falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
15543 Commits
32 Branches
935MB
Tree: d63f529df7
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'd63f529df7'
${ noResults }
FFmpeg/libavcodec/motion_est.c

1900 lines
70KB
Raw Blame History 

  1. /*

  2.  * Motion estimation

  3.  * Copyright (c) 2000,2001 Fabrice Bellard.

  4.  * Copyright (c) 2002-2004 Michael Niedermayer

  5.  *

  6.  * new motion estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>

  7.  *

  8.  * This file is part of FFmpeg.

  9.  *

  10.  * FFmpeg is free software; you can redistribute it and/or

  11.  * modify it under the terms of the GNU Lesser General Public

  12.  * License as published by the Free Software Foundation; either

  13.  * version 2.1 of the License, or (at your option) any later version.

  14.  *

  15.  * FFmpeg is distributed in the hope that it will be useful,

  16.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  17.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

  18.  * Lesser General Public License for more details.

  19.  *

  20.  * You should have received a copy of the GNU Lesser General Public

  21.  * License along with FFmpeg; if not, write to the Free Software

  22.  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

  23.  */

  24. 

  25. /**

  26.  * @file motion_est.c

  27.  * Motion estimation.

  28.  */

  29. 

  30. #include <stdlib.h>

  31. #include <stdio.h>

  32. #include <limits.h>

  33. #include "avcodec.h"

  34. #include "dsputil.h"

  35. #include "mpegvideo.h"

  36. 

  37. #undef NDEBUG

  38. #include <assert.h>

  39. 

  40. #define SQ(a) ((a)*(a))

  41. 

  42. #define P_LEFT P[1]

  43. #define P_TOP P[2]

  44. #define P_TOPRIGHT P[3]

  45. #define P_MEDIAN P[4]

  46. #define P_MV1 P[9]

  47. 

  48. static inline int sad_hpel_motion_search(MpegEncContext * s,

  49.                                   int *mx_ptr, int *my_ptr, int dmin,

  50.                                   int src_index, int ref_index,

  51.                                   int size, int h);

  52. 

  53. static inline int update_map_generation(MotionEstContext *c)

  54. {

  55.     c->map_generation+= 1<<(ME_MAP_MV_BITS*2);

  56.     if(c->map_generation==0){

  57.         c->map_generation= 1<<(ME_MAP_MV_BITS*2);

  58.         memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);

  59.     }

  60.     return c->map_generation;

  61. }

  62. 

  63. /* shape adaptive search stuff */

  64. typedef struct Minima{

  65.     int height;

  66.     int x, y;

  67.     int checked;

  68. }Minima;

  69. 

  70. static int minima_cmp(const void *a, const void *b){

  71.     const Minima *da = (const Minima *) a;

  72.     const Minima *db = (const Minima *) b;

  73. 

  74.     return da->height - db->height;

  75. }

  76. 

  77. #define FLAG_QPEL   1 //must be 1

  78. #define FLAG_CHROMA 2

  79. #define FLAG_DIRECT 4

  80. 

  81. 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){

  82.     const int offset[3]= {

  83.           y*c->  stride + x,

  84.         ((y*c->uvstride + x)>>1),

  85.         ((y*c->uvstride + x)>>1),

  86.     };

  87.     int i;

  88.     for(i=0; i<3; i++){

  89.         c->src[0][i]= src [i] + offset[i];

  90.         c->ref[0][i]= ref [i] + offset[i];

  91.     }

  92.     if(ref_index){

  93.         for(i=0; i<3; i++){

  94.             c->ref[ref_index][i]= ref2[i] + offset[i];

  95.         }

  96.     }

  97. }

  98. 

  99. static int get_flags(MotionEstContext *c, int direct, int chroma){

  100.     return   ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)

  101.            + (direct ? FLAG_DIRECT : 0)

  102.            + (chroma ? FLAG_CHROMA : 0);

  103. }

  104. 

  105. /*! \brief compares a block (either a full macroblock or a partition thereof)

  106.     against a proposed motion-compensated prediction of that block

  107.  */

  108. static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,

  109.                       const int size, const int h, int ref_index, int src_index,

  110.                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){

  111.     MotionEstContext * const c= &s->me;

  112.     const int stride= c->stride;

  113.     const int uvstride= c->uvstride;

  114.     const int qpel= flags&FLAG_QPEL;

  115.     const int chroma= flags&FLAG_CHROMA;

  116.     const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?

  117.     const int hx= subx + (x<<(1+qpel));

  118.     const int hy= suby + (y<<(1+qpel));

  119.     uint8_t * const * const ref= c->ref[ref_index];

  120.     uint8_t * const * const src= c->src[src_index];

  121.     int d;

  122.     //FIXME check chroma 4mv, (no crashes ...)

  123.     if(flags&FLAG_DIRECT){

  124.         assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));

  125.         if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){

  126.             const int time_pp= s->pp_time;

  127.             const int time_pb= s->pb_time;

  128.             const int mask= 2*qpel+1;

  129.             if(s->mv_type==MV_TYPE_8X8){

  130.                 int i;

  131.                 for(i=0; i<4; i++){

  132.                     int fx = c->direct_basis_mv[i][0] + hx;

  133.                     int fy = c->direct_basis_mv[i][1] + hy;

  134.                     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));

  135.                     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));

  136.                     int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));

  137.                     int bxy= (bx&mask) + ((by&mask)<<(qpel+1));

  138. 

  139.                     uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);

  140.                     if(qpel){

  141.                         c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);

  142.                         c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);

  143.                     }else{

  144.                         c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);

  145.                         c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);

  146.                     }

  147.                 }

  148.             }else{

  149.                 int fx = c->direct_basis_mv[0][0] + hx;

  150.                 int fy = c->direct_basis_mv[0][1] + hy;

  151.                 int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);

  152.                 int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);

  153.                 int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));

  154.                 int bxy= (bx&mask) + ((by&mask)<<(qpel+1));

  155. 

  156.                 if(qpel){

  157.                     c->qpel_put[1][fxy](c->temp               , ref[0] + (fx>>2) + (fy>>2)*stride               , stride);

  158.                     c->qpel_put[1][fxy](c->temp + 8           , ref[0] + (fx>>2) + (fy>>2)*stride + 8           , stride);

  159.                     c->qpel_put[1][fxy](c->temp     + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride     + 8*stride, stride);

  160.                     c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);

  161.                     c->qpel_avg[1][bxy](c->temp               , ref[8] + (bx>>2) + (by>>2)*stride               , stride);

  162.                     c->qpel_avg[1][bxy](c->temp + 8           , ref[8] + (bx>>2) + (by>>2)*stride + 8           , stride);

  163.                     c->qpel_avg[1][bxy](c->temp     + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride     + 8*stride, stride);

  164.                     c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);

  165.                 }else{

  166.                     assert((fx>>1) + 16*s->mb_x >= -16);

  167.                     assert((fy>>1) + 16*s->mb_y >= -16);

  168.                     assert((fx>>1) + 16*s->mb_x <= s->width);

  169.                     assert((fy>>1) + 16*s->mb_y <= s->height);

  170.                     assert((bx>>1) + 16*s->mb_x >= -16);

  171.                     assert((by>>1) + 16*s->mb_y >= -16);

  172.                     assert((bx>>1) + 16*s->mb_x <= s->width);

  173.                     assert((by>>1) + 16*s->mb_y <= s->height);

  174. 

  175.                     c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);

  176.                     c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);

  177.                 }

  178.             }

  179.             d = cmp_func(s, c->temp, src[0], stride, 16);

  180.         }else

  181.             d= 256*256*256*32;

  182.     }else{

  183.         int uvdxy;              /* no, it might not be used uninitialized */

  184.         if(dxy){

  185.             if(qpel){

  186.                 c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)

  187.                 if(chroma){

  188.                     int cx= hx/2;

  189.                     int cy= hy/2;

  190.                     cx= (cx>>1)|(cx&1);

  191.                     cy= (cy>>1)|(cy&1);

  192.                     uvdxy= (cx&1) + 2*(cy&1);

  193.                     //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264

  194.                 }

  195.             }else{

  196.                 c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);

  197.                 if(chroma)

  198.                     uvdxy= dxy | (x&1) | (2*(y&1));

  199.             }

  200.             d = cmp_func(s, c->temp, src[0], stride, h);

  201.         }else{

  202.             d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);

  203.             if(chroma)

  204.                 uvdxy= (x&1) + 2*(y&1);

  205.         }

  206.         if(chroma){

  207.             uint8_t * const uvtemp= c->temp + 16*stride;

  208.             c->hpel_put[size+1][uvdxy](uvtemp  , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);

  209.             c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);

  210.             d += chroma_cmp_func(s, uvtemp  , src[1], uvstride, h>>1);

  211.             d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);

  212.         }

  213.     }

  214. #if 0

  215.     if(full_pel){

  216.         const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);

  217.         score_map[index]= d;

  218.     }

  219. 

  220.     d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;

  221. #endif

  222.     return d;

  223. }

  224. 

  225. #include "motion_est_template.c"

  226. 

  227. static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){

  228.     return 0;

  229. }

  230. 

  231. static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){

  232. }

  233. 

  234. int ff_init_me(MpegEncContext *s){

  235.     MotionEstContext * const c= &s->me;

  236.     int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);

  237.     int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);

  238. 

  239.     if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){

  240.         av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");

  241.         return -1;

  242.     }

  243. 

  244.     c->avctx= s->avctx;

  245. 

  246.     if(cache_size < 2*dia_size && !c->stride){

  247.         av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");

  248.     }

  249. 

  250.     ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);

  251.     ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);

  252.     ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);

  253.     ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);

  254. 

  255.     c->flags    = get_flags(c, 0, c->avctx->me_cmp    &FF_CMP_CHROMA);

  256.     c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);

  257.     c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp    &FF_CMP_CHROMA);

  258. 

  259. /*FIXME s->no_rounding b_type*/

  260.     if(s->flags&CODEC_FLAG_QPEL){

  261.         c->sub_motion_search= qpel_motion_search;

  262.         c->qpel_avg= s->dsp.avg_qpel_pixels_tab;

  263.         if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;

  264.         else               c->qpel_put= s->dsp.put_qpel_pixels_tab;

  265.     }else{

  266.         if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)

  267.             c->sub_motion_search= hpel_motion_search;

  268.         else if(   c->avctx->me_sub_cmp == FF_CMP_SAD

  269.                 && c->avctx->    me_cmp == FF_CMP_SAD

  270.                 && c->avctx->    mb_cmp == FF_CMP_SAD)

  271.             c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles

  272.         else

  273.             c->sub_motion_search= hpel_motion_search;

  274.     }

  275.     c->hpel_avg= s->dsp.avg_pixels_tab;

  276.     if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;

  277.     else               c->hpel_put= s->dsp.put_pixels_tab;

  278. 

  279.     if(s->linesize){

  280.         c->stride  = s->linesize;

  281.         c->uvstride= s->uvlinesize;

  282.     }else{

  283.         c->stride  = 16*s->mb_width + 32;

  284.         c->uvstride=  8*s->mb_width + 16;

  285.     }

  286. 

  287.     /* 8x8 fullpel search would need a 4x4 chroma compare, which we do

  288.      * not have yet, and even if we had, the motion estimation code

  289.      * does not expect it. */

  290.     if(s->codec_id != CODEC_ID_SNOW){

  291.         if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){

  292.             s->dsp.me_cmp[2]= zero_cmp;

  293.         }

  294.         if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){

  295.             s->dsp.me_sub_cmp[2]= zero_cmp;

  296.         }

  297.         c->hpel_put[2][0]= c->hpel_put[2][1]=

  298.         c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;

  299.     }

  300. 

  301.     if(s->codec_id == CODEC_ID_H261){

  302.         c->sub_motion_search= no_sub_motion_search;

  303.     }

  304. 

  305.     c->temp= c->scratchpad;

  306. 

  307.     return 0;

  308. }

  309. 

  310. #if 0

  311. static int pix_dev(uint8_t * pix, int line_size, int mean)

  312. {

  313.     int s, i, j;

  314. 

  315.     s = 0;

  316.     for (i = 0; i < 16; i++) {

  317.         for (j = 0; j < 16; j += 8) {

  318.             s += FFABS(pix[0]-mean);

  319.             s += FFABS(pix[1]-mean);

  320.             s += FFABS(pix[2]-mean);

  321.             s += FFABS(pix[3]-mean);

  322.             s += FFABS(pix[4]-mean);

  323.             s += FFABS(pix[5]-mean);

  324.             s += FFABS(pix[6]-mean);

  325.             s += FFABS(pix[7]-mean);

  326.             pix += 8;

  327.         }

  328.         pix += line_size - 16;

  329.     }

  330.     return s;

  331. }

  332. #endif

  333. 

  334. static inline void no_motion_search(MpegEncContext * s,

  335.                                     int *mx_ptr, int *my_ptr)

  336. {

  337.     *mx_ptr = 16 * s->mb_x;

  338.     *my_ptr = 16 * s->mb_y;

  339. }

  340. 

  341. #define Z_THRESHOLD 256

  342. 

  343. #define CHECK_SAD_HALF_MV(suffix, x, y) \

  344. {\

  345.     d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\

  346.     d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\

  347.     COPY3_IF_LT(dminh, d, dx, x, dy, y)\

  348. }

  349. 

  350. static inline int sad_hpel_motion_search(MpegEncContext * s,

  351.                                   int *mx_ptr, int *my_ptr, int dmin,

  352.                                   int src_index, int ref_index,

  353.                                   int size, int h)

  354. {

  355.     MotionEstContext * const c= &s->me;

  356.     const int penalty_factor= c->sub_penalty_factor;

  357.     int mx, my, dminh;

  358.     uint8_t *pix, *ptr;

  359.     int stride= c->stride;

  360.     const int flags= c->sub_flags;

  361.     LOAD_COMMON

  362. 

  363.     assert(flags == 0);

  364. 

  365.     if(c->skip){

  366. //    printf("S");

  367.         *mx_ptr = 0;

  368.         *my_ptr = 0;

  369.         return dmin;

  370.     }

  371. //    printf("N");

  372. 

  373.     pix = c->src[src_index][0];

  374. 

  375.     mx = *mx_ptr;

  376.     my = *my_ptr;

  377.     ptr = c->ref[ref_index][0] + (my * stride) + mx;

  378. 

  379.     dminh = dmin;

  380. 

  381.     if (mx > xmin && mx < xmax &&

  382.         my > ymin && my < ymax) {

  383.         int dx=0, dy=0;

  384.         int d, pen_x, pen_y;

  385.         const int index= (my<<ME_MAP_SHIFT) + mx;

  386.         const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];

  387.         const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];

  388.         const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];

  389.         const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];

  390.         mx<<=1;

  391.         my<<=1;

  392. 

  393. 

  394.         pen_x= pred_x + mx;

  395.         pen_y= pred_y + my;

  396. 

  397.         ptr-= stride;

  398.         if(t<=b){

  399.             CHECK_SAD_HALF_MV(y2 , 0, -1)

  400.             if(l<=r){

  401.                 CHECK_SAD_HALF_MV(xy2, -1, -1)

  402.                 if(t+r<=b+l){

  403.                     CHECK_SAD_HALF_MV(xy2, +1, -1)

  404.                     ptr+= stride;

  405.                 }else{

  406.                     ptr+= stride;

  407.                     CHECK_SAD_HALF_MV(xy2, -1, +1)

  408.                 }

  409.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

  410.             }else{

  411.                 CHECK_SAD_HALF_MV(xy2, +1, -1)

  412.                 if(t+l<=b+r){

  413.                     CHECK_SAD_HALF_MV(xy2, -1, -1)

  414.                     ptr+= stride;

  415.                 }else{

  416.                     ptr+= stride;

  417.                     CHECK_SAD_HALF_MV(xy2, +1, +1)

  418.                 }

  419.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

  420.             }

  421.         }else{

  422.             if(l<=r){

  423.                 if(t+l<=b+r){

  424.                     CHECK_SAD_HALF_MV(xy2, -1, -1)

  425.                     ptr+= stride;

  426.                 }else{

  427.                     ptr+= stride;

  428.                     CHECK_SAD_HALF_MV(xy2, +1, +1)

  429.                 }

  430.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

  431.                 CHECK_SAD_HALF_MV(xy2, -1, +1)

  432.             }else{

  433.                 if(t+r<=b+l){

  434.                     CHECK_SAD_HALF_MV(xy2, +1, -1)

  435.                     ptr+= stride;

  436.                 }else{

  437.                     ptr+= stride;

  438.                     CHECK_SAD_HALF_MV(xy2, -1, +1)

  439.                 }

  440.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

  441.                 CHECK_SAD_HALF_MV(xy2, +1, +1)

  442.             }

  443.             CHECK_SAD_HALF_MV(y2 ,  0, +1)

  444.         }

  445.         mx+=dx;

  446.         my+=dy;

  447. 

  448.     }else{

  449.         mx<<=1;

  450.         my<<=1;

  451.     }

  452. 

  453.     *mx_ptr = mx;

  454.     *my_ptr = my;

  455.     return dminh;

  456. }

  457. 

  458. static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)

  459. {

  460.     const int xy= s->mb_x + s->mb_y*s->mb_stride;

  461. 

  462.     s->p_mv_table[xy][0] = mx;

  463.     s->p_mv_table[xy][1] = my;

  464. 

  465.     /* has already been set to the 4 MV if 4MV is done */

  466.     if(mv4){

  467.         int mot_xy= s->block_index[0];

  468. 

  469.         s->current_picture.motion_val[0][mot_xy  ][0]= mx;

  470.         s->current_picture.motion_val[0][mot_xy  ][1]= my;

  471.         s->current_picture.motion_val[0][mot_xy+1][0]= mx;

  472.         s->current_picture.motion_val[0][mot_xy+1][1]= my;

  473. 

  474.         mot_xy += s->b8_stride;

  475.         s->current_picture.motion_val[0][mot_xy  ][0]= mx;

  476.         s->current_picture.motion_val[0][mot_xy  ][1]= my;

  477.         s->current_picture.motion_val[0][mot_xy+1][0]= mx;

  478.         s->current_picture.motion_val[0][mot_xy+1][1]= my;

  479.     }

  480. }

  481. 

  482. /**

  483.  * get fullpel ME search limits.

  484.  */

  485. static inline void get_limits(MpegEncContext *s, int x, int y)

  486. {

  487.     MotionEstContext * const c= &s->me;

  488.     int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));

  489. /*

  490.     if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;

  491.     else                   c->range= 16;

  492. */

  493.     if (s->unrestricted_mv) {

  494.         c->xmin = - x - 16;

  495.         c->ymin = - y - 16;

  496.         c->xmax = - x + s->mb_width *16;

  497.         c->ymax = - y + s->mb_height*16;

  498.     } else if (s->out_format == FMT_H261){

  499.         // Search range of H261 is different from other codec standards

  500.         c->xmin = (x > 15) ? - 15 : 0;

  501.         c->ymin = (y > 15) ? - 15 : 0;

  502.         c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;

  503.         c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;

  504.     } else {

  505.         c->xmin = - x;

  506.         c->ymin = - y;

  507.         c->xmax = - x + s->mb_width *16 - 16;

  508.         c->ymax = - y + s->mb_height*16 - 16;

  509.     }

  510.     if(range){

  511.         c->xmin = FFMAX(c->xmin,-range);

  512.         c->xmax = FFMIN(c->xmax, range);

  513.         c->ymin = FFMAX(c->ymin,-range);

  514.         c->ymax = FFMIN(c->ymax, range);

  515.     }

  516. }

  517. 

  518. static inline void init_mv4_ref(MotionEstContext *c){

  519.     const int stride= c->stride;

  520. 

  521.     c->ref[1][0] = c->ref[0][0] + 8;

  522.     c->ref[2][0] = c->ref[0][0] + 8*stride;

  523.     c->ref[3][0] = c->ref[2][0] + 8;

  524.     c->src[1][0] = c->src[0][0] + 8;

  525.     c->src[2][0] = c->src[0][0] + 8*stride;

  526.     c->src[3][0] = c->src[2][0] + 8;

  527. }

  528. 

  529. static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)

  530. {

  531.     MotionEstContext * const c= &s->me;

  532.     const int size= 1;

  533.     const int h=8;

  534.     int block;

  535.     int P[10][2];

  536.     int dmin_sum=0, mx4_sum=0, my4_sum=0;

  537.     int same=1;

  538.     const int stride= c->stride;

  539.     uint8_t *mv_penalty= c->current_mv_penalty;

  540. 

  541.     init_mv4_ref(c);

  542. 

  543.     for(block=0; block<4; block++){

  544.         int mx4, my4;

  545.         int pred_x4, pred_y4;

  546.         int dmin4;

  547.         static const int off[4]= {2, 1, 1, -1};

  548.         const int mot_stride = s->b8_stride;

  549.         const int mot_xy = s->block_index[block];

  550. 

  551.         P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];

  552.         P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];

  553. 

  554.         if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);

  555. 

  556.         /* special case for first line */

  557.         if (s->first_slice_line && block<2) {

  558.             c->pred_x= pred_x4= P_LEFT[0];

  559.             c->pred_y= pred_y4= P_LEFT[1];

  560.         } else {

  561.             P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][0];

  562.             P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][1];

  563.             P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];

  564.             P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];

  565.             if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);

  566.             if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);

  567.             if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);

  568.             if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);

  569. 

  570.             P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  571.             P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  572. 

  573.             c->pred_x= pred_x4 = P_MEDIAN[0];

  574.             c->pred_y= pred_y4 = P_MEDIAN[1];

  575.         }

  576.         P_MV1[0]= mx;

  577.         P_MV1[1]= my;

  578. 

  579.         dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);

  580. 

  581.         dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);

  582. 

  583.         if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){

  584.             int dxy;

  585.             const int offset= ((block&1) + (block>>1)*stride)*8;

  586.             uint8_t *dest_y = c->scratchpad + offset;

  587.             if(s->quarter_sample){

  588.                 uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;

  589.                 dxy = ((my4 & 3) << 2) | (mx4 & 3);

  590. 

  591.                 if(s->no_rounding)

  592.                     s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , stride);

  593.                 else

  594.                     s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , stride);

  595.             }else{

  596.                 uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;

  597.                 dxy = ((my4 & 1) << 1) | (mx4 & 1);

  598. 

  599.                 if(s->no_rounding)

  600.                     s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , stride, h);

  601.                 else

  602.                     s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , stride, h);

  603.             }

  604.             dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;

  605.         }else

  606.             dmin_sum+= dmin4;

  607. 

  608.         if(s->quarter_sample){

  609.             mx4_sum+= mx4/2;

  610.             my4_sum+= my4/2;

  611.         }else{

  612.             mx4_sum+= mx4;

  613.             my4_sum+= my4;

  614.         }

  615. 

  616.         s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;

  617.         s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;

  618. 

  619.         if(mx4 != mx || my4 != my) same=0;

  620.     }

  621. 

  622.     if(same)

  623.         return INT_MAX;

  624. 

  625.     if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){

  626.         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);

  627.     }

  628. 

  629.     if(c->avctx->mb_cmp&FF_CMP_CHROMA){

  630.         int dxy;

  631.         int mx, my;

  632.         int offset;

  633. 

  634.         mx= ff_h263_round_chroma(mx4_sum);

  635.         my= ff_h263_round_chroma(my4_sum);

  636.         dxy = ((my & 1) << 1) | (mx & 1);

  637. 

  638.         offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;

  639. 

  640.         if(s->no_rounding){

  641.             s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);

  642.             s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);

  643.         }else{

  644.             s->dsp.put_pixels_tab       [1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);

  645.             s->dsp.put_pixels_tab       [1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);

  646.         }

  647. 

  648.         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);

  649.         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);

  650.     }

  651. 

  652.     c->pred_x= mx;

  653.     c->pred_y= my;

  654. 

  655.     switch(c->avctx->mb_cmp&0xFF){

  656.     /*case FF_CMP_SSE:

  657.         return dmin_sum+ 32*s->qscale*s->qscale;*/

  658.     case FF_CMP_RD:

  659.         return dmin_sum;

  660.     default:

  661.         return dmin_sum+ 11*c->mb_penalty_factor;

  662.     }

  663. }

  664. 

  665. static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){

  666.     MotionEstContext * const c= &s->me;

  667. 

  668.     c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;

  669.     c->src[1][0] = c->src[0][0] + s->linesize;

  670.     if(c->flags & FLAG_CHROMA){

  671.         c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;

  672.         c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;

  673.         c->src[1][1] = c->src[0][1] + s->uvlinesize;

  674.         c->src[1][2] = c->src[0][2] + s->uvlinesize;

  675.     }

  676. }

  677. 

  678. static int interlaced_search(MpegEncContext *s, int ref_index,

  679.                              int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)

  680. {

  681.     MotionEstContext * const c= &s->me;

  682.     const int size=0;

  683.     const int h=8;

  684.     int block;

  685.     int P[10][2];

  686.     uint8_t * const mv_penalty= c->current_mv_penalty;

  687.     int same=1;

  688.     const int stride= 2*s->linesize;

  689.     int dmin_sum= 0;

  690.     const int mot_stride= s->mb_stride;

  691.     const int xy= s->mb_x + s->mb_y*mot_stride;

  692. 

  693.     c->ymin>>=1;

  694.     c->ymax>>=1;

  695.     c->stride<<=1;

  696.     c->uvstride<<=1;

  697.     init_interlaced_ref(s, ref_index);

  698. 

  699.     for(block=0; block<2; block++){

  700.         int field_select;

  701.         int best_dmin= INT_MAX;

  702.         int best_field= -1;

  703. 

  704.         for(field_select=0; field_select<2; field_select++){

  705.             int dmin, mx_i, my_i;

  706.             int16_t (*mv_table)[2]= mv_tables[block][field_select];

  707. 

  708.             if(user_field_select){

  709.                 assert(field_select==0 || field_select==1);

  710.                 assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);

  711.                 if(field_select_tables[block][xy] != field_select)

  712.                     continue;

  713.             }

  714. 

  715.             P_LEFT[0] = mv_table[xy - 1][0];

  716.             P_LEFT[1] = mv_table[xy - 1][1];

  717.             if(P_LEFT[0]       > (c->xmax<<1)) P_LEFT[0]       = (c->xmax<<1);

  718. 

  719.             c->pred_x= P_LEFT[0];

  720.             c->pred_y= P_LEFT[1];

  721. 

  722.             if(!s->first_slice_line){

  723.                 P_TOP[0]      = mv_table[xy - mot_stride][0];

  724.                 P_TOP[1]      = mv_table[xy - mot_stride][1];

  725.                 P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];

  726.                 P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];

  727.                 if(P_TOP[1]      > (c->ymax<<1)) P_TOP[1]     = (c->ymax<<1);

  728.                 if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);

  729.                 if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);

  730.                 if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);

  731. 

  732.                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  733.                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  734.             }

  735.             P_MV1[0]= mx; //FIXME not correct if block != field_select

  736.             P_MV1[1]= my / 2;

  737. 

  738.             dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);

  739. 

  740.             dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);

  741. 

  742.             mv_table[xy][0]= mx_i;

  743.             mv_table[xy][1]= my_i;

  744. 

  745.             if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){

  746.                 int dxy;

  747. 

  748.                 //FIXME chroma ME

  749.                 uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;

  750.                 dxy = ((my_i & 1) << 1) | (mx_i & 1);

  751. 

  752.                 if(s->no_rounding){

  753.                     s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref    , stride, h);

  754.                 }else{

  755.                     s->dsp.put_pixels_tab       [size][dxy](c->scratchpad, ref    , stride, h);

  756.                 }

  757.                 dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);

  758.                 dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;

  759.             }else

  760.                 dmin+= c->mb_penalty_factor; //field_select bits

  761. 

  762.             dmin += field_select != block; //slightly prefer same field

  763. 

  764.             if(dmin < best_dmin){

  765.                 best_dmin= dmin;

  766.                 best_field= field_select;

  767.             }

  768.         }

  769.         {

  770.             int16_t (*mv_table)[2]= mv_tables[block][best_field];

  771. 

  772.             if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all

  773.             if(mv_table[xy][1]&1) same=0;

  774.             if(mv_table[xy][1]*2 != my) same=0;

  775.             if(best_field != block) same=0;

  776.         }

  777. 

  778.         field_select_tables[block][xy]= best_field;

  779.         dmin_sum += best_dmin;

  780.     }

  781. 

  782.     c->ymin<<=1;

  783.     c->ymax<<=1;

  784.     c->stride>>=1;

  785.     c->uvstride>>=1;

  786. 

  787.     if(same)

  788.         return INT_MAX;

  789. 

  790.     switch(c->avctx->mb_cmp&0xFF){

  791.     /*case FF_CMP_SSE:

  792.         return dmin_sum+ 32*s->qscale*s->qscale;*/

  793.     case FF_CMP_RD:

  794.         return dmin_sum;

  795.     default:

  796.         return dmin_sum+ 11*c->mb_penalty_factor;

  797.     }

  798. }

  799. 

  800. static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){

  801.     int ymax= s->me.ymax>>interlaced;

  802.     int ymin= s->me.ymin>>interlaced;

  803. 

  804.     if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;

  805.     if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;

  806.     if(mv[1] <       ymin) mv[1] =       ymin;

  807.     if(mv[1] >       ymax) mv[1] =       ymax;

  808. }

  809. 

  810. static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){

  811.     MotionEstContext * const c= &s->me;

  812.     Picture *p= s->current_picture_ptr;

  813.     int mb_xy= mb_x + mb_y*s->mb_stride;

  814.     int xy= 2*mb_x + 2*mb_y*s->b8_stride;

  815.     int mb_type= s->current_picture.mb_type[mb_xy];

  816.     int flags= c->flags;

  817.     int shift= (flags&FLAG_QPEL) + 1;

  818.     int mask= (1<<shift)-1;

  819.     int x, y, i;

  820.     int d=0;

  821.     me_cmp_func cmpf= s->dsp.sse[0];

  822.     me_cmp_func chroma_cmpf= s->dsp.sse[1];

  823. 

  824.     if(p_type && USES_LIST(mb_type, 1)){

  825.         av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");

  826.         return INT_MAX/2;

  827.     }

  828.     assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));

  829. 

  830.     for(i=0; i<4; i++){

  831.         int xy= s->block_index[i];

  832.         clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));

  833.         clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));

  834.     }

  835. 

  836.     if(IS_INTERLACED(mb_type)){

  837.         int xy2= xy  + s->b8_stride;

  838.         s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;

  839.         c->stride<<=1;

  840.         c->uvstride<<=1;

  841. 

  842.         if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){

  843.             av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");

  844.             return INT_MAX/2;

  845.         }

  846. 

  847.         if(USES_LIST(mb_type, 0)){

  848.             int field_select0= p->ref_index[0][xy ];

  849.             int field_select1= p->ref_index[0][xy2];

  850.             assert(field_select0==0 ||field_select0==1);

  851.             assert(field_select1==0 ||field_select1==1);

  852.             init_interlaced_ref(s, 0);

  853. 

  854.             if(p_type){

  855.                 s->p_field_select_table[0][mb_xy]= field_select0;

  856.                 s->p_field_select_table[1][mb_xy]= field_select1;

  857.                 *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];

  858.                 *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];

  859.                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;

  860.             }else{

  861.                 s->b_field_select_table[0][0][mb_xy]= field_select0;

  862.                 s->b_field_select_table[0][1][mb_xy]= field_select1;

  863.                 *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];

  864.                 *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];

  865.                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;

  866.             }

  867. 

  868.             x= p->motion_val[0][xy ][0];

  869.             y= p->motion_val[0][xy ][1];

  870.             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);

  871.             x= p->motion_val[0][xy2][0];

  872.             y= p->motion_val[0][xy2][1];

  873.             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);

  874.         }

  875.         if(USES_LIST(mb_type, 1)){

  876.             int field_select0= p->ref_index[1][xy ];

  877.             int field_select1= p->ref_index[1][xy2];

  878.             assert(field_select0==0 ||field_select0==1);

  879.             assert(field_select1==0 ||field_select1==1);

  880.             init_interlaced_ref(s, 2);

  881. 

  882.             s->b_field_select_table[1][0][mb_xy]= field_select0;

  883.             s->b_field_select_table[1][1][mb_xy]= field_select1;

  884.             *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];

  885.             *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];

  886.             if(USES_LIST(mb_type, 0)){

  887.                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;

  888.             }else{

  889.                 s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;

  890.             }

  891. 

  892.             x= p->motion_val[1][xy ][0];

  893.             y= p->motion_val[1][xy ][1];

  894.             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);

  895.             x= p->motion_val[1][xy2][0];

  896.             y= p->motion_val[1][xy2][1];

  897.             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);

  898.             //FIXME bidir scores

  899.         }

  900.         c->stride>>=1;

  901.         c->uvstride>>=1;

  902.     }else if(IS_8X8(mb_type)){

  903.         if(!(s->flags & CODEC_FLAG_4MV)){

  904.             av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");

  905.             return INT_MAX/2;

  906.         }

  907.         cmpf= s->dsp.sse[1];

  908.         chroma_cmpf= s->dsp.sse[1];

  909.         init_mv4_ref(c);

  910.         for(i=0; i<4; i++){

  911.             xy= s->block_index[i];

  912.             x= p->motion_val[0][xy][0];

  913.             y= p->motion_val[0][xy][1];

  914.             d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);

  915.         }

  916.         s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;

  917.     }else{

  918.         if(USES_LIST(mb_type, 0)){

  919.             if(p_type){

  920.                 *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];

  921.                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;

  922.             }else if(USES_LIST(mb_type, 1)){

  923.                 *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];

  924.                 *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];

  925.                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;

  926.             }else{

  927.                 *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];

  928.                 s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;

  929.             }

  930.             x= p->motion_val[0][xy][0];

  931.             y= p->motion_val[0][xy][1];

  932.             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);

  933.         }else if(USES_LIST(mb_type, 1)){

  934.             *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];

  935.             s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;

  936. 

  937.             x= p->motion_val[1][xy][0];

  938.             y= p->motion_val[1][xy][1];

  939.             d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);

  940.         }else

  941.             s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;

  942.     }

  943.     return d;

  944. }

  945. 

  946. void ff_estimate_p_frame_motion(MpegEncContext * s,

  947.                                 int mb_x, int mb_y)

  948. {

  949.     MotionEstContext * const c= &s->me;

  950.     uint8_t *pix, *ppix;

  951.     int sum, mx, my, dmin;

  952.     int varc;            ///< the variance of the block (sum of squared (p[y][x]-average))

  953.     int vard;            ///< sum of squared differences with the estimated motion vector

  954.     int P[10][2];

  955.     const int shift= 1+s->quarter_sample;

  956.     int mb_type=0;

  957.     Picture * const pic= &s->current_picture;

  958. 

  959.     init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);

  960. 

  961.     assert(s->quarter_sample==0 || s->quarter_sample==1);

  962.     assert(s->linesize == c->stride);

  963.     assert(s->uvlinesize == c->uvstride);

  964. 

  965.     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);

  966.     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);

  967.     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);

  968.     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

  969. 

  970.     get_limits(s, 16*mb_x, 16*mb_y);

  971.     c->skip=0;

  972. 

  973.     /* intra / predictive decision */

  974.     pix = c->src[0][0];

  975.     sum = s->dsp.pix_sum(pix, s->linesize);

  976.     varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;

  977. 

  978.     pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;

  979.     pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;

  980.     c->mb_var_sum_temp += (varc+128)>>8;

  981. 

  982.     if(c->avctx->me_threshold){

  983.         vard= check_input_motion(s, mb_x, mb_y, 1);

  984. 

  985.         if((vard+128)>>8 < c->avctx->me_threshold){

  986.             int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);

  987.             int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;

  988.             pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;

  989.             c->mc_mb_var_sum_temp += (vard+128)>>8;

  990.             c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);

  991.             return;

  992.         }

  993.         if((vard+128)>>8 < c->avctx->mb_threshold)

  994.             mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];

  995.     }

  996. 

  997.     switch(s->me_method) {

  998.     case ME_ZERO:

  999.     default:

  1000.         no_motion_search(s, &mx, &my);

  1001.         mx-= mb_x*16;

  1002.         my-= mb_y*16;

  1003.         dmin = 0;

  1004.         break;

  1005.     case ME_X1:

  1006.     case ME_EPZS:

  1007.        {

  1008.             const int mot_stride = s->b8_stride;

  1009.             const int mot_xy = s->block_index[0];

  1010. 

  1011.             P_LEFT[0]       = s->current_picture.motion_val[0][mot_xy - 1][0];

  1012.             P_LEFT[1]       = s->current_picture.motion_val[0][mot_xy - 1][1];

  1013. 

  1014.             if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);

  1015. 

  1016.             if(!s->first_slice_line) {

  1017.                 P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][0];

  1018.                 P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][1];

  1019.                 P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];

  1020.                 P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];

  1021.                 if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);

  1022.                 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);

  1023.                 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);

  1024. 

  1025.                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1026.                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1027. 

  1028.                 if(s->out_format == FMT_H263){

  1029.                     c->pred_x = P_MEDIAN[0];

  1030.                     c->pred_y = P_MEDIAN[1];

  1031.                 }else { /* mpeg1 at least */

  1032.                     c->pred_x= P_LEFT[0];

  1033.                     c->pred_y= P_LEFT[1];

  1034.                 }

  1035.             }else{

  1036.                 c->pred_x= P_LEFT[0];

  1037.                 c->pred_y= P_LEFT[1];

  1038.             }

  1039. 

  1040.         }

  1041.         dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);

  1042. 

  1043.         break;

  1044.     }

  1045. 

  1046.     /* At this point (mx,my) are full-pell and the relative displacement */

  1047.     ppix = c->ref[0][0] + (my * s->linesize) + mx;

  1048. 

  1049.     vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);

  1050. 

  1051.     pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;

  1052. //    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;

  1053.     c->mc_mb_var_sum_temp += (vard+128)>>8;

  1054. 

  1055. #if 0

  1056.     printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",

  1057.            varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);

  1058. #endif

  1059.     if(mb_type){

  1060.         int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);

  1061.         int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;

  1062.         c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);

  1063. 

  1064.         if(mb_type == CANDIDATE_MB_TYPE_INTER){

  1065.             c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1066.             set_p_mv_tables(s, mx, my, 1);

  1067.         }else{

  1068.             mx <<=shift;

  1069.             my <<=shift;

  1070.         }

  1071.         if(mb_type == CANDIDATE_MB_TYPE_INTER4V){

  1072.             h263_mv4_search(s, mx, my, shift);

  1073. 

  1074.             set_p_mv_tables(s, mx, my, 0);

  1075.         }

  1076.         if(mb_type == CANDIDATE_MB_TYPE_INTER_I){

  1077.             interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);

  1078.         }

  1079.     }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){

  1080.         int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);

  1081.         int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;

  1082.         c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);

  1083. 

  1084.         if (vard*2 + 200*256 > varc)

  1085.             mb_type|= CANDIDATE_MB_TYPE_INTRA;

  1086.         if (varc*2 + 200*256 > vard || s->qscale > 24){

  1087. //        if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){

  1088.             mb_type|= CANDIDATE_MB_TYPE_INTER;

  1089.             c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1090.             if(s->flags&CODEC_FLAG_MV0)

  1091.                 if(mx || my)

  1092.                     mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference

  1093.         }else{

  1094.             mx <<=shift;

  1095.             my <<=shift;

  1096.         }

  1097.         if((s->flags&CODEC_FLAG_4MV)

  1098.            && !c->skip && varc>50<<8 && vard>10<<8){

  1099.             if(h263_mv4_search(s, mx, my, shift) < INT_MAX)

  1100.                 mb_type|=CANDIDATE_MB_TYPE_INTER4V;

  1101. 

  1102.             set_p_mv_tables(s, mx, my, 0);

  1103.         }else

  1104.             set_p_mv_tables(s, mx, my, 1);

  1105.         if((s->flags&CODEC_FLAG_INTERLACED_ME)

  1106.            && !c->skip){ //FIXME varc/d checks

  1107.             if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)

  1108.                 mb_type |= CANDIDATE_MB_TYPE_INTER_I;

  1109.         }

  1110.     }else{

  1111.         int intra_score, i;

  1112.         mb_type= CANDIDATE_MB_TYPE_INTER;

  1113. 

  1114.         dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1115.         if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)

  1116.             dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);

  1117. 

  1118.         if((s->flags&CODEC_FLAG_4MV)

  1119.            && !c->skip && varc>50<<8 && vard>10<<8){

  1120.             int dmin4= h263_mv4_search(s, mx, my, shift);

  1121.             if(dmin4 < dmin){

  1122.                 mb_type= CANDIDATE_MB_TYPE_INTER4V;

  1123.                 dmin=dmin4;

  1124.             }

  1125.         }

  1126.         if((s->flags&CODEC_FLAG_INTERLACED_ME)

  1127.            && !c->skip){ //FIXME varc/d checks

  1128.             int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);

  1129.             if(dmin_i < dmin){

  1130.                 mb_type = CANDIDATE_MB_TYPE_INTER_I;

  1131.                 dmin= dmin_i;

  1132.             }

  1133.         }

  1134. 

  1135. //        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;

  1136.         set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);

  1137. 

  1138.         /* get intra luma score */

  1139.         if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){

  1140.             intra_score= varc - 500;

  1141.         }else{

  1142.             int mean= (sum+128)>>8;

  1143.             mean*= 0x01010101;

  1144. 

  1145.             for(i=0; i<16; i++){

  1146.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;

  1147.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;

  1148.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;

  1149.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;

  1150.             }

  1151. 

  1152.             intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);

  1153.         }

  1154. #if 0 //FIXME

  1155.         /* get chroma score */

  1156.         if(c->avctx->mb_cmp&FF_CMP_CHROMA){

  1157.             for(i=1; i<3; i++){

  1158.                 uint8_t *dest_c;

  1159.                 int mean;

  1160. 

  1161.                 if(s->out_format == FMT_H263){

  1162.                     mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)

  1163.                 }else{

  1164.                     mean= (s->last_dc[i] + 4)>>3;

  1165.                 }

  1166.                 dest_c = s->new_picture.data[i] + (mb_y * 8  * (s->uvlinesize)) + mb_x * 8;

  1167. 

  1168.                 mean*= 0x01010101;

  1169.                 for(i=0; i<8; i++){

  1170.                     *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;

  1171.                     *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;

  1172.                 }

  1173. 

  1174.                 intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);

  1175.             }

  1176.         }

  1177. #endif

  1178.         intra_score += c->mb_penalty_factor*16;

  1179. 

  1180.         if(intra_score < dmin){

  1181.             mb_type= CANDIDATE_MB_TYPE_INTRA;

  1182.             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup

  1183.         }else

  1184.             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;

  1185. 

  1186.         {

  1187.             int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);

  1188.             int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;

  1189.             c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);

  1190.         }

  1191.     }

  1192. 

  1193.     s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;

  1194. }

  1195. 

  1196. int ff_pre_estimate_p_frame_motion(MpegEncContext * s,

  1197.                                     int mb_x, int mb_y)

  1198. {

  1199.     MotionEstContext * const c= &s->me;

  1200.     int mx, my, dmin;

  1201.     int P[10][2];

  1202.     const int shift= 1+s->quarter_sample;

  1203.     const int xy= mb_x + mb_y*s->mb_stride;

  1204.     init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);

  1205. 

  1206.     assert(s->quarter_sample==0 || s->quarter_sample==1);

  1207. 

  1208.     c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);

  1209.     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

  1210. 

  1211.     get_limits(s, 16*mb_x, 16*mb_y);

  1212.     c->skip=0;

  1213. 

  1214.     P_LEFT[0]       = s->p_mv_table[xy + 1][0];

  1215.     P_LEFT[1]       = s->p_mv_table[xy + 1][1];

  1216. 

  1217.     if(P_LEFT[0]       < (c->xmin<<shift)) P_LEFT[0]       = (c->xmin<<shift);

  1218. 

  1219.     /* special case for first line */

  1220.     if (s->first_slice_line) {

  1221.         c->pred_x= P_LEFT[0];

  1222.         c->pred_y= P_LEFT[1];

  1223.         P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=

  1224.         P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME

  1225.     } else {

  1226.         P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];

  1227.         P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];

  1228.         P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];

  1229.         P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];

  1230.         if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);

  1231.         if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);

  1232.         if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);

  1233. 

  1234.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1235.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1236. 

  1237.         c->pred_x = P_MEDIAN[0];

  1238.         c->pred_y = P_MEDIAN[1];

  1239.     }

  1240. 

  1241.     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);

  1242. 

  1243.     s->p_mv_table[xy][0] = mx<<shift;

  1244.     s->p_mv_table[xy][1] = my<<shift;

  1245. 

  1246.     return dmin;

  1247. }

  1248. 

  1249. static int ff_estimate_motion_b(MpegEncContext * s,

  1250.                        int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)

  1251. {

  1252.     MotionEstContext * const c= &s->me;

  1253.     int mx, my, dmin;

  1254.     int P[10][2];

  1255.     const int shift= 1+s->quarter_sample;

  1256.     const int mot_stride = s->mb_stride;

  1257.     const int mot_xy = mb_y*mot_stride + mb_x;

  1258.     uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;

  1259.     int mv_scale;

  1260. 

  1261.     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);

  1262.     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);

  1263.     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);

  1264.     c->current_mv_penalty= mv_penalty;

  1265. 

  1266.     get_limits(s, 16*mb_x, 16*mb_y);

  1267. 

  1268.     switch(s->me_method) {

  1269.     case ME_ZERO:

  1270.     default:

  1271.         no_motion_search(s, &mx, &my);

  1272.         dmin = 0;

  1273.         mx-= mb_x*16;

  1274.         my-= mb_y*16;

  1275.         break;

  1276.     case ME_X1:

  1277.     case ME_EPZS:

  1278.        {

  1279.             P_LEFT[0]        = mv_table[mot_xy - 1][0];

  1280.             P_LEFT[1]        = mv_table[mot_xy - 1][1];

  1281. 

  1282.             if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);

  1283. 

  1284.             /* special case for first line */

  1285.             if (!s->first_slice_line) {

  1286.                 P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];

  1287.                 P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];

  1288.                 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];

  1289.                 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];

  1290.                 if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);

  1291.                 if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);

  1292.                 if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);

  1293. 

  1294.                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1295.                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1296.             }

  1297.             c->pred_x= P_LEFT[0];

  1298.             c->pred_y= P_LEFT[1];

  1299.         }

  1300. 

  1301.         if(mv_table == s->b_forw_mv_table){

  1302.             mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);

  1303.         }else{

  1304.             mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);

  1305.         }

  1306. 

  1307.         dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);

  1308. 

  1309.         break;

  1310.     }

  1311. 

  1312.     dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);

  1313. 

  1314.     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)

  1315.         dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);

  1316. 

  1317. //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);

  1318. //    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;

  1319.     mv_table[mot_xy][0]= mx;

  1320.     mv_table[mot_xy][1]= my;

  1321. 

  1322.     return dmin;

  1323. }

  1324. 

  1325. static inline int check_bidir_mv(MpegEncContext * s,

  1326.                    int motion_fx, int motion_fy,

  1327.                    int motion_bx, int motion_by,

  1328.                    int pred_fx, int pred_fy,

  1329.                    int pred_bx, int pred_by,

  1330.                    int size, int h)

  1331. {

  1332.     //FIXME optimize?

  1333.     //FIXME better f_code prediction (max mv & distance)

  1334.     //FIXME pointers

  1335.     MotionEstContext * const c= &s->me;

  1336.     uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame

  1337.     uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame

  1338.     int stride= c->stride;

  1339.     uint8_t *dest_y = c->scratchpad;

  1340.     uint8_t *ptr;

  1341.     int dxy;

  1342.     int src_x, src_y;

  1343.     int fbmin;

  1344.     uint8_t **src_data= c->src[0];

  1345.     uint8_t **ref_data= c->ref[0];

  1346.     uint8_t **ref2_data= c->ref[2];

  1347. 

  1348.     if(s->quarter_sample){

  1349.         dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);

  1350.         src_x = motion_fx >> 2;

  1351.         src_y = motion_fy >> 2;

  1352. 

  1353.         ptr = ref_data[0] + (src_y * stride) + src_x;

  1354.         s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , stride);

  1355. 

  1356.         dxy = ((motion_by & 3) << 2) | (motion_bx & 3);

  1357.         src_x = motion_bx >> 2;

  1358.         src_y = motion_by >> 2;

  1359. 

  1360.         ptr = ref2_data[0] + (src_y * stride) + src_x;

  1361.         s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);

  1362.     }else{

  1363.         dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);

  1364.         src_x = motion_fx >> 1;

  1365.         src_y = motion_fy >> 1;

  1366. 

  1367.         ptr = ref_data[0] + (src_y * stride) + src_x;

  1368.         s->dsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);

  1369. 

  1370.         dxy = ((motion_by & 1) << 1) | (motion_bx & 1);

  1371.         src_x = motion_bx >> 1;

  1372.         src_y = motion_by >> 1;

  1373. 

  1374.         ptr = ref2_data[0] + (src_y * stride) + src_x;

  1375.         s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);

  1376.     }

  1377. 

  1378.     fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor

  1379.            +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor

  1380.            + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic

  1381. 

  1382.     if(c->avctx->mb_cmp&FF_CMP_CHROMA){

  1383.     }

  1384.     //FIXME CHROMA !!!

  1385. 

  1386.     return fbmin;

  1387. }

  1388. 

  1389. /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/

  1390. static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)

  1391. {

  1392.     MotionEstContext * const c= &s->me;

  1393.     const int mot_stride = s->mb_stride;

  1394.     const int xy = mb_y *mot_stride + mb_x;

  1395.     int fbmin;

  1396.     int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];

  1397.     int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];

  1398.     int pred_bx= s->b_bidir_back_mv_table[xy-1][0];

  1399.     int pred_by= s->b_bidir_back_mv_table[xy-1][1];

  1400.     int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];

  1401.     int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];

  1402.     int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];

  1403.     int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];

  1404.     const int flags= c->sub_flags;

  1405.     const int qpel= flags&FLAG_QPEL;

  1406.     const int shift= 1+qpel;

  1407.     const int xmin= c->xmin<<shift;

  1408.     const int ymin= c->ymin<<shift;

  1409.     const int xmax= c->xmax<<shift;

  1410.     const int ymax= c->ymax<<shift;

  1411.     uint8_t map[8][8][8][8];

  1412. 

  1413.     memset(map,0,sizeof(map));

  1414. #define BIDIR_MAP(fx,fy,bx,by) \

  1415.     map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]

  1416.     BIDIR_MAP(0,0,0,0) = 1;

  1417. 

  1418.     fbmin= check_bidir_mv(s, motion_fx, motion_fy,

  1419.                           motion_bx, motion_by,

  1420.                           pred_fx, pred_fy,

  1421.                           pred_bx, pred_by,

  1422.                           0, 16);

  1423. 

  1424.     if(s->avctx->bidir_refine){

  1425.         int score, end;

  1426. #define CHECK_BIDIR(fx,fy,bx,by)\

  1427.     if( !BIDIR_MAP(fx,fy,bx,by)\

  1428.        &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\

  1429.        &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\

  1430.         BIDIR_MAP(fx,fy,bx,by) = 1;\

  1431.         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);\

  1432.         if(score < fbmin){\

  1433.             fbmin= score;\

  1434.             motion_fx+=fx;\

  1435.             motion_fy+=fy;\

  1436.             motion_bx+=bx;\

  1437.             motion_by+=by;\

  1438.             end=0;\

  1439.         }\

  1440.     }

  1441. #define CHECK_BIDIR2(a,b,c,d)\

  1442. CHECK_BIDIR(a,b,c,d)\

  1443. CHECK_BIDIR(-(a),-(b),-(c),-(d))

  1444. 

  1445. #define CHECK_BIDIRR(a,b,c,d)\

  1446. CHECK_BIDIR2(a,b,c,d)\

  1447. CHECK_BIDIR2(b,c,d,a)\

  1448. CHECK_BIDIR2(c,d,a,b)\

  1449. CHECK_BIDIR2(d,a,b,c)

  1450. 

  1451.         do{

  1452.             end=1;

  1453. 

  1454.             CHECK_BIDIRR( 0, 0, 0, 1)

  1455.             if(s->avctx->bidir_refine > 1){

  1456.                 CHECK_BIDIRR( 0, 0, 1, 1)

  1457.                 CHECK_BIDIR2( 0, 1, 0, 1)

  1458.                 CHECK_BIDIR2( 1, 0, 1, 0)

  1459.                 CHECK_BIDIRR( 0, 0,-1, 1)

  1460.                 CHECK_BIDIR2( 0,-1, 0, 1)

  1461.                 CHECK_BIDIR2(-1, 0, 1, 0)

  1462.                 if(s->avctx->bidir_refine > 2){

  1463.                     CHECK_BIDIRR( 0, 1, 1, 1)

  1464.                     CHECK_BIDIRR( 0,-1, 1, 1)

  1465.                     CHECK_BIDIRR( 0, 1,-1, 1)

  1466.                     CHECK_BIDIRR( 0, 1, 1,-1)

  1467.                     if(s->avctx->bidir_refine > 3){

  1468.                         CHECK_BIDIR2( 1, 1, 1, 1)

  1469.                         CHECK_BIDIRR( 1, 1, 1,-1)

  1470.                         CHECK_BIDIR2( 1, 1,-1,-1)

  1471.                         CHECK_BIDIR2( 1,-1,-1, 1)

  1472.                         CHECK_BIDIR2( 1,-1, 1,-1)

  1473.                     }

  1474.                 }

  1475.             }

  1476.         }while(!end);

  1477.     }

  1478. 

  1479.     s->b_bidir_forw_mv_table[xy][0]= motion_fx;

  1480.     s->b_bidir_forw_mv_table[xy][1]= motion_fy;

  1481.     s->b_bidir_back_mv_table[xy][0]= motion_bx;

  1482.     s->b_bidir_back_mv_table[xy][1]= motion_by;

  1483. 

  1484.     return fbmin;

  1485. }

  1486. 

  1487. static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)

  1488. {

  1489.     MotionEstContext * const c= &s->me;

  1490.     int P[10][2];

  1491.     const int mot_stride = s->mb_stride;

  1492.     const int mot_xy = mb_y*mot_stride + mb_x;

  1493.     const int shift= 1+s->quarter_sample;

  1494.     int dmin, i;

  1495.     const int time_pp= s->pp_time;

  1496.     const int time_pb= s->pb_time;

  1497.     int mx, my, xmin, xmax, ymin, ymax;

  1498.     int16_t (*mv_table)[2]= s->b_direct_mv_table;

  1499. 

  1500.     c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;

  1501.     ymin= xmin=(-32)>>shift;

  1502.     ymax= xmax=   31>>shift;

  1503. 

  1504.     if(IS_8X8(s->next_picture.mb_type[mot_xy])){

  1505.         s->mv_type= MV_TYPE_8X8;

  1506.     }else{

  1507.         s->mv_type= MV_TYPE_16X16;

  1508.     }

  1509. 

  1510.     for(i=0; i<4; i++){

  1511.         int index= s->block_index[i];

  1512.         int min, max;

  1513. 

  1514.         c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];

  1515.         c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];

  1516.         c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));

  1517.         c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));

  1518. //        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);

  1519. //        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);

  1520. 

  1521.         max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;

  1522.         min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;

  1523.         max+= 16*mb_x + 1; // +-1 is for the simpler rounding

  1524.         min+= 16*mb_x - 1;

  1525.         xmax= FFMIN(xmax, s->width - max);

  1526.         xmin= FFMAX(xmin, - 16     - min);

  1527. 

  1528.         max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;

  1529.         min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;

  1530.         max+= 16*mb_y + 1; // +-1 is for the simpler rounding

  1531.         min+= 16*mb_y - 1;

  1532.         ymax= FFMIN(ymax, s->height - max);

  1533.         ymin= FFMAX(ymin, - 16      - min);

  1534. 

  1535.         if(s->mv_type == MV_TYPE_16X16) break;

  1536.     }

  1537. 

  1538.     assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);

  1539. 

  1540.     if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){

  1541.         s->b_direct_mv_table[mot_xy][0]= 0;

  1542.         s->b_direct_mv_table[mot_xy][1]= 0;

  1543. 

  1544.         return 256*256*256*64;

  1545.     }

  1546. 

  1547.     c->xmin= xmin;

  1548.     c->ymin= ymin;

  1549.     c->xmax= xmax;

  1550.     c->ymax= ymax;

  1551.     c->flags     |= FLAG_DIRECT;

  1552.     c->sub_flags |= FLAG_DIRECT;

  1553.     c->pred_x=0;

  1554.     c->pred_y=0;

  1555. 

  1556.     P_LEFT[0]        = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);

  1557.     P_LEFT[1]        = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);

  1558. 

  1559.     /* special case for first line */

  1560.     if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped

  1561.         P_TOP[0]      = av_clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);

  1562.         P_TOP[1]      = av_clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);

  1563.         P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);

  1564.         P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);

  1565. 

  1566.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1567.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1568.     }

  1569. 

  1570.     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);

  1571.     if(c->sub_flags&FLAG_QPEL)

  1572.         dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1573.     else

  1574.         dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1575. 

  1576.     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)

  1577.         dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);

  1578. 

  1579.     get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed

  1580. 

  1581.     mv_table[mot_xy][0]= mx;

  1582.     mv_table[mot_xy][1]= my;

  1583.     c->flags     &= ~FLAG_DIRECT;

  1584.     c->sub_flags &= ~FLAG_DIRECT;

  1585. 

  1586.     return dmin;

  1587. }

  1588. 

  1589. void ff_estimate_b_frame_motion(MpegEncContext * s,

  1590.                              int mb_x, int mb_y)

  1591. {

  1592.     MotionEstContext * const c= &s->me;

  1593.     const int penalty_factor= c->mb_penalty_factor;

  1594.     int fmin, bmin, dmin, fbmin, bimin, fimin;

  1595.     int type=0;

  1596.     const int xy = mb_y*s->mb_stride + mb_x;

  1597.     init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);

  1598. 

  1599.     get_limits(s, 16*mb_x, 16*mb_y);

  1600. 

  1601.     c->skip=0;

  1602. 

  1603.     if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){

  1604.         int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0

  1605. 

  1606.         score= ((unsigned)(score*score + 128*256))>>16;

  1607.         c->mc_mb_var_sum_temp += score;

  1608.         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE

  1609.         s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;

  1610. 

  1611.         return;

  1612.     }

  1613. 

  1614.     if(c->avctx->me_threshold){

  1615.         int vard= check_input_motion(s, mb_x, mb_y, 0);

  1616. 

  1617.         if((vard+128)>>8 < c->avctx->me_threshold){

  1618. //            pix = c->src[0][0];

  1619. //            sum = s->dsp.pix_sum(pix, s->linesize);

  1620. //            varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;

  1621. 

  1622. //            pic->mb_var   [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;

  1623.              s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;

  1624. /*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;

  1625.             c->mb_var_sum_temp    += (varc+128)>>8;*/

  1626.             c->mc_mb_var_sum_temp += (vard+128)>>8;

  1627. /*            if (vard <= 64<<8 || vard < varc) {

  1628.                 c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);

  1629.             }else{

  1630.                 c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;

  1631.             }*/

  1632.             return;

  1633.         }

  1634.         if((vard+128)>>8 < c->avctx->mb_threshold){

  1635.             type= s->mb_type[mb_y*s->mb_stride + mb_x];

  1636.             if(type == CANDIDATE_MB_TYPE_DIRECT){

  1637.                 direct_search(s, mb_x, mb_y);

  1638.             }

  1639.             if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){

  1640.                 c->skip=0;

  1641.                 ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);

  1642.             }

  1643.             if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){

  1644.                 c->skip=0;

  1645.                 ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);

  1646.             }

  1647.             if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){

  1648.                 c->skip=0;

  1649.                 c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

  1650.                 interlaced_search(s, 0,

  1651.                                         s->b_field_mv_table[0], s->b_field_select_table[0],

  1652.                                         s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);

  1653.             }

  1654.             if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){

  1655.                 c->skip=0;

  1656.                 c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;

  1657.                 interlaced_search(s, 2,

  1658.                                         s->b_field_mv_table[1], s->b_field_select_table[1],

  1659.                                         s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);

  1660.             }

  1661.             return;

  1662.         }

  1663.     }

  1664. 

  1665.     if (s->codec_id == CODEC_ID_MPEG4)

  1666.         dmin= direct_search(s, mb_x, mb_y);

  1667.     else

  1668.         dmin= INT_MAX;

  1669. //FIXME penalty stuff for non mpeg4

  1670.     c->skip=0;

  1671.     fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;

  1672. 

  1673.     c->skip=0;

  1674.     bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;

  1675. //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

  1676. 

  1677.     c->skip=0;

  1678.     fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;

  1679. //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);

  1680. 

  1681.     if(s->flags & CODEC_FLAG_INTERLACED_ME){

  1682. //FIXME mb type penalty

  1683.         c->skip=0;

  1684.         c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

  1685.         fimin= interlaced_search(s, 0,

  1686.                                  s->b_field_mv_table[0], s->b_field_select_table[0],

  1687.                                  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);

  1688.         c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;

  1689.         bimin= interlaced_search(s, 2,

  1690.                                  s->b_field_mv_table[1], s->b_field_select_table[1],

  1691.                                  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);

  1692.     }else

  1693.         fimin= bimin= INT_MAX;

  1694. 

  1695.     {

  1696.         int score= fmin;

  1697.         type = CANDIDATE_MB_TYPE_FORWARD;

  1698. 

  1699.         if (dmin <= score){

  1700.             score = dmin;

  1701.             type = CANDIDATE_MB_TYPE_DIRECT;

  1702.         }

  1703.         if(bmin<score){

  1704.             score=bmin;

  1705.             type= CANDIDATE_MB_TYPE_BACKWARD;

  1706.         }

  1707.         if(fbmin<score){

  1708.             score=fbmin;

  1709.             type= CANDIDATE_MB_TYPE_BIDIR;

  1710.         }

  1711.         if(fimin<score){

  1712.             score=fimin;

  1713.             type= CANDIDATE_MB_TYPE_FORWARD_I;

  1714.         }

  1715.         if(bimin<score){

  1716.             score=bimin;

  1717.             type= CANDIDATE_MB_TYPE_BACKWARD_I;

  1718.         }

  1719. 

  1720.         score= ((unsigned)(score*score + 128*256))>>16;

  1721.         c->mc_mb_var_sum_temp += score;

  1722.         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE

  1723.     }

  1724. 

  1725.     if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){

  1726.         type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;

  1727.         if(fimin < INT_MAX)

  1728.             type |= CANDIDATE_MB_TYPE_FORWARD_I;

  1729.         if(bimin < INT_MAX)

  1730.             type |= CANDIDATE_MB_TYPE_BACKWARD_I;

  1731.         if(fimin < INT_MAX && bimin < INT_MAX){

  1732.             type |= CANDIDATE_MB_TYPE_BIDIR_I;

  1733.         }

  1734.          //FIXME something smarter

  1735.         if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB

  1736.         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])

  1737.             type |= CANDIDATE_MB_TYPE_DIRECT0;

  1738. #if 0

  1739.         if(s->out_format == FMT_MPEG1)

  1740.             type |= CANDIDATE_MB_TYPE_INTRA;

  1741. #endif

  1742.     }

  1743. 

  1744.     s->mb_type[mb_y*s->mb_stride + mb_x]= type;

  1745. }

  1746. 

  1747. /* find best f_code for ME which do unlimited searches */

  1748. int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)

  1749. {

  1750.     if(s->me_method>=ME_EPZS){

  1751.         int score[8];

  1752.         int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);

  1753.         uint8_t * fcode_tab= s->fcode_tab;

  1754.         int best_fcode=-1;

  1755.         int best_score=-10000000;

  1756. 

  1757.         if(s->msmpeg4_version)

  1758.             range= FFMIN(range, 16);

  1759.         else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)

  1760.             range= FFMIN(range, 256);

  1761. 

  1762.         for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);

  1763. 

  1764.         for(y=0; y<s->mb_height; y++){

  1765.             int x;

  1766.             int xy= y*s->mb_stride;

  1767.             for(x=0; x<s->mb_width; x++){

  1768.                 if(s->mb_type[xy] & type){

  1769.                     int mx= mv_table[xy][0];

  1770.                     int my= mv_table[xy][1];

  1771.                     int fcode= FFMAX(fcode_tab[mx + MAX_MV],

  1772.                                      fcode_tab[my + MAX_MV]);

  1773.                     int j;

  1774. 

  1775.                         if(mx >= range || mx < -range ||

  1776.                            my >= range || my < -range)

  1777.                             continue;

  1778. 

  1779.                     for(j=0; j<fcode && j<8; j++){

  1780.                         if(s->pict_type==FF_B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])

  1781.                             score[j]-= 170;

  1782.                     }

  1783.                 }

  1784.                 xy++;

  1785.             }

  1786.         }

  1787. 

  1788.         for(i=1; i<8; i++){

  1789.             if(score[i] > best_score){

  1790.                 best_score= score[i];

  1791.                 best_fcode= i;

  1792.             }

  1793. //            printf("%d %d\n", i, score[i]);

  1794.         }

  1795. 

  1796. //    printf("fcode: %d type: %d\n", i, s->pict_type);

  1797.         return best_fcode;

  1798. /*        for(i=0; i<=MAX_FCODE; i++){

  1799.             printf("%d ", mv_num[i]);

  1800.         }

  1801.         printf("\n");*/

  1802.     }else{

  1803.         return 1;

  1804.     }

  1805. }

  1806. 

  1807. void ff_fix_long_p_mvs(MpegEncContext * s)

  1808. {

  1809.     MotionEstContext * const c= &s->me;

  1810.     const int f_code= s->f_code;

  1811.     int y, range;

  1812.     assert(s->pict_type==FF_P_TYPE);

  1813. 

  1814.     range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);

  1815. 

  1816.     assert(range <= 16 || !s->msmpeg4_version);

  1817.     assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));

  1818. 

  1819.     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;

  1820. 

  1821. //printf("%d no:%d %d//\n", clip, noclip, f_code);

  1822.     if(s->flags&CODEC_FLAG_4MV){

  1823.         const int wrap= s->b8_stride;

  1824. 

  1825.         /* clip / convert to intra 8x8 type MVs */

  1826.         for(y=0; y<s->mb_height; y++){

  1827.             int xy= y*2*wrap;

  1828.             int i= y*s->mb_stride;

  1829.             int x;

  1830. 

  1831.             for(x=0; x<s->mb_width; x++){

  1832.                 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){

  1833.                     int block;

  1834.                     for(block=0; block<4; block++){

  1835.                         int off= (block& 1) + (block>>1)*wrap;

  1836.                         int mx= s->current_picture.motion_val[0][ xy + off ][0];

  1837.                         int my= s->current_picture.motion_val[0][ xy + off ][1];

  1838. 

  1839.                         if(   mx >=range || mx <-range

  1840.                            || my >=range || my <-range){

  1841.                             s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;

  1842.                             s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;

  1843.                             s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;

  1844.                         }

  1845.                     }

  1846.                 }

  1847.                 xy+=2;

  1848.                 i++;

  1849.             }

  1850.         }

  1851.     }

  1852. }

  1853. 

  1854. /**

  1855.  *

  1856.  * @param truncate 1 for truncation, 0 for using intra

  1857.  */

  1858. void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,

  1859.                      int16_t (*mv_table)[2], int f_code, int type, int truncate)

  1860. {

  1861.     MotionEstContext * const c= &s->me;

  1862.     int y, h_range, v_range;

  1863. 

  1864.     // RAL: 8 in MPEG-1, 16 in MPEG-4

  1865.     int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);

  1866. 

  1867.     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;

  1868. 

  1869.     h_range= range;

  1870.     v_range= field_select_table ? range>>1 : range;

  1871. 

  1872.     /* clip / convert to intra 16x16 type MVs */

  1873.     for(y=0; y<s->mb_height; y++){

  1874.         int x;

  1875.         int xy= y*s->mb_stride;

  1876.         for(x=0; x<s->mb_width; x++){

  1877.             if (s->mb_type[xy] & type){    // RAL: "type" test added...

  1878.                 if(field_select_table==NULL || field_select_table[xy] == field_select){

  1879.                     if(   mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range

  1880.                        || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){

  1881. 

  1882.                         if(truncate){

  1883.                             if     (mv_table[xy][0] > h_range-1) mv_table[xy][0]=  h_range-1;

  1884.                             else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;

  1885.                             if     (mv_table[xy][1] > v_range-1) mv_table[xy][1]=  v_range-1;

  1886.                             else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;

  1887.                         }else{

  1888.                             s->mb_type[xy] &= ~type;

  1889.                             s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;

  1890.                             mv_table[xy][0]=

  1891.                             mv_table[xy][1]= 0;

  1892.                         }

  1893.                     }

  1894.                 }

  1895.             }

  1896.             xy++;

  1897.         }

  1898.     }

  1899. }