Libav 0.7.1
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 int 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 #if 0
00378 static int pix_dev(uint8_t * pix, int line_size, int mean)
00379 {
00380     int s, i, j;
00381 
00382     s = 0;
00383     for (i = 0; i < 16; i++) {
00384         for (j = 0; j < 16; j += 8) {
00385             s += FFABS(pix[0]-mean);
00386             s += FFABS(pix[1]-mean);
00387             s += FFABS(pix[2]-mean);
00388             s += FFABS(pix[3]-mean);
00389             s += FFABS(pix[4]-mean);
00390             s += FFABS(pix[5]-mean);
00391             s += FFABS(pix[6]-mean);
00392             s += FFABS(pix[7]-mean);
00393             pix += 8;
00394         }
00395         pix += line_size - 16;
00396     }
00397     return s;
00398 }
00399 #endif
00400 
00401 static inline void no_motion_search(MpegEncContext * s,
00402                                     int *mx_ptr, int *my_ptr)
00403 {
00404     *mx_ptr = 16 * s->mb_x;
00405     *my_ptr = 16 * s->mb_y;
00406 }
00407 
00408 #define Z_THRESHOLD 256
00409 
00410 #define CHECK_SAD_HALF_MV(suffix, x, y) \
00411 {\
00412     d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
00413     d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
00414     COPY3_IF_LT(dminh, d, dx, x, dy, y)\
00415 }
00416 
00417 static inline int sad_hpel_motion_search(MpegEncContext * s,
00418                                   int *mx_ptr, int *my_ptr, int dmin,
00419                                   int src_index, int ref_index,
00420                                   int size, int h)
00421 {
00422     MotionEstContext * const c= &s->me;
00423     const int penalty_factor= c->sub_penalty_factor;
00424     int mx, my, dminh;
00425     uint8_t *pix, *ptr;
00426     int stride= c->stride;
00427     const int flags= c->sub_flags;
00428     LOAD_COMMON
00429 
00430     assert(flags == 0);
00431 
00432     if(c->skip){
00433 //    printf("S");
00434         *mx_ptr = 0;
00435         *my_ptr = 0;
00436         return dmin;
00437     }
00438 //    printf("N");
00439 
00440     pix = c->src[src_index][0];
00441 
00442     mx = *mx_ptr;
00443     my = *my_ptr;
00444     ptr = c->ref[ref_index][0] + (my * stride) + mx;
00445 
00446     dminh = dmin;
00447 
00448     if (mx > xmin && mx < xmax &&
00449         my > ymin && my < ymax) {
00450         int dx=0, dy=0;
00451         int d, pen_x, pen_y;
00452         const int index= (my<<ME_MAP_SHIFT) + mx;
00453         const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
00454         const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
00455         const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
00456         const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
00457         mx<<=1;
00458         my<<=1;
00459 
00460 
00461         pen_x= pred_x + mx;
00462         pen_y= pred_y + my;
00463 
00464         ptr-= stride;
00465         if(t<=b){
00466             CHECK_SAD_HALF_MV(y2 , 0, -1)
00467             if(l<=r){
00468                 CHECK_SAD_HALF_MV(xy2, -1, -1)
00469                 if(t+r<=b+l){
00470                     CHECK_SAD_HALF_MV(xy2, +1, -1)
00471                     ptr+= stride;
00472                 }else{
00473                     ptr+= stride;
00474                     CHECK_SAD_HALF_MV(xy2, -1, +1)
00475                 }
00476                 CHECK_SAD_HALF_MV(x2 , -1,  0)
00477             }else{
00478                 CHECK_SAD_HALF_MV(xy2, +1, -1)
00479                 if(t+l<=b+r){
00480                     CHECK_SAD_HALF_MV(xy2, -1, -1)
00481                     ptr+= stride;
00482                 }else{
00483                     ptr+= stride;
00484                     CHECK_SAD_HALF_MV(xy2, +1, +1)
00485                 }
00486                 CHECK_SAD_HALF_MV(x2 , +1,  0)
00487             }
00488         }else{
00489             if(l<=r){
00490                 if(t+l<=b+r){
00491                     CHECK_SAD_HALF_MV(xy2, -1, -1)
00492                     ptr+= stride;
00493                 }else{
00494                     ptr+= stride;
00495                     CHECK_SAD_HALF_MV(xy2, +1, +1)
00496                 }
00497                 CHECK_SAD_HALF_MV(x2 , -1,  0)
00498                 CHECK_SAD_HALF_MV(xy2, -1, +1)
00499             }else{
00500                 if(t+r<=b+l){
00501                     CHECK_SAD_HALF_MV(xy2, +1, -1)
00502                     ptr+= stride;
00503                 }else{
00504                     ptr+= stride;
00505                     CHECK_SAD_HALF_MV(xy2, -1, +1)
00506                 }
00507                 CHECK_SAD_HALF_MV(x2 , +1,  0)
00508                 CHECK_SAD_HALF_MV(xy2, +1, +1)
00509             }
00510             CHECK_SAD_HALF_MV(y2 ,  0, +1)
00511         }
00512         mx+=dx;
00513         my+=dy;
00514 
00515     }else{
00516         mx<<=1;
00517         my<<=1;
00518     }
00519 
00520     *mx_ptr = mx;
00521     *my_ptr = my;
00522     return dminh;
00523 }
00524 
00525 static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
00526 {
00527     const int xy= s->mb_x + s->mb_y*s->mb_stride;
00528 
00529     s->p_mv_table[xy][0] = mx;
00530     s->p_mv_table[xy][1] = my;
00531 
00532     /* has already been set to the 4 MV if 4MV is done */
00533     if(mv4){
00534         int mot_xy= s->block_index[0];
00535 
00536         s->current_picture.motion_val[0][mot_xy  ][0]= mx;
00537         s->current_picture.motion_val[0][mot_xy  ][1]= my;
00538         s->current_picture.motion_val[0][mot_xy+1][0]= mx;
00539         s->current_picture.motion_val[0][mot_xy+1][1]= my;
00540 
00541         mot_xy += s->b8_stride;
00542         s->current_picture.motion_val[0][mot_xy  ][0]= mx;
00543         s->current_picture.motion_val[0][mot_xy  ][1]= my;
00544         s->current_picture.motion_val[0][mot_xy+1][0]= mx;
00545         s->current_picture.motion_val[0][mot_xy+1][1]= my;
00546     }
00547 }
00548 
00552 static inline void get_limits(MpegEncContext *s, int x, int y)
00553 {
00554     MotionEstContext * const c= &s->me;
00555     int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
00556 /*
00557     if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
00558     else                   c->range= 16;
00559 */
00560     if (s->unrestricted_mv) {
00561         c->xmin = - x - 16;
00562         c->ymin = - y - 16;
00563         c->xmax = - x + s->mb_width *16;
00564         c->ymax = - y + s->mb_height*16;
00565     } else if (s->out_format == FMT_H261){
00566         // Search range of H261 is different from other codec standards
00567         c->xmin = (x > 15) ? - 15 : 0;
00568         c->ymin = (y > 15) ? - 15 : 0;
00569         c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
00570         c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
00571     } else {
00572         c->xmin = - x;
00573         c->ymin = - y;
00574         c->xmax = - x + s->mb_width *16 - 16;
00575         c->ymax = - y + s->mb_height*16 - 16;
00576     }
00577     if(range){
00578         c->xmin = FFMAX(c->xmin,-range);
00579         c->xmax = FFMIN(c->xmax, range);
00580         c->ymin = FFMAX(c->ymin,-range);
00581         c->ymax = FFMIN(c->ymax, range);
00582     }
00583 }
00584 
00585 static inline void init_mv4_ref(MotionEstContext *c){
00586     const int stride= c->stride;
00587 
00588     c->ref[1][0] = c->ref[0][0] + 8;
00589     c->ref[2][0] = c->ref[0][0] + 8*stride;
00590     c->ref[3][0] = c->ref[2][0] + 8;
00591     c->src[1][0] = c->src[0][0] + 8;
00592     c->src[2][0] = c->src[0][0] + 8*stride;
00593     c->src[3][0] = c->src[2][0] + 8;
00594 }
00595 
00596 static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
00597 {
00598     MotionEstContext * const c= &s->me;
00599     const int size= 1;
00600     const int h=8;
00601     int block;
00602     int P[10][2];
00603     int dmin_sum=0, mx4_sum=0, my4_sum=0;
00604     int same=1;
00605     const int stride= c->stride;
00606     uint8_t *mv_penalty= c->current_mv_penalty;
00607 
00608     init_mv4_ref(c);
00609 
00610     for(block=0; block<4; block++){
00611         int mx4, my4;
00612         int pred_x4, pred_y4;
00613         int dmin4;
00614         static const int off[4]= {2, 1, 1, -1};
00615         const int mot_stride = s->b8_stride;
00616         const int mot_xy = s->block_index[block];
00617 
00618         P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
00619         P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
00620 
00621         if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
00622 
00623         /* special case for first line */
00624         if (s->first_slice_line && block<2) {
00625             c->pred_x= pred_x4= P_LEFT[0];
00626             c->pred_y= pred_y4= P_LEFT[1];
00627         } else {
00628             P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][0];
00629             P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][1];
00630             P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
00631             P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
00632             if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
00633             if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
00634             if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
00635             if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
00636 
00637             P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
00638             P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
00639 
00640             c->pred_x= pred_x4 = P_MEDIAN[0];
00641             c->pred_y= pred_y4 = P_MEDIAN[1];
00642         }
00643         P_MV1[0]= mx;
00644         P_MV1[1]= my;
00645 
00646         dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
00647 
00648         dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
00649 
00650         if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
00651             int dxy;
00652             const int offset= ((block&1) + (block>>1)*stride)*8;
00653             uint8_t *dest_y = c->scratchpad + offset;
00654             if(s->quarter_sample){
00655                 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
00656                 dxy = ((my4 & 3) << 2) | (mx4 & 3);
00657 
00658                 if(s->no_rounding)
00659                     s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , stride);
00660                 else
00661                     s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , stride);
00662             }else{
00663                 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
00664                 dxy = ((my4 & 1) << 1) | (mx4 & 1);
00665 
00666                 if(s->no_rounding)
00667                     s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , stride, h);
00668                 else
00669                     s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , stride, h);
00670             }
00671             dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
00672         }else
00673             dmin_sum+= dmin4;
00674 
00675         if(s->quarter_sample){
00676             mx4_sum+= mx4/2;
00677             my4_sum+= my4/2;
00678         }else{
00679             mx4_sum+= mx4;
00680             my4_sum+= my4;
00681         }
00682 
00683         s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
00684         s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
00685 
00686         if(mx4 != mx || my4 != my) same=0;
00687     }
00688 
00689     if(same)
00690         return INT_MAX;
00691 
00692     if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
00693         dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
00694     }
00695 
00696     if(c->avctx->mb_cmp&FF_CMP_CHROMA){
00697         int dxy;
00698         int mx, my;
00699         int offset;
00700 
00701         mx= ff_h263_round_chroma(mx4_sum);
00702         my= ff_h263_round_chroma(my4_sum);
00703         dxy = ((my & 1) << 1) | (mx & 1);
00704 
00705         offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
00706 
00707         if(s->no_rounding){
00708             s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
00709             s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
00710         }else{
00711             s->dsp.put_pixels_tab       [1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
00712             s->dsp.put_pixels_tab       [1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
00713         }
00714 
00715         dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad  , s->uvlinesize, 8);
00716         dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad+8, s->uvlinesize, 8);
00717     }
00718 
00719     c->pred_x= mx;
00720     c->pred_y= my;
00721 
00722     switch(c->avctx->mb_cmp&0xFF){
00723     /*case FF_CMP_SSE:
00724         return dmin_sum+ 32*s->qscale*s->qscale;*/
00725     case FF_CMP_RD:
00726         return dmin_sum;
00727     default:
00728         return dmin_sum+ 11*c->mb_penalty_factor;
00729     }
00730 }
00731 
00732 static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
00733     MotionEstContext * const c= &s->me;
00734 
00735     c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
00736     c->src[1][0] = c->src[0][0] + s->linesize;
00737     if(c->flags & FLAG_CHROMA){
00738         c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
00739         c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
00740         c->src[1][1] = c->src[0][1] + s->uvlinesize;
00741         c->src[1][2] = c->src[0][2] + s->uvlinesize;
00742     }
00743 }
00744 
00745 static int interlaced_search(MpegEncContext *s, int ref_index,
00746                              int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
00747 {
00748     MotionEstContext * const c= &s->me;
00749     const int size=0;
00750     const int h=8;
00751     int block;
00752     int P[10][2];
00753     uint8_t * const mv_penalty= c->current_mv_penalty;
00754     int same=1;
00755     const int stride= 2*s->linesize;
00756     int dmin_sum= 0;
00757     const int mot_stride= s->mb_stride;
00758     const int xy= s->mb_x + s->mb_y*mot_stride;
00759 
00760     c->ymin>>=1;
00761     c->ymax>>=1;
00762     c->stride<<=1;
00763     c->uvstride<<=1;
00764     init_interlaced_ref(s, ref_index);
00765 
00766     for(block=0; block<2; block++){
00767         int field_select;
00768         int best_dmin= INT_MAX;
00769         int best_field= -1;
00770 
00771         for(field_select=0; field_select<2; field_select++){
00772             int dmin, mx_i, my_i;
00773             int16_t (*mv_table)[2]= mv_tables[block][field_select];
00774 
00775             if(user_field_select){
00776                 assert(field_select==0 || field_select==1);
00777                 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
00778                 if(field_select_tables[block][xy] != field_select)
00779                     continue;
00780             }
00781 
00782             P_LEFT[0] = mv_table[xy - 1][0];
00783             P_LEFT[1] = mv_table[xy - 1][1];
00784             if(P_LEFT[0]       > (c->xmax<<1)) P_LEFT[0]       = (c->xmax<<1);
00785 
00786             c->pred_x= P_LEFT[0];
00787             c->pred_y= P_LEFT[1];
00788 
00789             if(!s->first_slice_line){
00790                 P_TOP[0]      = mv_table[xy - mot_stride][0];
00791                 P_TOP[1]      = mv_table[xy - mot_stride][1];
00792                 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
00793                 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
00794                 if(P_TOP[1]      > (c->ymax<<1)) P_TOP[1]     = (c->ymax<<1);
00795                 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
00796                 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
00797                 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
00798 
00799                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
00800                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
00801             }
00802             P_MV1[0]= mx; //FIXME not correct if block != field_select
00803             P_MV1[1]= my / 2;
00804 
00805             dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
00806 
00807             dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
00808 
00809             mv_table[xy][0]= mx_i;
00810             mv_table[xy][1]= my_i;
00811 
00812             if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
00813                 int dxy;
00814 
00815                 //FIXME chroma ME
00816                 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
00817                 dxy = ((my_i & 1) << 1) | (mx_i & 1);
00818 
00819                 if(s->no_rounding){
00820                     s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref    , stride, h);
00821                 }else{
00822                     s->dsp.put_pixels_tab       [size][dxy](c->scratchpad, ref    , stride, h);
00823                 }
00824                 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
00825                 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
00826             }else
00827                 dmin+= c->mb_penalty_factor; //field_select bits
00828 
00829             dmin += field_select != block; //slightly prefer same field
00830 
00831             if(dmin < best_dmin){
00832                 best_dmin= dmin;
00833                 best_field= field_select;
00834             }
00835         }
00836         {
00837             int16_t (*mv_table)[2]= mv_tables[block][best_field];
00838 
00839             if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
00840             if(mv_table[xy][1]&1) same=0;
00841             if(mv_table[xy][1]*2 != my) same=0;
00842             if(best_field != block) same=0;
00843         }
00844 
00845         field_select_tables[block][xy]= best_field;
00846         dmin_sum += best_dmin;
00847     }
00848 
00849     c->ymin<<=1;
00850     c->ymax<<=1;
00851     c->stride>>=1;
00852     c->uvstride>>=1;
00853 
00854     if(same)
00855         return INT_MAX;
00856 
00857     switch(c->avctx->mb_cmp&0xFF){
00858     /*case FF_CMP_SSE:
00859         return dmin_sum+ 32*s->qscale*s->qscale;*/
00860     case FF_CMP_RD:
00861         return dmin_sum;
00862     default:
00863         return dmin_sum+ 11*c->mb_penalty_factor;
00864     }
00865 }
00866 
00867 static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
00868     int ymax= s->me.ymax>>interlaced;
00869     int ymin= s->me.ymin>>interlaced;
00870 
00871     if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
00872     if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
00873     if(mv[1] <       ymin) mv[1] =       ymin;
00874     if(mv[1] >       ymax) mv[1] =       ymax;
00875 }
00876 
00877 static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
00878     MotionEstContext * const c= &s->me;
00879     Picture *p= s->current_picture_ptr;
00880     int mb_xy= mb_x + mb_y*s->mb_stride;
00881     int xy= 2*mb_x + 2*mb_y*s->b8_stride;
00882     int mb_type= s->current_picture.mb_type[mb_xy];
00883     int flags= c->flags;
00884     int shift= (flags&FLAG_QPEL) + 1;
00885     int mask= (1<<shift)-1;
00886     int x, y, i;
00887     int d=0;
00888     me_cmp_func cmpf= s->dsp.sse[0];
00889     me_cmp_func chroma_cmpf= s->dsp.sse[1];
00890 
00891     if(p_type && USES_LIST(mb_type, 1)){
00892         av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
00893         return INT_MAX/2;
00894     }
00895     assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
00896 
00897     for(i=0; i<4; i++){
00898         int xy= s->block_index[i];
00899         clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
00900         clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
00901     }
00902 
00903     if(IS_INTERLACED(mb_type)){
00904         int xy2= xy  + s->b8_stride;
00905         s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
00906         c->stride<<=1;
00907         c->uvstride<<=1;
00908 
00909         if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
00910             av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
00911             return INT_MAX/2;
00912         }
00913 
00914         if(USES_LIST(mb_type, 0)){
00915             int field_select0= p->ref_index[0][4*mb_xy  ];
00916             int field_select1= p->ref_index[0][4*mb_xy+2];
00917             assert(field_select0==0 ||field_select0==1);
00918             assert(field_select1==0 ||field_select1==1);
00919             init_interlaced_ref(s, 0);
00920 
00921             if(p_type){
00922                 s->p_field_select_table[0][mb_xy]= field_select0;
00923                 s->p_field_select_table[1][mb_xy]= field_select1;
00924                 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
00925                 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
00926                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
00927             }else{
00928                 s->b_field_select_table[0][0][mb_xy]= field_select0;
00929                 s->b_field_select_table[0][1][mb_xy]= field_select1;
00930                 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
00931                 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
00932                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
00933             }
00934 
00935             x= p->motion_val[0][xy ][0];
00936             y= p->motion_val[0][xy ][1];
00937             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
00938             x= p->motion_val[0][xy2][0];
00939             y= p->motion_val[0][xy2][1];
00940             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
00941         }
00942         if(USES_LIST(mb_type, 1)){
00943             int field_select0= p->ref_index[1][4*mb_xy  ];
00944             int field_select1= p->ref_index[1][4*mb_xy+2];
00945             assert(field_select0==0 ||field_select0==1);
00946             assert(field_select1==0 ||field_select1==1);
00947             init_interlaced_ref(s, 2);
00948 
00949             s->b_field_select_table[1][0][mb_xy]= field_select0;
00950             s->b_field_select_table[1][1][mb_xy]= field_select1;
00951             *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
00952             *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
00953             if(USES_LIST(mb_type, 0)){
00954                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
00955             }else{
00956                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
00957             }
00958 
00959             x= p->motion_val[1][xy ][0];
00960             y= p->motion_val[1][xy ][1];
00961             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
00962             x= p->motion_val[1][xy2][0];
00963             y= p->motion_val[1][xy2][1];
00964             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
00965             //FIXME bidir scores
00966         }
00967         c->stride>>=1;
00968         c->uvstride>>=1;
00969     }else if(IS_8X8(mb_type)){
00970         if(!(s->flags & CODEC_FLAG_4MV)){
00971             av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
00972             return INT_MAX/2;
00973         }
00974         cmpf= s->dsp.sse[1];
00975         chroma_cmpf= s->dsp.sse[1];
00976         init_mv4_ref(c);
00977         for(i=0; i<4; i++){
00978             xy= s->block_index[i];
00979             x= p->motion_val[0][xy][0];
00980             y= p->motion_val[0][xy][1];
00981             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
00982         }
00983         s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
00984     }else{
00985         if(USES_LIST(mb_type, 0)){
00986             if(p_type){
00987                 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
00988                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
00989             }else if(USES_LIST(mb_type, 1)){
00990                 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
00991                 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
00992                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
00993             }else{
00994                 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
00995                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
00996             }
00997             x= p->motion_val[0][xy][0];
00998             y= p->motion_val[0][xy][1];
00999             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
01000         }else if(USES_LIST(mb_type, 1)){
01001             *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
01002             s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
01003 
01004             x= p->motion_val[1][xy][0];
01005             y= p->motion_val[1][xy][1];
01006             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
01007         }else
01008             s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
01009     }
01010     return d;
01011 }
01012 
01013 void ff_estimate_p_frame_motion(MpegEncContext * s,
01014                                 int mb_x, int mb_y)
01015 {
01016     MotionEstContext * const c= &s->me;
01017     uint8_t *pix, *ppix;
01018     int sum, mx, my, dmin;
01019     int varc;            
01020     int vard;            
01021     int P[10][2];
01022     const int shift= 1+s->quarter_sample;
01023     int mb_type=0;
01024     Picture * const pic= &s->current_picture;
01025 
01026     init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
01027 
01028     assert(s->quarter_sample==0 || s->quarter_sample==1);
01029     assert(s->linesize == c->stride);
01030     assert(s->uvlinesize == c->uvstride);
01031 
01032     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
01033     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
01034     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
01035     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
01036 
01037     get_limits(s, 16*mb_x, 16*mb_y);
01038     c->skip=0;
01039 
01040     /* intra / predictive decision */
01041     pix = c->src[0][0];
01042     sum = s->dsp.pix_sum(pix, s->linesize);
01043     varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
01044 
01045     pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
01046     pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
01047     c->mb_var_sum_temp += (varc+128)>>8;
01048 
01049     if(c->avctx->me_threshold){
01050         vard= check_input_motion(s, mb_x, mb_y, 1);
01051 
01052         if((vard+128)>>8 < c->avctx->me_threshold){
01053             int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
01054             int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
01055             pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
01056             c->mc_mb_var_sum_temp += (vard+128)>>8;
01057             c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
01058             return;
01059         }
01060         if((vard+128)>>8 < c->avctx->mb_threshold)
01061             mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
01062     }
01063 
01064     switch(s->me_method) {
01065     case ME_ZERO:
01066     default:
01067         no_motion_search(s, &mx, &my);
01068         mx-= mb_x*16;
01069         my-= mb_y*16;
01070         dmin = 0;
01071         break;
01072     case ME_X1:
01073     case ME_EPZS:
01074        {
01075             const int mot_stride = s->b8_stride;
01076             const int mot_xy = s->block_index[0];
01077 
01078             P_LEFT[0]       = s->current_picture.motion_val[0][mot_xy - 1][0];
01079             P_LEFT[1]       = s->current_picture.motion_val[0][mot_xy - 1][1];
01080 
01081             if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
01082 
01083             if(!s->first_slice_line) {
01084                 P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][0];
01085                 P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][1];
01086                 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
01087                 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
01088                 if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
01089                 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
01090                 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
01091 
01092                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
01093                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
01094 
01095                 if(s->out_format == FMT_H263){
01096                     c->pred_x = P_MEDIAN[0];
01097                     c->pred_y = P_MEDIAN[1];
01098                 }else { /* mpeg1 at least */
01099                     c->pred_x= P_LEFT[0];
01100                     c->pred_y= P_LEFT[1];
01101                 }
01102             }else{
01103                 c->pred_x= P_LEFT[0];
01104                 c->pred_y= P_LEFT[1];
01105             }
01106 
01107         }
01108         dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
01109 
01110         break;
01111     }
01112 
01113     /* At this point (mx,my) are full-pell and the relative displacement */
01114     ppix = c->ref[0][0] + (my * s->linesize) + mx;
01115 
01116     vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
01117 
01118     pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
01119 //    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
01120     c->mc_mb_var_sum_temp += (vard+128)>>8;
01121 
01122     if(mb_type){
01123         int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
01124         int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
01125         c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
01126 
01127         if(mb_type == CANDIDATE_MB_TYPE_INTER){
01128             c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01129             set_p_mv_tables(s, mx, my, 1);
01130         }else{
01131             mx <<=shift;
01132             my <<=shift;
01133         }
01134         if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
01135             h263_mv4_search(s, mx, my, shift);
01136 
01137             set_p_mv_tables(s, mx, my, 0);
01138         }
01139         if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
01140             interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
01141         }
01142     }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
01143         int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
01144         int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
01145         c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
01146 
01147         if (vard*2 + 200*256 > varc)
01148             mb_type|= CANDIDATE_MB_TYPE_INTRA;
01149         if (varc*2 + 200*256 > vard || s->qscale > 24){
01150 //        if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
01151             mb_type|= CANDIDATE_MB_TYPE_INTER;
01152             c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01153             if(s->flags&CODEC_FLAG_MV0)
01154                 if(mx || my)
01155                     mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
01156         }else{
01157             mx <<=shift;
01158             my <<=shift;
01159         }
01160         if((s->flags&CODEC_FLAG_4MV)
01161            && !c->skip && varc>50<<8 && vard>10<<8){
01162             if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
01163                 mb_type|=CANDIDATE_MB_TYPE_INTER4V;
01164 
01165             set_p_mv_tables(s, mx, my, 0);
01166         }else
01167             set_p_mv_tables(s, mx, my, 1);
01168         if((s->flags&CODEC_FLAG_INTERLACED_ME)
01169            && !c->skip){ //FIXME varc/d checks
01170             if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
01171                 mb_type |= CANDIDATE_MB_TYPE_INTER_I;
01172         }
01173     }else{
01174         int intra_score, i;
01175         mb_type= CANDIDATE_MB_TYPE_INTER;
01176 
01177         dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01178         if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
01179             dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
01180 
01181         if((s->flags&CODEC_FLAG_4MV)
01182            && !c->skip && varc>50<<8 && vard>10<<8){
01183             int dmin4= h263_mv4_search(s, mx, my, shift);
01184             if(dmin4 < dmin){
01185                 mb_type= CANDIDATE_MB_TYPE_INTER4V;
01186                 dmin=dmin4;
01187             }
01188         }
01189         if((s->flags&CODEC_FLAG_INTERLACED_ME)
01190            && !c->skip){ //FIXME varc/d checks
01191             int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
01192             if(dmin_i < dmin){
01193                 mb_type = CANDIDATE_MB_TYPE_INTER_I;
01194                 dmin= dmin_i;
01195             }
01196         }
01197 
01198 //        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
01199         set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
01200 
01201         /* get intra luma score */
01202         if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
01203             intra_score= varc - 500;
01204         }else{
01205             int mean= (sum+128)>>8;
01206             mean*= 0x01010101;
01207 
01208             for(i=0; i<16; i++){
01209                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
01210                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
01211                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
01212                 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
01213             }
01214 
01215             intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
01216         }
01217 #if 0 //FIXME
01218         /* get chroma score */
01219         if(c->avctx->mb_cmp&FF_CMP_CHROMA){
01220             for(i=1; i<3; i++){
01221                 uint8_t *dest_c;
01222                 int mean;
01223 
01224                 if(s->out_format == FMT_H263){
01225                     mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
01226                 }else{
01227                     mean= (s->last_dc[i] + 4)>>3;
01228                 }
01229                 dest_c = s->new_picture.data[i] + (mb_y * 8  * (s->uvlinesize)) + mb_x * 8;
01230 
01231                 mean*= 0x01010101;
01232                 for(i=0; i<8; i++){
01233                     *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
01234                     *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
01235                 }
01236 
01237                 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
01238             }
01239         }
01240 #endif
01241         intra_score += c->mb_penalty_factor*16;
01242 
01243         if(intra_score < dmin){
01244             mb_type= CANDIDATE_MB_TYPE_INTRA;
01245             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
01246         }else
01247             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
01248 
01249         {
01250             int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
01251             int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
01252             c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
01253         }
01254     }
01255 
01256     s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
01257 }
01258 
01259 int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
01260                                     int mb_x, int mb_y)
01261 {
01262     MotionEstContext * const c= &s->me;
01263     int mx, my, dmin;
01264     int P[10][2];
01265     const int shift= 1+s->quarter_sample;
01266     const int xy= mb_x + mb_y*s->mb_stride;
01267     init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
01268 
01269     assert(s->quarter_sample==0 || s->quarter_sample==1);
01270 
01271     c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
01272     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
01273 
01274     get_limits(s, 16*mb_x, 16*mb_y);
01275     c->skip=0;
01276 
01277     P_LEFT[0]       = s->p_mv_table[xy + 1][0];
01278     P_LEFT[1]       = s->p_mv_table[xy + 1][1];
01279 
01280     if(P_LEFT[0]       < (c->xmin<<shift)) P_LEFT[0]       = (c->xmin<<shift);
01281 
01282     /* special case for first line */
01283     if (s->first_slice_line) {
01284         c->pred_x= P_LEFT[0];
01285         c->pred_y= P_LEFT[1];
01286         P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
01287         P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
01288     } else {
01289         P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
01290         P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
01291         P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
01292         P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
01293         if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);
01294         if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
01295         if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
01296 
01297         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
01298         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
01299 
01300         c->pred_x = P_MEDIAN[0];
01301         c->pred_y = P_MEDIAN[1];
01302     }
01303 
01304     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
01305 
01306     s->p_mv_table[xy][0] = mx<<shift;
01307     s->p_mv_table[xy][1] = my<<shift;
01308 
01309     return dmin;
01310 }
01311 
01312 static int ff_estimate_motion_b(MpegEncContext * s,
01313                        int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
01314 {
01315     MotionEstContext * const c= &s->me;
01316     int mx, my, dmin;
01317     int P[10][2];
01318     const int shift= 1+s->quarter_sample;
01319     const int mot_stride = s->mb_stride;
01320     const int mot_xy = mb_y*mot_stride + mb_x;
01321     uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
01322     int mv_scale;
01323 
01324     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
01325     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
01326     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
01327     c->current_mv_penalty= mv_penalty;
01328 
01329     get_limits(s, 16*mb_x, 16*mb_y);
01330 
01331     switch(s->me_method) {
01332     case ME_ZERO:
01333     default:
01334         no_motion_search(s, &mx, &my);
01335         dmin = 0;
01336         mx-= mb_x*16;
01337         my-= mb_y*16;
01338         break;
01339     case ME_X1:
01340     case ME_EPZS:
01341        {
01342             P_LEFT[0]        = mv_table[mot_xy - 1][0];
01343             P_LEFT[1]        = mv_table[mot_xy - 1][1];
01344 
01345             if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
01346 
01347             /* special case for first line */
01348             if (!s->first_slice_line) {
01349                 P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
01350                 P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
01351                 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
01352                 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
01353                 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
01354                 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
01355                 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
01356 
01357                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
01358                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
01359             }
01360             c->pred_x= P_LEFT[0];
01361             c->pred_y= P_LEFT[1];
01362         }
01363 
01364         if(mv_table == s->b_forw_mv_table){
01365             mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
01366         }else{
01367             mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
01368         }
01369 
01370         dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
01371 
01372         break;
01373     }
01374 
01375     dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
01376 
01377     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
01378         dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
01379 
01380 //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
01381 //    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
01382     mv_table[mot_xy][0]= mx;
01383     mv_table[mot_xy][1]= my;
01384 
01385     return dmin;
01386 }
01387 
01388 static inline int check_bidir_mv(MpegEncContext * s,
01389                    int motion_fx, int motion_fy,
01390                    int motion_bx, int motion_by,
01391                    int pred_fx, int pred_fy,
01392                    int pred_bx, int pred_by,
01393                    int size, int h)
01394 {
01395     //FIXME optimize?
01396     //FIXME better f_code prediction (max mv & distance)
01397     //FIXME pointers
01398     MotionEstContext * const c= &s->me;
01399     uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
01400     uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
01401     int stride= c->stride;
01402     uint8_t *dest_y = c->scratchpad;
01403     uint8_t *ptr;
01404     int dxy;
01405     int src_x, src_y;
01406     int fbmin;
01407     uint8_t **src_data= c->src[0];
01408     uint8_t **ref_data= c->ref[0];
01409     uint8_t **ref2_data= c->ref[2];
01410 
01411     if(s->quarter_sample){
01412         dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
01413         src_x = motion_fx >> 2;
01414         src_y = motion_fy >> 2;
01415 
01416         ptr = ref_data[0] + (src_y * stride) + src_x;
01417         s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , stride);
01418 
01419         dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
01420         src_x = motion_bx >> 2;
01421         src_y = motion_by >> 2;
01422 
01423         ptr = ref2_data[0] + (src_y * stride) + src_x;
01424         s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);
01425     }else{
01426         dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
01427         src_x = motion_fx >> 1;
01428         src_y = motion_fy >> 1;
01429 
01430         ptr = ref_data[0] + (src_y * stride) + src_x;
01431         s->dsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
01432 
01433         dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
01434         src_x = motion_bx >> 1;
01435         src_y = motion_by >> 1;
01436 
01437         ptr = ref2_data[0] + (src_y * stride) + src_x;
01438         s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
01439     }
01440 
01441     fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
01442            +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
01443            + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
01444 
01445     if(c->avctx->mb_cmp&FF_CMP_CHROMA){
01446     }
01447     //FIXME CHROMA !!!
01448 
01449     return fbmin;
01450 }
01451 
01452 /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
01453 static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
01454 {
01455     MotionEstContext * const c= &s->me;
01456     const int mot_stride = s->mb_stride;
01457     const int xy = mb_y *mot_stride + mb_x;
01458     int fbmin;
01459     int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
01460     int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
01461     int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
01462     int pred_by= s->b_bidir_back_mv_table[xy-1][1];
01463     int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
01464     int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
01465     int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
01466     int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
01467     const int flags= c->sub_flags;
01468     const int qpel= flags&FLAG_QPEL;
01469     const int shift= 1+qpel;
01470     const int xmin= c->xmin<<shift;
01471     const int ymin= c->ymin<<shift;
01472     const int xmax= c->xmax<<shift;
01473     const int ymax= c->ymax<<shift;
01474 #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))
01475 #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))
01476     int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);
01477     uint8_t map[256];
01478 
01479     memset(map,0,sizeof(map));
01480     map[hashidx&255] = 1;
01481 
01482     fbmin= check_bidir_mv(s, motion_fx, motion_fy,
01483                           motion_bx, motion_by,
01484                           pred_fx, pred_fy,
01485                           pred_bx, pred_by,
01486                           0, 16);
01487 
01488     if(s->avctx->bidir_refine){
01489         int end;
01490         static const uint8_t limittab[5]={0,8,32,64,80};
01491         const int limit= limittab[s->avctx->bidir_refine];
01492         static const int8_t vect[][4]={
01493 { 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},
01494 
01495 { 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},
01496 { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},
01497 { 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},
01498 { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},
01499 
01500 { 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},
01501 { 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},
01502 { 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},
01503 { 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},
01504 
01505 { 1, 1, 1, 1}, {-1,-1,-1,-1},
01506 { 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},
01507 { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},
01508         };
01509         static const uint8_t hash[]={
01510 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),
01511 
01512 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),
01513 HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),
01514 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),
01515 HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),
01516 
01517 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),
01518 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),
01519 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),
01520 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),
01521 
01522 HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),
01523 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),
01524 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),
01525 };
01526 
01527 #define CHECK_BIDIR(fx,fy,bx,by)\
01528     if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\
01529        &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
01530        &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
01531         int score;\
01532         map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\
01533         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);\
01534         if(score < fbmin){\
01535             hashidx += HASH(fx,fy,bx,by);\
01536             fbmin= score;\
01537             motion_fx+=fx;\
01538             motion_fy+=fy;\
01539             motion_bx+=bx;\
01540             motion_by+=by;\
01541             end=0;\
01542         }\
01543     }
01544 #define CHECK_BIDIR2(a,b,c,d)\
01545 CHECK_BIDIR(a,b,c,d)\
01546 CHECK_BIDIR(-(a),-(b),-(c),-(d))
01547 
01548         do{
01549             int i;
01550             int borderdist=0;
01551             end=1;
01552 
01553             CHECK_BIDIR2(0,0,0,1)
01554             CHECK_BIDIR2(0,0,1,0)
01555             CHECK_BIDIR2(0,1,0,0)
01556             CHECK_BIDIR2(1,0,0,0)
01557 
01558             for(i=8; i<limit; i++){
01559                 int fx= motion_fx+vect[i][0];
01560                 int fy= motion_fy+vect[i][1];
01561                 int bx= motion_bx+vect[i][2];
01562                 int by= motion_by+vect[i][3];
01563                 if(borderdist<=0){
01564                     int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);
01565                     int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);
01566                     if((a|b) < 0)
01567                         map[(hashidx+hash[i])&255] = 1;
01568                 }
01569                 if(!map[(hashidx+hash[i])&255]){
01570                     int score;
01571                     map[(hashidx+hash[i])&255] = 1;
01572                     score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);
01573                     if(score < fbmin){
01574                         hashidx += hash[i];
01575                         fbmin= score;
01576                         motion_fx=fx;
01577                         motion_fy=fy;
01578                         motion_bx=bx;
01579                         motion_by=by;
01580                         end=0;
01581                         borderdist--;
01582                         if(borderdist<=0){
01583                             int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);
01584                             int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);
01585                             borderdist= FFMIN(a,b);
01586                         }
01587                     }
01588                 }
01589             }
01590         }while(!end);
01591     }
01592 
01593     s->b_bidir_forw_mv_table[xy][0]= motion_fx;
01594     s->b_bidir_forw_mv_table[xy][1]= motion_fy;
01595     s->b_bidir_back_mv_table[xy][0]= motion_bx;
01596     s->b_bidir_back_mv_table[xy][1]= motion_by;
01597 
01598     return fbmin;
01599 }
01600 
01601 static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
01602 {
01603     MotionEstContext * const c= &s->me;
01604     int P[10][2];
01605     const int mot_stride = s->mb_stride;
01606     const int mot_xy = mb_y*mot_stride + mb_x;
01607     const int shift= 1+s->quarter_sample;
01608     int dmin, i;
01609     const int time_pp= s->pp_time;
01610     const int time_pb= s->pb_time;
01611     int mx, my, xmin, xmax, ymin, ymax;
01612     int16_t (*mv_table)[2]= s->b_direct_mv_table;
01613 
01614     c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
01615     ymin= xmin=(-32)>>shift;
01616     ymax= xmax=   31>>shift;
01617 
01618     if(IS_8X8(s->next_picture.mb_type[mot_xy])){
01619         s->mv_type= MV_TYPE_8X8;
01620     }else{
01621         s->mv_type= MV_TYPE_16X16;
01622     }
01623 
01624     for(i=0; i<4; i++){
01625         int index= s->block_index[i];
01626         int min, max;
01627 
01628         c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
01629         c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
01630         c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
01631         c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
01632 //        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
01633 //        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
01634 
01635         max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
01636         min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
01637         max+= 16*mb_x + 1; // +-1 is for the simpler rounding
01638         min+= 16*mb_x - 1;
01639         xmax= FFMIN(xmax, s->width - max);
01640         xmin= FFMAX(xmin, - 16     - min);
01641 
01642         max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
01643         min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
01644         max+= 16*mb_y + 1; // +-1 is for the simpler rounding
01645         min+= 16*mb_y - 1;
01646         ymax= FFMIN(ymax, s->height - max);
01647         ymin= FFMAX(ymin, - 16      - min);
01648 
01649         if(s->mv_type == MV_TYPE_16X16) break;
01650     }
01651 
01652     assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
01653 
01654     if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
01655         s->b_direct_mv_table[mot_xy][0]= 0;
01656         s->b_direct_mv_table[mot_xy][1]= 0;
01657 
01658         return 256*256*256*64;
01659     }
01660 
01661     c->xmin= xmin;
01662     c->ymin= ymin;
01663     c->xmax= xmax;
01664     c->ymax= ymax;
01665     c->flags     |= FLAG_DIRECT;
01666     c->sub_flags |= FLAG_DIRECT;
01667     c->pred_x=0;
01668     c->pred_y=0;
01669 
01670     P_LEFT[0]        = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
01671     P_LEFT[1]        = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
01672 
01673     /* special case for first line */
01674     if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
01675         P_TOP[0]      = av_clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
01676         P_TOP[1]      = av_clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
01677         P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
01678         P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
01679 
01680         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
01681         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
01682     }
01683 
01684     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
01685     if(c->sub_flags&FLAG_QPEL)
01686         dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01687     else
01688         dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
01689 
01690     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
01691         dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
01692 
01693     get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
01694 
01695     mv_table[mot_xy][0]= mx;
01696     mv_table[mot_xy][1]= my;
01697     c->flags     &= ~FLAG_DIRECT;
01698     c->sub_flags &= ~FLAG_DIRECT;
01699 
01700     return dmin;
01701 }
01702 
01703 void ff_estimate_b_frame_motion(MpegEncContext * s,
01704                              int mb_x, int mb_y)
01705 {
01706     MotionEstContext * const c= &s->me;
01707     const int penalty_factor= c->mb_penalty_factor;
01708     int fmin, bmin, dmin, fbmin, bimin, fimin;
01709     int type=0;
01710     const int xy = mb_y*s->mb_stride + mb_x;
01711     init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
01712 
01713     get_limits(s, 16*mb_x, 16*mb_y);
01714 
01715     c->skip=0;
01716 
01717     if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
01718         int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
01719 
01720         score= ((unsigned)(score*score + 128*256))>>16;
01721         c->mc_mb_var_sum_temp += score;
01722         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
01723         s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
01724 
01725         return;
01726     }
01727 
01728     if(c->avctx->me_threshold){
01729         int vard= check_input_motion(s, mb_x, mb_y, 0);
01730 
01731         if((vard+128)>>8 < c->avctx->me_threshold){
01732 //            pix = c->src[0][0];
01733 //            sum = s->dsp.pix_sum(pix, s->linesize);
01734 //            varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
01735 
01736 //            pic->mb_var   [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
01737              s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
01738 /*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
01739             c->mb_var_sum_temp    += (varc+128)>>8;*/
01740             c->mc_mb_var_sum_temp += (vard+128)>>8;
01741 /*            if (vard <= 64<<8 || vard < varc) {
01742                 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
01743             }else{
01744                 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
01745             }*/
01746             return;
01747         }
01748         if((vard+128)>>8 < c->avctx->mb_threshold){
01749             type= s->mb_type[mb_y*s->mb_stride + mb_x];
01750             if(type == CANDIDATE_MB_TYPE_DIRECT){
01751                 direct_search(s, mb_x, mb_y);
01752             }
01753             if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
01754                 c->skip=0;
01755                 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
01756             }
01757             if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
01758                 c->skip=0;
01759                 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
01760             }
01761             if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
01762                 c->skip=0;
01763                 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
01764                 interlaced_search(s, 0,
01765                                         s->b_field_mv_table[0], s->b_field_select_table[0],
01766                                         s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
01767             }
01768             if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
01769                 c->skip=0;
01770                 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
01771                 interlaced_search(s, 2,
01772                                         s->b_field_mv_table[1], s->b_field_select_table[1],
01773                                         s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
01774             }
01775             return;
01776         }
01777     }
01778 
01779     if (s->codec_id == CODEC_ID_MPEG4)
01780         dmin= direct_search(s, mb_x, mb_y);
01781     else
01782         dmin= INT_MAX;
01783 //FIXME penalty stuff for non mpeg4
01784     c->skip=0;
01785     fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
01786 
01787     c->skip=0;
01788     bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
01789 //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
01790 
01791     c->skip=0;
01792     fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
01793 //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
01794 
01795     if(s->flags & CODEC_FLAG_INTERLACED_ME){
01796 //FIXME mb type penalty
01797         c->skip=0;
01798         c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
01799         fimin= interlaced_search(s, 0,
01800                                  s->b_field_mv_table[0], s->b_field_select_table[0],
01801                                  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
01802         c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
01803         bimin= interlaced_search(s, 2,
01804                                  s->b_field_mv_table[1], s->b_field_select_table[1],
01805                                  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
01806     }else
01807         fimin= bimin= INT_MAX;
01808 
01809     {
01810         int score= fmin;
01811         type = CANDIDATE_MB_TYPE_FORWARD;
01812 
01813         if (dmin <= score){
01814             score = dmin;
01815             type = CANDIDATE_MB_TYPE_DIRECT;
01816         }
01817         if(bmin<score){
01818             score=bmin;
01819             type= CANDIDATE_MB_TYPE_BACKWARD;
01820         }
01821         if(fbmin<score){
01822             score=fbmin;
01823             type= CANDIDATE_MB_TYPE_BIDIR;
01824         }
01825         if(fimin<score){
01826             score=fimin;
01827             type= CANDIDATE_MB_TYPE_FORWARD_I;
01828         }
01829         if(bimin<score){
01830             score=bimin;
01831             type= CANDIDATE_MB_TYPE_BACKWARD_I;
01832         }
01833 
01834         score= ((unsigned)(score*score + 128*256))>>16;
01835         c->mc_mb_var_sum_temp += score;
01836         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
01837     }
01838 
01839     if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
01840         type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
01841         if(fimin < INT_MAX)
01842             type |= CANDIDATE_MB_TYPE_FORWARD_I;
01843         if(bimin < INT_MAX)
01844             type |= CANDIDATE_MB_TYPE_BACKWARD_I;
01845         if(fimin < INT_MAX && bimin < INT_MAX){
01846             type |= CANDIDATE_MB_TYPE_BIDIR_I;
01847         }
01848          //FIXME something smarter
01849         if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
01850         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])
01851             type |= CANDIDATE_MB_TYPE_DIRECT0;
01852 #if 0
01853         if(s->out_format == FMT_MPEG1)
01854             type |= CANDIDATE_MB_TYPE_INTRA;
01855 #endif
01856     }
01857 
01858     s->mb_type[mb_y*s->mb_stride + mb_x]= type;
01859 }
01860 
01861 /* find best f_code for ME which do unlimited searches */
01862 int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
01863 {
01864     if(s->me_method>=ME_EPZS){
01865         int score[8];
01866         int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
01867         uint8_t * fcode_tab= s->fcode_tab;
01868         int best_fcode=-1;
01869         int best_score=-10000000;
01870 
01871         if(s->msmpeg4_version)
01872             range= FFMIN(range, 16);
01873         else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
01874             range= FFMIN(range, 256);
01875 
01876         for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
01877 
01878         for(y=0; y<s->mb_height; y++){
01879             int x;
01880             int xy= y*s->mb_stride;
01881             for(x=0; x<s->mb_width; x++){
01882                 if(s->mb_type[xy] & type){
01883                     int mx= mv_table[xy][0];
01884                     int my= mv_table[xy][1];
01885                     int fcode= FFMAX(fcode_tab[mx + MAX_MV],
01886                                      fcode_tab[my + MAX_MV]);
01887                     int j;
01888 
01889                         if(mx >= range || mx < -range ||
01890                            my >= range || my < -range)
01891                             continue;
01892 
01893                     for(j=0; j<fcode && j<8; j++){
01894                         if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
01895                             score[j]-= 170;
01896                     }
01897                 }
01898                 xy++;
01899             }
01900         }
01901 
01902         for(i=1; i<8; i++){
01903             if(score[i] > best_score){
01904                 best_score= score[i];
01905                 best_fcode= i;
01906             }
01907 //            printf("%d %d\n", i, score[i]);
01908         }
01909 
01910 //    printf("fcode: %d type: %d\n", i, s->pict_type);
01911         return best_fcode;
01912 /*        for(i=0; i<=MAX_FCODE; i++){
01913             printf("%d ", mv_num[i]);
01914         }
01915         printf("\n");*/
01916     }else{
01917         return 1;
01918     }
01919 }
01920 
01921 void ff_fix_long_p_mvs(MpegEncContext * s)
01922 {
01923     MotionEstContext * const c= &s->me;
01924     const int f_code= s->f_code;
01925     int y, range;
01926     assert(s->pict_type==AV_PICTURE_TYPE_P);
01927 
01928     range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
01929 
01930     assert(range <= 16 || !s->msmpeg4_version);
01931     assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
01932 
01933     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
01934 
01935 //printf("%d no:%d %d//\n", clip, noclip, f_code);
01936     if(s->flags&CODEC_FLAG_4MV){
01937         const int wrap= s->b8_stride;
01938 
01939         /* clip / convert to intra 8x8 type MVs */
01940         for(y=0; y<s->mb_height; y++){
01941             int xy= y*2*wrap;
01942             int i= y*s->mb_stride;
01943             int x;
01944 
01945             for(x=0; x<s->mb_width; x++){
01946                 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
01947                     int block;
01948                     for(block=0; block<4; block++){
01949                         int off= (block& 1) + (block>>1)*wrap;
01950                         int mx= s->current_picture.motion_val[0][ xy + off ][0];
01951                         int my= s->current_picture.motion_val[0][ xy + off ][1];
01952 
01953                         if(   mx >=range || mx <-range
01954                            || my >=range || my <-range){
01955                             s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
01956                             s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
01957                             s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
01958                         }
01959                     }
01960                 }
01961                 xy+=2;
01962                 i++;
01963             }
01964         }
01965     }
01966 }
01967 
01972 void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
01973                      int16_t (*mv_table)[2], int f_code, int type, int truncate)
01974 {
01975     MotionEstContext * const c= &s->me;
01976     int y, h_range, v_range;
01977 
01978     // RAL: 8 in MPEG-1, 16 in MPEG-4
01979     int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
01980 
01981     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
01982 
01983     h_range= range;
01984     v_range= field_select_table ? range>>1 : range;
01985 
01986     /* clip / convert to intra 16x16 type MVs */
01987     for(y=0; y<s->mb_height; y++){
01988         int x;
01989         int xy= y*s->mb_stride;
01990         for(x=0; x<s->mb_width; x++){
01991             if (s->mb_type[xy] & type){    // RAL: "type" test added...
01992                 if(field_select_table==NULL || field_select_table[xy] == field_select){
01993                     if(   mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
01994                        || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
01995 
01996                         if(truncate){
01997                             if     (mv_table[xy][0] > h_range-1) mv_table[xy][0]=  h_range-1;
01998                             else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
01999                             if     (mv_table[xy][1] > v_range-1) mv_table[xy][1]=  v_range-1;
02000                             else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
02001                         }else{
02002                             s->mb_type[xy] &= ~type;
02003                             s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
02004                             mv_table[xy][0]=
02005                             mv_table[xy][1]= 0;
02006                         }
02007                     }
02008                 }
02009             }
02010             xy++;
02011         }
02012     }
02013 }