Statistics
| Branch: | Revision:

ffmpeg / libavcodec / motion_est.c @ 6d93f194

History | View | Annotate | Download (56.2 KB)

1
/*
2
 * Motion estimation 
3
 * Copyright (c) 2000,2001 Fabrice Bellard.
4
 * Copyright (c) 2002-2003 Michael Niedermayer
5
 * 
6
 *
7
 * This library is free software; you can redistribute it and/or
8
 * modify it under the terms of the GNU Lesser General Public
9
 * License as published by the Free Software Foundation; either
10
 * version 2 of the License, or (at your option) any later version.
11
 *
12
 * This library is distributed in the hope that it will be useful,
13
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15
 * Lesser General Public License for more details.
16
 *
17
 * You should have received a copy of the GNU Lesser General Public
18
 * License along with this library; if not, write to the Free Software
19
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
20
 *
21
 * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
22
 */
23
 
24
/**
25
 * @file motion_est.c
26
 * Motion estimation.
27
 */
28
 
29
#include <stdlib.h>
30
#include <stdio.h>
31
#include "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
static int pix_dev(uint8_t * pix, int line_size, int mean)
397
{
398
    int s, i, j;
399

    
400
    s = 0;
401
    for (i = 0; i < 16; i++) {
402
        for (j = 0; j < 16; j += 8) {
403
            s += ABS(pix[0]-mean);
404
            s += ABS(pix[1]-mean);
405
            s += ABS(pix[2]-mean);
406
            s += ABS(pix[3]-mean);
407
            s += ABS(pix[4]-mean);
408
            s += ABS(pix[5]-mean);
409
            s += ABS(pix[6]-mean);
410
            s += ABS(pix[7]-mean);
411
            pix += 8;
412
        }
413
        pix += line_size - 16;
414
    }
415
    return s;
416
}
417

    
418
static inline void no_motion_search(MpegEncContext * s,
419
                                    int *mx_ptr, int *my_ptr)
420
{
421
    *mx_ptr = 16 * s->mb_x;
422
    *my_ptr = 16 * s->mb_y;
423
}
424

    
425
static int full_motion_search(MpegEncContext * s,
426
                              int *mx_ptr, int *my_ptr, int range,
427
                              int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
428
{
429
    int x1, y1, x2, y2, xx, yy, x, y;
430
    int mx, my, dmin, d;
431
    uint8_t *pix;
432

    
433
    xx = 16 * s->mb_x;
434
    yy = 16 * s->mb_y;
435
    x1 = xx - range + 1;        /* we loose one pixel to avoid boundary pb with half pixel pred */
436
    if (x1 < xmin)
437
        x1 = xmin;
438
    x2 = xx + range - 1;
439
    if (x2 > xmax)
440
        x2 = xmax;
441
    y1 = yy - range + 1;
442
    if (y1 < ymin)
443
        y1 = ymin;
444
    y2 = yy + range - 1;
445
    if (y2 > ymax)
446
        y2 = ymax;
447
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
448
    dmin = 0x7fffffff;
449
    mx = 0;
450
    my = 0;
451
    for (y = y1; y <= y2; y++) {
452
        for (x = x1; x <= x2; x++) {
453
            d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x,
454
                             s->linesize);
455
            if (d < dmin ||
456
                (d == dmin &&
457
                 (abs(x - xx) + abs(y - yy)) <
458
                 (abs(mx - xx) + abs(my - yy)))) {
459
                dmin = d;
460
                mx = x;
461
                my = y;
462
            }
463
        }
464
    }
465

    
466
    *mx_ptr = mx;
467
    *my_ptr = my;
468

    
469
#if 0
470
    if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
471
        *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
472
        fprintf(stderr, "error %d %d\n", *mx_ptr, *my_ptr);
473
    }
474
#endif
475
    return dmin;
476
}
477

    
478

    
479
static int log_motion_search(MpegEncContext * s,
480
                             int *mx_ptr, int *my_ptr, int range,
481
                             int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
482
{
483
    int x1, y1, x2, y2, xx, yy, x, y;
484
    int mx, my, dmin, d;
485
    uint8_t *pix;
486

    
487
    xx = s->mb_x << 4;
488
    yy = s->mb_y << 4;
489

    
490
    /* Left limit */
491
    x1 = xx - range;
492
    if (x1 < xmin)
493
        x1 = xmin;
494

    
495
    /* Right limit */
496
    x2 = xx + range;
497
    if (x2 > xmax)
498
        x2 = xmax;
499

    
500
    /* Upper limit */
501
    y1 = yy - range;
502
    if (y1 < ymin)
503
        y1 = ymin;
504

    
505
    /* Lower limit */
506
    y2 = yy + range;
507
    if (y2 > ymax)
508
        y2 = ymax;
509

    
510
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
511
    dmin = 0x7fffffff;
512
    mx = 0;
513
    my = 0;
514

    
515
    do {
516
        for (y = y1; y <= y2; y += range) {
517
            for (x = x1; x <= x2; x += range) {
518
                d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
519
                if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
520
                    dmin = d;
521
                    mx = x;
522
                    my = y;
523
                }
524
            }
525
        }
526

    
527
        range = range >> 1;
528

    
529
        x1 = mx - range;
530
        if (x1 < xmin)
531
            x1 = xmin;
532

    
533
        x2 = mx + range;
534
        if (x2 > xmax)
535
            x2 = xmax;
536

    
537
        y1 = my - range;
538
        if (y1 < ymin)
539
            y1 = ymin;
540

    
541
        y2 = my + range;
542
        if (y2 > ymax)
543
            y2 = ymax;
544

    
545
    } while (range >= 1);
546

    
547
#ifdef DEBUG
548
    fprintf(stderr, "log       - MX: %d\tMY: %d\n", mx, my);
549
#endif
550
    *mx_ptr = mx;
551
    *my_ptr = my;
552
    return dmin;
553
}
554

    
555
static int phods_motion_search(MpegEncContext * s,
556
                               int *mx_ptr, int *my_ptr, int range,
557
                               int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
558
{
559
    int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
560
    int mx, my, dminx, dminy;
561
    uint8_t *pix;
562

    
563
    xx = s->mb_x << 4;
564
    yy = s->mb_y << 4;
565

    
566
    /* Left limit */
567
    x1 = xx - range;
568
    if (x1 < xmin)
569
        x1 = xmin;
570

    
571
    /* Right limit */
572
    x2 = xx + range;
573
    if (x2 > xmax)
574
        x2 = xmax;
575

    
576
    /* Upper limit */
577
    y1 = yy - range;
578
    if (y1 < ymin)
579
        y1 = ymin;
580

    
581
    /* Lower limit */
582
    y2 = yy + range;
583
    if (y2 > ymax)
584
        y2 = ymax;
585

    
586
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
587
    mx = 0;
588
    my = 0;
589

    
590
    x = xx;
591
    y = yy;
592
    do {
593
        dminx = 0x7fffffff;
594
        dminy = 0x7fffffff;
595

    
596
        lastx = x;
597
        for (x = x1; x <= x2; x += range) {
598
            d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
599
            if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
600
                dminx = d;
601
                mx = x;
602
            }
603
        }
604

    
605
        x = lastx;
606
        for (y = y1; y <= y2; y += range) {
607
            d = s->dsp.pix_abs16x16(pix, ref_picture + (y * s->linesize) + x, s->linesize);
608
            if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
609
                dminy = d;
610
                my = y;
611
            }
612
        }
613

    
614
        range = range >> 1;
615

    
616
        x = mx;
617
        y = my;
618
        x1 = mx - range;
619
        if (x1 < xmin)
620
            x1 = xmin;
621

    
622
        x2 = mx + range;
623
        if (x2 > xmax)
624
            x2 = xmax;
625

    
626
        y1 = my - range;
627
        if (y1 < ymin)
628
            y1 = ymin;
629

    
630
        y2 = my + range;
631
        if (y2 > ymax)
632
            y2 = ymax;
633

    
634
    } while (range >= 1);
635

    
636
#ifdef DEBUG
637
    fprintf(stderr, "phods     - MX: %d\tMY: %d\n", mx, my);
638
#endif
639

    
640
    /* half pixel search */
641
    *mx_ptr = mx;
642
    *my_ptr = my;
643
    return dminy;
644
}
645

    
646

    
647
#define Z_THRESHOLD 256
648

    
649
#define CHECK_SAD_HALF_MV(suffix, x, y) \
650
{\
651
    d= pix_abs_ ## suffix(pix, ptr+((x)>>1), s->linesize);\
652
    d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
653
    COPY3_IF_LT(dminh, d, dx, x, dy, y)\
654
}
655

    
656
static inline int sad_hpel_motion_search(MpegEncContext * s,
657
                                  int *mx_ptr, int *my_ptr, int dmin,
658
                                  int xmin, int ymin, int xmax, int ymax,
659
                                  int pred_x, int pred_y, Picture *picture,
660
                                  int n, int size, uint8_t * const mv_penalty)
661
{
662
    uint8_t *ref_picture= picture->data[0];
663
    uint32_t *score_map= s->me.score_map;
664
    const int penalty_factor= s->me.sub_penalty_factor;
665
    int mx, my, xx, yy, dminh;
666
    uint8_t *pix, *ptr;
667
    op_pixels_abs_func pix_abs_x2;
668
    op_pixels_abs_func pix_abs_y2;
669
    op_pixels_abs_func pix_abs_xy2;
670
    
671
    if(size==0){
672
        pix_abs_x2 = s->dsp.pix_abs16x16_x2;
673
        pix_abs_y2 = s->dsp.pix_abs16x16_y2;
674
        pix_abs_xy2= s->dsp.pix_abs16x16_xy2;
675
    }else{
676
        pix_abs_x2 = s->dsp.pix_abs8x8_x2;
677
        pix_abs_y2 = s->dsp.pix_abs8x8_y2;
678
        pix_abs_xy2= s->dsp.pix_abs8x8_xy2;
679
    }
680

    
681
    if(s->me.skip){
682
//    printf("S");
683
        *mx_ptr = 0;
684
        *my_ptr = 0;
685
        return dmin;
686
    }
687
//    printf("N");
688
        
689
    xx = 16 * s->mb_x + 8*(n&1);
690
    yy = 16 * s->mb_y + 8*(n>>1);
691
    pix =  s->new_picture.data[0] + (yy * s->linesize) + xx;
692

    
693
    mx = *mx_ptr;
694
    my = *my_ptr;
695
    ptr = ref_picture + ((yy + my) * s->linesize) + (xx + mx);
696
    
697
    dminh = dmin;
698

    
699
    if (mx > xmin && mx < xmax && 
700
        my > ymin && my < ymax) {
701
        int dx=0, dy=0;
702
        int d, pen_x, pen_y; 
703
        const int index= (my<<ME_MAP_SHIFT) + mx;
704
        const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
705
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
706
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
707
        const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
708
        mx<<=1;
709
        my<<=1;
710

    
711
        
712
        pen_x= pred_x + mx;
713
        pen_y= pred_y + my;
714

    
715
        ptr-= s->linesize;
716
        if(t<=b){
717
            CHECK_SAD_HALF_MV(y2 , 0, -1)
718
            if(l<=r){
719
                CHECK_SAD_HALF_MV(xy2, -1, -1)
720
                if(t+r<=b+l){
721
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
722
                    ptr+= s->linesize;
723
                }else{
724
                    ptr+= s->linesize;
725
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
726
                }
727
                CHECK_SAD_HALF_MV(x2 , -1,  0)
728
            }else{
729
                CHECK_SAD_HALF_MV(xy2, +1, -1)
730
                if(t+l<=b+r){
731
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
732
                    ptr+= s->linesize;
733
                }else{
734
                    ptr+= s->linesize;
735
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
736
                }
737
                CHECK_SAD_HALF_MV(x2 , +1,  0)
738
            }
739
        }else{
740
            if(l<=r){
741
                if(t+l<=b+r){
742
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
743
                    ptr+= s->linesize;
744
                }else{
745
                    ptr+= s->linesize;
746
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
747
                }
748
                CHECK_SAD_HALF_MV(x2 , -1,  0)
749
                CHECK_SAD_HALF_MV(xy2, -1, +1)
750
            }else{
751
                if(t+r<=b+l){
752
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
753
                    ptr+= s->linesize;
754
                }else{
755
                    ptr+= s->linesize;
756
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
757
                }
758
                CHECK_SAD_HALF_MV(x2 , +1,  0)
759
                CHECK_SAD_HALF_MV(xy2, +1, +1)
760
            }
761
            CHECK_SAD_HALF_MV(y2 ,  0, +1)
762
        }
763
        mx+=dx;
764
        my+=dy;
765

    
766
    }else{
767
        mx<<=1;
768
        my<<=1;
769
    }
770

    
771
    *mx_ptr = mx;
772
    *my_ptr = my;
773
    return dminh;
774
}
775

    
776
static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
777
{
778
    const int xy= s->mb_x + s->mb_y*s->mb_stride;
779
    
780
    s->p_mv_table[xy][0] = mx;
781
    s->p_mv_table[xy][1] = my;
782

    
783
    /* has allready been set to the 4 MV if 4MV is done */
784
    if(mv4){
785
        int mot_xy= s->block_index[0];
786

    
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
        mot_xy += s->block_wrap[0];
793
        s->motion_val[mot_xy  ][0]= mx;
794
        s->motion_val[mot_xy  ][1]= my;
795
        s->motion_val[mot_xy+1][0]= mx;
796
        s->motion_val[mot_xy+1][1]= my;
797
    }
798
}
799

    
800
/**
801
 * get fullpel ME search limits.
802
 * @param range the approximate search range for the old ME code, unused for EPZS and newer
803
 */
804
static inline void get_limits(MpegEncContext *s, int *range, int *xmin, int *ymin, int *xmax, int *ymax)
805
{
806
    if(s->avctx->me_range) *range= s->avctx->me_range >> 1;
807
    else                   *range= 16;
808

    
809
    if (s->unrestricted_mv) {
810
        *xmin = -16;
811
        *ymin = -16;
812
        if(s->avctx->codec->id!=CODEC_ID_MPEG4){
813
            *xmax = s->mb_width*16;
814
            *ymax = s->mb_height*16;
815
        }else {
816
            *xmax = s->width;
817
            *ymax = s->height;
818
        }
819
    } else {
820
        *xmin = 0;
821
        *ymin = 0;
822
        *xmax = s->mb_width*16 - 16;
823
        *ymax = s->mb_height*16 - 16;
824
    }
825
    
826
    //FIXME try to limit x/y min/max if me_range is set
827
}
828

    
829
static inline int h263_mv4_search(MpegEncContext *s, int xmin, int ymin, int xmax, int ymax, int mx, int my, int shift)
830
{
831
    int block;
832
    int P[10][2];
833
    int dmin_sum=0, mx4_sum=0, my4_sum=0;
834
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
835

    
836
    for(block=0; block<4; block++){
837
        int mx4, my4;
838
        int pred_x4, pred_y4;
839
        int dmin4;
840
        static const int off[4]= {2, 1, 1, -1};
841
        const int mot_stride = s->block_wrap[0];
842
        const int mot_xy = s->block_index[block];
843
//        const int block_x= (block&1);
844
//        const int block_y= (block>>1);
845
#if 1 // this saves us a bit of cliping work and shouldnt affect compression in a negative way
846
        const int rel_xmin4= xmin;
847
        const int rel_xmax4= xmax;
848
        const int rel_ymin4= ymin;
849
        const int rel_ymax4= ymax;
850
#else
851
        const int rel_xmin4= xmin - block_x*8;
852
        const int rel_xmax4= xmax - block_x*8 + 8;
853
        const int rel_ymin4= ymin - block_y*8;
854
        const int rel_ymax4= ymax - block_y*8 + 8;
855
#endif
856
        P_LEFT[0] = s->motion_val[mot_xy - 1][0];
857
        P_LEFT[1] = s->motion_val[mot_xy - 1][1];
858

    
859
        if(P_LEFT[0]       > (rel_xmax4<<shift)) P_LEFT[0]       = (rel_xmax4<<shift);
860

    
861
        /* special case for first line */
862
        if (s->mb_y == 0 && block<2) {
863
            pred_x4= P_LEFT[0];
864
            pred_y4= P_LEFT[1];
865
        } else {
866
            P_TOP[0]      = s->motion_val[mot_xy - mot_stride             ][0];
867
            P_TOP[1]      = s->motion_val[mot_xy - mot_stride             ][1];
868
            P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + off[block]][0];
869
            P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + off[block]][1];
870
            if(P_TOP[1]      > (rel_ymax4<<shift)) P_TOP[1]     = (rel_ymax4<<shift);
871
            if(P_TOPRIGHT[0] < (rel_xmin4<<shift)) P_TOPRIGHT[0]= (rel_xmin4<<shift);
872
            if(P_TOPRIGHT[0] > (rel_xmax4<<shift)) P_TOPRIGHT[0]= (rel_xmax4<<shift);
873
            if(P_TOPRIGHT[1] > (rel_ymax4<<shift)) P_TOPRIGHT[1]= (rel_ymax4<<shift);
874
    
875
            P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
876
            P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
877

    
878
//            if(s->out_format == FMT_H263){
879
                pred_x4 = P_MEDIAN[0];
880
                pred_y4 = P_MEDIAN[1];
881
#if 0
882
            }else { /* mpeg1 at least */
883
                pred_x4= P_LEFT[0];
884
                pred_y4= P_LEFT[1];
885
            }
886
#endif
887
        }
888
        P_MV1[0]= mx;
889
        P_MV1[1]= my;
890

    
891
        dmin4 = s->me.motion_search[1](s, block, &mx4, &my4, P, pred_x4, pred_y4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
892
                                       &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
893

    
894
        dmin4= s->me.sub_motion_search(s, &mx4, &my4, dmin4, rel_xmin4, rel_ymin4, rel_xmax4, rel_ymax4, 
895
                                          pred_x4, pred_y4, &s->last_picture, block, 1, mv_penalty);
896
        
897
        if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
898
            int dxy;
899
            const int offset= ((block&1) + (block>>1)*s->linesize)*8;
900
            uint8_t *dest_y = s->me.scratchpad + offset;
901

    
902
            if(s->quarter_sample){
903
                uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>2)) + (s->mb_y*16 + (my4>>2))*s->linesize + offset;
904
                dxy = ((my4 & 3) << 2) | (mx4 & 3);
905

    
906
                if(s->no_rounding)
907
                    s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , s->linesize);
908
                else
909
                    s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , s->linesize);
910
            }else{
911
                uint8_t *ref= s->last_picture.data[0] + (s->mb_x*16 + (mx4>>1)) + (s->mb_y*16 + (my4>>1))*s->linesize + offset;
912
                dxy = ((my4 & 1) << 1) | (mx4 & 1);
913

    
914
                if(s->no_rounding)
915
                    s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , s->linesize, 8);
916
                else
917
                    s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , s->linesize, 8);
918
            }
919
            dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*s->me.mb_penalty_factor;
920
        }else
921
            dmin_sum+= dmin4;
922

    
923
        if(s->quarter_sample){
924
            mx4_sum+= mx4/2;
925
            my4_sum+= my4/2;
926
        }else{
927
            mx4_sum+= mx4;
928
            my4_sum+= my4;
929
        }
930
            
931
        s->motion_val[ s->block_index[block] ][0]= mx4;
932
        s->motion_val[ s->block_index[block] ][1]= my4;
933
    }
934
    
935
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
936
        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);
937
    }
938
    
939
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
940
        int dxy;
941
        int mx, my;
942
        int offset;
943

    
944
        mx= ff_h263_round_chroma(mx4_sum);
945
        my= ff_h263_round_chroma(my4_sum);
946
        dxy = ((my & 1) << 1) | (mx & 1);
947
        
948
        offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
949
       
950
        if(s->no_rounding){
951
            s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
952
            s->dsp.put_no_rnd_pixels_tab[1][dxy](s->me.scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
953
        }else{
954
            s->dsp.put_pixels_tab       [1][dxy](s->me.scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
955
            s->dsp.put_pixels_tab       [1][dxy](s->me.scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
956
        }
957

    
958
        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);
959
        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);
960
    }
961

    
962
    switch(s->avctx->mb_cmp&0xFF){
963
    /*case FF_CMP_SSE:
964
        return dmin_sum+ 32*s->qscale*s->qscale;*/
965
    case FF_CMP_RD:
966
        return dmin_sum;
967
    default:
968
        return dmin_sum+ 11*s->me.mb_penalty_factor;
969
    }
970
}
971

    
972
void ff_estimate_p_frame_motion(MpegEncContext * s,
973
                                int mb_x, int mb_y)
974
{
975
    uint8_t *pix, *ppix;
976
    int sum, varc, vard, mx, my, range, dmin, xx, yy;
977
    int xmin, ymin, xmax, ymax;
978
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
979
    int pred_x=0, pred_y=0;
980
    int P[10][2];
981
    const int shift= 1+s->quarter_sample;
982
    int mb_type=0;
983
    uint8_t *ref_picture= s->last_picture.data[0];
984
    Picture * const pic= &s->current_picture;
985
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
986
    
987
    assert(s->quarter_sample==0 || s->quarter_sample==1);
988

    
989
    s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
990
    s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
991
    s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
992

    
993
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
994
    rel_xmin= xmin - mb_x*16;
995
    rel_xmax= xmax - mb_x*16;
996
    rel_ymin= ymin - mb_y*16;
997
    rel_ymax= ymax - mb_y*16;
998
    s->me.skip=0;
999

    
1000
    switch(s->me_method) {
1001
    case ME_ZERO:
1002
    default:
1003
        no_motion_search(s, &mx, &my);
1004
        mx-= mb_x*16;
1005
        my-= mb_y*16;
1006
        dmin = 0;
1007
        break;
1008
    case ME_FULL:
1009
        dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1010
        mx-= mb_x*16;
1011
        my-= mb_y*16;
1012
        break;
1013
    case ME_LOG:
1014
        dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1015
        mx-= mb_x*16;
1016
        my-= mb_y*16;
1017
        break;
1018
    case ME_PHODS:
1019
        dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1020
        mx-= mb_x*16;
1021
        my-= mb_y*16;
1022
        break;
1023
    case ME_X1:
1024
    case ME_EPZS:
1025
       {
1026
            const int mot_stride = s->block_wrap[0];
1027
            const int mot_xy = s->block_index[0];
1028

    
1029
            P_LEFT[0]       = s->motion_val[mot_xy - 1][0];
1030
            P_LEFT[1]       = s->motion_val[mot_xy - 1][1];
1031

    
1032
            if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);
1033

    
1034
            if(mb_y) {
1035
                P_TOP[0]      = s->motion_val[mot_xy - mot_stride    ][0];
1036
                P_TOP[1]      = s->motion_val[mot_xy - mot_stride    ][1];
1037
                P_TOPRIGHT[0] = s->motion_val[mot_xy - mot_stride + 2][0];
1038
                P_TOPRIGHT[1] = s->motion_val[mot_xy - mot_stride + 2][1];
1039
                if(P_TOP[1]      > (rel_ymax<<shift)) P_TOP[1]     = (rel_ymax<<shift);
1040
                if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1041
                if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1042
        
1043
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1044
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1045

    
1046
                if(s->out_format == FMT_H263){
1047
                    pred_x = P_MEDIAN[0];
1048
                    pred_y = P_MEDIAN[1];
1049
                }else { /* mpeg1 at least */
1050
                    pred_x= P_LEFT[0];
1051
                    pred_y= P_LEFT[1];
1052
                }
1053
            }else{
1054
                pred_x= P_LEFT[0];
1055
                pred_y= P_LEFT[1];
1056
            }
1057

    
1058
        }
1059
        dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1060
                                      &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1061
 
1062
        break;
1063
    }
1064

    
1065
    /* intra / predictive decision */
1066
    xx = mb_x * 16;
1067
    yy = mb_y * 16;
1068

    
1069
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
1070
    /* At this point (mx,my) are full-pell and the relative displacement */
1071
    ppix = ref_picture + ((yy+my) * s->linesize) + (xx+mx);
1072
    
1073
    sum = s->dsp.pix_sum(pix, s->linesize);
1074
    
1075
    varc = (s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500 + 128)>>8;
1076
    vard = (s->dsp.sse[0](NULL, pix, ppix, s->linesize)+128)>>8;
1077

    
1078
//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);
1079
    pic->mb_var   [s->mb_stride * mb_y + mb_x] = varc;
1080
    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = vard;
1081
    pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1082
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
1083
    pic->mb_var_sum    += varc;
1084
    pic->mc_mb_var_sum += vard;
1085
//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);
1086
    
1087
#if 0
1088
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1089
           varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1090
#endif
1091
    if(s->flags&CODEC_FLAG_HQ){
1092
        if (vard <= 64 || vard < varc)
1093
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1094
        else
1095
            s->scene_change_score+= s->qscale;
1096

    
1097
        if (vard*2 + 200 > varc)
1098
            mb_type|= MB_TYPE_INTRA;
1099
        if (varc*2 + 200 > vard){
1100
            mb_type|= MB_TYPE_INTER;
1101
            s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1102
                                   pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1103
        }else{
1104
            mx <<=shift;
1105
            my <<=shift;
1106
        }
1107
        if((s->flags&CODEC_FLAG_4MV)
1108
           && !s->me.skip && varc>50 && vard>10){
1109
            h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1110
            mb_type|=MB_TYPE_INTER4V;
1111

    
1112
            set_p_mv_tables(s, mx, my, 0);
1113
        }else
1114
            set_p_mv_tables(s, mx, my, 1);
1115
    }else{
1116
        mb_type= MB_TYPE_INTER;
1117

    
1118
        dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1119
                                    pred_x, pred_y, &s->last_picture, 0, 0, mv_penalty);
1120
        
1121
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1122
            dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, &s->last_picture, mv_penalty);
1123

    
1124
        if((s->flags&CODEC_FLAG_4MV)
1125
           && !s->me.skip && varc>50 && vard>10){
1126
            int dmin4= h263_mv4_search(s, rel_xmin, rel_ymin, rel_xmax, rel_ymax, mx, my, shift);
1127
            if(dmin4 < dmin){
1128
                mb_type= MB_TYPE_INTER4V;
1129
                dmin=dmin4;
1130
            }
1131
        }
1132
        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin; 
1133
        set_p_mv_tables(s, mx, my, mb_type!=MB_TYPE_INTER4V);
1134
        
1135
        if (vard <= 64 || vard < varc) {
1136
            s->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1137
        }else{
1138
            s->scene_change_score+= s->qscale;
1139
        }
1140
    }
1141

    
1142
    s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1143
}
1144

    
1145
int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1146
                                    int mb_x, int mb_y)
1147
{
1148
    int mx, my, range, dmin;
1149
    int xmin, ymin, xmax, ymax;
1150
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1151
    int pred_x=0, pred_y=0;
1152
    int P[10][2];
1153
    const int shift= 1+s->quarter_sample;
1154
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV;
1155
    const int xy= mb_x + mb_y*s->mb_stride;
1156
    
1157
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1158

    
1159
    s->me.pre_penalty_factor    = get_penalty_factor(s, s->avctx->me_pre_cmp);
1160

    
1161
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1162
    rel_xmin= xmin - mb_x*16;
1163
    rel_xmax= xmax - mb_x*16;
1164
    rel_ymin= ymin - mb_y*16;
1165
    rel_ymax= ymax - mb_y*16;
1166
    s->me.skip=0;
1167

    
1168
    P_LEFT[0]       = s->p_mv_table[xy + 1][0];
1169
    P_LEFT[1]       = s->p_mv_table[xy + 1][1];
1170

    
1171
    if(P_LEFT[0]       < (rel_xmin<<shift)) P_LEFT[0]       = (rel_xmin<<shift);
1172

    
1173
    /* special case for first line */
1174
    if (mb_y == s->mb_height-1) {
1175
        pred_x= P_LEFT[0];
1176
        pred_y= P_LEFT[1];
1177
        P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1178
        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME 
1179
    } else {
1180
        P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
1181
        P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
1182
        P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1183
        P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1184
        if(P_TOP[1]      < (rel_ymin<<shift)) P_TOP[1]     = (rel_ymin<<shift);
1185
        if(P_TOPRIGHT[0] > (rel_xmax<<shift)) P_TOPRIGHT[0]= (rel_xmax<<shift);
1186
        if(P_TOPRIGHT[1] < (rel_ymin<<shift)) P_TOPRIGHT[1]= (rel_ymin<<shift);
1187
    
1188
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1189
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1190

    
1191
        pred_x = P_MEDIAN[0];
1192
        pred_y = P_MEDIAN[1];
1193
    }
1194
    dmin = s->me.pre_motion_search(s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1195
                                   &s->last_picture, s->p_mv_table, (1<<16)>>shift, mv_penalty);
1196

    
1197
    s->p_mv_table[xy][0] = mx<<shift;
1198
    s->p_mv_table[xy][1] = my<<shift;
1199
    
1200
    return dmin;
1201
}
1202

    
1203
static int ff_estimate_motion_b(MpegEncContext * s,
1204
                       int mb_x, int mb_y, int16_t (*mv_table)[2], Picture *picture, int f_code)
1205
{
1206
    int mx, my, range, dmin;
1207
    int xmin, ymin, xmax, ymax;
1208
    int rel_xmin, rel_ymin, rel_xmax, rel_ymax;
1209
    int pred_x=0, pred_y=0;
1210
    int P[10][2];
1211
    const int shift= 1+s->quarter_sample;
1212
    const int mot_stride = s->mb_stride;
1213
    const int mot_xy = mb_y*mot_stride + mb_x;
1214
    uint8_t * const ref_picture= picture->data[0];
1215
    uint8_t * const mv_penalty= s->me.mv_penalty[f_code] + MAX_MV;
1216
    int mv_scale;
1217
        
1218
    s->me.penalty_factor    = get_penalty_factor(s, s->avctx->me_cmp);
1219
    s->me.sub_penalty_factor= get_penalty_factor(s, s->avctx->me_sub_cmp);
1220
    s->me.mb_penalty_factor = get_penalty_factor(s, s->avctx->mb_cmp);
1221

    
1222
    get_limits(s, &range, &xmin, &ymin, &xmax, &ymax);
1223
    rel_xmin= xmin - mb_x*16;
1224
    rel_xmax= xmax - mb_x*16;
1225
    rel_ymin= ymin - mb_y*16;
1226
    rel_ymax= ymax - mb_y*16;
1227

    
1228
    switch(s->me_method) {
1229
    case ME_ZERO:
1230
    default:
1231
        no_motion_search(s, &mx, &my);
1232
        dmin = 0;
1233
        mx-= mb_x*16;
1234
        my-= mb_y*16;
1235
        break;
1236
    case ME_FULL:
1237
        dmin = full_motion_search(s, &mx, &my, range, xmin, ymin, xmax, ymax, ref_picture);
1238
        mx-= mb_x*16;
1239
        my-= mb_y*16;
1240
        break;
1241
    case ME_LOG:
1242
        dmin = log_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1243
        mx-= mb_x*16;
1244
        my-= mb_y*16;
1245
        break;
1246
    case ME_PHODS:
1247
        dmin = phods_motion_search(s, &mx, &my, range / 2, xmin, ymin, xmax, ymax, ref_picture);
1248
        mx-= mb_x*16;
1249
        my-= mb_y*16;
1250
        break;
1251
    case ME_X1:
1252
    case ME_EPZS:
1253
       {
1254
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
1255
            P_LEFT[1]        = mv_table[mot_xy - 1][1];
1256

    
1257
            if(P_LEFT[0]       > (rel_xmax<<shift)) P_LEFT[0]       = (rel_xmax<<shift);
1258

    
1259
            /* special case for first line */
1260
            if (mb_y) {
1261
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
1262
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
1263
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
1264
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
1265
                if(P_TOP[1] > (rel_ymax<<shift)) P_TOP[1]= (rel_ymax<<shift);
1266
                if(P_TOPRIGHT[0] < (rel_xmin<<shift)) P_TOPRIGHT[0]= (rel_xmin<<shift);
1267
                if(P_TOPRIGHT[1] > (rel_ymax<<shift)) P_TOPRIGHT[1]= (rel_ymax<<shift);
1268
        
1269
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1270
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1271
            }
1272
            pred_x= P_LEFT[0];
1273
            pred_y= P_LEFT[1];
1274
        }
1275
        
1276
        if(mv_table == s->b_forw_mv_table){
1277
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1278
        }else{
1279
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1280
        }
1281
        
1282
        dmin = s->me.motion_search[0](s, 0, &mx, &my, P, pred_x, pred_y, rel_xmin, rel_ymin, rel_xmax, rel_ymax, 
1283
                                      picture, s->p_mv_table, mv_scale, mv_penalty);
1284
 
1285
        break;
1286
    }
1287
    
1288
    dmin= s->me.sub_motion_search(s, &mx, &my, dmin, rel_xmin, rel_ymin, rel_xmax, rel_ymax,
1289
                                   pred_x, pred_y, picture, 0, 0, mv_penalty);
1290
                                   
1291
    if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1292
        dmin= s->me.get_mb_score(s, mx, my, pred_x, pred_y, picture, mv_penalty);
1293

    
1294
//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1295
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1296
    mv_table[mot_xy][0]= mx;
1297
    mv_table[mot_xy][1]= my;
1298

    
1299
    return dmin;
1300
}
1301

    
1302
static inline int check_bidir_mv(MpegEncContext * s,
1303
                   int mb_x, int mb_y,
1304
                   int motion_fx, int motion_fy,
1305
                   int motion_bx, int motion_by,
1306
                   int pred_fx, int pred_fy,
1307
                   int pred_bx, int pred_by)
1308
{
1309
    //FIXME optimize?
1310
    //FIXME move into template?
1311
    //FIXME better f_code prediction (max mv & distance)
1312
    uint8_t * const mv_penalty= s->me.mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1313
    uint8_t *dest_y = s->me.scratchpad;
1314
    uint8_t *ptr;
1315
    int dxy;
1316
    int src_x, src_y;
1317
    int fbmin;
1318

    
1319
    if(s->quarter_sample){
1320
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1321
        src_x = mb_x * 16 + (motion_fx >> 2);
1322
        src_y = mb_y * 16 + (motion_fy >> 2);
1323
        assert(src_x >=-16 && src_x<=s->width);
1324
        assert(src_y >=-16 && src_y<=s->height);
1325

    
1326
        ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1327
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);
1328

    
1329
        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1330
        src_x = mb_x * 16 + (motion_bx >> 2);
1331
        src_y = mb_y * 16 + (motion_by >> 2);
1332
        assert(src_x >=-16 && src_x<=s->width);
1333
        assert(src_y >=-16 && src_y<=s->height);
1334
    
1335
        ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1336
        s->dsp.avg_qpel_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize);
1337
    }else{
1338
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1339
        src_x = mb_x * 16 + (motion_fx >> 1);
1340
        src_y = mb_y * 16 + (motion_fy >> 1);
1341
        assert(src_x >=-16 && src_x<=s->width);
1342
        assert(src_y >=-16 && src_y<=s->height);
1343

    
1344
        ptr = s->last_picture.data[0] + (src_y * s->linesize) + src_x;
1345
        s->dsp.put_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
1346

    
1347
        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1348
        src_x = mb_x * 16 + (motion_bx >> 1);
1349
        src_y = mb_y * 16 + (motion_by >> 1);
1350
        assert(src_x >=-16 && src_x<=s->width);
1351
        assert(src_y >=-16 && src_y<=s->height);
1352
    
1353
        ptr = s->next_picture.data[0] + (src_y * s->linesize) + src_x;
1354
        s->dsp.avg_pixels_tab[0][dxy](dest_y    , ptr    , s->linesize, 16);
1355
    }
1356

    
1357
    fbmin = (mv_penalty[motion_fx-pred_fx] + mv_penalty[motion_fy-pred_fy])*s->me.mb_penalty_factor
1358
           +(mv_penalty[motion_bx-pred_bx] + mv_penalty[motion_by-pred_by])*s->me.mb_penalty_factor
1359
           + s->dsp.mb_cmp[0](s, s->new_picture.data[0] + mb_x*16 + mb_y*16*s->linesize, dest_y, s->linesize);
1360
           
1361
    if(s->avctx->mb_cmp&FF_CMP_CHROMA){
1362
    }
1363
    //FIXME CHROMA !!!
1364
           
1365
    return fbmin;
1366
}
1367

    
1368
/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1369
static inline int bidir_refine(MpegEncContext * s,
1370
                                  int mb_x, int mb_y)
1371
{
1372
    const int mot_stride = s->mb_stride;
1373
    const int xy = mb_y *mot_stride + mb_x;
1374
    int fbmin;
1375
    int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1376
    int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1377
    int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1378
    int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1379
    int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1380
    int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1381
    int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1382
    int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1383

    
1384
    //FIXME do refinement and add flag
1385
    
1386
    fbmin= check_bidir_mv(s, mb_x, mb_y, 
1387
                          motion_fx, motion_fy,
1388
                          motion_bx, motion_by,
1389
                          pred_fx, pred_fy,
1390
                          pred_bx, pred_by);
1391

    
1392
   return fbmin;
1393
}
1394

    
1395
static inline int direct_search(MpegEncContext * s,
1396
                                int mb_x, int mb_y)
1397
{
1398
    int P[10][2];
1399
    const int mot_stride = s->mb_stride;
1400
    const int mot_xy = mb_y*mot_stride + mb_x;
1401
    const int shift= 1+s->quarter_sample;
1402
    int dmin, i;
1403
    const int time_pp= s->pp_time;
1404
    const int time_pb= s->pb_time;
1405
    int mx, my, xmin, xmax, ymin, ymax;
1406
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
1407
    uint8_t * const mv_penalty= s->me.mv_penalty[1] + MAX_MV;
1408
    
1409
    ymin= xmin=(-32)>>shift;
1410
    ymax= xmax=   31>>shift;
1411

    
1412
    if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1413
        s->mv_type= MV_TYPE_8X8;
1414
    }else{
1415
        s->mv_type= MV_TYPE_16X16;
1416
    }
1417

    
1418
    for(i=0; i<4; i++){
1419
        int index= s->block_index[i];
1420
        int min, max;
1421
    
1422
        s->me.co_located_mv[i][0]= s->motion_val[index][0];
1423
        s->me.co_located_mv[i][1]= s->motion_val[index][1];
1424
        s->me.direct_basis_mv[i][0]= s->me.co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1425
        s->me.direct_basis_mv[i][1]= s->me.co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1426
//        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);
1427
//        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);
1428

    
1429
        max= FFMAX(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1430
        min= FFMIN(s->me.direct_basis_mv[i][0], s->me.direct_basis_mv[i][0] - s->me.co_located_mv[i][0])>>shift;
1431
        max+= (2*mb_x + (i& 1))*8 + 1; // +-1 is for the simpler rounding
1432
        min+= (2*mb_x + (i& 1))*8 - 1;
1433
        xmax= FFMIN(xmax, s->width - max);
1434
        xmin= FFMAX(xmin, - 16     - min);
1435

    
1436
        max= FFMAX(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1437
        min= FFMIN(s->me.direct_basis_mv[i][1], s->me.direct_basis_mv[i][1] - s->me.co_located_mv[i][1])>>shift;
1438
        max+= (2*mb_y + (i>>1))*8 + 1; // +-1 is for the simpler rounding
1439
        min+= (2*mb_y + (i>>1))*8 - 1;
1440
        ymax= FFMIN(ymax, s->height - max);
1441
        ymin= FFMAX(ymin, - 16      - min);
1442
        
1443
        if(s->mv_type == MV_TYPE_16X16) break;
1444
    }
1445
    
1446
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1447
    
1448
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1449
        s->b_direct_mv_table[mot_xy][0]= 0;
1450
        s->b_direct_mv_table[mot_xy][1]= 0;
1451

    
1452
        return 256*256*256*64;
1453
    }
1454

    
1455
    P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1456
    P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1457

    
1458
    /* special case for first line */
1459
    if (mb_y) {
1460
        P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
1461
        P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
1462
        P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
1463
        P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
1464
    
1465
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1466
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1467
    }
1468
 
1469
    //FIXME direct_search  ptr in context!!! (needed for chroma anyway or this will get messy)   
1470
    if(s->flags&CODEC_FLAG_QPEL){
1471
        dmin = simple_direct_qpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 
1472
                                                     &s->last_picture, mv_table, 1<<14, mv_penalty);
1473
        dmin = simple_direct_qpel_qpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1474
                                                0, 0, &s->last_picture, 0, 0, mv_penalty);
1475
        
1476
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1477
            dmin= simple_direct_qpel_qpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1478
    }else{
1479
        dmin = simple_direct_hpel_epzs_motion_search(s, 0, &mx, &my, P, 0, 0, xmin, ymin, xmax, ymax, 
1480
                                                     &s->last_picture, mv_table, 1<<15, mv_penalty);
1481
        dmin = simple_direct_hpel_hpel_motion_search(s, &mx, &my, dmin, xmin, ymin, xmax, ymax,
1482
                                                0, 0, &s->last_picture, 0, 0, mv_penalty);
1483

    
1484
        if(s->avctx->me_sub_cmp != s->avctx->mb_cmp && !s->me.skip)
1485
            dmin= simple_direct_hpel_hpel_get_mb_score(s, mx, my, 0, 0, &s->last_picture, mv_penalty);
1486
    }
1487

    
1488
    s->b_direct_mv_table[mot_xy][0]= mx;
1489
    s->b_direct_mv_table[mot_xy][1]= my;
1490
    return dmin;
1491
}
1492

    
1493
void ff_estimate_b_frame_motion(MpegEncContext * s,
1494
                             int mb_x, int mb_y)
1495
{
1496
    const int penalty_factor= s->me.mb_penalty_factor;
1497
    int fmin, bmin, dmin, fbmin;
1498
    int type=0;
1499
    
1500
    dmin= direct_search(s, mb_x, mb_y);
1501

    
1502
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, &s->last_picture, s->f_code) + 3*penalty_factor;
1503
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, &s->next_picture, s->b_code) + 2*penalty_factor;
1504
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1505

    
1506
    fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1507
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1508
    {
1509
        int score= fmin;
1510
        type = MB_TYPE_FORWARD;
1511
        
1512
        // RAL: No MB_TYPE_DIRECT in MPEG-1 video (only MPEG-4)
1513
        if (s->codec_id != CODEC_ID_MPEG1VIDEO && dmin <= score){
1514
            score = dmin;
1515
            type = MB_TYPE_DIRECT;
1516
        }
1517
        if(bmin<score){
1518
            score=bmin;
1519
            type= MB_TYPE_BACKWARD; 
1520
        }
1521
        if(fbmin<score){
1522
            score=fbmin;
1523
            type= MB_TYPE_BIDIR;
1524
        }
1525
        
1526
        score= ((unsigned)(score*score + 128*256))>>16;
1527
        s->current_picture.mc_mb_var_sum += score;
1528
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1529
    }
1530

    
1531
    if(s->flags&CODEC_FLAG_HQ){
1532
        type= MB_TYPE_FORWARD | MB_TYPE_BACKWARD | MB_TYPE_BIDIR | MB_TYPE_DIRECT; //FIXME something smarter
1533
        if(dmin>256*256*16) type&= ~MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1534
    }
1535

    
1536
    s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1537
}
1538

    
1539
/* find best f_code for ME which do unlimited searches */
1540
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1541
{
1542
    if(s->me_method>=ME_EPZS){
1543
        int score[8];
1544
        int i, y;
1545
        uint8_t * fcode_tab= s->fcode_tab;
1546
        int best_fcode=-1;
1547
        int best_score=-10000000;
1548

    
1549
        for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1550

    
1551
        for(y=0; y<s->mb_height; y++){
1552
            int x;
1553
            int xy= y*s->mb_stride;
1554
            for(x=0; x<s->mb_width; x++){
1555
                if(s->mb_type[xy] & type){
1556
                    int fcode= FFMAX(fcode_tab[mv_table[xy][0] + MAX_MV],
1557
                                     fcode_tab[mv_table[xy][1] + MAX_MV]);
1558
                    int j;
1559
                    
1560
                    for(j=0; j<fcode && j<8; j++){
1561
                        if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1562
                            score[j]-= 170;
1563
                    }
1564
                }
1565
                xy++;
1566
            }
1567
        }
1568
        
1569
        for(i=1; i<8; i++){
1570
            if(score[i] > best_score){
1571
                best_score= score[i];
1572
                best_fcode= i;
1573
            }
1574
//            printf("%d %d\n", i, score[i]);
1575
        }
1576

    
1577
//    printf("fcode: %d type: %d\n", i, s->pict_type);
1578
        return best_fcode;
1579
/*        for(i=0; i<=MAX_FCODE; i++){
1580
            printf("%d ", mv_num[i]);
1581
        }
1582
        printf("\n");*/
1583
    }else{
1584
        return 1;
1585
    }
1586
}
1587

    
1588
void ff_fix_long_p_mvs(MpegEncContext * s)
1589
{
1590
    const int f_code= s->f_code;
1591
    int y, range;
1592

    
1593
    range = (((s->codec_id == CODEC_ID_MPEG1VIDEO) ? 8 : 16) << f_code);
1594
    
1595
    if(s->msmpeg4_version) range= 16;
1596
    
1597
    if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1598
    
1599
    /* clip / convert to intra 16x16 type MVs */
1600
    for(y=0; y<s->mb_height; y++){
1601
        int x;
1602
        int xy= y*s->mb_stride;
1603
        for(x=0; x<s->mb_width; x++){
1604
            if(s->mb_type[xy]&MB_TYPE_INTER){
1605
                if(   s->p_mv_table[xy][0] >=range || s->p_mv_table[xy][0] <-range
1606
                   || s->p_mv_table[xy][1] >=range || s->p_mv_table[xy][1] <-range){
1607
                    s->mb_type[xy] &= ~MB_TYPE_INTER;
1608
                    s->mb_type[xy] |= MB_TYPE_INTRA;
1609
                    s->p_mv_table[xy][0] = 0;
1610
                    s->p_mv_table[xy][1] = 0;
1611
                }
1612
            }
1613
            xy++;
1614
        }
1615
    }
1616
//printf("%d no:%d %d//\n", clip, noclip, f_code);
1617
    if(s->flags&CODEC_FLAG_4MV){
1618
        const int wrap= 2+ s->mb_width*2;
1619

    
1620
        /* clip / convert to intra 8x8 type MVs */
1621
        for(y=0; y<s->mb_height; y++){
1622
            int xy= (y*2 + 1)*wrap + 1;
1623
            int i= y*s->mb_stride;
1624
            int x;
1625

    
1626
            for(x=0; x<s->mb_width; x++){
1627
                if(s->mb_type[i]&MB_TYPE_INTER4V){
1628
                    int block;
1629
                    for(block=0; block<4; block++){
1630
                        int off= (block& 1) + (block>>1)*wrap;
1631
                        int mx= s->motion_val[ xy + off ][0];
1632
                        int my= s->motion_val[ xy + off ][1];
1633

    
1634
                        if(   mx >=range || mx <-range
1635
                           || my >=range || my <-range){
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

    
1652
    // RAL: 8 in MPEG-1, 16 in MPEG-4
1653
    int range = (((s->codec_id == CODEC_ID_MPEG1VIDEO) ? 8 : 16) << f_code);
1654
    
1655
    if(s->avctx->me_range && range > s->avctx->me_range) range= s->avctx->me_range;
1656

    
1657
    /* clip / convert to intra 16x16 type MVs */
1658
    for(y=0; y<s->mb_height; y++){
1659
        int x;
1660
        int xy= y*s->mb_stride;
1661
        for(x=0; x<s->mb_width; x++){
1662
            if (s->mb_type[xy] & type){    // RAL: "type" test added...
1663
                if(   mv_table[xy][0] >=range || mv_table[xy][0] <-range
1664
                   || mv_table[xy][1] >=range || mv_table[xy][1] <-range){
1665

    
1666
                    if(s->codec_id == CODEC_ID_MPEG1VIDEO && 0){
1667
                    }else{
1668
                        if     (mv_table[xy][0] > range-1) mv_table[xy][0]=  range-1;
1669
                        else if(mv_table[xy][0] < -range ) mv_table[xy][0]= -range;
1670
                        if     (mv_table[xy][1] > range-1) mv_table[xy][1]=  range-1;
1671
                        else if(mv_table[xy][1] < -range ) mv_table[xy][1]= -range;
1672
                    }
1673
                }
1674
            }
1675
            xy++;
1676
        }
1677
    }
1678
}