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.
1756 Commits
32 Branches
353MB
Tree: d90cc6af20
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'd90cc6af20'
${ noResults }
FFmpeg/libavcodec/motion_est.c

1679 lines
56KB
Raw Blame History 

  1. /*

  2.  * Motion estimation 

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

  4.  * Copyright (c) 2002-2003 Michael Niedermayer

  5.  * 

  6.  *

  7.  * This library is free software; you can redistribute it and/or

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

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

  10.  * version 2 of the License, or (at your option) any later version.

  11.  *

  12.  * This library is distributed in the hope that it will be useful,

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

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

  15.  * Lesser General Public License for more details.

  16.  *

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

  18.  * License along with this library; if not, write to the Free Software

  19.  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA

  20.  *

  21.  * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>

  22.  */

  23. #include <stdlib.h>

  24. #include <stdio.h>

  25. #include "avcodec.h"

  26. #include "dsputil.h"

  27. #include "mpegvideo.h"

  28. 

  29. //#undef NDEBUG

  30. //#include <assert.h>

  31. 

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

  33. 

  34. #define P_LEFT P[1]

  35. #define P_TOP P[2]

  36. #define P_TOPRIGHT P[3]

  37. #define P_MEDIAN P[4]

  38. #define P_MV1 P[9]

  39. 

  40. static inline int sad_hpel_motion_search(MpegEncContext * s,

  41. 				  int *mx_ptr, int *my_ptr, int dmin,

  42. 				  int xmin, int ymin, int xmax, int ymax,

  43.                                   int pred_x, int pred_y, Picture *picture,

  44.                                   int n, int size, uint16_t * const mv_penalty);

  45. 

  46. static inline int update_map_generation(MpegEncContext * s)

  47. {

  48.     s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2);

  49.     if(s->me.map_generation==0){

  50.         s->me.map_generation= 1<<(ME_MAP_MV_BITS*2);

  51.         memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE);

  52.     }

  53.     return s->me.map_generation;

  54. }

  55. 

  56. /* shape adaptive search stuff */

  57. typedef struct Minima{

  58.     int height;

  59.     int x, y;

  60.     int checked;

  61. }Minima;

  62. 

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

  64.     Minima *da = (Minima *) a;

  65.     Minima *db = (Minima *) b;

  66.     

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

  68. }

  69.                                   

  70. /* SIMPLE */

  71. #define RENAME(a) simple_ ## a

  72. 

  73. #define CMP(d, x, y, size)\

  74. d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);

  75. 

  76. #define CMP_HPEL(d, dx, dy, x, y, size)\

  77. {\

  78.     const int dxy= (dx) + 2*(dy);\

  79.     hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\

  80.     d = cmp_sub(s, s->me.scratchpad, src_y, stride);\

  81. }

  82. 

  83. #define CMP_QPEL(d, dx, dy, x, y, size)\

  84. {\

  85.     const int dxy= (dx) + 4*(dy);\

  86.     qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\

  87.     d = cmp_sub(s, s->me.scratchpad, src_y, stride);\

  88. }

  89. 

  90. #include "motion_est_template.c"

  91. #undef RENAME

  92. #undef CMP

  93. #undef CMP_HPEL

  94. #undef CMP_QPEL

  95. #undef INIT

  96. 

  97. /* SIMPLE CHROMA */

  98. #define RENAME(a) simple_chroma_ ## a

  99. 

  100. #define CMP(d, x, y, size)\

  101. d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);\

  102. if(chroma_cmp){\

  103.     int dxy= ((x)&1) + 2*((y)&1);\

  104.     int c= ((x)>>1) + ((y)>>1)*uvstride;\

  105. \

  106.     chroma_hpel_put[0][dxy](s->me.scratchpad, ref_u + c, uvstride, 8);\

  107.     d += chroma_cmp(s, s->me.scratchpad, src_u, uvstride);\

  108.     chroma_hpel_put[0][dxy](s->me.scratchpad, ref_v + c, uvstride, 8);\

  109.     d += chroma_cmp(s, s->me.scratchpad, src_v, uvstride);\

  110. }

  111. 

  112. #define CMP_HPEL(d, dx, dy, x, y, size)\

  113. {\

  114.     const int dxy= (dx) + 2*(dy);\

  115.     hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\

  116.     d = cmp_sub(s, s->me.scratchpad, src_y, stride);\

  117.     if(chroma_cmp_sub){\

  118.         int cxy= (dxy) | ((x)&1) | (2*((y)&1));\

  119.         int c= ((x)>>1) + ((y)>>1)*uvstride;\

  120.         chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\

  121.         d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\

  122.         chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\

  123.         d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\

  124.     }\

  125. }

  126. 

  127. #define CMP_QPEL(d, dx, dy, x, y, size)\

  128. {\

  129.     const int dxy= (dx) + 4*(dy);\

  130.     qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\

  131.     d = cmp_sub(s, s->me.scratchpad, src_y, stride);\

  132.     if(chroma_cmp_sub){\

  133.         int cxy, c;\

  134.         int cx= (4*(x) + (dx))/2;\

  135.         int cy= (4*(y) + (dy))/2;\

  136.         cx= (cx>>1)|(cx&1);\

  137.         cy= (cy>>1)|(cy&1);\

  138.         cxy= (cx&1) + 2*(cy&1);\

  139.         c= ((cx)>>1) + ((cy)>>1)*uvstride;\

  140.         chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\

  141.         d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\

  142.         chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\

  143.         d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\

  144.     }\

  145. }

  146. 

  147. #include "motion_est_template.c"

  148. #undef RENAME

  149. #undef CMP

  150. #undef CMP_HPEL

  151. #undef CMP_QPEL

  152. #undef INIT

  153. 

  154. /* SIMPLE DIRECT HPEL */

  155. #define RENAME(a) simple_direct_hpel_ ## a

  156. //FIXME precalc divisions stuff

  157. 

  158. #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\

  159. if((x) >= xmin && 2*(x) + (dx) <= 2*xmax && (y) >= ymin && 2*(y) + (dy) <= 2*ymax){\

  160.     const int hx= 2*(x) + (dx);\

  161.     const int hy= 2*(y) + (dy);\

  162.     if(s->mv_type==MV_TYPE_8X8){\

  163.         int i;\

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

  165.             int fx = s->me.direct_basis_mv[i][0] + hx;\

  166.             int fy = s->me.direct_basis_mv[i][1] + hy;\

  167.             int bx = hx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\

  168.             int by = hy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\

  169.             int fxy= (fx&1) + 2*(fy&1);\

  170.             int bxy= (bx&1) + 2*(by&1);\

  171. \

  172.             uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\

  173.             hpel_put[1][fxy](dst, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 8);\

  174.             hpel_avg[1][bxy](dst, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 8);\

  175.         }\

  176.     }else{\

  177.         int fx = s->me.direct_basis_mv[0][0] + hx;\

  178.         int fy = s->me.direct_basis_mv[0][1] + hy;\

  179.         int bx = hx ? fx - s->me.co_located_mv[0][0] : (s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp);\

  180.         int by = hy ? fy - s->me.co_located_mv[0][1] : (s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp);\

  181.         int fxy= (fx&1) + 2*(fy&1);\

  182.         int bxy= (bx&1) + 2*(by&1);\

  183.         \

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

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

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

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

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

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

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

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

  192. \

  193.         hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\

  194.         hpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 16);\

  195.     }\

  196.     d = cmp_func(s, s->me.scratchpad, src_y, stride);\

  197. }else\

  198.     d= 256*256*256*32;

  199. 

  200. 

  201. #define CMP_HPEL(d, dx, dy, x, y, size)\

  202.     CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)

  203. 

  204. #define CMP(d, x, y, size)\

  205.     CMP_DIRECT(d, 0, 0, x, y, size, cmp)

  206.     

  207. #include "motion_est_template.c"

  208. #undef RENAME

  209. #undef CMP

  210. #undef CMP_HPEL

  211. #undef CMP_QPEL

  212. #undef INIT

  213. #undef CMP_DIRECT

  214. 

  215. /* SIMPLE DIRECT QPEL */

  216. #define RENAME(a) simple_direct_qpel_ ## a

  217. 

  218. #define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\

  219. if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\

  220.     const int qx= 4*(x) + (dx);\

  221.     const int qy= 4*(y) + (dy);\

  222.     if(s->mv_type==MV_TYPE_8X8){\

  223.         int i;\

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

  225.             int fx = s->me.direct_basis_mv[i][0] + qx;\

  226.             int fy = s->me.direct_basis_mv[i][1] + qy;\

  227.             int bx = qx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\

  228.             int by = qy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\

  229.             int fxy= (fx&3) + 4*(fy&3);\

  230.             int bxy= (bx&3) + 4*(by&3);\

  231. \

  232.             uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\

  233.             qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\

  234.             qpel_avg[1][bxy](dst, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\

  235.         }\

  236.     }else{\

  237.         int fx = s->me.direct_basis_mv[0][0] + qx;\

  238.         int fy = s->me.direct_basis_mv[0][1] + qy;\

  239.         int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\

  240.         int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\

  241.         int fxy= (fx&3) + 4*(fy&3);\

  242.         int bxy= (bx&3) + 4*(by&3);\

  243. \

  244.         qpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\

  245.         qpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\

  246.     }\

  247.     d = cmp_func(s, s->me.scratchpad, src_y, stride);\

  248. }else\

  249.     d= 256*256*256*32;

  250. 

  251. 

  252. #define CMP_QPEL(d, dx, dy, x, y, size)\

  253.     CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)

  254. 

  255. #define CMP(d, x, y, size)\

  256.     CMP_DIRECT(d, 0, 0, x, y, size, cmp)

  257. 

  258. #include "motion_est_template.c"

  259. #undef RENAME

  260. #undef CMP

  261. #undef CMP_HPEL

  262. #undef CMP_QPEL

  263. #undef INIT

  264. #undef CMP__DIRECT

  265. 

  266. 

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

  268.     return 0;

  269. }

  270. 

  271. static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){

  272.     DSPContext* c= &s->dsp;

  273.     int i;

  274.     

  275.     memset(cmp, 0, sizeof(void*)*11);

  276. 

  277.     switch(type&0xFF){

  278.     case FF_CMP_SAD:

  279.         cmp[0]= c->sad[0];

  280.         cmp[1]= c->sad[1];

  281.         break;

  282.     case FF_CMP_SATD:

  283.         cmp[0]= c->hadamard8_diff[0];

  284.         cmp[1]= c->hadamard8_diff[1];

  285.         break;

  286.     case FF_CMP_SSE:

  287.         cmp[0]= c->sse[0];

  288.         cmp[1]= c->sse[1];

  289.         break;

  290.     case FF_CMP_DCT:

  291.         cmp[0]= c->dct_sad[0];

  292.         cmp[1]= c->dct_sad[1];

  293.         break;

  294.     case FF_CMP_PSNR:

  295.         cmp[0]= c->quant_psnr[0];

  296.         cmp[1]= c->quant_psnr[1];

  297.         break;

  298.     case FF_CMP_BIT:

  299.         cmp[0]= c->bit[0];

  300.         cmp[1]= c->bit[1];

  301.         break;

  302.     case FF_CMP_RD:

  303.         cmp[0]= c->rd[0];

  304.         cmp[1]= c->rd[1];

  305.         break;

  306.     case FF_CMP_ZERO:

  307.         for(i=0; i<7; i++){

  308.             cmp[i]= zero_cmp;

  309.         }

  310.         break;

  311.     default:

  312.         fprintf(stderr,"internal error in cmp function selection\n");

  313.     }

  314. }

  315. 

  316. static inline int get_penalty_factor(MpegEncContext *s, int type){

  317.     switch(type&0xFF){

  318.     default:

  319.     case FF_CMP_SAD:

  320.         return s->qscale*2;

  321.     case FF_CMP_DCT:

  322.         return s->qscale*3;

  323.     case FF_CMP_SATD:

  324.         return s->qscale*6;

  325.     case FF_CMP_SSE:

  326.         return s->qscale*s->qscale*2;

  327.     case FF_CMP_BIT:

  328.         return 1;

  329.     case FF_CMP_RD:

  330.     case FF_CMP_PSNR:

  331.         return (s->qscale*s->qscale*185 + 64)>>7;

  332.     }

  333. }

  334. 

  335. void ff_init_me(MpegEncContext *s){

  336.     set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);

  337.     set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp);

  338.     set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);

  339.     set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp);

  340. 

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

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

  343.             s->me.sub_motion_search= simple_chroma_qpel_motion_search;

  344.         else

  345.             s->me.sub_motion_search= simple_qpel_motion_search;

  346.     }else{

  347.         if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)

  348.             s->me.sub_motion_search= simple_chroma_hpel_motion_search;

  349.         else if(   s->avctx->me_sub_cmp == FF_CMP_SAD 

  350.                 && s->avctx->    me_cmp == FF_CMP_SAD 

  351.                 && s->avctx->    mb_cmp == FF_CMP_SAD)

  352.             s->me.sub_motion_search= sad_hpel_motion_search;

  353.         else

  354.             s->me.sub_motion_search= simple_hpel_motion_search;

  355.     }

  356. 

  357.     if(s->avctx->me_cmp&FF_CMP_CHROMA){

  358.         s->me.motion_search[0]= simple_chroma_epzs_motion_search;

  359.         s->me.motion_search[1]= simple_chroma_epzs_motion_search4;

  360.     }else{

  361.         s->me.motion_search[0]= simple_epzs_motion_search;

  362.         s->me.motion_search[1]= simple_epzs_motion_search4;

  363.     }

  364.     

  365.     if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){

  366.         s->me.pre_motion_search= simple_chroma_epzs_motion_search;

  367.     }else{

  368.         s->me.pre_motion_search= simple_epzs_motion_search;

  369.     }

  370.     

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

  372.         if(s->avctx->mb_cmp&FF_CMP_CHROMA)

  373.             s->me.get_mb_score= simple_chroma_qpel_get_mb_score;

  374.         else

  375.             s->me.get_mb_score= simple_qpel_get_mb_score;

  376.     }else{

  377.         if(s->avctx->mb_cmp&FF_CMP_CHROMA)

  378.             s->me.get_mb_score= simple_chroma_hpel_get_mb_score;

  379.         else

  380.             s->me.get_mb_score= simple_hpel_get_mb_score;

  381.     }

  382. }

  383.       

  384. static int pix_dev(UINT8 * pix, int line_size, int mean)

  385. {

  386.     int s, i, j;

  387. 

  388.     s = 0;

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

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

  391. 	    s += ABS(pix[0]-mean);

  392. 	    s += ABS(pix[1]-mean);

  393. 	    s += ABS(pix[2]-mean);

  394. 	    s += ABS(pix[3]-mean);

  395. 	    s += ABS(pix[4]-mean);

  396. 	    s += ABS(pix[5]-mean);

  397. 	    s += ABS(pix[6]-mean);

  398. 	    s += ABS(pix[7]-mean);

  399. 	    pix += 8;

  400. 	}

  401. 	pix += line_size - 16;

  402.     }

  403.     return s;

  404. }

  405. 

  406. static inline void no_motion_search(MpegEncContext * s,

  407. 				    int *mx_ptr, int *my_ptr)

  408. {

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

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

  411. }

  412. 

  413. static int full_motion_search(MpegEncContext * s,

  414.                               int *mx_ptr, int *my_ptr, int range,

  415.                               int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)

  416. {

  417.     int x1, y1, x2, y2, xx, yy, x, y;

  418.     int mx, my, dmin, d;

  419.     UINT8 *pix;

  420. 

  421.     xx = 16 * s->mb_x;

  422.     yy = 16 * s->mb_y;

  423.     x1 = xx - range + 1;	/* we loose one pixel to avoid boundary pb with half pixel pred */

  424.     if (x1 < xmin)

  425. 	x1 = xmin;

  426.     x2 = xx + range - 1;

  427.     if (x2 > xmax)

  428. 	x2 = xmax;

  429.     y1 = yy - range + 1;

  430.     if (y1 < ymin)

  431. 	y1 = ymin;

  432.     y2 = yy + range - 1;

  433.     if (y2 > ymax)

  434. 	y2 = ymax;

  435.     pix = s->new_picture.data[0] + (yy * s->linesize) + xx;

  436.     dmin = 0x7fffffff;

  437.     mx = 0;

  438.     my = 0;

  439.     for (y = y1; y <= y2; y++) {

  440. 	for (x = x1; x <= x2; x++) {

  441. 	    d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,

  442. 			     s->linesize);

  443. 	    if (d < dmin ||

  444. 		(d == dmin &&

  445. 		 (abs(x - xx) + abs(y - yy)) <

  446. 		 (abs(mx - xx) + abs(my - yy)))) {

  447. 		dmin = d;

  448. 		mx = x;

  449. 		my = y;

  450. 	    }

  451. 	}

  452.     }

  453. 

  454.     *mx_ptr = mx;

  455.     *my_ptr = my;

  456. 

  457. #if 0

  458.     if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||

  459. 	*my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {

  460. 	fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);

  461.     }

  462. #endif

  463.     return dmin;

  464. }

  465. 

  466. 

  467. static int log_motion_search(MpegEncContext * s,

  468.                              int *mx_ptr, int *my_ptr, int range,

  469.                              int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)

  470. {

  471.     int x1, y1, x2, y2, xx, yy, x, y;

  472.     int mx, my, dmin, d;

  473.     UINT8 *pix;

  474. 

  475.     xx = s->mb_x << 4;

  476.     yy = s->mb_y << 4;

  477. 

  478.     /* Left limit */

  479.     x1 = xx - range;

  480.     if (x1 < xmin)

  481. 	x1 = xmin;

  482. 

  483.     /* Right limit */

  484.     x2 = xx + range;

  485.     if (x2 > xmax)

  486. 	x2 = xmax;

  487. 

  488.     /* Upper limit */

  489.     y1 = yy - range;

  490.     if (y1 < ymin)

  491. 	y1 = ymin;

  492. 

  493.     /* Lower limit */

  494.     y2 = yy + range;

  495.     if (y2 > ymax)

  496. 	y2 = ymax;

  497. 

  498.     pix = s->new_picture.data[0] + (yy * s->linesize) + xx;

  499.     dmin = 0x7fffffff;

  500.     mx = 0;

  501.     my = 0;

  502. 

  503.     do {

  504. 	for (y = y1; y <= y2; y += range) {

  505. 	    for (x = x1; x <= x2; x += range) {

  506. 		d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);

  507. 		if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {

  508. 		    dmin = d;

  509. 		    mx = x;

  510. 		    my = y;

  511. 		}

  512. 	    }

  513. 	}

  514. 

  515. 	range = range >> 1;

  516. 

  517. 	x1 = mx - range;

  518. 	if (x1 < xmin)

  519. 	    x1 = xmin;

  520. 

  521. 	x2 = mx + range;

  522. 	if (x2 > xmax)

  523. 	    x2 = xmax;

  524. 

  525. 	y1 = my - range;

  526. 	if (y1 < ymin)

  527. 	    y1 = ymin;

  528. 

  529. 	y2 = my + range;

  530. 	if (y2 > ymax)

  531. 	    y2 = ymax;

  532. 

  533.     } while (range >= 1);

  534. 

  535. #ifdef DEBUG

  536.     fprintf(stderr, "log       - MX: %d\tMY: %d\n", mx, my);

  537. #endif

  538.     *mx_ptr = mx;

  539.     *my_ptr = my;

  540.     return dmin;

  541. }

  542. 

  543. static int phods_motion_search(MpegEncContext * s,

  544.                                int *mx_ptr, int *my_ptr, int range,

  545.                                int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)

  546. {

  547.     int x1, y1, x2, y2, xx, yy, x, y, lastx, d;

  548.     int mx, my, dminx, dminy;

  549.     UINT8 *pix;

  550. 

  551.     xx = s->mb_x << 4;

  552.     yy = s->mb_y << 4;

  553. 

  554.     /* Left limit */

  555.     x1 = xx - range;

  556.     if (x1 < xmin)

  557. 	x1 = xmin;

  558. 

  559.     /* Right limit */

  560.     x2 = xx + range;

  561.     if (x2 > xmax)

  562. 	x2 = xmax;

  563. 

  564.     /* Upper limit */

  565.     y1 = yy - range;

  566.     if (y1 < ymin)

  567. 	y1 = ymin;

  568. 

  569.     /* Lower limit */

  570.     y2 = yy + range;

  571.     if (y2 > ymax)

  572. 	y2 = ymax;

  573. 

  574.     pix = s->new_picture.data[0] + (yy * s->linesize) + xx;

  575.     mx = 0;

  576.     my = 0;

  577. 

  578.     x = xx;

  579.     y = yy;

  580.     do {

  581.         dminx = 0x7fffffff;

  582.         dminy = 0x7fffffff;

  583. 

  584. 	lastx = x;

  585. 	for (x = x1; x <= x2; x += range) {

  586. 	    d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);

  587. 	    if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {

  588. 		dminx = d;

  589. 		mx = x;

  590. 	    }

  591. 	}

  592. 

  593. 	x = lastx;

  594. 	for (y = y1; y <= y2; y += range) {

  595. 	    d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);

  596. 	    if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {

  597. 		dminy = d;

  598. 		my = y;

  599. 	    }

  600. 	}

  601. 

  602. 	range = range >> 1;

  603. 

  604. 	x = mx;

  605. 	y = my;

  606. 	x1 = mx - range;

  607. 	if (x1 < xmin)

  608. 	    x1 = xmin;

  609. 

  610. 	x2 = mx + range;

  611. 	if (x2 > xmax)

  612. 	    x2 = xmax;

  613. 

  614. 	y1 = my - range;

  615. 	if (y1 < ymin)

  616. 	    y1 = ymin;

  617. 

  618. 	y2 = my + range;

  619. 	if (y2 > ymax)

  620. 	    y2 = ymax;

  621. 

  622.     } while (range >= 1);

  623. 

  624. #ifdef DEBUG

  625.     fprintf(stderr, "phods     - MX: %d\tMY: %d\n", mx, my);

  626. #endif

  627. 

  628.     /* half pixel search */

  629.     *mx_ptr = mx;

  630.     *my_ptr = my;

  631.     return dminy;

  632. }

  633. 

  634. 

  635. #define Z_THRESHOLD 256

  636. 

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

  638. {\

  639.     d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\

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

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

  642. }

  643. 

  644. static inline int sad_hpel_motion_search(MpegEncContext * s,

  645. 				  int *mx_ptr, int *my_ptr, int dmin,

  646. 				  int xmin, int ymin, int xmax, int ymax,

  647.                                   int pred_x, int pred_y, Picture *picture,

  648.                                   int n, int size, uint16_t * const mv_penalty)

  649. {

  650.     uint8_t *ref_picture= picture->data[0];

  651.     uint32_t *score_map= s->me.score_map;

  652.     const int penalty_factor= s->me.sub_penalty_factor;

  653.     int mx, my, xx, yy, dminh;

  654.     UINT8 *pix, *ptr;

  655.     op_pixels_abs_func pix_abs_x2;

  656.     op_pixels_abs_func pix_abs_y2;

  657.     op_pixels_abs_func pix_abs_xy2;

  658.     

  659.     if(size==0){

  660.         pix_abs_x2 = s->dsp.pix_abs16x16_x2;

  661.         pix_abs_y2 = s->dsp.pix_abs16x16_y2;

  662.         pix_abs_xy2= s->dsp.pix_abs16x16_xy2;

  663.     }else{

  664.         pix_abs_x2 = s->dsp.pix_abs8x8_x2;

  665.         pix_abs_y2 = s->dsp.pix_abs8x8_y2;

  666.         pix_abs_xy2= s->dsp.pix_abs8x8_xy2;

  667.     }

  668. 

  669.     if(s->me.skip){

  670. //    printf("S");

  671.         *mx_ptr = 0;

  672.         *my_ptr = 0;

  673.         return dmin;

  674.     }

  675. //    printf("N");

  676.         

  677.     xx = 16 * s->mb_x + 8*(n&1);

  678.     yy = 16 * s->mb_y + 8*(n>>1);

  679.     pix =  s->new_picture.data[0] + (yy * s->linesize) + xx;

  680. 

  681.     mx = *mx_ptr;

  682.     my = *my_ptr;

  683.     ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);

  684.     

  685.     dminh = dmin;

  686. 

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

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

  689.         int dx=0, dy=0;

  690.         int d, pen_x, pen_y; 

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

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

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

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

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

  696.         mx<<=1;

  697.         my<<=1;

  698. 

  699.         

  700.         pen_x= pred_x + mx;

  701.         pen_y= pred_y + my;

  702. 

  703.         ptr-= s->linesize;

  704.         if(t<=b){

  705.             CHECK_SAD_HALF_MV(y2 , 0, -1)

  706.             if(l<=r){

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

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

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

  710.                     ptr+= s->linesize;

  711.                 }else{

  712.                     ptr+= s->linesize;

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

  714.                 }

  715.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

  716.             }else{

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

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

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

  720.                     ptr+= s->linesize;

  721.                 }else{

  722.                     ptr+= s->linesize;

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

  724.                 }

  725.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

  726.             }

  727.         }else{

  728.             if(l<=r){

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

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

  731.                     ptr+= s->linesize;

  732.                 }else{

  733.                     ptr+= s->linesize;

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

  735.                 }

  736.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

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

  738.             }else{

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

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

  741.                     ptr+= s->linesize;

  742.                 }else{

  743.                     ptr+= s->linesize;

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

  745.                 }

  746.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

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

  748.             }

  749.             CHECK_SAD_HALF_MV(y2 ,  0, +1)

  750.         }

  751.         mx+=dx;

  752.         my+=dy;

  753. 

  754.     }else{

  755.         mx<<=1;

  756.         my<<=1;

  757.     }

  758. 

  759.     *mx_ptr = mx;

  760.     *my_ptr = my;

  761.     return dminh;

  762. }

  763. 

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

  765. {

  766.     const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2);

  767.     

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

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

  770. 

  771.     /* has allready been set to the 4 MV if 4MV is done */

  772.     if(mv4){

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

  774. 

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

  776.         s->motion_val[mot_xy  ][1]= my;

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

  778.         s->motion_val[mot_xy+1][1]= my;

  779. 

  780.         mot_xy += s->block_wrap[0];

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

  782.         s->motion_val[mot_xy  ][1]= my;

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

  784.         s->motion_val[mot_xy+1][1]= my;

  785.     }

  786. }

  787. 

  788. static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code)

  789. {

  790.     *range = 8 * (1 << (f_code - 1));

  791.     /* XXX: temporary kludge to avoid overflow for msmpeg4 */

  792.     if (s->out_format == FMT_H263 && !s->h263_msmpeg4)

  793. 	*range *= 2;

  794. 

  795.     if (s->unrestricted_mv) {

  796.         *xmin = -16;

  797.         *ymin = -16;

  798.         if (s->h263_plus)

  799.             *range *= 2;

  800.         if(s->avctx->codec->id!=CODEC_ID_MPEG4){

  801.             *xmax = s->mb_width*16;

  802.             *ymax = s->mb_height*16;

  803.         }else {

  804.             *xmax = s->width;

  805.             *ymax = s->height;

  806.         }

  807.     } else {

  808.         *xmin = 0;

  809.         *ymin = 0;

  810.         *xmax = s->mb_width*16 - 16;

  811.         *ymax = s->mb_height*16 - 16;

  812.     }

  813. }

  814. 

  815. static inline int h263_mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)

  816. {

  817.     int block;

  818.     int P[10][2];

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

  820.     uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;

  821. 

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

  823.         int mx4, my4;

  824.         int pred_x4, pred_y4;

  825.         int dmin4;

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

  827.         const int mot_stride = s->block_wrap[0];

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

  829. //        const int block_x= (block&1);

  830. //        const int block_y= (block>>1);

  831. #if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way

  832.         const int rel_xmin4= xmin;

  833.         const int rel_xmax4= xmax;

  834.         const int rel_ymin4= ymin;

  835.         const int rel_ymax4= ymax;

  836. #else

  837.         const int rel_xmin4= xmin - block_x*8;

  838.         const int rel_xmax4= xmax - block_x*8 + 8;

  839.         const int rel_ymin4= ymin - block_y*8;

  840.         const int rel_ymax4= ymax - block_y*8 + 8;

  841. #endif

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

  843.         P_LEFT[1] = s->motion_val[mot_xy - 1][1];

  844. 

  845.         if(P_LEFT[0]       > (rel_xmax4<<shift)) P_LEFT[0]       = (rel_xmax4<<shift);

  846. 

  847.         /* special case for first line */

  848.         if (s->mb_y == 0 && block<2) {

  849.             pred_x4= P_LEFT[0];

  850.             pred_y4= P_LEFT[1];

  851.         } else {

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

  853.             P_TOP[1]      = s->motion_val[mot_xy - mot_stride             ][1];

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

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

  856.             if(P_TOP[1]      > (rel_ymax4<<shift)) P_TOP[1]     = (rel_ymax4<<shift);

  857.             if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);

  858.             if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);

  859.             if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);

  860.     

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

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

  863. 

  864. //            if(s->out_format == FMT_H263){

  865.                 pred_x4 = P_MEDIAN[0];

  866.                 pred_y4 = P_MEDIAN[1];

  867. #if 0

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

  869.                 pred_x4= P_LEFT[0];

  870.                 pred_y4= P_LEFT[1];

  871.             }

  872. #endif

  873.         }

  874.         P_MV1[0]= mx;

  875.         P_MV1[1]= my;

  876. 

  877.         dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 

  878.                                        &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);

  879. 

  880.         dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 

  881. 					  pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);

  882.         

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

  884.             int dxy;

  885.             const int offset= ((block&1) + (block>>1)*s->linesize)*8;

  886.             uint8_t *dest_y = s->me.scratchpad + offset;

  887. 

  888.             if(s->quarter_sample){

  889.                 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>2)) + (s->mb_y*16 + (my4>>2))*s->linesize + offset;

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

  891. 

  892.                 if(s->no_rounding)

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

  894.                 else

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

  896.             }else{

  897.                 uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>1)) + (s->mb_y*16 + (my4>>1))*s->linesize + offset;

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

  899. 

  900.                 if(s->no_rounding)

  901.                     s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , s->linesize, 8);

  902.                 else

  903.                     s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , s->linesize, 8);

  904.             }

  905.             dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;

  906.         }else

  907.             dmin_sum+= dmin4;

  908. 

  909.         if(s->quarter_sample){

  910.             mx4_sum+= mx4/2;

  911.             my4_sum+= my4/2;

  912.         }else{

  913.             mx4_sum+= mx4;

  914.             my4_sum+= my4;

  915.         }

  916.             

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

  918.         s->motion_val[ s->block_index[block] ][1]= my4;

  919.     }

  920.     

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

  922.         dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*s->linesize, s->me.scratchpad, s->linesize);

  923.     }

  924.     

  925.     if(s->avctx->mb_cmp&FF_CMP_CHROMA){

  926.         int dxy;

  927.         int mx, my;

  928.         int offset;

  929. 

  930.         mx= ff_h263_round_chroma(mx4_sum);

  931.         my= ff_h263_round_chroma(my4_sum);

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

  933.         

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

  935.        

  936.         if(s->no_rounding){

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

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

  939.         }else{

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

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

  942.         }

  943. 

  944.         dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, s->me.scratchpad  , s->uvlinesize);

  945.         dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, s->me.scratchpad+8, s->uvlinesize);

  946.     }

  947. 

  948.     switch(s->avctx->mb_cmp&0xFF){

  949.     /*case FF_CMP_SSE:

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

  951.     case FF_CMP_RD:

  952.         return dmin_sum;

  953.     default:

  954.         return dmin_sum+ 11*s->me.mb_penalty_factor;

  955.     }

  956. }

  957. 

  958. void ff_estimate_p_frame_motion(MpegEncContext * s,

  959.                                 int mb_x, int mb_y)

  960. {

  961.     UINT8 *pix, *ppix;

  962.     int sum, varc, vard, mx, my, range, dmin, xx, yy;

  963.     int xmin, ymin, xmax, ymax;

  964.     int rel_xmin, rel_ymin, rel_xmax, rel_ymax;

  965.     int pred_x=0, pred_y=0;

  966.     int P[10][2];

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

  968.     int mb_type=0;

  969.     uint8_t *ref_picture= s->last_picture.data[0];

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

  971.     uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;

  972.     

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

  974. 

  975.     s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);

  976.     s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);

  977.     s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);

  978. 

  979.     get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);

  980.     rel_xmin= xmin - mb_x*16;

  981.     rel_xmax= xmax - mb_x*16;

  982.     rel_ymin= ymin - mb_y*16;

  983.     rel_ymax= ymax - mb_y*16;

  984.     s->me.skip=0;

  985. 

  986.     switch(s->me_method) {

  987.     case ME_ZERO:

  988.     default:

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

  990.         mx-= mb_x*16;

  991.         my-= mb_y*16;

  992.         dmin = 0;

  993.         break;

  994.     case ME_FULL:

  995. 	dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);

  996.         mx-= mb_x*16;

  997.         my-= mb_y*16;

  998.         break;

  999.     case ME_LOG:

  1000. 	dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1001.         mx-= mb_x*16;

  1002.         my-= mb_y*16;

  1003.         break;

  1004.     case ME_PHODS:

  1005. 	dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1006.         mx-= mb_x*16;

  1007.         my-= mb_y*16;

  1008.         break;

  1009.     case ME_X1:

  1010.     case ME_EPZS:

  1011.        {

  1012.             const int mot_stride = s->block_wrap[0];

  1013.             const int mot_xy = s->block_index[0];

  1014. 

  1015.             P_LEFT[0]       = s->motion_val[mot_xy - 1][0];

  1016.             P_LEFT[1]       = s->motion_val[mot_xy - 1][1];

  1017. 

  1018.             if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);

  1019. 

  1020.             if(mb_y) {

  1021.                 P_TOP[0]      = s->motion_val[mot_xy - mot_stride    ][0];

  1022.                 P_TOP[1]      = s->motion_val[mot_xy - mot_stride    ][1];

  1023.                 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];

  1024.                 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];

  1025.                 if(P_TOP[1]      > (rel_ymax<<shift)) P_TOP[1]     = (rel_ymax<<shift);

  1026.                 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);

  1027.                 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);

  1028.         

  1029.                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1030.                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1031. 

  1032.                 if(s->out_format == FMT_H263){

  1033.                     pred_x = P_MEDIAN[0];

  1034.                     pred_y = P_MEDIAN[1];

  1035.                 }else { /* mpeg1 at least */

  1036.                     pred_x= P_LEFT[0];

  1037.                     pred_y= P_LEFT[1];

  1038.                 }

  1039.             }else{

  1040.                 pred_x= P_LEFT[0];

  1041.                 pred_y= P_LEFT[1];

  1042.             }

  1043. 

  1044.         }

  1045.         dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 

  1046.                                       &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);

  1047.  

  1048.         break;

  1049.     }

  1050. 

  1051.     /* intra / predictive decision */

  1052.     xx = mb_x * 16;

  1053.     yy = mb_y * 16;

  1054. 

  1055.     pix = s->new_picture.data[0] + (yy * s->linesize) + xx;

  1056.     /* At this point (mx,my) are full-pell and the relative displacement */

  1057.     ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);

  1058.     

  1059.     sum = s->dsp.pix_sum(pix, s->linesize);

  1060.     

  1061.     varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;

  1062.     vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;

  1063. 

  1064. //printf("%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);

  1065.     pic->mb_var   [s->mb_width * mb_y + mb_x] = varc;

  1066.     pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard;

  1067.     pic->mb_mean  [s->mb_width * mb_y + mb_x] = (sum+128)>>8;

  1068. //    pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin; 

  1069.     pic->mb_var_sum    += varc;

  1070.     pic->mc_mb_var_sum += vard;

  1071. //printf("E%d %d %d %X %X %X\n", s->mb_width, mb_x, mb_y,(int)s, (int)s->mb_var, (int)s->mc_mb_var); fflush(stdout);

  1072.     

  1073. #if 0

  1074.     printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",

  1075. 	   varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);

  1076. #endif

  1077.     if(s->flags&CODEC_FLAG_HQ){

  1078.         if (vard <= 64 || vard < varc)

  1079.             s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);

  1080.         else

  1081.             s->scene_change_score+= s->qscale;

  1082. 

  1083.         if (vard*2 + 200 > varc)

  1084.             mb_type|= MB_TYPE_INTRA;

  1085.         if (varc*2 + 200 > vard){

  1086.             mb_type|= MB_TYPE_INTER;

  1087.             s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,

  1088. 				   pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);

  1089.         }else{

  1090.             mx <<=shift;

  1091.             my <<=shift;

  1092.         }

  1093.         if((s->flags&CODEC_FLAG_4MV)

  1094.            && !s->me.skip && varc>50 && vard>10){

  1095.             h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);

  1096.             mb_type|=MB_TYPE_INTER4V;

  1097. 

  1098.             set_p_mv_tables(s, mx, my, 0);

  1099.         }else

  1100.             set_p_mv_tables(s, mx, my, 1);

  1101.     }else{

  1102.         mb_type= MB_TYPE_INTER;

  1103. 

  1104.         dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,

  1105.                                     pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);

  1106.         

  1107.         if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)

  1108.             dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);

  1109. 

  1110.         if((s->flags&CODEC_FLAG_4MV)

  1111.            && !s->me.skip && varc>50 && vard>10){

  1112.             int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);

  1113.             if(dmin4 < dmin){

  1114.                 mb_type= MB_TYPE_INTER4V;

  1115.                 dmin=dmin4;

  1116.             }

  1117.         }

  1118.         pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin; 

  1119.         set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);

  1120.         

  1121.         if (vard <= 64 || vard < varc) {

  1122.             s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);

  1123.         }else{

  1124.             s->scene_change_score+= s->qscale;

  1125.         }

  1126.     }

  1127. 

  1128.     s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;

  1129. }

  1130. 

  1131. int ff_pre_estimate_p_frame_motion(MpegEncContext * s,

  1132.                                     int mb_x, int mb_y)

  1133. {

  1134.     int mx, my, range, dmin;

  1135.     int xmin, ymin, xmax, ymax;

  1136.     int rel_xmin, rel_ymin, rel_xmax, rel_ymax;

  1137.     int pred_x=0, pred_y=0;

  1138.     int P[10][2];

  1139.     const int shift= 1+s->quarter_sample;

  1140.     uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;

  1141.     const int mv_stride= s->mb_width + 2;

  1142.     const int xy= mb_x + 1 + (mb_y + 1)*mv_stride;

  1143.     

  1144.     assert(s->quarter_sample==0 || s->quarter_sample==1);

  1145. 

  1146.     s->me.pre_penalty_factor    = get_penalty_factor(s, s->avctx->me_pre_cmp);

  1147. 

  1148.     get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);

  1149.     rel_xmin= xmin - mb_x*16;

  1150.     rel_xmax= xmax - mb_x*16;

  1151.     rel_ymin= ymin - mb_y*16;

  1152.     rel_ymax= ymax - mb_y*16;

  1153.     s->me.skip=0;

  1154. 

  1155.     P_LEFT[0]       = s->p_mv_table[xy + 1][0];

  1156.     P_LEFT[1]       = s->p_mv_table[xy + 1][1];

  1157. 

  1158.     if(P_LEFT[0]       < (rel_xmin<<shift)) P_LEFT[0]       = (rel_xmin<<shift);

  1159. 

  1160.     /* special case for first line */

  1161.     if (mb_y == s->mb_height-1) {

  1162.         pred_x= P_LEFT[0];

  1163.         pred_y= P_LEFT[1];

  1164.         P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=

  1165.         P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME 

  1166.     } else {

  1167.         P_TOP[0]      = s->p_mv_table[xy + mv_stride    ][0];

  1168.         P_TOP[1]      = s->p_mv_table[xy + mv_stride    ][1];

  1169.         P_TOPRIGHT[0] = s->p_mv_table[xy + mv_stride - 1][0];

  1170.         P_TOPRIGHT[1] = s->p_mv_table[xy + mv_stride - 1][1];

  1171.         if(P_TOP[1]      < (rel_ymin<<shift)) P_TOP[1]     = (rel_ymin<<shift);

  1172.         if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);

  1173.         if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);

  1174.     

  1175.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1176.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1177. 

  1178.         pred_x = P_MEDIAN[0];

  1179.         pred_y = P_MEDIAN[1];

  1180.     }

  1181.     dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 

  1182.                                    &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);

  1183. 

  1184.     s->p_mv_table[xy][0] = mx<<shift;

  1185.     s->p_mv_table[xy][1] = my<<shift;

  1186.     

  1187.     return dmin;

  1188. }

  1189. 

  1190. int ff_estimate_motion_b(MpegEncContext * s,

  1191.                        int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)

  1192. {

  1193.     int mx, my, range, dmin;

  1194.     int xmin, ymin, xmax, ymax;

  1195.     int rel_xmin, rel_ymin, rel_xmax, rel_ymax;

  1196.     int pred_x=0, pred_y=0;

  1197.     int P[10][2];

  1198.     const int shift= 1+s->quarter_sample;

  1199.     const int mot_stride = s->mb_width + 2;

  1200.     const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;

  1201.     uint8_t * const ref_picture= picture->data[0];

  1202.     uint16_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;

  1203.     int mv_scale;

  1204.         

  1205.     s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);

  1206.     s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);

  1207.     s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);

  1208. 

  1209.     get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code);

  1210.     rel_xmin= xmin - mb_x*16;

  1211.     rel_xmax= xmax - mb_x*16;

  1212.     rel_ymin= ymin - mb_y*16;

  1213.     rel_ymax= ymax - mb_y*16;

  1214. 

  1215.     switch(s->me_method) {

  1216.     case ME_ZERO:

  1217.     default:

  1218. 	no_motion_search(s, &mx, &my);

  1219.         dmin = 0;

  1220.         mx-= mb_x*16;

  1221.         my-= mb_y*16;

  1222.         break;

  1223.     case ME_FULL:

  1224. 	dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);

  1225.         mx-= mb_x*16;

  1226.         my-= mb_y*16;

  1227.         break;

  1228.     case ME_LOG:

  1229. 	dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1230.         mx-= mb_x*16;

  1231.         my-= mb_y*16;

  1232.         break;

  1233.     case ME_PHODS:

  1234. 	dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1235.         mx-= mb_x*16;

  1236.         my-= mb_y*16;

  1237.         break;

  1238.     case ME_X1:

  1239.     case ME_EPZS:

  1240.        {

  1241.             P_LEFT[0]        = mv_table[mot_xy - 1][0];

  1242.             P_LEFT[1]        = mv_table[mot_xy - 1][1];

  1243. 

  1244.             if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);

  1245. 

  1246.             /* special case for first line */

  1247.             if (mb_y) {

  1248.                 P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];

  1249.                 P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];

  1250.                 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];

  1251.                 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];

  1252.                 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);

  1253.                 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);

  1254.                 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);

  1255.         

  1256.                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1257.                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1258.             }

  1259.             pred_x= P_LEFT[0];

  1260.             pred_y= P_LEFT[1];

  1261.         }

  1262.         

  1263.         if(mv_table == s->b_forw_mv_table){

  1264.             mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);

  1265.         }else{

  1266.             mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);

  1267.         }

  1268.         

  1269.         dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 

  1270.                                       picture, s->p_mv_table, mv_scale, mv_penalty);

  1271.  

  1272.         break;

  1273.     }

  1274.     

  1275.     dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,

  1276. 				   pred_x, pred_y, picture, 0, 0, mv_penalty);

  1277.                                    

  1278.     if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)

  1279.         dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);

  1280. 

  1281. //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);

  1282. //    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;

  1283.     mv_table[mot_xy][0]= mx;

  1284.     mv_table[mot_xy][1]= my;

  1285. 

  1286.     return dmin;

  1287. }

  1288. 

  1289. static inline int check_bidir_mv(MpegEncContext * s,

  1290.                    int mb_x, int mb_y,

  1291.                    int motion_fx, int motion_fy,

  1292.                    int motion_bx, int motion_by,

  1293.                    int pred_fx, int pred_fy,

  1294.                    int pred_bx, int pred_by)

  1295. {

  1296.     //FIXME optimize?

  1297.     //FIXME move into template?

  1298.     //FIXME better f_code prediction (max mv & distance)

  1299.     UINT16 *mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame

  1300.     uint8_t *dest_y = s->me.scratchpad;

  1301.     uint8_t *ptr;

  1302.     int dxy;

  1303.     int src_x, src_y;

  1304.     int fbmin;

  1305. 

  1306.     if(s->quarter_sample){

  1307.         dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);

  1308.         src_x = mb_x * 16 + (motion_fx >> 2);

  1309.         src_y = mb_y * 16 + (motion_fy >> 2);

  1310.         assert(src_x >=-16 && src_x<=s->width);

  1311.         assert(src_y >=-16 && src_y<=s->height);

  1312. 

  1313.         ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;

  1314.         s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);

  1315. 

  1316.         dxy = ((motion_by & 3) << 2) | (motion_bx & 3);

  1317.         src_x = mb_x * 16 + (motion_bx >> 2);

  1318.         src_y = mb_y * 16 + (motion_by >> 2);

  1319.         assert(src_x >=-16 && src_x<=s->width);

  1320.         assert(src_y >=-16 && src_y<=s->height);

  1321.     

  1322.         ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;

  1323.         s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);

  1324.     }else{

  1325.         dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);

  1326.         src_x = mb_x * 16 + (motion_fx >> 1);

  1327.         src_y = mb_y * 16 + (motion_fy >> 1);

  1328.         assert(src_x >=-16 && src_x<=s->width);

  1329.         assert(src_y >=-16 && src_y<=s->height);

  1330. 

  1331.         ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;

  1332.         s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

  1333. 

  1334.         dxy = ((motion_by & 1) << 1) | (motion_bx & 1);

  1335.         src_x = mb_x * 16 + (motion_bx >> 1);

  1336.         src_y = mb_y * 16 + (motion_by >> 1);

  1337.         assert(src_x >=-16 && src_x<=s->width);

  1338.         assert(src_y >=-16 && src_y<=s->height);

  1339.     

  1340.         ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;

  1341.         s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

  1342.     }

  1343. 

  1344.     fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor

  1345.            +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor

  1346.            + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);

  1347.            

  1348.     if(s->avctx->mb_cmp&FF_CMP_CHROMA){

  1349.     }

  1350.     //FIXME CHROMA !!!

  1351.            

  1352.     return fbmin;

  1353. }

  1354. 

  1355. /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/

  1356. static inline int bidir_refine(MpegEncContext * s,

  1357.                                   int mb_x, int mb_y)

  1358. {

  1359.     const int mot_stride = s->mb_width + 2;

  1360.     const int xy = (mb_y + 1)*mot_stride + mb_x + 1;

  1361.     int fbmin;

  1362.     int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];

  1363.     int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];

  1364.     int pred_bx= s->b_bidir_back_mv_table[xy-1][0];

  1365.     int pred_by= s->b_bidir_back_mv_table[xy-1][1];

  1366.     int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];

  1367.     int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];

  1368.     int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];

  1369.     int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];

  1370. 

  1371.     //FIXME do refinement and add flag

  1372.     

  1373.     fbmin= check_bidir_mv(s, mb_x, mb_y, 

  1374.                           motion_fx, motion_fy,

  1375.                           motion_bx, motion_by,

  1376.                           pred_fx, pred_fy,

  1377.                           pred_bx, pred_by);

  1378. 

  1379.    return fbmin;

  1380. }

  1381. 

  1382. static inline int direct_search(MpegEncContext * s,

  1383.                                 int mb_x, int mb_y)

  1384. {

  1385.     int P[10][2];

  1386.     const int mot_stride = s->mb_width + 2;

  1387.     const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;

  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.     uint16_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;

  1395.     

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

  1397.     ymax= xmax=   31>>shift;

  1398. 

  1399.     if(s->co_located_type_table[mb_x + mb_y*s->mb_width]==CO_LOCATED_TYPE_4MV){

  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.         s->me.co_located_mv[i][0]= s->motion_val[index][0];

  1410.         s->me.co_located_mv[i][1]= s->motion_val[index][1];

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

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

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

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

  1415. 

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

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

  1418.         max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding

  1419.         min+= (2*mb_x + (i& 1))*8 - 1;

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

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

  1422. 

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

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

  1425.         max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding

  1426.         min+= (2*mb_y + (i>>1))*8 - 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.     P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);

  1443.     P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);

  1444. 

  1445.     /* special case for first line */

  1446.     if (mb_y) {

  1447.         P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);

  1448.         P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);

  1449.         P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);

  1450.         P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);

  1451.     

  1452.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1453.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1454.     }

  1455.  

  1456.     //FIXME direct_search  ptr in context!!! (needed for chroma anyway or this will get messy)   

  1457.     if(s->flags&CODEC_FLAG_QPEL){

  1458.         dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 

  1459.                                                      &s->last_picture, mv_table, 1<<14, mv_penalty);

  1460.         dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,

  1461.                                                 0, 0, &s->last_picture, 0, 0, mv_penalty);

  1462.         

  1463.         if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)

  1464.             dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);

  1465.     }else{

  1466.         dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 

  1467.                                                      &s->last_picture, mv_table, 1<<15, mv_penalty);

  1468.         dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,

  1469.                                                 0, 0, &s->last_picture, 0, 0, mv_penalty);

  1470. 

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

  1472.             dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);

  1473.     }

  1474. 

  1475.     s->b_direct_mv_table[mot_xy][0]= mx;

  1476.     s->b_direct_mv_table[mot_xy][1]= my;

  1477.     return dmin;

  1478. }

  1479. 

  1480. void ff_estimate_b_frame_motion(MpegEncContext * s,

  1481.                              int mb_x, int mb_y)

  1482. {

  1483.     const int penalty_factor= s->me.mb_penalty_factor;

  1484.     int fmin, bmin, dmin, fbmin;

  1485.     int type=0;

  1486.     

  1487.     dmin= direct_search(s, mb_x, mb_y);

  1488. 

  1489.     fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;

  1490.     bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;

  1491. //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

  1492. 

  1493.     fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;

  1494. //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);

  1495.     {

  1496.         int score= fmin;

  1497.         type = MB_TYPE_FORWARD;

  1498.         

  1499.         // RAL: No MB_TYPE_DIRECT in MPEG-1 video (only MPEG-4)

  1500.         if (s->codec_id != CODEC_ID_MPEG1VIDEO && dmin <= score){

  1501.             score = dmin;

  1502.             type = MB_TYPE_DIRECT;

  1503.         }

  1504.         if(bmin<score){

  1505.             score=bmin;

  1506.             type= MB_TYPE_BACKWARD; 

  1507.         }

  1508.         if(fbmin<score){

  1509.             score=fbmin;

  1510.             type= MB_TYPE_BIDIR;

  1511.         }

  1512.         

  1513.         score= ((unsigned)(score*score + 128*256))>>16;

  1514.         s->current_picture.mc_mb_var_sum += score;

  1515.         s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSE

  1516.     }

  1517. 

  1518.     if(s->flags&CODEC_FLAG_HQ){

  1519.         type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter

  1520.         if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB

  1521.     }

  1522. 

  1523.     s->mb_type[mb_y*s->mb_width + mb_x]= type;

  1524. }

  1525. 

  1526. /* find best f_code for ME which do unlimited searches */

  1527. int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)

  1528. {

  1529.     if(s->me_method>=ME_EPZS){

  1530.         int score[8];

  1531.         int i, y;

  1532.         UINT8 * fcode_tab= s->fcode_tab;

  1533.         int best_fcode=-1;

  1534.         int best_score=-10000000;

  1535. 

  1536.         for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);

  1537. 

  1538.         for(y=0; y<s->mb_height; y++){

  1539.             int x;

  1540.             int xy= (y+1)* (s->mb_width+2) + 1;

  1541.             i= y*s->mb_width;

  1542.             for(x=0; x<s->mb_width; x++){

  1543.                 if(s->mb_type[i] & type){

  1544.                     int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],

  1545.                                      fcode_tab[mv_table[xy][1] + MAX_MV]);

  1546.                     int j;

  1547.                     

  1548.                     for(j=0; j<fcode && j<8; j++){

  1549.                         if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[i] < s->current_picture.mb_var[i])

  1550.                             score[j]-= 170;

  1551.                     }

  1552.                 }

  1553.                 i++;

  1554.                 xy++;

  1555.             }

  1556.         }

  1557.         

  1558.         for(i=1; i<8; i++){

  1559.             if(score[i] > best_score){

  1560.                 best_score= score[i];

  1561.                 best_fcode= i;

  1562.             }

  1563. //            printf("%d %d\n", i, score[i]);

  1564.         }

  1565. 

  1566. //    printf("fcode: %d type: %d\n", i, s->pict_type);

  1567.         return best_fcode;

  1568. /*        for(i=0; i<=MAX_FCODE; i++){

  1569.             printf("%d ", mv_num[i]);

  1570.         }

  1571.         printf("\n");*/

  1572.     }else{

  1573.         return 1;

  1574.     }

  1575. }

  1576. 

  1577. void ff_fix_long_p_mvs(MpegEncContext * s)

  1578. {

  1579.     const int f_code= s->f_code;

  1580.     int y;

  1581.     UINT8 * fcode_tab= s->fcode_tab;

  1582. //int clip=0;

  1583. //int noclip=0;

  1584.     /* clip / convert to intra 16x16 type MVs */

  1585.     for(y=0; y<s->mb_height; y++){

  1586.         int x;

  1587.         int xy= (y+1)* (s->mb_width+2)+1;

  1588.         int i= y*s->mb_width;

  1589.         for(x=0; x<s->mb_width; x++){

  1590.             if(s->mb_type[i]&MB_TYPE_INTER){

  1591.                 if(   fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code

  1592.                    || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0

  1593.                    || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code

  1594.                    || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){

  1595.                     s->mb_type[i] &= ~MB_TYPE_INTER;

  1596.                     s->mb_type[i] |= MB_TYPE_INTRA;

  1597.                     s->p_mv_table[xy][0] = 0;

  1598.                     s->p_mv_table[xy][1] = 0;

  1599. //clip++;

  1600.                 }

  1601. //else

  1602. //  noclip++;

  1603.             }

  1604.             xy++;

  1605.             i++;

  1606.         }

  1607.     }

  1608. //printf("%d no:%d %d//\n", clip, noclip, f_code);

  1609.     if(s->flags&CODEC_FLAG_4MV){

  1610.         const int wrap= 2+ s->mb_width*2;

  1611. 

  1612.         /* clip / convert to intra 8x8 type MVs */

  1613.         for(y=0; y<s->mb_height; y++){

  1614.             int xy= (y*2 + 1)*wrap + 1;

  1615.             int i= y*s->mb_width;

  1616.             int x;

  1617. 

  1618.             for(x=0; x<s->mb_width; x++){

  1619.                 if(s->mb_type[i]&MB_TYPE_INTER4V){

  1620.                     int block;

  1621.                     for(block=0; block<4; block++){

  1622.                         int off= (block& 1) + (block>>1)*wrap;

  1623.                         int mx= s->motion_val[ xy + off ][0];

  1624.                         int my= s->motion_val[ xy + off ][1];

  1625. 

  1626.                         if(   fcode_tab[mx + MAX_MV] > f_code

  1627.                            || fcode_tab[mx + MAX_MV] == 0

  1628.                            || fcode_tab[my + MAX_MV] > f_code

  1629.                            || fcode_tab[my + MAX_MV] == 0 ){

  1630.                             s->mb_type[i] &= ~MB_TYPE_INTER4V;

  1631.                             s->mb_type[i] |= MB_TYPE_INTRA;

  1632.                         }

  1633.                     }

  1634.                 }

  1635.                 xy+=2;

  1636.                 i++;

  1637.             }

  1638.         }

  1639.     }

  1640. }

  1641. 

  1642. void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)

  1643. {

  1644.     int y;

  1645.     UINT8 * fcode_tab= s->fcode_tab;

  1646. 

  1647.     // RAL: 8 in MPEG-1, 16 in MPEG-4

  1648.     int range = (((s->codec_id == CODEC_ID_MPEG1VIDEO) ? 8 : 16) << f_code);

  1649. 

  1650.     /* clip / convert to intra 16x16 type MVs */

  1651.     for(y=0; y<s->mb_height; y++){

  1652.         int x;

  1653.         int xy= (y+1)* (s->mb_width+2)+1;

  1654.         int i= y*s->mb_width;

  1655.         for(x=0; x<s->mb_width; x++)

  1656.             {

  1657.             if (s->mb_type[i] & type)    // RAL: "type" test added...

  1658.                 {

  1659.                 if (fcode_tab[mv_table[xy][0] + MAX_MV] > f_code || fcode_tab[mv_table[xy][0] + MAX_MV] == 0)

  1660.                     {

  1661.                     if(mv_table[xy][0]>0) 

  1662.                         mv_table[xy][0]=  range-1;

  1663.                     else

  1664.                         mv_table[xy][0]= -range;

  1665.                     }

  1666.                 if (fcode_tab[mv_table[xy][1] + MAX_MV] > f_code || fcode_tab[mv_table[xy][1] + MAX_MV] == 0)

  1667.                     {

  1668.                     if(mv_table[xy][1]>0) 

  1669.                         mv_table[xy][1]=  range-1;

  1670.                     else                  

  1671.                         mv_table[xy][1]= -range;

  1672.             }

  1673.             }

  1674.             xy++;

  1675.             i++;

  1676.         }

  1677.     }

  1678. }