Statistics
| Branch: | Revision:

ffmpeg / libavcodec / motion_est.c @ c192426f

History | View | Annotate | Download (76.6 KB)

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

    
25
/**
26
 * @file motion_est.c
27
 * Motion estimation.
28
 */
29

    
30
#include <stdlib.h>
31
#include <stdio.h>
32
#include <limits.h>
33
#include "avcodec.h"
34
#include "dsputil.h"
35
#include "mpegvideo.h"
36

    
37
#undef NDEBUG
38
#include <assert.h>
39

    
40
#define SQ(a) ((a)*(a))
41

    
42
#define P_LEFT P[1]
43
#define P_TOP P[2]
44
#define P_TOPRIGHT P[3]
45
#define P_MEDIAN P[4]
46
#define P_MV1 P[9]
47

    
48
static inline int sad_hpel_motion_search(MpegEncContext * s,
49
                                  int *mx_ptr, int *my_ptr, int dmin,
50
                                  int src_index, int ref_index,
51
                                  int size, int h);
52

    
53
static inline int update_map_generation(MotionEstContext *c)
54
{
55
    c->map_generation+= 1<<(ME_MAP_MV_BITS*2);
56
    if(c->map_generation==0){
57
        c->map_generation= 1<<(ME_MAP_MV_BITS*2);
58
        memset(c->map, 0, sizeof(uint32_t)*ME_MAP_SIZE);
59
    }
60
    return c->map_generation;
61
}
62

    
63
/* shape adaptive search stuff */
64
typedef struct Minima{
65
    int height;
66
    int x, y;
67
    int checked;
68
}Minima;
69

    
70
static int minima_cmp(const void *a, const void *b){
71
    const Minima *da = (const Minima *) a;
72
    const Minima *db = (const Minima *) b;
73

    
74
    return da->height - db->height;
75
}
76

    
77
#define FLAG_QPEL   1 //must be 1
78
#define FLAG_CHROMA 2
79
#define FLAG_DIRECT 4
80

    
81
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){
82
    const int offset[3]= {
83
          y*c->  stride + x,
84
        ((y*c->uvstride + x)>>1),
85
        ((y*c->uvstride + x)>>1),
86
    };
87
    int i;
88
    for(i=0; i<3; i++){
89
        c->src[0][i]= src [i] + offset[i];
90
        c->ref[0][i]= ref [i] + offset[i];
91
    }
92
    if(ref_index){
93
        for(i=0; i<3; i++){
94
            c->ref[ref_index][i]= ref2[i] + offset[i];
95
        }
96
    }
97
}
98

    
99
static int get_flags(MotionEstContext *c, int direct, int chroma){
100
    return   ((c->avctx->flags&CODEC_FLAG_QPEL) ? FLAG_QPEL : 0)
101
           + (direct ? FLAG_DIRECT : 0)
102
           + (chroma ? FLAG_CHROMA : 0);
103
}
104

    
105
/*! \brief compares a block (either a full macroblock or a partition thereof)
106
    against a proposed motion-compensated prediction of that block
107
 */
108
static av_always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
109
                      const int size, const int h, int ref_index, int src_index,
110
                      me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
111
    MotionEstContext * const c= &s->me;
112
    const int stride= c->stride;
113
    const int uvstride= c->uvstride;
114
    const int qpel= flags&FLAG_QPEL;
115
    const int chroma= flags&FLAG_CHROMA;
116
    const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
117
    const int hx= subx + (x<<(1+qpel));
118
    const int hy= suby + (y<<(1+qpel));
119
    uint8_t * const * const ref= c->ref[ref_index];
120
    uint8_t * const * const src= c->src[src_index];
121
    int d;
122
    //FIXME check chroma 4mv, (no crashes ...)
123
    if(flags&FLAG_DIRECT){
124
        assert(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1));
125
        if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
126
            const int time_pp= s->pp_time;
127
            const int time_pb= s->pb_time;
128
            const int mask= 2*qpel+1;
129
            if(s->mv_type==MV_TYPE_8X8){
130
                int i;
131
                for(i=0; i<4; i++){
132
                    int fx = c->direct_basis_mv[i][0] + hx;
133
                    int fy = c->direct_basis_mv[i][1] + hy;
134
                    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));
135
                    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));
136
                    int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
137
                    int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
138

    
139
                    uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
140
                    if(qpel){
141
                        c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
142
                        c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
143
                    }else{
144
                        c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
145
                        c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
146
                    }
147
                }
148
            }else{
149
                int fx = c->direct_basis_mv[0][0] + hx;
150
                int fy = c->direct_basis_mv[0][1] + hy;
151
                int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
152
                int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
153
                int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
154
                int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
155

    
156
                if(qpel){
157
                    c->qpel_put[1][fxy](c->temp               , ref[0] + (fx>>2) + (fy>>2)*stride               , stride);
158
                    c->qpel_put[1][fxy](c->temp + 8           , ref[0] + (fx>>2) + (fy>>2)*stride + 8           , stride);
159
                    c->qpel_put[1][fxy](c->temp     + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride     + 8*stride, stride);
160
                    c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
161
                    c->qpel_avg[1][bxy](c->temp               , ref[8] + (bx>>2) + (by>>2)*stride               , stride);
162
                    c->qpel_avg[1][bxy](c->temp + 8           , ref[8] + (bx>>2) + (by>>2)*stride + 8           , stride);
163
                    c->qpel_avg[1][bxy](c->temp     + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride     + 8*stride, stride);
164
                    c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
165
                }else{
166
                    assert((fx>>1) + 16*s->mb_x >= -16);
167
                    assert((fy>>1) + 16*s->mb_y >= -16);
168
                    assert((fx>>1) + 16*s->mb_x <= s->width);
169
                    assert((fy>>1) + 16*s->mb_y <= s->height);
170
                    assert((bx>>1) + 16*s->mb_x >= -16);
171
                    assert((by>>1) + 16*s->mb_y >= -16);
172
                    assert((bx>>1) + 16*s->mb_x <= s->width);
173
                    assert((by>>1) + 16*s->mb_y <= s->height);
174

    
175
                    c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
176
                    c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
177
                }
178
            }
179
            d = cmp_func(s, c->temp, src[0], stride, 16);
180
        }else
181
            d= 256*256*256*32;
182
    }else{
183
        int uvdxy;              /* no, it might not be used uninitialized */
184
        if(dxy){
185
            if(qpel){
186
                c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
187
                if(chroma){
188
                    int cx= hx/2;
189
                    int cy= hy/2;
190
                    cx= (cx>>1)|(cx&1);
191
                    cy= (cy>>1)|(cy&1);
192
                    uvdxy= (cx&1) + 2*(cy&1);
193
                    //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
194
                }
195
            }else{
196
                c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
197
                if(chroma)
198
                    uvdxy= dxy | (x&1) | (2*(y&1));
199
            }
200
            d = cmp_func(s, c->temp, src[0], stride, h);
201
        }else{
202
            d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
203
            if(chroma)
204
                uvdxy= (x&1) + 2*(y&1);
205
        }
206
        if(chroma){
207
            uint8_t * const uvtemp= c->temp + 16*stride;
208
            c->hpel_put[size+1][uvdxy](uvtemp  , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
209
            c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
210
            d += chroma_cmp_func(s, uvtemp  , src[1], uvstride, h>>1);
211
            d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
212
        }
213
    }
214
#if 0
215
    if(full_pel){
216
        const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
217
        score_map[index]= d;
218
    }
219

220
    d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
221
#endif
222
    return d;
223
}
224

    
225
#include "motion_est_template.c"
226

    
227
static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
228
    return 0;
229
}
230

    
231
static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
232
}
233

    
234
int ff_init_me(MpegEncContext *s){
235
    MotionEstContext * const c= &s->me;
236
    int cache_size= FFMIN(ME_MAP_SIZE>>ME_MAP_SHIFT, 1<<ME_MAP_SHIFT);
237
    int dia_size= FFMAX(FFABS(s->avctx->dia_size)&255, FFABS(s->avctx->pre_dia_size)&255);
238

    
239
    if(FFMIN(s->avctx->dia_size, s->avctx->pre_dia_size) < -ME_MAP_SIZE){
240
        av_log(s->avctx, AV_LOG_ERROR, "ME_MAP size is too small for SAB diamond\n");
241
        return -1;
242
    }
243

    
244
    c->avctx= s->avctx;
245

    
246
    if(cache_size < 2*dia_size && !c->stride){
247
        av_log(s->avctx, AV_LOG_INFO, "ME_MAP size may be a little small for the selected diamond size\n");
248
    }
249

    
250
    ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
251
    ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
252
    ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
253
    ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
254

    
255
    c->flags    = get_flags(c, 0, c->avctx->me_cmp    &FF_CMP_CHROMA);
256
    c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
257
    c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp    &FF_CMP_CHROMA);
258

    
259
/*FIXME s->no_rounding b_type*/
260
    if(s->flags&CODEC_FLAG_QPEL){
261
        c->sub_motion_search= qpel_motion_search;
262
        c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
263
        if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
264
        else               c->qpel_put= s->dsp.put_qpel_pixels_tab;
265
    }else{
266
        if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
267
            c->sub_motion_search= hpel_motion_search;
268
        else if(   c->avctx->me_sub_cmp == FF_CMP_SAD
269
                && c->avctx->    me_cmp == FF_CMP_SAD
270
                && c->avctx->    mb_cmp == FF_CMP_SAD)
271
            c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
272
        else
273
            c->sub_motion_search= hpel_motion_search;
274
    }
275
    c->hpel_avg= s->dsp.avg_pixels_tab;
276
    if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
277
    else               c->hpel_put= s->dsp.put_pixels_tab;
278

    
279
    if(s->linesize){
280
        c->stride  = s->linesize;
281
        c->uvstride= s->uvlinesize;
282
    }else{
283
        c->stride  = 16*s->mb_width + 32;
284
        c->uvstride=  8*s->mb_width + 16;
285
    }
286

    
287
    /* 8x8 fullpel search would need a 4x4 chroma compare, which we do
288
     * not have yet, and even if we had, the motion estimation code
289
     * does not expect it. */
290
    if(s->codec_id != CODEC_ID_SNOW){
291
        if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
292
            s->dsp.me_cmp[2]= zero_cmp;
293
        }
294
        if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
295
            s->dsp.me_sub_cmp[2]= zero_cmp;
296
        }
297
        c->hpel_put[2][0]= c->hpel_put[2][1]=
298
        c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
299
    }
300

    
301
    if(s->codec_id == CODEC_ID_H261){
302
        c->sub_motion_search= no_sub_motion_search;
303
    }
304

    
305
    c->temp= c->scratchpad;
306

    
307
    return 0;
308
}
309

    
310
#if 0
311
static int pix_dev(uint8_t * pix, int line_size, int mean)
312
{
313
    int s, i, j;
314

315
    s = 0;
316
    for (i = 0; i < 16; i++) {
317
        for (j = 0; j < 16; j += 8) {
318
            s += FFABS(pix[0]-mean);
319
            s += FFABS(pix[1]-mean);
320
            s += FFABS(pix[2]-mean);
321
            s += FFABS(pix[3]-mean);
322
            s += FFABS(pix[4]-mean);
323
            s += FFABS(pix[5]-mean);
324
            s += FFABS(pix[6]-mean);
325
            s += FFABS(pix[7]-mean);
326
            pix += 8;
327
        }
328
        pix += line_size - 16;
329
    }
330
    return s;
331
}
332
#endif
333

    
334
static inline void no_motion_search(MpegEncContext * s,
335
                                    int *mx_ptr, int *my_ptr)
336
{
337
    *mx_ptr = 16 * s->mb_x;
338
    *my_ptr = 16 * s->mb_y;
339
}
340

    
341
#if 0  /* the use of these functions is inside #if 0 */
342
static int full_motion_search(MpegEncContext * s,
343
                              int *mx_ptr, int *my_ptr, int range,
344
                              int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
345
{
346
    int x1, y1, x2, y2, xx, yy, x, y;
347
    int mx, my, dmin, d;
348
    uint8_t *pix;
349

350
    xx = 16 * s->mb_x;
351
    yy = 16 * s->mb_y;
352
    x1 = xx - range + 1;        /* we loose one pixel to avoid boundary pb with half pixel pred */
353
    if (x1 < xmin)
354
        x1 = xmin;
355
    x2 = xx + range - 1;
356
    if (x2 > xmax)
357
        x2 = xmax;
358
    y1 = yy - range + 1;
359
    if (y1 < ymin)
360
        y1 = ymin;
361
    y2 = yy + range - 1;
362
    if (y2 > ymax)
363
        y2 = ymax;
364
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
365
    dmin = 0x7fffffff;
366
    mx = 0;
367
    my = 0;
368
    for (y = y1; y <= y2; y++) {
369
        for (x = x1; x <= x2; x++) {
370
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
371
                             s->linesize, 16);
372
            if (d < dmin ||
373
                (d == dmin &&
374
                 (abs(x - xx) + abs(y - yy)) <
375
                 (abs(mx - xx) + abs(my - yy)))) {
376
                dmin = d;
377
                mx = x;
378
                my = y;
379
            }
380
        }
381
    }
382

383
    *mx_ptr = mx;
384
    *my_ptr = my;
385

386
#if 0
387
    if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
388
        *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
389
        av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
390
    }
391
#endif
392
    return dmin;
393
}
394

    
395

    
396
static int log_motion_search(MpegEncContext * s,
397
                             int *mx_ptr, int *my_ptr, int range,
398
                             int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
399
{
400
    int x1, y1, x2, y2, xx, yy, x, y;
401
    int mx, my, dmin, d;
402
    uint8_t *pix;
403

    
404
    xx = s->mb_x << 4;
405
    yy = s->mb_y << 4;
406

    
407
    /* Left limit */
408
    x1 = xx - range;
409
    if (x1 < xmin)
410
        x1 = xmin;
411

    
412
    /* Right limit */
413
    x2 = xx + range;
414
    if (x2 > xmax)
415
        x2 = xmax;
416

    
417
    /* Upper limit */
418
    y1 = yy - range;
419
    if (y1 < ymin)
420
        y1 = ymin;
421

    
422
    /* Lower limit */
423
    y2 = yy + range;
424
    if (y2 > ymax)
425
        y2 = ymax;
426

    
427
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
428
    dmin = 0x7fffffff;
429
    mx = 0;
430
    my = 0;
431

    
432
    do {
433
        for (y = y1; y <= y2; y += range) {
434
            for (x = x1; x <= x2; x += range) {
435
                d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
436
                if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
437
                    dmin = d;
438
                    mx = x;
439
                    my = y;
440
                }
441
            }
442
        }
443

    
444
        range = range >> 1;
445

    
446
        x1 = mx - range;
447
        if (x1 < xmin)
448
            x1 = xmin;
449

    
450
        x2 = mx + range;
451
        if (x2 > xmax)
452
            x2 = xmax;
453

    
454
        y1 = my - range;
455
        if (y1 < ymin)
456
            y1 = ymin;
457

    
458
        y2 = my + range;
459
        if (y2 > ymax)
460
            y2 = ymax;
461

    
462
    } while (range >= 1);
463

    
464
#ifdef DEBUG
465
    av_log(s->avctx, AV_LOG_DEBUG, "log       - MX: %d\tMY: %d\n", mx, my);
466
#endif
467
    *mx_ptr = mx;
468
    *my_ptr = my;
469
    return dmin;
470
}
471

    
472
static int phods_motion_search(MpegEncContext * s,
473
                               int *mx_ptr, int *my_ptr, int range,
474
                               int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
475
{
476
    int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
477
    int mx, my, dminx, dminy;
478
    uint8_t *pix;
479

    
480
    xx = s->mb_x << 4;
481
    yy = s->mb_y << 4;
482

    
483
    /* Left limit */
484
    x1 = xx - range;
485
    if (x1 < xmin)
486
        x1 = xmin;
487

    
488
    /* Right limit */
489
    x2 = xx + range;
490
    if (x2 > xmax)
491
        x2 = xmax;
492

    
493
    /* Upper limit */
494
    y1 = yy - range;
495
    if (y1 < ymin)
496
        y1 = ymin;
497

    
498
    /* Lower limit */
499
    y2 = yy + range;
500
    if (y2 > ymax)
501
        y2 = ymax;
502

    
503
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
504
    mx = 0;
505
    my = 0;
506

    
507
    x = xx;
508
    y = yy;
509
    do {
510
        dminx = 0x7fffffff;
511
        dminy = 0x7fffffff;
512

    
513
        lastx = x;
514
        for (x = x1; x <= x2; x += range) {
515
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
516
            if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
517
                dminx = d;
518
                mx = x;
519
            }
520
        }
521

    
522
        x = lastx;
523
        for (y = y1; y <= y2; y += range) {
524
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
525
            if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
526
                dminy = d;
527
                my = y;
528
            }
529
        }
530

    
531
        range = range >> 1;
532

    
533
        x = mx;
534
        y = my;
535
        x1 = mx - range;
536
        if (x1 < xmin)
537
            x1 = xmin;
538

    
539
        x2 = mx + range;
540
        if (x2 > xmax)
541
            x2 = xmax;
542

    
543
        y1 = my - range;
544
        if (y1 < ymin)
545
            y1 = ymin;
546

    
547
        y2 = my + range;
548
        if (y2 > ymax)
549
            y2 = ymax;
550

    
551
    } while (range >= 1);
552

    
553
#ifdef DEBUG
554
    av_log(s->avctx, AV_LOG_DEBUG, "phods     - MX: %d\tMY: %d\n", mx, my);
555
#endif
556

    
557
    /* half pixel search */
558
    *mx_ptr = mx;
559
    *my_ptr = my;
560
    return dminy;
561
}
562
#endif /* 0 */
563

    
564
#define Z_THRESHOLD 256
565

    
566
#define CHECK_SAD_HALF_MV(suffix, x, y) \
567
{\
568
    d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
569
    d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
570
    COPY3_IF_LT(dminh, d, dx, x, dy, y)\
571
}
572

    
573
static inline int sad_hpel_motion_search(MpegEncContext * s,
574
                                  int *mx_ptr, int *my_ptr, int dmin,
575
                                  int src_index, int ref_index,
576
                                  int size, int h)
577
{
578
    MotionEstContext * const c= &s->me;
579
    const int penalty_factor= c->sub_penalty_factor;
580
    int mx, my, dminh;
581
    uint8_t *pix, *ptr;
582
    int stride= c->stride;
583
    const int flags= c->sub_flags;
584
    LOAD_COMMON
585

    
586
    assert(flags == 0);
587

    
588
    if(c->skip){
589
//    printf("S");
590
        *mx_ptr = 0;
591
        *my_ptr = 0;
592
        return dmin;
593
    }
594
//    printf("N");
595

    
596
    pix = c->src[src_index][0];
597

    
598
    mx = *mx_ptr;
599
    my = *my_ptr;
600
    ptr = c->ref[ref_index][0] + (my * stride) + mx;
601

    
602
    dminh = dmin;
603

    
604
    if (mx > xmin && mx < xmax &&
605
        my > ymin && my < ymax) {
606
        int dx=0, dy=0;
607
        int d, pen_x, pen_y;
608
        const int index= (my<<ME_MAP_SHIFT) + mx;
609
        const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
610
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
611
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
612
        const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
613
        mx<<=1;
614
        my<<=1;
615

    
616

    
617
        pen_x= pred_x + mx;
618
        pen_y= pred_y + my;
619

    
620
        ptr-= stride;
621
        if(t<=b){
622
            CHECK_SAD_HALF_MV(y2 , 0, -1)
623
            if(l<=r){
624
                CHECK_SAD_HALF_MV(xy2, -1, -1)
625
                if(t+r<=b+l){
626
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
627
                    ptr+= stride;
628
                }else{
629
                    ptr+= stride;
630
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
631
                }
632
                CHECK_SAD_HALF_MV(x2 , -1,  0)
633
            }else{
634
                CHECK_SAD_HALF_MV(xy2, +1, -1)
635
                if(t+l<=b+r){
636
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
637
                    ptr+= stride;
638
                }else{
639
                    ptr+= stride;
640
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
641
                }
642
                CHECK_SAD_HALF_MV(x2 , +1,  0)
643
            }
644
        }else{
645
            if(l<=r){
646
                if(t+l<=b+r){
647
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
648
                    ptr+= stride;
649
                }else{
650
                    ptr+= stride;
651
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
652
                }
653
                CHECK_SAD_HALF_MV(x2 , -1,  0)
654
                CHECK_SAD_HALF_MV(xy2, -1, +1)
655
            }else{
656
                if(t+r<=b+l){
657
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
658
                    ptr+= stride;
659
                }else{
660
                    ptr+= stride;
661
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
662
                }
663
                CHECK_SAD_HALF_MV(x2 , +1,  0)
664
                CHECK_SAD_HALF_MV(xy2, +1, +1)
665
            }
666
            CHECK_SAD_HALF_MV(y2 ,  0, +1)
667
        }
668
        mx+=dx;
669
        my+=dy;
670

    
671
    }else{
672
        mx<<=1;
673
        my<<=1;
674
    }
675

    
676
    *mx_ptr = mx;
677
    *my_ptr = my;
678
    return dminh;
679
}
680

    
681
static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
682
{
683
    const int xy= s->mb_x + s->mb_y*s->mb_stride;
684

    
685
    s->p_mv_table[xy][0] = mx;
686
    s->p_mv_table[xy][1] = my;
687

    
688
    /* has already been set to the 4 MV if 4MV is done */
689
    if(mv4){
690
        int mot_xy= s->block_index[0];
691

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

    
697
        mot_xy += s->b8_stride;
698
        s->current_picture.motion_val[0][mot_xy  ][0]= mx;
699
        s->current_picture.motion_val[0][mot_xy  ][1]= my;
700
        s->current_picture.motion_val[0][mot_xy+1][0]= mx;
701
        s->current_picture.motion_val[0][mot_xy+1][1]= my;
702
    }
703
}
704

    
705
/**
706
 * get fullpel ME search limits.
707
 */
708
static inline void get_limits(MpegEncContext *s, int x, int y)
709
{
710
    MotionEstContext * const c= &s->me;
711
    int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
712
/*
713
    if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
714
    else                   c->range= 16;
715
*/
716
    if (s->unrestricted_mv) {
717
        c->xmin = - x - 16;
718
        c->ymin = - y - 16;
719
        c->xmax = - x + s->mb_width *16;
720
        c->ymax = - y + s->mb_height*16;
721
    } else if (s->out_format == FMT_H261){
722
        // Search range of H261 is different from other codec standards
723
        c->xmin = (x > 15) ? - 15 : 0;
724
        c->ymin = (y > 15) ? - 15 : 0;
725
        c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
726
        c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
727
    } else {
728
        c->xmin = - x;
729
        c->ymin = - y;
730
        c->xmax = - x + s->mb_width *16 - 16;
731
        c->ymax = - y + s->mb_height*16 - 16;
732
    }
733
    if(range){
734
        c->xmin = FFMAX(c->xmin,-range);
735
        c->xmax = FFMIN(c->xmax, range);
736
        c->ymin = FFMAX(c->ymin,-range);
737
        c->ymax = FFMIN(c->ymax, range);
738
    }
739
}
740

    
741
static inline void init_mv4_ref(MotionEstContext *c){
742
    const int stride= c->stride;
743

    
744
    c->ref[1][0] = c->ref[0][0] + 8;
745
    c->ref[2][0] = c->ref[0][0] + 8*stride;
746
    c->ref[3][0] = c->ref[2][0] + 8;
747
    c->src[1][0] = c->src[0][0] + 8;
748
    c->src[2][0] = c->src[0][0] + 8*stride;
749
    c->src[3][0] = c->src[2][0] + 8;
750
}
751

    
752
static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
753
{
754
    MotionEstContext * const c= &s->me;
755
    const int size= 1;
756
    const int h=8;
757
    int block;
758
    int P[10][2];
759
    int dmin_sum=0, mx4_sum=0, my4_sum=0;
760
    int same=1;
761
    const int stride= c->stride;
762
    uint8_t *mv_penalty= c->current_mv_penalty;
763

    
764
    init_mv4_ref(c);
765

    
766
    for(block=0; block<4; block++){
767
        int mx4, my4;
768
        int pred_x4, pred_y4;
769
        int dmin4;
770
        static const int off[4]= {2, 1, 1, -1};
771
        const int mot_stride = s->b8_stride;
772
        const int mot_xy = s->block_index[block];
773

    
774
        P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
775
        P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
776

    
777
        if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
778

    
779
        /* special case for first line */
780
        if (s->first_slice_line && block<2) {
781
            c->pred_x= pred_x4= P_LEFT[0];
782
            c->pred_y= pred_y4= P_LEFT[1];
783
        } else {
784
            P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][0];
785
            P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][1];
786
            P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
787
            P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
788
            if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
789
            if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
790
            if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
791
            if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
792

    
793
            P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
794
            P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
795

    
796
            c->pred_x= pred_x4 = P_MEDIAN[0];
797
            c->pred_y= pred_y4 = P_MEDIAN[1];
798
        }
799
        P_MV1[0]= mx;
800
        P_MV1[1]= my;
801

    
802
        dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
803

    
804
        dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
805

    
806
        if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
807
            int dxy;
808
            const int offset= ((block&1) + (block>>1)*stride)*8;
809
            uint8_t *dest_y = c->scratchpad + offset;
810
            if(s->quarter_sample){
811
                uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
812
                dxy = ((my4 & 3) << 2) | (mx4 & 3);
813

    
814
                if(s->no_rounding)
815
                    s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , stride);
816
                else
817
                    s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , stride);
818
            }else{
819
                uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
820
                dxy = ((my4 & 1) << 1) | (mx4 & 1);
821

    
822
                if(s->no_rounding)
823
                    s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , stride, h);
824
                else
825
                    s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , stride, h);
826
            }
827
            dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
828
        }else
829
            dmin_sum+= dmin4;
830

    
831
        if(s->quarter_sample){
832
            mx4_sum+= mx4/2;
833
            my4_sum+= my4/2;
834
        }else{
835
            mx4_sum+= mx4;
836
            my4_sum+= my4;
837
        }
838

    
839
        s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
840
        s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
841

    
842
        if(mx4 != mx || my4 != my) same=0;
843
    }
844

    
845
    if(same)
846
        return INT_MAX;
847

    
848
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
849
        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);
850
    }
851

    
852
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
853
        int dxy;
854
        int mx, my;
855
        int offset;
856

    
857
        mx= ff_h263_round_chroma(mx4_sum);
858
        my= ff_h263_round_chroma(my4_sum);
859
        dxy = ((my & 1) << 1) | (mx & 1);
860

    
861
        offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
862

    
863
        if(s->no_rounding){
864
            s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
865
            s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
866
        }else{
867
            s->dsp.put_pixels_tab       [1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
868
            s->dsp.put_pixels_tab       [1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
869
        }
870

    
871
        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);
872
        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);
873
    }
874

    
875
    c->pred_x= mx;
876
    c->pred_y= my;
877

    
878
    switch(c->avctx->mb_cmp&0xFF){
879
    /*case FF_CMP_SSE:
880
        return dmin_sum+ 32*s->qscale*s->qscale;*/
881
    case FF_CMP_RD:
882
        return dmin_sum;
883
    default:
884
        return dmin_sum+ 11*c->mb_penalty_factor;
885
    }
886
}
887

    
888
static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
889
    MotionEstContext * const c= &s->me;
890

    
891
    c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
892
    c->src[1][0] = c->src[0][0] + s->linesize;
893
    if(c->flags & FLAG_CHROMA){
894
        c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
895
        c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
896
        c->src[1][1] = c->src[0][1] + s->uvlinesize;
897
        c->src[1][2] = c->src[0][2] + s->uvlinesize;
898
    }
899
}
900

    
901
static int interlaced_search(MpegEncContext *s, int ref_index,
902
                             int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
903
{
904
    MotionEstContext * const c= &s->me;
905
    const int size=0;
906
    const int h=8;
907
    int block;
908
    int P[10][2];
909
    uint8_t * const mv_penalty= c->current_mv_penalty;
910
    int same=1;
911
    const int stride= 2*s->linesize;
912
    int dmin_sum= 0;
913
    const int mot_stride= s->mb_stride;
914
    const int xy= s->mb_x + s->mb_y*mot_stride;
915

    
916
    c->ymin>>=1;
917
    c->ymax>>=1;
918
    c->stride<<=1;
919
    c->uvstride<<=1;
920
    init_interlaced_ref(s, ref_index);
921

    
922
    for(block=0; block<2; block++){
923
        int field_select;
924
        int best_dmin= INT_MAX;
925
        int best_field= -1;
926

    
927
        for(field_select=0; field_select<2; field_select++){
928
            int dmin, mx_i, my_i;
929
            int16_t (*mv_table)[2]= mv_tables[block][field_select];
930

    
931
            if(user_field_select){
932
                assert(field_select==0 || field_select==1);
933
                assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
934
                if(field_select_tables[block][xy] != field_select)
935
                    continue;
936
            }
937

    
938
            P_LEFT[0] = mv_table[xy - 1][0];
939
            P_LEFT[1] = mv_table[xy - 1][1];
940
            if(P_LEFT[0]       > (c->xmax<<1)) P_LEFT[0]       = (c->xmax<<1);
941

    
942
            c->pred_x= P_LEFT[0];
943
            c->pred_y= P_LEFT[1];
944

    
945
            if(!s->first_slice_line){
946
                P_TOP[0]      = mv_table[xy - mot_stride][0];
947
                P_TOP[1]      = mv_table[xy - mot_stride][1];
948
                P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
949
                P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
950
                if(P_TOP[1]      > (c->ymax<<1)) P_TOP[1]     = (c->ymax<<1);
951
                if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
952
                if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
953
                if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
954

    
955
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
956
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
957
            }
958
            P_MV1[0]= mx; //FIXME not correct if block != field_select
959
            P_MV1[1]= my / 2;
960

    
961
            dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
962

    
963
            dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
964

    
965
            mv_table[xy][0]= mx_i;
966
            mv_table[xy][1]= my_i;
967

    
968
            if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
969
                int dxy;
970

    
971
                //FIXME chroma ME
972
                uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
973
                dxy = ((my_i & 1) << 1) | (mx_i & 1);
974

    
975
                if(s->no_rounding){
976
                    s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref    , stride, h);
977
                }else{
978
                    s->dsp.put_pixels_tab       [size][dxy](c->scratchpad, ref    , stride, h);
979
                }
980
                dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
981
                dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
982
            }else
983
                dmin+= c->mb_penalty_factor; //field_select bits
984

    
985
            dmin += field_select != block; //slightly prefer same field
986

    
987
            if(dmin < best_dmin){
988
                best_dmin= dmin;
989
                best_field= field_select;
990
            }
991
        }
992
        {
993
            int16_t (*mv_table)[2]= mv_tables[block][best_field];
994

    
995
            if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
996
            if(mv_table[xy][1]&1) same=0;
997
            if(mv_table[xy][1]*2 != my) same=0;
998
            if(best_field != block) same=0;
999
        }
1000

    
1001
        field_select_tables[block][xy]= best_field;
1002
        dmin_sum += best_dmin;
1003
    }
1004

    
1005
    c->ymin<<=1;
1006
    c->ymax<<=1;
1007
    c->stride>>=1;
1008
    c->uvstride>>=1;
1009

    
1010
    if(same)
1011
        return INT_MAX;
1012

    
1013
    switch(c->avctx->mb_cmp&0xFF){
1014
    /*case FF_CMP_SSE:
1015
        return dmin_sum+ 32*s->qscale*s->qscale;*/
1016
    case FF_CMP_RD:
1017
        return dmin_sum;
1018
    default:
1019
        return dmin_sum+ 11*c->mb_penalty_factor;
1020
    }
1021
}
1022

    
1023
static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
1024
    int ymax= s->me.ymax>>interlaced;
1025
    int ymin= s->me.ymin>>interlaced;
1026

    
1027
    if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
1028
    if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
1029
    if(mv[1] <       ymin) mv[1] =       ymin;
1030
    if(mv[1] >       ymax) mv[1] =       ymax;
1031
}
1032

    
1033
static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1034
    MotionEstContext * const c= &s->me;
1035
    Picture *p= s->current_picture_ptr;
1036
    int mb_xy= mb_x + mb_y*s->mb_stride;
1037
    int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1038
    int mb_type= s->current_picture.mb_type[mb_xy];
1039
    int flags= c->flags;
1040
    int shift= (flags&FLAG_QPEL) + 1;
1041
    int mask= (1<<shift)-1;
1042
    int x, y, i;
1043
    int d=0;
1044
    me_cmp_func cmpf= s->dsp.sse[0];
1045
    me_cmp_func chroma_cmpf= s->dsp.sse[1];
1046

    
1047
    if(p_type && USES_LIST(mb_type, 1)){
1048
        av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1049
        return INT_MAX/2;
1050
    }
1051
    assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1052

    
1053
    for(i=0; i<4; i++){
1054
        int xy= s->block_index[i];
1055
        clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1056
        clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1057
    }
1058

    
1059
    if(IS_INTERLACED(mb_type)){
1060
        int xy2= xy  + s->b8_stride;
1061
        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1062
        c->stride<<=1;
1063
        c->uvstride<<=1;
1064

    
1065
        if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1066
            av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1067
            return INT_MAX/2;
1068
        }
1069

    
1070
        if(USES_LIST(mb_type, 0)){
1071
            int field_select0= p->ref_index[0][xy ];
1072
            int field_select1= p->ref_index[0][xy2];
1073
            assert(field_select0==0 ||field_select0==1);
1074
            assert(field_select1==0 ||field_select1==1);
1075
            init_interlaced_ref(s, 0);
1076

    
1077
            if(p_type){
1078
                s->p_field_select_table[0][mb_xy]= field_select0;
1079
                s->p_field_select_table[1][mb_xy]= field_select1;
1080
                *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1081
                *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1082
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1083
            }else{
1084
                s->b_field_select_table[0][0][mb_xy]= field_select0;
1085
                s->b_field_select_table[0][1][mb_xy]= field_select1;
1086
                *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1087
                *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1088
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1089
            }
1090

    
1091
            x= p->motion_val[0][xy ][0];
1092
            y= p->motion_val[0][xy ][1];
1093
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1094
            x= p->motion_val[0][xy2][0];
1095
            y= p->motion_val[0][xy2][1];
1096
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1097
        }
1098
        if(USES_LIST(mb_type, 1)){
1099
            int field_select0= p->ref_index[1][xy ];
1100
            int field_select1= p->ref_index[1][xy2];
1101
            assert(field_select0==0 ||field_select0==1);
1102
            assert(field_select1==0 ||field_select1==1);
1103
            init_interlaced_ref(s, 2);
1104

    
1105
            s->b_field_select_table[1][0][mb_xy]= field_select0;
1106
            s->b_field_select_table[1][1][mb_xy]= field_select1;
1107
            *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1108
            *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1109
            if(USES_LIST(mb_type, 0)){
1110
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1111
            }else{
1112
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1113
            }
1114

    
1115
            x= p->motion_val[1][xy ][0];
1116
            y= p->motion_val[1][xy ][1];
1117
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1118
            x= p->motion_val[1][xy2][0];
1119
            y= p->motion_val[1][xy2][1];
1120
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1121
            //FIXME bidir scores
1122
        }
1123
        c->stride>>=1;
1124
        c->uvstride>>=1;
1125
    }else if(IS_8X8(mb_type)){
1126
        if(!(s->flags & CODEC_FLAG_4MV)){
1127
            av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1128
            return INT_MAX/2;
1129
        }
1130
        cmpf= s->dsp.sse[1];
1131
        chroma_cmpf= s->dsp.sse[1];
1132
        init_mv4_ref(c);
1133
        for(i=0; i<4; i++){
1134
            xy= s->block_index[i];
1135
            x= p->motion_val[0][xy][0];
1136
            y= p->motion_val[0][xy][1];
1137
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1138
        }
1139
        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1140
    }else{
1141
        if(USES_LIST(mb_type, 0)){
1142
            if(p_type){
1143
                *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1144
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1145
            }else if(USES_LIST(mb_type, 1)){
1146
                *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1147
                *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1148
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1149
            }else{
1150
                *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1151
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1152
            }
1153
            x= p->motion_val[0][xy][0];
1154
            y= p->motion_val[0][xy][1];
1155
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1156
        }else if(USES_LIST(mb_type, 1)){
1157
            *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1158
            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1159

    
1160
            x= p->motion_val[1][xy][0];
1161
            y= p->motion_val[1][xy][1];
1162
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1163
        }else
1164
            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1165
    }
1166
    return d;
1167
}
1168

    
1169
void ff_estimate_p_frame_motion(MpegEncContext * s,
1170
                                int mb_x, int mb_y)
1171
{
1172
    MotionEstContext * const c= &s->me;
1173
    uint8_t *pix, *ppix;
1174
    int sum, mx, my, dmin;
1175
    int varc;            ///< the variance of the block (sum of squared (p[y][x]-average))
1176
    int vard;            ///< sum of squared differences with the estimated motion vector
1177
    int P[10][2];
1178
    const int shift= 1+s->quarter_sample;
1179
    int mb_type=0;
1180
    Picture * const pic= &s->current_picture;
1181

    
1182
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1183

    
1184
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1185
    assert(s->linesize == c->stride);
1186
    assert(s->uvlinesize == c->uvstride);
1187

    
1188
    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1189
    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1190
    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1191
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1192

    
1193
    get_limits(s, 16*mb_x, 16*mb_y);
1194
    c->skip=0;
1195

    
1196
    /* intra / predictive decision */
1197
    pix = c->src[0][0];
1198
    sum = s->dsp.pix_sum(pix, s->linesize);
1199
    varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1200

    
1201
    pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1202
    pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1203
    c->mb_var_sum_temp += (varc+128)>>8;
1204

    
1205
    if(c->avctx->me_threshold){
1206
        vard= check_input_motion(s, mb_x, mb_y, 1);
1207

    
1208
        if((vard+128)>>8 < c->avctx->me_threshold){
1209
            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1210
            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1211
            pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1212
            c->mc_mb_var_sum_temp += (vard+128)>>8;
1213
            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1214
            return;
1215
        }
1216
        if((vard+128)>>8 < c->avctx->mb_threshold)
1217
            mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1218
    }
1219

    
1220
    switch(s->me_method) {
1221
    case ME_ZERO:
1222
    default:
1223
        no_motion_search(s, &mx, &my);
1224
        mx-= mb_x*16;
1225
        my-= mb_y*16;
1226
        dmin = 0;
1227
        break;
1228
#if 0
1229
    case ME_FULL:
1230
        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1231
        mx-= mb_x*16;
1232
        my-= mb_y*16;
1233
        break;
1234
    case ME_LOG:
1235
        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1236
        mx-= mb_x*16;
1237
        my-= mb_y*16;
1238
        break;
1239
    case ME_PHODS:
1240
        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1241
        mx-= mb_x*16;
1242
        my-= mb_y*16;
1243
        break;
1244
#endif
1245
    case ME_X1:
1246
    case ME_EPZS:
1247
       {
1248
            const int mot_stride = s->b8_stride;
1249
            const int mot_xy = s->block_index[0];
1250

    
1251
            P_LEFT[0]       = s->current_picture.motion_val[0][mot_xy - 1][0];
1252
            P_LEFT[1]       = s->current_picture.motion_val[0][mot_xy - 1][1];
1253

    
1254
            if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
1255

    
1256
            if(!s->first_slice_line) {
1257
                P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][0];
1258
                P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][1];
1259
                P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1260
                P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1261
                if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
1262
                if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1263
                if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1264

    
1265
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1266
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1267

    
1268
                if(s->out_format == FMT_H263){
1269
                    c->pred_x = P_MEDIAN[0];
1270
                    c->pred_y = P_MEDIAN[1];
1271
                }else { /* mpeg1 at least */
1272
                    c->pred_x= P_LEFT[0];
1273
                    c->pred_y= P_LEFT[1];
1274
                }
1275
            }else{
1276
                c->pred_x= P_LEFT[0];
1277
                c->pred_y= P_LEFT[1];
1278
            }
1279

    
1280
        }
1281
        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1282

    
1283
        break;
1284
    }
1285

    
1286
    /* At this point (mx,my) are full-pell and the relative displacement */
1287
    ppix = c->ref[0][0] + (my * s->linesize) + mx;
1288

    
1289
    vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1290

    
1291
    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1292
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1293
    c->mc_mb_var_sum_temp += (vard+128)>>8;
1294

    
1295
#if 0
1296
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1297
           varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1298
#endif
1299
    if(mb_type){
1300
        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1301
        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1302
        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1303

    
1304
        if(mb_type == CANDIDATE_MB_TYPE_INTER){
1305
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1306
            set_p_mv_tables(s, mx, my, 1);
1307
        }else{
1308
            mx <<=shift;
1309
            my <<=shift;
1310
        }
1311
        if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1312
            h263_mv4_search(s, mx, my, shift);
1313

    
1314
            set_p_mv_tables(s, mx, my, 0);
1315
        }
1316
        if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1317
            interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1318
        }
1319
    }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1320
        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1321
        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1322
        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1323

    
1324
        if (vard*2 + 200*256 > varc)
1325
            mb_type|= CANDIDATE_MB_TYPE_INTRA;
1326
        if (varc*2 + 200*256 > vard || s->qscale > 24){
1327
//        if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1328
            mb_type|= CANDIDATE_MB_TYPE_INTER;
1329
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1330
            if(s->flags&CODEC_FLAG_MV0)
1331
                if(mx || my)
1332
                    mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1333
        }else{
1334
            mx <<=shift;
1335
            my <<=shift;
1336
        }
1337
        if((s->flags&CODEC_FLAG_4MV)
1338
           && !c->skip && varc>50<<8 && vard>10<<8){
1339
            if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1340
                mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1341

    
1342
            set_p_mv_tables(s, mx, my, 0);
1343
        }else
1344
            set_p_mv_tables(s, mx, my, 1);
1345
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
1346
           && !c->skip){ //FIXME varc/d checks
1347
            if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1348
                mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1349
        }
1350
    }else{
1351
        int intra_score, i;
1352
        mb_type= CANDIDATE_MB_TYPE_INTER;
1353

    
1354
        dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1355
        if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1356
            dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1357

    
1358
        if((s->flags&CODEC_FLAG_4MV)
1359
           && !c->skip && varc>50<<8 && vard>10<<8){
1360
            int dmin4= h263_mv4_search(s, mx, my, shift);
1361
            if(dmin4 < dmin){
1362
                mb_type= CANDIDATE_MB_TYPE_INTER4V;
1363
                dmin=dmin4;
1364
            }
1365
        }
1366
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
1367
           && !c->skip){ //FIXME varc/d checks
1368
            int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1369
            if(dmin_i < dmin){
1370
                mb_type = CANDIDATE_MB_TYPE_INTER_I;
1371
                dmin= dmin_i;
1372
            }
1373
        }
1374

    
1375
//        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1376
        set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1377

    
1378
        /* get intra luma score */
1379
        if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1380
            intra_score= varc - 500;
1381
        }else{
1382
            int mean= (sum+128)>>8;
1383
            mean*= 0x01010101;
1384

    
1385
            for(i=0; i<16; i++){
1386
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1387
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1388
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1389
                *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1390
            }
1391

    
1392
            intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1393
        }
1394
#if 0 //FIXME
1395
        /* get chroma score */
1396
        if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1397
            for(i=1; i<3; i++){
1398
                uint8_t *dest_c;
1399
                int mean;
1400

1401
                if(s->out_format == FMT_H263){
1402
                    mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1403
                }else{
1404
                    mean= (s->last_dc[i] + 4)>>3;
1405
                }
1406
                dest_c = s->new_picture.data[i] + (mb_y * 8  * (s->uvlinesize)) + mb_x * 8;
1407

1408
                mean*= 0x01010101;
1409
                for(i=0; i<8; i++){
1410
                    *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1411
                    *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1412
                }
1413

1414
                intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1415
            }
1416
        }
1417
#endif
1418
        intra_score += c->mb_penalty_factor*16;
1419

    
1420
        if(intra_score < dmin){
1421
            mb_type= CANDIDATE_MB_TYPE_INTRA;
1422
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1423
        }else
1424
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1425

    
1426
        {
1427
            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1428
            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1429
            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1430
        }
1431
    }
1432

    
1433
    s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1434
}
1435

    
1436
int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1437
                                    int mb_x, int mb_y)
1438
{
1439
    MotionEstContext * const c= &s->me;
1440
    int mx, my, dmin;
1441
    int P[10][2];
1442
    const int shift= 1+s->quarter_sample;
1443
    const int xy= mb_x + mb_y*s->mb_stride;
1444
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1445

    
1446
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1447

    
1448
    c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1449
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1450

    
1451
    get_limits(s, 16*mb_x, 16*mb_y);
1452
    c->skip=0;
1453

    
1454
    P_LEFT[0]       = s->p_mv_table[xy + 1][0];
1455
    P_LEFT[1]       = s->p_mv_table[xy + 1][1];
1456

    
1457
    if(P_LEFT[0]       < (c->xmin<<shift)) P_LEFT[0]       = (c->xmin<<shift);
1458

    
1459
    /* special case for first line */
1460
    if (s->first_slice_line) {
1461
        c->pred_x= P_LEFT[0];
1462
        c->pred_y= P_LEFT[1];
1463
        P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1464
        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1465
    } else {
1466
        P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
1467
        P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
1468
        P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1469
        P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1470
        if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);
1471
        if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1472
        if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1473

    
1474
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1475
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1476

    
1477
        c->pred_x = P_MEDIAN[0];
1478
        c->pred_y = P_MEDIAN[1];
1479
    }
1480

    
1481
    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1482

    
1483
    s->p_mv_table[xy][0] = mx<<shift;
1484
    s->p_mv_table[xy][1] = my<<shift;
1485

    
1486
    return dmin;
1487
}
1488

    
1489
static int ff_estimate_motion_b(MpegEncContext * s,
1490
                       int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1491
{
1492
    MotionEstContext * const c= &s->me;
1493
    int mx, my, dmin;
1494
    int P[10][2];
1495
    const int shift= 1+s->quarter_sample;
1496
    const int mot_stride = s->mb_stride;
1497
    const int mot_xy = mb_y*mot_stride + mb_x;
1498
    uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1499
    int mv_scale;
1500

    
1501
    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1502
    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1503
    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1504
    c->current_mv_penalty= mv_penalty;
1505

    
1506
    get_limits(s, 16*mb_x, 16*mb_y);
1507

    
1508
    switch(s->me_method) {
1509
    case ME_ZERO:
1510
    default:
1511
        no_motion_search(s, &mx, &my);
1512
        dmin = 0;
1513
        mx-= mb_x*16;
1514
        my-= mb_y*16;
1515
        break;
1516
#if 0
1517
    case ME_FULL:
1518
        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1519
        mx-= mb_x*16;
1520
        my-= mb_y*16;
1521
        break;
1522
    case ME_LOG:
1523
        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1524
        mx-= mb_x*16;
1525
        my-= mb_y*16;
1526
        break;
1527
    case ME_PHODS:
1528
        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1529
        mx-= mb_x*16;
1530
        my-= mb_y*16;
1531
        break;
1532
#endif
1533
    case ME_X1:
1534
    case ME_EPZS:
1535
       {
1536
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
1537
            P_LEFT[1]        = mv_table[mot_xy - 1][1];
1538

    
1539
            if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
1540

    
1541
            /* special case for first line */
1542
            if (!s->first_slice_line) {
1543
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
1544
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
1545
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
1546
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
1547
                if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1548
                if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1549
                if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1550

    
1551
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1552
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1553
            }
1554
            c->pred_x= P_LEFT[0];
1555
            c->pred_y= P_LEFT[1];
1556
        }
1557

    
1558
        if(mv_table == s->b_forw_mv_table){
1559
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1560
        }else{
1561
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1562
        }
1563

    
1564
        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1565

    
1566
        break;
1567
    }
1568

    
1569
    dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1570

    
1571
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1572
        dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1573

    
1574
//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1575
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1576
    mv_table[mot_xy][0]= mx;
1577
    mv_table[mot_xy][1]= my;
1578

    
1579
    return dmin;
1580
}
1581

    
1582
static inline int check_bidir_mv(MpegEncContext * s,
1583
                   int motion_fx, int motion_fy,
1584
                   int motion_bx, int motion_by,
1585
                   int pred_fx, int pred_fy,
1586
                   int pred_bx, int pred_by,
1587
                   int size, int h)
1588
{
1589
    //FIXME optimize?
1590
    //FIXME better f_code prediction (max mv & distance)
1591
    //FIXME pointers
1592
    MotionEstContext * const c= &s->me;
1593
    uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1594
    uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1595
    int stride= c->stride;
1596
    uint8_t *dest_y = c->scratchpad;
1597
    uint8_t *ptr;
1598
    int dxy;
1599
    int src_x, src_y;
1600
    int fbmin;
1601
    uint8_t **src_data= c->src[0];
1602
    uint8_t **ref_data= c->ref[0];
1603
    uint8_t **ref2_data= c->ref[2];
1604

    
1605
    if(s->quarter_sample){
1606
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1607
        src_x = motion_fx >> 2;
1608
        src_y = motion_fy >> 2;
1609

    
1610
        ptr = ref_data[0] + (src_y * stride) + src_x;
1611
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , stride);
1612

    
1613
        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1614
        src_x = motion_bx >> 2;
1615
        src_y = motion_by >> 2;
1616

    
1617
        ptr = ref2_data[0] + (src_y * stride) + src_x;
1618
        s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);
1619
    }else{
1620
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1621
        src_x = motion_fx >> 1;
1622
        src_y = motion_fy >> 1;
1623

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

    
1627
        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1628
        src_x = motion_bx >> 1;
1629
        src_y = motion_by >> 1;
1630

    
1631
        ptr = ref2_data[0] + (src_y * stride) + src_x;
1632
        s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
1633
    }
1634

    
1635
    fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1636
           +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1637
           + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1638

    
1639
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1640
    }
1641
    //FIXME CHROMA !!!
1642

    
1643
    return fbmin;
1644
}
1645

    
1646
/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1647
static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1648
{
1649
    MotionEstContext * const c= &s->me;
1650
    const int mot_stride = s->mb_stride;
1651
    const int xy = mb_y *mot_stride + mb_x;
1652
    int fbmin;
1653
    int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1654
    int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1655
    int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1656
    int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1657
    int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1658
    int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1659
    int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1660
    int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1661
    const int flags= c->sub_flags;
1662
    const int qpel= flags&FLAG_QPEL;
1663
    const int shift= 1+qpel;
1664
    const int xmin= c->xmin<<shift;
1665
    const int ymin= c->ymin<<shift;
1666
    const int xmax= c->xmax<<shift;
1667
    const int ymax= c->ymax<<shift;
1668
    uint8_t map[8][8][8][8];
1669

    
1670
    memset(map,0,sizeof(map));
1671
#define BIDIR_MAP(fx,fy,bx,by) \
1672
    map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1673
    BIDIR_MAP(0,0,0,0) = 1;
1674

    
1675
    fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1676
                          motion_bx, motion_by,
1677
                          pred_fx, pred_fy,
1678
                          pred_bx, pred_by,
1679
                          0, 16);
1680

    
1681
    if(s->avctx->bidir_refine){
1682
        int score, end;
1683
#define CHECK_BIDIR(fx,fy,bx,by)\
1684
    if( !BIDIR_MAP(fx,fy,bx,by)\
1685
       &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1686
       &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1687
        BIDIR_MAP(fx,fy,bx,by) = 1;\
1688
        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);\
1689
        if(score < fbmin){\
1690
            fbmin= score;\
1691
            motion_fx+=fx;\
1692
            motion_fy+=fy;\
1693
            motion_bx+=bx;\
1694
            motion_by+=by;\
1695
            end=0;\
1696
        }\
1697
    }
1698
#define CHECK_BIDIR2(a,b,c,d)\
1699
CHECK_BIDIR(a,b,c,d)\
1700
CHECK_BIDIR(-(a),-(b),-(c),-(d))
1701

    
1702
#define CHECK_BIDIRR(a,b,c,d)\
1703
CHECK_BIDIR2(a,b,c,d)\
1704
CHECK_BIDIR2(b,c,d,a)\
1705
CHECK_BIDIR2(c,d,a,b)\
1706
CHECK_BIDIR2(d,a,b,c)
1707

    
1708
        do{
1709
            end=1;
1710

    
1711
            CHECK_BIDIRR( 0, 0, 0, 1)
1712
            if(s->avctx->bidir_refine > 1){
1713
                CHECK_BIDIRR( 0, 0, 1, 1)
1714
                CHECK_BIDIR2( 0, 1, 0, 1)
1715
                CHECK_BIDIR2( 1, 0, 1, 0)
1716
                CHECK_BIDIRR( 0, 0,-1, 1)
1717
                CHECK_BIDIR2( 0,-1, 0, 1)
1718
                CHECK_BIDIR2(-1, 0, 1, 0)
1719
                if(s->avctx->bidir_refine > 2){
1720
                    CHECK_BIDIRR( 0, 1, 1, 1)
1721
                    CHECK_BIDIRR( 0,-1, 1, 1)
1722
                    CHECK_BIDIRR( 0, 1,-1, 1)
1723
                    CHECK_BIDIRR( 0, 1, 1,-1)
1724
                    if(s->avctx->bidir_refine > 3){
1725
                        CHECK_BIDIR2( 1, 1, 1, 1)
1726
                        CHECK_BIDIRR( 1, 1, 1,-1)
1727
                        CHECK_BIDIR2( 1, 1,-1,-1)
1728
                        CHECK_BIDIR2( 1,-1,-1, 1)
1729
                        CHECK_BIDIR2( 1,-1, 1,-1)
1730
                    }
1731
                }
1732
            }
1733
        }while(!end);
1734
    }
1735

    
1736
    s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1737
    s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1738
    s->b_bidir_back_mv_table[xy][0]= motion_bx;
1739
    s->b_bidir_back_mv_table[xy][1]= motion_by;
1740

    
1741
    return fbmin;
1742
}
1743

    
1744
static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1745
{
1746
    MotionEstContext * const c= &s->me;
1747
    int P[10][2];
1748
    const int mot_stride = s->mb_stride;
1749
    const int mot_xy = mb_y*mot_stride + mb_x;
1750
    const int shift= 1+s->quarter_sample;
1751
    int dmin, i;
1752
    const int time_pp= s->pp_time;
1753
    const int time_pb= s->pb_time;
1754
    int mx, my, xmin, xmax, ymin, ymax;
1755
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
1756

    
1757
    c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1758
    ymin= xmin=(-32)>>shift;
1759
    ymax= xmax=   31>>shift;
1760

    
1761
    if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1762
        s->mv_type= MV_TYPE_8X8;
1763
    }else{
1764
        s->mv_type= MV_TYPE_16X16;
1765
    }
1766

    
1767
    for(i=0; i<4; i++){
1768
        int index= s->block_index[i];
1769
        int min, max;
1770

    
1771
        c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1772
        c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1773
        c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1774
        c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1775
//        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1776
//        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1777

    
1778
        max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1779
        min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1780
        max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1781
        min+= 16*mb_x - 1;
1782
        xmax= FFMIN(xmax, s->width - max);
1783
        xmin= FFMAX(xmin, - 16     - min);
1784

    
1785
        max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1786
        min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1787
        max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1788
        min+= 16*mb_y - 1;
1789
        ymax= FFMIN(ymax, s->height - max);
1790
        ymin= FFMAX(ymin, - 16      - min);
1791

    
1792
        if(s->mv_type == MV_TYPE_16X16) break;
1793
    }
1794

    
1795
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1796

    
1797
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1798
        s->b_direct_mv_table[mot_xy][0]= 0;
1799
        s->b_direct_mv_table[mot_xy][1]= 0;
1800

    
1801
        return 256*256*256*64;
1802
    }
1803

    
1804
    c->xmin= xmin;
1805
    c->ymin= ymin;
1806
    c->xmax= xmax;
1807
    c->ymax= ymax;
1808
    c->flags     |= FLAG_DIRECT;
1809
    c->sub_flags |= FLAG_DIRECT;
1810
    c->pred_x=0;
1811
    c->pred_y=0;
1812

    
1813
    P_LEFT[0]        = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1814
    P_LEFT[1]        = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1815

    
1816
    /* special case for first line */
1817
    if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as it is clipped
1818
        P_TOP[0]      = av_clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
1819
        P_TOP[1]      = av_clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
1820
        P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
1821
        P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
1822

    
1823
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1824
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1825
    }
1826

    
1827
    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1828
    if(c->sub_flags&FLAG_QPEL)
1829
        dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1830
    else
1831
        dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1832

    
1833
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1834
        dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1835

    
1836
    get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1837

    
1838
    mv_table[mot_xy][0]= mx;
1839
    mv_table[mot_xy][1]= my;
1840
    c->flags     &= ~FLAG_DIRECT;
1841
    c->sub_flags &= ~FLAG_DIRECT;
1842

    
1843
    return dmin;
1844
}
1845

    
1846
void ff_estimate_b_frame_motion(MpegEncContext * s,
1847
                             int mb_x, int mb_y)
1848
{
1849
    MotionEstContext * const c= &s->me;
1850
    const int penalty_factor= c->mb_penalty_factor;
1851
    int fmin, bmin, dmin, fbmin, bimin, fimin;
1852
    int type=0;
1853
    const int xy = mb_y*s->mb_stride + mb_x;
1854
    init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1855

    
1856
    get_limits(s, 16*mb_x, 16*mb_y);
1857

    
1858
    c->skip=0;
1859

    
1860
    if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
1861
        int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1862

    
1863
        score= ((unsigned)(score*score + 128*256))>>16;
1864
        c->mc_mb_var_sum_temp += score;
1865
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1866
        s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1867

    
1868
        return;
1869
    }
1870

    
1871
    if(c->avctx->me_threshold){
1872
        int vard= check_input_motion(s, mb_x, mb_y, 0);
1873

    
1874
        if((vard+128)>>8 < c->avctx->me_threshold){
1875
//            pix = c->src[0][0];
1876
//            sum = s->dsp.pix_sum(pix, s->linesize);
1877
//            varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1878

    
1879
//            pic->mb_var   [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1880
             s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1881
/*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1882
            c->mb_var_sum_temp    += (varc+128)>>8;*/
1883
            c->mc_mb_var_sum_temp += (vard+128)>>8;
1884
/*            if (vard <= 64<<8 || vard < varc) {
1885
                c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1886
            }else{
1887
                c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1888
            }*/
1889
            return;
1890
        }
1891
        if((vard+128)>>8 < c->avctx->mb_threshold){
1892
            type= s->mb_type[mb_y*s->mb_stride + mb_x];
1893
            if(type == CANDIDATE_MB_TYPE_DIRECT){
1894
                direct_search(s, mb_x, mb_y);
1895
            }
1896
            if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1897
                c->skip=0;
1898
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1899
            }
1900
            if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1901
                c->skip=0;
1902
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1903
            }
1904
            if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1905
                c->skip=0;
1906
                c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1907
                interlaced_search(s, 0,
1908
                                        s->b_field_mv_table[0], s->b_field_select_table[0],
1909
                                        s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1910
            }
1911
            if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1912
                c->skip=0;
1913
                c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1914
                interlaced_search(s, 2,
1915
                                        s->b_field_mv_table[1], s->b_field_select_table[1],
1916
                                        s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1917
            }
1918
            return;
1919
        }
1920
    }
1921

    
1922
    if (s->codec_id == CODEC_ID_MPEG4)
1923
        dmin= direct_search(s, mb_x, mb_y);
1924
    else
1925
        dmin= INT_MAX;
1926
//FIXME penalty stuff for non mpeg4
1927
    c->skip=0;
1928
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1929

    
1930
    c->skip=0;
1931
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1932
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1933

    
1934
    c->skip=0;
1935
    fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1936
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1937

    
1938
    if(s->flags & CODEC_FLAG_INTERLACED_ME){
1939
//FIXME mb type penalty
1940
        c->skip=0;
1941
        c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1942
        fimin= interlaced_search(s, 0,
1943
                                 s->b_field_mv_table[0], s->b_field_select_table[0],
1944
                                 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1945
        c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1946
        bimin= interlaced_search(s, 2,
1947
                                 s->b_field_mv_table[1], s->b_field_select_table[1],
1948
                                 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1949
    }else
1950
        fimin= bimin= INT_MAX;
1951

    
1952
    {
1953
        int score= fmin;
1954
        type = CANDIDATE_MB_TYPE_FORWARD;
1955

    
1956
        if (dmin <= score){
1957
            score = dmin;
1958
            type = CANDIDATE_MB_TYPE_DIRECT;
1959
        }
1960
        if(bmin<score){
1961
            score=bmin;
1962
            type= CANDIDATE_MB_TYPE_BACKWARD;
1963
        }
1964
        if(fbmin<score){
1965
            score=fbmin;
1966
            type= CANDIDATE_MB_TYPE_BIDIR;
1967
        }
1968
        if(fimin<score){
1969
            score=fimin;
1970
            type= CANDIDATE_MB_TYPE_FORWARD_I;
1971
        }
1972
        if(bimin<score){
1973
            score=bimin;
1974
            type= CANDIDATE_MB_TYPE_BACKWARD_I;
1975
        }
1976

    
1977
        score= ((unsigned)(score*score + 128*256))>>16;
1978
        c->mc_mb_var_sum_temp += score;
1979
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1980
    }
1981

    
1982
    if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1983
        type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1984
        if(fimin < INT_MAX)
1985
            type |= CANDIDATE_MB_TYPE_FORWARD_I;
1986
        if(bimin < INT_MAX)
1987
            type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1988
        if(fimin < INT_MAX && bimin < INT_MAX){
1989
            type |= CANDIDATE_MB_TYPE_BIDIR_I;
1990
        }
1991
         //FIXME something smarter
1992
        if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //do not try direct mode if it is invalid for this MB
1993
        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])
1994
            type |= CANDIDATE_MB_TYPE_DIRECT0;
1995
#if 0
1996
        if(s->out_format == FMT_MPEG1)
1997
            type |= CANDIDATE_MB_TYPE_INTRA;
1998
#endif
1999
    }
2000

    
2001
    s->mb_type[mb_y*s->mb_stride + mb_x]= type;
2002
}
2003

    
2004
/* find best f_code for ME which do unlimited searches */
2005
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
2006
{
2007
    if(s->me_method>=ME_EPZS){
2008
        int score[8];
2009
        int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
2010
        uint8_t * fcode_tab= s->fcode_tab;
2011
        int best_fcode=-1;
2012
        int best_score=-10000000;
2013

    
2014
        if(s->msmpeg4_version)
2015
            range= FFMIN(range, 16);
2016
        else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
2017
            range= FFMIN(range, 256);
2018

    
2019
        for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
2020

    
2021
        for(y=0; y<s->mb_height; y++){
2022
            int x;
2023
            int xy= y*s->mb_stride;
2024
            for(x=0; x<s->mb_width; x++){
2025
                if(s->mb_type[xy] & type){
2026
                    int mx= mv_table[xy][0];
2027
                    int my= mv_table[xy][1];
2028
                    int fcode= FFMAX(fcode_tab[mx + MAX_MV],
2029
                                     fcode_tab[my + MAX_MV]);
2030
                    int j;
2031

    
2032
                        if(mx >= range || mx < -range ||
2033
                           my >= range || my < -range)
2034
                            continue;
2035

    
2036
                    for(j=0; j<fcode && j<8; j++){
2037
                        if(s->pict_type==FF_B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
2038
                            score[j]-= 170;
2039
                    }
2040
                }
2041
                xy++;
2042
            }
2043
        }
2044

    
2045
        for(i=1; i<8; i++){
2046
            if(score[i] > best_score){
2047
                best_score= score[i];
2048
                best_fcode= i;
2049
            }
2050
//            printf("%d %d\n", i, score[i]);
2051
        }
2052

    
2053
//    printf("fcode: %d type: %d\n", i, s->pict_type);
2054
        return best_fcode;
2055
/*        for(i=0; i<=MAX_FCODE; i++){
2056
            printf("%d ", mv_num[i]);
2057
        }
2058
        printf("\n");*/
2059
    }else{
2060
        return 1;
2061
    }
2062
}
2063

    
2064
void ff_fix_long_p_mvs(MpegEncContext * s)
2065
{
2066
    MotionEstContext * const c= &s->me;
2067
    const int f_code= s->f_code;
2068
    int y, range;
2069
    assert(s->pict_type==FF_P_TYPE);
2070

    
2071
    range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2072

    
2073
    assert(range <= 16 || !s->msmpeg4_version);
2074
    assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
2075

    
2076
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2077

    
2078
//printf("%d no:%d %d//\n", clip, noclip, f_code);
2079
    if(s->flags&CODEC_FLAG_4MV){
2080
        const int wrap= s->b8_stride;
2081

    
2082
        /* clip / convert to intra 8x8 type MVs */
2083
        for(y=0; y<s->mb_height; y++){
2084
            int xy= y*2*wrap;
2085
            int i= y*s->mb_stride;
2086
            int x;
2087

    
2088
            for(x=0; x<s->mb_width; x++){
2089
                if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
2090
                    int block;
2091
                    for(block=0; block<4; block++){
2092
                        int off= (block& 1) + (block>>1)*wrap;
2093
                        int mx= s->current_picture.motion_val[0][ xy + off ][0];
2094
                        int my= s->current_picture.motion_val[0][ xy + off ][1];
2095

    
2096
                        if(   mx >=range || mx <-range
2097
                           || my >=range || my <-range){
2098
                            s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
2099
                            s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
2100
                            s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2101
                        }
2102
                    }
2103
                }
2104
                xy+=2;
2105
                i++;
2106
            }
2107
        }
2108
    }
2109
}
2110

    
2111
/**
2112
 *
2113
 * @param truncate 1 for truncation, 0 for using intra
2114
 */
2115
void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
2116
                     int16_t (*mv_table)[2], int f_code, int type, int truncate)
2117
{
2118
    MotionEstContext * const c= &s->me;
2119
    int y, h_range, v_range;
2120

    
2121
    // RAL: 8 in MPEG-1, 16 in MPEG-4
2122
    int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2123

    
2124
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2125

    
2126
    h_range= range;
2127
    v_range= field_select_table ? range>>1 : range;
2128

    
2129
    /* clip / convert to intra 16x16 type MVs */
2130
    for(y=0; y<s->mb_height; y++){
2131
        int x;
2132
        int xy= y*s->mb_stride;
2133
        for(x=0; x<s->mb_width; x++){
2134
            if (s->mb_type[xy] & type){    // RAL: "type" test added...
2135
                if(field_select_table==NULL || field_select_table[xy] == field_select){
2136
                    if(   mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2137
                       || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2138

    
2139
                        if(truncate){
2140
                            if     (mv_table[xy][0] > h_range-1) mv_table[xy][0]=  h_range-1;
2141
                            else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2142
                            if     (mv_table[xy][1] > v_range-1) mv_table[xy][1]=  v_range-1;
2143
                            else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2144
                        }else{
2145
                            s->mb_type[xy] &= ~type;
2146
                            s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
2147
                            mv_table[xy][0]=
2148
                            mv_table[xy][1]= 0;
2149
                        }
2150
                    }
2151
                }
2152
            }
2153
            xy++;
2154
        }
2155
    }
2156
}