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.
40315 Commits
32 Branches
570MB
Tree: a0ce85ac7d
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'a0ce85ac7d'
${ noResults }
FFmpeg/libavcodec/motion_est.c

1756 lines
67KB
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 Libav.

  9.  *

  10.  * Libav 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.  * Libav 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 Libav; if not, write to the Free Software

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

  23.  */

  24. 

  25. /**

  26.  * @file

  27.  * Motion estimation.

  28.  */

  29. 

  30. #include <stdlib.h>

  31. #include <stdio.h>

  32. #include <limits.h>

  33. 

  34. #include "avcodec.h"

  35. #include "mathops.h"

  36. #include "mpegutils.h"

  37. #include "mpegvideo.h"

  38. 

  39. #undef NDEBUG

  40. #include <assert.h>

  41. 

  42. #define P_LEFT P[1]

  43. #define P_TOP P[2]

  44. #define P_TOPRIGHT P[3]

  45. #define P_MEDIAN P[4]

  46. #define P_MV1 P[9]

  47. 

  48. #define ME_MAP_SHIFT 3

  49. #define ME_MAP_MV_BITS 11

  50. 

  51. static int sad_hpel_motion_search(MpegEncContext * s,

  52.                                   int *mx_ptr, int *my_ptr, int dmin,

  53.                                   int src_index, int ref_index,

  54.                                   int size, int h);

  55. 

  56. static inline unsigned update_map_generation(MotionEstContext *c)

  57. {

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

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

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

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

  62.     }

  63.     return c->map_generation;

  64. }

  65. 

  66. /* shape adaptive search stuff */

  67. typedef struct Minima{

  68.     int height;

  69.     int x, y;

  70.     int checked;

  71. }Minima;

  72. 

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

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

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

  76. 

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

  78. }

  79. 

  80. #define FLAG_QPEL   1 //must be 1

  81. #define FLAG_CHROMA 2

  82. #define FLAG_DIRECT 4

  83. 

  84. static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){

  85.     const int offset[3]= {

  86.           y*c->  stride + x,

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

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

  89.     };

  90.     int i;

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

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

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

  94.     }

  95.     if(ref_index){

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

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

  98.         }

  99.     }

  100. }

  101. 

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

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

  104.            + (direct ? FLAG_DIRECT : 0)

  105.            + (chroma ? FLAG_CHROMA : 0);

  106. }

  107. 

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

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

  110.                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel){

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

  112.     const int stride= c->stride;

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

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

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

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

  117.     int d;

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

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

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

  121.             const int time_pp= s->pp_time;

  122.             const int time_pb= s->pb_time;

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

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

  125.                 int i;

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

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

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

  129.                     int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4));

  130.                     int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));

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

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

  133. 

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

  135.                     if(qpel){

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

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

  138.                     }else{

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

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

  141.                     }

  142.                 }

  143.             }else{

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

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

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

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

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

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

  150. 

  151.                 if(qpel){

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

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

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

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

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

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

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

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

  160.                 }else{

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

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

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

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

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

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

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

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

  169. 

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

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

  172.                 }

  173.             }

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

  175.         }else

  176.             d= 256*256*256*32;

  177.     return d;

  178. }

  179. 

  180. static av_always_inline int cmp_inline(MpegEncContext *s, const int x, const int y, const int subx, const int suby,

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

  182.                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, int qpel, int chroma){

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

  184.     const int stride= c->stride;

  185.     const int uvstride= c->uvstride;

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

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

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

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

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

  191.     int d;

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

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

  194.         if(dxy){

  195.             if(qpel){

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

  197.                 if(chroma){

  198.                     int cx= hx/2;

  199.                     int cy= hy/2;

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

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

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

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

  204.                 }

  205.             }else{

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

  207.                 if(chroma)

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

  209.             }

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

  211.         }else{

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

  213.             if(chroma)

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

  215.         }

  216.         if(chroma){

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

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

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

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

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

  222.         }

  223.     return d;

  224. }

  225. 

  226. static int cmp_simple(MpegEncContext *s, const int x, const int y,

  227.                       int ref_index, int src_index,

  228.                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){

  229.     return cmp_inline(s,x,y,0,0,0,16,ref_index,src_index, cmp_func, chroma_cmp_func, 0, 0);

  230. }

  231. 

  232. static int cmp_fpel_internal(MpegEncContext *s, const int x, const int y,

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

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

  235.     if(flags&FLAG_DIRECT){

  236.         return cmp_direct_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);

  237.     }else{

  238.         return cmp_inline(s,x,y,0,0,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);

  239.     }

  240. }

  241. 

  242. static int cmp_internal(MpegEncContext *s, const int x, const int y, const int subx, const int suby,

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

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

  245.     if(flags&FLAG_DIRECT){

  246.         return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL);

  247.     }else{

  248.         return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags&FLAG_QPEL, flags&FLAG_CHROMA);

  249.     }

  250. }

  251. 

  252. /** @brief compares a block (either a full macroblock or a partition thereof)

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

  254.  */

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

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

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

  258.     if(av_builtin_constant_p(flags) && av_builtin_constant_p(h) && av_builtin_constant_p(size)

  259.        && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)

  260.        && flags==0 && h==16 && size==0 && subx==0 && suby==0){

  261.         return cmp_simple(s,x,y,ref_index,src_index, cmp_func, chroma_cmp_func);

  262.     }else if(av_builtin_constant_p(subx) && av_builtin_constant_p(suby)

  263.        && subx==0 && suby==0){

  264.         return cmp_fpel_internal(s,x,y,size,h,ref_index,src_index, cmp_func, chroma_cmp_func,flags);

  265.     }else{

  266.         return cmp_internal(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, flags);

  267.     }

  268. }

  269. 

  270. static int cmp_hpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,

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

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

  273.     if(flags&FLAG_DIRECT){

  274.         return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0);

  275.     }else{

  276.         return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 0, flags&FLAG_CHROMA);

  277.     }

  278. }

  279. 

  280. static int cmp_qpel(MpegEncContext *s, const int x, const int y, const int subx, const int suby,

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

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

  283.     if(flags&FLAG_DIRECT){

  284.         return cmp_direct_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1);

  285.     }else{

  286.         return cmp_inline(s,x,y,subx,suby,size,h,ref_index,src_index, cmp_func, chroma_cmp_func, 1, flags&FLAG_CHROMA);

  287.     }

  288. }

  289. 

  290. #include "motion_est_template.c"

  291. 

  292. static int zero_cmp(MpegEncContext *s, uint8_t *a, uint8_t *b,

  293.                     int stride, int h)

  294. {

  295.     return 0;

  296. }

  297. 

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

  299. }

  300. 

  301. int ff_init_me(MpegEncContext *s){

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

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

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

  305. 

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

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

  308.         return -1;

  309.     }

  310.     if (s->me_method != ME_ZERO &&

  311.         s->me_method != ME_EPZS &&

  312.         s->me_method != ME_X1) {

  313.         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");

  314.         return -1;

  315.     }

  316. 

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

  318. 

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

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

  321.     }

  322. 

  323.     ff_set_cmp(&s->mecc, s->mecc.me_pre_cmp, c->avctx->me_pre_cmp);

  324.     ff_set_cmp(&s->mecc, s->mecc.me_cmp,     c->avctx->me_cmp);

  325.     ff_set_cmp(&s->mecc, s->mecc.me_sub_cmp, c->avctx->me_sub_cmp);

  326.     ff_set_cmp(&s->mecc, s->mecc.mb_cmp,     c->avctx->mb_cmp);

  327. 

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

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

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

  331. 

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

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

  334.         c->sub_motion_search= qpel_motion_search;

  335.         c->qpel_avg = s->qdsp.avg_qpel_pixels_tab;

  336.         if (s->no_rounding)

  337.             c->qpel_put = s->qdsp.put_no_rnd_qpel_pixels_tab;

  338.         else

  339.             c->qpel_put = s->qdsp.put_qpel_pixels_tab;

  340.     }else{

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

  342.             c->sub_motion_search= hpel_motion_search;

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

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

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

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

  347.         else

  348.             c->sub_motion_search= hpel_motion_search;

  349.     }

  350.     c->hpel_avg = s->hdsp.avg_pixels_tab;

  351.     if (s->no_rounding)

  352.         c->hpel_put = s->hdsp.put_no_rnd_pixels_tab;

  353.     else

  354.         c->hpel_put = s->hdsp.put_pixels_tab;

  355. 

  356.     if(s->linesize){

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

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

  359.     }else{

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

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

  362.     }

  363. 

  364.     /* 8x8 fullpel search would need a 4x4 chroma compare, which we do

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

  366.      * does not expect it. */

  367.     if ((c->avctx->me_cmp & FF_CMP_CHROMA) /* && !s->mecc.me_cmp[2] */)

  368.         s->mecc.me_cmp[2] = zero_cmp;

  369.     if ((c->avctx->me_sub_cmp & FF_CMP_CHROMA) && !s->mecc.me_sub_cmp[2])

  370.         s->mecc.me_sub_cmp[2] = zero_cmp;

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

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

  373. 

  374.     if(s->codec_id == AV_CODEC_ID_H261){

  375.         c->sub_motion_search= no_sub_motion_search;

  376.     }

  377. 

  378.     return 0;

  379. }

  380. 

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

  382. {\

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

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

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

  386. }

  387. 

  388. static int sad_hpel_motion_search(MpegEncContext * s,

  389.                                   int *mx_ptr, int *my_ptr, int dmin,

  390.                                   int src_index, int ref_index,

  391.                                   int size, int h)

  392. {

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

  394.     const int penalty_factor= c->sub_penalty_factor;

  395.     int mx, my, dminh;

  396.     uint8_t *pix, *ptr;

  397.     int stride= c->stride;

  398.     const int flags= c->sub_flags;

  399.     LOAD_COMMON

  400. 

  401.     assert(flags == 0);

  402. 

  403.     if(c->skip){

  404.         *mx_ptr = 0;

  405.         *my_ptr = 0;

  406.         return dmin;

  407.     }

  408. 

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

  410. 

  411.     mx = *mx_ptr;

  412.     my = *my_ptr;

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

  414. 

  415.     dminh = dmin;

  416. 

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

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

  419.         int dx=0, dy=0;

  420.         int d, pen_x, pen_y;

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

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

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

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

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

  426.         mx<<=1;

  427.         my<<=1;

  428. 

  429. 

  430.         pen_x= pred_x + mx;

  431.         pen_y= pred_y + my;

  432. 

  433.         ptr-= stride;

  434.         if(t<=b){

  435.             CHECK_SAD_HALF_MV(y2 , 0, -1)

  436.             if(l<=r){

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

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

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

  440.                     ptr+= stride;

  441.                 }else{

  442.                     ptr+= stride;

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

  444.                 }

  445.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

  446.             }else{

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

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

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

  450.                     ptr+= stride;

  451.                 }else{

  452.                     ptr+= stride;

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

  454.                 }

  455.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

  456.             }

  457.         }else{

  458.             if(l<=r){

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

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

  461.                     ptr+= stride;

  462.                 }else{

  463.                     ptr+= stride;

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

  465.                 }

  466.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

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

  468.             }else{

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

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

  471.                     ptr+= stride;

  472.                 }else{

  473.                     ptr+= stride;

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

  475.                 }

  476.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

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

  478.             }

  479.             CHECK_SAD_HALF_MV(y2 ,  0, +1)

  480.         }

  481.         mx+=dx;

  482.         my+=dy;

  483. 

  484.     }else{

  485.         mx<<=1;

  486.         my<<=1;

  487.     }

  488. 

  489.     *mx_ptr = mx;

  490.     *my_ptr = my;

  491.     return dminh;

  492. }

  493. 

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

  495. {

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

  497. 

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

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

  500. 

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

  502.     if(mv4){

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

  504. 

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

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

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

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

  509. 

  510.         mot_xy += s->b8_stride;

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

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

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

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

  515.     }

  516. }

  517. 

  518. /**

  519.  * get fullpel ME search limits.

  520.  */

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

  522. {

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

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

  525. /*

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

  527.     else                   c->range= 16;

  528. */

  529.     if (s->unrestricted_mv) {

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

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

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

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

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

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

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

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

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

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

  540.     } else {

  541.         c->xmin = - x;

  542.         c->ymin = - y;

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

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

  545.     }

  546.     if(range){

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

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

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

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

  551.     }

  552. }

  553. 

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

  555.     const int stride= c->stride;

  556. 

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

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

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

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

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

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

  563. }

  564. 

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

  566. {

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

  568.     const int size= 1;

  569.     const int h=8;

  570.     int block;

  571.     int P[10][2];

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

  573.     int same=1;

  574.     const int stride= c->stride;

  575.     uint8_t *mv_penalty= c->current_mv_penalty;

  576. 

  577.     init_mv4_ref(c);

  578. 

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

  580.         int mx4, my4;

  581.         int pred_x4, pred_y4;

  582.         int dmin4;

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

  584.         const int mot_stride = s->b8_stride;

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

  586. 

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

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

  589. 

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

  591. 

  592.         /* special case for first line */

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

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

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

  596.         } else {

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

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

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

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

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

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

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

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

  605. 

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

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

  608. 

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

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

  611.         }

  612.         P_MV1[0]= mx;

  613.         P_MV1[1]= my;

  614. 

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

  616. 

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

  618. 

  619.         if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {

  620.             int dxy;

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

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

  623.             if(s->quarter_sample){

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

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

  626. 

  627.                 if(s->no_rounding)

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

  629.                 else

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

  631.             }else{

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

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

  634. 

  635.                 if(s->no_rounding)

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

  637.                 else

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

  639.             }

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

  641.         }else

  642.             dmin_sum+= dmin4;

  643. 

  644.         if(s->quarter_sample){

  645.             mx4_sum+= mx4/2;

  646.             my4_sum+= my4/2;

  647.         }else{

  648.             mx4_sum+= mx4;

  649.             my4_sum+= my4;

  650.         }

  651. 

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

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

  654. 

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

  656.     }

  657. 

  658.     if(same)

  659.         return INT_MAX;

  660. 

  661.     if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {

  662.         dmin_sum += s->mecc.mb_cmp[0](s,

  663.                                       s->new_picture.f->data[0] +

  664.                                       s->mb_x * 16 + s->mb_y * 16 * stride,

  665.                                       c->scratchpad, stride, 16);

  666.     }

  667. 

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

  669.         int dxy;

  670.         int mx, my;

  671.         int offset;

  672. 

  673.         mx= ff_h263_round_chroma(mx4_sum);

  674.         my= ff_h263_round_chroma(my4_sum);

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

  676. 

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

  678. 

  679.         if(s->no_rounding){

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

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

  682.         }else{

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

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

  685.         }

  686. 

  687.         dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture.f->data[1] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad,     s->uvlinesize, 8);

  688.         dmin_sum += s->mecc.mb_cmp[1](s, s->new_picture.f->data[2] + s->mb_x * 8 + s->mb_y * 8 * s->uvlinesize, c->scratchpad + 8, s->uvlinesize, 8);

  689.     }

  690. 

  691.     c->pred_x= mx;

  692.     c->pred_y= my;

  693. 

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

  695.     /*case FF_CMP_SSE:

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

  697.     case FF_CMP_RD:

  698.         return dmin_sum;

  699.     default:

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

  701.     }

  702. }

  703. 

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

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

  706. 

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

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

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

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

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

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

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

  714.     }

  715. }

  716. 

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

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

  719. {

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

  721.     const int size=0;

  722.     const int h=8;

  723.     int block;

  724.     int P[10][2];

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

  726.     int same=1;

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

  728.     int dmin_sum= 0;

  729.     const int mot_stride= s->mb_stride;

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

  731. 

  732.     c->ymin>>=1;

  733.     c->ymax>>=1;

  734.     c->stride<<=1;

  735.     c->uvstride<<=1;

  736.     init_interlaced_ref(s, ref_index);

  737. 

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

  739.         int field_select;

  740.         int best_dmin= INT_MAX;

  741.         int best_field= -1;

  742. 

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

  744.             int dmin, mx_i, my_i;

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

  746. 

  747.             if(user_field_select){

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

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

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

  751.                     continue;

  752.             }

  753. 

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

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

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

  757. 

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

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

  760. 

  761.             if(!s->first_slice_line){

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

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

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

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

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

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

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

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

  770. 

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

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

  773.             }

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

  775.             P_MV1[1]= my / 2;

  776. 

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

  778. 

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

  780. 

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

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

  783. 

  784.             if (s->mecc.me_sub_cmp[0] != s->mecc.mb_cmp[0]) {

  785.                 int dxy;

  786. 

  787.                 //FIXME chroma ME

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

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

  790. 

  791.                 if(s->no_rounding){

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

  793.                 }else{

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

  795.                 }

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

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

  798.             }else

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

  800. 

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

  802. 

  803.             if(dmin < best_dmin){

  804.                 best_dmin= dmin;

  805.                 best_field= field_select;

  806.             }

  807.         }

  808.         {

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

  810. 

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

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

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

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

  815.         }

  816. 

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

  818.         dmin_sum += best_dmin;

  819.     }

  820. 

  821.     c->ymin<<=1;

  822.     c->ymax<<=1;

  823.     c->stride>>=1;

  824.     c->uvstride>>=1;

  825. 

  826.     if(same)

  827.         return INT_MAX;

  828. 

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

  830.     /*case FF_CMP_SSE:

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

  832.     case FF_CMP_RD:

  833.         return dmin_sum;

  834.     default:

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

  836.     }

  837. }

  838. 

  839. static inline int get_penalty_factor(int lambda, int lambda2, int type){

  840.     switch(type&0xFF){

  841.     default:

  842.     case FF_CMP_SAD:

  843.         return lambda>>FF_LAMBDA_SHIFT;

  844.     case FF_CMP_DCT:

  845.         return (3*lambda)>>(FF_LAMBDA_SHIFT+1);

  846.     case FF_CMP_SATD:

  847.     case FF_CMP_DCT264:

  848.         return (2*lambda)>>FF_LAMBDA_SHIFT;

  849.     case FF_CMP_RD:

  850.     case FF_CMP_PSNR:

  851.     case FF_CMP_SSE:

  852.     case FF_CMP_NSSE:

  853.         return lambda2>>FF_LAMBDA_SHIFT;

  854.     case FF_CMP_BIT:

  855.         return 1;

  856.     }

  857. }

  858. 

  859. void ff_estimate_p_frame_motion(MpegEncContext * s,

  860.                                 int mb_x, int mb_y)

  861. {

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

  863.     uint8_t *pix, *ppix;

  864.     int sum, mx, my, dmin;

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

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

  867.     int P[10][2];

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

  869.     int mb_type=0;

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

  871. 

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

  873. 

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

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

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

  877. 

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

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

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

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

  882. 

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

  884.     c->skip=0;

  885. 

  886.     /* intra / predictive decision */

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

  888.     sum  = s->mpvencdsp.pix_sum(pix, s->linesize);

  889.     varc = s->mpvencdsp.pix_norm1(pix, s->linesize) -

  890.            (((unsigned) sum * sum) >> 8) + 500;

  891. 

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

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

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

  895. 

  896.     switch(s->me_method) {

  897.     case ME_ZERO:

  898.     default:

  899.         mx   = 0;

  900.         my   = 0;

  901.         dmin = 0;

  902.         break;

  903.     case ME_X1:

  904.     case ME_EPZS:

  905.        {

  906.             const int mot_stride = s->b8_stride;

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

  908. 

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

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

  911. 

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

  913. 

  914.             if(!s->first_slice_line) {

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

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

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

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

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

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

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

  922. 

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

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

  925. 

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

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

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

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

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

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

  932.                 }

  933.             }else{

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

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

  936.             }

  937. 

  938.         }

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

  940. 

  941.         break;

  942.     }

  943. 

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

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

  946. 

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

  948. 

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

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

  951. 

  952.     if(mb_type){

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

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

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

  956. 

  957.         if(mb_type == CANDIDATE_MB_TYPE_INTER){

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

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

  960.         }else{

  961.             mx <<=shift;

  962.             my <<=shift;

  963.         }

  964.         if(mb_type == CANDIDATE_MB_TYPE_INTER4V){

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

  966. 

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

  968.         }

  969.         if(mb_type == CANDIDATE_MB_TYPE_INTER_I){

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

  971.         }

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

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

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

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

  976. 

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

  978.             mb_type|= CANDIDATE_MB_TYPE_INTRA;

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

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

  981.             mb_type|= CANDIDATE_MB_TYPE_INTER;

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

  983.             if (s->mpv_flags & FF_MPV_FLAG_MV0)

  984.                 if(mx || my)

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

  986.         }else{

  987.             mx <<=shift;

  988.             my <<=shift;

  989.         }

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

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

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

  993.                 mb_type|=CANDIDATE_MB_TYPE_INTER4V;

  994. 

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

  996.         }else

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

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

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

  1000.             if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)

  1001.                 mb_type |= CANDIDATE_MB_TYPE_INTER_I;

  1002.         }

  1003.     }else{

  1004.         int intra_score, i;

  1005.         mb_type= CANDIDATE_MB_TYPE_INTER;

  1006. 

  1007.         dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1008.         if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)

  1009.             dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);

  1010. 

  1011.         if((s->flags&CODEC_FLAG_4MV)

  1012.            && !c->skip && varc>50<<8 && vard>10<<8){

  1013.             int dmin4= h263_mv4_search(s, mx, my, shift);

  1014.             if(dmin4 < dmin){

  1015.                 mb_type= CANDIDATE_MB_TYPE_INTER4V;

  1016.                 dmin=dmin4;

  1017.             }

  1018.         }

  1019.         if((s->flags&CODEC_FLAG_INTERLACED_ME)

  1020.            && !c->skip){ //FIXME varc/d checks

  1021.             int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);

  1022.             if(dmin_i < dmin){

  1023.                 mb_type = CANDIDATE_MB_TYPE_INTER_I;

  1024.                 dmin= dmin_i;

  1025.             }

  1026.         }

  1027. 

  1028.         set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);

  1029. 

  1030.         /* get intra luma score */

  1031.         if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){

  1032.             intra_score= varc - 500;

  1033.         }else{

  1034.             unsigned mean = (sum+128)>>8;

  1035.             mean*= 0x01010101;

  1036. 

  1037.             for(i=0; i<16; i++){

  1038.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;

  1039.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;

  1040.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;

  1041.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;

  1042.             }

  1043. 

  1044.             intra_score= s->mecc.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);

  1045.         }

  1046.         intra_score += c->mb_penalty_factor*16;

  1047. 

  1048.         if(intra_score < dmin){

  1049.             mb_type= CANDIDATE_MB_TYPE_INTRA;

  1050.             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup

  1051.         }else

  1052.             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;

  1053. 

  1054.         {

  1055.             int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);

  1056.             int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;

  1057.             c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);

  1058.         }

  1059.     }

  1060. 

  1061.     s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;

  1062. }

  1063. 

  1064. int ff_pre_estimate_p_frame_motion(MpegEncContext * s,

  1065.                                     int mb_x, int mb_y)

  1066. {

  1067.     MotionEstContext * const c= &s->me;

  1068.     int mx, my, dmin;

  1069.     int P[10][2];

  1070.     const int shift= 1+s->quarter_sample;

  1071.     const int xy= mb_x + mb_y*s->mb_stride;

  1072.     init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);

  1073. 

  1074.     assert(s->quarter_sample==0 || s->quarter_sample==1);

  1075. 

  1076.     c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);

  1077.     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

  1078. 

  1079.     get_limits(s, 16*mb_x, 16*mb_y);

  1080.     c->skip=0;

  1081. 

  1082.     P_LEFT[0]       = s->p_mv_table[xy + 1][0];

  1083.     P_LEFT[1]       = s->p_mv_table[xy + 1][1];

  1084. 

  1085.     if(P_LEFT[0]       < (c->xmin<<shift)) P_LEFT[0]       = (c->xmin<<shift);

  1086. 

  1087.     /* special case for first line */

  1088.     if (s->first_slice_line) {

  1089.         c->pred_x= P_LEFT[0];

  1090.         c->pred_y= P_LEFT[1];

  1091.         P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=

  1092.         P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME

  1093.     } else {

  1094.         P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];

  1095.         P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];

  1096.         P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];

  1097.         P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];

  1098.         if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);

  1099.         if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);

  1100.         if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);

  1101. 

  1102.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1103.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1104. 

  1105.         c->pred_x = P_MEDIAN[0];

  1106.         c->pred_y = P_MEDIAN[1];

  1107.     }

  1108. 

  1109.     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);

  1110. 

  1111.     s->p_mv_table[xy][0] = mx<<shift;

  1112.     s->p_mv_table[xy][1] = my<<shift;

  1113. 

  1114.     return dmin;

  1115. }

  1116. 

  1117. static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,

  1118.                              int16_t (*mv_table)[2], int ref_index, int f_code)

  1119. {

  1120.     MotionEstContext * const c= &s->me;

  1121.     int mx, my, dmin;

  1122.     int P[10][2];

  1123.     const int shift= 1+s->quarter_sample;

  1124.     const int mot_stride = s->mb_stride;

  1125.     const int mot_xy = mb_y*mot_stride + mb_x;

  1126.     uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;

  1127.     int mv_scale;

  1128. 

  1129.     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);

  1130.     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);

  1131.     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);

  1132.     c->current_mv_penalty= mv_penalty;

  1133. 

  1134.     get_limits(s, 16*mb_x, 16*mb_y);

  1135. 

  1136.     switch(s->me_method) {

  1137.     case ME_ZERO:

  1138.     default:

  1139.         mx   = 0;

  1140.         my   = 0;

  1141.         dmin = 0;

  1142.         break;

  1143.     case ME_X1:

  1144.     case ME_EPZS:

  1145.         P_LEFT[0] = mv_table[mot_xy - 1][0];

  1146.         P_LEFT[1] = mv_table[mot_xy - 1][1];

  1147. 

  1148.         if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);

  1149. 

  1150.         /* special case for first line */

  1151.         if (!s->first_slice_line) {

  1152.             P_TOP[0]      = mv_table[mot_xy - mot_stride    ][0];

  1153.             P_TOP[1]      = mv_table[mot_xy - mot_stride    ][1];

  1154.             P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];

  1155.             P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];

  1156.             if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);

  1157.             if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);

  1158.             if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);

  1159. 

  1160.             P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1161.             P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1162.         }

  1163.         c->pred_x = P_LEFT[0];

  1164.         c->pred_y = P_LEFT[1];

  1165. 

  1166.         if(mv_table == s->b_forw_mv_table){

  1167.             mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);

  1168.         }else{

  1169.             mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);

  1170.         }

  1171. 

  1172.         dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);

  1173. 

  1174.         break;

  1175.     }

  1176. 

  1177.     dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);

  1178. 

  1179.     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)

  1180.         dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);

  1181. 

  1182. //    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;

  1183.     mv_table[mot_xy][0]= mx;

  1184.     mv_table[mot_xy][1]= my;

  1185. 

  1186.     return dmin;

  1187. }

  1188. 

  1189. static inline int check_bidir_mv(MpegEncContext * s,

  1190.                    int motion_fx, int motion_fy,

  1191.                    int motion_bx, int motion_by,

  1192.                    int pred_fx, int pred_fy,

  1193.                    int pred_bx, int pred_by,

  1194.                    int size, int h)

  1195. {

  1196.     //FIXME optimize?

  1197.     //FIXME better f_code prediction (max mv & distance)

  1198.     //FIXME pointers

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

  1200.     uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame

  1201.     uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame

  1202.     int stride= c->stride;

  1203.     uint8_t *dest_y = c->scratchpad;

  1204.     uint8_t *ptr;

  1205.     int dxy;

  1206.     int src_x, src_y;

  1207.     int fbmin;

  1208.     uint8_t **src_data= c->src[0];

  1209.     uint8_t **ref_data= c->ref[0];

  1210.     uint8_t **ref2_data= c->ref[2];

  1211. 

  1212.     if(s->quarter_sample){

  1213.         dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);

  1214.         src_x = motion_fx >> 2;

  1215.         src_y = motion_fy >> 2;

  1216. 

  1217.         ptr = ref_data[0] + (src_y * stride) + src_x;

  1218.         s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride);

  1219. 

  1220.         dxy = ((motion_by & 3) << 2) | (motion_bx & 3);

  1221.         src_x = motion_bx >> 2;

  1222.         src_y = motion_by >> 2;

  1223. 

  1224.         ptr = ref2_data[0] + (src_y * stride) + src_x;

  1225.         s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride);

  1226.     }else{

  1227.         dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);

  1228.         src_x = motion_fx >> 1;

  1229.         src_y = motion_fy >> 1;

  1230. 

  1231.         ptr = ref_data[0] + (src_y * stride) + src_x;

  1232.         s->hdsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);

  1233. 

  1234.         dxy = ((motion_by & 1) << 1) | (motion_bx & 1);

  1235.         src_x = motion_bx >> 1;

  1236.         src_y = motion_by >> 1;

  1237. 

  1238.         ptr = ref2_data[0] + (src_y * stride) + src_x;

  1239.         s->hdsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);

  1240.     }

  1241. 

  1242.     fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor

  1243.            +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor

  1244.            + s->mecc.mb_cmp[size](s, src_data[0], dest_y, stride, h); // FIXME new_pic

  1245. 

  1246.     if(c->avctx->mb_cmp&FF_CMP_CHROMA){

  1247.     }

  1248.     //FIXME CHROMA !!!

  1249. 

  1250.     return fbmin;

  1251. }

  1252. 

  1253. /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/

  1254. static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)

  1255. {

  1256.     MotionEstContext * const c= &s->me;

  1257.     const int mot_stride = s->mb_stride;

  1258.     const int xy = mb_y *mot_stride + mb_x;

  1259.     int fbmin;

  1260.     int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];

  1261.     int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];

  1262.     int pred_bx= s->b_bidir_back_mv_table[xy-1][0];

  1263.     int pred_by= s->b_bidir_back_mv_table[xy-1][1];

  1264.     int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];

  1265.     int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];

  1266.     int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];

  1267.     int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];

  1268.     const int flags= c->sub_flags;

  1269.     const int qpel= flags&FLAG_QPEL;

  1270.     const int shift= 1+qpel;

  1271.     const int xmin= c->xmin<<shift;

  1272.     const int ymin= c->ymin<<shift;

  1273.     const int xmax= c->xmax<<shift;

  1274.     const int ymax= c->ymax<<shift;

  1275. #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))

  1276. #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))

  1277.     int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);

  1278.     uint8_t map[256] = { 0 };

  1279. 

  1280.     map[hashidx&255] = 1;

  1281. 

  1282.     fbmin= check_bidir_mv(s, motion_fx, motion_fy,

  1283.                           motion_bx, motion_by,

  1284.                           pred_fx, pred_fy,

  1285.                           pred_bx, pred_by,

  1286.                           0, 16);

  1287. 

  1288.     if(s->avctx->bidir_refine){

  1289.         int end;

  1290.         static const uint8_t limittab[5]={0,8,32,64,80};

  1291.         const int limit= limittab[s->avctx->bidir_refine];

  1292.         static const int8_t vect[][4]={

  1293. { 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},

  1294. 

  1295. { 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},

  1296. { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},

  1297. { 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},

  1298. { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},

  1299. 

  1300. { 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},

  1301. { 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},

  1302. { 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},

  1303. { 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},

  1304. 

  1305. { 1, 1, 1, 1}, {-1,-1,-1,-1},

  1306. { 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},

  1307. { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},

  1308.         };

  1309.         static const uint8_t hash[]={

  1310. 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),

  1311. 

  1312. 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),

  1313. HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),

  1314. 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),

  1315. HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),

  1316. 

  1317. 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),

  1318. 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),

  1319. 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),

  1320. 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),

  1321. 

  1322. HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),

  1323. 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),

  1324. 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),

  1325. };

  1326. 

  1327. #define CHECK_BIDIR(fx,fy,bx,by)\

  1328.     if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\

  1329.        &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\

  1330.        &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\

  1331.         int score;\

  1332.         map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\

  1333.         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);\

  1334.         if(score < fbmin){\

  1335.             hashidx += HASH(fx,fy,bx,by);\

  1336.             fbmin= score;\

  1337.             motion_fx+=fx;\

  1338.             motion_fy+=fy;\

  1339.             motion_bx+=bx;\

  1340.             motion_by+=by;\

  1341.             end=0;\

  1342.         }\

  1343.     }

  1344. #define CHECK_BIDIR2(a,b,c,d)\

  1345. CHECK_BIDIR(a,b,c,d)\

  1346. CHECK_BIDIR(-(a),-(b),-(c),-(d))

  1347. 

  1348.         do{

  1349.             int i;

  1350.             int borderdist=0;

  1351.             end=1;

  1352. 

  1353.             CHECK_BIDIR2(0,0,0,1)

  1354.             CHECK_BIDIR2(0,0,1,0)

  1355.             CHECK_BIDIR2(0,1,0,0)

  1356.             CHECK_BIDIR2(1,0,0,0)

  1357. 

  1358.             for(i=8; i<limit; i++){

  1359.                 int fx= motion_fx+vect[i][0];

  1360.                 int fy= motion_fy+vect[i][1];

  1361.                 int bx= motion_bx+vect[i][2];

  1362.                 int by= motion_by+vect[i][3];

  1363.                 if(borderdist<=0){

  1364.                     int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);

  1365.                     int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);

  1366.                     if((a|b) < 0)

  1367.                         map[(hashidx+hash[i])&255] = 1;

  1368.                 }

  1369.                 if(!map[(hashidx+hash[i])&255]){

  1370.                     int score;

  1371.                     map[(hashidx+hash[i])&255] = 1;

  1372.                     score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);

  1373.                     if(score < fbmin){

  1374.                         hashidx += hash[i];

  1375.                         fbmin= score;

  1376.                         motion_fx=fx;

  1377.                         motion_fy=fy;

  1378.                         motion_bx=bx;

  1379.                         motion_by=by;

  1380.                         end=0;

  1381.                         borderdist--;

  1382.                         if(borderdist<=0){

  1383.                             int a= FFMIN(xmax - FFMAX(fx,bx), FFMIN(fx,bx) - xmin);

  1384.                             int b= FFMIN(ymax - FFMAX(fy,by), FFMIN(fy,by) - ymin);

  1385.                             borderdist= FFMIN(a,b);

  1386.                         }

  1387.                     }

  1388.                 }

  1389.             }

  1390.         }while(!end);

  1391.     }

  1392. 

  1393.     s->b_bidir_forw_mv_table[xy][0]= motion_fx;

  1394.     s->b_bidir_forw_mv_table[xy][1]= motion_fy;

  1395.     s->b_bidir_back_mv_table[xy][0]= motion_bx;

  1396.     s->b_bidir_back_mv_table[xy][1]= motion_by;

  1397. 

  1398.     return fbmin;

  1399. }

  1400. 

  1401. static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)

  1402. {

  1403.     MotionEstContext * const c= &s->me;

  1404.     int P[10][2];

  1405.     const int mot_stride = s->mb_stride;

  1406.     const int mot_xy = mb_y*mot_stride + mb_x;

  1407.     const int shift= 1+s->quarter_sample;

  1408.     int dmin, i;

  1409.     const int time_pp= s->pp_time;

  1410.     const int time_pb= s->pb_time;

  1411.     int mx, my, xmin, xmax, ymin, ymax;

  1412.     int16_t (*mv_table)[2]= s->b_direct_mv_table;

  1413. 

  1414.     c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;

  1415.     ymin= xmin=(-32)>>shift;

  1416.     ymax= xmax=   31>>shift;

  1417. 

  1418.     if (IS_8X8(s->next_picture.mb_type[mot_xy])) {

  1419.         s->mv_type= MV_TYPE_8X8;

  1420.     }else{

  1421.         s->mv_type= MV_TYPE_16X16;

  1422.     }

  1423. 

  1424.     for(i=0; i<4; i++){

  1425.         int index= s->block_index[i];

  1426.         int min, max;

  1427. 

  1428.         c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];

  1429.         c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];

  1430.         c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));

  1431.         c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));

  1432. //        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);

  1433. //        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);

  1434. 

  1435.         max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;

  1436.         min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;

  1437.         max+= 16*mb_x + 1; // +-1 is for the simpler rounding

  1438.         min+= 16*mb_x - 1;

  1439.         xmax= FFMIN(xmax, s->width - max);

  1440.         xmin= FFMAX(xmin, - 16     - min);

  1441. 

  1442.         max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;

  1443.         min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;

  1444.         max+= 16*mb_y + 1; // +-1 is for the simpler rounding

  1445.         min+= 16*mb_y - 1;

  1446.         ymax= FFMIN(ymax, s->height - max);

  1447.         ymin= FFMAX(ymin, - 16      - min);

  1448. 

  1449.         if(s->mv_type == MV_TYPE_16X16) break;

  1450.     }

  1451. 

  1452.     assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);

  1453. 

  1454.     if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){

  1455.         s->b_direct_mv_table[mot_xy][0]= 0;

  1456.         s->b_direct_mv_table[mot_xy][1]= 0;

  1457. 

  1458.         return 256*256*256*64;

  1459.     }

  1460. 

  1461.     c->xmin= xmin;

  1462.     c->ymin= ymin;

  1463.     c->xmax= xmax;

  1464.     c->ymax= ymax;

  1465.     c->flags     |= FLAG_DIRECT;

  1466.     c->sub_flags |= FLAG_DIRECT;

  1467.     c->pred_x=0;

  1468.     c->pred_y=0;

  1469. 

  1470.     P_LEFT[0]        = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);

  1471.     P_LEFT[1]        = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);

  1472. 

  1473.     /* special case for first line */

  1474.     if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped

  1475.         P_TOP[0]      = av_clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);

  1476.         P_TOP[1]      = av_clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);

  1477.         P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);

  1478.         P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);

  1479. 

  1480.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1481.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1482.     }

  1483. 

  1484.     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);

  1485.     if(c->sub_flags&FLAG_QPEL)

  1486.         dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1487.     else

  1488.         dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1489. 

  1490.     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)

  1491.         dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);

  1492. 

  1493.     get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed

  1494. 

  1495.     mv_table[mot_xy][0]= mx;

  1496.     mv_table[mot_xy][1]= my;

  1497.     c->flags     &= ~FLAG_DIRECT;

  1498.     c->sub_flags &= ~FLAG_DIRECT;

  1499. 

  1500.     return dmin;

  1501. }

  1502. 

  1503. void ff_estimate_b_frame_motion(MpegEncContext * s,

  1504.                              int mb_x, int mb_y)

  1505. {

  1506.     MotionEstContext * const c= &s->me;

  1507.     const int penalty_factor= c->mb_penalty_factor;

  1508.     int fmin, bmin, dmin, fbmin, bimin, fimin;

  1509.     int type=0;

  1510.     const int xy = mb_y*s->mb_stride + mb_x;

  1511.     init_ref(c, s->new_picture.f->data, s->last_picture.f->data,

  1512.              s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2);

  1513. 

  1514.     get_limits(s, 16*mb_x, 16*mb_y);

  1515. 

  1516.     c->skip=0;

  1517. 

  1518.     if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {

  1519.         int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0

  1520. 

  1521.         score= ((unsigned)(score*score + 128*256))>>16;

  1522.         c->mc_mb_var_sum_temp += score;

  1523.         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE

  1524.         s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;

  1525. 

  1526.         return;

  1527.     }

  1528. 

  1529.     if (s->codec_id == AV_CODEC_ID_MPEG4)

  1530.         dmin= direct_search(s, mb_x, mb_y);

  1531.     else

  1532.         dmin= INT_MAX;

  1533. //FIXME penalty stuff for non mpeg4

  1534.     c->skip=0;

  1535.     fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +

  1536.            3 * penalty_factor;

  1537. 

  1538.     c->skip=0;

  1539.     bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +

  1540.            2 * penalty_factor;

  1541.     av_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

  1542. 

  1543.     c->skip=0;

  1544.     fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;

  1545.     av_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);

  1546. 

  1547.     if(s->flags & CODEC_FLAG_INTERLACED_ME){

  1548. //FIXME mb type penalty

  1549.         c->skip=0;

  1550.         c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

  1551.         fimin= interlaced_search(s, 0,

  1552.                                  s->b_field_mv_table[0], s->b_field_select_table[0],

  1553.                                  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);

  1554.         c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;

  1555.         bimin= interlaced_search(s, 2,

  1556.                                  s->b_field_mv_table[1], s->b_field_select_table[1],

  1557.                                  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);

  1558.     }else

  1559.         fimin= bimin= INT_MAX;

  1560. 

  1561.     {

  1562.         int score= fmin;

  1563.         type = CANDIDATE_MB_TYPE_FORWARD;

  1564. 

  1565.         if (dmin <= score){

  1566.             score = dmin;

  1567.             type = CANDIDATE_MB_TYPE_DIRECT;

  1568.         }

  1569.         if(bmin<score){

  1570.             score=bmin;

  1571.             type= CANDIDATE_MB_TYPE_BACKWARD;

  1572.         }

  1573.         if(fbmin<score){

  1574.             score=fbmin;

  1575.             type= CANDIDATE_MB_TYPE_BIDIR;

  1576.         }

  1577.         if(fimin<score){

  1578.             score=fimin;

  1579.             type= CANDIDATE_MB_TYPE_FORWARD_I;

  1580.         }

  1581.         if(bimin<score){

  1582.             score=bimin;

  1583.             type= CANDIDATE_MB_TYPE_BACKWARD_I;

  1584.         }

  1585. 

  1586.         score= ((unsigned)(score*score + 128*256))>>16;

  1587.         c->mc_mb_var_sum_temp += score;

  1588.         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE

  1589.     }

  1590. 

  1591.     if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){

  1592.         type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;

  1593.         if(fimin < INT_MAX)

  1594.             type |= CANDIDATE_MB_TYPE_FORWARD_I;

  1595.         if(bimin < INT_MAX)

  1596.             type |= CANDIDATE_MB_TYPE_BACKWARD_I;

  1597.         if(fimin < INT_MAX && bimin < INT_MAX){

  1598.             type |= CANDIDATE_MB_TYPE_BIDIR_I;

  1599.         }

  1600.          //FIXME something smarter

  1601.         if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB

  1602.         if (s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT &&

  1603.             s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])

  1604.             type |= CANDIDATE_MB_TYPE_DIRECT0;

  1605.     }

  1606. 

  1607.     s->mb_type[mb_y*s->mb_stride + mb_x]= type;

  1608. }

  1609. 

  1610. /* find best f_code for ME which do unlimited searches */

  1611. int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)

  1612. {

  1613.     if(s->me_method>=ME_EPZS){

  1614.         int score[8];

  1615.         int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);

  1616.         uint8_t * fcode_tab= s->fcode_tab;

  1617.         int best_fcode=-1;

  1618.         int best_score=-10000000;

  1619. 

  1620.         if(s->msmpeg4_version)

  1621.             range= FFMIN(range, 16);

  1622.         else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)

  1623.             range= FFMIN(range, 256);

  1624. 

  1625.         for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);

  1626. 

  1627.         for(y=0; y<s->mb_height; y++){

  1628.             int x;

  1629.             int xy= y*s->mb_stride;

  1630.             for(x=0; x<s->mb_width; x++){

  1631.                 if(s->mb_type[xy] & type){

  1632.                     int mx= mv_table[xy][0];

  1633.                     int my= mv_table[xy][1];

  1634.                     int fcode= FFMAX(fcode_tab[mx + MAX_MV],

  1635.                                      fcode_tab[my + MAX_MV]);

  1636.                     int j;

  1637. 

  1638.                         if(mx >= range || mx < -range ||

  1639.                            my >= range || my < -range)

  1640.                             continue;

  1641. 

  1642.                     for(j=0; j<fcode && j<8; j++){

  1643.                         if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])

  1644.                             score[j]-= 170;

  1645.                     }

  1646.                 }

  1647.                 xy++;

  1648.             }

  1649.         }

  1650. 

  1651.         for(i=1; i<8; i++){

  1652.             if(score[i] > best_score){

  1653.                 best_score= score[i];

  1654.                 best_fcode= i;

  1655.             }

  1656.         }

  1657. 

  1658.         return best_fcode;

  1659.     }else{

  1660.         return 1;

  1661.     }

  1662. }

  1663. 

  1664. void ff_fix_long_p_mvs(MpegEncContext * s)

  1665. {

  1666.     MotionEstContext * const c= &s->me;

  1667.     const int f_code= s->f_code;

  1668.     int y, range;

  1669.     assert(s->pict_type==AV_PICTURE_TYPE_P);

  1670. 

  1671.     range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);

  1672. 

  1673.     assert(range <= 16 || !s->msmpeg4_version);

  1674.     assert(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));

  1675. 

  1676.     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;

  1677. 

  1678.     if(s->flags&CODEC_FLAG_4MV){

  1679.         const int wrap= s->b8_stride;

  1680. 

  1681.         /* clip / convert to intra 8x8 type MVs */

  1682.         for(y=0; y<s->mb_height; y++){

  1683.             int xy= y*2*wrap;

  1684.             int i= y*s->mb_stride;

  1685.             int x;

  1686. 

  1687.             for(x=0; x<s->mb_width; x++){

  1688.                 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){

  1689.                     int block;

  1690.                     for(block=0; block<4; block++){

  1691.                         int off= (block& 1) + (block>>1)*wrap;

  1692.                         int mx = s->current_picture.motion_val[0][ xy + off ][0];

  1693.                         int my = s->current_picture.motion_val[0][ xy + off ][1];

  1694. 

  1695.                         if(   mx >=range || mx <-range

  1696.                            || my >=range || my <-range){

  1697.                             s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;

  1698.                             s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;

  1699.                             s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;

  1700.                         }

  1701.                     }

  1702.                 }

  1703.                 xy+=2;

  1704.                 i++;

  1705.             }

  1706.         }

  1707.     }

  1708. }

  1709. 

  1710. /**

  1711.  *

  1712.  * @param truncate 1 for truncation, 0 for using intra

  1713.  */

  1714. void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,

  1715.                      int16_t (*mv_table)[2], int f_code, int type, int truncate)

  1716. {

  1717.     MotionEstContext * const c= &s->me;

  1718.     int y, h_range, v_range;

  1719. 

  1720.     // RAL: 8 in MPEG-1, 16 in MPEG-4

  1721.     int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);

  1722. 

  1723.     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;

  1724. 

  1725.     h_range= range;

  1726.     v_range= field_select_table ? range>>1 : range;

  1727. 

  1728.     /* clip / convert to intra 16x16 type MVs */

  1729.     for(y=0; y<s->mb_height; y++){

  1730.         int x;

  1731.         int xy= y*s->mb_stride;

  1732.         for(x=0; x<s->mb_width; x++){

  1733.             if (s->mb_type[xy] & type){    // RAL: "type" test added...

  1734.                 if(field_select_table==NULL || field_select_table[xy] == field_select){

  1735.                     if(   mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range

  1736.                        || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){

  1737. 

  1738.                         if(truncate){

  1739.                             if     (mv_table[xy][0] > h_range-1) mv_table[xy][0]=  h_range-1;

  1740.                             else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;

  1741.                             if     (mv_table[xy][1] > v_range-1) mv_table[xy][1]=  v_range-1;

  1742.                             else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;

  1743.                         }else{

  1744.                             s->mb_type[xy] &= ~type;

  1745.                             s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;

  1746.                             mv_table[xy][0]=

  1747.                             mv_table[xy][1]= 0;

  1748.                         }

  1749.                     }

  1750.                 }

  1751.             }

  1752.             xy++;

  1753.         }

  1754.     }

  1755. }