f.motion_val[1][xy], !!IS_INTERLACED(mb_type));
00877     }
00878 
00879     if(IS_INTERLACED(mb_type)){
00880         int xy2= xy  + s->b8_stride;
00881         s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
00882         c->stride<<=1;
00883         c->uvstride<<=1;
00884 
00885         if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
00886             av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
00887             return INT_MAX/2;
00888         }
00889 
00890         if(USES_LIST(mb_type, 0)){
00891             int field_select0= p->f.ref_index[0][4*mb_xy  ];
00892             int field_select1= p->f.ref_index[0][4*mb_xy+2];
00893             assert(field_select0==0 ||field_select0==1);
00894             assert(field_select1==0 ||field_select1==1);
00895             init_interlaced_ref(s, 0);
00896 
00897             if(p_type){
00898                 s->p_field_select_table[0][mb_xy]= field_select0;
00899                 s->p_field_select_table[1][mb_xy]= field_select1;
00900                 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
00901                 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
00902                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
00903             }else{
00904                 s->b_field_select_table[0][0][mb_xy]= field_select0;
00905                 s->b_field_select_table[0][1][mb_xy]= field_select1;
00906                 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy ];
00907                 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[0][xy2];
00908                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
00909             }
00910 
00911             x = p->f.motion_val[0][xy ][0];
00912             y = p->f.motion_val[0][xy ][1];
00913             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
00914             x = p->f.motion_val[0][xy2][0];
00915             y = p->f.motion_val[0][xy2][1];
00916             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
00917         }
00918         if(USES_LIST(mb_type, 1)){
00919             int field_select0 = p->f.ref_index[1][4 * mb_xy    ];
00920             int field_select1 = p->f.ref_index[1][4 * mb_xy + 2];
00921             assert(field_select0==0 ||field_select0==1);
00922             assert(field_select1==0 ||field_select1==1);
00923             init_interlaced_ref(s, 2);
00924 
00925             s->b_field_select_table[1][0][mb_xy]= field_select0;
00926             s->b_field_select_table[1][1][mb_xy]= field_select1;
00927             *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy ];
00928             *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy] = *(uint32_t*)p->f.motion_val[1][xy2];
00929             if(USES_LIST(mb_type, 0)){
00930                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
00931             }else{
00932                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
00933             }
00934 
00935             x = p->f.motion_val[1][xy ][0];
00936             y = p->f.motion_val[1][xy ][1];
00937             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
00938             x = p->f.motion_val[1][xy2][0];
00939             y = p->f.motion_val[1][xy2][1];
00940             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
00941             //FIXME bidir scores
00942         }
00943         c->stride>>=1;
00944         c->uvstride>>=1;
00945     }else if(IS_8X8(mb_type)){
00946         if(!(s->flags & CODEC_FLAG_4MV)){
00947             av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
00948             return INT_MAX/2;
00949         }
00950         cmpf= s->dsp.sse[1];
00951         chroma_cmpf= s->dsp.sse[1];
00952         init_mv4_ref(c);
00953         for(i=0; i<4; i++){
00954             xy= s->block_index[i];
00955             x= p->f.motion_val[0][xy][0];
00956             y= p->f.motion_val[0][xy][1];
00957             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
00958         }
00959         s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
00960     }else{
00961         if(USES_LIST(mb_type, 0)){
00962             if(p_type){
00963                 *(uint32_t*)s->p_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
00964                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
00965             }else if(USES_LIST(mb_type, 1)){
00966                 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
00967                 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
00968                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
00969             }else{
00970                 *(uint32_t*)s->b_forw_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[0][xy];
00971                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
00972             }
00973             x = p->f.motion_val[0][xy][0];
00974             y = p->f.motion_val[0][xy][1];
00975             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
00976         }else if(USES_LIST(mb_type, 1)){
00977             *(uint32_t*)s->b_back_mv_table[mb_xy] = *(uint32_t*)p->f.motion_val[1][xy];
00978             s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
00979 
00980             x = p->f.motion_val[1][xy][0];
00981             y = p->f.motion_val[1][xy][1];
00982             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
00983         }else
00984             s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
00985     }
00986     return d;
00987 }
00988 
00989 void ff_estimate_p_frame_motion(MpegEncContext * s,
00990                                 int mb_x, int mb_y)
00991 {
00992     MotionEstContext * const c= &s->me;
00993     uint8_t *pix, *ppix;
00994     int sum, mx, my, dmin;
00995     int varc;            
00996     int vard;            
00997     int P[10][2];
00998     const int shift= 1+s->quarter_sample;
00999     int mb_type=0;
01000     Picture * const pic= &s->current_picture;
01001 
01002     init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
01003 
01004     assert(s->quarter_sample==0 || s->quarter_sample==1);
01005     assert(s->linesize == c->stride);
01006     assert(s->uvlinesize == c->uvstride);
01007 
01008     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
01009     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
01010     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
01011     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
01012 
01013     get_limits(s, 16*mb_x, 16*mb_y);
01014     c->skip=0;
01015 
01016     /* intra / predictive decision */
01017     pix = c->src[0][0];
01018     sum = s->dsp.pix_sum(pix, s->linesize);
01019     varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)sum*sum)>>8) + 500;
01020 
01021     pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
01022     pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
01023     c->mb_var_sum_temp += (varc+128)>>8;
01024 
01025     if(c->avctx->me_threshold){
01026         vard= check_input_motion(s, mb_x, mb_y, 1);
01027 
01028         if((vard+128)>>8 < c->avctx->me_threshold){
01029             int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
01030             int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
01031             pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
01032             c->mc_mb_var_sum_temp += (vard+128)>>8;
01033             c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
01034             return;
01035         }
01036         if((vard+128)>>8 < c->avctx->mb_threshold)
01037             mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
01038     }
01039 
01040     switch(s->me_method) {
01041     case ME_ZERO:
01042     default:
01043         no_motion_search(s, &mx, &my);
01044         mx-= mb_x*16;
01045         my-= mb_y*16;
01046         dmin = 0;
01047         break;
01048     case ME_X1:
01049     case ME_EPZS:
01050        {
01051             const int mot_stride = s->b8_stride;
01052             const int mot_xy = s->block_index[0];
01053 
01054             P_LEFT[0] = s->current_picture.f.motion_val[0][mot_xy - 1][0];
01055             P_LEFT[1] = s->current_picture.f.motion_val[0][mot_xy - 1][1];
01056 
01057             if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
01058 
01059             if(!s->first_slice_line) {
01060                 P_TOP[0]      = s->current_picture.f.motion_val[0][mot_xy - mot_stride    ][0];
01061                 P_TOP[1]      = s->current_picture.f.motion_val[0][mot_xy - mot_stride    ][1];
01062                 P_TOPRIGHT[0] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][0];
01063                 P_TOPRIGHT[1] = s->current_picture.f.motion_val[0][mot_xy - mot_stride + 2][1];
01064                 if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
01065                 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
01066                 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
01067 
01068                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
01069                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
01070 
01071                 if(s->out_format == FMT_H263){
01072                     c->pred_x = P_MEDIAN[0];
01073                     c->pred_y = P_MEDIAN[1];
01074                 }else { /* mpeg1 at least */
01075                     c->pred_x= P_LEFT[0];
01076                     c->pred_y= P_LEFT[1];
01077                 }
01078             }else{
01079                 c->pred_x= P_LEFT[0];
01080                 c->pred_y= P_LEFT[1];
01081             }
01082 
01083         }
01084         dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
01085 
01086         break;
01087     }
01088 
01089     /* At this point (mx,my) are full-pell and the relative displacement */
01090     ppix = c->ref[0][0] + (my * s->linesize) + mx;
01091 
01092     vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
01093 
01094     pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
01095 //    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
01096     c->mc_mb_var_sum_temp += (vard+128)>>8;
01097 
01098     if(mb_type){
01099         int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
01100         int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
01101         c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
01102 
01103         if(mb_type == CANDIDATE_MB_TYPE_INTER){
01104             c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01105             set_p_mv_tables(s, mx, my, 1);
01106         }else{
01107             mx <<=shift;
01108             my <<=shift;
01109         }
01110         if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
01111             h263_mv4_search(s, mx, my, shift);
01112 
01113             set_p_mv_tables(s, mx, my, 0);
01114         }
01115         if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
01116             interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
01117         }
01118     }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
01119         int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
01120         int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
01121         c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
01122 
01123         if (vard*2 + 200*256 > varc)
01124             mb_type|= CANDIDATE_MB_TYPE_INTRA;
01125         if (varc*2 + 200*256 > vard || s->qscale > 24){
01126 //        if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
01127             mb_type|= CANDIDATE_MB_TYPE_INTER;
01128             c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01129             if(s->flags&CODEC_FLAG_MV0)
01130                 if(mx || my)
01131                     mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
01132         }else{
01133             mx <<=shift;
01134             my <<=shift;
01135         }
01136         if((s->flags&CODEC_FLAG_4MV)
01137            && !c->skip && varc>50<<8 && vard>10<<8){
01138             if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
01139                 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
01140 
01141             set_p_mv_tables(s, mx, my, 0);
01142         }else
01143             set_p_mv_tables(s, mx, my, 1);
01144         if((s->flags&CODEC_FLAG_INTERLACED_ME)
01145            && !c->skip){ //FIXME varc/d checks
01146             if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
01147                 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
01148         }
01149     }else{
01150         int intra_score, i;
01151         mb_type= CANDIDATE_MB_TYPE_INTER;
01152 
01153         dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01154         if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
01155             dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
01156 
01157         if((s->flags&CODEC_FLAG_4MV)
01158            && !c->skip && varc>50<<8 && vard>10<<8){
01159             int dmin4= h263_mv4_search(s, mx, my, shift);
01160             if(dmin4 < dmin){
01161                 mb_type= CANDIDATE_MB_TYPE_INTER4V;
01162                 dmin=dmin4;
01163             }
01164         }
01165         if((s->flags&CODEC_FLAG_INTERLACED_ME)
01166            && !c->skip){ //FIXME varc/d checks
01167             int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
01168             if(dmin_i < dmin){
01169                 mb_type = CANDIDATE_MB_TYPE_INTER_I;
01170                 dmin= dmin_i;
01171             }
01172         }
01173 
01174 //        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
01175         set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
01176 
01177         /* get intra luma score */
01178         if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
01179             intra_score= varc - 500;
01180         }else{
01181             unsigned mean = (sum+128)>>8;
01182             mean*= 0x01010101;
01183 
01184             for(i=0; i<16; i++){
01185                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
01186                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
01187                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
01188                 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
01189             }
01190 
01191             intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
01192         }
01193         intra_score += c->mb_penalty_factor*16;
01194 
01195         if(intra_score < dmin){
01196             mb_type= CANDIDATE_MB_TYPE_INTRA;
01197             s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
01198         }else
01199             s->current_picture.f.mb_type[mb_y*s->mb_stride + mb_x] = 0;
01200 
01201         {
01202             int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
01203             int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
01204             c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
01205         }
01206     }
01207 
01208     s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
01209 }
01210 
01211 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
01212                                     int mb_x, int mb_y)
01213 {
01214     MotionEstContext * const c= &s->me;
01215     int mx, my, dmin;
01216     int P[10][2];
01217     const int shift= 1+s->quarter_sample;
01218     const int xy= mb_x + mb_y*s->mb_stride;
01219     init_ref(c, s->new_picture.f.data, s->last_picture.f.data, NULL, 16*mb_x, 16*mb_y, 0);
01220 
01221     assert(s->quarter_sample==0 || s->quarter_sample==1);
01222 
01223     c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
01224     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
01225 
01226     get_limits(s, 16*mb_x, 16*mb_y);
01227     c->skip=0;
01228 
01229     P_LEFT[0]       = s->p_mv_table[xy + 1][0];
01230     P_LEFT[1]       = s->p_mv_table[xy + 1][1];
01231 
01232     if(P_LEFT[0]       < (c->xmin<<shift)) P_LEFT[0]       = (c->xmin<<shift);
01233 
01234     /* special case for first line */
01235     if (s->first_slice_line) {
01236         c->pred_x= P_LEFT[0];
01237         c->pred_y= P_LEFT[1];
01238         P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
01239         P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
01240     } else {
01241         P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
01242         P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
01243         P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
01244         P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
01245         if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);
01246         if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
01247         if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
01248 
01249         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
01250         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
01251 
01252         c->pred_x = P_MEDIAN[0];
01253         c->pred_y = P_MEDIAN[1];
01254     }
01255 
01256     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
01257 
01258     s->p_mv_table[xy][0] = mx<<shift;
01259     s->p_mv_table[xy][1] = my<<shift;
01260 
01261     return dmin;
01262 }
01263 
01264 static int ff_estimate_motion_b(MpegEncContext * s,
01265                        int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
01266 {
01267     MotionEstContext * const c= &s->me;
01268     int mx, my, dmin;
01269     int P[10][2];
01270     const int shift= 1+s->quarter_sample;
01271     const int mot_stride = s->mb_stride;
01272     const int mot_xy = mb_y*mot_stride + mb_x;
01273     uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
01274     int mv_scale;
01275 
01276     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
01277     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
01278     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
01279     c->current_mv_penalty= mv_penalty;
01280 
01281     get_limits(s, 16*mb_x, 16*mb_y);
01282 
01283     switch(s->me_method) {
01284     case ME_ZERO:
01285     default:
01286         no_motion_search(s, &mx, &my);
01287         dmin = 0;
01288         mx-= mb_x*16;
01289         my-= mb_y*16;
01290         break;
01291     case ME_X1:
01292     case ME_EPZS:
01293        {
01294             P_LEFT[0]        = mv_table[mot_xy - 1][0];
01295             P_LEFT[1]        = mv_table[mot_xy - 1][1];
01296 
01297             if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
01298 
01299             /* special case for first line */
01300             if (!s->first_slice_line) {
01301                 P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
01302                 P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
01303                 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
01304                 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
01305                 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
01306                 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
01307                 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
01308 
01309                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
01310                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
01311             }
01312             c->pred_x= P_LEFT[0];
01313             c->pred_y= P_LEFT[1];
01314         }
01315 
01316         if(mv_table == s->b_forw_mv_table){
01317             mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
01318         }else{
01319             mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
01320         }
01321 
01322         dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
01323 
01324         break;
01325     }
01326 
01327     dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
01328 
01329     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
01330         dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
01331 
01332 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
01333 //    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
01334     mv_table[mot_xy][0]= mx;
01335     mv_table[mot_xy][1]= my;
01336 
01337     return dmin;
01338 }
01339 
01340 static inline int check_bidir_mv(MpegEncContext * s,
01341                    int motion_fx, int motion_fy,
01342                    int motion_bx, int motion_by,
01343                    int pred_fx, int pred_fy,
01344                    int pred_bx, int pred_by,
01345                    int size, int h)
01346 {
01347     //FIXME optimize?
01348     //FIXME better f_code prediction (max mv & distance)
01349     //FIXME pointers
01350     MotionEstContext * const c= &s->me;
01351     uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
01352     uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
01353     int stride= c->stride;
01354     uint8_t *dest_y = c->scratchpad;
01355     uint8_t *ptr;
01356     int dxy;
01357     int src_x, src_y;
01358     int fbmin;
01359     uint8_t **src_data= c->src[0];
01360     uint8_t **ref_data= c->ref[0];
01361     uint8_t **ref2_data= c->ref[2];
01362 
01363     if(s->quarter_sample){
01364         dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
01365         src_x = motion_fx >> 2;
01366         src_y = motion_fy >> 2;
01367 
01368         ptr = ref_data[0] + (src_y * stride) + src_x;
01369         s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , stride);
01370 
01371         dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
01372         src_x = motion_bx >> 2;
01373         src_y = motion_by >> 2;
01374 
01375         ptr = ref2_data[0] + (src_y * stride) + src_x;
01376         s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);
01377     }else{
01378         dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
01379         src_x = motion_fx >> 1;
01380         src_y = motion_fy >> 1;
01381 
01382         ptr = ref_data[0] + (src_y * stride) + src_x;
01383         s->dsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
01384 
01385         dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
01386         src_x = motion_bx >> 1;
01387         src_y = motion_by >> 1;
01388 
01389         ptr = ref2_data[0] + (src_y * stride) + src_x;
01390         s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
01391     }
01392 
01393     fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
01394            +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
01395            + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
01396 
01397     if(c->avctx->mb_cmp&FF_CMP_CHROMA){
01398     }
01399     //FIXME CHROMA !!!
01400 
01401     return fbmin;
01402 }
01403 
01404 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
01405 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
01406 {
01407     MotionEstContext * const c= &s->me;
01408     const int mot_stride = s->mb_stride;
01409     const int xy = mb_y *mot_stride + mb_x;
01410     int fbmin;
01411     int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
01412     int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
01413     int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
01414     int pred_by= s->b_bidir_back_mv_table[xy-1][1];
01415     int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
01416     int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
01417     int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
01418     int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
01419     const int flags= c->sub_flags;
01420     const int qpel= flags&FLAG_QPEL;
01421     const int shift= 1+qpel;
01422     const int xmin= c->xmin<<shift;
01423     const int ymin= c->ymin<<shift;
01424     const int xmax= c->xmax<<shift;
01425     const int ymax= c->ymax<<shift;
01426 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
01427 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
01428     int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
01429     uint8_t map[256];
01430 
01431     memset(map,0,sizeof(map));
01432     map[hashidx&255] = 1;
01433 
01434     fbmin= check_bidir_mv(s, motion_fx, motion_fy,
01435                           motion_bx, motion_by,
01436                           pred_fx, pred_fy,
01437                           pred_bx, pred_by,
01438                           0, 16);
01439 
01440     if(s->avctx->bidir_refine){
01441         int end;
01442         static const uint8_t limittab[5]={0,8,32,64,80};
01443         const int limit= limittab[s->avctx->bidir_refine];
01444         static const int8_t vect[][4]={
01445 { 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},
01446 
01447 { 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},
01448 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
01449 { 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},
01450 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
01451 
01452 { 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},
01453 { 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},
01454 { 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},
01455 { 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},
01456 
01457 { 1, 1, 1, 1}, {-1,-1,-1,-1},
01458 { 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},
01459 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
01460         };
01461         static const uint8_t hash[]={
01462 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),
01463 
01464 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),
01465 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
01466 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),
01467 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
01468 
01469 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),
01470 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),
01471 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),
01472 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),
01473 
01474 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
01475 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),
01476 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),
01477 };
01478 
01479 #define CHECK_BIDIR(fx,fy,bx,by)\
01480     if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
01481        &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
01482        &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
01483         int score;\
01484         map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
01485         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);\
01486         if(score < fbmin){\
01487             hashidx += HASH(fx,fy,bx,by);\
01488             fbmin= score;\
01489             motion_fx+=fx;\
01490             motion_fy+=fy;\
01491             motion_bx+=bx;\
01492             motion_by+=by;\
01493             end=0;\
01494         }\
01495     }
01496 #define CHECK_BIDIR2(a,b,c,d)\
01497 CHECK_BIDIR(a,b,c,d)\
01498 CHECK_BIDIR(-(a),-(b),-(c),-(d))
01499 
01500         do{
01501             int i;
01502             int borderdist=0;
01503             end=1;
01504 
01505             CHECK_BIDIR2(0,0,0,1)
01506             CHECK_BIDIR2(0,0,1,0)
01507             CHECK_BIDIR2(0,1,0,0)
01508             CHECK_BIDIR2(1,0,0,0)
01509 
01510             for(i=8; i<limit; i++){
01511                 int fx= motion_fx+vect[i][0];
01512                 int fy= motion_fy+vect[i][1];
01513                 int bx= motion_bx+vect[i][2];
01514                 int by= motion_by+vect[i][3];
01515                 if(borderdist<=0){
01516                     int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
01517                     int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
01518                     if((a|b) < 0)
01519                         map[(hashidx+hash[i])&255] = 1;
01520                 }
01521                 if(!map[(hashidx+hash[i])&255]){
01522                     int score;
01523                     map[(hashidx+hash[i])&255] = 1;
01524                     score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
01525                     if(score < fbmin){
01526                         hashidx += hash[i];
01527                         fbmin= score;
01528                         motion_fx=fx;
01529                         motion_fy=fy;
01530                         motion_bx=bx;
01531                         motion_by=by;
01532                         end=0;
01533                         borderdist--;
01534                         if(borderdist<=0){
01535                             int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
01536                             int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
01537                             borderdist= FFMIN(a,b);
01538                         }
01539                     }
01540                 }
01541             }
01542         }while(!end);
01543     }
01544 
01545     s->b_bidir_forw_mv_table[xy][0]= motion_fx;
01546     s->b_bidir_forw_mv_table[xy][1]= motion_fy;
01547     s->b_bidir_back_mv_table[xy][0]= motion_bx;
01548     s->b_bidir_back_mv_table[xy][1]= motion_by;
01549 
01550     return fbmin;
01551 }
01552 
01553 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
01554 {
01555     MotionEstContext * const c= &s->me;
01556     int P[10][2];
01557     const int mot_stride = s->mb_stride;
01558     const int mot_xy = mb_y*mot_stride + mb_x;
01559     const int shift= 1+s->quarter_sample;
01560     int dmin, i;
01561     const int time_pp= s->pp_time;
01562     const int time_pb= s->pb_time;
01563     int mx, my, xmin, xmax, ymin, ymax;
01564     int16_t (*mv_table)[2]= s->b_direct_mv_table;
01565 
01566     c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
01567     ymin= xmin=(-32)>>shift;
01568     ymax= xmax=   31>>shift;
01569 
01570     if (IS_8X8(s->next_picture.f.mb_type[mot_xy])) {
01571         s->mv_type= MV_TYPE_8X8;
01572     }else{
01573         s->mv_type= MV_TYPE_16X16;
01574     }
01575 
01576     for(i=0; i<4; i++){
01577         int index= s->block_index[i];
01578         int min, max;
01579 
01580         c->co_located_mv[i][0] = s->next_picture.f.motion_val[0][index][0];
01581         c->co_located_mv[i][1] = s->next_picture.f.motion_val[0][index][1];
01582         c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
01583         c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
01584 //        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
01585 //        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
01586 
01587         max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
01588         min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
01589         max+= 16*mb_x + 1; // +-1 is for the simpler rounding
01590         min+= 16*mb_x - 1;
01591         xmax= FFMIN(xmax, s->width - max);
01592         xmin= FFMAX(xmin, - 16     - min);
01593 
01594         max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
01595         min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
01596         max+= 16*mb_y + 1; // +-1 is for the simpler rounding
01597         min+= 16*mb_y - 1;
01598         ymax= FFMIN(ymax, s->height - max);
01599         ymin= FFMAX(ymin, - 16      - min);
01600 
01601         if(s->mv_type == MV_TYPE_16X16) break;
01602     }
01603 
01604     assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
01605 
01606     if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
01607         s->b_direct_mv_table[mot_xy][0]= 0;
01608         s->b_direct_mv_table[mot_xy][1]= 0;
01609 
01610         return 256*256*256*64;
01611     }
01612 
01613     c->xmin= xmin;
01614     c->ymin= ymin;
01615     c->xmax= xmax;
01616     c->ymax= ymax;
01617     c->flags     |= FLAG_DIRECT;
01618     c->sub_flags |= FLAG_DIRECT;
01619     c->pred_x=0;
01620     c->pred_y=0;
01621 
01622     P_LEFT[0]        = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
01623     P_LEFT[1]        = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
01624 
01625     /* special case for first line */
01626     if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
01627         P_TOP[0]      = av_clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
01628         P_TOP[1]      = av_clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
01629         P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
01630         P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
01631 
01632         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
01633         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
01634     }
01635 
01636     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
01637     if(c->sub_flags&FLAG_QPEL)
01638         dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01639     else
01640         dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01641 
01642     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
01643         dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
01644 
01645     get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
01646 
01647     mv_table[mot_xy][0]= mx;
01648     mv_table[mot_xy][1]= my;
01649     c->flags     &= ~FLAG_DIRECT;
01650     c->sub_flags &= ~FLAG_DIRECT;
01651 
01652     return dmin;
01653 }
01654 
01655 void ff_estimate_b_frame_motion(MpegEncContext * s,
01656                              int mb_x, int mb_y)
01657 {
01658     MotionEstContext * const c= &s->me;
01659     const int penalty_factor= c->mb_penalty_factor;
01660     int fmin, bmin, dmin, fbmin, bimin, fimin;
01661     int type=0;
01662     const int xy = mb_y*s->mb_stride + mb_x;
01663     init_ref(c, s->new_picture.f.data, s->last_picture.f.data,
01664              s->next_picture.f.data, 16 * mb_x, 16 * mb_y, 2);
01665 
01666     get_limits(s, 16*mb_x, 16*mb_y);
01667 
01668     c->skip=0;
01669 
01670     if (s->codec_id == CODEC_ID_MPEG4 && s->next_picture.f.mbskip_table[xy]) {
01671         int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
01672 
01673         score= ((unsigned)(score*score + 128*256))>>16;
01674         c->mc_mb_var_sum_temp += score;
01675         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
01676         s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
01677 
01678         return;
01679     }
01680 
01681     if(c->avctx->me_threshold){
01682         int vard= check_input_motion(s, mb_x, mb_y, 0);
01683 
01684         if((vard+128)>>8 < c->avctx->me_threshold){
01685 //            pix = c->src[0][0];
01686 //            sum = s->dsp.pix_sum(pix, s->linesize);
01687 //            varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
01688 
01689 //            pic->mb_var   [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
01690              s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
01691 /*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
01692             c->mb_var_sum_temp    += (varc+128)>>8;*/
01693             c->mc_mb_var_sum_temp += (vard+128)>>8;
01694 /*            if (vard <= 64<<8 || vard < varc) {
01695                 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
01696             }else{
01697                 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
01698             }*/
01699             return;
01700         }
01701         if((vard+128)>>8 < c->avctx->mb_threshold){
01702             type= s->mb_type[mb_y*s->mb_stride + mb_x];
01703             if(type == CANDIDATE_MB_TYPE_DIRECT){
01704                 direct_search(s, mb_x, mb_y);
01705             }
01706             if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
01707                 c->skip=0;
01708                 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
01709             }
01710             if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
01711                 c->skip=0;
01712                 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
01713             }
01714             if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
01715                 c->skip=0;
01716                 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
01717                 interlaced_search(s, 0,
01718                                         s->b_field_mv_table[0], s->b_field_select_table[0],
01719                                         s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
01720             }
01721             if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
01722                 c->skip=0;
01723                 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
01724                 interlaced_search(s, 2,
01725                                         s->b_field_mv_table[1], s->b_field_select_table[1],
01726                                         s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
01727             }
01728             return;
01729         }
01730     }
01731 
01732     if (s->codec_id == CODEC_ID_MPEG4)
01733         dmin= direct_search(s, mb_x, mb_y);
01734     else
01735         dmin= INT_MAX;
01736 //FIXME penalty stuff for non mpeg4
01737     c->skip=0;
01738     fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
01739 
01740     c->skip=0;
01741     bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
01742 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
01743 
01744     c->skip=0;
01745     fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
01746 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
01747 
01748     if(s->flags & CODEC_FLAG_INTERLACED_ME){
01749 //FIXME mb type penalty
01750         c->skip=0;
01751         c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
01752         fimin= interlaced_search(s, 0,
01753                                  s->b_field_mv_table[0], s->b_field_select_table[0],
01754                                  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
01755         c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
01756         bimin= interlaced_search(s, 2,
01757                                  s->b_field_mv_table[1], s->b_field_select_table[1],
01758                                  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
01759     }else
01760         fimin= bimin= INT_MAX;
01761 
01762     {
01763         int score= fmin;
01764         type = CANDIDATE_MB_TYPE_FORWARD;
01765 
01766         if (dmin <= score){
01767             score = dmin;
01768             type = CANDIDATE_MB_TYPE_DIRECT;
01769         }
01770         if(bmin<score){
01771             score=bmin;
01772             type= CANDIDATE_MB_TYPE_BACKWARD;
01773         }
01774         if(fbmin<score){
01775             score=fbmin;
01776             type= CANDIDATE_MB_TYPE_BIDIR;
01777         }
01778         if(fimin<score){
01779             score=fimin;
01780             type= CANDIDATE_MB_TYPE_FORWARD_I;
01781         }
01782         if(bimin<score){
01783             score=bimin;
01784             type= CANDIDATE_MB_TYPE_BACKWARD_I;
01785         }
01786 
01787         score= ((unsigned)(score*score + 128*256))>>16;
01788         c->mc_mb_var_sum_temp += score;
01789         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
01790     }
01791 
01792     if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
01793         type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
01794         if(fimin < INT_MAX)
01795             type |= CANDIDATE_MB_TYPE_FORWARD_I;
01796         if(bimin < INT_MAX)
01797             type |= CANDIDATE_MB_TYPE_BACKWARD_I;
01798         if(fimin < INT_MAX && bimin < INT_MAX){
01799             type |= CANDIDATE_MB_TYPE_BIDIR_I;
01800         }
01801          //FIXME something smarter
01802         if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
01803         if(s->codec_id == CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
01804             type |= CANDIDATE_MB_TYPE_DIRECT0;
01805     }
01806 
01807     s->mb_type[mb_y*s->mb_stride + mb_x]= type;
01808 }
01809 
01810 /* find best f_code for ME which do unlimited searches */
01811 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
01812 {
01813     if(s->me_method>=ME_EPZS){
01814         int score[8];
01815         int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
01816         uint8_t * fcode_tab= s->fcode_tab;
01817         int best_fcode=-1;
01818         int best_score=-10000000;
01819 
01820         if(s->msmpeg4_version)
01821             range= FFMIN(range, 16);
01822         else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
01823             range= FFMIN(range, 256);
01824 
01825         for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
01826 
01827         for(y=0; y<s->mb_height; y++){
01828             int x;
01829             int xy= y*s->mb_stride;
01830             for(x=0; x<s->mb_width; x++){
01831                 if(s->mb_type[xy] & type){
01832                     int mx= mv_table[xy][0];
01833                     int my= mv_table[xy][1];
01834                     int fcode= FFMAX(fcode_tab[mx + MAX_MV],
01835                                      fcode_tab[my + MAX_MV]);
01836                     int j;
01837 
01838                         if(mx >= range || mx < -range ||
01839                            my >= range || my < -range)
01840                             continue;
01841 
01842                     for(j=0; j<fcode && j<8; j++){
01843                         if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
01844                             score[j]-= 170;
01845                     }
01846                 }
01847                 xy++;
01848             }
01849         }
01850 
01851         for(i=1; i<8; i++){
01852             if(score[i] > best_score){
01853                 best_score= score[i];
01854                 best_fcode= i;
01855             }
01856 //            printf("%d %d\n", i, score[i]);
01857         }
01858 
01859 //    printf("fcode: %d type: %d\n", i, s->pict_type);
01860         return best_fcode;
01861 /*        for(i=0; i<=MAX_FCODE; i++){
01862             printf("%d ", mv_num[i]);
01863         }
01864         printf("\n");*/
01865     }else{
01866         return 1;
01867     }
01868 }
01869 
01870 void ff_fix_long_p_mvs(MpegEncContext * s)
01871 {
01872     MotionEstContext * const c= &s->me;
01873     const int f_code= s->f_code;
01874     int y, range;
01875     assert(s->pict_type==AV_PICTURE_TYPE_P);
01876 
01877     range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
01878 
01879     assert(range <= 16 || !s->msmpeg4_version);
01880     assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
01881 
01882     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
01883 
01884 //printf("%d no:%d %d//\n", clip, noclip, f_code);
01885     if(s->flags&CODEC_FLAG_4MV){
01886         const int wrap= s->b8_stride;
01887 
01888         /* clip / convert to intra 8x8 type MVs */
01889         for(y=0; y<s->mb_height; y++){
01890             int xy= y*2*wrap;
01891             int i= y*s->mb_stride;
01892             int x;
01893 
01894             for(x=0; x<s->mb_width; x++){
01895                 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
01896                     int block;
01897                     for(block=0; block<4; block++){
01898                         int off= (block& 1) + (block>>1)*wrap;
01899                         int mx = s->current_picture.f.motion_val[0][ xy + off ][0];
01900                         int my = s->current_picture.f.motion_val[0][ xy + off ][1];
01901 
01902                         if(   mx >=range || mx <-range
01903                            || my >=range || my <-range){
01904                             s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
01905                             s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
01906                             s->current_picture.f.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;
01907                         }
01908                     }
01909                 }
01910                 xy+=2;
01911                 i++;
01912             }
01913         }
01914     }
01915 }
01916 
01921 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
01922                      int16_t (*mv_table)[2], int f_code, int type, int truncate)
01923 {
01924     MotionEstContext * const c= &s->me;
01925     int y, h_range, v_range;
01926 
01927     // RAL: 8 in MPEG-1, 16 in MPEG-4
01928     int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
01929 
01930     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
01931 
01932     h_range= range;
01933     v_range= field_select_table ? range>>1 : range;
01934 
01935     /* clip / convert to intra 16x16 type MVs */
01936     for(y=0; y<s->mb_height; y++){
01937         int x;
01938         int xy= y*s->mb_stride;
01939         for(x=0; x<s->mb_width; x++){
01940             if (s->mb_type[xy] & type){    // RAL: "type" test added...
01941                 if(field_select_table==NULL || field_select_table[xy] == field_select){
01942                     if(   mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
01943                        || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
01944 
01945                         if(truncate){
01946                             if     (mv_table[xy][0] > h_range-1) mv_table[xy][0]=  h_range-1;
01947                             else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
01948                             if     (mv_table[xy][1] > v_range-1) mv_table[xy][1]=  v_range-1;
01949                             else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
01950                         }else{
01951                             s->mb_type[xy] &= ~type;
01952                             s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
01953                             mv_table[xy][0]=
01954                             mv_table[xy][1]= 0;
01955                         }
01956                     }
01957                 }
01958             }
01959             xy++;
01960         }
01961     }
01962 }
Generated on Sun Apr 22 2012 21:54:02 for Libav by doxygen 1.7.1