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.
554 Commits
32 Branches
325MB
Tree: 9dbcbd92e8
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '9dbcbd92e8'
${ noResults }
FFmpeg/libavcodec/motion_est.c

1290 lines
37KB
Raw Blame History 

  1. /*

  2.  * Motion estimation 

  3.  * Copyright (c) 2000,2001 Gerard Lantau.

  4.  * 

  5.  *

  6.  * This program is free software; you can redistribute it and/or modify

  7.  * it under the terms of the GNU General Public License as published by

  8.  * the Free Software Foundation; either version 2 of the License, or

  9.  * (at your option) any later version.

  10.  *

  11.  * This program is distributed in the hope that it will be useful,

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

  13.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the

  14.  * GNU General Public License for more details.

  15.  *

  16.  * You should have received a copy of the GNU General Public License

  17.  * along with this program; if not, write to the Free Software

  18.  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.

  19.  *

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

  21.  */

  22. #include <stdlib.h>

  23. #include <stdio.h>

  24. #include "avcodec.h"

  25. #include "dsputil.h"

  26. #include "mpegvideo.h"

  27. 

  28. #define ABS(a) ((a)>0 ? (a) : -(a))

  29. #define MAX(a,b) ((a) > (b) ? (a) : (b))

  30. #define INTER_BIAS	257

  31. 

  32. static void halfpel_motion_search(MpegEncContext * s,

  33. 				  int *mx_ptr, int *my_ptr, int dmin,

  34. 				  int xmin, int ymin, int xmax, int ymax,

  35.                                   int pred_x, int pred_y, uint8_t *ref_picture);

  36. 

  37. static int pix_sum(UINT8 * pix, int line_size)

  38. {

  39.     int s, i, j;

  40. 

  41.     s = 0;

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

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

  44. 	    s += pix[0];

  45. 	    s += pix[1];

  46. 	    s += pix[2];

  47. 	    s += pix[3];

  48. 	    s += pix[4];

  49. 	    s += pix[5];

  50. 	    s += pix[6];

  51. 	    s += pix[7];

  52. 	    pix += 8;

  53. 	}

  54. 	pix += line_size - 16;

  55.     }

  56.     return s;

  57. }

  58. 

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

  60. {

  61.     int s, i, j;

  62. 

  63.     s = 0;

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

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

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

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

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

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

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

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

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

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

  74. 	    pix += 8;

  75. 	}

  76. 	pix += line_size - 16;

  77.     }

  78.     return s;

  79. }

  80. 

  81. static int pix_norm1(UINT8 * pix, int line_size)

  82. {

  83.     int s, i, j;

  84.     UINT32 *sq = squareTbl + 256;

  85. 

  86.     s = 0;

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

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

  89. 	    s += sq[pix[0]];

  90. 	    s += sq[pix[1]];

  91. 	    s += sq[pix[2]];

  92. 	    s += sq[pix[3]];

  93. 	    s += sq[pix[4]];

  94. 	    s += sq[pix[5]];

  95. 	    s += sq[pix[6]];

  96. 	    s += sq[pix[7]];

  97. 	    pix += 8;

  98. 	}

  99. 	pix += line_size - 16;

  100.     }

  101.     return s;

  102. }

  103. 

  104. static int pix_norm(UINT8 * pix1, UINT8 * pix2, int line_size)

  105. {

  106.     int s, i, j;

  107.     UINT32 *sq = squareTbl + 256;

  108. 

  109.     s = 0;

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

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

  112. 	    s += sq[pix1[0] - pix2[0]];

  113. 	    s += sq[pix1[1] - pix2[1]];

  114. 	    s += sq[pix1[2] - pix2[2]];

  115. 	    s += sq[pix1[3] - pix2[3]];

  116. 	    s += sq[pix1[4] - pix2[4]];

  117. 	    s += sq[pix1[5] - pix2[5]];

  118. 	    s += sq[pix1[6] - pix2[6]];

  119. 	    s += sq[pix1[7] - pix2[7]];

  120. 	    pix1 += 8;

  121. 	    pix2 += 8;

  122. 	}

  123. 	pix1 += line_size - 16;

  124. 	pix2 += line_size - 16;

  125.     }

  126.     return s;

  127. }

  128. 

  129. static void no_motion_search(MpegEncContext * s,

  130. 			     int *mx_ptr, int *my_ptr)

  131. {

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

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

  134. }

  135. 

  136. static int full_motion_search(MpegEncContext * s,

  137.                               int *mx_ptr, int *my_ptr, int range,

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

  139. {

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

  141.     int mx, my, dmin, d;

  142.     UINT8 *pix;

  143. 

  144.     xx = 16 * s->mb_x;

  145.     yy = 16 * s->mb_y;

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

  147.     if (x1 < xmin)

  148. 	x1 = xmin;

  149.     x2 = xx + range - 1;

  150.     if (x2 > xmax)

  151. 	x2 = xmax;

  152.     y1 = yy - range + 1;

  153.     if (y1 < ymin)

  154. 	y1 = ymin;

  155.     y2 = yy + range - 1;

  156.     if (y2 > ymax)

  157. 	y2 = ymax;

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

  159.     dmin = 0x7fffffff;

  160.     mx = 0;

  161.     my = 0;

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

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

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

  165. 			     s->linesize);

  166. 	    if (d < dmin ||

  167. 		(d == dmin &&

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

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

  170. 		dmin = d;

  171. 		mx = x;

  172. 		my = y;

  173. 	    }

  174. 	}

  175.     }

  176. 

  177.     *mx_ptr = mx;

  178.     *my_ptr = my;

  179. 

  180. #if 0

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

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

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

  184.     }

  185. #endif

  186.     return dmin;

  187. }

  188. 

  189. 

  190. static int log_motion_search(MpegEncContext * s,

  191.                              int *mx_ptr, int *my_ptr, int range,

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

  193. {

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

  195.     int mx, my, dmin, d;

  196.     UINT8 *pix;

  197. 

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

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

  200. 

  201.     /* Left limit */

  202.     x1 = xx - range;

  203.     if (x1 < xmin)

  204. 	x1 = xmin;

  205. 

  206.     /* Right limit */

  207.     x2 = xx + range;

  208.     if (x2 > xmax)

  209. 	x2 = xmax;

  210. 

  211.     /* Upper limit */

  212.     y1 = yy - range;

  213.     if (y1 < ymin)

  214. 	y1 = ymin;

  215. 

  216.     /* Lower limit */

  217.     y2 = yy + range;

  218.     if (y2 > ymax)

  219. 	y2 = ymax;

  220. 

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

  222.     dmin = 0x7fffffff;

  223.     mx = 0;

  224.     my = 0;

  225. 

  226.     do {

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

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

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

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

  231. 		    dmin = d;

  232. 		    mx = x;

  233. 		    my = y;

  234. 		}

  235. 	    }

  236. 	}

  237. 

  238. 	range = range >> 1;

  239. 

  240. 	x1 = mx - range;

  241. 	if (x1 < xmin)

  242. 	    x1 = xmin;

  243. 

  244. 	x2 = mx + range;

  245. 	if (x2 > xmax)

  246. 	    x2 = xmax;

  247. 

  248. 	y1 = my - range;

  249. 	if (y1 < ymin)

  250. 	    y1 = ymin;

  251. 

  252. 	y2 = my + range;

  253. 	if (y2 > ymax)

  254. 	    y2 = ymax;

  255. 

  256.     } while (range >= 1);

  257. 

  258. #ifdef DEBUG

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

  260. #endif

  261.     *mx_ptr = mx;

  262.     *my_ptr = my;

  263.     return dmin;

  264. }

  265. 

  266. static int phods_motion_search(MpegEncContext * s,

  267.                                int *mx_ptr, int *my_ptr, int range,

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

  269. {

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

  271.     int mx, my, dminx, dminy;

  272.     UINT8 *pix;

  273. 

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

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

  276. 

  277.     /* Left limit */

  278.     x1 = xx - range;

  279.     if (x1 < xmin)

  280. 	x1 = xmin;

  281. 

  282.     /* Right limit */

  283.     x2 = xx + range;

  284.     if (x2 > xmax)

  285. 	x2 = xmax;

  286. 

  287.     /* Upper limit */

  288.     y1 = yy - range;

  289.     if (y1 < ymin)

  290. 	y1 = ymin;

  291. 

  292.     /* Lower limit */

  293.     y2 = yy + range;

  294.     if (y2 > ymax)

  295. 	y2 = ymax;

  296. 

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

  298.     mx = 0;

  299.     my = 0;

  300. 

  301.     x = xx;

  302.     y = yy;

  303.     do {

  304.         dminx = 0x7fffffff;

  305.         dminy = 0x7fffffff;

  306. 

  307. 	lastx = x;

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

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

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

  311. 		dminx = d;

  312. 		mx = x;

  313. 	    }

  314. 	}

  315. 

  316. 	x = lastx;

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

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

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

  320. 		dminy = d;

  321. 		my = y;

  322. 	    }

  323. 	}

  324. 

  325. 	range = range >> 1;

  326. 

  327. 	x = mx;

  328. 	y = my;

  329. 	x1 = mx - range;

  330. 	if (x1 < xmin)

  331. 	    x1 = xmin;

  332. 

  333. 	x2 = mx + range;

  334. 	if (x2 > xmax)

  335. 	    x2 = xmax;

  336. 

  337. 	y1 = my - range;

  338. 	if (y1 < ymin)

  339. 	    y1 = ymin;

  340. 

  341. 	y2 = my + range;

  342. 	if (y2 > ymax)

  343. 	    y2 = ymax;

  344. 

  345.     } while (range >= 1);

  346. 

  347. #ifdef DEBUG

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

  349. #endif

  350. 

  351.     /* half pixel search */

  352.     *mx_ptr = mx;

  353.     *my_ptr = my;

  354.     return dminy;

  355. }

  356. 

  357. 

  358. #define Z_THRESHOLD 256

  359. 

  360. #define CHECK_MV(x,y)\

  361. {\

  362.     d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\

  363.     d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\

  364.     if(d<dmin){\

  365.         best[0]=x;\

  366.         best[1]=y;\

  367.         dmin=d;\

  368.     }\

  369. }

  370. 

  371. #define CHECK_MV_DIR(x,y,new_dir)\

  372. {\

  373.     d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\

  374.     d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\

  375.     if(d<dmin){\

  376.         best[0]=x;\

  377.         best[1]=y;\

  378.         dmin=d;\

  379.         next_dir= new_dir;\

  380.     }\

  381. }

  382. 

  383. #define CHECK_MV4(x,y)\

  384. {\

  385.     d = pix_abs8x8(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\

  386.     d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\

  387.     if(d<dmin){\

  388.         best[0]=x;\

  389.         best[1]=y;\

  390.         dmin=d;\

  391.     }\

  392. }

  393. 

  394. #define CHECK_MV4_DIR(x,y,new_dir)\

  395. {\

  396.     d = pix_abs8x8(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);\

  397.     d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;\

  398.     if(d<dmin){\

  399.         best[0]=x;\

  400.         best[1]=y;\

  401.         dmin=d;\

  402.         next_dir= new_dir;\

  403.     }\

  404. }

  405. 

  406. 

  407. #define check(x,y,S,v)\

  408. if( (x)<(xmin<<(S)) ) printf("%d %d %d %d xmin" #v, (x), (y), s->mb_x, s->mb_y);\

  409. if( (x)>(xmax<<(S)) ) printf("%d %d %d %d xmax" #v, (x), (y), s->mb_x, s->mb_y);\

  410. if( (y)<(ymin<<(S)) ) printf("%d %d %d %d ymin" #v, (x), (y), s->mb_x, s->mb_y);\

  411. if( (y)>(ymax<<(S)) ) printf("%d %d %d %d ymax" #v, (x), (y), s->mb_x, s->mb_y);\

  412. 

  413. 

  414. static inline int small_diamond_search(MpegEncContext * s, int *best, int dmin,

  415.                                        UINT8 *new_pic, UINT8 *old_pic, int pic_stride,

  416.                                        int pred_x, int pred_y, UINT16 *mv_penalty, int quant,

  417.                                        int xmin, int ymin, int xmax, int ymax, int shift)

  418. {

  419.     int next_dir=-1;

  420. 

  421.     for(;;){

  422.         int d;

  423.         const int dir= next_dir;

  424.         const int x= best[0];

  425.         const int y= best[1];

  426.         next_dir=-1;

  427. 

  428. //printf("%d", dir);

  429.         if(dir!=2 && x>xmin) CHECK_MV_DIR(x-1, y  , 0)

  430.         if(dir!=3 && y>ymin) CHECK_MV_DIR(x  , y-1, 1)

  431.         if(dir!=0 && x<xmax) CHECK_MV_DIR(x+1, y  , 2)

  432.         if(dir!=1 && y<ymax) CHECK_MV_DIR(x  , y+1, 3)

  433. 

  434.         if(next_dir==-1){

  435.             return dmin;

  436.         }

  437.     }

  438. 

  439. /*    for(;;){

  440.         int d;

  441.         const int x= best[0];

  442.         const int y= best[1];

  443.         const int last_min=dmin;

  444.         if(x>xmin) CHECK_MV(x-1, y  )

  445.         if(y>xmin) CHECK_MV(x  , y-1)

  446.         if(x<xmax) CHECK_MV(x+1, y  )

  447.         if(y<xmax) CHECK_MV(x  , y+1)

  448.         if(x>xmin && y>ymin) CHECK_MV(x-1, y-1)

  449.         if(x>xmin && y<ymax) CHECK_MV(x-1, y+1)

  450.         if(x<xmax && y>ymin) CHECK_MV(x+1, y-1)

  451.         if(x<xmax && y<ymax) CHECK_MV(x+1, y+1)

  452.         if(x-1>xmin) CHECK_MV(x-2, y  )

  453.         if(y-1>xmin) CHECK_MV(x  , y-2)

  454.         if(x+1<xmax) CHECK_MV(x+2, y  )

  455.         if(y+1<xmax) CHECK_MV(x  , y+2)

  456.         if(x-1>xmin && y-1>ymin) CHECK_MV(x-2, y-2)

  457.         if(x-1>xmin && y+1<ymax) CHECK_MV(x-2, y+2)

  458.         if(x+1<xmax && y-1>ymin) CHECK_MV(x+2, y-2)

  459.         if(x+1<xmax && y+1<ymax) CHECK_MV(x+2, y+2)

  460.         if(dmin==last_min) return dmin;

  461.     }

  462.     */

  463. }

  464. 

  465. static inline int small_diamond_search4MV(MpegEncContext * s, int *best, int dmin,

  466.                                        UINT8 *new_pic, UINT8 *old_pic, int pic_stride,

  467.                                        int pred_x, int pred_y, UINT16 *mv_penalty, int quant,

  468.                                        int xmin, int ymin, int xmax, int ymax, int shift)

  469. {

  470.     int next_dir=-1;

  471. 

  472.     for(;;){

  473.         int d;

  474.         const int dir= next_dir;

  475.         const int x= best[0];

  476.         const int y= best[1];

  477.         next_dir=-1;

  478. 

  479. //printf("%d", dir);

  480.         if(dir!=2 && x>xmin) CHECK_MV4_DIR(x-1, y  , 0)

  481.         if(dir!=3 && y>ymin) CHECK_MV4_DIR(x  , y-1, 1)

  482.         if(dir!=0 && x<xmax) CHECK_MV4_DIR(x+1, y  , 2)

  483.         if(dir!=1 && y<ymax) CHECK_MV4_DIR(x  , y+1, 3)

  484. 

  485.         if(next_dir==-1){

  486.             return dmin;

  487.         }

  488.     }

  489. }

  490. 

  491. static inline int snake_search(MpegEncContext * s, int *best, int dmin,

  492.                                        UINT8 *new_pic, UINT8 *old_pic, int pic_stride,

  493.                                        int pred_x, int pred_y, UINT16 *mv_penalty, int quant,

  494.                                        int xmin, int ymin, int xmax, int ymax, int shift)

  495. {

  496.     int dir=0;

  497.     int c=1;

  498.     static int x_dir[8]= {1,1,0,-1,-1,-1, 0, 1};

  499.     static int y_dir[8]= {0,1,1, 1, 0,-1,-1,-1};

  500.     int fails=0;

  501.     int last_d[2]={dmin, dmin};

  502. 

  503. /*static int good=0;

  504. static int bad=0;

  505. static int point=0;

  506. 

  507. point++;

  508. if(256*256*256*64%point==0)

  509. {

  510.     printf("%d %d %d\n", good, bad, point);

  511. }*/

  512. 

  513.     for(;;){

  514.         int x= best[0];

  515.         int y= best[1];

  516.         int d;

  517.         x+=x_dir[dir];

  518.         y+=y_dir[dir];

  519.         if(x>=xmin && x<=xmax && y>=ymin && y<=ymax){

  520.             d = pix_abs16x16(new_pic, old_pic + (x) + (y)*pic_stride, pic_stride);

  521.             d += (mv_penalty[((x)<<shift)-pred_x] + mv_penalty[((y)<<shift)-pred_y])*quant;

  522.         }else{

  523.             d = dmin + 10000; //FIXME smarter boundary handling

  524.         }

  525.         if(d<dmin){

  526.             best[0]=x;

  527.             best[1]=y;

  528.             dmin=d;

  529. 

  530.             if(last_d[1] - last_d[0] > last_d[0] - d) c= -c;

  531.             dir+=c;

  532. 

  533.             fails=0;

  534. //good++;

  535.             last_d[1]=last_d[0];

  536.             last_d[0]=d;

  537.         }else{

  538. //bad++;

  539.             if(fails){

  540.                 if(fails>=3) return dmin;

  541.             }else{

  542.                 c= -c;

  543.             }

  544.             dir+=c*2;

  545.             fails++;

  546.         }

  547.         dir&=7;

  548.     }

  549. }

  550. 

  551. static int epzs_motion_search(MpegEncContext * s,

  552.                              int *mx_ptr, int *my_ptr,

  553.                              int P[5][2], int pred_x, int pred_y,

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

  555. {

  556.     int best[2]={0, 0};

  557.     int d, dmin; 

  558.     UINT8 *new_pic, *old_pic;

  559.     const int pic_stride= s->linesize;

  560.     const int pic_xy= (s->mb_y*pic_stride + s->mb_x)*16;

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

  562.     int quant= s->qscale; // qscale of the prev frame

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

  564. 

  565.     new_pic = s->new_picture[0] + pic_xy;

  566.     old_pic = ref_picture + pic_xy;

  567.    

  568.     dmin = pix_abs16x16(new_pic, old_pic, pic_stride);

  569.     if(dmin<Z_THRESHOLD){

  570.         *mx_ptr= 0;

  571.         *my_ptr= 0;

  572. //printf("Z");

  573.         return dmin;

  574.     }

  575. 

  576.     /* first line */

  577.     if ((s->mb_y == 0 || s->first_slice_line || s->first_gob_line)) {

  578.         CHECK_MV(P[1][0]>>shift, P[1][1]>>shift)

  579.     }else{

  580.         CHECK_MV(P[4][0]>>shift, P[4][1]>>shift)

  581.         if(dmin<Z_THRESHOLD){

  582.             *mx_ptr= P[4][0]>>shift;

  583.             *my_ptr= P[4][1]>>shift;

  584. //printf("M\n");

  585.             return dmin;

  586.         }

  587.         CHECK_MV(P[1][0]>>shift, P[1][1]>>shift)

  588.         CHECK_MV(P[2][0]>>shift, P[2][1]>>shift)

  589.         CHECK_MV(P[3][0]>>shift, P[3][1]>>shift)

  590.     }

  591.     CHECK_MV(P[0][0]>>shift, P[0][1]>>shift)

  592. 

  593. //check(best[0],best[1],0, b0)

  594.     if(s->me_method==ME_EPZS)

  595.         dmin= small_diamond_search(s, best, dmin, new_pic, old_pic, pic_stride, 

  596.                                    pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, shift);

  597.     else

  598.         dmin=         snake_search(s, best, dmin, new_pic, old_pic, pic_stride, 

  599.                                    pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, shift);

  600. //check(best[0],best[1],0, b1)

  601.     *mx_ptr= best[0];

  602.     *my_ptr= best[1];    

  603. 

  604. //    printf("%d %d %d \n", best[0], best[1], dmin);

  605.     return dmin;

  606. }

  607. 

  608. static int epzs_motion_search4(MpegEncContext * s, int block,

  609.                              int *mx_ptr, int *my_ptr,

  610.                              int P[6][2], int pred_x, int pred_y,

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

  612. {

  613.     int best[2]={0, 0};

  614.     int d, dmin; 

  615.     UINT8 *new_pic, *old_pic;

  616.     const int pic_stride= s->linesize;

  617.     const int pic_xy= ((s->mb_y*2 + (block>>1))*pic_stride + s->mb_x*2 + (block&1))*8;

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

  619.     int quant= s->qscale; // qscale of the prev frame

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

  621. 

  622.     new_pic = s->new_picture[0] + pic_xy;

  623.     old_pic = ref_picture + pic_xy;

  624.    

  625.     dmin = pix_abs8x8(new_pic, old_pic, pic_stride);

  626. 

  627.     /* first line */

  628.     if ((s->mb_y == 0 || s->first_slice_line || s->first_gob_line) && block<2) {

  629.         CHECK_MV4(P[1][0]>>shift, P[1][1]>>shift)

  630.     }else{

  631.         CHECK_MV4(P[4][0]>>shift, P[4][1]>>shift)

  632.         if(dmin<Z_THRESHOLD){

  633.             *mx_ptr= P[4][0]>>shift;

  634.             *my_ptr= P[4][1]>>shift;

  635. //printf("M\n");

  636.             return dmin;

  637.         }

  638.         CHECK_MV4(P[1][0]>>shift, P[1][1]>>shift)

  639.         CHECK_MV4(P[2][0]>>shift, P[2][1]>>shift)

  640.         CHECK_MV4(P[3][0]>>shift, P[3][1]>>shift)

  641.     }

  642.     CHECK_MV4(P[0][0]>>shift, P[0][1]>>shift)

  643.     CHECK_MV4(P[5][0]>>shift, P[5][1]>>shift)

  644. 

  645. //check(best[0],best[1],0, b0)

  646.     dmin= small_diamond_search4MV(s, best, dmin, new_pic, old_pic, pic_stride, 

  647.                                    pred_x, pred_y, mv_penalty, quant, xmin, ymin, xmax, ymax, shift);

  648. //check(best[0],best[1],0, b1)

  649.     *mx_ptr= best[0];

  650.     *my_ptr= best[1];    

  651. 

  652. //    printf("%d %d %d \n", best[0], best[1], dmin);

  653.     return dmin;

  654. }

  655. 

  656. #define CHECK_HALF_MV(suffix, x, y) \

  657.     d= pix_abs16x16_ ## suffix(pix, ptr+((x)>>1), s->linesize);\

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

  659.     if(d<dminh){\

  660.         dminh= d;\

  661.         mx= mx1 + x;\

  662.         my= my1 + y;\

  663.     }

  664. 

  665. #define CHECK_HALF_MV4(suffix, x, y) \

  666.     d= pix_abs8x8_ ## suffix(pix, ptr+((x)>>1), s->linesize);\

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

  668.     if(d<dminh){\

  669.         dminh= d;\

  670.         mx= mx1 + x;\

  671.         my= my1 + y;\

  672.     }

  673.     

  674. /* The idea would be to make half pel ME after Inter/Intra decision to 

  675.    save time. */

  676. static inline void halfpel_motion_search(MpegEncContext * s,

  677. 				  int *mx_ptr, int *my_ptr, int dmin,

  678. 				  int xmin, int ymin, int xmax, int ymax,

  679.                                   int pred_x, int pred_y, uint8_t *ref_picture)

  680. {

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

  682.     const int quant= s->qscale;

  683.     int pen_x, pen_y;

  684.     int mx, my, mx1, my1, d, xx, yy, dminh;

  685.     UINT8 *pix, *ptr;

  686. 

  687.     mx = *mx_ptr;

  688.     my = *my_ptr;

  689.     ptr = ref_picture + (my * s->linesize) + mx;

  690. 

  691.     xx = 16 * s->mb_x;

  692.     yy = 16 * s->mb_y;

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

  694.     

  695.     dminh = dmin;

  696. 

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

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

  699. 

  700.         mx= mx1= 2*(mx - xx);

  701.         my= my1= 2*(my - yy);

  702.         if(dmin < Z_THRESHOLD && mx==0 && my==0){

  703.             *mx_ptr = 0;

  704.             *my_ptr = 0;

  705.             return;

  706.         }

  707.         

  708.         pen_x= pred_x + mx;

  709.         pen_y= pred_y + my;

  710. 

  711.         ptr-= s->linesize;

  712.         CHECK_HALF_MV(xy2, -1, -1)

  713.         CHECK_HALF_MV(y2 ,  0, -1)

  714.         CHECK_HALF_MV(xy2, +1, -1)

  715.         

  716.         ptr+= s->linesize;

  717.         CHECK_HALF_MV(x2 , -1,  0)

  718.         CHECK_HALF_MV(x2 , +1,  0)

  719.         CHECK_HALF_MV(xy2, -1, +1)

  720.         CHECK_HALF_MV(y2 ,  0, +1)

  721.         CHECK_HALF_MV(xy2, +1, +1)

  722. 

  723.     }else{

  724.         mx= 2*(mx - xx);

  725.         my= 2*(my - yy);

  726.     }

  727. 

  728.     *mx_ptr = mx;

  729.     *my_ptr = my;

  730. }

  731. 

  732. static inline void halfpel_motion_search4(MpegEncContext * s,

  733. 				  int *mx_ptr, int *my_ptr, int dmin,

  734. 				  int xmin, int ymin, int xmax, int ymax,

  735.                                   int pred_x, int pred_y, int block_x, int block_y,

  736.                                   uint8_t *ref_picture)

  737. {

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

  739.     const int quant= s->qscale;

  740.     int pen_x, pen_y;

  741.     int mx, my, mx1, my1, d, xx, yy, dminh;

  742.     UINT8 *pix, *ptr;

  743. 

  744.     xx = 8 * block_x;

  745.     yy = 8 * block_y;

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

  747.     

  748.     mx = *mx_ptr;

  749.     my = *my_ptr;

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

  751. 

  752.     dminh = dmin;

  753. 

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

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

  756. 

  757.         mx= mx1= 2*mx;

  758.         my= my1= 2*my;

  759.         if(dmin < Z_THRESHOLD && mx==0 && my==0){

  760.             *mx_ptr = 0;

  761.             *my_ptr = 0;

  762.             return;

  763.         }

  764.         

  765.         pen_x= pred_x + mx;

  766.         pen_y= pred_y + my;

  767. 

  768.         ptr-= s->linesize;

  769.         CHECK_HALF_MV4(xy2, -1, -1)

  770.         CHECK_HALF_MV4(y2 ,  0, -1)

  771.         CHECK_HALF_MV4(xy2, +1, -1)

  772.         

  773.         ptr+= s->linesize;

  774.         CHECK_HALF_MV4(x2 , -1,  0)

  775.         CHECK_HALF_MV4(x2 , +1,  0)

  776.         CHECK_HALF_MV4(xy2, -1, +1)

  777.         CHECK_HALF_MV4(y2 ,  0, +1)

  778.         CHECK_HALF_MV4(xy2, +1, +1)

  779. 

  780.     }else{

  781.         mx*=2;

  782.         my*=2;

  783.     }

  784. 

  785.     *mx_ptr = mx;

  786.     *my_ptr = my;

  787. }

  788. 

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

  790. {

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

  792.     

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

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

  795. 

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

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

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

  799. 

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

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

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

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

  804. 

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

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

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

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

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

  810.     }

  811. }

  812. 

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

  814. {

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

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

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

  818. 	*range *= 2;

  819. 

  820.     if (s->unrestricted_mv) {

  821.         *xmin = -16;

  822.         *ymin = -16;

  823.         if (s->h263_plus)

  824.             *range *= 2;

  825.         if(s->avctx==NULL || s->avctx->codec->id!=CODEC_ID_MPEG4){

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

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

  828.         }else {

  829.             /* XXX: dunno if this is correct but ffmpeg4 decoder wont like it otherwise 

  830. 	            (cuz the drawn edge isnt large enough))*/

  831.             *xmax = s->width;

  832.             *ymax = s->height;

  833.         }

  834.     } else {

  835.         *xmin = 0;

  836.         *ymin = 0;

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

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

  839.     }

  840. }

  841. 

  842. void ff_estimate_p_frame_motion(MpegEncContext * s,

  843.                                 int mb_x, int mb_y)

  844. {

  845.     UINT8 *pix, *ppix;

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

  847.     int xmin, ymin, xmax, ymax;

  848.     int rel_xmin, rel_ymin, rel_xmax, rel_ymax;

  849.     int pred_x=0, pred_y=0;

  850.     int P[6][2];

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

  852.     int mb_type=0;

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

  854. 

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

  856. 

  857.     switch(s->me_method) {

  858.     case ME_ZERO:

  859.     default:

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

  861.         dmin = 0;

  862.         break;

  863.     case ME_FULL:

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

  865.         break;

  866.     case ME_LOG:

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

  868.         break;

  869.     case ME_PHODS:

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

  871.         break;

  872.     case ME_X1:

  873.     case ME_EPZS:

  874.        {

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

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

  877. 

  878.             rel_xmin= xmin - mb_x*16;

  879.             rel_xmax= xmax - mb_x*16;

  880.             rel_ymin= ymin - mb_y*16;

  881.             rel_ymax= ymax - mb_y*16;

  882. 

  883.             P[0][0] = s->motion_val[mot_xy    ][0];

  884.             P[0][1] = s->motion_val[mot_xy    ][1];

  885.             P[1][0] = s->motion_val[mot_xy - 1][0];

  886.             P[1][1] = s->motion_val[mot_xy - 1][1];

  887.             if(P[1][0] > (rel_xmax<<shift)) P[1][0]= (rel_xmax<<shift);

  888. 

  889.             /* special case for first line */

  890.             if ((mb_y == 0 || s->first_slice_line || s->first_gob_line)) {

  891.                 P[4][0] = P[1][0];

  892.                 P[4][1] = P[1][1];

  893.             } else {

  894.                 P[2][0] = s->motion_val[mot_xy - mot_stride             ][0];

  895.                 P[2][1] = s->motion_val[mot_xy - mot_stride             ][1];

  896.                 P[3][0] = s->motion_val[mot_xy - mot_stride + 2         ][0];

  897.                 P[3][1] = s->motion_val[mot_xy - mot_stride + 2         ][1];

  898.                 if(P[2][1] > (rel_ymax<<shift)) P[2][1]= (rel_ymax<<shift);

  899.                 if(P[3][0] < (rel_xmin<<shift)) P[3][0]= (rel_xmin<<shift);

  900.                 if(P[3][1] > (rel_ymax<<shift)) P[3][1]= (rel_ymax<<shift);

  901.         

  902.                 P[4][0]= mid_pred(P[1][0], P[2][0], P[3][0]);

  903.                 P[4][1]= mid_pred(P[1][1], P[2][1], P[3][1]);

  904.             }

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

  906.                 pred_x = P[4][0];

  907.                 pred_y = P[4][1];

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

  909.                 pred_x= P[1][0];

  910.                 pred_y= P[1][1];

  911.             }

  912.         }

  913.         dmin = epzs_motion_search(s, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, ref_picture);

  914.  

  915.         mx+= mb_x*16;

  916.         my+= mb_y*16;

  917.         break;

  918.     }

  919.     

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

  921.         int block;

  922. 

  923.         mb_type|= MB_TYPE_INTER4V;

  924. 

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

  926.             int mx4, my4;

  927.             int pred_x4, pred_y4;

  928.             int dmin4;

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

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

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

  932.             const int block_x= mb_x*2 + (block&1);

  933.             const int block_y= mb_y*2 + (block>>1);

  934. 

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

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

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

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

  939. 

  940.             P[0][0] = s->motion_val[mot_xy    ][0];

  941.             P[0][1] = s->motion_val[mot_xy    ][1];

  942.             P[1][0] = s->motion_val[mot_xy - 1][0];

  943.             P[1][1] = s->motion_val[mot_xy - 1][1];

  944.             if(P[1][0] > (rel_xmax4<<shift)) P[1][0]= (rel_xmax4<<shift);

  945. 

  946.             /* special case for first line */

  947.             if ((mb_y == 0 || s->first_slice_line || s->first_gob_line) && block<2) {

  948.                 P[4][0] = P[1][0];

  949.                 P[4][1] = P[1][1];

  950.             } else {

  951.                 P[2][0] = s->motion_val[mot_xy - mot_stride             ][0];

  952.                 P[2][1] = s->motion_val[mot_xy - mot_stride             ][1];

  953.                 P[3][0] = s->motion_val[mot_xy - mot_stride + off[block]][0];

  954.                 P[3][1] = s->motion_val[mot_xy - mot_stride + off[block]][1];

  955.                 if(P[2][1] > (rel_ymax4<<shift)) P[2][1]= (rel_ymax4<<shift);

  956.                 if(P[3][0] < (rel_xmin4<<shift)) P[3][0]= (rel_xmin4<<shift);

  957.                 if(P[3][0] > (rel_xmax4<<shift)) P[3][0]= (rel_xmax4<<shift);

  958.                 if(P[3][1] > (rel_ymax4<<shift)) P[3][1]= (rel_ymax4<<shift);

  959.         

  960.                 P[4][0]= mid_pred(P[1][0], P[2][0], P[3][0]);

  961.                 P[4][1]= mid_pred(P[1][1], P[2][1], P[3][1]);

  962.             }

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

  964.                 pred_x4 = P[4][0];

  965.                 pred_y4 = P[4][1];

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

  967.                 pred_x4= P[1][0];

  968.                 pred_y4= P[1][1];

  969.             }

  970.             P[5][0]= mx - mb_x*16;

  971.             P[5][1]= my - mb_y*16;

  972. 

  973.             dmin4 = epzs_motion_search4(s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, ref_picture);

  974. 

  975.             halfpel_motion_search4(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 

  976.                                    pred_x4, pred_y4, block_x, block_y, ref_picture);

  977.      

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

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

  980.         }

  981.     }

  982. 

  983.     /* intra / predictive decision */

  984.     xx = mb_x * 16;

  985.     yy = mb_y * 16;

  986. 

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

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

  989.     ppix = ref_picture + (my * s->linesize) + mx;

  990.     

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

  992. #if 0

  993.     varc = pix_dev(pix, s->linesize, (sum+128)>>8) + INTER_BIAS;

  994.     vard = pix_abs16x16(pix, ppix, s->linesize);

  995. #else

  996.     sum= (sum+8)>>4;

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

  998.     vard = (pix_norm(pix, ppix, s->linesize)+128)>>8;

  999. #endif

  1000. 

  1001.     s->mb_var[s->mb_width * mb_y + mb_x] = varc;

  1002.     s->avg_mb_var+= varc;

  1003.     s->mc_mb_var += vard;

  1004. 

  1005.     

  1006. #if 0

  1007.     printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",

  1008. 	   varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);

  1009. #endif

  1010.     if(s->flags&CODEC_FLAG_HQ){

  1011.         if (vard*2 + 200 > varc)

  1012.             mb_type|= MB_TYPE_INTRA;

  1013.         if (varc*2 + 200 > vard){

  1014.             mb_type|= MB_TYPE_INTER;

  1015.             halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y, ref_picture);

  1016.         }else{

  1017.             mx = mx*2 - mb_x*32;

  1018.             my = my*2 - mb_y*32;

  1019.         }

  1020.     }else{

  1021.         if (vard <= 64 || vard < varc) {

  1022.             mb_type|= MB_TYPE_INTER;

  1023.             if (s->me_method != ME_ZERO) {

  1024.                 halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y, ref_picture);

  1025.             } else {

  1026.                 mx -= 16 * mb_x;

  1027.                 my -= 16 * mb_y;

  1028.             }

  1029. #if 0

  1030.             if (vard < 10) {

  1031.                 skip++;

  1032.                 fprintf(stderr,"\nEarly skip: %d vard: %2d varc: %5d dmin: %d", 

  1033.                                 skip, vard, varc, dmin);

  1034.             }

  1035. #endif

  1036.         }else{

  1037.             mb_type|= MB_TYPE_INTRA;

  1038.             mx = 0;//mx*2 - 32 * mb_x;

  1039.             my = 0;//my*2 - 32 * mb_y;

  1040.         }

  1041.     }

  1042. 

  1043.     s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;

  1044.     set_p_mv_tables(s, mx, my);

  1045. }

  1046. 

  1047. void ff_estimate_motion_b(MpegEncContext * s,

  1048.                        int mb_x, int mb_y, int16_t (*mv_table)[2], uint8_t *ref_picture, int f_code)

  1049. {

  1050.     UINT8 *pix, *ppix;

  1051.     int sum, varc, vard, mx, my, range, dmin, xx, yy;

  1052.     int xmin, ymin, xmax, ymax;

  1053.     int rel_xmin, rel_ymin, rel_xmax, rel_ymax;

  1054.     int pred_x=0, pred_y=0;

  1055.     int P[6][2];

  1056.     const int shift= 1+s->quarter_sample;

  1057.     int mb_type=0;

  1058.     const int mot_stride = s->mb_width + 2;

  1059.     const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;

  1060.     

  1061.     get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code);

  1062. 

  1063.     switch(s->me_method) {

  1064.     case ME_ZERO:

  1065.     default:

  1066. 	no_motion_search(s, &mx, &my);

  1067.         dmin = 0;

  1068.         break;

  1069.     case ME_FULL:

  1070. 	dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);

  1071.         break;

  1072.     case ME_LOG:

  1073. 	dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1074.         break;

  1075.     case ME_PHODS:

  1076. 	dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);

  1077.         break;

  1078.     case ME_X1:

  1079.     case ME_EPZS:

  1080.        {

  1081. 

  1082.             rel_xmin= xmin - mb_x*16;

  1083.             rel_xmax= xmax - mb_x*16;

  1084.             rel_ymin= ymin - mb_y*16;

  1085.             rel_ymax= ymax - mb_y*16;

  1086. 

  1087.             P[0][0] = mv_table[mot_xy    ][0];

  1088.             P[0][1] = mv_table[mot_xy    ][1];

  1089.             P[1][0] = mv_table[mot_xy - 1][0];

  1090.             P[1][1] = mv_table[mot_xy - 1][1];

  1091.             if(P[1][0] > (rel_xmax<<shift)) P[1][0]= (rel_xmax<<shift);

  1092. 

  1093.             /* special case for first line */

  1094.             if ((mb_y == 0 || s->first_slice_line || s->first_gob_line)) {

  1095.                 P[4][0] = P[1][0];

  1096.                 P[4][1] = P[1][1];

  1097.             } else {

  1098.                 P[2][0] = mv_table[mot_xy - mot_stride             ][0];

  1099.                 P[2][1] = mv_table[mot_xy - mot_stride             ][1];

  1100.                 P[3][0] = mv_table[mot_xy - mot_stride + 1         ][0];

  1101.                 P[3][1] = mv_table[mot_xy - mot_stride + 1         ][1];

  1102.                 if(P[2][1] > (rel_ymax<<shift)) P[2][1]= (rel_ymax<<shift);

  1103.                 if(P[3][0] < (rel_xmin<<shift)) P[3][0]= (rel_xmin<<shift);

  1104.                 if(P[3][1] > (rel_ymax<<shift)) P[3][1]= (rel_ymax<<shift);

  1105.         

  1106.                 P[4][0]= mid_pred(P[1][0], P[2][0], P[3][0]);

  1107.                 P[4][1]= mid_pred(P[1][1], P[2][1], P[3][1]);

  1108.             }

  1109.             pred_x= P[1][0];

  1110.             pred_y= P[1][1];

  1111.         }

  1112.         dmin = epzs_motion_search(s, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, ref_picture);

  1113.  

  1114.         mx+= mb_x*16;

  1115.         my+= mb_y*16;

  1116.         break;

  1117.     }

  1118.     

  1119.     /* intra / predictive decision */

  1120. //    xx = mb_x * 16;

  1121. //    yy = mb_y * 16;

  1122. 

  1123. //    pix = s->new_picture[0] + (yy * s->linesize) + xx;

  1124.     /* At this point (mx,my) are full-pell and the absolute displacement */

  1125. //    ppix = ref_picture + (my * s->linesize) + mx;

  1126.     

  1127.     halfpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax, pred_x, pred_y, ref_picture);

  1128. 

  1129. //    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;

  1130.     mv_table[mot_xy][0]= mx;

  1131.     mv_table[mot_xy][1]= my;

  1132. }

  1133. 

  1134. 

  1135. int ff_decide_type(MpegEncContext * s,

  1136.                 int mb_x, int mb_y)

  1137. {

  1138. 

  1139. }

  1140. 

  1141. void ff_estimate_b_frame_motion(MpegEncContext * s,

  1142.                              int mb_x, int mb_y)

  1143. {

  1144.     const int mot_stride = s->mb_width + 2;

  1145.     const int xy = (mb_y + 1)*mot_stride + mb_x + 1;

  1146. 

  1147.     ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, s->last_picture[0], s->f_code);

  1148.     ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, s->next_picture[0], s->b_code);

  1149. //printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);

  1150.     s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];

  1151.     s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];

  1152.     s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];

  1153.     s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];

  1154.     

  1155.     s->mb_type[mb_y*s->mb_width + mb_x]= MB_TYPE_FORWARD; //FIXME

  1156. }

  1157. 

  1158. /* find best f_code for ME which do unlimited searches */

  1159. int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)

  1160. {

  1161.     int f_code;

  1162. 

  1163.     if(s->me_method>=ME_EPZS){

  1164.         int mv_num[8];

  1165.         int i, y;

  1166.         int loose=0;

  1167.         UINT8 * fcode_tab= s->fcode_tab;

  1168. 

  1169.         for(i=0; i<8; i++) mv_num[i]=0;

  1170. 

  1171.         for(y=0; y<s->mb_height; y++){

  1172.             int x;

  1173.             int xy= (y+1)* (s->mb_width+2) + 1;

  1174.             i= y*s->mb_width;

  1175.             for(x=0; x<s->mb_width; x++){

  1176.                 if(s->mb_type[i] & type){

  1177.                     mv_num[ fcode_tab[mv_table[xy][0] + MAX_MV] ]++;

  1178.                     mv_num[ fcode_tab[mv_table[xy][1] + MAX_MV] ]++;

  1179. //printf("%d %d %d\n", s->mv_table[0][i], fcode_tab[s->mv_table[0][i] + MAX_MV], i);

  1180.                 }

  1181.                 i++;

  1182.                 xy++;

  1183.             }

  1184.         }

  1185. 

  1186.         for(i=MAX_FCODE; i>1; i--){

  1187.             loose+= mv_num[i];

  1188.             if(loose > s->mb_num/20) break; //FIXME this is pretty ineffective

  1189.         }

  1190. //    printf("fcode: %d type: %d\n", i, s->pict_type);

  1191.         return i;

  1192. /*        for(i=0; i<=MAX_FCODE; i++){

  1193.             printf("%d ", mv_num[i]);

  1194.         }

  1195.         printf("\n");*/

  1196.     }else{

  1197.         return 1;

  1198.     }

  1199. }

  1200. 

  1201. void ff_fix_long_p_mvs(MpegEncContext * s)

  1202. {

  1203.     const int f_code= s->f_code;

  1204.     int y;

  1205.     UINT8 * fcode_tab= s->fcode_tab;

  1206. 

  1207.     /* clip / convert to intra 16x16 type MVs */

  1208.     for(y=0; y<s->mb_height; y++){

  1209.         int x;

  1210.         int xy= (y+1)* (s->mb_width+2)+1;

  1211.         int i= y*s->mb_width;

  1212.         for(x=0; x<s->mb_width; x++){

  1213.             if(s->mb_type[i]&MB_TYPE_INTER){

  1214.                 if(   fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code

  1215.                    || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0

  1216.                    || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code

  1217.                    || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){

  1218.                     s->mb_type[i] &= ~MB_TYPE_INTER;

  1219.                     s->mb_type[i] |= MB_TYPE_INTRA;

  1220.                     s->p_mv_table[xy][0] = 0;

  1221.                     s->p_mv_table[xy][1] = 0;

  1222.                 }

  1223.             }

  1224.             xy++;

  1225.             i++;

  1226.         }

  1227.     }

  1228. 

  1229.     if(s->flags&CODEC_FLAG_4MV){

  1230.         const int wrap= 2+ s->mb_width*2;

  1231. 

  1232.         /* clip / convert to intra 8x8 type MVs */

  1233.         for(y=0; y<s->mb_height; y++){

  1234.             int xy= (y*2 + 1)*wrap + 1;

  1235.             int i= y*s->mb_width;

  1236.             int x;

  1237. 

  1238.             for(x=0; x<s->mb_width; x++){

  1239.                 if(s->mb_type[i]&MB_TYPE_INTER4V){

  1240.                     int block;

  1241.                     for(block=0; block<4; block++){

  1242.                         int off= (block& 1) + (block>>1)*wrap;

  1243.                         int mx= s->motion_val[ xy + off ][0];

  1244.                         int my= s->motion_val[ xy + off ][1];

  1245. 

  1246.                         if(   fcode_tab[mx + MAX_MV] > f_code

  1247.                            || fcode_tab[mx + MAX_MV] == 0

  1248.                            || fcode_tab[my + MAX_MV] > f_code

  1249.                            || fcode_tab[my + MAX_MV] == 0 ){

  1250.                             s->mb_type[i] &= ~MB_TYPE_INTER4V;

  1251.                             s->mb_type[i] |= MB_TYPE_INTRA;

  1252.                         }

  1253.                     }

  1254.                     xy+=2;

  1255.                     i++;

  1256.                 }

  1257.             }

  1258.         }

  1259.     }

  1260. }

  1261. 

  1262. void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)

  1263. {

  1264.     int y;

  1265.     UINT8 * fcode_tab= s->fcode_tab;

  1266. 

  1267.     /* clip / convert to intra 16x16 type MVs */

  1268.     for(y=0; y<s->mb_height; y++){

  1269.         int x;

  1270.         int xy= (y+1)* (s->mb_width+2)+1;

  1271.         int i= y*s->mb_width;

  1272.         for(x=0; x<s->mb_width; x++){

  1273.             if(s->mb_type[i]&type){

  1274.                 if(   fcode_tab[mv_table[xy][0] + MAX_MV] > f_code

  1275.                    || fcode_tab[mv_table[xy][0] + MAX_MV] == 0

  1276.                    || fcode_tab[mv_table[xy][1] + MAX_MV] > f_code

  1277.                    || fcode_tab[mv_table[xy][1] + MAX_MV] == 0 ){

  1278.                     s->mb_type[i] &= ~type;

  1279.                     if(s->mb_type[i]==0) s->mb_type[i]= MB_TYPE_FORWARD; //FIXME 

  1280.                     mv_table[xy][0] = 0;

  1281.                     mv_table[xy][1] = 0;

  1282.                     //this is certainly bad FIXME

  1283.                 }

  1284.             }

  1285.             xy++;

  1286.             i++;

  1287.         }

  1288.     }

  1289. }