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.
2804 Commits
32 Branches
484MB
Tree: 3615e2be84
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '3615e2be84'
${ noResults }
FFmpeg/libavcodec/motion_est.c

1746 lines
59KB
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.  

  24. /**

  25.  * @file motion_est.c

  26.  * Motion estimation.

  27.  */

  28.  

  29. #include <stdlib.h>

  30. #include <stdio.h>

  31. #include <limits.h>

  32. #include "avcodec.h"

  33. #include "dsputil.h"

  34. #include "mpegvideo.h"

  35. 

  36. //#undef NDEBUG

  37. //#include <assert.h>

  38. 

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

  40. 

  41. #define P_LEFT P[1]

  42. #define P_TOP P[2]

  43. #define P_TOPRIGHT P[3]

  44. #define P_MEDIAN P[4]

  45. #define P_MV1 P[9]

  46. 

  47. static inline int sad_hpel_motion_search(MpegEncContext * s,

  48. 				  int *mx_ptr, int *my_ptr, int dmin,

  49. 				  int xmin, int ymin, int xmax, int ymax,

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

  51.                                   int n, int size, uint8_t * const mv_penalty);

  52. 

  53. static inline int update_map_generation(MpegEncContext * s)

  54. {

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

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

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

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

  59.     }

  60.     return s->me.map_generation;

  61. }

  62. 

  63. /* shape adaptive search stuff */

  64. typedef struct Minima{

  65.     int height;

  66.     int x, y;

  67.     int checked;

  68. }Minima;

  69. 

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

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

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

  73.     

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

  75. }

  76.                                   

  77. /* SIMPLE */

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

  79. 

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

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

  82. 

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

  84. {\

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

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

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

  88. }

  89. 

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

  91. {\

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

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

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

  95. }

  96. 

  97. #include "motion_est_template.c"

  98. #undef RENAME

  99. #undef CMP

  100. #undef CMP_HPEL

  101. #undef CMP_QPEL

  102. #undef INIT

  103. 

  104. /* SIMPLE CHROMA */

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

  106. 

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

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

  109. if(chroma_cmp){\

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

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

  112. \

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

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

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

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

  117. }

  118. 

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

  120. {\

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

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

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

  124.     if(chroma_cmp_sub){\

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

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

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

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

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

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

  131.     }\

  132. }

  133. 

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

  135. {\

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

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

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

  139.     if(chroma_cmp_sub){\

  140.         int cxy, c;\

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

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

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

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

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

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

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

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

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

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

  151.     }\

  152. }

  153. 

  154. #include "motion_est_template.c"

  155. #undef RENAME

  156. #undef CMP

  157. #undef CMP_HPEL

  158. #undef CMP_QPEL

  159. #undef INIT

  160. 

  161. /* SIMPLE DIRECT HPEL */

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

  163. //FIXME precalc divisions stuff

  164. 

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

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

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

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

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

  170.         int i;\

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

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

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

  174.             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;\

  175.             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;\

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

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

  178. \

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

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

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

  182.         }\

  183.     }else{\

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

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

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

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

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

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

  190.         \

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

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

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

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

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

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

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

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

  199. \

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

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

  202.     }\

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

  204. }else\

  205.     d= 256*256*256*32;

  206. 

  207. 

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

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

  210. 

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

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

  213.     

  214. #include "motion_est_template.c"

  215. #undef RENAME

  216. #undef CMP

  217. #undef CMP_HPEL

  218. #undef CMP_QPEL

  219. #undef INIT

  220. #undef CMP_DIRECT

  221. 

  222. /* SIMPLE DIRECT QPEL */

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

  224. 

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

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

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

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

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

  230.         int i;\

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

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

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

  234.             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;\

  235.             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;\

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

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

  238. \

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

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

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

  242.         }\

  243.     }else{\

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

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

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

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

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

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

  250. \

  251.         qpel_put[1][fxy](s->me.scratchpad               , (ref_y ) + (fx>>2) + (fy>>2)*(stride)               , stride);\

  252.         qpel_put[1][fxy](s->me.scratchpad + 8           , (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8           , stride);\

  253.         qpel_put[1][fxy](s->me.scratchpad     + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride)     + 8*stride, stride);\

  254.         qpel_put[1][fxy](s->me.scratchpad + 8 + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 + 8*stride, stride);\

  255.         qpel_avg[1][bxy](s->me.scratchpad               , (ref2_y) + (bx>>2) + (by>>2)*(stride)               , stride);\

  256.         qpel_avg[1][bxy](s->me.scratchpad + 8           , (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8           , stride);\

  257.         qpel_avg[1][bxy](s->me.scratchpad     + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride)     + 8*stride, stride);\

  258.         qpel_avg[1][bxy](s->me.scratchpad + 8 + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8 + 8*stride, stride);\

  259.     }\

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

  261. }else\

  262.     d= 256*256*256*32;

  263. 

  264. 

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

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

  267. 

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

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

  270. 

  271. #include "motion_est_template.c"

  272. #undef RENAME

  273. #undef CMP

  274. #undef CMP_HPEL

  275. #undef CMP_QPEL

  276. #undef INIT

  277. #undef CMP__DIRECT

  278. 

  279. 

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

  281.     return 0;

  282. }

  283. 

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

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

  286.     int i;

  287.     

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

  289. 

  290.     switch(type&0xFF){

  291.     case FF_CMP_SAD:

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

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

  294.         break;

  295.     case FF_CMP_SATD:

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

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

  298.         break;

  299.     case FF_CMP_SSE:

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

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

  302.         break;

  303.     case FF_CMP_DCT:

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

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

  306.         break;

  307.     case FF_CMP_PSNR:

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

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

  310.         break;

  311.     case FF_CMP_BIT:

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

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

  314.         break;

  315.     case FF_CMP_RD:

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

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

  318.         break;

  319.     case FF_CMP_ZERO:

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

  321.             cmp[i]= zero_cmp;

  322.         }

  323.         break;

  324.     default:

  325.         av_log(s->avctx, AV_LOG_ERROR,"internal error in cmp function selection\n");

  326.     }

  327. }

  328. 

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

  330.     switch(type&0xFF){

  331.     default:

  332.     case FF_CMP_SAD:

  333.         return s->qscale*2;

  334.     case FF_CMP_DCT:

  335.         return s->qscale*3;

  336.     case FF_CMP_SATD:

  337.         return s->qscale*6;

  338.     case FF_CMP_SSE:

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

  340.     case FF_CMP_BIT:

  341.         return 1;

  342.     case FF_CMP_RD:

  343.     case FF_CMP_PSNR:

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

  345.     }

  346. }

  347. 

  348. void ff_init_me(MpegEncContext *s){

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

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

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

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

  353. 

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

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

  356.             s->me.sub_motion_search= simple_chroma_qpel_motion_search;

  357.         else

  358.             s->me.sub_motion_search= simple_qpel_motion_search;

  359.     }else{

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

  361.             s->me.sub_motion_search= simple_chroma_hpel_motion_search;

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

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

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

  365.             s->me.sub_motion_search= sad_hpel_motion_search;

  366.         else

  367.             s->me.sub_motion_search= simple_hpel_motion_search;

  368.     }

  369. 

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

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

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

  373.     }else{

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

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

  376.     }

  377.     

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

  379.         s->me.pre_motion_search= simple_chroma_epzs_motion_search;

  380.     }else{

  381.         s->me.pre_motion_search= simple_epzs_motion_search;

  382.     }

  383.     

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

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

  386.             s->me.get_mb_score= simple_chroma_qpel_get_mb_score;

  387.         else

  388.             s->me.get_mb_score= simple_qpel_get_mb_score;

  389.     }else{

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

  391.             s->me.get_mb_score= simple_chroma_hpel_get_mb_score;

  392.         else

  393.             s->me.get_mb_score= simple_hpel_get_mb_score;

  394.     }

  395. }

  396.       

  397. #if 0

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

  399. {

  400.     int s, i, j;

  401. 

  402.     s = 0;

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

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

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

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

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

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

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

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

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

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

  413. 	    pix += 8;

  414. 	}

  415. 	pix += line_size - 16;

  416.     }

  417.     return s;

  418. }

  419. #endif

  420. 

  421. static inline void no_motion_search(MpegEncContext * s,

  422. 				    int *mx_ptr, int *my_ptr)

  423. {

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

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

  426. }

  427. 

  428. static int full_motion_search(MpegEncContext * s,

  429.                               int *mx_ptr, int *my_ptr, int range,

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

  431. {

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

  433.     int mx, my, dmin, d;

  434.     uint8_t *pix;

  435. 

  436.     xx = 16 * s->mb_x;

  437.     yy = 16 * s->mb_y;

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

  439.     if (x1 < xmin)

  440. 	x1 = xmin;

  441.     x2 = xx + range - 1;

  442.     if (x2 > xmax)

  443. 	x2 = xmax;

  444.     y1 = yy - range + 1;

  445.     if (y1 < ymin)

  446. 	y1 = ymin;

  447.     y2 = yy + range - 1;

  448.     if (y2 > ymax)

  449. 	y2 = ymax;

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

  451.     dmin = 0x7fffffff;

  452.     mx = 0;

  453.     my = 0;

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

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

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

  457. 			     s->linesize);

  458. 	    if (d < dmin ||

  459. 		(d == dmin &&

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

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

  462. 		dmin = d;

  463. 		mx = x;

  464. 		my = y;

  465. 	    }

  466. 	}

  467.     }

  468. 

  469.     *mx_ptr = mx;

  470.     *my_ptr = my;

  471. 

  472. #if 0

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

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

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

  476.     }

  477. #endif

  478.     return dmin;

  479. }

  480. 

  481. 

  482. static int log_motion_search(MpegEncContext * s,

  483.                              int *mx_ptr, int *my_ptr, int range,

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

  485. {

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

  487.     int mx, my, dmin, d;

  488.     uint8_t *pix;

  489. 

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

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

  492. 

  493.     /* Left limit */

  494.     x1 = xx - range;

  495.     if (x1 < xmin)

  496. 	x1 = xmin;

  497. 

  498.     /* Right limit */

  499.     x2 = xx + range;

  500.     if (x2 > xmax)

  501. 	x2 = xmax;

  502. 

  503.     /* Upper limit */

  504.     y1 = yy - range;

  505.     if (y1 < ymin)

  506. 	y1 = ymin;

  507. 

  508.     /* Lower limit */

  509.     y2 = yy + range;

  510.     if (y2 > ymax)

  511. 	y2 = ymax;

  512. 

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

  514.     dmin = 0x7fffffff;

  515.     mx = 0;

  516.     my = 0;

  517. 

  518.     do {

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

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

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

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

  523. 		    dmin = d;

  524. 		    mx = x;

  525. 		    my = y;

  526. 		}

  527. 	    }

  528. 	}

  529. 

  530. 	range = range >> 1;

  531. 

  532. 	x1 = mx - range;

  533. 	if (x1 < xmin)

  534. 	    x1 = xmin;

  535. 

  536. 	x2 = mx + range;

  537. 	if (x2 > xmax)

  538. 	    x2 = xmax;

  539. 

  540. 	y1 = my - range;

  541. 	if (y1 < ymin)

  542. 	    y1 = ymin;

  543. 

  544. 	y2 = my + range;

  545. 	if (y2 > ymax)

  546. 	    y2 = ymax;

  547. 

  548.     } while (range >= 1);

  549. 

  550. #ifdef DEBUG

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

  552. #endif

  553.     *mx_ptr = mx;

  554.     *my_ptr = my;

  555.     return dmin;

  556. }

  557. 

  558. static int phods_motion_search(MpegEncContext * s,

  559.                                int *mx_ptr, int *my_ptr, int range,

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

  561. {

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

  563.     int mx, my, dminx, dminy;

  564.     uint8_t *pix;

  565. 

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

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

  568. 

  569.     /* Left limit */

  570.     x1 = xx - range;

  571.     if (x1 < xmin)

  572. 	x1 = xmin;

  573. 

  574.     /* Right limit */

  575.     x2 = xx + range;

  576.     if (x2 > xmax)

  577. 	x2 = xmax;

  578. 

  579.     /* Upper limit */

  580.     y1 = yy - range;

  581.     if (y1 < ymin)

  582. 	y1 = ymin;

  583. 

  584.     /* Lower limit */

  585.     y2 = yy + range;

  586.     if (y2 > ymax)

  587. 	y2 = ymax;

  588. 

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

  590.     mx = 0;

  591.     my = 0;

  592. 

  593.     x = xx;

  594.     y = yy;

  595.     do {

  596.         dminx = 0x7fffffff;

  597.         dminy = 0x7fffffff;

  598. 

  599. 	lastx = x;

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

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

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

  603. 		dminx = d;

  604. 		mx = x;

  605. 	    }

  606. 	}

  607. 

  608. 	x = lastx;

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

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

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

  612. 		dminy = d;

  613. 		my = y;

  614. 	    }

  615. 	}

  616. 

  617. 	range = range >> 1;

  618. 

  619. 	x = mx;

  620. 	y = my;

  621. 	x1 = mx - range;

  622. 	if (x1 < xmin)

  623. 	    x1 = xmin;

  624. 

  625. 	x2 = mx + range;

  626. 	if (x2 > xmax)

  627. 	    x2 = xmax;

  628. 

  629. 	y1 = my - range;

  630. 	if (y1 < ymin)

  631. 	    y1 = ymin;

  632. 

  633. 	y2 = my + range;

  634. 	if (y2 > ymax)

  635. 	    y2 = ymax;

  636. 

  637.     } while (range >= 1);

  638. 

  639. #ifdef DEBUG

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

  641. #endif

  642. 

  643.     /* half pixel search */

  644.     *mx_ptr = mx;

  645.     *my_ptr = my;

  646.     return dminy;

  647. }

  648. 

  649. 

  650. #define Z_THRESHOLD 256

  651. 

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

  653. {\

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

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

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

  657. }

  658. 

  659. static inline int sad_hpel_motion_search(MpegEncContext * s,

  660. 				  int *mx_ptr, int *my_ptr, int dmin,

  661. 				  int xmin, int ymin, int xmax, int ymax,

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

  663.                                   int n, int size, uint8_t * const mv_penalty)

  664. {

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

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

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

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

  669.     uint8_t *pix, *ptr;

  670.     op_pixels_abs_func pix_abs_x2;

  671.     op_pixels_abs_func pix_abs_y2;

  672.     op_pixels_abs_func pix_abs_xy2;

  673.     

  674.     if(size==0){

  675.         pix_abs_x2 = s->dsp.pix_abs16x16_x2;

  676.         pix_abs_y2 = s->dsp.pix_abs16x16_y2;

  677.         pix_abs_xy2= s->dsp.pix_abs16x16_xy2;

  678.     }else{

  679.         pix_abs_x2 = s->dsp.pix_abs8x8_x2;

  680.         pix_abs_y2 = s->dsp.pix_abs8x8_y2;

  681.         pix_abs_xy2= s->dsp.pix_abs8x8_xy2;

  682.     }

  683. 

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

  685. //    printf("S");

  686.         *mx_ptr = 0;

  687.         *my_ptr = 0;

  688.         return dmin;

  689.     }

  690. //    printf("N");

  691.         

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

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

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

  695. 

  696.     mx = *mx_ptr;

  697.     my = *my_ptr;

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

  699.     

  700.     dminh = dmin;

  701. 

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

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

  704.         int dx=0, dy=0;

  705.         int d, pen_x, pen_y; 

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

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

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

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

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

  711.         mx<<=1;

  712.         my<<=1;

  713. 

  714.         

  715.         pen_x= pred_x + mx;

  716.         pen_y= pred_y + my;

  717. 

  718.         ptr-= s->linesize;

  719.         if(t<=b){

  720.             CHECK_SAD_HALF_MV(y2 , 0, -1)

  721.             if(l<=r){

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

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

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

  725.                     ptr+= s->linesize;

  726.                 }else{

  727.                     ptr+= s->linesize;

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

  729.                 }

  730.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

  731.             }else{

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

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

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

  735.                     ptr+= s->linesize;

  736.                 }else{

  737.                     ptr+= s->linesize;

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

  739.                 }

  740.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

  741.             }

  742.         }else{

  743.             if(l<=r){

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

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

  746.                     ptr+= s->linesize;

  747.                 }else{

  748.                     ptr+= s->linesize;

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

  750.                 }

  751.                 CHECK_SAD_HALF_MV(x2 , -1,  0)

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

  753.             }else{

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

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

  756.                     ptr+= s->linesize;

  757.                 }else{

  758.                     ptr+= s->linesize;

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

  760.                 }

  761.                 CHECK_SAD_HALF_MV(x2 , +1,  0)

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

  763.             }

  764.             CHECK_SAD_HALF_MV(y2 ,  0, +1)

  765.         }

  766.         mx+=dx;

  767.         my+=dy;

  768. 

  769.     }else{

  770.         mx<<=1;

  771.         my<<=1;

  772.     }

  773. 

  774.     *mx_ptr = mx;

  775.     *my_ptr = my;

  776.     return dminh;

  777. }

  778. 

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

  780. {

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

  782.     

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

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

  785. 

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

  787.     if(mv4){

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

  789. 

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

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

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

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

  794. 

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

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

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

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

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

  800.     }

  801. }

  802. 

  803. /**

  804.  * get fullpel ME search limits.

  805.  * @param range the approximate search range for the old ME code, unused for EPZS and newer

  806.  */

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

  808. {

  809.     if(s->avctx->me_range) *range= s->avctx->me_range >> 1;

  810.     else                   *range= 16;

  811. 

  812.     if (s->unrestricted_mv) {

  813.         *xmin = -16;

  814.         *ymin = -16;

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

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

  817.     } else {

  818.         *xmin = 0;

  819.         *ymin = 0;

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

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

  822.     }

  823.     

  824.     //FIXME try to limit x/y min/max if me_range is set

  825. }

  826. 

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

  828. {

  829.     int block;

  830.     int P[10][2];

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

  832.     uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;

  833.     int same=1;

  834. 

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

  836.         int mx4, my4;

  837.         int pred_x4, pred_y4;

  838.         int dmin4;

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

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

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

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

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

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

  845.         const int rel_xmin4= xmin;

  846.         const int rel_xmax4= xmax;

  847.         const int rel_ymin4= ymin;

  848.         const int rel_ymax4= ymax;

  849. #else

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

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

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

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

  854. #endif

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

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

  857. 

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

  859. 

  860.         /* special case for first line */

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

  862.             pred_x4= P_LEFT[0];

  863.             pred_y4= P_LEFT[1];

  864.         } else {

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

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

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

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

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

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

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

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

  873.     

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

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

  876. 

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

  878.                 pred_x4 = P_MEDIAN[0];

  879.                 pred_y4 = P_MEDIAN[1];

  880. #if 0

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

  882.                 pred_x4= P_LEFT[0];

  883.                 pred_y4= P_LEFT[1];

  884.             }

  885. #endif

  886.         }

  887.         P_MV1[0]= mx;

  888.         P_MV1[1]= my;

  889. 

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

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

  892. 

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

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

  895.         

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

  897.             int dxy;

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

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

  900. 

  901.             if(s->quarter_sample){

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

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

  904. 

  905.                 if(s->no_rounding)

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

  907.                 else

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

  909.             }else{

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

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

  912. 

  913.                 if(s->no_rounding)

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

  915.                 else

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

  917.             }

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

  919.         }else

  920.             dmin_sum+= dmin4;

  921. 

  922.         if(s->quarter_sample){

  923.             mx4_sum+= mx4/2;

  924.             my4_sum+= my4/2;

  925.         }else{

  926.             mx4_sum+= mx4;

  927.             my4_sum+= my4;

  928.         }

  929.             

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

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

  932. 

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

  934.     }

  935.     

  936.     if(same)

  937.         return INT_MAX;

  938.     

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

  940.         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);

  941.     }

  942.     

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

  944.         int dxy;

  945.         int mx, my;

  946.         int offset;

  947. 

  948.         mx= ff_h263_round_chroma(mx4_sum);

  949.         my= ff_h263_round_chroma(my4_sum);

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

  951.         

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

  953.        

  954.         if(s->no_rounding){

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

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

  957.         }else{

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

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

  960.         }

  961. 

  962.         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);

  963.         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);

  964.     }

  965. 

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

  967.     /*case FF_CMP_SSE:

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

  969.     case FF_CMP_RD:

  970.         return dmin_sum;

  971.     default:

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

  973.     }

  974. }

  975. 

  976. void ff_estimate_p_frame_motion(MpegEncContext * s,

  977.                                 int mb_x, int mb_y)

  978. {

  979.     uint8_t *pix, *ppix;

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

  981.     int xmin, ymin, xmax, ymax;

  982.     int rel_xmin, rel_ymin, rel_xmax, rel_ymax;

  983.     int pred_x=0, pred_y=0;

  984.     int P[10][2];

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

  986.     int mb_type=0;

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

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

  989.     uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;

  990.     

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

  992. 

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

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

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

  996. 

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

  998.     rel_xmin= xmin - mb_x*16;

  999.     rel_xmax= xmax - mb_x*16;

  1000.     rel_ymin= ymin - mb_y*16;

  1001.     rel_ymax= ymax - mb_y*16;

  1002.     s->me.skip=0;

  1003. 

  1004.     switch(s->me_method) {

  1005.     case ME_ZERO:

  1006.     default:

  1007. 	no_motion_search(s, &mx, &my);

  1008.         mx-= mb_x*16;

  1009.         my-= mb_y*16;

  1010.         dmin = 0;

  1011.         break;

  1012.     case ME_FULL:

  1013. 	dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);

  1014.         mx-= mb_x*16;

  1015.         my-= mb_y*16;

  1016.         break;

  1017.     case ME_LOG:

  1018. 	dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1019.         mx-= mb_x*16;

  1020.         my-= mb_y*16;

  1021.         break;

  1022.     case ME_PHODS:

  1023. 	dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1024.         mx-= mb_x*16;

  1025.         my-= mb_y*16;

  1026.         break;

  1027.     case ME_X1:

  1028.     case ME_EPZS:

  1029.        {

  1030.             const int mot_stride = s->block_wrap[0];

  1031.             const int mot_xy = s->block_index[0];

  1032. 

  1033.             P_LEFT[0]       = s->motion_val[mot_xy - 1][0];

  1034.             P_LEFT[1]       = s->motion_val[mot_xy - 1][1];

  1035. 

  1036.             if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);

  1037. 

  1038.             if(mb_y) {

  1039.                 P_TOP[0]      = s->motion_val[mot_xy - mot_stride    ][0];

  1040.                 P_TOP[1]      = s->motion_val[mot_xy - mot_stride    ][1];

  1041.                 P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];

  1042.                 P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];

  1043.                 if(P_TOP[1]      > (rel_ymax<<shift)) P_TOP[1]     = (rel_ymax<<shift);

  1044.                 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);

  1045.                 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);

  1046.         

  1047.                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1048.                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1049. 

  1050.                 if(s->out_format == FMT_H263){

  1051.                     pred_x = P_MEDIAN[0];

  1052.                     pred_y = P_MEDIAN[1];

  1053.                 }else { /* mpeg1 at least */

  1054.                     pred_x= P_LEFT[0];

  1055.                     pred_y= P_LEFT[1];

  1056.                 }

  1057.             }else{

  1058.                 pred_x= P_LEFT[0];

  1059.                 pred_y= P_LEFT[1];

  1060.             }

  1061. 

  1062.         }

  1063.         dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 

  1064.                                       &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);

  1065.  

  1066.         break;

  1067.     }

  1068. 

  1069.     /* intra / predictive decision */

  1070.     xx = mb_x * 16;

  1071.     yy = mb_y * 16;

  1072. 

  1073.     pix = s->new_picture.data[0] + (yy * s->linesize) + xx;

  1074.     /* At this point (mx,my) are full-pell and the relative displacement */

  1075.     ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);

  1076.     

  1077.     sum = s->dsp.pix_sum(pix, s->linesize);

  1078.     

  1079.     varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;

  1080.     vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;

  1081. 

  1082. //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);

  1083.     pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;

  1084.     pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;

  1085.     pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;

  1086. //    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 

  1087.     pic->mb_var_sum    += varc;

  1088.     pic->mc_mb_var_sum += vard;

  1089. //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);

  1090.     

  1091. #if 0

  1092.     printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",

  1093. 	   varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);

  1094. #endif

  1095.     if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){

  1096.         if (vard <= 64 || vard < varc)

  1097.             s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);

  1098.         else

  1099.             s->scene_change_score+= s->qscale;

  1100. 

  1101.         if (vard*2 + 200 > varc)

  1102.             mb_type|= MB_TYPE_INTRA;

  1103.         if (varc*2 + 200 > vard){

  1104.             mb_type|= MB_TYPE_INTER;

  1105.             s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,

  1106. 				   pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);

  1107.             if(s->flags&CODEC_FLAG_MV0)

  1108.                 if(mx || my)

  1109.                     mb_type |= MB_TYPE_SKIPED; //FIXME check difference

  1110.         }else{

  1111.             mx <<=shift;

  1112.             my <<=shift;

  1113.         }

  1114.         if((s->flags&CODEC_FLAG_4MV)

  1115.            && !s->me.skip && varc>50 && vard>10){

  1116.             if(h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift) < INT_MAX)

  1117.                 mb_type|=MB_TYPE_INTER4V;

  1118. 

  1119.             set_p_mv_tables(s, mx, my, 0);

  1120.         }else

  1121.             set_p_mv_tables(s, mx, my, 1);

  1122.     }else{

  1123.         int intra_score, i;

  1124.         mb_type= MB_TYPE_INTER;

  1125. 

  1126.         dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,

  1127.                                     pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);

  1128.         

  1129.         if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)

  1130.             dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);

  1131. 

  1132.         if((s->flags&CODEC_FLAG_4MV)

  1133.            && !s->me.skip && varc>50 && vard>10){

  1134.             int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);

  1135.             if(dmin4 < dmin){

  1136.                 mb_type= MB_TYPE_INTER4V;

  1137.                 dmin=dmin4;

  1138.             }

  1139.         }

  1140.                 

  1141. //        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 

  1142.         set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);

  1143. 

  1144.         /* get intra luma score */

  1145.         if((s->avctx->mb_cmp&0xFF)==FF_CMP_SSE){

  1146.             intra_score= (varc<<8) - 500; //FIXME dont scale it down so we dont have to fix it

  1147.         }else{

  1148.             int mean= (sum+128)>>8;

  1149.             mean*= 0x01010101;

  1150.             

  1151.             for(i=0; i<16; i++){

  1152.                 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 0]) = mean;

  1153.                 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 4]) = mean;

  1154.                 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+ 8]) = mean;

  1155.                 *(uint32_t*)(&s->me.scratchpad[i*s->linesize+12]) = mean;

  1156.             }

  1157. 

  1158.             intra_score= s->dsp.mb_cmp[0](s, s->me.scratchpad, pix, s->linesize);

  1159.         }

  1160. #if 0 //FIXME

  1161.         /* get chroma score */

  1162.         if(s->avctx->mb_cmp&FF_CMP_CHROMA){

  1163.             for(i=1; i<3; i++){

  1164.                 uint8_t *dest_c;

  1165.                 int mean;

  1166.                 

  1167.                 if(s->out_format == FMT_H263){

  1168.                     mean= (s->dc_val[i][mb_x + (mb_y+1)*(s->mb_width+2)] + 4)>>3; //FIXME not exact but simple ;)

  1169.                 }else{

  1170.                     mean= (s->last_dc[i] + 4)>>3;

  1171.                 }

  1172.                 dest_c = s->new_picture.data[i] + (mb_y * 8  * (s->uvlinesize)) + mb_x * 8;

  1173.                 

  1174.                 mean*= 0x01010101;

  1175.                 for(i=0; i<8; i++){

  1176.                     *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 0]) = mean;

  1177.                     *(uint32_t*)(&s->me.scratchpad[i*s->uvlinesize+ 4]) = mean;

  1178.                 }

  1179.                 

  1180.                 intra_score+= s->dsp.mb_cmp[1](s, s->me.scratchpad, dest_c, s->uvlinesize);

  1181.             }                

  1182.         }

  1183. #endif

  1184.         intra_score += s->me.mb_penalty_factor*16;

  1185.         

  1186.         if(intra_score < dmin){

  1187.             mb_type= MB_TYPE_INTRA;

  1188.             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= MB_TYPE_INTRA; //FIXME cleanup

  1189.         }else

  1190.             s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;

  1191.         

  1192.         if (vard <= 64 || vard < varc) { //FIXME

  1193.             s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);

  1194.         }else{

  1195.             s->scene_change_score+= s->qscale;

  1196.         }

  1197.     }

  1198. 

  1199.     s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;

  1200. }

  1201. 

  1202. int ff_pre_estimate_p_frame_motion(MpegEncContext * s,

  1203.                                     int mb_x, int mb_y)

  1204. {

  1205.     int mx, my, range, dmin;

  1206.     int xmin, ymin, xmax, ymax;

  1207.     int rel_xmin, rel_ymin, rel_xmax, rel_ymax;

  1208.     int pred_x=0, pred_y=0;

  1209.     int P[10][2];

  1210.     const int shift= 1+s->quarter_sample;

  1211.     uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;

  1212.     const int xy= mb_x + mb_y*s->mb_stride;

  1213.     

  1214.     assert(s->quarter_sample==0 || s->quarter_sample==1);

  1215. 

  1216.     s->me.pre_penalty_factor    = get_penalty_factor(s, s->avctx->me_pre_cmp);

  1217. 

  1218.     get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);

  1219.     rel_xmin= xmin - mb_x*16;

  1220.     rel_xmax= xmax - mb_x*16;

  1221.     rel_ymin= ymin - mb_y*16;

  1222.     rel_ymax= ymax - mb_y*16;

  1223.     s->me.skip=0;

  1224. 

  1225.     P_LEFT[0]       = s->p_mv_table[xy + 1][0];

  1226.     P_LEFT[1]       = s->p_mv_table[xy + 1][1];

  1227. 

  1228.     if(P_LEFT[0]       < (rel_xmin<<shift)) P_LEFT[0]       = (rel_xmin<<shift);

  1229. 

  1230.     /* special case for first line */

  1231.     if (mb_y == s->mb_height-1) {

  1232.         pred_x= P_LEFT[0];

  1233.         pred_y= P_LEFT[1];

  1234.         P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=

  1235.         P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME 

  1236.     } else {

  1237.         P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];

  1238.         P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];

  1239.         P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];

  1240.         P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];

  1241.         if(P_TOP[1]      < (rel_ymin<<shift)) P_TOP[1]     = (rel_ymin<<shift);

  1242.         if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);

  1243.         if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);

  1244.     

  1245.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1246.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1247. 

  1248.         pred_x = P_MEDIAN[0];

  1249.         pred_y = P_MEDIAN[1];

  1250.     }

  1251.     dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 

  1252.                                    &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);

  1253. 

  1254.     s->p_mv_table[xy][0] = mx<<shift;

  1255.     s->p_mv_table[xy][1] = my<<shift;

  1256.     

  1257.     return dmin;

  1258. }

  1259. 

  1260. static int ff_estimate_motion_b(MpegEncContext * s,

  1261.                        int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)

  1262. {

  1263.     int mx, my, range, dmin;

  1264.     int xmin, ymin, xmax, ymax;

  1265.     int rel_xmin, rel_ymin, rel_xmax, rel_ymax;

  1266.     int pred_x=0, pred_y=0;

  1267.     int P[10][2];

  1268.     const int shift= 1+s->quarter_sample;

  1269.     const int mot_stride = s->mb_stride;

  1270.     const int mot_xy = mb_y*mot_stride + mb_x;

  1271.     uint8_t * const ref_picture= picture->data[0];

  1272.     uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;

  1273.     int mv_scale;

  1274.         

  1275.     s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);

  1276.     s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);

  1277.     s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);

  1278. 

  1279.     get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);

  1280.     rel_xmin= xmin - mb_x*16;

  1281.     rel_xmax= xmax - mb_x*16;

  1282.     rel_ymin= ymin - mb_y*16;

  1283.     rel_ymax= ymax - mb_y*16;

  1284. 

  1285.     switch(s->me_method) {

  1286.     case ME_ZERO:

  1287.     default:

  1288. 	no_motion_search(s, &mx, &my);

  1289.         dmin = 0;

  1290.         mx-= mb_x*16;

  1291.         my-= mb_y*16;

  1292.         break;

  1293.     case ME_FULL:

  1294. 	dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);

  1295.         mx-= mb_x*16;

  1296.         my-= mb_y*16;

  1297.         break;

  1298.     case ME_LOG:

  1299. 	dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1300.         mx-= mb_x*16;

  1301.         my-= mb_y*16;

  1302.         break;

  1303.     case ME_PHODS:

  1304. 	dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1305.         mx-= mb_x*16;

  1306.         my-= mb_y*16;

  1307.         break;

  1308.     case ME_X1:

  1309.     case ME_EPZS:

  1310.        {

  1311.             P_LEFT[0]        = mv_table[mot_xy - 1][0];

  1312.             P_LEFT[1]        = mv_table[mot_xy - 1][1];

  1313. 

  1314.             if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);

  1315. 

  1316.             /* special case for first line */

  1317.             if (mb_y) {

  1318.                 P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];

  1319.                 P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];

  1320.                 P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];

  1321.                 P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];

  1322.                 if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);

  1323.                 if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);

  1324.                 if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);

  1325.         

  1326.                 P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1327.                 P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1328.             }

  1329.             pred_x= P_LEFT[0];

  1330.             pred_y= P_LEFT[1];

  1331.         }

  1332.         

  1333.         if(mv_table == s->b_forw_mv_table){

  1334.             mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);

  1335.         }else{

  1336.             mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);

  1337.         }

  1338.         

  1339.         dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 

  1340.                                       picture, s->p_mv_table, mv_scale, mv_penalty);

  1341.  

  1342.         break;

  1343.     }

  1344.     

  1345.     dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,

  1346. 				   pred_x, pred_y, picture, 0, 0, mv_penalty);

  1347.                                    

  1348.     if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)

  1349.         dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);

  1350. 

  1351. //printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);

  1352. //    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;

  1353.     mv_table[mot_xy][0]= mx;

  1354.     mv_table[mot_xy][1]= my;

  1355. 

  1356.     return dmin;

  1357. }

  1358. 

  1359. static inline int check_bidir_mv(MpegEncContext * s,

  1360.                    int mb_x, int mb_y,

  1361.                    int motion_fx, int motion_fy,

  1362.                    int motion_bx, int motion_by,

  1363.                    int pred_fx, int pred_fy,

  1364.                    int pred_bx, int pred_by)

  1365. {

  1366.     //FIXME optimize?

  1367.     //FIXME move into template?

  1368.     //FIXME better f_code prediction (max mv & distance)

  1369.     uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame

  1370.     uint8_t *dest_y = s->me.scratchpad;

  1371.     uint8_t *ptr;

  1372.     int dxy;

  1373.     int src_x, src_y;

  1374.     int fbmin;

  1375. 

  1376.     if(s->quarter_sample){

  1377.         dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);

  1378.         src_x = mb_x * 16 + (motion_fx >> 2);

  1379.         src_y = mb_y * 16 + (motion_fy >> 2);

  1380.         assert(src_x >=-16 && src_x<=s->h_edge_pos);

  1381.         assert(src_y >=-16 && src_y<=s->v_edge_pos);

  1382. 

  1383.         ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;

  1384.         s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);

  1385. 

  1386.         dxy = ((motion_by & 3) << 2) | (motion_bx & 3);

  1387.         src_x = mb_x * 16 + (motion_bx >> 2);

  1388.         src_y = mb_y * 16 + (motion_by >> 2);

  1389.         assert(src_x >=-16 && src_x<=s->h_edge_pos);

  1390.         assert(src_y >=-16 && src_y<=s->v_edge_pos);

  1391.     

  1392.         ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;

  1393.         s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);

  1394.     }else{

  1395.         dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);

  1396.         src_x = mb_x * 16 + (motion_fx >> 1);

  1397.         src_y = mb_y * 16 + (motion_fy >> 1);

  1398.         assert(src_x >=-16 && src_x<=s->h_edge_pos);

  1399.         assert(src_y >=-16 && src_y<=s->v_edge_pos);

  1400. 

  1401.         ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;

  1402.         s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

  1403. 

  1404.         dxy = ((motion_by & 1) << 1) | (motion_bx & 1);

  1405.         src_x = mb_x * 16 + (motion_bx >> 1);

  1406.         src_y = mb_y * 16 + (motion_by >> 1);

  1407.         assert(src_x >=-16 && src_x<=s->h_edge_pos);

  1408.         assert(src_y >=-16 && src_y<=s->v_edge_pos);

  1409.     

  1410.         ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;

  1411.         s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);

  1412.     }

  1413. 

  1414.     fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor

  1415.            +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor

  1416.            + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);

  1417.            

  1418.     if(s->avctx->mb_cmp&FF_CMP_CHROMA){

  1419.     }

  1420.     //FIXME CHROMA !!!

  1421.            

  1422.     return fbmin;

  1423. }

  1424. 

  1425. /* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/

  1426. static inline int bidir_refine(MpegEncContext * s,

  1427.                                   int mb_x, int mb_y)

  1428. {

  1429.     const int mot_stride = s->mb_stride;

  1430.     const int xy = mb_y *mot_stride + mb_x;

  1431.     int fbmin;

  1432.     int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];

  1433.     int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];

  1434.     int pred_bx= s->b_bidir_back_mv_table[xy-1][0];

  1435.     int pred_by= s->b_bidir_back_mv_table[xy-1][1];

  1436.     int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];

  1437.     int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];

  1438.     int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];

  1439.     int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];

  1440. 

  1441.     //FIXME do refinement and add flag

  1442.     

  1443.     fbmin= check_bidir_mv(s, mb_x, mb_y, 

  1444.                           motion_fx, motion_fy,

  1445.                           motion_bx, motion_by,

  1446.                           pred_fx, pred_fy,

  1447.                           pred_bx, pred_by);

  1448. 

  1449.    return fbmin;

  1450. }

  1451. 

  1452. static inline int direct_search(MpegEncContext * s,

  1453.                                 int mb_x, int mb_y)

  1454. {

  1455.     int P[10][2];

  1456.     const int mot_stride = s->mb_stride;

  1457.     const int mot_xy = mb_y*mot_stride + mb_x;

  1458.     const int shift= 1+s->quarter_sample;

  1459.     int dmin, i;

  1460.     const int time_pp= s->pp_time;

  1461.     const int time_pb= s->pb_time;

  1462.     int mx, my, xmin, xmax, ymin, ymax;

  1463.     int16_t (*mv_table)[2]= s->b_direct_mv_table;

  1464.     uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;

  1465.     

  1466.     ymin= xmin=(-32)>>shift;

  1467.     ymax= xmax=   31>>shift;

  1468. 

  1469.     if(IS_8X8(s->next_picture.mb_type[mot_xy])){

  1470.         s->mv_type= MV_TYPE_8X8;

  1471.     }else{

  1472.         s->mv_type= MV_TYPE_16X16;

  1473.     }

  1474. 

  1475.     for(i=0; i<4; i++){

  1476.         int index= s->block_index[i];

  1477.         int min, max;

  1478.     

  1479.         s->me.co_located_mv[i][0]= s->motion_val[index][0];

  1480.         s->me.co_located_mv[i][1]= s->motion_val[index][1];

  1481.         s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));

  1482.         s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));

  1483. //        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);

  1484. //        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);

  1485. 

  1486.         max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;

  1487.         min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;

  1488.         max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding

  1489.         min+= (2*mb_x + (i& 1))*8 - 1;

  1490.         xmax= FFMIN(xmax, s->width - max);

  1491.         xmin= FFMAX(xmin, - 16     - min);

  1492. 

  1493.         max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;

  1494.         min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;

  1495.         max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding

  1496.         min+= (2*mb_y + (i>>1))*8 - 1;

  1497.         ymax= FFMIN(ymax, s->height - max);

  1498.         ymin= FFMAX(ymin, - 16      - min);

  1499.         

  1500.         if(s->mv_type == MV_TYPE_16X16) break;

  1501.     }

  1502.     

  1503.     assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);

  1504.     

  1505.     if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){

  1506.         s->b_direct_mv_table[mot_xy][0]= 0;

  1507.         s->b_direct_mv_table[mot_xy][1]= 0;

  1508. 

  1509.         return 256*256*256*64;

  1510.     }

  1511. 

  1512.     P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);

  1513.     P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);

  1514. 

  1515.     /* special case for first line */

  1516.     if (mb_y) {

  1517.         P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);

  1518.         P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);

  1519.         P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);

  1520.         P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);

  1521.     

  1522.         P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);

  1523.         P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);

  1524.     }

  1525.  

  1526.     //FIXME direct_search  ptr in context!!! (needed for chroma anyway or this will get messy)   

  1527.     if(s->flags&CODEC_FLAG_QPEL){

  1528.         dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 

  1529.                                                      &s->last_picture, mv_table, 1<<14, mv_penalty);

  1530.         dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,

  1531.                                                 0, 0, &s->last_picture, 0, 0, mv_penalty);

  1532.         

  1533.         if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)

  1534.             dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);

  1535.     }else{

  1536.         dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 

  1537.                                                      &s->last_picture, mv_table, 1<<15, mv_penalty);

  1538.         dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,

  1539.                                                 0, 0, &s->last_picture, 0, 0, mv_penalty);

  1540. 

  1541.         if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)

  1542.             dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);

  1543.     }

  1544. 

  1545.     s->b_direct_mv_table[mot_xy][0]= mx;

  1546.     s->b_direct_mv_table[mot_xy][1]= my;

  1547.     return dmin;

  1548. }

  1549. 

  1550. void ff_estimate_b_frame_motion(MpegEncContext * s,

  1551.                              int mb_x, int mb_y)

  1552. {

  1553.     const int penalty_factor= s->me.mb_penalty_factor;

  1554.     int fmin, bmin, dmin, fbmin;

  1555.     int type=0;

  1556.     

  1557.     s->me.skip=0;

  1558.     if (s->codec_id == CODEC_ID_MPEG4)

  1559.         dmin= direct_search(s, mb_x, mb_y);

  1560.     else

  1561.         dmin= INT_MAX;

  1562. 

  1563.     s->me.skip=0;

  1564.     fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;

  1565.     

  1566.     s->me.skip=0;

  1567.     bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;

  1568. //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

  1569. 

  1570.     s->me.skip=0;

  1571.     fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;

  1572. //printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);

  1573.     {

  1574.         int score= fmin;

  1575.         type = MB_TYPE_FORWARD;

  1576.         

  1577.         if (dmin <= score){

  1578.             score = dmin;

  1579.             type = MB_TYPE_DIRECT;

  1580.         }

  1581.         if(bmin<score){

  1582.             score=bmin;

  1583.             type= MB_TYPE_BACKWARD; 

  1584.         }

  1585.         if(fbmin<score){

  1586.             score=fbmin;

  1587.             type= MB_TYPE_BIDIR;

  1588.         }

  1589.         

  1590.         score= ((unsigned)(score*score + 128*256))>>16;

  1591.         s->current_picture.mc_mb_var_sum += score;

  1592.         s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE

  1593.     }

  1594. 

  1595.     if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){

  1596.         type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter

  1597.         if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB

  1598.     }

  1599. 

  1600.     s->mb_type[mb_y*s->mb_stride + mb_x]= type;

  1601. }

  1602. 

  1603. /* find best f_code for ME which do unlimited searches */

  1604. int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)

  1605. {

  1606.     if(s->me_method>=ME_EPZS){

  1607.         int score[8];

  1608.         int i, y;

  1609.         uint8_t * fcode_tab= s->fcode_tab;

  1610.         int best_fcode=-1;

  1611.         int best_score=-10000000;

  1612. 

  1613.         for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);

  1614. 

  1615.         for(y=0; y<s->mb_height; y++){

  1616.             int x;

  1617.             int xy= y*s->mb_stride;

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

  1619.                 if(s->mb_type[xy] & type){

  1620.                     int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],

  1621.                                      fcode_tab[mv_table[xy][1] + MAX_MV]);

  1622.                     int j;

  1623.                     

  1624.                     for(j=0; j<fcode && j<8; j++){

  1625.                         if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])

  1626.                             score[j]-= 170;

  1627.                     }

  1628.                 }

  1629.                 xy++;

  1630.             }

  1631.         }

  1632.         

  1633.         for(i=1; i<8; i++){

  1634.             if(score[i] > best_score){

  1635.                 best_score= score[i];

  1636.                 best_fcode= i;

  1637.             }

  1638. //            printf("%d %d\n", i, score[i]);

  1639.         }

  1640. 

  1641. //    printf("fcode: %d type: %d\n", i, s->pict_type);

  1642.         return best_fcode;

  1643. /*        for(i=0; i<=MAX_FCODE; i++){

  1644.             printf("%d ", mv_num[i]);

  1645.         }

  1646.         printf("\n");*/

  1647.     }else{

  1648.         return 1;

  1649.     }

  1650. }

  1651. 

  1652. void ff_fix_long_p_mvs(MpegEncContext * s)

  1653. {

  1654.     const int f_code= s->f_code;

  1655.     int y, range;

  1656.     assert(s->pict_type==P_TYPE);

  1657. 

  1658.     range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);

  1659.     

  1660.     if(s->msmpeg4_version) range= 16;

  1661.     

  1662.     if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;

  1663.     

  1664.     /* clip / convert to intra 16x16 type MVs */

  1665.     for(y=0; y<s->mb_height; y++){

  1666.         int x;

  1667.         int xy= y*s->mb_stride;

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

  1669.             if(s->mb_type[xy]&MB_TYPE_INTER){

  1670.                 if(   s->p_mv_table[xy][0] >=range || s->p_mv_table[xy][0] <-range

  1671.                    || s->p_mv_table[xy][1] >=range || s->p_mv_table[xy][1] <-range){

  1672.                     s->mb_type[xy] &= ~MB_TYPE_INTER;

  1673.                     s->mb_type[xy] |= MB_TYPE_INTRA;

  1674.                     s->current_picture.mb_type[xy]= MB_TYPE_INTRA;

  1675.                     s->p_mv_table[xy][0] = 0;

  1676.                     s->p_mv_table[xy][1] = 0;

  1677.                 }

  1678.             }

  1679.             xy++;

  1680.         }

  1681.     }

  1682. //printf("%d no:%d %d//\n", clip, noclip, f_code);

  1683.     if(s->flags&CODEC_FLAG_4MV){

  1684.         const int wrap= 2+ s->mb_width*2;

  1685. 

  1686.         /* clip / convert to intra 8x8 type MVs */

  1687.         for(y=0; y<s->mb_height; y++){

  1688.             int xy= (y*2 + 1)*wrap + 1;

  1689.             int i= y*s->mb_stride;

  1690.             int x;

  1691. 

  1692.             for(x=0; x<s->mb_width; x++){

  1693.                 if(s->mb_type[i]&MB_TYPE_INTER4V){

  1694.                     int block;

  1695.                     for(block=0; block<4; block++){

  1696.                         int off= (block& 1) + (block>>1)*wrap;

  1697.                         int mx= s->motion_val[ xy + off ][0];

  1698.                         int my= s->motion_val[ xy + off ][1];

  1699. 

  1700.                         if(   mx >=range || mx <-range

  1701.                            || my >=range || my <-range){

  1702.                             s->mb_type[i] &= ~MB_TYPE_INTER4V;

  1703.                             s->mb_type[i] |= MB_TYPE_INTRA;

  1704.                             s->current_picture.mb_type[i]= MB_TYPE_INTRA;

  1705.                         }

  1706.                     }

  1707.                 }

  1708.                 xy+=2;

  1709.                 i++;

  1710.             }

  1711.         }

  1712.     }

  1713. }

  1714. 

  1715. void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)

  1716. {

  1717.     int y;

  1718. 

  1719.     // RAL: 8 in MPEG-1, 16 in MPEG-4

  1720.     int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);

  1721.     

  1722.     if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;

  1723. 

  1724.     /* clip / convert to intra 16x16 type MVs */

  1725.     for(y=0; y<s->mb_height; y++){

  1726.         int x;

  1727.         int xy= y*s->mb_stride;

  1728.         for(x=0; x<s->mb_width; x++){

  1729.             if (s->mb_type[xy] & type){    // RAL: "type" test added...

  1730.                 if(   mv_table[xy][0] >=range || mv_table[xy][0] <-range

  1731.                    || mv_table[xy][1] >=range || mv_table[xy][1] <-range){

  1732. 

  1733.                     if(s->codec_id == CODEC_ID_MPEG1VIDEO && 0){

  1734.                     }else{

  1735.                         if     (mv_table[xy][0] > range-1) mv_table[xy][0]=  range-1;

  1736.                         else if(mv_table[xy][0] < -range ) mv_table[xy][0]= -range;

  1737.                         if     (mv_table[xy][1] > range-1) mv_table[xy][1]=  range-1;

  1738.                         else if(mv_table[xy][1] < -range ) mv_table[xy][1]= -range;

  1739.                     }

  1740.                 }

  1741.             }

  1742.             xy++;

  1743.         }

  1744.     }

  1745. }