Statistics
| Branch: | Revision:

ffmpeg / libavcodec / motion_est.c @ 755bfeab

History | View | Annotate | Download (76.4 KB)

1
/*
2
 * Motion estimation
3
 * Copyright (c) 2000,2001 Fabrice Bellard.
4
 * Copyright (c) 2002-2004 Michael Niedermayer
5
 *
6
 *
7
 * This file is part of FFmpeg.
8
 *
9
 * FFmpeg is free software; you can redistribute it and/or
10
 * modify it under the terms of the GNU Lesser General Public
11
 * License as published by the Free Software Foundation; either
12
 * version 2.1 of the License, or (at your option) any later version.
13
 *
14
 * FFmpeg is distributed in the hope that it will be useful,
15
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
17
 * Lesser General Public License for more details.
18
 *
19
 * You should have received a copy of the GNU Lesser General Public
20
 * License along with FFmpeg; if not, write to the Free Software
21
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22
 *
23
 * new Motion Estimation (X1/EPZS) by Michael Niedermayer <michaelni@gmx.at>
24
 */
25

    
26
/**
27
 * @file motion_est.c
28
 * Motion estimation.
29
 */
30

    
31
#include <stdlib.h>
32
#include <stdio.h>
33
#include <limits.h>
34
#include "avcodec.h"
35
#include "dsputil.h"
36
#include "mpegvideo.h"
37

    
38
#undef NDEBUG
39
#include <assert.h>
40

    
41
#define SQ(a) ((a)*(a))
42

    
43
#define P_LEFT P[1]
44
#define P_TOP P[2]
45
#define P_TOPRIGHT P[3]
46
#define P_MEDIAN P[4]
47
#define P_MV1 P[9]
48

    
49
static inline int sad_hpel_motion_search(MpegEncContext * s,
50
                                  int *mx_ptr, int *my_ptr, int dmin,
51
                                  int src_index, int ref_index,
52
                                  int size, int h);
53

    
54
static inline int update_map_generation(MotionEstContext *c)
55
{
56
    c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
57
    if(c->map_generation==0){
58
        c->map_generation= 1<<(ME_MAP_MV_BITS*2);
59
        memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
60
    }
61
    return c->map_generation;
62
}
63

    
64
/* shape adaptive search stuff */
65
typedef struct Minima{
66
    int height;
67
    int x, y;
68
    int checked;
69
}Minima;
70

    
71
static int minima_cmp(const void *a, const void *b){
72
    const Minima *da = (const Minima *) a;
73
    const Minima *db = (const Minima *) b;
74

    
75
    return da->height - db->height;
76
}
77

    
78
#define FLAG_QPEL   1 //must be 1
79
#define FLAG_CHROMA 2
80
#define FLAG_DIRECT 4
81

    
82
static inline void init_ref(MotionEstContext *c, uint8_t *src[3], uint8_t *ref[3], uint8_t *ref2[3], int x, int y, int ref_index){
83
    const int offset[3]= {
84
          y*c->  stride + x,
85
        ((y*c->uvstride + x)>>1),
86
        ((y*c->uvstride + x)>>1),
87
    };
88
    int i;
89
    for(i=0; i<3; i++){
90
        c->src[0][i]= src [i] + offset[i];
91
        c->ref[0][i]= ref [i] + offset[i];
92
    }
93
    if(ref_index){
94
        for(i=0; i<3; i++){
95
            c->ref[ref_index][i]= ref2[i] + offset[i];
96
        }
97
    }
98
}
99

    
100
static int get_flags(MotionEstContext *c, int direct, int chroma){
101
    return   ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
102
           + (direct ? FLAG_DIRECT : 0)
103
           + (chroma ? FLAG_CHROMA : 0);
104
}
105

    
106
/*! \brief compares a block (either a full macroblock or a partition thereof)
107
    against a proposed motion-compensated prediction of that block
108
 */
109
static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
110
                      const int size, const int h, int ref_index, int src_index,
111
                      me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
112
    MotionEstContext * const c= &s->me;
113
    const int stride= c->stride;
114
    const int uvstride= c->uvstride;
115
    const int qpel= flags&FLAG_QPEL;
116
    const int chroma= flags&FLAG_CHROMA;
117
    const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
118
    const int hx= subx + (x<<(1+qpel));
119
    const int hy= suby + (y<<(1+qpel));
120
    uint8_t * const * const ref= c->ref[ref_index];
121
    uint8_t * const * const src= c->src[src_index];
122
    int d;
123
    //FIXME check chroma 4mv, (no crashes ...)
124
    if(flags&FLAG_DIRECT){
125
        assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
126
        if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
127
            const int time_pp= s->pp_time;
128
            const int time_pb= s->pb_time;
129
            const int mask= 2*qpel+1;
130
            if(s->mv_type==MV_TYPE_8X8){
131
                int i;
132
                for(i=0; i<4; i++){
133
                    int fx = c->direct_basis_mv[i][0] + hx;
134
                    int fy = c->direct_basis_mv[i][1] + hy;
135
                    int bx = hx ? fx - c->co_located_mv[i][0] : c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(qpel+4));
136
                    int by = hy ? fy - c->co_located_mv[i][1] : c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(qpel+4));
137
                    int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
138
                    int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
139

    
140
                    uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
141
                    if(qpel){
142
                        c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
143
                        c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
144
                    }else{
145
                        c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
146
                        c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
147
                    }
148
                }
149
            }else{
150
                int fx = c->direct_basis_mv[0][0] + hx;
151
                int fy = c->direct_basis_mv[0][1] + hy;
152
                int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
153
                int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
154
                int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
155
                int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
156

    
157
                if(qpel){
158
                    c->qpel_put[1][fxy](c->temp               , ref[0] + (fx>>2) + (fy>>2)*stride               , stride);
159
                    c->qpel_put[1][fxy](c->temp + 8           , ref[0] + (fx>>2) + (fy>>2)*stride + 8           , stride);
160
                    c->qpel_put[1][fxy](c->temp     + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride     + 8*stride, stride);
161
                    c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
162
                    c->qpel_avg[1][bxy](c->temp               , ref[8] + (bx>>2) + (by>>2)*stride               , stride);
163
                    c->qpel_avg[1][bxy](c->temp + 8           , ref[8] + (bx>>2) + (by>>2)*stride + 8           , stride);
164
                    c->qpel_avg[1][bxy](c->temp     + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride     + 8*stride, stride);
165
                    c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
166
                }else{
167
                    assert((fx>>1) + 16*s->mb_x >= -16);
168
                    assert((fy>>1) + 16*s->mb_y >= -16);
169
                    assert((fx>>1) + 16*s->mb_x <= s->width);
170
                    assert((fy>>1) + 16*s->mb_y <= s->height);
171
                    assert((bx>>1) + 16*s->mb_x >= -16);
172
                    assert((by>>1) + 16*s->mb_y >= -16);
173
                    assert((bx>>1) + 16*s->mb_x <= s->width);
174
                    assert((by>>1) + 16*s->mb_y <= s->height);
175

    
176
                    c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
177
                    c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
178
                }
179
            }
180
            d = cmp_func(s, c->temp, src[0], stride, 16);
181
        }else
182
            d= 256*256*256*32;
183
    }else{
184
        int uvdxy;              /* no, it might not be used uninitialized */
185
        if(dxy){
186
            if(qpel){
187
                c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
188
                if(chroma){
189
                    int cx= hx/2;
190
                    int cy= hy/2;
191
                    cx= (cx>>1)|(cx&1);
192
                    cy= (cy>>1)|(cy&1);
193
                    uvdxy= (cx&1) + 2*(cy&1);
194
                    //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
195
                }
196
            }else{
197
                c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
198
                if(chroma)
199
                    uvdxy= dxy | (x&1) | (2*(y&1));
200
            }
201
            d = cmp_func(s, c->temp, src[0], stride, h);
202
        }else{
203
            d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
204
            if(chroma)
205
                uvdxy= (x&1) + 2*(y&1);
206
        }
207
        if(chroma){
208
            uint8_t * const uvtemp= c->temp + 16*stride;
209
            c->hpel_put[size+1][uvdxy](uvtemp  , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
210
            c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
211
            d += chroma_cmp_func(s, uvtemp  , src[1], uvstride, h>>1);
212
            d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
213
        }
214
    }
215
#if 0
216
    if(full_pel){
217
        const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
218
        score_map[index]= d;
219
    }
220

221
    d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
222
#endif
223
    return d;
224
}
225

    
226
#include "motion_est_template.c"
227

    
228
static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
229
    return 0;
230
}
231

    
232
static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
233
}
234

    
235
void ff_init_me(MpegEncContext *s){
236
    MotionEstContext * const c= &s->me;
237
    int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
238
    int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
239
    c->avctx= s->avctx;
240

    
241
    if(cache_size < 2*dia_size && !c->stride){
242
        av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
243
    }
244

    
245
    ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
246
    ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
247
    ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
248
    ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
249

    
250
    c->flags    = get_flags(c, 0, c->avctx->me_cmp    &FF_CMP_CHROMA);
251
    c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
252
    c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp    &FF_CMP_CHROMA);
253

    
254
/*FIXME s->no_rounding b_type*/
255
    if(s->flags&CODEC_FLAG_QPEL){
256
        c->sub_motion_search= qpel_motion_search;
257
        c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
258
        if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
259
        else               c->qpel_put= s->dsp.put_qpel_pixels_tab;
260
    }else{
261
        if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
262
            c->sub_motion_search= hpel_motion_search;
263
        else if(   c->avctx->me_sub_cmp == FF_CMP_SAD
264
                && c->avctx->    me_cmp == FF_CMP_SAD
265
                && c->avctx->    mb_cmp == FF_CMP_SAD)
266
            c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
267
        else
268
            c->sub_motion_search= hpel_motion_search;
269
    }
270
    c->hpel_avg= s->dsp.avg_pixels_tab;
271
    if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
272
    else               c->hpel_put= s->dsp.put_pixels_tab;
273

    
274
    if(s->linesize){
275
        c->stride  = s->linesize;
276
        c->uvstride= s->uvlinesize;
277
    }else{
278
        c->stride  = 16*s->mb_width + 32;
279
        c->uvstride=  8*s->mb_width + 16;
280
    }
281

    
282
    /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
283
     * not have yet, and even if we had, the motion estimation code
284
     * does not expect it. */
285
    if(s->codec_id != CODEC_ID_SNOW){
286
        if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
287
            s->dsp.me_cmp[2]= zero_cmp;
288
        }
289
        if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
290
            s->dsp.me_sub_cmp[2]= zero_cmp;
291
        }
292
        c->hpel_put[2][0]= c->hpel_put[2][1]=
293
        c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
294
    }
295

    
296
    if(s->codec_id == CODEC_ID_H261){
297
        c->sub_motion_search= no_sub_motion_search;
298
    }
299

    
300
    c->temp= c->scratchpad;
301
}
302

    
303
#if 0
304
static int pix_dev(uint8_t * pix, int line_size, int mean)
305
{
306
    int s, i, j;
307

308
    s = 0;
309
    for (i = 0; i < 16; i++) {
310
        for (j = 0; j < 16; j += 8) {
311
            s += FFABS(pix[0]-mean);
312
            s += FFABS(pix[1]-mean);
313
            s += FFABS(pix[2]-mean);
314
            s += FFABS(pix[3]-mean);
315
            s += FFABS(pix[4]-mean);
316
            s += FFABS(pix[5]-mean);
317
            s += FFABS(pix[6]-mean);
318
            s += FFABS(pix[7]-mean);
319
            pix += 8;
320
        }
321
        pix += line_size - 16;
322
    }
323
    return s;
324
}
325
#endif
326

    
327
static inline void no_motion_search(MpegEncContext * s,
328
                                    int *mx_ptr, int *my_ptr)
329
{
330
    *mx_ptr = 16 * s->mb_x;
331
    *my_ptr = 16 * s->mb_y;
332
}
333

    
334
#if 0  /* the use of these functions is inside #if 0 */
335
static int full_motion_search(MpegEncContext * s,
336
                              int *mx_ptr, int *my_ptr, int range,
337
                              int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
338
{
339
    int x1, y1, x2, y2, xx, yy, x, y;
340
    int mx, my, dmin, d;
341
    uint8_t *pix;
342

343
    xx = 16 * s->mb_x;
344
    yy = 16 * s->mb_y;
345
    x1 = xx - range + 1;        /* we loose one pixel to avoid boundary pb with half pixel pred */
346
    if (x1 < xmin)
347
        x1 = xmin;
348
    x2 = xx + range - 1;
349
    if (x2 > xmax)
350
        x2 = xmax;
351
    y1 = yy - range + 1;
352
    if (y1 < ymin)
353
        y1 = ymin;
354
    y2 = yy + range - 1;
355
    if (y2 > ymax)
356
        y2 = ymax;
357
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
358
    dmin = 0x7fffffff;
359
    mx = 0;
360
    my = 0;
361
    for (y = y1; y <= y2; y++) {
362
        for (x = x1; x <= x2; x++) {
363
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
364
                             s->linesize, 16);
365
            if (d < dmin ||
366
                (d == dmin &&
367
                 (abs(x - xx) + abs(y - yy)) <
368
                 (abs(mx - xx) + abs(my - yy)))) {
369
                dmin = d;
370
                mx = x;
371
                my = y;
372
            }
373
        }
374
    }
375

376
    *mx_ptr = mx;
377
    *my_ptr = my;
378

379
#if 0
380
    if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
381
        *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
382
        av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
383
    }
384
#endif
385
    return dmin;
386
}
387

    
388

    
389
static int log_motion_search(MpegEncContext * s,
390
                             int *mx_ptr, int *my_ptr, int range,
391
                             int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
392
{
393
    int x1, y1, x2, y2, xx, yy, x, y;
394
    int mx, my, dmin, d;
395
    uint8_t *pix;
396

    
397
    xx = s->mb_x << 4;
398
    yy = s->mb_y << 4;
399

    
400
    /* Left limit */
401
    x1 = xx - range;
402
    if (x1 < xmin)
403
        x1 = xmin;
404

    
405
    /* Right limit */
406
    x2 = xx + range;
407
    if (x2 > xmax)
408
        x2 = xmax;
409

    
410
    /* Upper limit */
411
    y1 = yy - range;
412
    if (y1 < ymin)
413
        y1 = ymin;
414

    
415
    /* Lower limit */
416
    y2 = yy + range;
417
    if (y2 > ymax)
418
        y2 = ymax;
419

    
420
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
421
    dmin = 0x7fffffff;
422
    mx = 0;
423
    my = 0;
424

    
425
    do {
426
        for (y = y1; y <= y2; y += range) {
427
            for (x = x1; x <= x2; x += range) {
428
                d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
429
                if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
430
                    dmin = d;
431
                    mx = x;
432
                    my = y;
433
                }
434
            }
435
        }
436

    
437
        range = range >> 1;
438

    
439
        x1 = mx - range;
440
        if (x1 < xmin)
441
            x1 = xmin;
442

    
443
        x2 = mx + range;
444
        if (x2 > xmax)
445
            x2 = xmax;
446

    
447
        y1 = my - range;
448
        if (y1 < ymin)
449
            y1 = ymin;
450

    
451
        y2 = my + range;
452
        if (y2 > ymax)
453
            y2 = ymax;
454

    
455
    } while (range >= 1);
456

    
457
#ifdef DEBUG
458
    av_log(s->avctx, AV_LOG_DEBUG, "log       - MX: %d\tMY: %d\n", mx, my);
459
#endif
460
    *mx_ptr = mx;
461
    *my_ptr = my;
462
    return dmin;
463
}
464

    
465
static int phods_motion_search(MpegEncContext * s,
466
                               int *mx_ptr, int *my_ptr, int range,
467
                               int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
468
{
469
    int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
470
    int mx, my, dminx, dminy;
471
    uint8_t *pix;
472

    
473
    xx = s->mb_x << 4;
474
    yy = s->mb_y << 4;
475

    
476
    /* Left limit */
477
    x1 = xx - range;
478
    if (x1 < xmin)
479
        x1 = xmin;
480

    
481
    /* Right limit */
482
    x2 = xx + range;
483
    if (x2 > xmax)
484
        x2 = xmax;
485

    
486
    /* Upper limit */
487
    y1 = yy - range;
488
    if (y1 < ymin)
489
        y1 = ymin;
490

    
491
    /* Lower limit */
492
    y2 = yy + range;
493
    if (y2 > ymax)
494
        y2 = ymax;
495

    
496
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
497
    mx = 0;
498
    my = 0;
499

    
500
    x = xx;
501
    y = yy;
502
    do {
503
        dminx = 0x7fffffff;
504
        dminy = 0x7fffffff;
505

    
506
        lastx = x;
507
        for (x = x1; x <= x2; x += range) {
508
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
509
            if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
510
                dminx = d;
511
                mx = x;
512
            }
513
        }
514

    
515
        x = lastx;
516
        for (y = y1; y <= y2; y += range) {
517
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
518
            if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
519
                dminy = d;
520
                my = y;
521
            }
522
        }
523

    
524
        range = range >> 1;
525

    
526
        x = mx;
527
        y = my;
528
        x1 = mx - range;
529
        if (x1 < xmin)
530
            x1 = xmin;
531

    
532
        x2 = mx + range;
533
        if (x2 > xmax)
534
            x2 = xmax;
535

    
536
        y1 = my - range;
537
        if (y1 < ymin)
538
            y1 = ymin;
539

    
540
        y2 = my + range;
541
        if (y2 > ymax)
542
            y2 = ymax;
543

    
544
    } while (range >= 1);
545

    
546
#ifdef DEBUG
547
    av_log(s->avctx, AV_LOG_DEBUG, "phods     - MX: %d\tMY: %d\n", mx, my);
548
#endif
549

    
550
    /* half pixel search */
551
    *mx_ptr = mx;
552
    *my_ptr = my;
553
    return dminy;
554
}
555
#endif /* 0 */
556

    
557
#define Z_THRESHOLD 256
558

    
559
#define CHECK_SAD_HALF_MV(suffix, x, y) \
560
{\
561
    d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
562
    d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
563
    COPY3_IF_LT(dminh, d, dx, x, dy, y)\
564
}
565

    
566
static inline int sad_hpel_motion_search(MpegEncContext * s,
567
                                  int *mx_ptr, int *my_ptr, int dmin,
568
                                  int src_index, int ref_index,
569
                                  int size, int h)
570
{
571
    MotionEstContext * const c= &s->me;
572
    const int penalty_factor= c->sub_penalty_factor;
573
    int mx, my, dminh;
574
    uint8_t *pix, *ptr;
575
    int stride= c->stride;
576
    const int flags= c->sub_flags;
577
    LOAD_COMMON
578

    
579
    assert(flags == 0);
580

    
581
    if(c->skip){
582
//    printf("S");
583
        *mx_ptr = 0;
584
        *my_ptr = 0;
585
        return dmin;
586
    }
587
//    printf("N");
588

    
589
    pix = c->src[src_index][0];
590

    
591
    mx = *mx_ptr;
592
    my = *my_ptr;
593
    ptr = c->ref[ref_index][0] + (my * stride) + mx;
594

    
595
    dminh = dmin;
596

    
597
    if (mx > xmin && mx < xmax &&
598
        my > ymin && my < ymax) {
599
        int dx=0, dy=0;
600
        int d, pen_x, pen_y;
601
        const int index= (my<<ME_MAP_SHIFT) + mx;
602
        const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
603
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
604
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
605
        const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
606
        mx<<=1;
607
        my<<=1;
608

    
609

    
610
        pen_x= pred_x + mx;
611
        pen_y= pred_y + my;
612

    
613
        ptr-= stride;
614
        if(t<=b){
615
            CHECK_SAD_HALF_MV(y2 , 0, -1)
616
            if(l<=r){
617
                CHECK_SAD_HALF_MV(xy2, -1, -1)
618
                if(t+r<=b+l){
619
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
620
                    ptr+= stride;
621
                }else{
622
                    ptr+= stride;
623
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
624
                }
625
                CHECK_SAD_HALF_MV(x2 , -1,  0)
626
            }else{
627
                CHECK_SAD_HALF_MV(xy2, +1, -1)
628
                if(t+l<=b+r){
629
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
630
                    ptr+= stride;
631
                }else{
632
                    ptr+= stride;
633
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
634
                }
635
                CHECK_SAD_HALF_MV(x2 , +1,  0)
636
            }
637
        }else{
638
            if(l<=r){
639
                if(t+l<=b+r){
640
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
641
                    ptr+= stride;
642
                }else{
643
                    ptr+= stride;
644
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
645
                }
646
                CHECK_SAD_HALF_MV(x2 , -1,  0)
647
                CHECK_SAD_HALF_MV(xy2, -1, +1)
648
            }else{
649
                if(t+r<=b+l){
650
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
651
                    ptr+= stride;
652
                }else{
653
                    ptr+= stride;
654
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
655
                }
656
                CHECK_SAD_HALF_MV(x2 , +1,  0)
657
                CHECK_SAD_HALF_MV(xy2, +1, +1)
658
            }
659
            CHECK_SAD_HALF_MV(y2 ,  0, +1)
660
        }
661
        mx+=dx;
662
        my+=dy;
663

    
664
    }else{
665
        mx<<=1;
666
        my<<=1;
667
    }
668

    
669
    *mx_ptr = mx;
670
    *my_ptr = my;
671
    return dminh;
672
}
673

    
674
static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
675
{
676
    const int xy= s->mb_x + s->mb_y*s->mb_stride;
677

    
678
    s->p_mv_table[xy][0] = mx;
679
    s->p_mv_table[xy][1] = my;
680

    
681
    /* has already been set to the 4 MV if 4MV is done */
682
    if(mv4){
683
        int mot_xy= s->block_index[0];
684

    
685
        s->current_picture.motion_val[0][mot_xy  ][0]= mx;
686
        s->current_picture.motion_val[0][mot_xy  ][1]= my;
687
        s->current_picture.motion_val[0][mot_xy+1][0]= mx;
688
        s->current_picture.motion_val[0][mot_xy+1][1]= my;
689

    
690
        mot_xy += s->b8_stride;
691
        s->current_picture.motion_val[0][mot_xy  ][0]= mx;
692
        s->current_picture.motion_val[0][mot_xy  ][1]= my;
693
        s->current_picture.motion_val[0][mot_xy+1][0]= mx;
694
        s->current_picture.motion_val[0][mot_xy+1][1]= my;
695
    }
696
}
697

    
698
/**
699
 * get fullpel ME search limits.
700
 */
701
static inline void get_limits(MpegEncContext *s, int x, int y)
702
{
703
    MotionEstContext * const c= &s->me;
704
    int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
705
/*
706
    if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
707
    else                   c->range= 16;
708
*/
709
    if (s->unrestricted_mv) {
710
        c->xmin = - x - 16;
711
        c->ymin = - y - 16;
712
        c->xmax = - x + s->mb_width *16;
713
        c->ymax = - y + s->mb_height*16;
714
    } else if (s->out_format == FMT_H261){
715
        // Search range of H261 is different from other codec standards
716
        c->xmin = (x > 15) ? - 15 : 0;
717
        c->ymin = (y > 15) ? - 15 : 0;
718
        c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
719
        c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
720
    } else {
721
        c->xmin = - x;
722
        c->ymin = - y;
723
        c->xmax = - x + s->mb_width *16 - 16;
724
        c->ymax = - y + s->mb_height*16 - 16;
725
    }
726
    if(range){
727
        c->xmin = FFMAX(c->xmin,-range);
728
        c->xmax = FFMIN(c->xmax, range);
729
        c->ymin = FFMAX(c->ymin,-range);
730
        c->ymax = FFMIN(c->ymax, range);
731
    }
732
}
733

    
734
static inline void init_mv4_ref(MotionEstContext *c){
735
    const int stride= c->stride;
736

    
737
    c->ref[1][0] = c->ref[0][0] + 8;
738
    c->ref[2][0] = c->ref[0][0] + 8*stride;
739
    c->ref[3][0] = c->ref[2][0] + 8;
740
    c->src[1][0] = c->src[0][0] + 8;
741
    c->src[2][0] = c->src[0][0] + 8*stride;
742
    c->src[3][0] = c->src[2][0] + 8;
743
}
744

    
745
static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
746
{
747
    MotionEstContext * const c= &s->me;
748
    const int size= 1;
749
    const int h=8;
750
    int block;
751
    int P[10][2];
752
    int dmin_sum=0, mx4_sum=0, my4_sum=0;
753
    int same=1;
754
    const int stride= c->stride;
755
    uint8_t *mv_penalty= c->current_mv_penalty;
756

    
757
    init_mv4_ref(c);
758

    
759
    for(block=0; block<4; block++){
760
        int mx4, my4;
761
        int pred_x4, pred_y4;
762
        int dmin4;
763
        static const int off[4]= {2, 1, 1, -1};
764
        const int mot_stride = s->b8_stride;
765
        const int mot_xy = s->block_index[block];
766

    
767
        P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
768
        P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
769

    
770
        if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
771

    
772
        /* special case for first line */
773
        if (s->first_slice_line && block<2) {
774
            c->pred_x= pred_x4= P_LEFT[0];
775
            c->pred_y= pred_y4= P_LEFT[1];
776
        } else {
777
            P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][0];
778
            P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][1];
779
            P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
780
            P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
781
            if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
782
            if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
783
            if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
784
            if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
785

    
786
            P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
787
            P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
788

    
789
            c->pred_x= pred_x4 = P_MEDIAN[0];
790
            c->pred_y= pred_y4 = P_MEDIAN[1];
791
        }
792
        P_MV1[0]= mx;
793
        P_MV1[1]= my;
794

    
795
        dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
796

    
797
        dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
798

    
799
        if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
800
            int dxy;
801
            const int offset= ((block&1) + (block>>1)*stride)*8;
802
            uint8_t *dest_y = c->scratchpad + offset;
803
            if(s->quarter_sample){
804
                uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
805
                dxy = ((my4 & 3) << 2) | (mx4 & 3);
806

    
807
                if(s->no_rounding)
808
                    s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , stride);
809
                else
810
                    s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , stride);
811
            }else{
812
                uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
813
                dxy = ((my4 & 1) << 1) | (mx4 & 1);
814

    
815
                if(s->no_rounding)
816
                    s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , stride, h);
817
                else
818
                    s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , stride, h);
819
            }
820
            dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
821
        }else
822
            dmin_sum+= dmin4;
823

    
824
        if(s->quarter_sample){
825
            mx4_sum+= mx4/2;
826
            my4_sum+= my4/2;
827
        }else{
828
            mx4_sum+= mx4;
829
            my4_sum+= my4;
830
        }
831

    
832
        s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
833
        s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
834

    
835
        if(mx4 != mx || my4 != my) same=0;
836
    }
837

    
838
    if(same)
839
        return INT_MAX;
840

    
841
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
842
        dmin_sum += s->dsp.mb_cmp[0](s, s->new_picture.data[0] + s->mb_x*16 + s->mb_y*16*stride, c->scratchpad, stride, 16);
843
    }
844

    
845
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
846
        int dxy;
847
        int mx, my;
848
        int offset;
849

    
850
        mx= ff_h263_round_chroma(mx4_sum);
851
        my= ff_h263_round_chroma(my4_sum);
852
        dxy = ((my & 1) << 1) | (mx & 1);
853

    
854
        offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
855

    
856
        if(s->no_rounding){
857
            s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
858
            s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
859
        }else{
860
            s->dsp.put_pixels_tab       [1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
861
            s->dsp.put_pixels_tab       [1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
862
        }
863

    
864
        dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[1] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad  , s->uvlinesize, 8);
865
        dmin_sum += s->dsp.mb_cmp[1](s, s->new_picture.data[2] + s->mb_x*8 + s->mb_y*8*s->uvlinesize, c->scratchpad+8, s->uvlinesize, 8);
866
    }
867

    
868
    c->pred_x= mx;
869
    c->pred_y= my;
870

    
871
    switch(c->avctx->mb_cmp&0xFF){
872
    /*case FF_CMP_SSE:
873
        return dmin_sum+ 32*s->qscale*s->qscale;*/
874
    case FF_CMP_RD:
875
        return dmin_sum;
876
    default:
877
        return dmin_sum+ 11*c->mb_penalty_factor;
878
    }
879
}
880

    
881
static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
882
    MotionEstContext * const c= &s->me;
883

    
884
    c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
885
    c->src[1][0] = c->src[0][0] + s->linesize;
886
    if(c->flags & FLAG_CHROMA){
887
        c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
888
        c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
889
        c->src[1][1] = c->src[0][1] + s->uvlinesize;
890
        c->src[1][2] = c->src[0][2] + s->uvlinesize;
891
    }
892
}
893

    
894
static int interlaced_search(MpegEncContext *s, int ref_index,
895
                             int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
896
{
897
    MotionEstContext * const c= &s->me;
898
    const int size=0;
899
    const int h=8;
900
    int block;
901
    int P[10][2];
902
    uint8_t * const mv_penalty= c->current_mv_penalty;
903
    int same=1;
904
    const int stride= 2*s->linesize;
905
    int dmin_sum= 0;
906
    const int mot_stride= s->mb_stride;
907
    const int xy= s->mb_x + s->mb_y*mot_stride;
908

    
909
    c->ymin>>=1;
910
    c->ymax>>=1;
911
    c->stride<<=1;
912
    c->uvstride<<=1;
913
    init_interlaced_ref(s, ref_index);
914

    
915
    for(block=0; block<2; block++){
916
        int field_select;
917
        int best_dmin= INT_MAX;
918
        int best_field= -1;
919

    
920
        for(field_select=0; field_select<2; field_select++){
921
            int dmin, mx_i, my_i;
922
            int16_t (*mv_table)[2]= mv_tables[block][field_select];
923

    
924
            if(user_field_select){
925
                assert(field_select==0 || field_select==1);
926
                assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
927
                if(field_select_tables[block][xy] != field_select)
928
                    continue;
929
            }
930

    
931
            P_LEFT[0] = mv_table[xy - 1][0];
932
            P_LEFT[1] = mv_table[xy - 1][1];
933
            if(P_LEFT[0]       > (c->xmax<<1)) P_LEFT[0]       = (c->xmax<<1);
934

    
935
            c->pred_x= P_LEFT[0];
936
            c->pred_y= P_LEFT[1];
937

    
938
            if(!s->first_slice_line){
939
                P_TOP[0]      = mv_table[xy - mot_stride][0];
940
                P_TOP[1]      = mv_table[xy - mot_stride][1];
941
                P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
942
                P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
943
                if(P_TOP[1]      > (c->ymax<<1)) P_TOP[1]     = (c->ymax<<1);
944
                if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
945
                if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
946
                if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
947

    
948
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
949
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
950
            }
951
            P_MV1[0]= mx; //FIXME not correct if block != field_select
952
            P_MV1[1]= my / 2;
953

    
954
            dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
955

    
956
            dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
957

    
958
            mv_table[xy][0]= mx_i;
959
            mv_table[xy][1]= my_i;
960

    
961
            if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
962
                int dxy;
963

    
964
                //FIXME chroma ME
965
                uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
966
                dxy = ((my_i & 1) << 1) | (mx_i & 1);
967

    
968
                if(s->no_rounding){
969
                    s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref    , stride, h);
970
                }else{
971
                    s->dsp.put_pixels_tab       [size][dxy](c->scratchpad, ref    , stride, h);
972
                }
973
                dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
974
                dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
975
            }else
976
                dmin+= c->mb_penalty_factor; //field_select bits
977

    
978
            dmin += field_select != block; //slightly prefer same field
979

    
980
            if(dmin < best_dmin){
981
                best_dmin= dmin;
982
                best_field= field_select;
983
            }
984
        }
985
        {
986
            int16_t (*mv_table)[2]= mv_tables[block][best_field];
987

    
988
            if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
989
            if(mv_table[xy][1]&1) same=0;
990
            if(mv_table[xy][1]*2 != my) same=0;
991
            if(best_field != block) same=0;
992
        }
993

    
994
        field_select_tables[block][xy]= best_field;
995
        dmin_sum += best_dmin;
996
    }
997

    
998
    c->ymin<<=1;
999
    c->ymax<<=1;
1000
    c->stride>>=1;
1001
    c->uvstride>>=1;
1002

    
1003
    if(same)
1004
        return INT_MAX;
1005

    
1006
    switch(c->avctx->mb_cmp&0xFF){
1007
    /*case FF_CMP_SSE:
1008
        return dmin_sum+ 32*s->qscale*s->qscale;*/
1009
    case FF_CMP_RD:
1010
        return dmin_sum;
1011
    default:
1012
        return dmin_sum+ 11*c->mb_penalty_factor;
1013
    }
1014
}
1015

    
1016
static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
1017
    int ymax= s->me.ymax>>interlaced;
1018
    int ymin= s->me.ymin>>interlaced;
1019

    
1020
    if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
1021
    if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
1022
    if(mv[1] <       ymin) mv[1] =       ymin;
1023
    if(mv[1] >       ymax) mv[1] =       ymax;
1024
}
1025

    
1026
static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1027
    MotionEstContext * const c= &s->me;
1028
    Picture *p= s->current_picture_ptr;
1029
    int mb_xy= mb_x + mb_y*s->mb_stride;
1030
    int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1031
    int mb_type= s->current_picture.mb_type[mb_xy];
1032
    int flags= c->flags;
1033
    int shift= (flags&FLAG_QPEL) + 1;
1034
    int mask= (1<<shift)-1;
1035
    int x, y, i;
1036
    int d=0;
1037
    me_cmp_func cmpf= s->dsp.sse[0];
1038
    me_cmp_func chroma_cmpf= s->dsp.sse[1];
1039

    
1040
    if(p_type && USES_LIST(mb_type, 1)){
1041
        av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1042
        return INT_MAX/2;
1043
    }
1044
    assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1045

    
1046
    for(i=0; i<4; i++){
1047
        int xy= s->block_index[i];
1048
        clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1049
        clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1050
    }
1051

    
1052
    if(IS_INTERLACED(mb_type)){
1053
        int xy2= xy  + s->b8_stride;
1054
        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1055
        c->stride<<=1;
1056
        c->uvstride<<=1;
1057

    
1058
        if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1059
            av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1060
            return INT_MAX/2;
1061
        }
1062

    
1063
        if(USES_LIST(mb_type, 0)){
1064
            int field_select0= p->ref_index[0][xy ];
1065
            int field_select1= p->ref_index[0][xy2];
1066
            assert(field_select0==0 ||field_select0==1);
1067
            assert(field_select1==0 ||field_select1==1);
1068
            init_interlaced_ref(s, 0);
1069

    
1070
            if(p_type){
1071
                s->p_field_select_table[0][mb_xy]= field_select0;
1072
                s->p_field_select_table[1][mb_xy]= field_select1;
1073
                *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1074
                *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1075
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1076
            }else{
1077
                s->b_field_select_table[0][0][mb_xy]= field_select0;
1078
                s->b_field_select_table[0][1][mb_xy]= field_select1;
1079
                *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1080
                *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1081
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1082
            }
1083

    
1084
            x= p->motion_val[0][xy ][0];
1085
            y= p->motion_val[0][xy ][1];
1086
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1087
            x= p->motion_val[0][xy2][0];
1088
            y= p->motion_val[0][xy2][1];
1089
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1090
        }
1091
        if(USES_LIST(mb_type, 1)){
1092
            int field_select0= p->ref_index[1][xy ];
1093
            int field_select1= p->ref_index[1][xy2];
1094
            assert(field_select0==0 ||field_select0==1);
1095
            assert(field_select1==0 ||field_select1==1);
1096
            init_interlaced_ref(s, 2);
1097

    
1098
            s->b_field_select_table[1][0][mb_xy]= field_select0;
1099
            s->b_field_select_table[1][1][mb_xy]= field_select1;
1100
            *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1101
            *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1102
            if(USES_LIST(mb_type, 0)){
1103
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1104
            }else{
1105
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1106
            }
1107

    
1108
            x= p->motion_val[1][xy ][0];
1109
            y= p->motion_val[1][xy ][1];
1110
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1111
            x= p->motion_val[1][xy2][0];
1112
            y= p->motion_val[1][xy2][1];
1113
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1114
            //FIXME bidir scores
1115
        }
1116
        c->stride>>=1;
1117
        c->uvstride>>=1;
1118
    }else if(IS_8X8(mb_type)){
1119
        if(!(s->flags & CODEC_FLAG_4MV)){
1120
            av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1121
            return INT_MAX/2;
1122
        }
1123
        cmpf= s->dsp.sse[1];
1124
        chroma_cmpf= s->dsp.sse[1];
1125
        init_mv4_ref(c);
1126
        for(i=0; i<4; i++){
1127
            xy= s->block_index[i];
1128
            x= p->motion_val[0][xy][0];
1129
            y= p->motion_val[0][xy][1];
1130
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1131
        }
1132
        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1133
    }else{
1134
        if(USES_LIST(mb_type, 0)){
1135
            if(p_type){
1136
                *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1137
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1138
            }else if(USES_LIST(mb_type, 1)){
1139
                *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1140
                *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1141
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1142
            }else{
1143
                *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1144
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1145
            }
1146
            x= p->motion_val[0][xy][0];
1147
            y= p->motion_val[0][xy][1];
1148
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1149
        }else if(USES_LIST(mb_type, 1)){
1150
            *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1151
            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1152

    
1153
            x= p->motion_val[1][xy][0];
1154
            y= p->motion_val[1][xy][1];
1155
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1156
        }else
1157
            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1158
    }
1159
    return d;
1160
}
1161

    
1162
void ff_estimate_p_frame_motion(MpegEncContext * s,
1163
                                int mb_x, int mb_y)
1164
{
1165
    MotionEstContext * const c= &s->me;
1166
    uint8_t *pix, *ppix;
1167
    int sum, mx, my, dmin;
1168
    int varc;            ///< the variance of the block (sum of squared (p[y][x]-average))
1169
    int vard;            ///< sum of squared differences with the estimated motion vector
1170
    int P[10][2];
1171
    const int shift= 1+s->quarter_sample;
1172
    int mb_type=0;
1173
    Picture * const pic= &s->current_picture;
1174

    
1175
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1176

    
1177
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1178
    assert(s->linesize == c->stride);
1179
    assert(s->uvlinesize == c->uvstride);
1180

    
1181
    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1182
    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1183
    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1184
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1185

    
1186
    get_limits(s, 16*mb_x, 16*mb_y);
1187
    c->skip=0;
1188

    
1189
    /* intra / predictive decision */
1190
    pix = c->src[0][0];
1191
    sum = s->dsp.pix_sum(pix, s->linesize);
1192
    varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1193

    
1194
    pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1195
    pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1196
    c->mb_var_sum_temp += (varc+128)>>8;
1197

    
1198
    if(c->avctx->me_threshold){
1199
        vard= check_input_motion(s, mb_x, mb_y, 1);
1200

    
1201
        if((vard+128)>>8 < c->avctx->me_threshold){
1202
            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1203
            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1204
            pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1205
            c->mc_mb_var_sum_temp += (vard+128)>>8;
1206
            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1207
            return;
1208
        }
1209
        if((vard+128)>>8 < c->avctx->mb_threshold)
1210
            mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1211
    }
1212

    
1213
    switch(s->me_method) {
1214
    case ME_ZERO:
1215
    default:
1216
        no_motion_search(s, &mx, &my);
1217
        mx-= mb_x*16;
1218
        my-= mb_y*16;
1219
        dmin = 0;
1220
        break;
1221
#if 0
1222
    case ME_FULL:
1223
        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1224
        mx-= mb_x*16;
1225
        my-= mb_y*16;
1226
        break;
1227
    case ME_LOG:
1228
        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1229
        mx-= mb_x*16;
1230
        my-= mb_y*16;
1231
        break;
1232
    case ME_PHODS:
1233
        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1234
        mx-= mb_x*16;
1235
        my-= mb_y*16;
1236
        break;
1237
#endif
1238
    case ME_X1:
1239
    case ME_EPZS:
1240
       {
1241
            const int mot_stride = s->b8_stride;
1242
            const int mot_xy = s->block_index[0];
1243

    
1244
            P_LEFT[0]       = s->current_picture.motion_val[0][mot_xy - 1][0];
1245
            P_LEFT[1]       = s->current_picture.motion_val[0][mot_xy - 1][1];
1246

    
1247
            if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
1248

    
1249
            if(!s->first_slice_line) {
1250
                P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][0];
1251
                P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][1];
1252
                P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1253
                P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1254
                if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
1255
                if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1256
                if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1257

    
1258
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1259
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1260

    
1261
                if(s->out_format == FMT_H263){
1262
                    c->pred_x = P_MEDIAN[0];
1263
                    c->pred_y = P_MEDIAN[1];
1264
                }else { /* mpeg1 at least */
1265
                    c->pred_x= P_LEFT[0];
1266
                    c->pred_y= P_LEFT[1];
1267
                }
1268
            }else{
1269
                c->pred_x= P_LEFT[0];
1270
                c->pred_y= P_LEFT[1];
1271
            }
1272

    
1273
        }
1274
        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1275

    
1276
        break;
1277
    }
1278

    
1279
    /* At this point (mx,my) are full-pell and the relative displacement */
1280
    ppix = c->ref[0][0] + (my * s->linesize) + mx;
1281

    
1282
    vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1283

    
1284
    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1285
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1286
    c->mc_mb_var_sum_temp += (vard+128)>>8;
1287

    
1288
#if 0
1289
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1290
           varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1291
#endif
1292
    if(mb_type){
1293
        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1294
        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1295
        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1296

    
1297
        if(mb_type == CANDIDATE_MB_TYPE_INTER){
1298
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1299
            set_p_mv_tables(s, mx, my, 1);
1300
        }else{
1301
            mx <<=shift;
1302
            my <<=shift;
1303
        }
1304
        if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1305
            h263_mv4_search(s, mx, my, shift);
1306

    
1307
            set_p_mv_tables(s, mx, my, 0);
1308
        }
1309
        if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1310
            interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1311
        }
1312
    }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1313
        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1314
        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1315
        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1316

    
1317
        if (vard*2 + 200*256 > varc)
1318
            mb_type|= CANDIDATE_MB_TYPE_INTRA;
1319
        if (varc*2 + 200*256 > vard || s->qscale > 24){
1320
//        if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1321
            mb_type|= CANDIDATE_MB_TYPE_INTER;
1322
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1323
            if(s->flags&CODEC_FLAG_MV0)
1324
                if(mx || my)
1325
                    mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1326
        }else{
1327
            mx <<=shift;
1328
            my <<=shift;
1329
        }
1330
        if((s->flags&CODEC_FLAG_4MV)
1331
           && !c->skip && varc>50<<8 && vard>10<<8){
1332
            if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1333
                mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1334

    
1335
            set_p_mv_tables(s, mx, my, 0);
1336
        }else
1337
            set_p_mv_tables(s, mx, my, 1);
1338
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
1339
           && !c->skip){ //FIXME varc/d checks
1340
            if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1341
                mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1342
        }
1343
    }else{
1344
        int intra_score, i;
1345
        mb_type= CANDIDATE_MB_TYPE_INTER;
1346

    
1347
        dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1348
        if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1349
            dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1350

    
1351
        if((s->flags&CODEC_FLAG_4MV)
1352
           && !c->skip && varc>50<<8 && vard>10<<8){
1353
            int dmin4= h263_mv4_search(s, mx, my, shift);
1354
            if(dmin4 < dmin){
1355
                mb_type= CANDIDATE_MB_TYPE_INTER4V;
1356
                dmin=dmin4;
1357
            }
1358
        }
1359
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
1360
           && !c->skip){ //FIXME varc/d checks
1361
            int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1362
            if(dmin_i < dmin){
1363
                mb_type = CANDIDATE_MB_TYPE_INTER_I;
1364
                dmin= dmin_i;
1365
            }
1366
        }
1367

    
1368
//        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1369
        set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1370

    
1371
        /* get intra luma score */
1372
        if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1373
            intra_score= varc - 500;
1374
        }else{
1375
            int mean= (sum+128)>>8;
1376
            mean*= 0x01010101;
1377

    
1378
            for(i=0; i<16; i++){
1379
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1380
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1381
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1382
                *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1383
            }
1384

    
1385
            intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1386
        }
1387
#if 0 //FIXME
1388
        /* get chroma score */
1389
        if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1390
            for(i=1; i<3; i++){
1391
                uint8_t *dest_c;
1392
                int mean;
1393

1394
                if(s->out_format == FMT_H263){
1395
                    mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1396
                }else{
1397
                    mean= (s->last_dc[i] + 4)>>3;
1398
                }
1399
                dest_c = s->new_picture.data[i] + (mb_y * 8  * (s->uvlinesize)) + mb_x * 8;
1400

1401
                mean*= 0x01010101;
1402
                for(i=0; i<8; i++){
1403
                    *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1404
                    *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1405
                }
1406

1407
                intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1408
            }
1409
        }
1410
#endif
1411
        intra_score += c->mb_penalty_factor*16;
1412

    
1413
        if(intra_score < dmin){
1414
            mb_type= CANDIDATE_MB_TYPE_INTRA;
1415
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1416
        }else
1417
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1418

    
1419
        {
1420
            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1421
            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1422
            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1423
        }
1424
    }
1425

    
1426
    s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1427
}
1428

    
1429
int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1430
                                    int mb_x, int mb_y)
1431
{
1432
    MotionEstContext * const c= &s->me;
1433
    int mx, my, dmin;
1434
    int P[10][2];
1435
    const int shift= 1+s->quarter_sample;
1436
    const int xy= mb_x + mb_y*s->mb_stride;
1437
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1438

    
1439
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1440

    
1441
    c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1442
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1443

    
1444
    get_limits(s, 16*mb_x, 16*mb_y);
1445
    c->skip=0;
1446

    
1447
    P_LEFT[0]       = s->p_mv_table[xy + 1][0];
1448
    P_LEFT[1]       = s->p_mv_table[xy + 1][1];
1449

    
1450
    if(P_LEFT[0]       < (c->xmin<<shift)) P_LEFT[0]       = (c->xmin<<shift);
1451

    
1452
    /* special case for first line */
1453
    if (s->first_slice_line) {
1454
        c->pred_x= P_LEFT[0];
1455
        c->pred_y= P_LEFT[1];
1456
        P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1457
        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1458
    } else {
1459
        P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
1460
        P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
1461
        P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1462
        P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1463
        if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);
1464
        if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1465
        if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1466

    
1467
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1468
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1469

    
1470
        c->pred_x = P_MEDIAN[0];
1471
        c->pred_y = P_MEDIAN[1];
1472
    }
1473

    
1474
    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1475

    
1476
    s->p_mv_table[xy][0] = mx<<shift;
1477
    s->p_mv_table[xy][1] = my<<shift;
1478

    
1479
    return dmin;
1480
}
1481

    
1482
static int ff_estimate_motion_b(MpegEncContext * s,
1483
                       int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1484
{
1485
    MotionEstContext * const c= &s->me;
1486
    int mx, my, dmin;
1487
    int P[10][2];
1488
    const int shift= 1+s->quarter_sample;
1489
    const int mot_stride = s->mb_stride;
1490
    const int mot_xy = mb_y*mot_stride + mb_x;
1491
    uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1492
    int mv_scale;
1493

    
1494
    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1495
    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1496
    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1497
    c->current_mv_penalty= mv_penalty;
1498

    
1499
    get_limits(s, 16*mb_x, 16*mb_y);
1500

    
1501
    switch(s->me_method) {
1502
    case ME_ZERO:
1503
    default:
1504
        no_motion_search(s, &mx, &my);
1505
        dmin = 0;
1506
        mx-= mb_x*16;
1507
        my-= mb_y*16;
1508
        break;
1509
#if 0
1510
    case ME_FULL:
1511
        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1512
        mx-= mb_x*16;
1513
        my-= mb_y*16;
1514
        break;
1515
    case ME_LOG:
1516
        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1517
        mx-= mb_x*16;
1518
        my-= mb_y*16;
1519
        break;
1520
    case ME_PHODS:
1521
        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1522
        mx-= mb_x*16;
1523
        my-= mb_y*16;
1524
        break;
1525
#endif
1526
    case ME_X1:
1527
    case ME_EPZS:
1528
       {
1529
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
1530
            P_LEFT[1]        = mv_table[mot_xy - 1][1];
1531

    
1532
            if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
1533

    
1534
            /* special case for first line */
1535
            if (!s->first_slice_line) {
1536
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
1537
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
1538
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
1539
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
1540
                if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1541
                if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1542
                if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1543

    
1544
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1545
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1546
            }
1547
            c->pred_x= P_LEFT[0];
1548
            c->pred_y= P_LEFT[1];
1549
        }
1550

    
1551
        if(mv_table == s->b_forw_mv_table){
1552
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1553
        }else{
1554
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1555
        }
1556

    
1557
        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1558

    
1559
        break;
1560
    }
1561

    
1562
    dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1563

    
1564
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1565
        dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1566

    
1567
//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1568
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1569
    mv_table[mot_xy][0]= mx;
1570
    mv_table[mot_xy][1]= my;
1571

    
1572
    return dmin;
1573
}
1574

    
1575
static inline int check_bidir_mv(MpegEncContext * s,
1576
                   int motion_fx, int motion_fy,
1577
                   int motion_bx, int motion_by,
1578
                   int pred_fx, int pred_fy,
1579
                   int pred_bx, int pred_by,
1580
                   int size, int h)
1581
{
1582
    //FIXME optimize?
1583
    //FIXME better f_code prediction (max mv & distance)
1584
    //FIXME pointers
1585
    MotionEstContext * const c= &s->me;
1586
    uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1587
    uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1588
    int stride= c->stride;
1589
    uint8_t *dest_y = c->scratchpad;
1590
    uint8_t *ptr;
1591
    int dxy;
1592
    int src_x, src_y;
1593
    int fbmin;
1594
    uint8_t **src_data= c->src[0];
1595
    uint8_t **ref_data= c->ref[0];
1596
    uint8_t **ref2_data= c->ref[2];
1597

    
1598
    if(s->quarter_sample){
1599
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1600
        src_x = motion_fx >> 2;
1601
        src_y = motion_fy >> 2;
1602

    
1603
        ptr = ref_data[0] + (src_y * stride) + src_x;
1604
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , stride);
1605

    
1606
        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1607
        src_x = motion_bx >> 2;
1608
        src_y = motion_by >> 2;
1609

    
1610
        ptr = ref2_data[0] + (src_y * stride) + src_x;
1611
        s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);
1612
    }else{
1613
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1614
        src_x = motion_fx >> 1;
1615
        src_y = motion_fy >> 1;
1616

    
1617
        ptr = ref_data[0] + (src_y * stride) + src_x;
1618
        s->dsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
1619

    
1620
        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1621
        src_x = motion_bx >> 1;
1622
        src_y = motion_by >> 1;
1623

    
1624
        ptr = ref2_data[0] + (src_y * stride) + src_x;
1625
        s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
1626
    }
1627

    
1628
    fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1629
           +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1630
           + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1631

    
1632
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1633
    }
1634
    //FIXME CHROMA !!!
1635

    
1636
    return fbmin;
1637
}
1638

    
1639
/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1640
static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1641
{
1642
    MotionEstContext * const c= &s->me;
1643
    const int mot_stride = s->mb_stride;
1644
    const int xy = mb_y *mot_stride + mb_x;
1645
    int fbmin;
1646
    int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1647
    int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1648
    int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1649
    int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1650
    int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1651
    int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1652
    int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1653
    int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1654
    const int flags= c->sub_flags;
1655
    const int qpel= flags&FLAG_QPEL;
1656
    const int shift= 1+qpel;
1657
    const int xmin= c->xmin<<shift;
1658
    const int ymin= c->ymin<<shift;
1659
    const int xmax= c->xmax<<shift;
1660
    const int ymax= c->ymax<<shift;
1661
    uint8_t map[8][8][8][8];
1662

    
1663
    memset(map,0,sizeof(map));
1664
#define BIDIR_MAP(fx,fy,bx,by) \
1665
    map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1666
    BIDIR_MAP(0,0,0,0) = 1;
1667

    
1668
    fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1669
                          motion_bx, motion_by,
1670
                          pred_fx, pred_fy,
1671
                          pred_bx, pred_by,
1672
                          0, 16);
1673

    
1674
    if(s->avctx->bidir_refine){
1675
        int score, end;
1676
#define CHECK_BIDIR(fx,fy,bx,by)\
1677
    if( !BIDIR_MAP(fx,fy,bx,by)\
1678
       &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1679
       &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1680
        BIDIR_MAP(fx,fy,bx,by) = 1;\
1681
        score= check_bidir_mv(s, motion_fx+fx, motion_fy+fy, motion_bx+bx, motion_by+by, pred_fx, pred_fy, pred_bx, pred_by, 0, 16);\
1682
        if(score < fbmin){\
1683
            fbmin= score;\
1684
            motion_fx+=fx;\
1685
            motion_fy+=fy;\
1686
            motion_bx+=bx;\
1687
            motion_by+=by;\
1688
            end=0;\
1689
        }\
1690
    }
1691
#define CHECK_BIDIR2(a,b,c,d)\
1692
CHECK_BIDIR(a,b,c,d)\
1693
CHECK_BIDIR(-(a),-(b),-(c),-(d))
1694

    
1695
#define CHECK_BIDIRR(a,b,c,d)\
1696
CHECK_BIDIR2(a,b,c,d)\
1697
CHECK_BIDIR2(b,c,d,a)\
1698
CHECK_BIDIR2(c,d,a,b)\
1699
CHECK_BIDIR2(d,a,b,c)
1700

    
1701
        do{
1702
            end=1;
1703

    
1704
            CHECK_BIDIRR( 0, 0, 0, 1)
1705
            if(s->avctx->bidir_refine > 1){
1706
                CHECK_BIDIRR( 0, 0, 1, 1)
1707
                CHECK_BIDIR2( 0, 1, 0, 1)
1708
                CHECK_BIDIR2( 1, 0, 1, 0)
1709
                CHECK_BIDIRR( 0, 0,-1, 1)
1710
                CHECK_BIDIR2( 0,-1, 0, 1)
1711
                CHECK_BIDIR2(-1, 0, 1, 0)
1712
                if(s->avctx->bidir_refine > 2){
1713
                    CHECK_BIDIRR( 0, 1, 1, 1)
1714
                    CHECK_BIDIRR( 0,-1, 1, 1)
1715
                    CHECK_BIDIRR( 0, 1,-1, 1)
1716
                    CHECK_BIDIRR( 0, 1, 1,-1)
1717
                    if(s->avctx->bidir_refine > 3){
1718
                        CHECK_BIDIR2( 1, 1, 1, 1)
1719
                        CHECK_BIDIRR( 1, 1, 1,-1)
1720
                        CHECK_BIDIR2( 1, 1,-1,-1)
1721
                        CHECK_BIDIR2( 1,-1,-1, 1)
1722
                        CHECK_BIDIR2( 1,-1, 1,-1)
1723
                    }
1724
                }
1725
            }
1726
        }while(!end);
1727
    }
1728

    
1729
    s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1730
    s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1731
    s->b_bidir_back_mv_table[xy][0]= motion_bx;
1732
    s->b_bidir_back_mv_table[xy][1]= motion_by;
1733

    
1734
    return fbmin;
1735
}
1736

    
1737
static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1738
{
1739
    MotionEstContext * const c= &s->me;
1740
    int P[10][2];
1741
    const int mot_stride = s->mb_stride;
1742
    const int mot_xy = mb_y*mot_stride + mb_x;
1743
    const int shift= 1+s->quarter_sample;
1744
    int dmin, i;
1745
    const int time_pp= s->pp_time;
1746
    const int time_pb= s->pb_time;
1747
    int mx, my, xmin, xmax, ymin, ymax;
1748
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
1749

    
1750
    c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1751
    ymin= xmin=(-32)>>shift;
1752
    ymax= xmax=   31>>shift;
1753

    
1754
    if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1755
        s->mv_type= MV_TYPE_8X8;
1756
    }else{
1757
        s->mv_type= MV_TYPE_16X16;
1758
    }
1759

    
1760
    for(i=0; i<4; i++){
1761
        int index= s->block_index[i];
1762
        int min, max;
1763

    
1764
        c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1765
        c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1766
        c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1767
        c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1768
//        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1769
//        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1770

    
1771
        max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1772
        min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1773
        max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1774
        min+= 16*mb_x - 1;
1775
        xmax= FFMIN(xmax, s->width - max);
1776
        xmin= FFMAX(xmin, - 16     - min);
1777

    
1778
        max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1779
        min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1780
        max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1781
        min+= 16*mb_y - 1;
1782
        ymax= FFMIN(ymax, s->height - max);
1783
        ymin= FFMAX(ymin, - 16      - min);
1784

    
1785
        if(s->mv_type == MV_TYPE_16X16) break;
1786
    }
1787

    
1788
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1789

    
1790
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1791
        s->b_direct_mv_table[mot_xy][0]= 0;
1792
        s->b_direct_mv_table[mot_xy][1]= 0;
1793

    
1794
        return 256*256*256*64;
1795
    }
1796

    
1797
    c->xmin= xmin;
1798
    c->ymin= ymin;
1799
    c->xmax= xmax;
1800
    c->ymax= ymax;
1801
    c->flags     |= FLAG_DIRECT;
1802
    c->sub_flags |= FLAG_DIRECT;
1803
    c->pred_x=0;
1804
    c->pred_y=0;
1805

    
1806
    P_LEFT[0]        = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1807
    P_LEFT[1]        = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1808

    
1809
    /* special case for first line */
1810
    if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its clipped
1811
        P_TOP[0]      = av_clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
1812
        P_TOP[1]      = av_clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
1813
        P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
1814
        P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
1815

    
1816
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1817
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1818
    }
1819

    
1820
    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1821
    if(c->sub_flags&FLAG_QPEL)
1822
        dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1823
    else
1824
        dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1825

    
1826
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1827
        dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1828

    
1829
    get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1830

    
1831
    mv_table[mot_xy][0]= mx;
1832
    mv_table[mot_xy][1]= my;
1833
    c->flags     &= ~FLAG_DIRECT;
1834
    c->sub_flags &= ~FLAG_DIRECT;
1835

    
1836
    return dmin;
1837
}
1838

    
1839
void ff_estimate_b_frame_motion(MpegEncContext * s,
1840
                             int mb_x, int mb_y)
1841
{
1842
    MotionEstContext * const c= &s->me;
1843
    const int penalty_factor= c->mb_penalty_factor;
1844
    int fmin, bmin, dmin, fbmin, bimin, fimin;
1845
    int type=0;
1846
    const int xy = mb_y*s->mb_stride + mb_x;
1847
    init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1848

    
1849
    get_limits(s, 16*mb_x, 16*mb_y);
1850

    
1851
    c->skip=0;
1852

    
1853
    if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
1854
        int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1855

    
1856
        score= ((unsigned)(score*score + 128*256))>>16;
1857
        c->mc_mb_var_sum_temp += score;
1858
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1859
        s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1860

    
1861
        return;
1862
    }
1863

    
1864
    if(c->avctx->me_threshold){
1865
        int vard= check_input_motion(s, mb_x, mb_y, 0);
1866

    
1867
        if((vard+128)>>8 < c->avctx->me_threshold){
1868
//            pix = c->src[0][0];
1869
//            sum = s->dsp.pix_sum(pix, s->linesize);
1870
//            varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1871

    
1872
//            pic->mb_var   [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1873
             s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1874
/*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1875
            c->mb_var_sum_temp    += (varc+128)>>8;*/
1876
            c->mc_mb_var_sum_temp += (vard+128)>>8;
1877
/*            if (vard <= 64<<8 || vard < varc) {
1878
                c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1879
            }else{
1880
                c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1881
            }*/
1882
            return;
1883
        }
1884
        if((vard+128)>>8 < c->avctx->mb_threshold){
1885
            type= s->mb_type[mb_y*s->mb_stride + mb_x];
1886
            if(type == CANDIDATE_MB_TYPE_DIRECT){
1887
                direct_search(s, mb_x, mb_y);
1888
            }
1889
            if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1890
                c->skip=0;
1891
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1892
            }
1893
            if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1894
                c->skip=0;
1895
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1896
            }
1897
            if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1898
                c->skip=0;
1899
                c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1900
                interlaced_search(s, 0,
1901
                                        s->b_field_mv_table[0], s->b_field_select_table[0],
1902
                                        s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1903
            }
1904
            if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1905
                c->skip=0;
1906
                c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1907
                interlaced_search(s, 2,
1908
                                        s->b_field_mv_table[1], s->b_field_select_table[1],
1909
                                        s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1910
            }
1911
            return;
1912
        }
1913
    }
1914

    
1915
    if (s->codec_id == CODEC_ID_MPEG4)
1916
        dmin= direct_search(s, mb_x, mb_y);
1917
    else
1918
        dmin= INT_MAX;
1919
//FIXME penalty stuff for non mpeg4
1920
    c->skip=0;
1921
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1922

    
1923
    c->skip=0;
1924
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1925
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1926

    
1927
    c->skip=0;
1928
    fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1929
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1930

    
1931
    if(s->flags & CODEC_FLAG_INTERLACED_ME){
1932
//FIXME mb type penalty
1933
        c->skip=0;
1934
        c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1935
        fimin= interlaced_search(s, 0,
1936
                                 s->b_field_mv_table[0], s->b_field_select_table[0],
1937
                                 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1938
        c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1939
        bimin= interlaced_search(s, 2,
1940
                                 s->b_field_mv_table[1], s->b_field_select_table[1],
1941
                                 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1942
    }else
1943
        fimin= bimin= INT_MAX;
1944

    
1945
    {
1946
        int score= fmin;
1947
        type = CANDIDATE_MB_TYPE_FORWARD;
1948

    
1949
        if (dmin <= score){
1950
            score = dmin;
1951
            type = CANDIDATE_MB_TYPE_DIRECT;
1952
        }
1953
        if(bmin<score){
1954
            score=bmin;
1955
            type= CANDIDATE_MB_TYPE_BACKWARD;
1956
        }
1957
        if(fbmin<score){
1958
            score=fbmin;
1959
            type= CANDIDATE_MB_TYPE_BIDIR;
1960
        }
1961
        if(fimin<score){
1962
            score=fimin;
1963
            type= CANDIDATE_MB_TYPE_FORWARD_I;
1964
        }
1965
        if(bimin<score){
1966
            score=bimin;
1967
            type= CANDIDATE_MB_TYPE_BACKWARD_I;
1968
        }
1969

    
1970
        score= ((unsigned)(score*score + 128*256))>>16;
1971
        c->mc_mb_var_sum_temp += score;
1972
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1973
    }
1974

    
1975
    if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1976
        type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1977
        if(fimin < INT_MAX)
1978
            type |= CANDIDATE_MB_TYPE_FORWARD_I;
1979
        if(bimin < INT_MAX)
1980
            type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1981
        if(fimin < INT_MAX && bimin < INT_MAX){
1982
            type |= CANDIDATE_MB_TYPE_BIDIR_I;
1983
        }
1984
         //FIXME something smarter
1985
        if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1986
        if(s->codec_id == CODEC_ID_MPEG4 && type&CANDIDATE_MB_TYPE_DIRECT && s->flags&CODEC_FLAG_MV0 && *(uint32_t*)s->b_direct_mv_table[xy])
1987
            type |= CANDIDATE_MB_TYPE_DIRECT0;
1988
#if 0
1989
        if(s->out_format == FMT_MPEG1)
1990
            type |= CANDIDATE_MB_TYPE_INTRA;
1991
#endif
1992
    }
1993

    
1994
    s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1995
}
1996

    
1997
/* find best f_code for ME which do unlimited searches */
1998
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1999
{
2000
    if(s->me_method>=ME_EPZS){
2001
        int score[8];
2002
        int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
2003
        uint8_t * fcode_tab= s->fcode_tab;
2004
        int best_fcode=-1;
2005
        int best_score=-10000000;
2006

    
2007
        if(s->msmpeg4_version)
2008
            range= FFMIN(range, 16);
2009
        else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
2010
            range= FFMIN(range, 256);
2011

    
2012
        for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
2013

    
2014
        for(y=0; y<s->mb_height; y++){
2015
            int x;
2016
            int xy= y*s->mb_stride;
2017
            for(x=0; x<s->mb_width; x++){
2018
                if(s->mb_type[xy] & type){
2019
                    int mx= mv_table[xy][0];
2020
                    int my= mv_table[xy][1];
2021
                    int fcode= FFMAX(fcode_tab[mx + MAX_MV],
2022
                                     fcode_tab[my + MAX_MV]);
2023
                    int j;
2024

    
2025
                        if(mx >= range || mx < -range ||
2026
                           my >= range || my < -range)
2027
                            continue;
2028

    
2029
                    for(j=0; j<fcode && j<8; j++){
2030
                        if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
2031
                            score[j]-= 170;
2032
                    }
2033
                }
2034
                xy++;
2035
            }
2036
        }
2037

    
2038
        for(i=1; i<8; i++){
2039
            if(score[i] > best_score){
2040
                best_score= score[i];
2041
                best_fcode= i;
2042
            }
2043
//            printf("%d %d\n", i, score[i]);
2044
        }
2045

    
2046
//    printf("fcode: %d type: %d\n", i, s->pict_type);
2047
        return best_fcode;
2048
/*        for(i=0; i<=MAX_FCODE; i++){
2049
            printf("%d ", mv_num[i]);
2050
        }
2051
        printf("\n");*/
2052
    }else{
2053
        return 1;
2054
    }
2055
}
2056

    
2057
void ff_fix_long_p_mvs(MpegEncContext * s)
2058
{
2059
    MotionEstContext * const c= &s->me;
2060
    const int f_code= s->f_code;
2061
    int y, range;
2062
    assert(s->pict_type==P_TYPE);
2063

    
2064
    range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2065

    
2066
    assert(range <= 16 || !s->msmpeg4_version);
2067
    assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
2068

    
2069
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2070

    
2071
//printf("%d no:%d %d//\n", clip, noclip, f_code);
2072
    if(s->flags&CODEC_FLAG_4MV){
2073
        const int wrap= s->b8_stride;
2074

    
2075
        /* clip / convert to intra 8x8 type MVs */
2076
        for(y=0; y<s->mb_height; y++){
2077
            int xy= y*2*wrap;
2078
            int i= y*s->mb_stride;
2079
            int x;
2080

    
2081
            for(x=0; x<s->mb_width; x++){
2082
                if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
2083
                    int block;
2084
                    for(block=0; block<4; block++){
2085
                        int off= (block& 1) + (block>>1)*wrap;
2086
                        int mx= s->current_picture.motion_val[0][ xy + off ][0];
2087
                        int my= s->current_picture.motion_val[0][ xy + off ][1];
2088

    
2089
                        if(   mx >=range || mx <-range
2090
                           || my >=range || my <-range){
2091
                            s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
2092
                            s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
2093
                            s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2094
                        }
2095
                    }
2096
                }
2097
                xy+=2;
2098
                i++;
2099
            }
2100
        }
2101
    }
2102
}
2103

    
2104
/**
2105
 *
2106
 * @param truncate 1 for truncation, 0 for using intra
2107
 */
2108
void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
2109
                     int16_t (*mv_table)[2], int f_code, int type, int truncate)
2110
{
2111
    MotionEstContext * const c= &s->me;
2112
    int y, h_range, v_range;
2113

    
2114
    // RAL: 8 in MPEG-1, 16 in MPEG-4
2115
    int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2116

    
2117
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2118

    
2119
    h_range= range;
2120
    v_range= field_select_table ? range>>1 : range;
2121

    
2122
    /* clip / convert to intra 16x16 type MVs */
2123
    for(y=0; y<s->mb_height; y++){
2124
        int x;
2125
        int xy= y*s->mb_stride;
2126
        for(x=0; x<s->mb_width; x++){
2127
            if (s->mb_type[xy] & type){    // RAL: "type" test added...
2128
                if(field_select_table==NULL || field_select_table[xy] == field_select){
2129
                    if(   mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2130
                       || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2131

    
2132
                        if(truncate){
2133
                            if     (mv_table[xy][0] > h_range-1) mv_table[xy][0]=  h_range-1;
2134
                            else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2135
                            if     (mv_table[xy][1] > v_range-1) mv_table[xy][1]=  v_range-1;
2136
                            else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2137
                        }else{
2138
                            s->mb_type[xy] &= ~type;
2139
                            s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
2140
                            mv_table[xy][0]=
2141
                            mv_table[xy][1]= 0;
2142
                        }
2143
                    }
2144
                }
2145
            }
2146
            xy++;
2147
        }
2148
    }
2149
}