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.
41838 Commits
32 Branches
283MB
Tree: 3058872c29
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '3058872c29'
${ noResults }
FFmpeg/libavcodec/motion_est.c

1737 lines
66KB
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 "internal.h"

  36. #include "mathops.h"

  37. #include "mpegutils.h"

  38. #include "mpegvideo.h"

  39. 

  40. #undef NDEBUG

  41. #include <assert.h>

  42. 

  43. #define P_LEFT P[1]

  44. #define P_TOP P[2]

  45. #define P_TOPRIGHT P[3]

  46. #define P_MEDIAN P[4]

  47. #define P_MV1 P[9]

  48. 

  49. #define ME_MAP_SHIFT 3

  50. #define ME_MAP_MV_BITS 11

  51. 

  52. static int sad_hpel_motion_search(MpegEncContext * s,

  53.                                   int *mx_ptr, int *my_ptr, int dmin,

  54.                                   int src_index, int ref_index,

  55.                                   int size, int h);

  56. 

  57. static inline unsigned update_map_generation(MotionEstContext *c)

  58. {

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

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

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

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

  63.     }

  64.     return c->map_generation;

  65. }

  66. 

  67. /* shape adaptive search stuff */

  68. typedef struct Minima{

  69.     int height;

  70.     int x, y;

  71.     int checked;

  72. }Minima;

  73. 

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

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

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

  77. 

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

  79. }

  80. 

  81. #define FLAG_QPEL   1 //must be 1

  82. #define FLAG_CHROMA 2

  83. #define FLAG_DIRECT 4

  84. 

  85. 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){

  86.     const int offset[3]= {

  87.           y*c->  stride + x,

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

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

  90.     };

  91.     int i;

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

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

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

  95.     }

  96.     if(ref_index){

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

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

  99.         }

  100.     }

  101. }

  102. 

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

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

  105.            + (direct ? FLAG_DIRECT : 0)

  106.            + (chroma ? FLAG_CHROMA : 0);

  107. }

  108. 

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

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

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

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

  113.     const int stride= c->stride;

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

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

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

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

  118.     int d;

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

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

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

  122.             const int time_pp= s->pp_time;

  123.             const int time_pb= s->pb_time;

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

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

  126.                 int i;

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

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

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

  130.                     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));

  131.                     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));

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

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

  134. 

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

  136.                     if(qpel){

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

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

  139.                     }else{

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

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

  142.                     }

  143.                 }

  144.             }else{

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

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

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

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

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

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

  151. 

  152.                 if(qpel){

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

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

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

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

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

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

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

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

  161.                 }else{

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

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

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

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

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

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

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

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

  170. 

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

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

  173.                 }

  174.             }

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

  176.         }else

  177.             d= 256*256*256*32;

  178.     return d;

  179. }

  180. 

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

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

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

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

  185.     const int stride= c->stride;

  186.     const int uvstride= c->uvstride;

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

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

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

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

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

  192.     int d;

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

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

  195.         if(dxy){

  196.             if(qpel){

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

  198.                 if(chroma){

  199.                     int cx= hx/2;

  200.                     int cy= hy/2;

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

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

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

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

  205.                 }

  206.             }else{

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

  208.                 if(chroma)

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

  210.             }

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

  212.         }else{

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

  214.             if(chroma)

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

  216.         }

  217.         if(chroma){

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

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

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

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

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

  223.         }

  224.     return d;

  225. }

  226. 

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

  228.                       int ref_index, int src_index,

  229.                       me_cmp_func cmp_func, me_cmp_func chroma_cmp_func){

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

  231. }

  232. 

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

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

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

  236.     if(flags&FLAG_DIRECT){

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

  238.     }else{

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

  240.     }

  241. }

  242. 

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

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

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

  246.     if(flags&FLAG_DIRECT){

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

  248.     }else{

  249.         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);

  250.     }

  251. }

  252. 

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

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

  255.  */

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

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

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

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

  260.        && av_builtin_constant_p(subx) && av_builtin_constant_p(suby)

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

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

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

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

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

  266.     }else{

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

  268.     }

  269. }

  270. 

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

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

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

  274.     if(flags&FLAG_DIRECT){

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

  276.     }else{

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

  278.     }

  279. }

  280. 

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

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

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

  284.     if(flags&FLAG_DIRECT){

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

  286.     }else{

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

  288.     }

  289. }

  290. 

  291. #include "motion_est_template.c"

  292. 

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

  294.                     ptrdiff_t stride, int h)

  295. {

  296.     return 0;

  297. }

  298. 

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

  300. }

  301. 

  302. int ff_init_me(MpegEncContext *s){

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

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

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

  306. 

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

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

  309.         return -1;

  310.     }

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

  312.         s->me_method != ME_EPZS &&

  313.         s->me_method != ME_X1) {

  314.         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");

  315.         return -1;

  316.     }

  317. 

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

  319. 

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

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

  322.     }

  323. 

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

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

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

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

  328. 

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

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

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

  332. 

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

  334.     if (s->avctx->flags & CODEC_FLAG_QPEL) {

  335.         c->sub_motion_search= qpel_motion_search;

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

  337.         if (s->no_rounding)

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

  339.         else

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

  341.     }else{

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

  343.             c->sub_motion_search= hpel_motion_search;

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

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

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

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

  348.         else

  349.             c->sub_motion_search= hpel_motion_search;

  350.     }

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

  352.     if (s->no_rounding)

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

  354.     else

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

  356. 

  357.     if(s->linesize){

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

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

  360.     }else{

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

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

  363.     }

  364. 

  365.     /* 8x8 fullpel search would need a 4x4 chroma compare, which we do

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

  367.      * does not expect it. */

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

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

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

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

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

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

  374. 

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

  376.         c->sub_motion_search= no_sub_motion_search;

  377.     }

  378. 

  379.     return 0;

  380. }

  381. 

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

  383. {\

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

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

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

  387. }

  388. 

  389. static int sad_hpel_motion_search(MpegEncContext * s,

  390.                                   int *mx_ptr, int *my_ptr, int dmin,

  391.                                   int src_index, int ref_index,

  392.                                   int size, int h)

  393. {

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

  395.     const int penalty_factor= c->sub_penalty_factor;

  396.     int mx, my, dminh;

  397.     uint8_t *pix, *ptr;

  398.     int stride= c->stride;

  399.     const int flags= c->sub_flags;

  400.     LOAD_COMMON

  401. 

  402.     assert(flags == 0);

  403. 

  404.     if(c->skip){

  405.         *mx_ptr = 0;

  406.         *my_ptr = 0;

  407.         return dmin;

  408.     }

  409. 

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

  411. 

  412.     mx = *mx_ptr;

  413.     my = *my_ptr;

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

  415. 

  416.     dminh = dmin;

  417. 

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

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

  420.         int dx=0, dy=0;

  421.         int d, pen_x, pen_y;

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

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

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

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

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

  427.         mx<<=1;

  428.         my<<=1;

  429. 

  430. 

  431.         pen_x= pred_x + mx;

  432.         pen_y= pred_y + my;

  433. 

  434.         ptr-= stride;

  435.         if(t<=b){

  436.             CHECK_SAD_HALF_MV(y2 , 0, -1)

  437.             if(l<=r){

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

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

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

  441.                     ptr+= stride;

  442.                 }else{

  443.                     ptr+= stride;

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

  445.                 }

  446.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

  447.             }else{

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

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

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

  451.                     ptr+= stride;

  452.                 }else{

  453.                     ptr+= stride;

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

  455.                 }

  456.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

  457.             }

  458.         }else{

  459.             if(l<=r){

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

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

  462.                     ptr+= stride;

  463.                 }else{

  464.                     ptr+= stride;

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

  466.                 }

  467.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

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

  469.             }else{

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

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

  472.                     ptr+= stride;

  473.                 }else{

  474.                     ptr+= stride;

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

  476.                 }

  477.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

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

  479.             }

  480.             CHECK_SAD_HALF_MV(y2 ,  0, +1)

  481.         }

  482.         mx+=dx;

  483.         my+=dy;

  484. 

  485.     }else{

  486.         mx<<=1;

  487.         my<<=1;

  488.     }

  489. 

  490.     *mx_ptr = mx;

  491.     *my_ptr = my;

  492.     return dminh;

  493. }

  494. 

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

  496. {

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

  498. 

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

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

  501. 

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

  503.     if(mv4){

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

  505. 

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

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

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

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

  510. 

  511.         mot_xy += s->b8_stride;

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

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

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

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

  516.     }

  517. }

  518. 

  519. /**

  520.  * get fullpel ME search limits.

  521.  */

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

  523. {

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

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

  526. /*

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

  528.     else                   c->range= 16;

  529. */

  530.     if (s->unrestricted_mv) {

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

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

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

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

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

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

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

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

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

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

  541.     } else {

  542.         c->xmin = - x;

  543.         c->ymin = - y;

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

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

  546.     }

  547.     if(range){

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

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

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

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

  552.     }

  553. }

  554. 

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

  556.     const int stride= c->stride;

  557. 

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

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

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

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

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

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

  564. }

  565. 

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

  567. {

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

  569.     const int size= 1;

  570.     const int h=8;

  571.     int block;

  572.     int P[10][2];

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

  574.     int same=1;

  575.     const int stride= c->stride;

  576.     uint8_t *mv_penalty= c->current_mv_penalty;

  577. 

  578.     init_mv4_ref(c);

  579. 

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

  581.         int mx4, my4;

  582.         int pred_x4, pred_y4;

  583.         int dmin4;

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

  585.         const int mot_stride = s->b8_stride;

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

  587. 

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

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

  590. 

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

  592. 

  593.         /* special case for first line */

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

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

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

  597.         } else {

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

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

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

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

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

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

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

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

  606. 

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

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

  609. 

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

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

  612.         }

  613.         P_MV1[0]= mx;

  614.         P_MV1[1]= my;

  615. 

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

  617. 

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

  619. 

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

  621.             int dxy;

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

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

  624.             if(s->quarter_sample){

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

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

  627. 

  628.                 if(s->no_rounding)

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

  630.                 else

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

  632.             }else{

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

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

  635. 

  636.                 if(s->no_rounding)

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

  638.                 else

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

  640.             }

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

  642.         }else

  643.             dmin_sum+= dmin4;

  644. 

  645.         if(s->quarter_sample){

  646.             mx4_sum+= mx4/2;

  647.             my4_sum+= my4/2;

  648.         }else{

  649.             mx4_sum+= mx4;

  650.             my4_sum+= my4;

  651.         }

  652. 

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

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

  655. 

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

  657.     }

  658. 

  659.     if(same)

  660.         return INT_MAX;

  661. 

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

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

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

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

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

  667.     }

  668. 

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

  670.         int dxy;

  671.         int mx, my;

  672.         int offset;

  673. 

  674.         mx= ff_h263_round_chroma(mx4_sum);

  675.         my= ff_h263_round_chroma(my4_sum);

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

  677. 

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

  679. 

  680.         if(s->no_rounding){

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

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

  683.         }else{

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

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

  686.         }

  687. 

  688.         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);

  689.         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);

  690.     }

  691. 

  692.     c->pred_x= mx;

  693.     c->pred_y= my;

  694. 

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

  696.     /*case FF_CMP_SSE:

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

  698.     case FF_CMP_RD:

  699.         return dmin_sum;

  700.     default:

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

  702.     }

  703. }

  704. 

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

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

  707. 

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

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

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

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

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

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

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

  715.     }

  716. }

  717. 

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

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

  720. {

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

  722.     const int size=0;

  723.     const int h=8;

  724.     int block;

  725.     int P[10][2];

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

  727.     int same=1;

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

  729.     int dmin_sum= 0;

  730.     const int mot_stride= s->mb_stride;

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

  732. 

  733.     c->ymin>>=1;

  734.     c->ymax>>=1;

  735.     c->stride<<=1;

  736.     c->uvstride<<=1;

  737.     init_interlaced_ref(s, ref_index);

  738. 

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

  740.         int field_select;

  741.         int best_dmin= INT_MAX;

  742.         int best_field= -1;

  743. 

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

  745.             int dmin, mx_i, my_i;

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

  747. 

  748.             if(user_field_select){

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

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

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

  752.                     continue;

  753.             }

  754. 

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

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

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

  758. 

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

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

  761. 

  762.             if(!s->first_slice_line){

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

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

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

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

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

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

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

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

  771. 

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

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

  774.             }

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

  776.             P_MV1[1]= my / 2;

  777. 

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

  779. 

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

  781. 

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

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

  784. 

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

  786.                 int dxy;

  787. 

  788.                 //FIXME chroma ME

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

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

  791. 

  792.                 if(s->no_rounding){

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

  794.                 }else{

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

  796.                 }

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

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

  799.             }else

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

  801. 

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

  803. 

  804.             if(dmin < best_dmin){

  805.                 best_dmin= dmin;

  806.                 best_field= field_select;

  807.             }

  808.         }

  809.         {

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

  811. 

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

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

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

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

  816.         }

  817. 

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

  819.         dmin_sum += best_dmin;

  820.     }

  821. 

  822.     c->ymin<<=1;

  823.     c->ymax<<=1;

  824.     c->stride>>=1;

  825.     c->uvstride>>=1;

  826. 

  827.     if(same)

  828.         return INT_MAX;

  829. 

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

  831.     /*case FF_CMP_SSE:

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

  833.     case FF_CMP_RD:

  834.         return dmin_sum;

  835.     default:

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

  837.     }

  838. }

  839. 

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

  841.     switch(type&0xFF){

  842.     default:

  843.     case FF_CMP_SAD:

  844.         return lambda>>FF_LAMBDA_SHIFT;

  845.     case FF_CMP_DCT:

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

  847.     case FF_CMP_SATD:

  848.     case FF_CMP_DCT264:

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

  850.     case FF_CMP_RD:

  851.     case FF_CMP_PSNR:

  852.     case FF_CMP_SSE:

  853.     case FF_CMP_NSSE:

  854.         return lambda2>>FF_LAMBDA_SHIFT;

  855.     case FF_CMP_BIT:

  856.         return 1;

  857.     }

  858. }

  859. 

  860. void ff_estimate_p_frame_motion(MpegEncContext * s,

  861.                                 int mb_x, int mb_y)

  862. {

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

  864.     uint8_t *pix, *ppix;

  865.     int sum, mx, my, dmin;

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

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

  868.     int P[10][2];

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

  870.     int mb_type=0;

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

  872. 

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

  874. 

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

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

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

  878. 

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

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

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

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

  883. 

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

  885.     c->skip=0;

  886. 

  887.     /* intra / predictive decision */

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

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

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

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

  892. 

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

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

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

  896. 

  897.     switch(s->me_method) {

  898.     case ME_ZERO:

  899.     default:

  900.         mx   = 0;

  901.         my   = 0;

  902.         dmin = 0;

  903.         break;

  904.     case ME_X1:

  905.     case ME_EPZS:

  906.        {

  907.             const int mot_stride = s->b8_stride;

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

  909. 

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

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

  912. 

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

  914. 

  915.             if(!s->first_slice_line) {

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

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

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

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

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

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

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

  923. 

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

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

  926. 

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

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

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

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

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

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

  933.                 }

  934.             }else{

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

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

  937.             }

  938. 

  939.         }

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

  941. 

  942.         break;

  943.     }

  944. 

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

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

  947. 

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

  949. 

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

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

  952. 

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

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

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

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

  957. 

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

  959.             mb_type|= CANDIDATE_MB_TYPE_INTRA;

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

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

  962.             mb_type|= CANDIDATE_MB_TYPE_INTER;

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

  964.             if (s->mpv_flags & FF_MPV_FLAG_MV0)

  965.                 if(mx || my)

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

  967.         }else{

  968.             mx <<=shift;

  969.             my <<=shift;

  970.         }

  971.         if ((s->avctx->flags & CODEC_FLAG_4MV)

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

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

  974.                 mb_type|=CANDIDATE_MB_TYPE_INTER4V;

  975. 

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

  977.         }else

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

  979.         if ((s->avctx->flags & CODEC_FLAG_INTERLACED_ME)

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

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

  982.                 mb_type |= CANDIDATE_MB_TYPE_INTER_I;

  983.         }

  984.     }else{

  985.         int intra_score, i;

  986.         mb_type= CANDIDATE_MB_TYPE_INTER;

  987. 

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

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

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

  991. 

  992.         if ((s->avctx->flags & CODEC_FLAG_4MV)

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

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

  995.             if(dmin4 < dmin){

  996.                 mb_type= CANDIDATE_MB_TYPE_INTER4V;

  997.                 dmin=dmin4;

  998.             }

  999.         }

  1000.         if ((s->avctx->flags & CODEC_FLAG_INTERLACED_ME)

  1001.            && !c->skip){ //FIXME varc/d checks

  1002.             int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);

  1003.             if(dmin_i < dmin){

  1004.                 mb_type = CANDIDATE_MB_TYPE_INTER_I;

  1005.                 dmin= dmin_i;

  1006.             }

  1007.         }

  1008. 

  1009.         set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);

  1010. 

  1011.         /* get intra luma score */

  1012.         if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){

  1013.             intra_score= varc - 500;

  1014.         }else{

  1015.             unsigned mean = (sum+128)>>8;

  1016.             mean*= 0x01010101;

  1017. 

  1018.             for(i=0; i<16; i++){

  1019.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;

  1020.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;

  1021.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;

  1022.                 *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;

  1023.             }

  1024. 

  1025.             intra_score= s->mecc.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);

  1026.         }

  1027.         intra_score += c->mb_penalty_factor*16;

  1028. 

  1029.         if(intra_score < dmin){

  1030.             mb_type= CANDIDATE_MB_TYPE_INTRA;

  1031.             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup

  1032.         }else

  1033.             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x] = 0;

  1034. 

  1035.         {

  1036.             int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);

  1037.             int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;

  1038.             c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);

  1039.         }

  1040.     }

  1041. 

  1042.     s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;

  1043. }

  1044. 

  1045. int ff_pre_estimate_p_frame_motion(MpegEncContext * s,

  1046.                                     int mb_x, int mb_y)

  1047. {

  1048.     MotionEstContext * const c= &s->me;

  1049.     int mx, my, dmin;

  1050.     int P[10][2];

  1051.     const int shift= 1+s->quarter_sample;

  1052.     const int xy= mb_x + mb_y*s->mb_stride;

  1053.     init_ref(c, s->new_picture.f->data, s->last_picture.f->data, NULL, 16*mb_x, 16*mb_y, 0);

  1054. 

  1055.     assert(s->quarter_sample==0 || s->quarter_sample==1);

  1056. 

  1057.     c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);

  1058.     c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

  1059. 

  1060.     get_limits(s, 16*mb_x, 16*mb_y);

  1061.     c->skip=0;

  1062. 

  1063.     P_LEFT[0]       = s->p_mv_table[xy + 1][0];

  1064.     P_LEFT[1]       = s->p_mv_table[xy + 1][1];

  1065. 

  1066.     if(P_LEFT[0]       < (c->xmin<<shift)) P_LEFT[0]       = (c->xmin<<shift);

  1067. 

  1068.     /* special case for first line */

  1069.     if (s->first_slice_line) {

  1070.         c->pred_x= P_LEFT[0];

  1071.         c->pred_y= P_LEFT[1];

  1072.         P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=

  1073.         P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME

  1074.     } else {

  1075.         P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];

  1076.         P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];

  1077.         P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];

  1078.         P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];

  1079.         if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);

  1080.         if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);

  1081.         if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);

  1082. 

  1083.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1084.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1085. 

  1086.         c->pred_x = P_MEDIAN[0];

  1087.         c->pred_y = P_MEDIAN[1];

  1088.     }

  1089. 

  1090.     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);

  1091. 

  1092.     s->p_mv_table[xy][0] = mx<<shift;

  1093.     s->p_mv_table[xy][1] = my<<shift;

  1094. 

  1095.     return dmin;

  1096. }

  1097. 

  1098. static int estimate_motion_b(MpegEncContext *s, int mb_x, int mb_y,

  1099.                              int16_t (*mv_table)[2], int ref_index, int f_code)

  1100. {

  1101.     MotionEstContext * const c= &s->me;

  1102.     int mx, my, dmin;

  1103.     int P[10][2];

  1104.     const int shift= 1+s->quarter_sample;

  1105.     const int mot_stride = s->mb_stride;

  1106.     const int mot_xy = mb_y*mot_stride + mb_x;

  1107.     uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;

  1108.     int mv_scale;

  1109. 

  1110.     c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);

  1111.     c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);

  1112.     c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);

  1113.     c->current_mv_penalty= mv_penalty;

  1114. 

  1115.     get_limits(s, 16*mb_x, 16*mb_y);

  1116. 

  1117.     switch(s->me_method) {

  1118.     case ME_ZERO:

  1119.     default:

  1120.         mx   = 0;

  1121.         my   = 0;

  1122.         dmin = 0;

  1123.         break;

  1124.     case ME_X1:

  1125.     case ME_EPZS:

  1126.         P_LEFT[0] = mv_table[mot_xy - 1][0];

  1127.         P_LEFT[1] = mv_table[mot_xy - 1][1];

  1128. 

  1129.         if (P_LEFT[0] > (c->xmax << shift)) P_LEFT[0] = (c->xmax << shift);

  1130. 

  1131.         /* special case for first line */

  1132.         if (!s->first_slice_line) {

  1133.             P_TOP[0]      = mv_table[mot_xy - mot_stride    ][0];

  1134.             P_TOP[1]      = mv_table[mot_xy - mot_stride    ][1];

  1135.             P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1][0];

  1136.             P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1][1];

  1137.             if (P_TOP[1] > (c->ymax << shift)) P_TOP[1] = (c->ymax << shift);

  1138.             if (P_TOPRIGHT[0] < (c->xmin << shift)) P_TOPRIGHT[0] = (c->xmin << shift);

  1139.             if (P_TOPRIGHT[1] > (c->ymax << shift)) P_TOPRIGHT[1] = (c->ymax << shift);

  1140. 

  1141.             P_MEDIAN[0] = mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1142.             P_MEDIAN[1] = mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1143.         }

  1144.         c->pred_x = P_LEFT[0];

  1145.         c->pred_y = P_LEFT[1];

  1146. 

  1147.         if(mv_table == s->b_forw_mv_table){

  1148.             mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);

  1149.         }else{

  1150.             mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);

  1151.         }

  1152. 

  1153.         dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);

  1154. 

  1155.         break;

  1156.     }

  1157. 

  1158.     dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);

  1159. 

  1160.     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)

  1161.         dmin= get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);

  1162. 

  1163. //    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;

  1164.     mv_table[mot_xy][0]= mx;

  1165.     mv_table[mot_xy][1]= my;

  1166. 

  1167.     return dmin;

  1168. }

  1169. 

  1170. static inline int check_bidir_mv(MpegEncContext * s,

  1171.                    int motion_fx, int motion_fy,

  1172.                    int motion_bx, int motion_by,

  1173.                    int pred_fx, int pred_fy,

  1174.                    int pred_bx, int pred_by,

  1175.                    int size, int h)

  1176. {

  1177.     //FIXME optimize?

  1178.     //FIXME better f_code prediction (max mv & distance)

  1179.     //FIXME pointers

  1180.     MotionEstContext * const c= &s->me;

  1181.     uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame

  1182.     uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame

  1183.     int stride= c->stride;

  1184.     uint8_t *dest_y = c->scratchpad;

  1185.     uint8_t *ptr;

  1186.     int dxy;

  1187.     int src_x, src_y;

  1188.     int fbmin;

  1189.     uint8_t **src_data= c->src[0];

  1190.     uint8_t **ref_data= c->ref[0];

  1191.     uint8_t **ref2_data= c->ref[2];

  1192. 

  1193.     if(s->quarter_sample){

  1194.         dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);

  1195.         src_x = motion_fx >> 2;

  1196.         src_y = motion_fy >> 2;

  1197. 

  1198.         ptr = ref_data[0] + (src_y * stride) + src_x;

  1199.         s->qdsp.put_qpel_pixels_tab[0][dxy](dest_y, ptr, stride);

  1200. 

  1201.         dxy = ((motion_by & 3) << 2) | (motion_bx & 3);

  1202.         src_x = motion_bx >> 2;

  1203.         src_y = motion_by >> 2;

  1204. 

  1205.         ptr = ref2_data[0] + (src_y * stride) + src_x;

  1206.         s->qdsp.avg_qpel_pixels_tab[size][dxy](dest_y, ptr, stride);

  1207.     }else{

  1208.         dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);

  1209.         src_x = motion_fx >> 1;

  1210.         src_y = motion_fy >> 1;

  1211. 

  1212.         ptr = ref_data[0] + (src_y * stride) + src_x;

  1213.         s->hdsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);

  1214. 

  1215.         dxy = ((motion_by & 1) << 1) | (motion_bx & 1);

  1216.         src_x = motion_bx >> 1;

  1217.         src_y = motion_by >> 1;

  1218. 

  1219.         ptr = ref2_data[0] + (src_y * stride) + src_x;

  1220.         s->hdsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);

  1221.     }

  1222. 

  1223.     fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor

  1224.            +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor

  1225.            + s->mecc.mb_cmp[size](s, src_data[0], dest_y, stride, h); // FIXME new_pic

  1226. 

  1227.     if(c->avctx->mb_cmp&FF_CMP_CHROMA){

  1228.     }

  1229.     //FIXME CHROMA !!!

  1230. 

  1231.     return fbmin;

  1232. }

  1233. 

  1234. /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/

  1235. static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)

  1236. {

  1237.     MotionEstContext * const c= &s->me;

  1238.     const int mot_stride = s->mb_stride;

  1239.     const int xy = mb_y *mot_stride + mb_x;

  1240.     int fbmin;

  1241.     int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];

  1242.     int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];

  1243.     int pred_bx= s->b_bidir_back_mv_table[xy-1][0];

  1244.     int pred_by= s->b_bidir_back_mv_table[xy-1][1];

  1245.     int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];

  1246.     int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];

  1247.     int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];

  1248.     int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];

  1249.     const int flags= c->sub_flags;

  1250.     const int qpel= flags&FLAG_QPEL;

  1251.     const int shift= 1+qpel;

  1252.     const int xmin= c->xmin<<shift;

  1253.     const int ymin= c->ymin<<shift;

  1254.     const int xmax= c->xmax<<shift;

  1255.     const int ymax= c->ymax<<shift;

  1256. #define HASH(fx,fy,bx,by) ((fx)+17*(fy)+63*(bx)+117*(by))

  1257. #define HASH8(fx,fy,bx,by) ((uint8_t)HASH(fx,fy,bx,by))

  1258.     int hashidx= HASH(motion_fx,motion_fy, motion_bx, motion_by);

  1259.     uint8_t map[256] = { 0 };

  1260. 

  1261.     map[hashidx&255] = 1;

  1262. 

  1263.     fbmin= check_bidir_mv(s, motion_fx, motion_fy,

  1264.                           motion_bx, motion_by,

  1265.                           pred_fx, pred_fy,

  1266.                           pred_bx, pred_by,

  1267.                           0, 16);

  1268. 

  1269.     if(s->avctx->bidir_refine){

  1270.         int end;

  1271.         static const uint8_t limittab[5]={0,8,32,64,80};

  1272.         const int limit= limittab[s->avctx->bidir_refine];

  1273.         static const int8_t vect[][4]={

  1274. { 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},

  1275. 

  1276. { 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},

  1277. { 0, 1, 0, 1}, { 0,-1, 0,-1}, { 1, 0, 1, 0}, {-1, 0,-1, 0},

  1278. { 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},

  1279. { 0,-1, 0, 1}, { 0, 1, 0,-1}, {-1, 0, 1, 0}, { 1, 0,-1, 0},

  1280. 

  1281. { 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},

  1282. { 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},

  1283. { 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},

  1284. { 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},

  1285. 

  1286. { 1, 1, 1, 1}, {-1,-1,-1,-1},

  1287. { 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},

  1288. { 1, 1,-1,-1}, {-1,-1, 1, 1}, { 1,-1,-1, 1}, {-1, 1, 1,-1}, { 1,-1, 1,-1}, {-1, 1,-1, 1},

  1289.         };

  1290.         static const uint8_t hash[]={

  1291. 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),

  1292. 

  1293. 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),

  1294. HASH8( 0, 1, 0, 1), HASH8( 0,-1, 0,-1), HASH8( 1, 0, 1, 0), HASH8(-1, 0,-1, 0),

  1295. 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),

  1296. HASH8( 0,-1, 0, 1), HASH8( 0, 1, 0,-1), HASH8(-1, 0, 1, 0), HASH8( 1, 0,-1, 0),

  1297. 

  1298. 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),

  1299. 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),

  1300. 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),

  1301. 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),

  1302. 

  1303. HASH8( 1, 1, 1, 1), HASH8(-1,-1,-1,-1),

  1304. 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),

  1305. 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),

  1306. };

  1307. 

  1308. #define CHECK_BIDIR(fx,fy,bx,by)\

  1309.     if( !map[(hashidx+HASH(fx,fy,bx,by))&255]\

  1310.        &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\

  1311.        &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\

  1312.         int score;\

  1313.         map[(hashidx+HASH(fx,fy,bx,by))&255] = 1;\

  1314.         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);\

  1315.         if(score < fbmin){\

  1316.             hashidx += HASH(fx,fy,bx,by);\

  1317.             fbmin= score;\

  1318.             motion_fx+=fx;\

  1319.             motion_fy+=fy;\

  1320.             motion_bx+=bx;\

  1321.             motion_by+=by;\

  1322.             end=0;\

  1323.         }\

  1324.     }

  1325. #define CHECK_BIDIR2(a,b,c,d)\

  1326. CHECK_BIDIR(a,b,c,d)\

  1327. CHECK_BIDIR(-(a),-(b),-(c),-(d))

  1328. 

  1329.         do{

  1330.             int i;

  1331.             int borderdist=0;

  1332.             end=1;

  1333. 

  1334.             CHECK_BIDIR2(0,0,0,1)

  1335.             CHECK_BIDIR2(0,0,1,0)

  1336.             CHECK_BIDIR2(0,1,0,0)

  1337.             CHECK_BIDIR2(1,0,0,0)

  1338. 

  1339.             for(i=8; i<limit; i++){

  1340.                 int fx= motion_fx+vect[i][0];

  1341.                 int fy= motion_fy+vect[i][1];

  1342.                 int bx= motion_bx+vect[i][2];

  1343.                 int by= motion_by+vect[i][3];

  1344.                 if(borderdist<=0){

  1345.                     int a= (xmax - FFMAX(fx,bx))|(FFMIN(fx,bx) - xmin);

  1346.                     int b= (ymax - FFMAX(fy,by))|(FFMIN(fy,by) - ymin);

  1347.                     if((a|b) < 0)

  1348.                         map[(hashidx+hash[i])&255] = 1;

  1349.                 }

  1350.                 if(!map[(hashidx+hash[i])&255]){

  1351.                     int score;

  1352.                     map[(hashidx+hash[i])&255] = 1;

  1353.                     score= check_bidir_mv(s, fx, fy, bx, by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);

  1354.                     if(score < fbmin){

  1355.                         hashidx += hash[i];

  1356.                         fbmin= score;

  1357.                         motion_fx=fx;

  1358.                         motion_fy=fy;

  1359.                         motion_bx=bx;

  1360.                         motion_by=by;

  1361.                         end=0;

  1362.                         borderdist--;

  1363.                         if(borderdist<=0){

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

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

  1366.                             borderdist= FFMIN(a,b);

  1367.                         }

  1368.                     }

  1369.                 }

  1370.             }

  1371.         }while(!end);

  1372.     }

  1373. 

  1374.     s->b_bidir_forw_mv_table[xy][0]= motion_fx;

  1375.     s->b_bidir_forw_mv_table[xy][1]= motion_fy;

  1376.     s->b_bidir_back_mv_table[xy][0]= motion_bx;

  1377.     s->b_bidir_back_mv_table[xy][1]= motion_by;

  1378. 

  1379.     return fbmin;

  1380. }

  1381. 

  1382. static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)

  1383. {

  1384.     MotionEstContext * const c= &s->me;

  1385.     int P[10][2];

  1386.     const int mot_stride = s->mb_stride;

  1387.     const int mot_xy = mb_y*mot_stride + mb_x;

  1388.     const int shift= 1+s->quarter_sample;

  1389.     int dmin, i;

  1390.     const int time_pp= s->pp_time;

  1391.     const int time_pb= s->pb_time;

  1392.     int mx, my, xmin, xmax, ymin, ymax;

  1393.     int16_t (*mv_table)[2]= s->b_direct_mv_table;

  1394. 

  1395.     c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;

  1396.     ymin= xmin=(-32)>>shift;

  1397.     ymax= xmax=   31>>shift;

  1398. 

  1399.     if (IS_8X8(s->next_picture.mb_type[mot_xy])) {

  1400.         s->mv_type= MV_TYPE_8X8;

  1401.     }else{

  1402.         s->mv_type= MV_TYPE_16X16;

  1403.     }

  1404. 

  1405.     for(i=0; i<4; i++){

  1406.         int index= s->block_index[i];

  1407.         int min, max;

  1408. 

  1409.         c->co_located_mv[i][0] = s->next_picture.motion_val[0][index][0];

  1410.         c->co_located_mv[i][1] = s->next_picture.motion_val[0][index][1];

  1411.         c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));

  1412.         c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));

  1413. //        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);

  1414. //        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);

  1415. 

  1416.         max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;

  1417.         min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;

  1418.         max+= 16*mb_x + 1; // +-1 is for the simpler rounding

  1419.         min+= 16*mb_x - 1;

  1420.         xmax= FFMIN(xmax, s->width - max);

  1421.         xmin= FFMAX(xmin, - 16     - min);

  1422. 

  1423.         max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;

  1424.         min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;

  1425.         max+= 16*mb_y + 1; // +-1 is for the simpler rounding

  1426.         min+= 16*mb_y - 1;

  1427.         ymax= FFMIN(ymax, s->height - max);

  1428.         ymin= FFMAX(ymin, - 16      - min);

  1429. 

  1430.         if(s->mv_type == MV_TYPE_16X16) break;

  1431.     }

  1432. 

  1433.     assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);

  1434. 

  1435.     if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){

  1436.         s->b_direct_mv_table[mot_xy][0]= 0;

  1437.         s->b_direct_mv_table[mot_xy][1]= 0;

  1438. 

  1439.         return 256*256*256*64;

  1440.     }

  1441. 

  1442.     c->xmin= xmin;

  1443.     c->ymin= ymin;

  1444.     c->xmax= xmax;

  1445.     c->ymax= ymax;

  1446.     c->flags     |= FLAG_DIRECT;

  1447.     c->sub_flags |= FLAG_DIRECT;

  1448.     c->pred_x=0;

  1449.     c->pred_y=0;

  1450. 

  1451.     P_LEFT[0]        = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);

  1452.     P_LEFT[1]        = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);

  1453. 

  1454.     /* special case for first line */

  1455.     if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped

  1456.         P_TOP[0]      = av_clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);

  1457.         P_TOP[1]      = av_clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);

  1458.         P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);

  1459.         P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);

  1460. 

  1461.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1462.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1463.     }

  1464. 

  1465.     dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);

  1466.     if(c->sub_flags&FLAG_QPEL)

  1467.         dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1468.     else

  1469.         dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);

  1470. 

  1471.     if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)

  1472.         dmin= get_mb_score(s, mx, my, 0, 0, 0, 16, 1);

  1473. 

  1474.     get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed

  1475. 

  1476.     mv_table[mot_xy][0]= mx;

  1477.     mv_table[mot_xy][1]= my;

  1478.     c->flags     &= ~FLAG_DIRECT;

  1479.     c->sub_flags &= ~FLAG_DIRECT;

  1480. 

  1481.     return dmin;

  1482. }

  1483. 

  1484. void ff_estimate_b_frame_motion(MpegEncContext * s,

  1485.                              int mb_x, int mb_y)

  1486. {

  1487.     MotionEstContext * const c= &s->me;

  1488.     const int penalty_factor= c->mb_penalty_factor;

  1489.     int fmin, bmin, dmin, fbmin, bimin, fimin;

  1490.     int type=0;

  1491.     const int xy = mb_y*s->mb_stride + mb_x;

  1492.     init_ref(c, s->new_picture.f->data, s->last_picture.f->data,

  1493.              s->next_picture.f->data, 16 * mb_x, 16 * mb_y, 2);

  1494. 

  1495.     get_limits(s, 16*mb_x, 16*mb_y);

  1496. 

  1497.     c->skip=0;

  1498. 

  1499.     if (s->codec_id == AV_CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]) {

  1500.         int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0

  1501. 

  1502.         score= ((unsigned)(score*score + 128*256))>>16;

  1503.         c->mc_mb_var_sum_temp += score;

  1504.         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE

  1505.         s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;

  1506. 

  1507.         return;

  1508.     }

  1509. 

  1510.     if (s->codec_id == AV_CODEC_ID_MPEG4)

  1511.         dmin= direct_search(s, mb_x, mb_y);

  1512.     else

  1513.         dmin= INT_MAX;

  1514. //FIXME penalty stuff for non mpeg4

  1515.     c->skip=0;

  1516.     fmin = estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) +

  1517.            3 * penalty_factor;

  1518. 

  1519.     c->skip=0;

  1520.     bmin = estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) +

  1521.            2 * penalty_factor;

  1522.     ff_dlog(s, " %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

  1523. 

  1524.     c->skip=0;

  1525.     fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;

  1526.     ff_dlog(s, "%d %d %d %d\n", dmin, fmin, bmin, fbmin);

  1527. 

  1528.     if (s->avctx->flags & CODEC_FLAG_INTERLACED_ME) {

  1529. //FIXME mb type penalty

  1530.         c->skip=0;

  1531.         c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;

  1532.         fimin= interlaced_search(s, 0,

  1533.                                  s->b_field_mv_table[0], s->b_field_select_table[0],

  1534.                                  s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);

  1535.         c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;

  1536.         bimin= interlaced_search(s, 2,

  1537.                                  s->b_field_mv_table[1], s->b_field_select_table[1],

  1538.                                  s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);

  1539.     }else

  1540.         fimin= bimin= INT_MAX;

  1541. 

  1542.     {

  1543.         int score= fmin;

  1544.         type = CANDIDATE_MB_TYPE_FORWARD;

  1545. 

  1546.         if (dmin <= score){

  1547.             score = dmin;

  1548.             type = CANDIDATE_MB_TYPE_DIRECT;

  1549.         }

  1550.         if(bmin<score){

  1551.             score=bmin;

  1552.             type= CANDIDATE_MB_TYPE_BACKWARD;

  1553.         }

  1554.         if(fbmin<score){

  1555.             score=fbmin;

  1556.             type= CANDIDATE_MB_TYPE_BIDIR;

  1557.         }

  1558.         if(fimin<score){

  1559.             score=fimin;

  1560.             type= CANDIDATE_MB_TYPE_FORWARD_I;

  1561.         }

  1562.         if(bimin<score){

  1563.             score=bimin;

  1564.             type= CANDIDATE_MB_TYPE_BACKWARD_I;

  1565.         }

  1566. 

  1567.         score= ((unsigned)(score*score + 128*256))>>16;

  1568.         c->mc_mb_var_sum_temp += score;

  1569.         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE

  1570.     }

  1571. 

  1572.     if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){

  1573.         type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;

  1574.         if(fimin < INT_MAX)

  1575.             type |= CANDIDATE_MB_TYPE_FORWARD_I;

  1576.         if(bimin < INT_MAX)

  1577.             type |= CANDIDATE_MB_TYPE_BACKWARD_I;

  1578.         if(fimin < INT_MAX && bimin < INT_MAX){

  1579.             type |= CANDIDATE_MB_TYPE_BIDIR_I;

  1580.         }

  1581.          //FIXME something smarter

  1582.         if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB

  1583.         if (s->codec_id == AV_CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT &&

  1584.             s->mpv_flags & FF_MPV_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])

  1585.             type |= CANDIDATE_MB_TYPE_DIRECT0;

  1586.     }

  1587. 

  1588.     s->mb_type[mb_y*s->mb_stride + mb_x]= type;

  1589. }

  1590. 

  1591. /* find best f_code for ME which do unlimited searches */

  1592. int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)

  1593. {

  1594.     if(s->me_method>=ME_EPZS){

  1595.         int score[8];

  1596.         int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);

  1597.         uint8_t * fcode_tab= s->fcode_tab;

  1598.         int best_fcode=-1;

  1599.         int best_score=-10000000;

  1600. 

  1601.         if(s->msmpeg4_version)

  1602.             range= FFMIN(range, 16);

  1603.         else if(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)

  1604.             range= FFMIN(range, 256);

  1605. 

  1606.         for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);

  1607. 

  1608.         for(y=0; y<s->mb_height; y++){

  1609.             int x;

  1610.             int xy= y*s->mb_stride;

  1611.             for(x=0; x<s->mb_width; x++){

  1612.                 if(s->mb_type[xy] & type){

  1613.                     int mx= mv_table[xy][0];

  1614.                     int my= mv_table[xy][1];

  1615.                     int fcode= FFMAX(fcode_tab[mx + MAX_MV],

  1616.                                      fcode_tab[my + MAX_MV]);

  1617.                     int j;

  1618. 

  1619.                         if(mx >= range || mx < -range ||

  1620.                            my >= range || my < -range)

  1621.                             continue;

  1622. 

  1623.                     for(j=0; j<fcode && j<8; j++){

  1624.                         if(s->pict_type==AV_PICTURE_TYPE_B || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])

  1625.                             score[j]-= 170;

  1626.                     }

  1627.                 }

  1628.                 xy++;

  1629.             }

  1630.         }

  1631. 

  1632.         for(i=1; i<8; i++){

  1633.             if(score[i] > best_score){

  1634.                 best_score= score[i];

  1635.                 best_fcode= i;

  1636.             }

  1637.         }

  1638. 

  1639.         return best_fcode;

  1640.     }else{

  1641.         return 1;

  1642.     }

  1643. }

  1644. 

  1645. void ff_fix_long_p_mvs(MpegEncContext * s)

  1646. {

  1647.     MotionEstContext * const c= &s->me;

  1648.     const int f_code= s->f_code;

  1649.     int y, range;

  1650.     assert(s->pict_type==AV_PICTURE_TYPE_P);

  1651. 

  1652.     range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);

  1653. 

  1654.     assert(range <= 16 || !s->msmpeg4_version);

  1655.     assert(range <=256 || !(s->codec_id == AV_CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));

  1656. 

  1657.     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;

  1658. 

  1659.     if (s->avctx->flags & CODEC_FLAG_4MV) {

  1660.         const int wrap= s->b8_stride;

  1661. 

  1662.         /* clip / convert to intra 8x8 type MVs */

  1663.         for(y=0; y<s->mb_height; y++){

  1664.             int xy= y*2*wrap;

  1665.             int i= y*s->mb_stride;

  1666.             int x;

  1667. 

  1668.             for(x=0; x<s->mb_width; x++){

  1669.                 if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){

  1670.                     int block;

  1671.                     for(block=0; block<4; block++){

  1672.                         int off= (block& 1) + (block>>1)*wrap;

  1673.                         int mx = s->current_picture.motion_val[0][ xy + off ][0];

  1674.                         int my = s->current_picture.motion_val[0][ xy + off ][1];

  1675. 

  1676.                         if(   mx >=range || mx <-range

  1677.                            || my >=range || my <-range){

  1678.                             s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;

  1679.                             s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;

  1680.                             s->current_picture.mb_type[i] = CANDIDATE_MB_TYPE_INTRA;

  1681.                         }

  1682.                     }

  1683.                 }

  1684.                 xy+=2;

  1685.                 i++;

  1686.             }

  1687.         }

  1688.     }

  1689. }

  1690. 

  1691. /**

  1692.  *

  1693.  * @param truncate 1 for truncation, 0 for using intra

  1694.  */

  1695. void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,

  1696.                      int16_t (*mv_table)[2], int f_code, int type, int truncate)

  1697. {

  1698.     MotionEstContext * const c= &s->me;

  1699.     int y, h_range, v_range;

  1700. 

  1701.     // RAL: 8 in MPEG-1, 16 in MPEG-4

  1702.     int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);

  1703. 

  1704.     if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;

  1705. 

  1706.     h_range= range;

  1707.     v_range= field_select_table ? range>>1 : range;

  1708. 

  1709.     /* clip / convert to intra 16x16 type MVs */

  1710.     for(y=0; y<s->mb_height; y++){

  1711.         int x;

  1712.         int xy= y*s->mb_stride;

  1713.         for(x=0; x<s->mb_width; x++){

  1714.             if (s->mb_type[xy] & type){    // RAL: "type" test added...

  1715.                 if (!field_select_table || field_select_table[xy] == field_select) {

  1716.                     if(   mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range

  1717.                        || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){

  1718. 

  1719.                         if(truncate){

  1720.                             if     (mv_table[xy][0] > h_range-1) mv_table[xy][0]=  h_range-1;

  1721.                             else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;

  1722.                             if     (mv_table[xy][1] > v_range-1) mv_table[xy][1]=  v_range-1;

  1723.                             else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;

  1724.                         }else{

  1725.                             s->mb_type[xy] &= ~type;

  1726.                             s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;

  1727.                             mv_table[xy][0]=

  1728.                             mv_table[xy][1]= 0;

  1729.                         }

  1730.                     }

  1731.                 }

  1732.             }

  1733.             xy++;

  1734.         }

  1735.     }

  1736. }