Statistics
| Branch: | Revision:

ffmpeg / libavcodec / motion_est.c @ 186447f8

History | View | Annotate | Download (56.2 KB)

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
        fprintf(stderr,"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

    
834
    for(block=0; block<4; block++){
835
        int mx4, my4;
836
        int pred_x4, pred_y4;
837
        int dmin4;
838
        static const int off[4]= {2, 1, 1, -1};
839
        const int mot_stride = s->block_wrap[0];
840
        const int mot_xy = s->block_index[block];
841
//        const int block_x= (block&1);
842
//        const int block_y= (block>>1);
843
#if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way
844
        const int rel_xmin4= xmin;
845
        const int rel_xmax4= xmax;
846
        const int rel_ymin4= ymin;
847
        const int rel_ymax4= ymax;
848
#else
849
        const int rel_xmin4= xmin - block_x*8;
850
        const int rel_xmax4= xmax - block_x*8 + 8;
851
        const int rel_ymin4= ymin - block_y*8;
852
        const int rel_ymax4= ymax - block_y*8 + 8;
853
#endif
854
        P_LEFT[0] = s->motion_val[mot_xy - 1][0];
855
        P_LEFT[1] = s->motion_val[mot_xy - 1][1];
856

    
857
        if(P_LEFT[0]       > (rel_xmax4<<shift)) P_LEFT[0]       = (rel_xmax4<<shift);
858

    
859
        /* special case for first line */
860
        if (s->mb_y == 0 && block<2) {
861
            pred_x4= P_LEFT[0];
862
            pred_y4= P_LEFT[1];
863
        } else {
864
            P_TOP[0]      = s->motion_val[mot_xy - mot_stride             ][0];
865
            P_TOP[1]      = s->motion_val[mot_xy - mot_stride             ][1];
866
            P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
867
            P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
868
            if(P_TOP[1]      > (rel_ymax4<<shift)) P_TOP[1]     = (rel_ymax4<<shift);
869
            if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);
870
            if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);
871
            if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);
872
    
873
            P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
874
            P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
875

    
876
//            if(s->out_format == FMT_H263){
877
                pred_x4 = P_MEDIAN[0];
878
                pred_y4 = P_MEDIAN[1];
879
#if 0
880
            }else { /* mpeg1 at least */
881
                pred_x4= P_LEFT[0];
882
                pred_y4= P_LEFT[1];
883
            }
884
#endif
885
        }
886
        P_MV1[0]= mx;
887
        P_MV1[1]= my;
888

    
889
        dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
890
                                       &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
891

    
892
        dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
893
                                          pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);
894
        
895
        if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
896
            int dxy;
897
            const int offset= ((block&1) + (block>>1)*s->linesize)*8;
898
            uint8_t *dest_y = s->me.scratchpad + offset;
899

    
900
            if(s->quarter_sample){
901
                uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>2)) + (s->mb_y*16 + (my4>>2))*s->linesize + offset;
902
                dxy = ((my4 & 3) << 2) | (mx4 & 3);
903

    
904
                if(s->no_rounding)
905
                    s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , s->linesize);
906
                else
907
                    s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , s->linesize);
908
            }else{
909
                uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>1)) + (s->mb_y*16 + (my4>>1))*s->linesize + offset;
910
                dxy = ((my4 & 1) << 1) | (mx4 & 1);
911

    
912
                if(s->no_rounding)
913
                    s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , s->linesize, 8);
914
                else
915
                    s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , s->linesize, 8);
916
            }
917
            dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
918
        }else
919
            dmin_sum+= dmin4;
920

    
921
        if(s->quarter_sample){
922
            mx4_sum+= mx4/2;
923
            my4_sum+= my4/2;
924
        }else{
925
            mx4_sum+= mx4;
926
            my4_sum+= my4;
927
        }
928
            
929
        s->motion_val[ s->block_index[block] ][0]= mx4;
930
        s->motion_val[ s->block_index[block] ][1]= my4;
931
    }
932
    
933
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
934
        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);
935
    }
936
    
937
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
938
        int dxy;
939
        int mx, my;
940
        int offset;
941

    
942
        mx= ff_h263_round_chroma(mx4_sum);
943
        my= ff_h263_round_chroma(my4_sum);
944
        dxy = ((my & 1) << 1) | (mx & 1);
945
        
946
        offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
947
       
948
        if(s->no_rounding){
949
            s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
950
            s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
951
        }else{
952
            s->dsp.put_pixels_tab       [1][dxy](s->me.scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
953
            s->dsp.put_pixels_tab       [1][dxy](s->me.scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
954
        }
955

    
956
        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);
957
        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);
958
    }
959

    
960
    switch(s->avctx->mb_cmp&0xFF){
961
    /*case FF_CMP_SSE:
962
        return dmin_sum+ 32*s->qscale*s->qscale;*/
963
    case FF_CMP_RD:
964
        return dmin_sum;
965
    default:
966
        return dmin_sum+ 11*s->me.mb_penalty_factor;
967
    }
968
}
969

    
970
void ff_estimate_p_frame_motion(MpegEncContext * s,
971
                                int mb_x, int mb_y)
972
{
973
    uint8_t *pix, *ppix;
974
    int sum, varc, vard, mx, my, range, dmin, xx, yy;
975
    int xmin, ymin, xmax, ymax;
976
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
977
    int pred_x=0, pred_y=0;
978
    int P[10][2];
979
    const int shift= 1+s->quarter_sample;
980
    int mb_type=0;
981
    uint8_t *ref_picture= s->last_picture.data[0];
982
    Picture * const pic= &s->current_picture;
983
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
984
    
985
    assert(s->quarter_sample==0 || s->quarter_sample==1);
986

    
987
    s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
988
    s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
989
    s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
990

    
991
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
992
    rel_xmin= xmin - mb_x*16;
993
    rel_xmax= xmax - mb_x*16;
994
    rel_ymin= ymin - mb_y*16;
995
    rel_ymax= ymax - mb_y*16;
996
    s->me.skip=0;
997

    
998
    switch(s->me_method) {
999
    case ME_ZERO:
1000
    default:
1001
        no_motion_search(s, &mx, &my);
1002
        mx-= mb_x*16;
1003
        my-= mb_y*16;
1004
        dmin = 0;
1005
        break;
1006
    case ME_FULL:
1007
        dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1008
        mx-= mb_x*16;
1009
        my-= mb_y*16;
1010
        break;
1011
    case ME_LOG:
1012
        dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1013
        mx-= mb_x*16;
1014
        my-= mb_y*16;
1015
        break;
1016
    case ME_PHODS:
1017
        dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1018
        mx-= mb_x*16;
1019
        my-= mb_y*16;
1020
        break;
1021
    case ME_X1:
1022
    case ME_EPZS:
1023
       {
1024
            const int mot_stride = s->block_wrap[0];
1025
            const int mot_xy = s->block_index[0];
1026

    
1027
            P_LEFT[0]       = s->motion_val[mot_xy - 1][0];
1028
            P_LEFT[1]       = s->motion_val[mot_xy - 1][1];
1029

    
1030
            if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);
1031

    
1032
            if(mb_y) {
1033
                P_TOP[0]      = s->motion_val[mot_xy - mot_stride    ][0];
1034
                P_TOP[1]      = s->motion_val[mot_xy - mot_stride    ][1];
1035
                P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];
1036
                P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];
1037
                if(P_TOP[1]      > (rel_ymax<<shift)) P_TOP[1]     = (rel_ymax<<shift);
1038
                if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1039
                if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1040
        
1041
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1042
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1043

    
1044
                if(s->out_format == FMT_H263){
1045
                    pred_x = P_MEDIAN[0];
1046
                    pred_y = P_MEDIAN[1];
1047
                }else { /* mpeg1 at least */
1048
                    pred_x= P_LEFT[0];
1049
                    pred_y= P_LEFT[1];
1050
                }
1051
            }else{
1052
                pred_x= P_LEFT[0];
1053
                pred_y= P_LEFT[1];
1054
            }
1055

    
1056
        }
1057
        dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1058
                                      &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1059
 
1060
        break;
1061
    }
1062

    
1063
    /* intra / predictive decision */
1064
    xx = mb_x * 16;
1065
    yy = mb_y * 16;
1066

    
1067
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
1068
    /* At this point (mx,my) are full-pell and the relative displacement */
1069
    ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
1070
    
1071
    sum = s->dsp.pix_sum(pix, s->linesize);
1072
    
1073
    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1074
    vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
1075

    
1076
//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);
1077
    pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;
1078
    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1079
    pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1080
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
1081
    pic->mb_var_sum    += varc;
1082
    pic->mc_mb_var_sum += vard;
1083
//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);
1084
    
1085
#if 0
1086
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1087
           varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1088
#endif
1089
    if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1090
        if (vard <= 64 || vard < varc)
1091
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1092
        else
1093
            s->scene_change_score+= s->qscale;
1094

    
1095
        if (vard*2 + 200 > varc)
1096
            mb_type|= MB_TYPE_INTRA;
1097
        if (varc*2 + 200 > vard){
1098
            mb_type|= MB_TYPE_INTER;
1099
            s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1100
                                   pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1101
        }else{
1102
            mx <<=shift;
1103
            my <<=shift;
1104
        }
1105
        if((s->flags&CODEC_FLAG_4MV)
1106
           && !s->me.skip && varc>50 && vard>10){
1107
            h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1108
            mb_type|=MB_TYPE_INTER4V;
1109

    
1110
            set_p_mv_tables(s, mx, my, 0);
1111
        }else
1112
            set_p_mv_tables(s, mx, my, 1);
1113
    }else{
1114
        mb_type= MB_TYPE_INTER;
1115

    
1116
        dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1117
                                    pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1118
        
1119
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1120
            dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);
1121

    
1122
        if((s->flags&CODEC_FLAG_4MV)
1123
           && !s->me.skip && varc>50 && vard>10){
1124
            int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1125
            if(dmin4 < dmin){
1126
                mb_type= MB_TYPE_INTER4V;
1127
                dmin=dmin4;
1128
            }
1129
        }
1130
        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
1131
        set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
1132
        
1133
        if (vard <= 64 || vard < varc) {
1134
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1135
        }else{
1136
            s->scene_change_score+= s->qscale;
1137
        }
1138
    }
1139

    
1140
    s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1141
}
1142

    
1143
int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1144
                                    int mb_x, int mb_y)
1145
{
1146
    int mx, my, range, dmin;
1147
    int xmin, ymin, xmax, ymax;
1148
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1149
    int pred_x=0, pred_y=0;
1150
    int P[10][2];
1151
    const int shift= 1+s->quarter_sample;
1152
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1153
    const int xy= mb_x + mb_y*s->mb_stride;
1154
    
1155
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1156

    
1157
    s->me.pre_penalty_factor    = get_penalty_factor(s, s->avctx->me_pre_cmp);
1158

    
1159
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1160
    rel_xmin= xmin - mb_x*16;
1161
    rel_xmax= xmax - mb_x*16;
1162
    rel_ymin= ymin - mb_y*16;
1163
    rel_ymax= ymax - mb_y*16;
1164
    s->me.skip=0;
1165

    
1166
    P_LEFT[0]       = s->p_mv_table[xy + 1][0];
1167
    P_LEFT[1]       = s->p_mv_table[xy + 1][1];
1168

    
1169
    if(P_LEFT[0]       < (rel_xmin<<shift)) P_LEFT[0]       = (rel_xmin<<shift);
1170

    
1171
    /* special case for first line */
1172
    if (mb_y == s->mb_height-1) {
1173
        pred_x= P_LEFT[0];
1174
        pred_y= P_LEFT[1];
1175
        P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1176
        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME 
1177
    } else {
1178
        P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
1179
        P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
1180
        P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1181
        P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1182
        if(P_TOP[1]      < (rel_ymin<<shift)) P_TOP[1]     = (rel_ymin<<shift);
1183
        if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
1184
        if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
1185
    
1186
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1187
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1188

    
1189
        pred_x = P_MEDIAN[0];
1190
        pred_y = P_MEDIAN[1];
1191
    }
1192
    dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1193
                                   &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1194

    
1195
    s->p_mv_table[xy][0] = mx<<shift;
1196
    s->p_mv_table[xy][1] = my<<shift;
1197
    
1198
    return dmin;
1199
}
1200

    
1201
static int ff_estimate_motion_b(MpegEncContext * s,
1202
                       int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
1203
{
1204
    int mx, my, range, dmin;
1205
    int xmin, ymin, xmax, ymax;
1206
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1207
    int pred_x=0, pred_y=0;
1208
    int P[10][2];
1209
    const int shift= 1+s->quarter_sample;
1210
    const int mot_stride = s->mb_stride;
1211
    const int mot_xy = mb_y*mot_stride + mb_x;
1212
    uint8_t * const ref_picture= picture->data[0];
1213
    uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1214
    int mv_scale;
1215
        
1216
    s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
1217
    s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1218
    s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1219

    
1220
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1221
    rel_xmin= xmin - mb_x*16;
1222
    rel_xmax= xmax - mb_x*16;
1223
    rel_ymin= ymin - mb_y*16;
1224
    rel_ymax= ymax - mb_y*16;
1225

    
1226
    switch(s->me_method) {
1227
    case ME_ZERO:
1228
    default:
1229
        no_motion_search(s, &mx, &my);
1230
        dmin = 0;
1231
        mx-= mb_x*16;
1232
        my-= mb_y*16;
1233
        break;
1234
    case ME_FULL:
1235
        dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1236
        mx-= mb_x*16;
1237
        my-= mb_y*16;
1238
        break;
1239
    case ME_LOG:
1240
        dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1241
        mx-= mb_x*16;
1242
        my-= mb_y*16;
1243
        break;
1244
    case ME_PHODS:
1245
        dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1246
        mx-= mb_x*16;
1247
        my-= mb_y*16;
1248
        break;
1249
    case ME_X1:
1250
    case ME_EPZS:
1251
       {
1252
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
1253
            P_LEFT[1]        = mv_table[mot_xy - 1][1];
1254

    
1255
            if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);
1256

    
1257
            /* special case for first line */
1258
            if (mb_y) {
1259
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
1260
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
1261
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
1262
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
1263
                if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);
1264
                if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1265
                if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1266
        
1267
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1268
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1269
            }
1270
            pred_x= P_LEFT[0];
1271
            pred_y= P_LEFT[1];
1272
        }
1273
        
1274
        if(mv_table == s->b_forw_mv_table){
1275
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1276
        }else{
1277
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1278
        }
1279
        
1280
        dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1281
                                      picture, s->p_mv_table, mv_scale, mv_penalty);
1282
 
1283
        break;
1284
    }
1285
    
1286
    dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1287
                                   pred_x, pred_y, picture, 0, 0, mv_penalty);
1288
                                   
1289
    if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1290
        dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);
1291

    
1292
//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1293
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1294
    mv_table[mot_xy][0]= mx;
1295
    mv_table[mot_xy][1]= my;
1296

    
1297
    return dmin;
1298
}
1299

    
1300
static inline int check_bidir_mv(MpegEncContext * s,
1301
                   int mb_x, int mb_y,
1302
                   int motion_fx, int motion_fy,
1303
                   int motion_bx, int motion_by,
1304
                   int pred_fx, int pred_fy,
1305
                   int pred_bx, int pred_by)
1306
{
1307
    //FIXME optimize?
1308
    //FIXME move into template?
1309
    //FIXME better f_code prediction (max mv & distance)
1310
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1311
    uint8_t *dest_y = s->me.scratchpad;
1312
    uint8_t *ptr;
1313
    int dxy;
1314
    int src_x, src_y;
1315
    int fbmin;
1316

    
1317
    if(s->quarter_sample){
1318
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1319
        src_x = mb_x * 16 + (motion_fx >> 2);
1320
        src_y = mb_y * 16 + (motion_fy >> 2);
1321
        assert(src_x >=-16 && src_x<=s->h_edge_pos);
1322
        assert(src_y >=-16 && src_y<=s->v_edge_pos);
1323

    
1324
        ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1325
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);
1326

    
1327
        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1328
        src_x = mb_x * 16 + (motion_bx >> 2);
1329
        src_y = mb_y * 16 + (motion_by >> 2);
1330
        assert(src_x >=-16 && src_x<=s->h_edge_pos);
1331
        assert(src_y >=-16 && src_y<=s->v_edge_pos);
1332
    
1333
        ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1334
        s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);
1335
    }else{
1336
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1337
        src_x = mb_x * 16 + (motion_fx >> 1);
1338
        src_y = mb_y * 16 + (motion_fy >> 1);
1339
        assert(src_x >=-16 && src_x<=s->h_edge_pos);
1340
        assert(src_y >=-16 && src_y<=s->v_edge_pos);
1341

    
1342
        ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1343
        s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
1344

    
1345
        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1346
        src_x = mb_x * 16 + (motion_bx >> 1);
1347
        src_y = mb_y * 16 + (motion_by >> 1);
1348
        assert(src_x >=-16 && src_x<=s->h_edge_pos);
1349
        assert(src_y >=-16 && src_y<=s->v_edge_pos);
1350
    
1351
        ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1352
        s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
1353
    }
1354

    
1355
    fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1356
           +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1357
           + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
1358
           
1359
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1360
    }
1361
    //FIXME CHROMA !!!
1362
           
1363
    return fbmin;
1364
}
1365

    
1366
/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1367
static inline int bidir_refine(MpegEncContext * s,
1368
                                  int mb_x, int mb_y)
1369
{
1370
    const int mot_stride = s->mb_stride;
1371
    const int xy = mb_y *mot_stride + mb_x;
1372
    int fbmin;
1373
    int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1374
    int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1375
    int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1376
    int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1377
    int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1378
    int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1379
    int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1380
    int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1381

    
1382
    //FIXME do refinement and add flag
1383
    
1384
    fbmin= check_bidir_mv(s, mb_x, mb_y, 
1385
                          motion_fx, motion_fy,
1386
                          motion_bx, motion_by,
1387
                          pred_fx, pred_fy,
1388
                          pred_bx, pred_by);
1389

    
1390
   return fbmin;
1391
}
1392

    
1393
static inline int direct_search(MpegEncContext * s,
1394
                                int mb_x, int mb_y)
1395
{
1396
    int P[10][2];
1397
    const int mot_stride = s->mb_stride;
1398
    const int mot_xy = mb_y*mot_stride + mb_x;
1399
    const int shift= 1+s->quarter_sample;
1400
    int dmin, i;
1401
    const int time_pp= s->pp_time;
1402
    const int time_pb= s->pb_time;
1403
    int mx, my, xmin, xmax, ymin, ymax;
1404
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
1405
    uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1406
    
1407
    ymin= xmin=(-32)>>shift;
1408
    ymax= xmax=   31>>shift;
1409

    
1410
    if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1411
        s->mv_type= MV_TYPE_8X8;
1412
    }else{
1413
        s->mv_type= MV_TYPE_16X16;
1414
    }
1415

    
1416
    for(i=0; i<4; i++){
1417
        int index= s->block_index[i];
1418
        int min, max;
1419
    
1420
        s->me.co_located_mv[i][0]= s->motion_val[index][0];
1421
        s->me.co_located_mv[i][1]= s->motion_val[index][1];
1422
        s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1423
        s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1424
//        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);
1425
//        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);
1426

    
1427
        max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1428
        min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1429
        max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding
1430
        min+= (2*mb_x + (i& 1))*8 - 1;
1431
        xmax= FFMIN(xmax, s->width - max);
1432
        xmin= FFMAX(xmin, - 16     - min);
1433

    
1434
        max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1435
        min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1436
        max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding
1437
        min+= (2*mb_y + (i>>1))*8 - 1;
1438
        ymax= FFMIN(ymax, s->height - max);
1439
        ymin= FFMAX(ymin, - 16      - min);
1440
        
1441
        if(s->mv_type == MV_TYPE_16X16) break;
1442
    }
1443
    
1444
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1445
    
1446
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1447
        s->b_direct_mv_table[mot_xy][0]= 0;
1448
        s->b_direct_mv_table[mot_xy][1]= 0;
1449

    
1450
        return 256*256*256*64;
1451
    }
1452

    
1453
    P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1454
    P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1455

    
1456
    /* special case for first line */
1457
    if (mb_y) {
1458
        P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
1459
        P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
1460
        P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
1461
        P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
1462
    
1463
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1464
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1465
    }
1466
 
1467
    //FIXME direct_search  ptr in context!!! (needed for chroma anyway or this will get messy)   
1468
    if(s->flags&CODEC_FLAG_QPEL){
1469
        dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 
1470
                                                     &s->last_picture, mv_table, 1<<14, mv_penalty);
1471
        dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1472
                                                0, 0, &s->last_picture, 0, 0, mv_penalty);
1473
        
1474
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1475
            dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1476
    }else{
1477
        dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 
1478
                                                     &s->last_picture, mv_table, 1<<15, mv_penalty);
1479
        dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1480
                                                0, 0, &s->last_picture, 0, 0, mv_penalty);
1481

    
1482
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1483
            dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1484
    }
1485

    
1486
    s->b_direct_mv_table[mot_xy][0]= mx;
1487
    s->b_direct_mv_table[mot_xy][1]= my;
1488
    return dmin;
1489
}
1490

    
1491
void ff_estimate_b_frame_motion(MpegEncContext * s,
1492
                             int mb_x, int mb_y)
1493
{
1494
    const int penalty_factor= s->me.mb_penalty_factor;
1495
    int fmin, bmin, dmin, fbmin;
1496
    int type=0;
1497
    
1498
    s->me.skip=0;
1499
    if (s->codec_id == CODEC_ID_MPEG4)
1500
        dmin= direct_search(s, mb_x, mb_y);
1501
    else
1502
        dmin= INT_MAX;
1503

    
1504
    s->me.skip=0;
1505
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;
1506
    
1507
    s->me.skip=0;
1508
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;
1509
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1510

    
1511
    s->me.skip=0;
1512
    fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1513
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1514
    {
1515
        int score= fmin;
1516
        type = MB_TYPE_FORWARD;
1517
        
1518
        if (dmin <= score){
1519
            score = dmin;
1520
            type = MB_TYPE_DIRECT;
1521
        }
1522
        if(bmin<score){
1523
            score=bmin;
1524
            type= MB_TYPE_BACKWARD; 
1525
        }
1526
        if(fbmin<score){
1527
            score=fbmin;
1528
            type= MB_TYPE_BIDIR;
1529
        }
1530
        
1531
        score= ((unsigned)(score*score + 128*256))>>16;
1532
        s->current_picture.mc_mb_var_sum += score;
1533
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1534
    }
1535

    
1536
    if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1537
        type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
1538
        if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1539
    }
1540

    
1541
    s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1542
}
1543

    
1544
/* find best f_code for ME which do unlimited searches */
1545
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1546
{
1547
    if(s->me_method>=ME_EPZS){
1548
        int score[8];
1549
        int i, y;
1550
        uint8_t * fcode_tab= s->fcode_tab;
1551
        int best_fcode=-1;
1552
        int best_score=-10000000;
1553

    
1554
        for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1555

    
1556
        for(y=0; y<s->mb_height; y++){
1557
            int x;
1558
            int xy= y*s->mb_stride;
1559
            for(x=0; x<s->mb_width; x++){
1560
                if(s->mb_type[xy] & type){
1561
                    int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1562
                                     fcode_tab[mv_table[xy][1] + MAX_MV]);
1563
                    int j;
1564
                    
1565
                    for(j=0; j<fcode && j<8; j++){
1566
                        if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1567
                            score[j]-= 170;
1568
                    }
1569
                }
1570
                xy++;
1571
            }
1572
        }
1573
        
1574
        for(i=1; i<8; i++){
1575
            if(score[i] > best_score){
1576
                best_score= score[i];
1577
                best_fcode= i;
1578
            }
1579
//            printf("%d %d\n", i, score[i]);
1580
        }
1581

    
1582
//    printf("fcode: %d type: %d\n", i, s->pict_type);
1583
        return best_fcode;
1584
/*        for(i=0; i<=MAX_FCODE; i++){
1585
            printf("%d ", mv_num[i]);
1586
        }
1587
        printf("\n");*/
1588
    }else{
1589
        return 1;
1590
    }
1591
}
1592

    
1593
void ff_fix_long_p_mvs(MpegEncContext * s)
1594
{
1595
    const int f_code= s->f_code;
1596
    int y, range;
1597
    assert(s->pict_type==P_TYPE);
1598

    
1599
    range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1600
    
1601
    if(s->msmpeg4_version) range= 16;
1602
    
1603
    if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1604
    
1605
    /* clip / convert to intra 16x16 type MVs */
1606
    for(y=0; y<s->mb_height; y++){
1607
        int x;
1608
        int xy= y*s->mb_stride;
1609
        for(x=0; x<s->mb_width; x++){
1610
            if(s->mb_type[xy]&MB_TYPE_INTER){
1611
                if(   s->p_mv_table[xy][0] >=range || s->p_mv_table[xy][0] <-range
1612
                   || s->p_mv_table[xy][1] >=range || s->p_mv_table[xy][1] <-range){
1613
                    s->mb_type[xy] &= ~MB_TYPE_INTER;
1614
                    s->mb_type[xy] |= MB_TYPE_INTRA;
1615
                    s->p_mv_table[xy][0] = 0;
1616
                    s->p_mv_table[xy][1] = 0;
1617
                }
1618
            }
1619
            xy++;
1620
        }
1621
    }
1622
//printf("%d no:%d %d//\n", clip, noclip, f_code);
1623
    if(s->flags&CODEC_FLAG_4MV){
1624
        const int wrap= 2+ s->mb_width*2;
1625

    
1626
        /* clip / convert to intra 8x8 type MVs */
1627
        for(y=0; y<s->mb_height; y++){
1628
            int xy= (y*2 + 1)*wrap + 1;
1629
            int i= y*s->mb_stride;
1630
            int x;
1631

    
1632
            for(x=0; x<s->mb_width; x++){
1633
                if(s->mb_type[i]&MB_TYPE_INTER4V){
1634
                    int block;
1635
                    for(block=0; block<4; block++){
1636
                        int off= (block& 1) + (block>>1)*wrap;
1637
                        int mx= s->motion_val[ xy + off ][0];
1638
                        int my= s->motion_val[ xy + off ][1];
1639

    
1640
                        if(   mx >=range || mx <-range
1641
                           || my >=range || my <-range){
1642
                            s->mb_type[i] &= ~MB_TYPE_INTER4V;
1643
                            s->mb_type[i] |= MB_TYPE_INTRA;
1644
                        }
1645
                    }
1646
                }
1647
                xy+=2;
1648
                i++;
1649
            }
1650
        }
1651
    }
1652
}
1653

    
1654
void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
1655
{
1656
    int y;
1657

    
1658
    // RAL: 8 in MPEG-1, 16 in MPEG-4
1659
    int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1660
    
1661
    if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1662

    
1663
    /* clip / convert to intra 16x16 type MVs */
1664
    for(y=0; y<s->mb_height; y++){
1665
        int x;
1666
        int xy= y*s->mb_stride;
1667
        for(x=0; x<s->mb_width; x++){
1668
            if (s->mb_type[xy] & type){    // RAL: "type" test added...
1669
                if(   mv_table[xy][0] >=range || mv_table[xy][0] <-range
1670
                   || mv_table[xy][1] >=range || mv_table[xy][1] <-range){
1671

    
1672
                    if(s->codec_id == CODEC_ID_MPEG1VIDEO && 0){
1673
                    }else{
1674
                        if     (mv_table[xy][0] > range-1) mv_table[xy][0]=  range-1;
1675
                        else if(mv_table[xy][0] < -range ) mv_table[xy][0]= -range;
1676
                        if     (mv_table[xy][1] > range-1) mv_table[xy][1]=  range-1;
1677
                        else if(mv_table[xy][1] < -range ) mv_table[xy][1]= -range;
1678
                    }
1679
                }
1680
            }
1681
            xy++;
1682
        }
1683
    }
1684
}