Statistics
| Branch: | Revision:

ffmpeg / libavcodec / motion_est.c @ 029911d1

History | View | Annotate | Download (57.3 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 "avcodec.h"
32
#include "dsputil.h"
33
#include "mpegvideo.h"
34

    
35
#undef NDEBUG
36
#include <assert.h>
37

    
38
#define SQ(a) ((a)*(a))
39

    
40
#define P_LEFT P[1]
41
#define P_TOP P[2]
42
#define P_TOPRIGHT P[3]
43
#define P_MEDIAN P[4]
44
#define P_MV1 P[9]
45

    
46
static inline int sad_hpel_motion_search(MpegEncContext * s,
47
                                  int *mx_ptr, int *my_ptr, int dmin,
48
                                  int xmin, int ymin, int xmax, int ymax,
49
                                  int pred_x, int pred_y, Picture *picture,
50
                                  int n, int size, uint8_t * const mv_penalty);
51

    
52
static inline int update_map_generation(MpegEncContext * s)
53
{
54
    s->me.map_generation+= 1<<(ME_MAP_MV_BITS*2);
55
    if(s->me.map_generation==0){
56
        s->me.map_generation= 1<<(ME_MAP_MV_BITS*2);
57
        memset(s->me.map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
58
    }
59
    return s->me.map_generation;
60
}
61

    
62
/* shape adaptive search stuff */
63
typedef struct Minima{
64
    int height;
65
    int x, y;
66
    int checked;
67
}Minima;
68

    
69
static int minima_cmp(const void *a, const void *b){
70
    const Minima *da = (const Minima *) a;
71
    const Minima *db = (const Minima *) b;
72
    
73
    return da->height - db->height;
74
}
75
                                  
76
/* SIMPLE */
77
#define RENAME(a) simple_ ## a
78

    
79
#define CMP(d, x, y, size)\
80
d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);
81

    
82
#define CMP_HPEL(d, dx, dy, x, y, size)\
83
{\
84
    const int dxy= (dx) + 2*(dy);\
85
    hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\
86
    d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
87
}
88

    
89
#define CMP_QPEL(d, dx, dy, x, y, size)\
90
{\
91
    const int dxy= (dx) + 4*(dy);\
92
    qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
93
    d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
94
}
95

    
96
#include "motion_est_template.c"
97
#undef RENAME
98
#undef CMP
99
#undef CMP_HPEL
100
#undef CMP_QPEL
101
#undef INIT
102

    
103
/* SIMPLE CHROMA */
104
#define RENAME(a) simple_chroma_ ## a
105

    
106
#define CMP(d, x, y, size)\
107
d = cmp(s, src_y, (ref_y) + (x) + (y)*(stride), stride);\
108
if(chroma_cmp){\
109
    int dxy= ((x)&1) + 2*((y)&1);\
110
    int c= ((x)>>1) + ((y)>>1)*uvstride;\
111
\
112
    chroma_hpel_put[0][dxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
113
    d += chroma_cmp(s, s->me.scratchpad, src_u, uvstride);\
114
    chroma_hpel_put[0][dxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
115
    d += chroma_cmp(s, s->me.scratchpad, src_v, uvstride);\
116
}
117

    
118
#define CMP_HPEL(d, dx, dy, x, y, size)\
119
{\
120
    const int dxy= (dx) + 2*(dy);\
121
    hpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride, (16>>size));\
122
    d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
123
    if(chroma_cmp_sub){\
124
        int cxy= (dxy) | ((x)&1) | (2*((y)&1));\
125
        int c= ((x)>>1) + ((y)>>1)*uvstride;\
126
        chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
127
        d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
128
        chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
129
        d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
130
    }\
131
}
132

    
133
#define CMP_QPEL(d, dx, dy, x, y, size)\
134
{\
135
    const int dxy= (dx) + 4*(dy);\
136
    qpel_put[0][dxy](s->me.scratchpad, (ref_y) + (x) + (y)*(stride), stride);\
137
    d = cmp_sub(s, s->me.scratchpad, src_y, stride);\
138
    if(chroma_cmp_sub){\
139
        int cxy, c;\
140
        int cx= (4*(x) + (dx))/2;\
141
        int cy= (4*(y) + (dy))/2;\
142
        cx= (cx>>1)|(cx&1);\
143
        cy= (cy>>1)|(cy&1);\
144
        cxy= (cx&1) + 2*(cy&1);\
145
        c= ((cx)>>1) + ((cy)>>1)*uvstride;\
146
        chroma_hpel_put[0][cxy](s->me.scratchpad, ref_u + c, uvstride, 8);\
147
        d += chroma_cmp_sub(s, s->me.scratchpad, src_u, uvstride);\
148
        chroma_hpel_put[0][cxy](s->me.scratchpad, ref_v + c, uvstride, 8);\
149
        d += chroma_cmp_sub(s, s->me.scratchpad, src_v, uvstride);\
150
    }\
151
}
152

    
153
#include "motion_est_template.c"
154
#undef RENAME
155
#undef CMP
156
#undef CMP_HPEL
157
#undef CMP_QPEL
158
#undef INIT
159

    
160
/* SIMPLE DIRECT HPEL */
161
#define RENAME(a) simple_direct_hpel_ ## a
162
//FIXME precalc divisions stuff
163

    
164
#define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
165
if((x) >= xmin && 2*(x) + (dx) <= 2*xmax && (y) >= ymin && 2*(y) + (dy) <= 2*ymax){\
166
    const int hx= 2*(x) + (dx);\
167
    const int hy= 2*(y) + (dy);\
168
    if(s->mv_type==MV_TYPE_8X8){\
169
        int i;\
170
        for(i=0; i<4; i++){\
171
            int fx = s->me.direct_basis_mv[i][0] + hx;\
172
            int fy = s->me.direct_basis_mv[i][1] + hy;\
173
            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;\
174
            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;\
175
            int fxy= (fx&1) + 2*(fy&1);\
176
            int bxy= (bx&1) + 2*(by&1);\
177
\
178
            uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
179
            hpel_put[1][fxy](dst, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 8);\
180
            hpel_avg[1][bxy](dst, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 8);\
181
        }\
182
    }else{\
183
        int fx = s->me.direct_basis_mv[0][0] + hx;\
184
        int fy = s->me.direct_basis_mv[0][1] + hy;\
185
        int bx = hx ? fx - s->me.co_located_mv[0][0] : (s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp);\
186
        int by = hy ? fy - s->me.co_located_mv[0][1] : (s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp);\
187
        int fxy= (fx&1) + 2*(fy&1);\
188
        int bxy= (bx&1) + 2*(by&1);\
189
        \
190
        assert((fx>>1) + 16*s->mb_x >= -16);\
191
        assert((fy>>1) + 16*s->mb_y >= -16);\
192
        assert((fx>>1) + 16*s->mb_x <= s->width);\
193
        assert((fy>>1) + 16*s->mb_y <= s->height);\
194
        assert((bx>>1) + 16*s->mb_x >= -16);\
195
        assert((by>>1) + 16*s->mb_y >= -16);\
196
        assert((bx>>1) + 16*s->mb_x <= s->width);\
197
        assert((by>>1) + 16*s->mb_y <= s->height);\
198
\
199
        hpel_put[0][fxy](s->me.scratchpad, (ref_y ) + (fx>>1) + (fy>>1)*(stride), stride, 16);\
200
        hpel_avg[0][bxy](s->me.scratchpad, (ref2_y) + (bx>>1) + (by>>1)*(stride), stride, 16);\
201
    }\
202
    d = cmp_func(s, s->me.scratchpad, src_y, stride);\
203
}else\
204
    d= 256*256*256*32;
205

    
206

    
207
#define CMP_HPEL(d, dx, dy, x, y, size)\
208
    CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
209

    
210
#define CMP(d, x, y, size)\
211
    CMP_DIRECT(d, 0, 0, x, y, size, cmp)
212
    
213
#include "motion_est_template.c"
214
#undef RENAME
215
#undef CMP
216
#undef CMP_HPEL
217
#undef CMP_QPEL
218
#undef INIT
219
#undef CMP_DIRECT
220

    
221
/* SIMPLE DIRECT QPEL */
222
#define RENAME(a) simple_direct_qpel_ ## a
223

    
224
#define CMP_DIRECT(d, dx, dy, x, y, size, cmp_func)\
225
if((x) >= xmin && 4*(x) + (dx) <= 4*xmax && (y) >= ymin && 4*(y) + (dy) <= 4*ymax){\
226
    const int qx= 4*(x) + (dx);\
227
    const int qy= 4*(y) + (dy);\
228
    if(s->mv_type==MV_TYPE_8X8){\
229
        int i;\
230
        for(i=0; i<4; i++){\
231
            int fx = s->me.direct_basis_mv[i][0] + qx;\
232
            int fy = s->me.direct_basis_mv[i][1] + qy;\
233
            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;\
234
            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;\
235
            int fxy= (fx&3) + 4*(fy&3);\
236
            int bxy= (bx&3) + 4*(by&3);\
237
\
238
            uint8_t *dst= s->me.scratchpad + 8*(i&1) + 8*stride*(i>>1);\
239
            qpel_put[1][fxy](dst, (ref_y ) + (fx>>2) + (fy>>2)*(stride), stride);\
240
            qpel_avg[1][bxy](dst, (ref2_y) + (bx>>2) + (by>>2)*(stride), stride);\
241
        }\
242
    }else{\
243
        int fx = s->me.direct_basis_mv[0][0] + qx;\
244
        int fy = s->me.direct_basis_mv[0][1] + qy;\
245
        int bx = qx ? fx - s->me.co_located_mv[0][0] : s->me.co_located_mv[0][0]*(time_pb - time_pp)/time_pp;\
246
        int by = qy ? fy - s->me.co_located_mv[0][1] : s->me.co_located_mv[0][1]*(time_pb - time_pp)/time_pp;\
247
        int fxy= (fx&3) + 4*(fy&3);\
248
        int bxy= (bx&3) + 4*(by&3);\
249
\
250
        qpel_put[1][fxy](s->me.scratchpad               , (ref_y ) + (fx>>2) + (fy>>2)*(stride)               , stride);\
251
        qpel_put[1][fxy](s->me.scratchpad + 8           , (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8           , stride);\
252
        qpel_put[1][fxy](s->me.scratchpad     + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride)     + 8*stride, stride);\
253
        qpel_put[1][fxy](s->me.scratchpad + 8 + 8*stride, (ref_y ) + (fx>>2) + (fy>>2)*(stride) + 8 + 8*stride, stride);\
254
        qpel_avg[1][bxy](s->me.scratchpad               , (ref2_y) + (bx>>2) + (by>>2)*(stride)               , stride);\
255
        qpel_avg[1][bxy](s->me.scratchpad + 8           , (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8           , stride);\
256
        qpel_avg[1][bxy](s->me.scratchpad     + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride)     + 8*stride, stride);\
257
        qpel_avg[1][bxy](s->me.scratchpad + 8 + 8*stride, (ref2_y) + (bx>>2) + (by>>2)*(stride) + 8 + 8*stride, stride);\
258
    }\
259
    d = cmp_func(s, s->me.scratchpad, src_y, stride);\
260
}else\
261
    d= 256*256*256*32;
262

    
263

    
264
#define CMP_QPEL(d, dx, dy, x, y, size)\
265
    CMP_DIRECT(d, dx, dy, x, y, size, cmp_sub)
266

    
267
#define CMP(d, x, y, size)\
268
    CMP_DIRECT(d, 0, 0, x, y, size, cmp)
269

    
270
#include "motion_est_template.c"
271
#undef RENAME
272
#undef CMP
273
#undef CMP_HPEL
274
#undef CMP_QPEL
275
#undef INIT
276
#undef CMP__DIRECT
277

    
278

    
279
static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride){
280
    return 0;
281
}
282

    
283
static void set_cmp(MpegEncContext *s, me_cmp_func *cmp, int type){
284
    DSPContext* c= &s->dsp;
285
    int i;
286
    
287
    memset(cmp, 0, sizeof(void*)*11);
288

    
289
    switch(type&0xFF){
290
    case FF_CMP_SAD:
291
        cmp[0]= c->sad[0];
292
        cmp[1]= c->sad[1];
293
        break;
294
    case FF_CMP_SATD:
295
        cmp[0]= c->hadamard8_diff[0];
296
        cmp[1]= c->hadamard8_diff[1];
297
        break;
298
    case FF_CMP_SSE:
299
        cmp[0]= c->sse[0];
300
        cmp[1]= c->sse[1];
301
        break;
302
    case FF_CMP_DCT:
303
        cmp[0]= c->dct_sad[0];
304
        cmp[1]= c->dct_sad[1];
305
        break;
306
    case FF_CMP_PSNR:
307
        cmp[0]= c->quant_psnr[0];
308
        cmp[1]= c->quant_psnr[1];
309
        break;
310
    case FF_CMP_BIT:
311
        cmp[0]= c->bit[0];
312
        cmp[1]= c->bit[1];
313
        break;
314
    case FF_CMP_RD:
315
        cmp[0]= c->rd[0];
316
        cmp[1]= c->rd[1];
317
        break;
318
    case FF_CMP_ZERO:
319
        for(i=0; i<7; i++){
320
            cmp[i]= zero_cmp;
321
        }
322
        break;
323
    default:
324
        fprintf(stderr,"internal error in cmp function selection\n");
325
    }
326
}
327

    
328
static inline int get_penalty_factor(MpegEncContext *s, int type){
329
    switch(type&0xFF){
330
    default:
331
    case FF_CMP_SAD:
332
        return s->qscale*2;
333
    case FF_CMP_DCT:
334
        return s->qscale*3;
335
    case FF_CMP_SATD:
336
        return s->qscale*6;
337
    case FF_CMP_SSE:
338
        return s->qscale*s->qscale*2;
339
    case FF_CMP_BIT:
340
        return 1;
341
    case FF_CMP_RD:
342
    case FF_CMP_PSNR:
343
        return (s->qscale*s->qscale*185 + 64)>>7;
344
    }
345
}
346

    
347
void ff_init_me(MpegEncContext *s){
348
    set_cmp(s, s->dsp.me_pre_cmp, s->avctx->me_pre_cmp);
349
    set_cmp(s, s->dsp.me_cmp, s->avctx->me_cmp);
350
    set_cmp(s, s->dsp.me_sub_cmp, s->avctx->me_sub_cmp);
351
    set_cmp(s, s->dsp.mb_cmp, s->avctx->mb_cmp);
352

    
353
    if(s->flags&CODEC_FLAG_QPEL){
354
        if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
355
            s->me.sub_motion_search= simple_chroma_qpel_motion_search;
356
        else
357
            s->me.sub_motion_search= simple_qpel_motion_search;
358
    }else{
359
        if(s->avctx->me_sub_cmp&FF_CMP_CHROMA)
360
            s->me.sub_motion_search= simple_chroma_hpel_motion_search;
361
        else if(   s->avctx->me_sub_cmp == FF_CMP_SAD 
362
                && s->avctx->    me_cmp == FF_CMP_SAD 
363
                && s->avctx->    mb_cmp == FF_CMP_SAD)
364
            s->me.sub_motion_search= sad_hpel_motion_search;
365
        else
366
            s->me.sub_motion_search= simple_hpel_motion_search;
367
    }
368

    
369
    if(s->avctx->me_cmp&FF_CMP_CHROMA){
370
        s->me.motion_search[0]= simple_chroma_epzs_motion_search;
371
        s->me.motion_search[1]= simple_chroma_epzs_motion_search4;
372
    }else{
373
        s->me.motion_search[0]= simple_epzs_motion_search;
374
        s->me.motion_search[1]= simple_epzs_motion_search4;
375
    }
376
    
377
    if(s->avctx->me_pre_cmp&FF_CMP_CHROMA){
378
        s->me.pre_motion_search= simple_chroma_epzs_motion_search;
379
    }else{
380
        s->me.pre_motion_search= simple_epzs_motion_search;
381
    }
382
    
383
    if(s->flags&CODEC_FLAG_QPEL){
384
        if(s->avctx->mb_cmp&FF_CMP_CHROMA)
385
            s->me.get_mb_score= simple_chroma_qpel_get_mb_score;
386
        else
387
            s->me.get_mb_score= simple_qpel_get_mb_score;
388
    }else{
389
        if(s->avctx->mb_cmp&FF_CMP_CHROMA)
390
            s->me.get_mb_score= simple_chroma_hpel_get_mb_score;
391
        else
392
            s->me.get_mb_score= simple_hpel_get_mb_score;
393
    }
394
}
395
      
396
#if 0
397
static int pix_dev(uint8_t * pix, int line_size, int mean)
398
{
399
    int s, i, j;
400

401
    s = 0;
402
    for (i = 0; i < 16; i++) {
403
        for (j = 0; j < 16; j += 8) {
404
            s += ABS(pix[0]-mean);
405
            s += ABS(pix[1]-mean);
406
            s += ABS(pix[2]-mean);
407
            s += ABS(pix[3]-mean);
408
            s += ABS(pix[4]-mean);
409
            s += ABS(pix[5]-mean);
410
            s += ABS(pix[6]-mean);
411
            s += ABS(pix[7]-mean);
412
            pix += 8;
413
        }
414
        pix += line_size - 16;
415
    }
416
    return s;
417
}
418
#endif
419

    
420
static inline void no_motion_search(MpegEncContext * s,
421
                                    int *mx_ptr, int *my_ptr)
422
{
423
    *mx_ptr = 16 * s->mb_x;
424
    *my_ptr = 16 * s->mb_y;
425
}
426

    
427
static int full_motion_search(MpegEncContext * s,
428
                              int *mx_ptr, int *my_ptr, int range,
429
                              int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
430
{
431
    int x1, y1, x2, y2, xx, yy, x, y;
432
    int mx, my, dmin, d;
433
    uint8_t *pix;
434

    
435
    xx = 16 * s->mb_x;
436
    yy = 16 * s->mb_y;
437
    x1 = xx - range + 1;        /* we loose one pixel to avoid boundary pb with half pixel pred */
438
    if (x1 < xmin)
439
        x1 = xmin;
440
    x2 = xx + range - 1;
441
    if (x2 > xmax)
442
        x2 = xmax;
443
    y1 = yy - range + 1;
444
    if (y1 < ymin)
445
        y1 = ymin;
446
    y2 = yy + range - 1;
447
    if (y2 > ymax)
448
        y2 = ymax;
449
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
450
    dmin = 0x7fffffff;
451
    mx = 0;
452
    my = 0;
453
    for (y = y1; y <= y2; y++) {
454
        for (x = x1; x <= x2; x++) {
455
            d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
456
                             s->linesize);
457
            if (d < dmin ||
458
                (d == dmin &&
459
                 (abs(x - xx) + abs(y - yy)) <
460
                 (abs(mx - xx) + abs(my - yy)))) {
461
                dmin = d;
462
                mx = x;
463
                my = y;
464
            }
465
        }
466
    }
467

    
468
    *mx_ptr = mx;
469
    *my_ptr = my;
470

    
471
#if 0
472
    if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
473
        *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
474
        fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
475
    }
476
#endif
477
    return dmin;
478
}
479

    
480

    
481
static int log_motion_search(MpegEncContext * s,
482
                             int *mx_ptr, int *my_ptr, int range,
483
                             int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
484
{
485
    int x1, y1, x2, y2, xx, yy, x, y;
486
    int mx, my, dmin, d;
487
    uint8_t *pix;
488

    
489
    xx = s->mb_x << 4;
490
    yy = s->mb_y << 4;
491

    
492
    /* Left limit */
493
    x1 = xx - range;
494
    if (x1 < xmin)
495
        x1 = xmin;
496

    
497
    /* Right limit */
498
    x2 = xx + range;
499
    if (x2 > xmax)
500
        x2 = xmax;
501

    
502
    /* Upper limit */
503
    y1 = yy - range;
504
    if (y1 < ymin)
505
        y1 = ymin;
506

    
507
    /* Lower limit */
508
    y2 = yy + range;
509
    if (y2 > ymax)
510
        y2 = ymax;
511

    
512
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
513
    dmin = 0x7fffffff;
514
    mx = 0;
515
    my = 0;
516

    
517
    do {
518
        for (y = y1; y <= y2; y += range) {
519
            for (x = x1; x <= x2; x += range) {
520
                d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
521
                if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
522
                    dmin = d;
523
                    mx = x;
524
                    my = y;
525
                }
526
            }
527
        }
528

    
529
        range = range >> 1;
530

    
531
        x1 = mx - range;
532
        if (x1 < xmin)
533
            x1 = xmin;
534

    
535
        x2 = mx + range;
536
        if (x2 > xmax)
537
            x2 = xmax;
538

    
539
        y1 = my - range;
540
        if (y1 < ymin)
541
            y1 = ymin;
542

    
543
        y2 = my + range;
544
        if (y2 > ymax)
545
            y2 = ymax;
546

    
547
    } while (range >= 1);
548

    
549
#ifdef DEBUG
550
    fprintf(stderr, "log       - MX: %d\tMY: %d\n", mx, my);
551
#endif
552
    *mx_ptr = mx;
553
    *my_ptr = my;
554
    return dmin;
555
}
556

    
557
static int phods_motion_search(MpegEncContext * s,
558
                               int *mx_ptr, int *my_ptr, int range,
559
                               int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
560
{
561
    int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
562
    int mx, my, dminx, dminy;
563
    uint8_t *pix;
564

    
565
    xx = s->mb_x << 4;
566
    yy = s->mb_y << 4;
567

    
568
    /* Left limit */
569
    x1 = xx - range;
570
    if (x1 < xmin)
571
        x1 = xmin;
572

    
573
    /* Right limit */
574
    x2 = xx + range;
575
    if (x2 > xmax)
576
        x2 = xmax;
577

    
578
    /* Upper limit */
579
    y1 = yy - range;
580
    if (y1 < ymin)
581
        y1 = ymin;
582

    
583
    /* Lower limit */
584
    y2 = yy + range;
585
    if (y2 > ymax)
586
        y2 = ymax;
587

    
588
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
589
    mx = 0;
590
    my = 0;
591

    
592
    x = xx;
593
    y = yy;
594
    do {
595
        dminx = 0x7fffffff;
596
        dminy = 0x7fffffff;
597

    
598
        lastx = x;
599
        for (x = x1; x <= x2; x += range) {
600
            d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
601
            if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
602
                dminx = d;
603
                mx = x;
604
            }
605
        }
606

    
607
        x = lastx;
608
        for (y = y1; y <= y2; y += range) {
609
            d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
610
            if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
611
                dminy = d;
612
                my = y;
613
            }
614
        }
615

    
616
        range = range >> 1;
617

    
618
        x = mx;
619
        y = my;
620
        x1 = mx - range;
621
        if (x1 < xmin)
622
            x1 = xmin;
623

    
624
        x2 = mx + range;
625
        if (x2 > xmax)
626
            x2 = xmax;
627

    
628
        y1 = my - range;
629
        if (y1 < ymin)
630
            y1 = ymin;
631

    
632
        y2 = my + range;
633
        if (y2 > ymax)
634
            y2 = ymax;
635

    
636
    } while (range >= 1);
637

    
638
#ifdef DEBUG
639
    fprintf(stderr, "phods     - MX: %d\tMY: %d\n", mx, my);
640
#endif
641

    
642
    /* half pixel search */
643
    *mx_ptr = mx;
644
    *my_ptr = my;
645
    return dminy;
646
}
647

    
648

    
649
#define Z_THRESHOLD 256
650

    
651
#define CHECK_SAD_HALF_MV(suffix, x, y) \
652
{\
653
    d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
654
    d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
655
    COPY3_IF_LT(dminh, d, dx, x, dy, y)\
656
}
657

    
658
static inline int sad_hpel_motion_search(MpegEncContext * s,
659
                                  int *mx_ptr, int *my_ptr, int dmin,
660
                                  int xmin, int ymin, int xmax, int ymax,
661
                                  int pred_x, int pred_y, Picture *picture,
662
                                  int n, int size, uint8_t * const mv_penalty)
663
{
664
    uint8_t *ref_picture= picture->data[0];
665
    uint32_t *score_map= s->me.score_map;
666
    const int penalty_factor= s->me.sub_penalty_factor;
667
    int mx, my, xx, yy, dminh;
668
    uint8_t *pix, *ptr;
669
    op_pixels_abs_func pix_abs_x2;
670
    op_pixels_abs_func pix_abs_y2;
671
    op_pixels_abs_func pix_abs_xy2;
672
    
673
    if(size==0){
674
        pix_abs_x2 = s->dsp.pix_abs16x16_x2;
675
        pix_abs_y2 = s->dsp.pix_abs16x16_y2;
676
        pix_abs_xy2= s->dsp.pix_abs16x16_xy2;
677
    }else{
678
        pix_abs_x2 = s->dsp.pix_abs8x8_x2;
679
        pix_abs_y2 = s->dsp.pix_abs8x8_y2;
680
        pix_abs_xy2= s->dsp.pix_abs8x8_xy2;
681
    }
682

    
683
    if(s->me.skip){
684
//    printf("S");
685
        *mx_ptr = 0;
686
        *my_ptr = 0;
687
        return dmin;
688
    }
689
//    printf("N");
690
        
691
    xx = 16 * s->mb_x + 8*(n&1);
692
    yy = 16 * s->mb_y + 8*(n>>1);
693
    pix =  s->new_picture.data[0] + (yy * s->linesize) + xx;
694

    
695
    mx = *mx_ptr;
696
    my = *my_ptr;
697
    ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
698
    
699
    dminh = dmin;
700

    
701
    if (mx > xmin && mx < xmax && 
702
        my > ymin && my < ymax) {
703
        int dx=0, dy=0;
704
        int d, pen_x, pen_y; 
705
        const int index= (my<<ME_MAP_SHIFT) + mx;
706
        const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
707
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
708
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
709
        const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
710
        mx<<=1;
711
        my<<=1;
712

    
713
        
714
        pen_x= pred_x + mx;
715
        pen_y= pred_y + my;
716

    
717
        ptr-= s->linesize;
718
        if(t<=b){
719
            CHECK_SAD_HALF_MV(y2 , 0, -1)
720
            if(l<=r){
721
                CHECK_SAD_HALF_MV(xy2, -1, -1)
722
                if(t+r<=b+l){
723
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
724
                    ptr+= s->linesize;
725
                }else{
726
                    ptr+= s->linesize;
727
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
728
                }
729
                CHECK_SAD_HALF_MV(x2 , -1,  0)
730
            }else{
731
                CHECK_SAD_HALF_MV(xy2, +1, -1)
732
                if(t+l<=b+r){
733
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
734
                    ptr+= s->linesize;
735
                }else{
736
                    ptr+= s->linesize;
737
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
738
                }
739
                CHECK_SAD_HALF_MV(x2 , +1,  0)
740
            }
741
        }else{
742
            if(l<=r){
743
                if(t+l<=b+r){
744
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
745
                    ptr+= s->linesize;
746
                }else{
747
                    ptr+= s->linesize;
748
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
749
                }
750
                CHECK_SAD_HALF_MV(x2 , -1,  0)
751
                CHECK_SAD_HALF_MV(xy2, -1, +1)
752
            }else{
753
                if(t+r<=b+l){
754
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
755
                    ptr+= s->linesize;
756
                }else{
757
                    ptr+= s->linesize;
758
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
759
                }
760
                CHECK_SAD_HALF_MV(x2 , +1,  0)
761
                CHECK_SAD_HALF_MV(xy2, +1, +1)
762
            }
763
            CHECK_SAD_HALF_MV(y2 ,  0, +1)
764
        }
765
        mx+=dx;
766
        my+=dy;
767

    
768
    }else{
769
        mx<<=1;
770
        my<<=1;
771
    }
772

    
773
    *mx_ptr = mx;
774
    *my_ptr = my;
775
    return dminh;
776
}
777

    
778
static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
779
{
780
    const int xy= s->mb_x + s->mb_y*s->mb_stride;
781
    
782
    s->p_mv_table[xy][0] = mx;
783
    s->p_mv_table[xy][1] = my;
784

    
785
    /* has allready been set to the 4 MV if 4MV is done */
786
    if(mv4){
787
        int mot_xy= s->block_index[0];
788

    
789
        s->motion_val[mot_xy  ][0]= mx;
790
        s->motion_val[mot_xy  ][1]= my;
791
        s->motion_val[mot_xy+1][0]= mx;
792
        s->motion_val[mot_xy+1][1]= my;
793

    
794
        mot_xy += s->block_wrap[0];
795
        s->motion_val[mot_xy  ][0]= mx;
796
        s->motion_val[mot_xy  ][1]= my;
797
        s->motion_val[mot_xy+1][0]= mx;
798
        s->motion_val[mot_xy+1][1]= my;
799
    }
800
}
801

    
802
/**
803
 * get fullpel ME search limits.
804
 * @param range the approximate search range for the old ME code, unused for EPZS and newer
805
 */
806
static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax)
807
{
808
    if(s->avctx->me_range) *range= s->avctx->me_range >> 1;
809
    else                   *range= 16;
810

    
811
    if (s->unrestricted_mv) {
812
        *xmin = -16;
813
        *ymin = -16;
814
        *xmax = s->mb_width*16;
815
        *ymax = s->mb_height*16;
816
    } else {
817
        *xmin = 0;
818
        *ymin = 0;
819
        *xmax = s->mb_width*16 - 16;
820
        *ymax = s->mb_height*16 - 16;
821
    }
822
    
823
    //FIXME try to limit x/y min/max if me_range is set
824
}
825

    
826
static inline int h263_mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
827
{
828
    int block;
829
    int P[10][2];
830
    int dmin_sum=0, mx4_sum=0, my4_sum=0;
831
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
832

    
833
    for(block=0; block<4; block++){
834
        int mx4, my4;
835
        int pred_x4, pred_y4;
836
        int dmin4;
837
        static const int off[4]= {2, 1, 1, -1};
838
        const int mot_stride = s->block_wrap[0];
839
        const int mot_xy = s->block_index[block];
840
//        const int block_x= (block&1);
841
//        const int block_y= (block>>1);
842
#if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way
843
        const int rel_xmin4= xmin;
844
        const int rel_xmax4= xmax;
845
        const int rel_ymin4= ymin;
846
        const int rel_ymax4= ymax;
847
#else
848
        const int rel_xmin4= xmin - block_x*8;
849
        const int rel_xmax4= xmax - block_x*8 + 8;
850
        const int rel_ymin4= ymin - block_y*8;
851
        const int rel_ymax4= ymax - block_y*8 + 8;
852
#endif
853
        P_LEFT[0] = s->motion_val[mot_xy - 1][0];
854
        P_LEFT[1] = s->motion_val[mot_xy - 1][1];
855

    
856
        if(P_LEFT[0]       > (rel_xmax4<<shift)) P_LEFT[0]       = (rel_xmax4<<shift);
857

    
858
        /* special case for first line */
859
        if (s->mb_y == 0 && block<2) {
860
            pred_x4= P_LEFT[0];
861
            pred_y4= P_LEFT[1];
862
        } else {
863
            P_TOP[0]      = s->motion_val[mot_xy - mot_stride             ][0];
864
            P_TOP[1]      = s->motion_val[mot_xy - mot_stride             ][1];
865
            P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
866
            P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
867
            if(P_TOP[1]      > (rel_ymax4<<shift)) P_TOP[1]     = (rel_ymax4<<shift);
868
            if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);
869
            if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);
870
            if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);
871
    
872
            P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
873
            P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
874

    
875
//            if(s->out_format == FMT_H263){
876
                pred_x4 = P_MEDIAN[0];
877
                pred_y4 = P_MEDIAN[1];
878
#if 0
879
            }else { /* mpeg1 at least */
880
                pred_x4= P_LEFT[0];
881
                pred_y4= P_LEFT[1];
882
            }
883
#endif
884
        }
885
        P_MV1[0]= mx;
886
        P_MV1[1]= my;
887

    
888
        dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
889
                                       &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
890

    
891
        dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
892
                                          pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);
893
        
894
        if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
895
            int dxy;
896
            const int offset= ((block&1) + (block>>1)*s->linesize)*8;
897
            uint8_t *dest_y = s->me.scratchpad + offset;
898

    
899
            if(s->quarter_sample){
900
                uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>2)) + (s->mb_y*16 + (my4>>2))*s->linesize + offset;
901
                dxy = ((my4 & 3) << 2) | (mx4 & 3);
902

    
903
                if(s->no_rounding)
904
                    s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , s->linesize);
905
                else
906
                    s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , s->linesize);
907
            }else{
908
                uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>1)) + (s->mb_y*16 + (my4>>1))*s->linesize + offset;
909
                dxy = ((my4 & 1) << 1) | (mx4 & 1);
910

    
911
                if(s->no_rounding)
912
                    s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , s->linesize, 8);
913
                else
914
                    s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , s->linesize, 8);
915
            }
916
            dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
917
        }else
918
            dmin_sum+= dmin4;
919

    
920
        if(s->quarter_sample){
921
            mx4_sum+= mx4/2;
922
            my4_sum+= my4/2;
923
        }else{
924
            mx4_sum+= mx4;
925
            my4_sum+= my4;
926
        }
927
            
928
        s->motion_val[ s->block_index[block] ][0]= mx4;
929
        s->motion_val[ s->block_index[block] ][1]= my4;
930
    }
931
    
932
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
933
        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);
934
    }
935
    
936
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
937
        int dxy;
938
        int mx, my;
939
        int offset;
940

    
941
        mx= ff_h263_round_chroma(mx4_sum);
942
        my= ff_h263_round_chroma(my4_sum);
943
        dxy = ((my & 1) << 1) | (mx & 1);
944
        
945
        offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
946
       
947
        if(s->no_rounding){
948
            s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
949
            s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
950
        }else{
951
            s->dsp.put_pixels_tab       [1][dxy](s->me.scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
952
            s->dsp.put_pixels_tab       [1][dxy](s->me.scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
953
        }
954

    
955
        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);
956
        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);
957
    }
958

    
959
    switch(s->avctx->mb_cmp&0xFF){
960
    /*case FF_CMP_SSE:
961
        return dmin_sum+ 32*s->qscale*s->qscale;*/
962
    case FF_CMP_RD:
963
        return dmin_sum;
964
    default:
965
        return dmin_sum+ 11*s->me.mb_penalty_factor;
966
    }
967
}
968

    
969
void ff_estimate_p_frame_motion(MpegEncContext * s,
970
                                int mb_x, int mb_y)
971
{
972
    uint8_t *pix, *ppix;
973
    int sum, varc, vard, mx, my, range, dmin, xx, yy;
974
    int xmin, ymin, xmax, ymax;
975
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
976
    int pred_x=0, pred_y=0;
977
    int P[10][2];
978
    const int shift= 1+s->quarter_sample;
979
    int mb_type=0;
980
    uint8_t *ref_picture= s->last_picture.data[0];
981
    Picture * const pic= &s->current_picture;
982
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
983
    
984
    assert(s->quarter_sample==0 || s->quarter_sample==1);
985

    
986
    s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
987
    s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
988
    s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
989

    
990
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
991
    rel_xmin= xmin - mb_x*16;
992
    rel_xmax= xmax - mb_x*16;
993
    rel_ymin= ymin - mb_y*16;
994
    rel_ymax= ymax - mb_y*16;
995
    s->me.skip=0;
996

    
997
    switch(s->me_method) {
998
    case ME_ZERO:
999
    default:
1000
        no_motion_search(s, &mx, &my);
1001
        mx-= mb_x*16;
1002
        my-= mb_y*16;
1003
        dmin = 0;
1004
        break;
1005
    case ME_FULL:
1006
        dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1007
        mx-= mb_x*16;
1008
        my-= mb_y*16;
1009
        break;
1010
    case ME_LOG:
1011
        dmin = log_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_PHODS:
1016
        dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1017
        mx-= mb_x*16;
1018
        my-= mb_y*16;
1019
        break;
1020
    case ME_X1:
1021
    case ME_EPZS:
1022
       {
1023
            const int mot_stride = s->block_wrap[0];
1024
            const int mot_xy = s->block_index[0];
1025

    
1026
            P_LEFT[0]       = s->motion_val[mot_xy - 1][0];
1027
            P_LEFT[1]       = s->motion_val[mot_xy - 1][1];
1028

    
1029
            if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);
1030

    
1031
            if(mb_y) {
1032
                P_TOP[0]      = s->motion_val[mot_xy - mot_stride    ][0];
1033
                P_TOP[1]      = s->motion_val[mot_xy - mot_stride    ][1];
1034
                P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];
1035
                P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];
1036
                if(P_TOP[1]      > (rel_ymax<<shift)) P_TOP[1]     = (rel_ymax<<shift);
1037
                if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1038
                if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1039
        
1040
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1041
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1042

    
1043
                if(s->out_format == FMT_H263){
1044
                    pred_x = P_MEDIAN[0];
1045
                    pred_y = P_MEDIAN[1];
1046
                }else { /* mpeg1 at least */
1047
                    pred_x= P_LEFT[0];
1048
                    pred_y= P_LEFT[1];
1049
                }
1050
            }else{
1051
                pred_x= P_LEFT[0];
1052
                pred_y= P_LEFT[1];
1053
            }
1054

    
1055
        }
1056
        dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1057
                                      &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1058
 
1059
        break;
1060
    }
1061

    
1062
    /* intra / predictive decision */
1063
    xx = mb_x * 16;
1064
    yy = mb_y * 16;
1065

    
1066
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
1067
    /* At this point (mx,my) are full-pell and the relative displacement */
1068
    ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
1069
    
1070
    sum = s->dsp.pix_sum(pix, s->linesize);
1071
    
1072
    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1073
    vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
1074

    
1075
//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);
1076
    pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;
1077
    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1078
    pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1079
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
1080
    pic->mb_var_sum    += varc;
1081
    pic->mc_mb_var_sum += vard;
1082
//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);
1083
    
1084
#if 0
1085
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1086
           varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1087
#endif
1088
    if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1089
        if (vard <= 64 || vard < varc)
1090
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1091
        else
1092
            s->scene_change_score+= s->qscale;
1093

    
1094
        if (vard*2 + 200 > varc)
1095
            mb_type|= MB_TYPE_INTRA;
1096
        if (varc*2 + 200 > vard){
1097
            mb_type|= MB_TYPE_INTER;
1098
            s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1099
                                   pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1100
        }else{
1101
            mx <<=shift;
1102
            my <<=shift;
1103
        }
1104
        if((s->flags&CODEC_FLAG_4MV)
1105
           && !s->me.skip && varc>50 && vard>10){
1106
            h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1107
            mb_type|=MB_TYPE_INTER4V;
1108

    
1109
            set_p_mv_tables(s, mx, my, 0);
1110
        }else
1111
            set_p_mv_tables(s, mx, my, 1);
1112
    }else{
1113
        mb_type= MB_TYPE_INTER;
1114

    
1115
        dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1116
                                    pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1117
        
1118
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1119
            dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);
1120

    
1121
        if((s->flags&CODEC_FLAG_4MV)
1122
           && !s->me.skip && varc>50 && vard>10){
1123
            int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1124
            if(dmin4 < dmin){
1125
                mb_type= MB_TYPE_INTER4V;
1126
                dmin=dmin4;
1127
            }
1128
        }
1129
        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
1130
        set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
1131
        
1132
        if (vard <= 64 || vard < varc) {
1133
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1134
        }else{
1135
            s->scene_change_score+= s->qscale;
1136
        }
1137
    }
1138

    
1139
    s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1140
}
1141

    
1142
int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1143
                                    int mb_x, int mb_y)
1144
{
1145
    int mx, my, range, dmin;
1146
    int xmin, ymin, xmax, ymax;
1147
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1148
    int pred_x=0, pred_y=0;
1149
    int P[10][2];
1150
    const int shift= 1+s->quarter_sample;
1151
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1152
    const int xy= mb_x + mb_y*s->mb_stride;
1153
    
1154
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1155

    
1156
    s->me.pre_penalty_factor    = get_penalty_factor(s, s->avctx->me_pre_cmp);
1157

    
1158
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1159
    rel_xmin= xmin - mb_x*16;
1160
    rel_xmax= xmax - mb_x*16;
1161
    rel_ymin= ymin - mb_y*16;
1162
    rel_ymax= ymax - mb_y*16;
1163
    s->me.skip=0;
1164

    
1165
    P_LEFT[0]       = s->p_mv_table[xy + 1][0];
1166
    P_LEFT[1]       = s->p_mv_table[xy + 1][1];
1167

    
1168
    if(P_LEFT[0]       < (rel_xmin<<shift)) P_LEFT[0]       = (rel_xmin<<shift);
1169

    
1170
    /* special case for first line */
1171
    if (mb_y == s->mb_height-1) {
1172
        pred_x= P_LEFT[0];
1173
        pred_y= P_LEFT[1];
1174
        P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1175
        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME 
1176
    } else {
1177
        P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
1178
        P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
1179
        P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1180
        P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1181
        if(P_TOP[1]      < (rel_ymin<<shift)) P_TOP[1]     = (rel_ymin<<shift);
1182
        if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
1183
        if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
1184
    
1185
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1186
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1187

    
1188
        pred_x = P_MEDIAN[0];
1189
        pred_y = P_MEDIAN[1];
1190
    }
1191
    dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1192
                                   &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1193

    
1194
    s->p_mv_table[xy][0] = mx<<shift;
1195
    s->p_mv_table[xy][1] = my<<shift;
1196
    
1197
    return dmin;
1198
}
1199

    
1200
static int ff_estimate_motion_b(MpegEncContext * s,
1201
                       int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
1202
{
1203
    int mx, my, range, dmin;
1204
    int xmin, ymin, xmax, ymax;
1205
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1206
    int pred_x=0, pred_y=0;
1207
    int P[10][2];
1208
    const int shift= 1+s->quarter_sample;
1209
    const int mot_stride = s->mb_stride;
1210
    const int mot_xy = mb_y*mot_stride + mb_x;
1211
    uint8_t * const ref_picture= picture->data[0];
1212
    uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1213
    int mv_scale;
1214
        
1215
    s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
1216
    s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1217
    s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1218

    
1219
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1220
    rel_xmin= xmin - mb_x*16;
1221
    rel_xmax= xmax - mb_x*16;
1222
    rel_ymin= ymin - mb_y*16;
1223
    rel_ymax= ymax - mb_y*16;
1224

    
1225
    switch(s->me_method) {
1226
    case ME_ZERO:
1227
    default:
1228
        no_motion_search(s, &mx, &my);
1229
        dmin = 0;
1230
        mx-= mb_x*16;
1231
        my-= mb_y*16;
1232
        break;
1233
    case ME_FULL:
1234
        dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1235
        mx-= mb_x*16;
1236
        my-= mb_y*16;
1237
        break;
1238
    case ME_LOG:
1239
        dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1240
        mx-= mb_x*16;
1241
        my-= mb_y*16;
1242
        break;
1243
    case ME_PHODS:
1244
        dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1245
        mx-= mb_x*16;
1246
        my-= mb_y*16;
1247
        break;
1248
    case ME_X1:
1249
    case ME_EPZS:
1250
       {
1251
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
1252
            P_LEFT[1]        = mv_table[mot_xy - 1][1];
1253

    
1254
            if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);
1255

    
1256
            /* special case for first line */
1257
            if (mb_y) {
1258
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
1259
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
1260
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
1261
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
1262
                if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);
1263
                if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1264
                if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1265
        
1266
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1267
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1268
            }
1269
            pred_x= P_LEFT[0];
1270
            pred_y= P_LEFT[1];
1271
        }
1272
        
1273
        if(mv_table == s->b_forw_mv_table){
1274
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1275
        }else{
1276
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1277
        }
1278
        
1279
        dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1280
                                      picture, s->p_mv_table, mv_scale, mv_penalty);
1281
 
1282
        break;
1283
    }
1284
    
1285
    dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1286
                                   pred_x, pred_y, picture, 0, 0, mv_penalty);
1287
                                   
1288
    if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1289
        dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);
1290

    
1291
//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1292
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1293
    mv_table[mot_xy][0]= mx;
1294
    mv_table[mot_xy][1]= my;
1295

    
1296
    return dmin;
1297
}
1298

    
1299
static inline int check_bidir_mv(MpegEncContext * s,
1300
                   int mb_x, int mb_y,
1301
                   int motion_fx, int motion_fy,
1302
                   int motion_bx, int motion_by,
1303
                   int pred_fx, int pred_fy,
1304
                   int pred_bx, int pred_by)
1305
{
1306
    //FIXME optimize?
1307
    //FIXME move into template?
1308
    //FIXME better f_code prediction (max mv & distance)
1309
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1310
    uint8_t *dest_y = s->me.scratchpad;
1311
    uint8_t *ptr;
1312
    int dxy;
1313
    int src_x, src_y;
1314
    int fbmin;
1315

    
1316
    if(s->quarter_sample){
1317
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1318
        src_x = mb_x * 16 + (motion_fx >> 2);
1319
        src_y = mb_y * 16 + (motion_fy >> 2);
1320
        assert(src_x >=-16 && src_x<=s->h_edge_pos);
1321
        assert(src_y >=-16 && src_y<=s->v_edge_pos);
1322

    
1323
        ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1324
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);
1325

    
1326
        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1327
        src_x = mb_x * 16 + (motion_bx >> 2);
1328
        src_y = mb_y * 16 + (motion_by >> 2);
1329
        assert(src_x >=-16 && src_x<=s->h_edge_pos);
1330
        assert(src_y >=-16 && src_y<=s->v_edge_pos);
1331
    
1332
        ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1333
        s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);
1334
    }else{
1335
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1336
        src_x = mb_x * 16 + (motion_fx >> 1);
1337
        src_y = mb_y * 16 + (motion_fy >> 1);
1338
        assert(src_x >=-16 && src_x<=s->h_edge_pos);
1339
        assert(src_y >=-16 && src_y<=s->v_edge_pos);
1340

    
1341
        ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1342
        s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
1343

    
1344
        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1345
        src_x = mb_x * 16 + (motion_bx >> 1);
1346
        src_y = mb_y * 16 + (motion_by >> 1);
1347
        assert(src_x >=-16 && src_x<=s->h_edge_pos);
1348
        assert(src_y >=-16 && src_y<=s->v_edge_pos);
1349
    
1350
        ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1351
        s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
1352
    }
1353

    
1354
    fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1355
           +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1356
           + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
1357
           
1358
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1359
    }
1360
    //FIXME CHROMA !!!
1361
           
1362
    return fbmin;
1363
}
1364

    
1365
/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1366
static inline int bidir_refine(MpegEncContext * s,
1367
                                  int mb_x, int mb_y)
1368
{
1369
    const int mot_stride = s->mb_stride;
1370
    const int xy = mb_y *mot_stride + mb_x;
1371
    int fbmin;
1372
    int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1373
    int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1374
    int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1375
    int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1376
    int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1377
    int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1378
    int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1379
    int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1380

    
1381
    //FIXME do refinement and add flag
1382
    
1383
    fbmin= check_bidir_mv(s, mb_x, mb_y, 
1384
                          motion_fx, motion_fy,
1385
                          motion_bx, motion_by,
1386
                          pred_fx, pred_fy,
1387
                          pred_bx, pred_by);
1388

    
1389
   return fbmin;
1390
}
1391

    
1392
static inline int direct_search(MpegEncContext * s,
1393
                                int mb_x, int mb_y)
1394
{
1395
    int P[10][2];
1396
    const int mot_stride = s->mb_stride;
1397
    const int mot_xy = mb_y*mot_stride + mb_x;
1398
    const int shift= 1+s->quarter_sample;
1399
    int dmin, i;
1400
    const int time_pp= s->pp_time;
1401
    const int time_pb= s->pb_time;
1402
    int mx, my, xmin, xmax, ymin, ymax;
1403
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
1404
    uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1405
    
1406
    ymin= xmin=(-32)>>shift;
1407
    ymax= xmax=   31>>shift;
1408

    
1409
    if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1410
        s->mv_type= MV_TYPE_8X8;
1411
    }else{
1412
        s->mv_type= MV_TYPE_16X16;
1413
    }
1414

    
1415
    for(i=0; i<4; i++){
1416
        int index= s->block_index[i];
1417
        int min, max;
1418
    
1419
        s->me.co_located_mv[i][0]= s->motion_val[index][0];
1420
        s->me.co_located_mv[i][1]= s->motion_val[index][1];
1421
        s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1422
        s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1423
//        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);
1424
//        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);
1425

    
1426
        max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1427
        min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1428
        max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding
1429
        min+= (2*mb_x + (i& 1))*8 - 1;
1430
        xmax= FFMIN(xmax, s->width - max);
1431
        xmin= FFMAX(xmin, - 16     - min);
1432

    
1433
        max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1434
        min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1435
        max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding
1436
        min+= (2*mb_y + (i>>1))*8 - 1;
1437
        ymax= FFMIN(ymax, s->height - max);
1438
        ymin= FFMAX(ymin, - 16      - min);
1439
        
1440
        if(s->mv_type == MV_TYPE_16X16) break;
1441
    }
1442
    
1443
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1444
    
1445
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1446
        s->b_direct_mv_table[mot_xy][0]= 0;
1447
        s->b_direct_mv_table[mot_xy][1]= 0;
1448

    
1449
        return 256*256*256*64;
1450
    }
1451

    
1452
    P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1453
    P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1454

    
1455
    /* special case for first line */
1456
    if (mb_y) {
1457
        P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
1458
        P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
1459
        P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
1460
        P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
1461
    
1462
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1463
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1464
    }
1465
 
1466
    //FIXME direct_search  ptr in context!!! (needed for chroma anyway or this will get messy)   
1467
    if(s->flags&CODEC_FLAG_QPEL){
1468
        dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 
1469
                                                     &s->last_picture, mv_table, 1<<14, mv_penalty);
1470
        dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1471
                                                0, 0, &s->last_picture, 0, 0, mv_penalty);
1472
        
1473
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1474
            dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1475
    }else{
1476
        dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 
1477
                                                     &s->last_picture, mv_table, 1<<15, mv_penalty);
1478
        dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1479
                                                0, 0, &s->last_picture, 0, 0, mv_penalty);
1480

    
1481
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1482
            dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1483
    }
1484

    
1485
    s->b_direct_mv_table[mot_xy][0]= mx;
1486
    s->b_direct_mv_table[mot_xy][1]= my;
1487
    return dmin;
1488
}
1489

    
1490
void ff_estimate_b_frame_motion(MpegEncContext * s,
1491
                             int mb_x, int mb_y)
1492
{
1493
    const int penalty_factor= s->me.mb_penalty_factor;
1494
    int fmin, bmin, dmin, fbmin;
1495
    int type=0;
1496
    
1497
    dmin= direct_search(s, mb_x, mb_y);
1498

    
1499
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;
1500
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;
1501
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1502

    
1503
    fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1504
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1505
    {
1506
        int score= fmin;
1507
        type = MB_TYPE_FORWARD;
1508
        
1509
        // RAL: No MB_TYPE_DIRECT in MPEG-1 video (only MPEG-4)
1510
        if (s->codec_id == CODEC_ID_MPEG4 && dmin <= score){
1511
            score = dmin;
1512
            type = MB_TYPE_DIRECT;
1513
        }
1514
        if(bmin<score){
1515
            score=bmin;
1516
            type= MB_TYPE_BACKWARD; 
1517
        }
1518
        if(fbmin<score){
1519
            score=fbmin;
1520
            type= MB_TYPE_BIDIR;
1521
        }
1522
        
1523
        score= ((unsigned)(score*score + 128*256))>>16;
1524
        s->current_picture.mc_mb_var_sum += score;
1525
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1526
    }
1527

    
1528
    if(s->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1529
        type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
1530
        if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1531
    }
1532

    
1533
    s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1534
}
1535

    
1536
/* find best f_code for ME which do unlimited searches */
1537
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1538
{
1539
    if(s->me_method>=ME_EPZS){
1540
        int score[8];
1541
        int i, y;
1542
        uint8_t * fcode_tab= s->fcode_tab;
1543
        int best_fcode=-1;
1544
        int best_score=-10000000;
1545

    
1546
        for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1547

    
1548
        for(y=0; y<s->mb_height; y++){
1549
            int x;
1550
            int xy= y*s->mb_stride;
1551
            for(x=0; x<s->mb_width; x++){
1552
                if(s->mb_type[xy] & type){
1553
                    int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1554
                                     fcode_tab[mv_table[xy][1] + MAX_MV]);
1555
                    int j;
1556
                    
1557
                    for(j=0; j<fcode && j<8; j++){
1558
                        if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1559
                            score[j]-= 170;
1560
                    }
1561
                }
1562
                xy++;
1563
            }
1564
        }
1565
        
1566
        for(i=1; i<8; i++){
1567
            if(score[i] > best_score){
1568
                best_score= score[i];
1569
                best_fcode= i;
1570
            }
1571
//            printf("%d %d\n", i, score[i]);
1572
        }
1573

    
1574
//    printf("fcode: %d type: %d\n", i, s->pict_type);
1575
        return best_fcode;
1576
/*        for(i=0; i<=MAX_FCODE; i++){
1577
            printf("%d ", mv_num[i]);
1578
        }
1579
        printf("\n");*/
1580
    }else{
1581
        return 1;
1582
    }
1583
}
1584

    
1585
void ff_fix_long_p_mvs(MpegEncContext * s)
1586
{
1587
    const int f_code= s->f_code;
1588
    int y, range;
1589
    assert(s->pict_type==P_TYPE);
1590

    
1591
    range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1592
    
1593
    if(s->msmpeg4_version) range= 16;
1594
    
1595
    if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1596
    
1597
    /* clip / convert to intra 16x16 type MVs */
1598
    for(y=0; y<s->mb_height; y++){
1599
        int x;
1600
        int xy= y*s->mb_stride;
1601
        for(x=0; x<s->mb_width; x++){
1602
            if(s->mb_type[xy]&MB_TYPE_INTER){
1603
                if(   s->p_mv_table[xy][0] >=range || s->p_mv_table[xy][0] <-range
1604
                   || s->p_mv_table[xy][1] >=range || s->p_mv_table[xy][1] <-range){
1605
                    s->mb_type[xy] &= ~MB_TYPE_INTER;
1606
                    s->mb_type[xy] |= MB_TYPE_INTRA;
1607
                    s->p_mv_table[xy][0] = 0;
1608
                    s->p_mv_table[xy][1] = 0;
1609
                }
1610
            }
1611
            xy++;
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_stride;
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(   mx >=range || mx <-range
1633
                           || my >=range || my <-range){
1634
                            s->mb_type[i] &= ~MB_TYPE_INTER4V;
1635
                            s->mb_type[i] |= MB_TYPE_INTRA;
1636
                        }
1637
                    }
1638
                }
1639
                xy+=2;
1640
                i++;
1641
            }
1642
        }
1643
    }
1644
}
1645

    
1646
void ff_fix_long_b_mvs(MpegEncContext * s, int16_t (*mv_table)[2], int f_code, int type)
1647
{
1648
    int y;
1649

    
1650
    // RAL: 8 in MPEG-1, 16 in MPEG-4
1651
    int range = (((s->out_format == FMT_MPEG1) ? 8 : 16) << f_code);
1652
    
1653
    if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1654

    
1655
    /* clip / convert to intra 16x16 type MVs */
1656
    for(y=0; y<s->mb_height; y++){
1657
        int x;
1658
        int xy= y*s->mb_stride;
1659
        for(x=0; x<s->mb_width; x++){
1660
            if (s->mb_type[xy] & type){    // RAL: "type" test added...
1661
                if(   mv_table[xy][0] >=range || mv_table[xy][0] <-range
1662
                   || mv_table[xy][1] >=range || mv_table[xy][1] <-range){
1663

    
1664
                    if(s->codec_id == CODEC_ID_MPEG1VIDEO && 0){
1665
                    }else{
1666
                        if     (mv_table[xy][0] > range-1) mv_table[xy][0]=  range-1;
1667
                        else if(mv_table[xy][0] < -range ) mv_table[xy][0]= -range;
1668
                        if     (mv_table[xy][1] > range-1) mv_table[xy][1]=  range-1;
1669
                        else if(mv_table[xy][1] < -range ) mv_table[xy][1]= -range;
1670
                    }
1671
                }
1672
            }
1673
            xy++;
1674
        }
1675
    }
1676
}
1677
#if 0
1678
/**
1679
 * estimates global motion and inits sprite_ref
1680
 */
1681
void ff_estimate_global_motion(MpegEncContext *s, int sprite_ref[3][2]){
1682
    int y;
1683
    int num= 16<<s->f_code;
1684
    int score[2][num];
1685
    int best_i[2]={0,0};
1686
    int best_score[2]={0,0};
1687
    
1688
    memset(score, 0, 2*num*sizeof(int));
1689
    
1690
    for(y=0; y<s->mb_height; y++){
1691
        int x;
1692
        int xy= (y+1)* (s->mb_width+2)+1;
1693
        int i= y*s->mb_width;
1694
        for(x=0; x<s->mb_width; x++){
1695
            int mv[2];
1696
           
1697
            if(!(s->mb_type[i]&MB_TYPE_INTER))
1698
                continue;
1699
            
1700
            mv[0]= s->p_mv_table[xy][0];
1701
            mv[1]= s->p_mv_table[xy][1];
1702

1703
            if(mv[0]==0 && mv[1]==0) continue;
1704
            
1705
            score[0][mv[0] + num/2]++;
1706
            score[1][mv[1] + num/2]++;
1707
        }
1708
    }
1709

1710
    for(n=0; n<2; n++){
1711
        for(i=1; i<num-1; i++){
1712
            int s= score[n][i-1] + score[n][i]*2 + score[n][i+1];
1713
    
1714
            if(s > best_score[n]){
1715
                best_score[n]= s;
1716
                best_i[n]= i;
1717
            }
1718
        }
1719
    }
1720
    
1721
    sprite_ref[0][0]= best_i[0] - num/2;
1722
    sprite_ref[0][1]= best_i[1] - num/2;
1723
    
1724
    // decide block type
1725
}
1726
#endif