Statistics
| Branch: | Revision:

ffmpeg / libavcodec / motion_est.c @ c40c3482

History | View | Annotate | Download (56.4 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
#include <stdlib.h>
24
#include <stdio.h>
25
#include "avcodec.h"
26
#include "dsputil.h"
27
#include "mpegvideo.h"
28

    
29
//#undef NDEBUG
30
//#include <assert.h>
31

    
32
#define SQ(a) ((a)*(a))
33

    
34
#define P_LEFT P[1]
35
#define P_TOP P[2]
36
#define P_TOPRIGHT P[3]
37
#define P_MEDIAN P[4]
38
#define P_MV1 P[9]
39

    
40
static inline int sad_hpel_motion_search(MpegEncContext * s,
41
                                  int *mx_ptr, int *my_ptr, int dmin,
42
                                  int xmin, int ymin, int xmax, int ymax,
43
                                  int pred_x, int pred_y, Picture *picture,
44
                                  int n, int size, uint16_t * const mv_penalty);
45

    
46
static inline int update_map_generation(MpegEncContext * s)
47
{
48
    s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2);
49
    if(s->me.map_generation==0){
50
        s->me.map_generation= 1<<(ME_MAP_MV_BITS*2);
51
        memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
52
    }
53
    return s->me.map_generation;
54
}
55

    
56
/* shape adaptive search stuff */
57
typedef struct Minima{
58
    int height;
59
    int x, y;
60
    int checked;
61
}Minima;
62

    
63
static int minima_cmp(const void *a, const void *b){
64
    Minima *da = (Minima *) a;
65
    Minima *db = (Minima *) b;
66
    
67
    return da->height - db->height;
68
}
69
                                  
70
/* SIMPLE */
71
#define RENAME(a) simple_ ## a
72

    
73
#define CMP(d, x, y, size)\
74
d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);
75

    
76
#define CMP_HPEL(d, dx, dy, x, y, size)\
77
{\
78
    const int dxy= (dx) + 2*(dy);\
79
    hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\
80
    d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
81
}
82

    
83
#define CMP_QPEL(d, dx, dy, x, y, size)\
84
{\
85
    const int dxy= (dx) + 4*(dy);\
86
    qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
87
    d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
88
}
89

    
90
#include "motion_est_template.c"
91
#undef RENAME
92
#undef CMP
93
#undef CMP_HPEL
94
#undef CMP_QPEL
95
#undef INIT
96

    
97
/* SIMPLE CHROMA */
98
#define RENAME(a) simple_chroma_ ## a
99

    
100
#define CMP(d, x, y, size)\
101
d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);\
102
if(chroma_cmp){\
103
    int dxy= ((x)&1) + 2*((y)&1);\
104
    int c= ((x)>>1) + ((y)>>1)*uvstride;\
105
\
106
    chroma_hpel_put[0][dxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
107
    d += chroma_cmp(s, s->me.scratchpad, src_u, uvstride);\
108
    chroma_hpel_put[0][dxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
109
    d += chroma_cmp(s, s->me.scratchpad, src_v, uvstride);\
110
}
111

    
112
#define CMP_HPEL(d, dx, dy, x, y, size)\
113
{\
114
    const int dxy= (dx) + 2*(dy);\
115
    hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\
116
    d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
117
    if(chroma_cmp_sub){\
118
        int cxy= (dxy) | ((x)&1) | (2*((y)&1));\
119
        int c= ((x)>>1) + ((y)>>1)*uvstride;\
120
        chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
121
        d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
122
        chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
123
        d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
124
    }\
125
}
126

    
127
#define CMP_QPEL(d, dx, dy, x, y, size)\
128
{\
129
    const int dxy= (dx) + 4*(dy);\
130
    qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
131
    d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
132
    if(chroma_cmp_sub){\
133
        int cxy, c;\
134
        int cx= (4*(x) + (dx))/2;\
135
        int cy= (4*(y) + (dy))/2;\
136
        cx= (cx>>1)|(cx&1);\
137
        cy= (cy>>1)|(cy&1);\
138
        cxy= (cx&1) + 2*(cy&1);\
139
        c= ((cx)>>1) + ((cy)>>1)*uvstride;\
140
        chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
141
        d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
142
        chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
143
        d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
144
    }\
145
}
146

    
147
#include "motion_est_template.c"
148
#undef RENAME
149
#undef CMP
150
#undef CMP_HPEL
151
#undef CMP_QPEL
152
#undef INIT
153

    
154
/* SIMPLE DIRECT HPEL */
155
#define RENAME(a) simple_direct_hpel_ ## a
156
//FIXME precalc divisions stuff
157

    
158
#define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
159
if((x) >= xmin && 2*(x) + (dx) <= 2*xmax && (y) >= ymin && 2*(y) + (dy) <= 2*ymax){\
160
    const int hx= 2*(x) + (dx);\
161
    const int hy= 2*(y) + (dy);\
162
    if(s->mv_type==MV_TYPE_8X8){\
163
        int i;\
164
        for(i=0; i<4; i++){\
165
            int fx = s->me.direct_basis_mv[i][0] + hx;\
166
            int fy = s->me.direct_basis_mv[i][1] + hy;\
167
            int bx = hx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\
168
            int by = hy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\
169
            int fxy= (fx&1) + 2*(fy&1);\
170
            int bxy= (bx&1) + 2*(by&1);\
171
\
172
            uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
173
            hpel_put[1][fxy](dst, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 8);\
174
            hpel_avg[1][bxy](dst, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 8);\
175
        }\
176
    }else{\
177
        int fx = s->me.direct_basis_mv[0][0] + hx;\
178
        int fy = s->me.direct_basis_mv[0][1] + hy;\
179
        int bx = hx ? fx - s->me.co_located_mv[0][0] : (s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp);\
180
        int by = hy ? fy - s->me.co_located_mv[0][1] : (s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp);\
181
        int fxy= (fx&1) + 2*(fy&1);\
182
        int bxy= (bx&1) + 2*(by&1);\
183
        \
184
        assert((fx>>1) + 16*s->mb_x >= -16);\
185
        assert((fy>>1) + 16*s->mb_y >= -16);\
186
        assert((fx>>1) + 16*s->mb_x <= s->width);\
187
        assert((fy>>1) + 16*s->mb_y <= s->height);\
188
        assert((bx>>1) + 16*s->mb_x >= -16);\
189
        assert((by>>1) + 16*s->mb_y >= -16);\
190
        assert((bx>>1) + 16*s->mb_x <= s->width);\
191
        assert((by>>1) + 16*s->mb_y <= s->height);\
192
\
193
        hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\
194
        hpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 16);\
195
    }\
196
    d = cmp_func(s, s->me.scratchpad, src_y, stride);\
197
}else\
198
    d= 256*256*256*32;
199

    
200

    
201
#define CMP_HPEL(d, dx, dy, x, y, size)\
202
    CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
203

    
204
#define CMP(d, x, y, size)\
205
    CMP_DIRECT(d, 0, 0, x, y, size, cmp)
206
    
207
#include "motion_est_template.c"
208
#undef RENAME
209
#undef CMP
210
#undef CMP_HPEL
211
#undef CMP_QPEL
212
#undef INIT
213
#undef CMP_DIRECT
214

    
215
/* SIMPLE DIRECT QPEL */
216
#define RENAME(a) simple_direct_qpel_ ## a
217

    
218
#define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
219
if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\
220
    const int qx= 4*(x) + (dx);\
221
    const int qy= 4*(y) + (dy);\
222
    if(s->mv_type==MV_TYPE_8X8){\
223
        int i;\
224
        for(i=0; i<4; i++){\
225
            int fx = s->me.direct_basis_mv[i][0] + qx;\
226
            int fy = s->me.direct_basis_mv[i][1] + qy;\
227
            int bx = qx ? fx - s->me.co_located_mv[i][0] : s->me.co_located_mv[i][0]*(time_pb - time_pp)/time_pp + (i &1)*16;\
228
            int by = qy ? fy - s->me.co_located_mv[i][1] : s->me.co_located_mv[i][1]*(time_pb - time_pp)/time_pp + (i>>1)*16;\
229
            int fxy= (fx&3) + 4*(fy&3);\
230
            int bxy= (bx&3) + 4*(by&3);\
231
\
232
            uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
233
            qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\
234
            qpel_avg[1][bxy](dst, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\
235
        }\
236
    }else{\
237
        int fx = s->me.direct_basis_mv[0][0] + qx;\
238
        int fy = s->me.direct_basis_mv[0][1] + qy;\
239
        int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\
240
        int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\
241
        int fxy= (fx&3) + 4*(fy&3);\
242
        int bxy= (bx&3) + 4*(by&3);\
243
\
244
        qpel_put[1][fxy](s->me.scratchpad               , (ref_y ) + (fx>>2) + (fy>>2)*(stride)               , stride);\
245
        qpel_put[1][fxy](s->me.scratchpad + 8           , (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8           , stride);\
246
        qpel_put[1][fxy](s->me.scratchpad     + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride)     + 8*stride, stride);\
247
        qpel_put[1][fxy](s->me.scratchpad + 8 + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 + 8*stride, stride);\
248
        qpel_avg[1][bxy](s->me.scratchpad               , (ref2_y) + (bx>>2) + (by>>2)*(stride)               , stride);\
249
        qpel_avg[1][bxy](s->me.scratchpad + 8           , (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8           , stride);\
250
        qpel_avg[1][bxy](s->me.scratchpad     + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride)     + 8*stride, stride);\
251
        qpel_avg[1][bxy](s->me.scratchpad + 8 + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8 + 8*stride, stride);\
252
    }\
253
    d = cmp_func(s, s->me.scratchpad, src_y, stride);\
254
}else\
255
    d= 256*256*256*32;
256

    
257

    
258
#define CMP_QPEL(d, dx, dy, x, y, size)\
259
    CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
260

    
261
#define CMP(d, x, y, size)\
262
    CMP_DIRECT(d, 0, 0, x, y, size, cmp)
263

    
264
#include "motion_est_template.c"
265
#undef RENAME
266
#undef CMP
267
#undef CMP_HPEL
268
#undef CMP_QPEL
269
#undef INIT
270
#undef CMP__DIRECT
271

    
272

    
273
static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride){
274
    return 0;
275
}
276

    
277
static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){
278
    DSPContext* c= &s->dsp;
279
    int i;
280
    
281
    memset(cmp, 0, sizeof(void*)*11);
282

    
283
    switch(type&0xFF){
284
    case FF_CMP_SAD:
285
        cmp[0]= c->sad[0];
286
        cmp[1]= c->sad[1];
287
        break;
288
    case FF_CMP_SATD:
289
        cmp[0]= c->hadamard8_diff[0];
290
        cmp[1]= c->hadamard8_diff[1];
291
        break;
292
    case FF_CMP_SSE:
293
        cmp[0]= c->sse[0];
294
        cmp[1]= c->sse[1];
295
        break;
296
    case FF_CMP_DCT:
297
        cmp[0]= c->dct_sad[0];
298
        cmp[1]= c->dct_sad[1];
299
        break;
300
    case FF_CMP_PSNR:
301
        cmp[0]= c->quant_psnr[0];
302
        cmp[1]= c->quant_psnr[1];
303
        break;
304
    case FF_CMP_BIT:
305
        cmp[0]= c->bit[0];
306
        cmp[1]= c->bit[1];
307
        break;
308
    case FF_CMP_RD:
309
        cmp[0]= c->rd[0];
310
        cmp[1]= c->rd[1];
311
        break;
312
    case FF_CMP_ZERO:
313
        for(i=0; i<7; i++){
314
            cmp[i]= zero_cmp;
315
        }
316
        break;
317
    default:
318
        fprintf(stderr,"internal error in cmp function selection\n");
319
    }
320
}
321

    
322
static inline int get_penalty_factor(MpegEncContext *s, int type){
323
    switch(type&0xFF){
324
    default:
325
    case FF_CMP_SAD:
326
        return s->qscale*2;
327
    case FF_CMP_DCT:
328
        return s->qscale*3;
329
    case FF_CMP_SATD:
330
        return s->qscale*6;
331
    case FF_CMP_SSE:
332
        return s->qscale*s->qscale*2;
333
    case FF_CMP_BIT:
334
        return 1;
335
    case FF_CMP_RD:
336
    case FF_CMP_PSNR:
337
        return (s->qscale*s->qscale*185 + 64)>>7;
338
    }
339
}
340

    
341
void ff_init_me(MpegEncContext *s){
342
    set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
343
    set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp);
344
    set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
345
    set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp);
346

    
347
    if(s->flags&CODEC_FLAG_QPEL){
348
        if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
349
            s->me.sub_motion_search= simple_chroma_qpel_motion_search;
350
        else
351
            s->me.sub_motion_search= simple_qpel_motion_search;
352
    }else{
353
        if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
354
            s->me.sub_motion_search= simple_chroma_hpel_motion_search;
355
        else if(   s->avctx->me_sub_cmp == FF_CMP_SAD 
356
                && s->avctx->    me_cmp == FF_CMP_SAD 
357
                && s->avctx->    mb_cmp == FF_CMP_SAD)
358
            s->me.sub_motion_search= sad_hpel_motion_search;
359
        else
360
            s->me.sub_motion_search= simple_hpel_motion_search;
361
    }
362

    
363
    if(s->avctx->me_cmp&FF_CMP_CHROMA){
364
        s->me.motion_search[0]= simple_chroma_epzs_motion_search;
365
        s->me.motion_search[1]= simple_chroma_epzs_motion_search4;
366
    }else{
367
        s->me.motion_search[0]= simple_epzs_motion_search;
368
        s->me.motion_search[1]= simple_epzs_motion_search4;
369
    }
370
    
371
    if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){
372
        s->me.pre_motion_search= simple_chroma_epzs_motion_search;
373
    }else{
374
        s->me.pre_motion_search= simple_epzs_motion_search;
375
    }
376
    
377
    if(s->flags&CODEC_FLAG_QPEL){
378
        if(s->avctx->mb_cmp&FF_CMP_CHROMA)
379
            s->me.get_mb_score= simple_chroma_qpel_get_mb_score;
380
        else
381
            s->me.get_mb_score= simple_qpel_get_mb_score;
382
    }else{
383
        if(s->avctx->mb_cmp&FF_CMP_CHROMA)
384
            s->me.get_mb_score= simple_chroma_hpel_get_mb_score;
385
        else
386
            s->me.get_mb_score= simple_hpel_get_mb_score;
387
    }
388
}
389
      
390
static int pix_dev(UINT8 * pix, int line_size, int mean)
391
{
392
    int s, i, j;
393

    
394
    s = 0;
395
    for (i = 0; i < 16; i++) {
396
        for (j = 0; j < 16; j += 8) {
397
            s += ABS(pix[0]-mean);
398
            s += ABS(pix[1]-mean);
399
            s += ABS(pix[2]-mean);
400
            s += ABS(pix[3]-mean);
401
            s += ABS(pix[4]-mean);
402
            s += ABS(pix[5]-mean);
403
            s += ABS(pix[6]-mean);
404
            s += ABS(pix[7]-mean);
405
            pix += 8;
406
        }
407
        pix += line_size - 16;
408
    }
409
    return s;
410
}
411

    
412
static inline void no_motion_search(MpegEncContext * s,
413
                                    int *mx_ptr, int *my_ptr)
414
{
415
    *mx_ptr = 16 * s->mb_x;
416
    *my_ptr = 16 * s->mb_y;
417
}
418

    
419
static int full_motion_search(MpegEncContext * s,
420
                              int *mx_ptr, int *my_ptr, int range,
421
                              int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
422
{
423
    int x1, y1, x2, y2, xx, yy, x, y;
424
    int mx, my, dmin, d;
425
    UINT8 *pix;
426

    
427
    xx = 16 * s->mb_x;
428
    yy = 16 * s->mb_y;
429
    x1 = xx - range + 1;        /* we loose one pixel to avoid boundary pb with half pixel pred */
430
    if (x1 < xmin)
431
        x1 = xmin;
432
    x2 = xx + range - 1;
433
    if (x2 > xmax)
434
        x2 = xmax;
435
    y1 = yy - range + 1;
436
    if (y1 < ymin)
437
        y1 = ymin;
438
    y2 = yy + range - 1;
439
    if (y2 > ymax)
440
        y2 = ymax;
441
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
442
    dmin = 0x7fffffff;
443
    mx = 0;
444
    my = 0;
445
    for (y = y1; y <= y2; y++) {
446
        for (x = x1; x <= x2; x++) {
447
            d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
448
                             s->linesize);
449
            if (d < dmin ||
450
                (d == dmin &&
451
                 (abs(x - xx) + abs(y - yy)) <
452
                 (abs(mx - xx) + abs(my - yy)))) {
453
                dmin = d;
454
                mx = x;
455
                my = y;
456
            }
457
        }
458
    }
459

    
460
    *mx_ptr = mx;
461
    *my_ptr = my;
462

    
463
#if 0
464
    if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
465
        *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
466
        fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
467
    }
468
#endif
469
    return dmin;
470
}
471

    
472

    
473
static int log_motion_search(MpegEncContext * s,
474
                             int *mx_ptr, int *my_ptr, int range,
475
                             int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
476
{
477
    int x1, y1, x2, y2, xx, yy, x, y;
478
    int mx, my, dmin, d;
479
    UINT8 *pix;
480

    
481
    xx = s->mb_x << 4;
482
    yy = s->mb_y << 4;
483

    
484
    /* Left limit */
485
    x1 = xx - range;
486
    if (x1 < xmin)
487
        x1 = xmin;
488

    
489
    /* Right limit */
490
    x2 = xx + range;
491
    if (x2 > xmax)
492
        x2 = xmax;
493

    
494
    /* Upper limit */
495
    y1 = yy - range;
496
    if (y1 < ymin)
497
        y1 = ymin;
498

    
499
    /* Lower limit */
500
    y2 = yy + range;
501
    if (y2 > ymax)
502
        y2 = ymax;
503

    
504
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
505
    dmin = 0x7fffffff;
506
    mx = 0;
507
    my = 0;
508

    
509
    do {
510
        for (y = y1; y <= y2; y += range) {
511
            for (x = x1; x <= x2; x += range) {
512
                d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
513
                if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
514
                    dmin = d;
515
                    mx = x;
516
                    my = y;
517
                }
518
            }
519
        }
520

    
521
        range = range >> 1;
522

    
523
        x1 = mx - range;
524
        if (x1 < xmin)
525
            x1 = xmin;
526

    
527
        x2 = mx + range;
528
        if (x2 > xmax)
529
            x2 = xmax;
530

    
531
        y1 = my - range;
532
        if (y1 < ymin)
533
            y1 = ymin;
534

    
535
        y2 = my + range;
536
        if (y2 > ymax)
537
            y2 = ymax;
538

    
539
    } while (range >= 1);
540

    
541
#ifdef DEBUG
542
    fprintf(stderr, "log       - MX: %d\tMY: %d\n", mx, my);
543
#endif
544
    *mx_ptr = mx;
545
    *my_ptr = my;
546
    return dmin;
547
}
548

    
549
static int phods_motion_search(MpegEncContext * s,
550
                               int *mx_ptr, int *my_ptr, int range,
551
                               int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
552
{
553
    int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
554
    int mx, my, dminx, dminy;
555
    UINT8 *pix;
556

    
557
    xx = s->mb_x << 4;
558
    yy = s->mb_y << 4;
559

    
560
    /* Left limit */
561
    x1 = xx - range;
562
    if (x1 < xmin)
563
        x1 = xmin;
564

    
565
    /* Right limit */
566
    x2 = xx + range;
567
    if (x2 > xmax)
568
        x2 = xmax;
569

    
570
    /* Upper limit */
571
    y1 = yy - range;
572
    if (y1 < ymin)
573
        y1 = ymin;
574

    
575
    /* Lower limit */
576
    y2 = yy + range;
577
    if (y2 > ymax)
578
        y2 = ymax;
579

    
580
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
581
    mx = 0;
582
    my = 0;
583

    
584
    x = xx;
585
    y = yy;
586
    do {
587
        dminx = 0x7fffffff;
588
        dminy = 0x7fffffff;
589

    
590
        lastx = x;
591
        for (x = x1; x <= x2; x += range) {
592
            d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
593
            if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
594
                dminx = d;
595
                mx = x;
596
            }
597
        }
598

    
599
        x = lastx;
600
        for (y = y1; y <= y2; y += range) {
601
            d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
602
            if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
603
                dminy = d;
604
                my = y;
605
            }
606
        }
607

    
608
        range = range >> 1;
609

    
610
        x = mx;
611
        y = my;
612
        x1 = mx - range;
613
        if (x1 < xmin)
614
            x1 = xmin;
615

    
616
        x2 = mx + range;
617
        if (x2 > xmax)
618
            x2 = xmax;
619

    
620
        y1 = my - range;
621
        if (y1 < ymin)
622
            y1 = ymin;
623

    
624
        y2 = my + range;
625
        if (y2 > ymax)
626
            y2 = ymax;
627

    
628
    } while (range >= 1);
629

    
630
#ifdef DEBUG
631
    fprintf(stderr, "phods     - MX: %d\tMY: %d\n", mx, my);
632
#endif
633

    
634
    /* half pixel search */
635
    *mx_ptr = mx;
636
    *my_ptr = my;
637
    return dminy;
638
}
639

    
640

    
641
#define Z_THRESHOLD 256
642

    
643
#define CHECK_SAD_HALF_MV(suffix, x, y) \
644
{\
645
    d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
646
    d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
647
    COPY3_IF_LT(dminh, d, dx, x, dy, y)\
648
}
649

    
650
static inline int sad_hpel_motion_search(MpegEncContext * s,
651
                                  int *mx_ptr, int *my_ptr, int dmin,
652
                                  int xmin, int ymin, int xmax, int ymax,
653
                                  int pred_x, int pred_y, Picture *picture,
654
                                  int n, int size, uint16_t * const mv_penalty)
655
{
656
    uint8_t *ref_picture= picture->data[0];
657
    uint32_t *score_map= s->me.score_map;
658
    const int penalty_factor= s->me.sub_penalty_factor;
659
    int mx, my, xx, yy, dminh;
660
    UINT8 *pix, *ptr;
661
    op_pixels_abs_func pix_abs_x2;
662
    op_pixels_abs_func pix_abs_y2;
663
    op_pixels_abs_func pix_abs_xy2;
664
    
665
    if(size==0){
666
        pix_abs_x2 = s->dsp.pix_abs16x16_x2;
667
        pix_abs_y2 = s->dsp.pix_abs16x16_y2;
668
        pix_abs_xy2= s->dsp.pix_abs16x16_xy2;
669
    }else{
670
        pix_abs_x2 = s->dsp.pix_abs8x8_x2;
671
        pix_abs_y2 = s->dsp.pix_abs8x8_y2;
672
        pix_abs_xy2= s->dsp.pix_abs8x8_xy2;
673
    }
674

    
675
    if(s->me.skip){
676
//    printf("S");
677
        *mx_ptr = 0;
678
        *my_ptr = 0;
679
        return dmin;
680
    }
681
//    printf("N");
682
        
683
    xx = 16 * s->mb_x + 8*(n&1);
684
    yy = 16 * s->mb_y + 8*(n>>1);
685
    pix =  s->new_picture.data[0] + (yy * s->linesize) + xx;
686

    
687
    mx = *mx_ptr;
688
    my = *my_ptr;
689
    ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
690
    
691
    dminh = dmin;
692

    
693
    if (mx > xmin && mx < xmax && 
694
        my > ymin && my < ymax) {
695
        int dx=0, dy=0;
696
        int d, pen_x, pen_y; 
697
        const int index= (my<<ME_MAP_SHIFT) + mx;
698
        const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
699
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
700
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
701
        const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
702
        mx<<=1;
703
        my<<=1;
704

    
705
        
706
        pen_x= pred_x + mx;
707
        pen_y= pred_y + my;
708

    
709
        ptr-= s->linesize;
710
        if(t<=b){
711
            CHECK_SAD_HALF_MV(y2 , 0, -1)
712
            if(l<=r){
713
                CHECK_SAD_HALF_MV(xy2, -1, -1)
714
                if(t+r<=b+l){
715
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
716
                    ptr+= s->linesize;
717
                }else{
718
                    ptr+= s->linesize;
719
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
720
                }
721
                CHECK_SAD_HALF_MV(x2 , -1,  0)
722
            }else{
723
                CHECK_SAD_HALF_MV(xy2, +1, -1)
724
                if(t+l<=b+r){
725
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
726
                    ptr+= s->linesize;
727
                }else{
728
                    ptr+= s->linesize;
729
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
730
                }
731
                CHECK_SAD_HALF_MV(x2 , +1,  0)
732
            }
733
        }else{
734
            if(l<=r){
735
                if(t+l<=b+r){
736
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
737
                    ptr+= s->linesize;
738
                }else{
739
                    ptr+= s->linesize;
740
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
741
                }
742
                CHECK_SAD_HALF_MV(x2 , -1,  0)
743
                CHECK_SAD_HALF_MV(xy2, -1, +1)
744
            }else{
745
                if(t+r<=b+l){
746
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
747
                    ptr+= s->linesize;
748
                }else{
749
                    ptr+= s->linesize;
750
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
751
                }
752
                CHECK_SAD_HALF_MV(x2 , +1,  0)
753
                CHECK_SAD_HALF_MV(xy2, +1, +1)
754
            }
755
            CHECK_SAD_HALF_MV(y2 ,  0, +1)
756
        }
757
        mx+=dx;
758
        my+=dy;
759

    
760
    }else{
761
        mx<<=1;
762
        my<<=1;
763
    }
764

    
765
    *mx_ptr = mx;
766
    *my_ptr = my;
767
    return dminh;
768
}
769

    
770
static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
771
{
772
    const int xy= s->mb_x + 1 + (s->mb_y + 1)*(s->mb_width + 2);
773
    
774
    s->p_mv_table[xy][0] = mx;
775
    s->p_mv_table[xy][1] = my;
776

    
777
    /* has allready been set to the 4 MV if 4MV is done */
778
    if(mv4){
779
        int mot_xy= s->block_index[0];
780

    
781
        s->motion_val[mot_xy  ][0]= mx;
782
        s->motion_val[mot_xy  ][1]= my;
783
        s->motion_val[mot_xy+1][0]= mx;
784
        s->motion_val[mot_xy+1][1]= my;
785

    
786
        mot_xy += s->block_wrap[0];
787
        s->motion_val[mot_xy  ][0]= mx;
788
        s->motion_val[mot_xy  ][1]= my;
789
        s->motion_val[mot_xy+1][0]= mx;
790
        s->motion_val[mot_xy+1][1]= my;
791
    }
792
}
793

    
794
static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax, int f_code)
795
{
796
    *range = 8 * (1 << (f_code - 1));
797
    /* XXX: temporary kludge to avoid overflow for msmpeg4 */
798
    if (s->out_format == FMT_H263 && !s->h263_msmpeg4)
799
        *range *= 2;
800

    
801
    if (s->unrestricted_mv) {
802
        *xmin = -16;
803
        *ymin = -16;
804
        if (s->h263_plus)
805
            *range *= 2;
806
        if(s->avctx->codec->id!=CODEC_ID_MPEG4){
807
            *xmax = s->mb_width*16;
808
            *ymax = s->mb_height*16;
809
        }else {
810
            *xmax = s->width;
811
            *ymax = s->height;
812
        }
813
    } else {
814
        *xmin = 0;
815
        *ymin = 0;
816
        *xmax = s->mb_width*16 - 16;
817
        *ymax = s->mb_height*16 - 16;
818
    }
819
}
820

    
821
static inline int h263_mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
822
{
823
    int block;
824
    int P[10][2];
825
    int dmin_sum=0, mx4_sum=0, my4_sum=0;
826
    uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
827

    
828
    for(block=0; block<4; block++){
829
        int mx4, my4;
830
        int pred_x4, pred_y4;
831
        int dmin4;
832
        static const int off[4]= {2, 1, 1, -1};
833
        const int mot_stride = s->block_wrap[0];
834
        const int mot_xy = s->block_index[block];
835
//        const int block_x= (block&1);
836
//        const int block_y= (block>>1);
837
#if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way
838
        const int rel_xmin4= xmin;
839
        const int rel_xmax4= xmax;
840
        const int rel_ymin4= ymin;
841
        const int rel_ymax4= ymax;
842
#else
843
        const int rel_xmin4= xmin - block_x*8;
844
        const int rel_xmax4= xmax - block_x*8 + 8;
845
        const int rel_ymin4= ymin - block_y*8;
846
        const int rel_ymax4= ymax - block_y*8 + 8;
847
#endif
848
        P_LEFT[0] = s->motion_val[mot_xy - 1][0];
849
        P_LEFT[1] = s->motion_val[mot_xy - 1][1];
850

    
851
        if(P_LEFT[0]       > (rel_xmax4<<shift)) P_LEFT[0]       = (rel_xmax4<<shift);
852

    
853
        /* special case for first line */
854
        if (s->mb_y == 0 && block<2) {
855
            pred_x4= P_LEFT[0];
856
            pred_y4= P_LEFT[1];
857
        } else {
858
            P_TOP[0]      = s->motion_val[mot_xy - mot_stride             ][0];
859
            P_TOP[1]      = s->motion_val[mot_xy - mot_stride             ][1];
860
            P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
861
            P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
862
            if(P_TOP[1]      > (rel_ymax4<<shift)) P_TOP[1]     = (rel_ymax4<<shift);
863
            if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);
864
            if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);
865
            if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);
866
    
867
            P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
868
            P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
869

    
870
//            if(s->out_format == FMT_H263){
871
                pred_x4 = P_MEDIAN[0];
872
                pred_y4 = P_MEDIAN[1];
873
#if 0
874
            }else { /* mpeg1 at least */
875
                pred_x4= P_LEFT[0];
876
                pred_y4= P_LEFT[1];
877
            }
878
#endif
879
        }
880
        P_MV1[0]= mx;
881
        P_MV1[1]= my;
882

    
883
        dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
884
                                       &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
885

    
886
        dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
887
                                          pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);
888
        
889
        if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
890
            int dxy;
891
            const int offset= ((block&1) + (block>>1)*s->linesize)*8;
892
            uint8_t *dest_y = s->me.scratchpad + offset;
893

    
894
            if(s->quarter_sample){
895
                uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>2)) + (s->mb_y*16 + (my4>>2))*s->linesize + offset;
896
                dxy = ((my4 & 3) << 2) | (mx4 & 3);
897

    
898
                if(s->no_rounding)
899
                    s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , s->linesize);
900
                else
901
                    s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , s->linesize);
902
            }else{
903
                uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>1)) + (s->mb_y*16 + (my4>>1))*s->linesize + offset;
904
                dxy = ((my4 & 1) << 1) | (mx4 & 1);
905

    
906
                if(s->no_rounding)
907
                    s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , s->linesize, 8);
908
                else
909
                    s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , s->linesize, 8);
910
            }
911
            dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
912
        }else
913
            dmin_sum+= dmin4;
914

    
915
        if(s->quarter_sample){
916
            mx4_sum+= mx4/2;
917
            my4_sum+= my4/2;
918
        }else{
919
            mx4_sum+= mx4;
920
            my4_sum+= my4;
921
        }
922
            
923
        s->motion_val[ s->block_index[block] ][0]= mx4;
924
        s->motion_val[ s->block_index[block] ][1]= my4;
925
    }
926
    
927
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
928
        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);
929
    }
930
    
931
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
932
        int dxy;
933
        int mx, my;
934
        int offset;
935

    
936
        mx= ff_h263_round_chroma(mx4_sum);
937
        my= ff_h263_round_chroma(my4_sum);
938
        dxy = ((my & 1) << 1) | (mx & 1);
939
        
940
        offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
941
       
942
        if(s->no_rounding){
943
            s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
944
            s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
945
        }else{
946
            s->dsp.put_pixels_tab       [1][dxy](s->me.scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
947
            s->dsp.put_pixels_tab       [1][dxy](s->me.scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
948
        }
949

    
950
        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);
951
        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);
952
    }
953

    
954
    switch(s->avctx->mb_cmp&0xFF){
955
    /*case FF_CMP_SSE:
956
        return dmin_sum+ 32*s->qscale*s->qscale;*/
957
    case FF_CMP_RD:
958
        return dmin_sum;
959
    default:
960
        return dmin_sum+ 11*s->me.mb_penalty_factor;
961
    }
962
}
963

    
964
void ff_estimate_p_frame_motion(MpegEncContext * s,
965
                                int mb_x, int mb_y)
966
{
967
    UINT8 *pix, *ppix;
968
    int sum, varc, vard, mx, my, range, dmin, xx, yy;
969
    int xmin, ymin, xmax, ymax;
970
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
971
    int pred_x=0, pred_y=0;
972
    int P[10][2];
973
    const int shift= 1+s->quarter_sample;
974
    int mb_type=0;
975
    uint8_t *ref_picture= s->last_picture.data[0];
976
    Picture * const pic= &s->current_picture;
977
    uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
978
    
979
    assert(s->quarter_sample==0 || s->quarter_sample==1);
980

    
981
    s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
982
    s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
983
    s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
984

    
985
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
986
    rel_xmin= xmin - mb_x*16;
987
    rel_xmax= xmax - mb_x*16;
988
    rel_ymin= ymin - mb_y*16;
989
    rel_ymax= ymax - mb_y*16;
990
    s->me.skip=0;
991

    
992
    switch(s->me_method) {
993
    case ME_ZERO:
994
    default:
995
        no_motion_search(s, &mx, &my);
996
        mx-= mb_x*16;
997
        my-= mb_y*16;
998
        dmin = 0;
999
        break;
1000
    case ME_FULL:
1001
        dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1002
        mx-= mb_x*16;
1003
        my-= mb_y*16;
1004
        break;
1005
    case ME_LOG:
1006
        dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1007
        mx-= mb_x*16;
1008
        my-= mb_y*16;
1009
        break;
1010
    case ME_PHODS:
1011
        dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1012
        mx-= mb_x*16;
1013
        my-= mb_y*16;
1014
        break;
1015
    case ME_X1:
1016
    case ME_EPZS:
1017
       {
1018
            const int mot_stride = s->block_wrap[0];
1019
            const int mot_xy = s->block_index[0];
1020

    
1021
            P_LEFT[0]       = s->motion_val[mot_xy - 1][0];
1022
            P_LEFT[1]       = s->motion_val[mot_xy - 1][1];
1023

    
1024
            if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);
1025

    
1026
            if(mb_y) {
1027
                P_TOP[0]      = s->motion_val[mot_xy - mot_stride    ][0];
1028
                P_TOP[1]      = s->motion_val[mot_xy - mot_stride    ][1];
1029
                P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];
1030
                P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];
1031
                if(P_TOP[1]      > (rel_ymax<<shift)) P_TOP[1]     = (rel_ymax<<shift);
1032
                if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1033
                if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1034
        
1035
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1036
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1037

    
1038
                if(s->out_format == FMT_H263){
1039
                    pred_x = P_MEDIAN[0];
1040
                    pred_y = P_MEDIAN[1];
1041
                }else { /* mpeg1 at least */
1042
                    pred_x= P_LEFT[0];
1043
                    pred_y= P_LEFT[1];
1044
                }
1045
            }else{
1046
                pred_x= P_LEFT[0];
1047
                pred_y= P_LEFT[1];
1048
            }
1049

    
1050
        }
1051
        dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1052
                                      &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1053
 
1054
        break;
1055
    }
1056

    
1057
    /* intra / predictive decision */
1058
    xx = mb_x * 16;
1059
    yy = mb_y * 16;
1060

    
1061
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
1062
    /* At this point (mx,my) are full-pell and the relative displacement */
1063
    ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
1064
    
1065
    sum = s->dsp.pix_sum(pix, s->linesize);
1066
    
1067
    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1068
    vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
1069

    
1070
//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);
1071
    pic->mb_var   [s->mb_width * mb_y + mb_x] = varc;
1072
    pic->mc_mb_var[s->mb_width * mb_y + mb_x] = vard;
1073
    pic->mb_mean  [s->mb_width * mb_y + mb_x] = (sum+128)>>8;
1074
//    pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin; 
1075
    pic->mb_var_sum    += varc;
1076
    pic->mc_mb_var_sum += vard;
1077
//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);
1078
    
1079
#if 0
1080
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1081
           varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1082
#endif
1083
    if(s->flags&CODEC_FLAG_HQ){
1084
        if (vard <= 64 || vard < varc)
1085
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1086
        else
1087
            s->scene_change_score+= s->qscale;
1088

    
1089
        if (vard*2 + 200 > varc)
1090
            mb_type|= MB_TYPE_INTRA;
1091
        if (varc*2 + 200 > vard){
1092
            mb_type|= MB_TYPE_INTER;
1093
            s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1094
                                   pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1095
        }else{
1096
            mx <<=shift;
1097
            my <<=shift;
1098
        }
1099
        if((s->flags&CODEC_FLAG_4MV)
1100
           && !s->me.skip && varc>50 && vard>10){
1101
            h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1102
            mb_type|=MB_TYPE_INTER4V;
1103

    
1104
            set_p_mv_tables(s, mx, my, 0);
1105
        }else
1106
            set_p_mv_tables(s, mx, my, 1);
1107
    }else{
1108
        mb_type= MB_TYPE_INTER;
1109

    
1110
        dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1111
                                    pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1112
        
1113
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1114
            dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);
1115

    
1116
        if((s->flags&CODEC_FLAG_4MV)
1117
           && !s->me.skip && varc>50 && vard>10){
1118
            int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1119
            if(dmin4 < dmin){
1120
                mb_type= MB_TYPE_INTER4V;
1121
                dmin=dmin4;
1122
            }
1123
        }
1124
        pic->mb_cmp_score[s->mb_width * mb_y + mb_x] = dmin; 
1125
        set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
1126
        
1127
        if (vard <= 64 || vard < varc) {
1128
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1129
        }else{
1130
            s->scene_change_score+= s->qscale;
1131
        }
1132
    }
1133

    
1134
    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1135
}
1136

    
1137
int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1138
                                    int mb_x, int mb_y)
1139
{
1140
    int mx, my, range, dmin;
1141
    int xmin, ymin, xmax, ymax;
1142
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1143
    int pred_x=0, pred_y=0;
1144
    int P[10][2];
1145
    const int shift= 1+s->quarter_sample;
1146
    uint16_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1147
    const int mv_stride= s->mb_width + 2;
1148
    const int xy= mb_x + 1 + (mb_y + 1)*mv_stride;
1149
    
1150
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1151

    
1152
    s->me.pre_penalty_factor    = get_penalty_factor(s, s->avctx->me_pre_cmp);
1153

    
1154
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, s->f_code);
1155
    rel_xmin= xmin - mb_x*16;
1156
    rel_xmax= xmax - mb_x*16;
1157
    rel_ymin= ymin - mb_y*16;
1158
    rel_ymax= ymax - mb_y*16;
1159
    s->me.skip=0;
1160

    
1161
    P_LEFT[0]       = s->p_mv_table[xy + 1][0];
1162
    P_LEFT[1]       = s->p_mv_table[xy + 1][1];
1163

    
1164
    if(P_LEFT[0]       < (rel_xmin<<shift)) P_LEFT[0]       = (rel_xmin<<shift);
1165

    
1166
    /* special case for first line */
1167
    if (mb_y == s->mb_height-1) {
1168
        pred_x= P_LEFT[0];
1169
        pred_y= P_LEFT[1];
1170
        P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1171
        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME 
1172
    } else {
1173
        P_TOP[0]      = s->p_mv_table[xy + mv_stride    ][0];
1174
        P_TOP[1]      = s->p_mv_table[xy + mv_stride    ][1];
1175
        P_TOPRIGHT[0] = s->p_mv_table[xy + mv_stride - 1][0];
1176
        P_TOPRIGHT[1] = s->p_mv_table[xy + mv_stride - 1][1];
1177
        if(P_TOP[1]      < (rel_ymin<<shift)) P_TOP[1]     = (rel_ymin<<shift);
1178
        if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
1179
        if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
1180
    
1181
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1182
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1183

    
1184
        pred_x = P_MEDIAN[0];
1185
        pred_y = P_MEDIAN[1];
1186
    }
1187
    dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1188
                                   &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1189

    
1190
    s->p_mv_table[xy][0] = mx<<shift;
1191
    s->p_mv_table[xy][1] = my<<shift;
1192
    
1193
    return dmin;
1194
}
1195

    
1196
int ff_estimate_motion_b(MpegEncContext * s,
1197
                       int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
1198
{
1199
    int mx, my, range, dmin;
1200
    int xmin, ymin, xmax, ymax;
1201
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1202
    int pred_x=0, pred_y=0;
1203
    int P[10][2];
1204
    const int shift= 1+s->quarter_sample;
1205
    const int mot_stride = s->mb_width + 2;
1206
    const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
1207
    uint8_t * const ref_picture= picture->data[0];
1208
    uint16_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1209
    int mv_scale;
1210
        
1211
    s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
1212
    s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1213
    s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1214

    
1215
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax, f_code);
1216
    rel_xmin= xmin - mb_x*16;
1217
    rel_xmax= xmax - mb_x*16;
1218
    rel_ymin= ymin - mb_y*16;
1219
    rel_ymax= ymax - mb_y*16;
1220

    
1221
    switch(s->me_method) {
1222
    case ME_ZERO:
1223
    default:
1224
        no_motion_search(s, &mx, &my);
1225
        dmin = 0;
1226
        mx-= mb_x*16;
1227
        my-= mb_y*16;
1228
        break;
1229
    case ME_FULL:
1230
        dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1231
        mx-= mb_x*16;
1232
        my-= mb_y*16;
1233
        break;
1234
    case ME_LOG:
1235
        dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1236
        mx-= mb_x*16;
1237
        my-= mb_y*16;
1238
        break;
1239
    case ME_PHODS:
1240
        dmin = phods_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_X1:
1245
    case ME_EPZS:
1246
       {
1247
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
1248
            P_LEFT[1]        = mv_table[mot_xy - 1][1];
1249

    
1250
            if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);
1251

    
1252
            /* special case for first line */
1253
            if (mb_y) {
1254
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
1255
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
1256
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
1257
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
1258
                if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);
1259
                if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1260
                if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1261
        
1262
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1263
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1264
            }
1265
            pred_x= P_LEFT[0];
1266
            pred_y= P_LEFT[1];
1267
        }
1268
        
1269
        if(mv_table == s->b_forw_mv_table){
1270
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1271
        }else{
1272
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1273
        }
1274
        
1275
        dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1276
                                      picture, s->p_mv_table, mv_scale, mv_penalty);
1277
 
1278
        break;
1279
    }
1280
    
1281
    dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1282
                                   pred_x, pred_y, picture, 0, 0, mv_penalty);
1283
                                   
1284
    if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1285
        dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);
1286

    
1287
//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1288
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1289
    mv_table[mot_xy][0]= mx;
1290
    mv_table[mot_xy][1]= my;
1291

    
1292
    return dmin;
1293
}
1294

    
1295
static inline int check_bidir_mv(MpegEncContext * s,
1296
                   int mb_x, int mb_y,
1297
                   int motion_fx, int motion_fy,
1298
                   int motion_bx, int motion_by,
1299
                   int pred_fx, int pred_fy,
1300
                   int pred_bx, int pred_by)
1301
{
1302
    //FIXME optimize?
1303
    //FIXME move into template?
1304
    //FIXME better f_code prediction (max mv & distance)
1305
    UINT16 *mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1306
    uint8_t *dest_y = s->me.scratchpad;
1307
    uint8_t *ptr;
1308
    int dxy;
1309
    int src_x, src_y;
1310
    int fbmin;
1311

    
1312
    if(s->quarter_sample){
1313
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1314
        src_x = mb_x * 16 + (motion_fx >> 2);
1315
        src_y = mb_y * 16 + (motion_fy >> 2);
1316
        assert(src_x >=-16 && src_x<=s->width);
1317
        assert(src_y >=-16 && src_y<=s->height);
1318

    
1319
        ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1320
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);
1321

    
1322
        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1323
        src_x = mb_x * 16 + (motion_bx >> 2);
1324
        src_y = mb_y * 16 + (motion_by >> 2);
1325
        assert(src_x >=-16 && src_x<=s->width);
1326
        assert(src_y >=-16 && src_y<=s->height);
1327
    
1328
        ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1329
        s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);
1330
    }else{
1331
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1332
        src_x = mb_x * 16 + (motion_fx >> 1);
1333
        src_y = mb_y * 16 + (motion_fy >> 1);
1334
        assert(src_x >=-16 && src_x<=s->width);
1335
        assert(src_y >=-16 && src_y<=s->height);
1336

    
1337
        ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1338
        s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
1339

    
1340
        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1341
        src_x = mb_x * 16 + (motion_bx >> 1);
1342
        src_y = mb_y * 16 + (motion_by >> 1);
1343
        assert(src_x >=-16 && src_x<=s->width);
1344
        assert(src_y >=-16 && src_y<=s->height);
1345
    
1346
        ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1347
        s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
1348
    }
1349

    
1350
    fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1351
           +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1352
           + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
1353
           
1354
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1355
    }
1356
    //FIXME CHROMA !!!
1357
           
1358
    return fbmin;
1359
}
1360

    
1361
/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1362
static inline int bidir_refine(MpegEncContext * s,
1363
                                  int mb_x, int mb_y)
1364
{
1365
    const int mot_stride = s->mb_width + 2;
1366
    const int xy = (mb_y + 1)*mot_stride + mb_x + 1;
1367
    int fbmin;
1368
    int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1369
    int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1370
    int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1371
    int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1372
    int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1373
    int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1374
    int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1375
    int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1376

    
1377
    //FIXME do refinement and add flag
1378
    
1379
    fbmin= check_bidir_mv(s, mb_x, mb_y, 
1380
                          motion_fx, motion_fy,
1381
                          motion_bx, motion_by,
1382
                          pred_fx, pred_fy,
1383
                          pred_bx, pred_by);
1384

    
1385
   return fbmin;
1386
}
1387

    
1388
static inline int direct_search(MpegEncContext * s,
1389
                                int mb_x, int mb_y)
1390
{
1391
    int P[10][2];
1392
    const int mot_stride = s->mb_width + 2;
1393
    const int mot_xy = (mb_y + 1)*mot_stride + mb_x + 1;
1394
    const int shift= 1+s->quarter_sample;
1395
    int dmin, i;
1396
    const int time_pp= s->pp_time;
1397
    const int time_pb= s->pb_time;
1398
    int mx, my, xmin, xmax, ymin, ymax;
1399
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
1400
    uint16_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1401
    
1402
    ymin= xmin=(-32)>>shift;
1403
    ymax= xmax=   31>>shift;
1404

    
1405
    if(s->co_located_type_table[mb_x + mb_y*s->mb_width]==CO_LOCATED_TYPE_4MV){
1406
        s->mv_type= MV_TYPE_8X8;
1407
    }else{
1408
        s->mv_type= MV_TYPE_16X16;
1409
    }
1410

    
1411
    for(i=0; i<4; i++){
1412
        int index= s->block_index[i];
1413
        int min, max;
1414
    
1415
        s->me.co_located_mv[i][0]= s->motion_val[index][0];
1416
        s->me.co_located_mv[i][1]= s->motion_val[index][1];
1417
        s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1418
        s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1419
//        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);
1420
//        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);
1421

    
1422
        max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1423
        min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1424
        max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding
1425
        min+= (2*mb_x + (i& 1))*8 - 1;
1426
        xmax= FFMIN(xmax, s->width - max);
1427
        xmin= FFMAX(xmin, - 16     - min);
1428

    
1429
        max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1430
        min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1431
        max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding
1432
        min+= (2*mb_y + (i>>1))*8 - 1;
1433
        ymax= FFMIN(ymax, s->height - max);
1434
        ymin= FFMAX(ymin, - 16      - min);
1435
        
1436
        if(s->mv_type == MV_TYPE_16X16) break;
1437
    }
1438
    
1439
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1440
    
1441
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1442
        s->b_direct_mv_table[mot_xy][0]= 0;
1443
        s->b_direct_mv_table[mot_xy][1]= 0;
1444

    
1445
        return 256*256*256*64;
1446
    }
1447

    
1448
    P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1449
    P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1450

    
1451
    /* special case for first line */
1452
    if (mb_y) {
1453
        P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
1454
        P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
1455
        P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
1456
        P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
1457
    
1458
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1459
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1460
    }
1461
 
1462
    //FIXME direct_search  ptr in context!!! (needed for chroma anyway or this will get messy)   
1463
    if(s->flags&CODEC_FLAG_QPEL){
1464
        dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 
1465
                                                     &s->last_picture, mv_table, 1<<14, mv_penalty);
1466
        dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1467
                                                0, 0, &s->last_picture, 0, 0, mv_penalty);
1468
        
1469
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1470
            dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1471
    }else{
1472
        dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 
1473
                                                     &s->last_picture, mv_table, 1<<15, mv_penalty);
1474
        dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1475
                                                0, 0, &s->last_picture, 0, 0, mv_penalty);
1476

    
1477
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1478
            dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1479
    }
1480

    
1481
    s->b_direct_mv_table[mot_xy][0]= mx;
1482
    s->b_direct_mv_table[mot_xy][1]= my;
1483
    return dmin;
1484
}
1485

    
1486
void ff_estimate_b_frame_motion(MpegEncContext * s,
1487
                             int mb_x, int mb_y)
1488
{
1489
    const int penalty_factor= s->me.mb_penalty_factor;
1490
    int fmin, bmin, dmin, fbmin;
1491
    int type=0;
1492
    
1493
    dmin= direct_search(s, mb_x, mb_y);
1494

    
1495
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;
1496
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;
1497
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1498

    
1499
    fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1500
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1501
    {
1502
        int score= fmin;
1503
        type = MB_TYPE_FORWARD;
1504
        
1505
        // RAL: No MB_TYPE_DIRECT in MPEG-1 video (only MPEG-4)
1506
        if (s->codec_id != CODEC_ID_MPEG1VIDEO && dmin <= score){
1507
            score = dmin;
1508
            type = MB_TYPE_DIRECT;
1509
        }
1510
        if(bmin<score){
1511
            score=bmin;
1512
            type= MB_TYPE_BACKWARD; 
1513
        }
1514
        if(fbmin<score){
1515
            score=fbmin;
1516
            type= MB_TYPE_BIDIR;
1517
        }
1518
        
1519
        score= ((unsigned)(score*score + 128*256))>>16;
1520
        s->current_picture.mc_mb_var_sum += score;
1521
        s->current_picture.mc_mb_var[mb_y*s->mb_width + mb_x] = score; //FIXME use SSE
1522
    }
1523

    
1524
    if(s->flags&CODEC_FLAG_HQ){
1525
        type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
1526
        if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1527
    }
1528

    
1529
    s->mb_type[mb_y*s->mb_width + mb_x]= type;
1530
}
1531

    
1532
/* find best f_code for ME which do unlimited searches */
1533
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1534
{
1535
    if(s->me_method>=ME_EPZS){
1536
        int score[8];
1537
        int i, y;
1538
        UINT8 * fcode_tab= s->fcode_tab;
1539
        int best_fcode=-1;
1540
        int best_score=-10000000;
1541

    
1542
        for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1543

    
1544
        for(y=0; y<s->mb_height; y++){
1545
            int x;
1546
            int xy= (y+1)* (s->mb_width+2) + 1;
1547
            i= y*s->mb_width;
1548
            for(x=0; x<s->mb_width; x++){
1549
                if(s->mb_type[i] & type){
1550
                    int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1551
                                     fcode_tab[mv_table[xy][1] + MAX_MV]);
1552
                    int j;
1553
                    
1554
                    for(j=0; j<fcode && j<8; j++){
1555
                        if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[i] < s->current_picture.mb_var[i])
1556
                            score[j]-= 170;
1557
                    }
1558
                }
1559
                i++;
1560
                xy++;
1561
            }
1562
        }
1563
        
1564
        for(i=1; i<8; i++){
1565
            if(score[i] > best_score){
1566
                best_score= score[i];
1567
                best_fcode= i;
1568
            }
1569
//            printf("%d %d\n", i, score[i]);
1570
        }
1571

    
1572
//    printf("fcode: %d type: %d\n", i, s->pict_type);
1573
        return best_fcode;
1574
/*        for(i=0; i<=MAX_FCODE; i++){
1575
            printf("%d ", mv_num[i]);
1576
        }
1577
        printf("\n");*/
1578
    }else{
1579
        return 1;
1580
    }
1581
}
1582

    
1583
void ff_fix_long_p_mvs(MpegEncContext * s)
1584
{
1585
    const int f_code= s->f_code;
1586
    int y;
1587
    UINT8 * fcode_tab= s->fcode_tab;
1588
//int clip=0;
1589
//int noclip=0;
1590
    /* clip / convert to intra 16x16 type MVs */
1591
    for(y=0; y<s->mb_height; y++){
1592
        int x;
1593
        int xy= (y+1)* (s->mb_width+2)+1;
1594
        int i= y*s->mb_width;
1595
        for(x=0; x<s->mb_width; x++){
1596
            if(s->mb_type[i]&MB_TYPE_INTER){
1597
                if(   fcode_tab[s->p_mv_table[xy][0] + MAX_MV] > f_code
1598
                   || fcode_tab[s->p_mv_table[xy][0] + MAX_MV] == 0
1599
                   || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] > f_code
1600
                   || fcode_tab[s->p_mv_table[xy][1] + MAX_MV] == 0 ){
1601
                    s->mb_type[i] &= ~MB_TYPE_INTER;
1602
                    s->mb_type[i] |= MB_TYPE_INTRA;
1603
                    s->p_mv_table[xy][0] = 0;
1604
                    s->p_mv_table[xy][1] = 0;
1605
//clip++;
1606
                }
1607
//else
1608
//  noclip++;
1609
            }
1610
            xy++;
1611
            i++;
1612
        }
1613
    }
1614
//printf("%d no:%d %d//\n", clip, noclip, f_code);
1615
    if(s->flags&CODEC_FLAG_4MV){
1616
        const int wrap= 2+ s->mb_width*2;
1617

    
1618
        /* clip / convert to intra 8x8 type MVs */
1619
        for(y=0; y<s->mb_height; y++){
1620
            int xy= (y*2 + 1)*wrap + 1;
1621
            int i= y*s->mb_width;
1622
            int x;
1623

    
1624
            for(x=0; x<s->mb_width; x++){
1625
                if(s->mb_type[i]&MB_TYPE_INTER4V){
1626
                    int block;
1627
                    for(block=0; block<4; block++){
1628
                        int off= (block& 1) + (block>>1)*wrap;
1629
                        int mx= s->motion_val[ xy + off ][0];
1630
                        int my= s->motion_val[ xy + off ][1];
1631

    
1632
                        if(   fcode_tab[mx + MAX_MV] > f_code
1633
                           || fcode_tab[mx + MAX_MV] == 0
1634
                           || fcode_tab[my + MAX_MV] > f_code
1635
                           || fcode_tab[my + MAX_MV] == 0 ){
1636
                            s->mb_type[i] &= ~MB_TYPE_INTER4V;
1637
                            s->mb_type[i] |= MB_TYPE_INTRA;
1638
                        }
1639
                    }
1640
                }
1641
                xy+=2;
1642
                i++;
1643
            }
1644
        }
1645
    }
1646
}
1647

    
1648
void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
1649
{
1650
    int y;
1651
    UINT8 * fcode_tab= s->fcode_tab;
1652

    
1653
    // RAL: 8 in MPEG-1, 16 in MPEG-4
1654
    int range = (((s->codec_id == CODEC_ID_MPEG1VIDEO) ? 8 : 16) << f_code);
1655

    
1656
    /* clip / convert to intra 16x16 type MVs */
1657
    for(y=0; y<s->mb_height; y++){
1658
        int x;
1659
        int xy= (y+1)* (s->mb_width+2)+1;
1660
        int i= y*s->mb_width;
1661
        for(x=0; x<s->mb_width; x++)
1662
            {
1663
            if (s->mb_type[i] & type)    // RAL: "type" test added...
1664
                {
1665
                if (fcode_tab[mv_table[xy][0] + MAX_MV] > f_code || fcode_tab[mv_table[xy][0] + MAX_MV] == 0)
1666
                    {
1667
                    if(mv_table[xy][0]>0) 
1668
                        mv_table[xy][0]=  range-1;
1669
                    else
1670
                        mv_table[xy][0]= -range;
1671
                    }
1672
                if (fcode_tab[mv_table[xy][1] + MAX_MV] > f_code || fcode_tab[mv_table[xy][1] + MAX_MV] == 0)
1673
                    {
1674
                    if(mv_table[xy][1]>0) 
1675
                        mv_table[xy][1]=  range-1;
1676
                    else                  
1677
                        mv_table[xy][1]= -range;
1678
            }
1679
            }
1680
            xy++;
1681
            i++;
1682
        }
1683
    }
1684
}