Statistics
| Branch: | Revision:

ffmpeg / libavcodec / motion_est.c @ c6c36725

History | View | Annotate | Download (74.8 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
static always_inline int cmp(MpegEncContext *s, const int x, const int y, const int subx, const int suby,
107
                      const int size, const int h, int ref_index, int src_index,
108
                      me_cmp_func cmp_func, me_cmp_func chroma_cmp_func, const int flags){
109
    MotionEstContext * const c= &s->me;
110
    const int stride= c->stride;
111
    const int uvstride= c->uvstride;
112
    const int qpel= flags&FLAG_QPEL;
113
    const int chroma= flags&FLAG_CHROMA;
114
    const int dxy= subx + (suby<<(1+qpel)); //FIXME log2_subpel?
115
    const int hx= subx + (x<<(1+qpel));
116
    const int hy= suby + (y<<(1+qpel));
117
    uint8_t * const * const ref= c->ref[ref_index];
118
    uint8_t * const * const src= c->src[src_index];
119
    int d;
120
    //FIXME check chroma 4mv, (no crashes ...)
121
    if(flags&FLAG_DIRECT){
122
        if(x >= c->xmin && hx <= c->xmax<<(qpel+1) && y >= c->ymin && hy <= c->ymax<<(qpel+1)){
123
            const int time_pp= s->pp_time;
124
            const int time_pb= s->pb_time;
125
            const int mask= 2*qpel+1;
126
            if(s->mv_type==MV_TYPE_8X8){
127
                int i;
128
                for(i=0; i<4; i++){
129
                    int fx = c->direct_basis_mv[i][0] + hx;
130
                    int fy = c->direct_basis_mv[i][1] + hy;
131
                    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));
132
                    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));
133
                    int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
134
                    int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
135

    
136
                    uint8_t *dst= c->temp + 8*(i&1) + 8*stride*(i>>1);
137
                    if(qpel){
138
                        c->qpel_put[1][fxy](dst, ref[0] + (fx>>2) + (fy>>2)*stride, stride);
139
                        c->qpel_avg[1][bxy](dst, ref[8] + (bx>>2) + (by>>2)*stride, stride);
140
                    }else{
141
                        c->hpel_put[1][fxy](dst, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 8);
142
                        c->hpel_avg[1][bxy](dst, ref[8] + (bx>>1) + (by>>1)*stride, stride, 8);
143
                    }
144
                }
145
            }else{
146
                int fx = c->direct_basis_mv[0][0] + hx;
147
                int fy = c->direct_basis_mv[0][1] + hy;
148
                int bx = hx ? fx - c->co_located_mv[0][0] : (c->co_located_mv[0][0]*(time_pb - time_pp)/time_pp);
149
                int by = hy ? fy - c->co_located_mv[0][1] : (c->co_located_mv[0][1]*(time_pb - time_pp)/time_pp);
150
                int fxy= (fx&mask) + ((fy&mask)<<(qpel+1));
151
                int bxy= (bx&mask) + ((by&mask)<<(qpel+1));
152

    
153
                if(qpel){
154
                    c->qpel_put[1][fxy](c->temp               , ref[0] + (fx>>2) + (fy>>2)*stride               , stride);
155
                    c->qpel_put[1][fxy](c->temp + 8           , ref[0] + (fx>>2) + (fy>>2)*stride + 8           , stride);
156
                    c->qpel_put[1][fxy](c->temp     + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride     + 8*stride, stride);
157
                    c->qpel_put[1][fxy](c->temp + 8 + 8*stride, ref[0] + (fx>>2) + (fy>>2)*stride + 8 + 8*stride, stride);
158
                    c->qpel_avg[1][bxy](c->temp               , ref[8] + (bx>>2) + (by>>2)*stride               , stride);
159
                    c->qpel_avg[1][bxy](c->temp + 8           , ref[8] + (bx>>2) + (by>>2)*stride + 8           , stride);
160
                    c->qpel_avg[1][bxy](c->temp     + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride     + 8*stride, stride);
161
                    c->qpel_avg[1][bxy](c->temp + 8 + 8*stride, ref[8] + (bx>>2) + (by>>2)*stride + 8 + 8*stride, stride);
162
                }else{
163
                    assert((fx>>1) + 16*s->mb_x >= -16);
164
                    assert((fy>>1) + 16*s->mb_y >= -16);
165
                    assert((fx>>1) + 16*s->mb_x <= s->width);
166
                    assert((fy>>1) + 16*s->mb_y <= s->height);
167
                    assert((bx>>1) + 16*s->mb_x >= -16);
168
                    assert((by>>1) + 16*s->mb_y >= -16);
169
                    assert((bx>>1) + 16*s->mb_x <= s->width);
170
                    assert((by>>1) + 16*s->mb_y <= s->height);
171

    
172
                    c->hpel_put[0][fxy](c->temp, ref[0] + (fx>>1) + (fy>>1)*stride, stride, 16);
173
                    c->hpel_avg[0][bxy](c->temp, ref[8] + (bx>>1) + (by>>1)*stride, stride, 16);
174
                }
175
            }
176
            d = cmp_func(s, c->temp, src[0], stride, 16);
177
        }else
178
            d= 256*256*256*32;
179
    }else{
180
        int uvdxy;              /* no, it might not be used uninitialized */
181
        if(dxy){
182
            if(qpel){
183
                c->qpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride); //FIXME prototype (add h)
184
                if(chroma){
185
                    int cx= hx/2;
186
                    int cy= hy/2;
187
                    cx= (cx>>1)|(cx&1);
188
                    cy= (cy>>1)|(cy&1);
189
                    uvdxy= (cx&1) + 2*(cy&1);
190
                    //FIXME x/y wrong, but mpeg4 qpel is sick anyway, we should drop as much of it as possible in favor for h264
191
                }
192
            }else{
193
                c->hpel_put[size][dxy](c->temp, ref[0] + x + y*stride, stride, h);
194
                if(chroma)
195
                    uvdxy= dxy | (x&1) | (2*(y&1));
196
            }
197
            d = cmp_func(s, c->temp, src[0], stride, h);
198
        }else{
199
            d = cmp_func(s, src[0], ref[0] + x + y*stride, stride, h);
200
            if(chroma)
201
                uvdxy= (x&1) + 2*(y&1);
202
        }
203
        if(chroma){
204
            uint8_t * const uvtemp= c->temp + 16*stride;
205
            c->hpel_put[size+1][uvdxy](uvtemp  , ref[1] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
206
            c->hpel_put[size+1][uvdxy](uvtemp+8, ref[2] + (x>>1) + (y>>1)*uvstride, uvstride, h>>1);
207
            d += chroma_cmp_func(s, uvtemp  , src[1], uvstride, h>>1);
208
            d += chroma_cmp_func(s, uvtemp+8, src[2], uvstride, h>>1);
209
        }
210
    }
211
#if 0
212
    if(full_pel){
213
        const int index= (((y)<<ME_MAP_SHIFT) + (x))&(ME_MAP_SIZE-1);
214
        score_map[index]= d;
215
    }
216

217
    d += (c->mv_penalty[hx - c->pred_x] + c->mv_penalty[hy - c->pred_y])*c->penalty_factor;
218
#endif
219
    return d;
220
}
221

    
222
#include "motion_est_template.c"
223

    
224
static int zero_cmp(void *s, uint8_t *a, uint8_t *b, int stride, int h){
225
    return 0;
226
}
227

    
228
static void zero_hpel(uint8_t *a, const uint8_t *b, int stride, int h){
229
}
230

    
231
void ff_init_me(MpegEncContext *s){
232
    MotionEstContext * const c= &s->me;
233
    c->avctx= s->avctx;
234

    
235
    ff_set_cmp(&s->dsp, s->dsp.me_pre_cmp, c->avctx->me_pre_cmp);
236
    ff_set_cmp(&s->dsp, s->dsp.me_cmp, c->avctx->me_cmp);
237
    ff_set_cmp(&s->dsp, s->dsp.me_sub_cmp, c->avctx->me_sub_cmp);
238
    ff_set_cmp(&s->dsp, s->dsp.mb_cmp, c->avctx->mb_cmp);
239

    
240
    c->flags    = get_flags(c, 0, c->avctx->me_cmp    &FF_CMP_CHROMA);
241
    c->sub_flags= get_flags(c, 0, c->avctx->me_sub_cmp&FF_CMP_CHROMA);
242
    c->mb_flags = get_flags(c, 0, c->avctx->mb_cmp    &FF_CMP_CHROMA);
243

    
244
/*FIXME s->no_rounding b_type*/
245
    if(s->flags&CODEC_FLAG_QPEL){
246
        c->sub_motion_search= qpel_motion_search;
247
        c->qpel_avg= s->dsp.avg_qpel_pixels_tab;
248
        if(s->no_rounding) c->qpel_put= s->dsp.put_no_rnd_qpel_pixels_tab;
249
        else               c->qpel_put= s->dsp.put_qpel_pixels_tab;
250
    }else{
251
        if(c->avctx->me_sub_cmp&FF_CMP_CHROMA)
252
            c->sub_motion_search= hpel_motion_search;
253
        else if(   c->avctx->me_sub_cmp == FF_CMP_SAD
254
                && c->avctx->    me_cmp == FF_CMP_SAD
255
                && c->avctx->    mb_cmp == FF_CMP_SAD)
256
            c->sub_motion_search= sad_hpel_motion_search; // 2050 vs. 2450 cycles
257
        else
258
            c->sub_motion_search= hpel_motion_search;
259
    }
260
    c->hpel_avg= s->dsp.avg_pixels_tab;
261
    if(s->no_rounding) c->hpel_put= s->dsp.put_no_rnd_pixels_tab;
262
    else               c->hpel_put= s->dsp.put_pixels_tab;
263

    
264
    if(s->linesize){
265
        c->stride  = s->linesize;
266
        c->uvstride= s->uvlinesize;
267
    }else{
268
        c->stride  = 16*s->mb_width + 32;
269
        c->uvstride=  8*s->mb_width + 16;
270
    }
271

    
272
    // 8x8 fullpel search would need a 4x4 chroma compare, which we dont have yet, and even if we had the motion estimation code doesnt expect it
273
    if(s->codec_id != CODEC_ID_SNOW){
274
        if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
275
            s->dsp.me_cmp[2]= zero_cmp;
276
        }
277
        if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
278
            s->dsp.me_sub_cmp[2]= zero_cmp;
279
        }
280
        c->hpel_put[2][0]= c->hpel_put[2][1]=
281
        c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
282
    }
283

    
284
    if(s->codec_id == CODEC_ID_H261){
285
        c->sub_motion_search= no_sub_motion_search;
286
    }
287

    
288
    c->temp= c->scratchpad;
289
}
290

    
291
#if 0
292
static int pix_dev(uint8_t * pix, int line_size, int mean)
293
{
294
    int s, i, j;
295

296
    s = 0;
297
    for (i = 0; i < 16; i++) {
298
        for (j = 0; j < 16; j += 8) {
299
            s += FFABS(pix[0]-mean);
300
            s += FFABS(pix[1]-mean);
301
            s += FFABS(pix[2]-mean);
302
            s += FFABS(pix[3]-mean);
303
            s += FFABS(pix[4]-mean);
304
            s += FFABS(pix[5]-mean);
305
            s += FFABS(pix[6]-mean);
306
            s += FFABS(pix[7]-mean);
307
            pix += 8;
308
        }
309
        pix += line_size - 16;
310
    }
311
    return s;
312
}
313
#endif
314

    
315
static inline void no_motion_search(MpegEncContext * s,
316
                                    int *mx_ptr, int *my_ptr)
317
{
318
    *mx_ptr = 16 * s->mb_x;
319
    *my_ptr = 16 * s->mb_y;
320
}
321

    
322
#if 0  /* the use of these functions is inside #if 0 */
323
static int full_motion_search(MpegEncContext * s,
324
                              int *mx_ptr, int *my_ptr, int range,
325
                              int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
326
{
327
    int x1, y1, x2, y2, xx, yy, x, y;
328
    int mx, my, dmin, d;
329
    uint8_t *pix;
330

331
    xx = 16 * s->mb_x;
332
    yy = 16 * s->mb_y;
333
    x1 = xx - range + 1;        /* we loose one pixel to avoid boundary pb with half pixel pred */
334
    if (x1 < xmin)
335
        x1 = xmin;
336
    x2 = xx + range - 1;
337
    if (x2 > xmax)
338
        x2 = xmax;
339
    y1 = yy - range + 1;
340
    if (y1 < ymin)
341
        y1 = ymin;
342
    y2 = yy + range - 1;
343
    if (y2 > ymax)
344
        y2 = ymax;
345
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
346
    dmin = 0x7fffffff;
347
    mx = 0;
348
    my = 0;
349
    for (y = y1; y <= y2; y++) {
350
        for (x = x1; x <= x2; x++) {
351
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
352
                             s->linesize, 16);
353
            if (d < dmin ||
354
                (d == dmin &&
355
                 (abs(x - xx) + abs(y - yy)) <
356
                 (abs(mx - xx) + abs(my - yy)))) {
357
                dmin = d;
358
                mx = x;
359
                my = y;
360
            }
361
        }
362
    }
363

364
    *mx_ptr = mx;
365
    *my_ptr = my;
366

367
#if 0
368
    if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
369
        *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
370
        av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
371
    }
372
#endif
373
    return dmin;
374
}
375

    
376

    
377
static int log_motion_search(MpegEncContext * s,
378
                             int *mx_ptr, int *my_ptr, int range,
379
                             int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
380
{
381
    int x1, y1, x2, y2, xx, yy, x, y;
382
    int mx, my, dmin, d;
383
    uint8_t *pix;
384

    
385
    xx = s->mb_x << 4;
386
    yy = s->mb_y << 4;
387

    
388
    /* Left limit */
389
    x1 = xx - range;
390
    if (x1 < xmin)
391
        x1 = xmin;
392

    
393
    /* Right limit */
394
    x2 = xx + range;
395
    if (x2 > xmax)
396
        x2 = xmax;
397

    
398
    /* Upper limit */
399
    y1 = yy - range;
400
    if (y1 < ymin)
401
        y1 = ymin;
402

    
403
    /* Lower limit */
404
    y2 = yy + range;
405
    if (y2 > ymax)
406
        y2 = ymax;
407

    
408
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
409
    dmin = 0x7fffffff;
410
    mx = 0;
411
    my = 0;
412

    
413
    do {
414
        for (y = y1; y <= y2; y += range) {
415
            for (x = x1; x <= x2; x += range) {
416
                d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
417
                if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
418
                    dmin = d;
419
                    mx = x;
420
                    my = y;
421
                }
422
            }
423
        }
424

    
425
        range = range >> 1;
426

    
427
        x1 = mx - range;
428
        if (x1 < xmin)
429
            x1 = xmin;
430

    
431
        x2 = mx + range;
432
        if (x2 > xmax)
433
            x2 = xmax;
434

    
435
        y1 = my - range;
436
        if (y1 < ymin)
437
            y1 = ymin;
438

    
439
        y2 = my + range;
440
        if (y2 > ymax)
441
            y2 = ymax;
442

    
443
    } while (range >= 1);
444

    
445
#ifdef DEBUG
446
    av_log(s->avctx, AV_LOG_DEBUG, "log       - MX: %d\tMY: %d\n", mx, my);
447
#endif
448
    *mx_ptr = mx;
449
    *my_ptr = my;
450
    return dmin;
451
}
452

    
453
static int phods_motion_search(MpegEncContext * s,
454
                               int *mx_ptr, int *my_ptr, int range,
455
                               int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
456
{
457
    int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
458
    int mx, my, dminx, dminy;
459
    uint8_t *pix;
460

    
461
    xx = s->mb_x << 4;
462
    yy = s->mb_y << 4;
463

    
464
    /* Left limit */
465
    x1 = xx - range;
466
    if (x1 < xmin)
467
        x1 = xmin;
468

    
469
    /* Right limit */
470
    x2 = xx + range;
471
    if (x2 > xmax)
472
        x2 = xmax;
473

    
474
    /* Upper limit */
475
    y1 = yy - range;
476
    if (y1 < ymin)
477
        y1 = ymin;
478

    
479
    /* Lower limit */
480
    y2 = yy + range;
481
    if (y2 > ymax)
482
        y2 = ymax;
483

    
484
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
485
    mx = 0;
486
    my = 0;
487

    
488
    x = xx;
489
    y = yy;
490
    do {
491
        dminx = 0x7fffffff;
492
        dminy = 0x7fffffff;
493

    
494
        lastx = x;
495
        for (x = x1; x <= x2; x += range) {
496
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
497
            if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
498
                dminx = d;
499
                mx = x;
500
            }
501
        }
502

    
503
        x = lastx;
504
        for (y = y1; y <= y2; y += range) {
505
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
506
            if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
507
                dminy = d;
508
                my = y;
509
            }
510
        }
511

    
512
        range = range >> 1;
513

    
514
        x = mx;
515
        y = my;
516
        x1 = mx - range;
517
        if (x1 < xmin)
518
            x1 = xmin;
519

    
520
        x2 = mx + range;
521
        if (x2 > xmax)
522
            x2 = xmax;
523

    
524
        y1 = my - range;
525
        if (y1 < ymin)
526
            y1 = ymin;
527

    
528
        y2 = my + range;
529
        if (y2 > ymax)
530
            y2 = ymax;
531

    
532
    } while (range >= 1);
533

    
534
#ifdef DEBUG
535
    av_log(s->avctx, AV_LOG_DEBUG, "phods     - MX: %d\tMY: %d\n", mx, my);
536
#endif
537

    
538
    /* half pixel search */
539
    *mx_ptr = mx;
540
    *my_ptr = my;
541
    return dminy;
542
}
543
#endif /* 0 */
544

    
545
#define Z_THRESHOLD 256
546

    
547
#define CHECK_SAD_HALF_MV(suffix, x, y) \
548
{\
549
    d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
550
    d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
551
    COPY3_IF_LT(dminh, d, dx, x, dy, y)\
552
}
553

    
554
static inline int sad_hpel_motion_search(MpegEncContext * s,
555
                                  int *mx_ptr, int *my_ptr, int dmin,
556
                                  int src_index, int ref_index,
557
                                  int size, int h)
558
{
559
    MotionEstContext * const c= &s->me;
560
    const int penalty_factor= c->sub_penalty_factor;
561
    int mx, my, dminh;
562
    uint8_t *pix, *ptr;
563
    int stride= c->stride;
564
    const int flags= c->sub_flags;
565
    LOAD_COMMON
566

    
567
    assert(flags == 0);
568

    
569
    if(c->skip){
570
//    printf("S");
571
        *mx_ptr = 0;
572
        *my_ptr = 0;
573
        return dmin;
574
    }
575
//    printf("N");
576

    
577
    pix = c->src[src_index][0];
578

    
579
    mx = *mx_ptr;
580
    my = *my_ptr;
581
    ptr = c->ref[ref_index][0] + (my * stride) + mx;
582

    
583
    dminh = dmin;
584

    
585
    if (mx > xmin && mx < xmax &&
586
        my > ymin && my < ymax) {
587
        int dx=0, dy=0;
588
        int d, pen_x, pen_y;
589
        const int index= (my<<ME_MAP_SHIFT) + mx;
590
        const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
591
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
592
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
593
        const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
594
        mx<<=1;
595
        my<<=1;
596

    
597

    
598
        pen_x= pred_x + mx;
599
        pen_y= pred_y + my;
600

    
601
        ptr-= stride;
602
        if(t<=b){
603
            CHECK_SAD_HALF_MV(y2 , 0, -1)
604
            if(l<=r){
605
                CHECK_SAD_HALF_MV(xy2, -1, -1)
606
                if(t+r<=b+l){
607
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
608
                    ptr+= stride;
609
                }else{
610
                    ptr+= stride;
611
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
612
                }
613
                CHECK_SAD_HALF_MV(x2 , -1,  0)
614
            }else{
615
                CHECK_SAD_HALF_MV(xy2, +1, -1)
616
                if(t+l<=b+r){
617
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
618
                    ptr+= stride;
619
                }else{
620
                    ptr+= stride;
621
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
622
                }
623
                CHECK_SAD_HALF_MV(x2 , +1,  0)
624
            }
625
        }else{
626
            if(l<=r){
627
                if(t+l<=b+r){
628
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
629
                    ptr+= stride;
630
                }else{
631
                    ptr+= stride;
632
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
633
                }
634
                CHECK_SAD_HALF_MV(x2 , -1,  0)
635
                CHECK_SAD_HALF_MV(xy2, -1, +1)
636
            }else{
637
                if(t+r<=b+l){
638
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
639
                    ptr+= stride;
640
                }else{
641
                    ptr+= stride;
642
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
643
                }
644
                CHECK_SAD_HALF_MV(x2 , +1,  0)
645
                CHECK_SAD_HALF_MV(xy2, +1, +1)
646
            }
647
            CHECK_SAD_HALF_MV(y2 ,  0, +1)
648
        }
649
        mx+=dx;
650
        my+=dy;
651

    
652
    }else{
653
        mx<<=1;
654
        my<<=1;
655
    }
656

    
657
    *mx_ptr = mx;
658
    *my_ptr = my;
659
    return dminh;
660
}
661

    
662
static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
663
{
664
    const int xy= s->mb_x + s->mb_y*s->mb_stride;
665

    
666
    s->p_mv_table[xy][0] = mx;
667
    s->p_mv_table[xy][1] = my;
668

    
669
    /* has already been set to the 4 MV if 4MV is done */
670
    if(mv4){
671
        int mot_xy= s->block_index[0];
672

    
673
        s->current_picture.motion_val[0][mot_xy  ][0]= mx;
674
        s->current_picture.motion_val[0][mot_xy  ][1]= my;
675
        s->current_picture.motion_val[0][mot_xy+1][0]= mx;
676
        s->current_picture.motion_val[0][mot_xy+1][1]= my;
677

    
678
        mot_xy += s->b8_stride;
679
        s->current_picture.motion_val[0][mot_xy  ][0]= mx;
680
        s->current_picture.motion_val[0][mot_xy  ][1]= my;
681
        s->current_picture.motion_val[0][mot_xy+1][0]= mx;
682
        s->current_picture.motion_val[0][mot_xy+1][1]= my;
683
    }
684
}
685

    
686
/**
687
 * get fullpel ME search limits.
688
 */
689
static inline void get_limits(MpegEncContext *s, int x, int y)
690
{
691
    MotionEstContext * const c= &s->me;
692
/*
693
    if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
694
    else                   c->range= 16;
695
*/
696
    if (s->unrestricted_mv) {
697
        c->xmin = - x - 16;
698
        c->ymin = - y - 16;
699
        c->xmax = - x + s->mb_width *16;
700
        c->ymax = - y + s->mb_height*16;
701
    } else if (s->out_format == FMT_H261){
702
        // Search range of H261 is different from other codec standards
703
        c->xmin = (x > 15) ? - 15 : 0;
704
        c->ymin = (y > 15) ? - 15 : 0;
705
        c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
706
        c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
707
    } else {
708
        c->xmin = - x;
709
        c->ymin = - y;
710
        c->xmax = - x + s->mb_width *16 - 16;
711
        c->ymax = - y + s->mb_height*16 - 16;
712
    }
713
}
714

    
715
static inline void init_mv4_ref(MotionEstContext *c){
716
    const int stride= c->stride;
717

    
718
    c->ref[1][0] = c->ref[0][0] + 8;
719
    c->ref[2][0] = c->ref[0][0] + 8*stride;
720
    c->ref[3][0] = c->ref[2][0] + 8;
721
    c->src[1][0] = c->src[0][0] + 8;
722
    c->src[2][0] = c->src[0][0] + 8*stride;
723
    c->src[3][0] = c->src[2][0] + 8;
724
}
725

    
726
static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
727
{
728
    MotionEstContext * const c= &s->me;
729
    const int size= 1;
730
    const int h=8;
731
    int block;
732
    int P[10][2];
733
    int dmin_sum=0, mx4_sum=0, my4_sum=0;
734
    int same=1;
735
    const int stride= c->stride;
736
    uint8_t *mv_penalty= c->current_mv_penalty;
737

    
738
    init_mv4_ref(c);
739

    
740
    for(block=0; block<4; block++){
741
        int mx4, my4;
742
        int pred_x4, pred_y4;
743
        int dmin4;
744
        static const int off[4]= {2, 1, 1, -1};
745
        const int mot_stride = s->b8_stride;
746
        const int mot_xy = s->block_index[block];
747

    
748
        P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
749
        P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
750

    
751
        if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
752

    
753
        /* special case for first line */
754
        if (s->first_slice_line && block<2) {
755
            c->pred_x= pred_x4= P_LEFT[0];
756
            c->pred_y= pred_y4= P_LEFT[1];
757
        } else {
758
            P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][0];
759
            P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][1];
760
            P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
761
            P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
762
            if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
763
            if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
764
            if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
765
            if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
766

    
767
            P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
768
            P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
769

    
770
            c->pred_x= pred_x4 = P_MEDIAN[0];
771
            c->pred_y= pred_y4 = P_MEDIAN[1];
772
        }
773
        P_MV1[0]= mx;
774
        P_MV1[1]= my;
775

    
776
        dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
777

    
778
        dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
779

    
780
        if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
781
            int dxy;
782
            const int offset= ((block&1) + (block>>1)*stride)*8;
783
            uint8_t *dest_y = c->scratchpad + offset;
784
            if(s->quarter_sample){
785
                uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
786
                dxy = ((my4 & 3) << 2) | (mx4 & 3);
787

    
788
                if(s->no_rounding)
789
                    s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , stride);
790
                else
791
                    s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , stride);
792
            }else{
793
                uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
794
                dxy = ((my4 & 1) << 1) | (mx4 & 1);
795

    
796
                if(s->no_rounding)
797
                    s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , stride, h);
798
                else
799
                    s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , stride, h);
800
            }
801
            dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
802
        }else
803
            dmin_sum+= dmin4;
804

    
805
        if(s->quarter_sample){
806
            mx4_sum+= mx4/2;
807
            my4_sum+= my4/2;
808
        }else{
809
            mx4_sum+= mx4;
810
            my4_sum+= my4;
811
        }
812

    
813
        s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
814
        s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
815

    
816
        if(mx4 != mx || my4 != my) same=0;
817
    }
818

    
819
    if(same)
820
        return INT_MAX;
821

    
822
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
823
        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);
824
    }
825

    
826
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
827
        int dxy;
828
        int mx, my;
829
        int offset;
830

    
831
        mx= ff_h263_round_chroma(mx4_sum);
832
        my= ff_h263_round_chroma(my4_sum);
833
        dxy = ((my & 1) << 1) | (mx & 1);
834

    
835
        offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
836

    
837
        if(s->no_rounding){
838
            s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
839
            s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
840
        }else{
841
            s->dsp.put_pixels_tab       [1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
842
            s->dsp.put_pixels_tab       [1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
843
        }
844

    
845
        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);
846
        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);
847
    }
848

    
849
    c->pred_x= mx;
850
    c->pred_y= my;
851

    
852
    switch(c->avctx->mb_cmp&0xFF){
853
    /*case FF_CMP_SSE:
854
        return dmin_sum+ 32*s->qscale*s->qscale;*/
855
    case FF_CMP_RD:
856
        return dmin_sum;
857
    default:
858
        return dmin_sum+ 11*c->mb_penalty_factor;
859
    }
860
}
861

    
862
static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
863
    MotionEstContext * const c= &s->me;
864

    
865
    c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
866
    c->src[1][0] = c->src[0][0] + s->linesize;
867
    if(c->flags & FLAG_CHROMA){
868
        c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
869
        c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
870
        c->src[1][1] = c->src[0][1] + s->uvlinesize;
871
        c->src[1][2] = c->src[0][2] + s->uvlinesize;
872
    }
873
}
874

    
875
static int interlaced_search(MpegEncContext *s, int ref_index,
876
                             int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
877
{
878
    MotionEstContext * const c= &s->me;
879
    const int size=0;
880
    const int h=8;
881
    int block;
882
    int P[10][2];
883
    uint8_t * const mv_penalty= c->current_mv_penalty;
884
    int same=1;
885
    const int stride= 2*s->linesize;
886
    int dmin_sum= 0;
887
    const int mot_stride= s->mb_stride;
888
    const int xy= s->mb_x + s->mb_y*mot_stride;
889

    
890
    c->ymin>>=1;
891
    c->ymax>>=1;
892
    c->stride<<=1;
893
    c->uvstride<<=1;
894
    init_interlaced_ref(s, ref_index);
895

    
896
    for(block=0; block<2; block++){
897
        int field_select;
898
        int best_dmin= INT_MAX;
899
        int best_field= -1;
900

    
901
        for(field_select=0; field_select<2; field_select++){
902
            int dmin, mx_i, my_i;
903
            int16_t (*mv_table)[2]= mv_tables[block][field_select];
904

    
905
            if(user_field_select){
906
                assert(field_select==0 || field_select==1);
907
                assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
908
                if(field_select_tables[block][xy] != field_select)
909
                    continue;
910
            }
911

    
912
            P_LEFT[0] = mv_table[xy - 1][0];
913
            P_LEFT[1] = mv_table[xy - 1][1];
914
            if(P_LEFT[0]       > (c->xmax<<1)) P_LEFT[0]       = (c->xmax<<1);
915

    
916
            c->pred_x= P_LEFT[0];
917
            c->pred_y= P_LEFT[1];
918

    
919
            if(!s->first_slice_line){
920
                P_TOP[0]      = mv_table[xy - mot_stride][0];
921
                P_TOP[1]      = mv_table[xy - mot_stride][1];
922
                P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
923
                P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
924
                if(P_TOP[1]      > (c->ymax<<1)) P_TOP[1]     = (c->ymax<<1);
925
                if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
926
                if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
927
                if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
928

    
929
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
930
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
931
            }
932
            P_MV1[0]= mx; //FIXME not correct if block != field_select
933
            P_MV1[1]= my / 2;
934

    
935
            dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
936

    
937
            dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
938

    
939
            mv_table[xy][0]= mx_i;
940
            mv_table[xy][1]= my_i;
941

    
942
            if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
943
                int dxy;
944

    
945
                //FIXME chroma ME
946
                uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
947
                dxy = ((my_i & 1) << 1) | (mx_i & 1);
948

    
949
                if(s->no_rounding){
950
                    s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref    , stride, h);
951
                }else{
952
                    s->dsp.put_pixels_tab       [size][dxy](c->scratchpad, ref    , stride, h);
953
                }
954
                dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
955
                dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
956
            }else
957
                dmin+= c->mb_penalty_factor; //field_select bits
958

    
959
            dmin += field_select != block; //slightly prefer same field
960

    
961
            if(dmin < best_dmin){
962
                best_dmin= dmin;
963
                best_field= field_select;
964
            }
965
        }
966
        {
967
            int16_t (*mv_table)[2]= mv_tables[block][best_field];
968

    
969
            if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
970
            if(mv_table[xy][1]&1) same=0;
971
            if(mv_table[xy][1]*2 != my) same=0;
972
            if(best_field != block) same=0;
973
        }
974

    
975
        field_select_tables[block][xy]= best_field;
976
        dmin_sum += best_dmin;
977
    }
978

    
979
    c->ymin<<=1;
980
    c->ymax<<=1;
981
    c->stride>>=1;
982
    c->uvstride>>=1;
983

    
984
    if(same)
985
        return INT_MAX;
986

    
987
    switch(c->avctx->mb_cmp&0xFF){
988
    /*case FF_CMP_SSE:
989
        return dmin_sum+ 32*s->qscale*s->qscale;*/
990
    case FF_CMP_RD:
991
        return dmin_sum;
992
    default:
993
        return dmin_sum+ 11*c->mb_penalty_factor;
994
    }
995
}
996

    
997
static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
998
    int ymax= s->me.ymax>>interlaced;
999
    int ymin= s->me.ymin>>interlaced;
1000

    
1001
    if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
1002
    if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
1003
    if(mv[1] <       ymin) mv[1] =       ymin;
1004
    if(mv[1] >       ymax) mv[1] =       ymax;
1005
}
1006

    
1007
static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1008
    MotionEstContext * const c= &s->me;
1009
    Picture *p= s->current_picture_ptr;
1010
    int mb_xy= mb_x + mb_y*s->mb_stride;
1011
    int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1012
    int mb_type= s->current_picture.mb_type[mb_xy];
1013
    int flags= c->flags;
1014
    int shift= (flags&FLAG_QPEL) + 1;
1015
    int mask= (1<<shift)-1;
1016
    int x, y, i;
1017
    int d=0;
1018
    me_cmp_func cmpf= s->dsp.sse[0];
1019
    me_cmp_func chroma_cmpf= s->dsp.sse[1];
1020

    
1021
    if(p_type && USES_LIST(mb_type, 1)){
1022
        av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1023
        return INT_MAX/2;
1024
    }
1025
    assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1026

    
1027
    for(i=0; i<4; i++){
1028
        int xy= s->block_index[i];
1029
        clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1030
        clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1031
    }
1032

    
1033
    if(IS_INTERLACED(mb_type)){
1034
        int xy2= xy  + s->b8_stride;
1035
        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1036
        c->stride<<=1;
1037
        c->uvstride<<=1;
1038

    
1039
        if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1040
            av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1041
            return INT_MAX/2;
1042
        }
1043

    
1044
        if(USES_LIST(mb_type, 0)){
1045
            int field_select0= p->ref_index[0][xy ];
1046
            int field_select1= p->ref_index[0][xy2];
1047
            assert(field_select0==0 ||field_select0==1);
1048
            assert(field_select1==0 ||field_select1==1);
1049
            init_interlaced_ref(s, 0);
1050

    
1051
            if(p_type){
1052
                s->p_field_select_table[0][mb_xy]= field_select0;
1053
                s->p_field_select_table[1][mb_xy]= field_select1;
1054
                *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1055
                *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1056
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1057
            }else{
1058
                s->b_field_select_table[0][0][mb_xy]= field_select0;
1059
                s->b_field_select_table[0][1][mb_xy]= field_select1;
1060
                *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1061
                *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1062
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1063
            }
1064

    
1065
            x= p->motion_val[0][xy ][0];
1066
            y= p->motion_val[0][xy ][1];
1067
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1068
            x= p->motion_val[0][xy2][0];
1069
            y= p->motion_val[0][xy2][1];
1070
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1071
        }
1072
        if(USES_LIST(mb_type, 1)){
1073
            int field_select0= p->ref_index[1][xy ];
1074
            int field_select1= p->ref_index[1][xy2];
1075
            assert(field_select0==0 ||field_select0==1);
1076
            assert(field_select1==0 ||field_select1==1);
1077
            init_interlaced_ref(s, 2);
1078

    
1079
            s->b_field_select_table[1][0][mb_xy]= field_select0;
1080
            s->b_field_select_table[1][1][mb_xy]= field_select1;
1081
            *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1082
            *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1083
            if(USES_LIST(mb_type, 0)){
1084
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1085
            }else{
1086
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1087
            }
1088

    
1089
            x= p->motion_val[1][xy ][0];
1090
            y= p->motion_val[1][xy ][1];
1091
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1092
            x= p->motion_val[1][xy2][0];
1093
            y= p->motion_val[1][xy2][1];
1094
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1095
            //FIXME bidir scores
1096
        }
1097
        c->stride>>=1;
1098
        c->uvstride>>=1;
1099
    }else if(IS_8X8(mb_type)){
1100
        if(!(s->flags & CODEC_FLAG_4MV)){
1101
            av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1102
            return INT_MAX/2;
1103
        }
1104
        cmpf= s->dsp.sse[1];
1105
        chroma_cmpf= s->dsp.sse[1];
1106
        init_mv4_ref(c);
1107
        for(i=0; i<4; i++){
1108
            xy= s->block_index[i];
1109
            x= p->motion_val[0][xy][0];
1110
            y= p->motion_val[0][xy][1];
1111
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1112
        }
1113
        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1114
    }else{
1115
        if(USES_LIST(mb_type, 0)){
1116
            if(p_type){
1117
                *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1118
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1119
            }else if(USES_LIST(mb_type, 1)){
1120
                *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1121
                *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1122
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1123
            }else{
1124
                *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1125
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1126
            }
1127
            x= p->motion_val[0][xy][0];
1128
            y= p->motion_val[0][xy][1];
1129
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1130
        }else if(USES_LIST(mb_type, 1)){
1131
            *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1132
            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1133

    
1134
            x= p->motion_val[1][xy][0];
1135
            y= p->motion_val[1][xy][1];
1136
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1137
        }else
1138
            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1139
    }
1140
    return d;
1141
}
1142

    
1143
void ff_estimate_p_frame_motion(MpegEncContext * s,
1144
                                int mb_x, int mb_y)
1145
{
1146
    MotionEstContext * const c= &s->me;
1147
    uint8_t *pix, *ppix;
1148
    int sum, varc, vard, mx, my, dmin;
1149
    int P[10][2];
1150
    const int shift= 1+s->quarter_sample;
1151
    int mb_type=0;
1152
    Picture * const pic= &s->current_picture;
1153

    
1154
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1155

    
1156
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1157
    assert(s->linesize == c->stride);
1158
    assert(s->uvlinesize == c->uvstride);
1159

    
1160
    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1161
    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1162
    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1163
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1164

    
1165
    get_limits(s, 16*mb_x, 16*mb_y);
1166
    c->skip=0;
1167

    
1168
    /* intra / predictive decision */
1169
    pix = c->src[0][0];
1170
    sum = s->dsp.pix_sum(pix, s->linesize);
1171
    varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1172

    
1173
    pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1174
    pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1175
    c->mb_var_sum_temp += (varc+128)>>8;
1176

    
1177
    if(c->avctx->me_threshold){
1178
        vard= check_input_motion(s, mb_x, mb_y, 1);
1179

    
1180
        if((vard+128)>>8 < c->avctx->me_threshold){
1181
            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1182
            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1183
            pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1184
            c->mc_mb_var_sum_temp += (vard+128)>>8;
1185
            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1186
            return;
1187
        }
1188
        if((vard+128)>>8 < c->avctx->mb_threshold)
1189
            mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1190
    }
1191

    
1192
    switch(s->me_method) {
1193
    case ME_ZERO:
1194
    default:
1195
        no_motion_search(s, &mx, &my);
1196
        mx-= mb_x*16;
1197
        my-= mb_y*16;
1198
        dmin = 0;
1199
        break;
1200
#if 0
1201
    case ME_FULL:
1202
        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1203
        mx-= mb_x*16;
1204
        my-= mb_y*16;
1205
        break;
1206
    case ME_LOG:
1207
        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1208
        mx-= mb_x*16;
1209
        my-= mb_y*16;
1210
        break;
1211
    case ME_PHODS:
1212
        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1213
        mx-= mb_x*16;
1214
        my-= mb_y*16;
1215
        break;
1216
#endif
1217
    case ME_X1:
1218
    case ME_EPZS:
1219
       {
1220
            const int mot_stride = s->b8_stride;
1221
            const int mot_xy = s->block_index[0];
1222

    
1223
            P_LEFT[0]       = s->current_picture.motion_val[0][mot_xy - 1][0];
1224
            P_LEFT[1]       = s->current_picture.motion_val[0][mot_xy - 1][1];
1225

    
1226
            if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
1227

    
1228
            if(!s->first_slice_line) {
1229
                P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][0];
1230
                P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][1];
1231
                P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1232
                P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1233
                if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
1234
                if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1235
                if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1236

    
1237
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1238
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1239

    
1240
                if(s->out_format == FMT_H263){
1241
                    c->pred_x = P_MEDIAN[0];
1242
                    c->pred_y = P_MEDIAN[1];
1243
                }else { /* mpeg1 at least */
1244
                    c->pred_x= P_LEFT[0];
1245
                    c->pred_y= P_LEFT[1];
1246
                }
1247
            }else{
1248
                c->pred_x= P_LEFT[0];
1249
                c->pred_y= P_LEFT[1];
1250
            }
1251

    
1252
        }
1253
        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1254

    
1255
        break;
1256
    }
1257

    
1258
    /* At this point (mx,my) are full-pell and the relative displacement */
1259
    ppix = c->ref[0][0] + (my * s->linesize) + mx;
1260

    
1261
    vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1262

    
1263
    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1264
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1265
    c->mc_mb_var_sum_temp += (vard+128)>>8;
1266

    
1267
#if 0
1268
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1269
           varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1270
#endif
1271
    if(mb_type){
1272
        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1273
        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1274
        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1275

    
1276
        if(mb_type == CANDIDATE_MB_TYPE_INTER){
1277
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1278
            set_p_mv_tables(s, mx, my, 1);
1279
        }else{
1280
            mx <<=shift;
1281
            my <<=shift;
1282
        }
1283
        if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1284
            h263_mv4_search(s, mx, my, shift);
1285

    
1286
            set_p_mv_tables(s, mx, my, 0);
1287
        }
1288
        if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1289
            interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1290
        }
1291
    }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1292
        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1293
        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1294
        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1295

    
1296
        if (vard*2 + 200*256 > varc)
1297
            mb_type|= CANDIDATE_MB_TYPE_INTRA;
1298
        if (varc*2 + 200*256 > vard || s->qscale > 24){
1299
//        if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1300
            mb_type|= CANDIDATE_MB_TYPE_INTER;
1301
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1302
            if(s->flags&CODEC_FLAG_MV0)
1303
                if(mx || my)
1304
                    mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1305
        }else{
1306
            mx <<=shift;
1307
            my <<=shift;
1308
        }
1309
        if((s->flags&CODEC_FLAG_4MV)
1310
           && !c->skip && varc>50<<8 && vard>10<<8){
1311
            if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1312
                mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1313

    
1314
            set_p_mv_tables(s, mx, my, 0);
1315
        }else
1316
            set_p_mv_tables(s, mx, my, 1);
1317
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
1318
           && !c->skip){ //FIXME varc/d checks
1319
            if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1320
                mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1321
        }
1322
    }else{
1323
        int intra_score, i;
1324
        mb_type= CANDIDATE_MB_TYPE_INTER;
1325

    
1326
        dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1327
        if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1328
            dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1329

    
1330
        if((s->flags&CODEC_FLAG_4MV)
1331
           && !c->skip && varc>50<<8 && vard>10<<8){
1332
            int dmin4= h263_mv4_search(s, mx, my, shift);
1333
            if(dmin4 < dmin){
1334
                mb_type= CANDIDATE_MB_TYPE_INTER4V;
1335
                dmin=dmin4;
1336
            }
1337
        }
1338
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
1339
           && !c->skip){ //FIXME varc/d checks
1340
            int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1341
            if(dmin_i < dmin){
1342
                mb_type = CANDIDATE_MB_TYPE_INTER_I;
1343
                dmin= dmin_i;
1344
            }
1345
        }
1346

    
1347
//        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1348
        set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1349

    
1350
        /* get intra luma score */
1351
        if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1352
            intra_score= varc - 500;
1353
        }else{
1354
            int mean= (sum+128)>>8;
1355
            mean*= 0x01010101;
1356

    
1357
            for(i=0; i<16; i++){
1358
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1359
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1360
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1361
                *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1362
            }
1363

    
1364
            intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1365
        }
1366
#if 0 //FIXME
1367
        /* get chroma score */
1368
        if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1369
            for(i=1; i<3; i++){
1370
                uint8_t *dest_c;
1371
                int mean;
1372

1373
                if(s->out_format == FMT_H263){
1374
                    mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1375
                }else{
1376
                    mean= (s->last_dc[i] + 4)>>3;
1377
                }
1378
                dest_c = s->new_picture.data[i] + (mb_y * 8  * (s->uvlinesize)) + mb_x * 8;
1379

1380
                mean*= 0x01010101;
1381
                for(i=0; i<8; i++){
1382
                    *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1383
                    *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1384
                }
1385

1386
                intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1387
            }
1388
        }
1389
#endif
1390
        intra_score += c->mb_penalty_factor*16;
1391

    
1392
        if(intra_score < dmin){
1393
            mb_type= CANDIDATE_MB_TYPE_INTRA;
1394
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1395
        }else
1396
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1397

    
1398
        {
1399
            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1400
            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1401
            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1402
        }
1403
    }
1404

    
1405
    s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1406
}
1407

    
1408
int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1409
                                    int mb_x, int mb_y)
1410
{
1411
    MotionEstContext * const c= &s->me;
1412
    int mx, my, dmin;
1413
    int P[10][2];
1414
    const int shift= 1+s->quarter_sample;
1415
    const int xy= mb_x + mb_y*s->mb_stride;
1416
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1417

    
1418
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1419

    
1420
    c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1421
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1422

    
1423
    get_limits(s, 16*mb_x, 16*mb_y);
1424
    c->skip=0;
1425

    
1426
    P_LEFT[0]       = s->p_mv_table[xy + 1][0];
1427
    P_LEFT[1]       = s->p_mv_table[xy + 1][1];
1428

    
1429
    if(P_LEFT[0]       < (c->xmin<<shift)) P_LEFT[0]       = (c->xmin<<shift);
1430

    
1431
    /* special case for first line */
1432
    if (s->first_slice_line) {
1433
        c->pred_x= P_LEFT[0];
1434
        c->pred_y= P_LEFT[1];
1435
        P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1436
        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1437
    } else {
1438
        P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
1439
        P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
1440
        P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1441
        P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1442
        if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);
1443
        if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1444
        if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1445

    
1446
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1447
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1448

    
1449
        c->pred_x = P_MEDIAN[0];
1450
        c->pred_y = P_MEDIAN[1];
1451
    }
1452

    
1453
    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1454

    
1455
    s->p_mv_table[xy][0] = mx<<shift;
1456
    s->p_mv_table[xy][1] = my<<shift;
1457

    
1458
    return dmin;
1459
}
1460

    
1461
static int ff_estimate_motion_b(MpegEncContext * s,
1462
                       int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1463
{
1464
    MotionEstContext * const c= &s->me;
1465
    int mx, my, dmin;
1466
    int P[10][2];
1467
    const int shift= 1+s->quarter_sample;
1468
    const int mot_stride = s->mb_stride;
1469
    const int mot_xy = mb_y*mot_stride + mb_x;
1470
    uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1471
    int mv_scale;
1472

    
1473
    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1474
    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1475
    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1476
    c->current_mv_penalty= mv_penalty;
1477

    
1478
    get_limits(s, 16*mb_x, 16*mb_y);
1479

    
1480
    switch(s->me_method) {
1481
    case ME_ZERO:
1482
    default:
1483
        no_motion_search(s, &mx, &my);
1484
        dmin = 0;
1485
        mx-= mb_x*16;
1486
        my-= mb_y*16;
1487
        break;
1488
#if 0
1489
    case ME_FULL:
1490
        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1491
        mx-= mb_x*16;
1492
        my-= mb_y*16;
1493
        break;
1494
    case ME_LOG:
1495
        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1496
        mx-= mb_x*16;
1497
        my-= mb_y*16;
1498
        break;
1499
    case ME_PHODS:
1500
        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1501
        mx-= mb_x*16;
1502
        my-= mb_y*16;
1503
        break;
1504
#endif
1505
    case ME_X1:
1506
    case ME_EPZS:
1507
       {
1508
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
1509
            P_LEFT[1]        = mv_table[mot_xy - 1][1];
1510

    
1511
            if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
1512

    
1513
            /* special case for first line */
1514
            if (!s->first_slice_line) {
1515
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
1516
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
1517
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
1518
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
1519
                if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1520
                if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1521
                if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1522

    
1523
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1524
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1525
            }
1526
            c->pred_x= P_LEFT[0];
1527
            c->pred_y= P_LEFT[1];
1528
        }
1529

    
1530
        if(mv_table == s->b_forw_mv_table){
1531
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1532
        }else{
1533
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1534
        }
1535

    
1536
        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1537

    
1538
        break;
1539
    }
1540

    
1541
    dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1542

    
1543
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1544
        dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1545

    
1546
//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1547
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1548
    mv_table[mot_xy][0]= mx;
1549
    mv_table[mot_xy][1]= my;
1550

    
1551
    return dmin;
1552
}
1553

    
1554
static inline int check_bidir_mv(MpegEncContext * s,
1555
                   int motion_fx, int motion_fy,
1556
                   int motion_bx, int motion_by,
1557
                   int pred_fx, int pred_fy,
1558
                   int pred_bx, int pred_by,
1559
                   int size, int h)
1560
{
1561
    //FIXME optimize?
1562
    //FIXME better f_code prediction (max mv & distance)
1563
    //FIXME pointers
1564
    MotionEstContext * const c= &s->me;
1565
    uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1566
    uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1567
    int stride= c->stride;
1568
    uint8_t *dest_y = c->scratchpad;
1569
    uint8_t *ptr;
1570
    int dxy;
1571
    int src_x, src_y;
1572
    int fbmin;
1573
    uint8_t **src_data= c->src[0];
1574
    uint8_t **ref_data= c->ref[0];
1575
    uint8_t **ref2_data= c->ref[2];
1576

    
1577
    if(s->quarter_sample){
1578
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1579
        src_x = motion_fx >> 2;
1580
        src_y = motion_fy >> 2;
1581

    
1582
        ptr = ref_data[0] + (src_y * stride) + src_x;
1583
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , stride);
1584

    
1585
        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1586
        src_x = motion_bx >> 2;
1587
        src_y = motion_by >> 2;
1588

    
1589
        ptr = ref2_data[0] + (src_y * stride) + src_x;
1590
        s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);
1591
    }else{
1592
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1593
        src_x = motion_fx >> 1;
1594
        src_y = motion_fy >> 1;
1595

    
1596
        ptr = ref_data[0] + (src_y * stride) + src_x;
1597
        s->dsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
1598

    
1599
        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1600
        src_x = motion_bx >> 1;
1601
        src_y = motion_by >> 1;
1602

    
1603
        ptr = ref2_data[0] + (src_y * stride) + src_x;
1604
        s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
1605
    }
1606

    
1607
    fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1608
           +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1609
           + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1610

    
1611
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1612
    }
1613
    //FIXME CHROMA !!!
1614

    
1615
    return fbmin;
1616
}
1617

    
1618
/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1619
static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1620
{
1621
    MotionEstContext * const c= &s->me;
1622
    const int mot_stride = s->mb_stride;
1623
    const int xy = mb_y *mot_stride + mb_x;
1624
    int fbmin;
1625
    int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1626
    int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1627
    int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1628
    int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1629
    int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1630
    int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1631
    int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1632
    int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1633
    const int flags= c->sub_flags;
1634
    const int qpel= flags&FLAG_QPEL;
1635
    const int shift= 1+qpel;
1636
    const int xmin= c->xmin<<shift;
1637
    const int ymin= c->ymin<<shift;
1638
    const int xmax= c->xmax<<shift;
1639
    const int ymax= c->ymax<<shift;
1640
    uint8_t map[8][8][8][8];
1641

    
1642
    memset(map,0,sizeof(map));
1643
#define BIDIR_MAP(fx,fy,bx,by) \
1644
    map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1645
    BIDIR_MAP(0,0,0,0) = 1;
1646

    
1647
    fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1648
                          motion_bx, motion_by,
1649
                          pred_fx, pred_fy,
1650
                          pred_bx, pred_by,
1651
                          0, 16);
1652

    
1653
    if(s->avctx->bidir_refine){
1654
        int score, end;
1655
#define CHECK_BIDIR(fx,fy,bx,by)\
1656
    if( !BIDIR_MAP(fx,fy,bx,by)\
1657
       &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1658
       &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1659
        BIDIR_MAP(fx,fy,bx,by) = 1;\
1660
        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);\
1661
        if(score < fbmin){\
1662
            fbmin= score;\
1663
            motion_fx+=fx;\
1664
            motion_fy+=fy;\
1665
            motion_bx+=bx;\
1666
            motion_by+=by;\
1667
            end=0;\
1668
        }\
1669
    }
1670
#define CHECK_BIDIR2(a,b,c,d)\
1671
CHECK_BIDIR(a,b,c,d)\
1672
CHECK_BIDIR(-(a),-(b),-(c),-(d))
1673

    
1674
#define CHECK_BIDIRR(a,b,c,d)\
1675
CHECK_BIDIR2(a,b,c,d)\
1676
CHECK_BIDIR2(b,c,d,a)\
1677
CHECK_BIDIR2(c,d,a,b)\
1678
CHECK_BIDIR2(d,a,b,c)
1679

    
1680
        do{
1681
            end=1;
1682

    
1683
            CHECK_BIDIRR( 0, 0, 0, 1)
1684
            if(s->avctx->bidir_refine > 1){
1685
                CHECK_BIDIRR( 0, 0, 1, 1)
1686
                CHECK_BIDIR2( 0, 1, 0, 1)
1687
                CHECK_BIDIR2( 1, 0, 1, 0)
1688
                CHECK_BIDIRR( 0, 0,-1, 1)
1689
                CHECK_BIDIR2( 0,-1, 0, 1)
1690
                CHECK_BIDIR2(-1, 0, 1, 0)
1691
                if(s->avctx->bidir_refine > 2){
1692
                    CHECK_BIDIRR( 0, 1, 1, 1)
1693
                    CHECK_BIDIRR( 0,-1, 1, 1)
1694
                    CHECK_BIDIRR( 0, 1,-1, 1)
1695
                    CHECK_BIDIRR( 0, 1, 1,-1)
1696
                    if(s->avctx->bidir_refine > 3){
1697
                        CHECK_BIDIR2( 1, 1, 1, 1)
1698
                        CHECK_BIDIRR( 1, 1, 1,-1)
1699
                        CHECK_BIDIR2( 1, 1,-1,-1)
1700
                        CHECK_BIDIR2( 1,-1,-1, 1)
1701
                        CHECK_BIDIR2( 1,-1, 1,-1)
1702
                    }
1703
                }
1704
            }
1705
        }while(!end);
1706
    }
1707

    
1708
    s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1709
    s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1710
    s->b_bidir_back_mv_table[xy][0]= motion_bx;
1711
    s->b_bidir_back_mv_table[xy][1]= motion_by;
1712

    
1713
    return fbmin;
1714
}
1715

    
1716
static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1717
{
1718
    MotionEstContext * const c= &s->me;
1719
    int P[10][2];
1720
    const int mot_stride = s->mb_stride;
1721
    const int mot_xy = mb_y*mot_stride + mb_x;
1722
    const int shift= 1+s->quarter_sample;
1723
    int dmin, i;
1724
    const int time_pp= s->pp_time;
1725
    const int time_pb= s->pb_time;
1726
    int mx, my, xmin, xmax, ymin, ymax;
1727
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
1728

    
1729
    c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1730
    ymin= xmin=(-32)>>shift;
1731
    ymax= xmax=   31>>shift;
1732

    
1733
    if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1734
        s->mv_type= MV_TYPE_8X8;
1735
    }else{
1736
        s->mv_type= MV_TYPE_16X16;
1737
    }
1738

    
1739
    for(i=0; i<4; i++){
1740
        int index= s->block_index[i];
1741
        int min, max;
1742

    
1743
        c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1744
        c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1745
        c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1746
        c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1747
//        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1748
//        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1749

    
1750
        max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1751
        min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1752
        max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1753
        min+= 16*mb_x - 1;
1754
        xmax= FFMIN(xmax, s->width - max);
1755
        xmin= FFMAX(xmin, - 16     - min);
1756

    
1757
        max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1758
        min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1759
        max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1760
        min+= 16*mb_y - 1;
1761
        ymax= FFMIN(ymax, s->height - max);
1762
        ymin= FFMAX(ymin, - 16      - min);
1763

    
1764
        if(s->mv_type == MV_TYPE_16X16) break;
1765
    }
1766

    
1767
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1768

    
1769
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1770
        s->b_direct_mv_table[mot_xy][0]= 0;
1771
        s->b_direct_mv_table[mot_xy][1]= 0;
1772

    
1773
        return 256*256*256*64;
1774
    }
1775

    
1776
    c->xmin= xmin;
1777
    c->ymin= ymin;
1778
    c->xmax= xmax;
1779
    c->ymax= ymax;
1780
    c->flags     |= FLAG_DIRECT;
1781
    c->sub_flags |= FLAG_DIRECT;
1782
    c->pred_x=0;
1783
    c->pred_y=0;
1784

    
1785
    P_LEFT[0]        = clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1786
    P_LEFT[1]        = clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1787

    
1788
    /* special case for first line */
1789
    if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its clipped
1790
        P_TOP[0]      = clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
1791
        P_TOP[1]      = clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
1792
        P_TOPRIGHT[0] = clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
1793
        P_TOPRIGHT[1] = clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
1794

    
1795
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1796
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1797
    }
1798

    
1799
    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1800
    if(c->sub_flags&FLAG_QPEL)
1801
        dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1802
    else
1803
        dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1804

    
1805
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1806
        dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1807

    
1808
    get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1809

    
1810
    s->b_direct_mv_table[mot_xy][0]= mx;
1811
    s->b_direct_mv_table[mot_xy][1]= my;
1812
    c->flags     &= ~FLAG_DIRECT;
1813
    c->sub_flags &= ~FLAG_DIRECT;
1814

    
1815
    return dmin;
1816
}
1817

    
1818
void ff_estimate_b_frame_motion(MpegEncContext * s,
1819
                             int mb_x, int mb_y)
1820
{
1821
    MotionEstContext * const c= &s->me;
1822
    const int penalty_factor= c->mb_penalty_factor;
1823
    int fmin, bmin, dmin, fbmin, bimin, fimin;
1824
    int type=0;
1825
    const int xy = mb_y*s->mb_stride + mb_x;
1826
    init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1827

    
1828
    get_limits(s, 16*mb_x, 16*mb_y);
1829

    
1830
    c->skip=0;
1831
    if(c->avctx->me_threshold){
1832
        int vard= check_input_motion(s, mb_x, mb_y, 0);
1833

    
1834
        if((vard+128)>>8 < c->avctx->me_threshold){
1835
//            pix = c->src[0][0];
1836
//            sum = s->dsp.pix_sum(pix, s->linesize);
1837
//            varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1838

    
1839
//            pic->mb_var   [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1840
             s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1841
/*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1842
            c->mb_var_sum_temp    += (varc+128)>>8;*/
1843
            c->mc_mb_var_sum_temp += (vard+128)>>8;
1844
/*            if (vard <= 64<<8 || vard < varc) {
1845
                c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1846
            }else{
1847
                c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1848
            }*/
1849
            return;
1850
        }
1851
        if((vard+128)>>8 < c->avctx->mb_threshold){
1852
            type= s->mb_type[mb_y*s->mb_stride + mb_x];
1853
            if(type == CANDIDATE_MB_TYPE_DIRECT){
1854
                direct_search(s, mb_x, mb_y);
1855
            }
1856
            if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1857
                c->skip=0;
1858
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1859
            }
1860
            if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1861
                c->skip=0;
1862
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1863
            }
1864
            if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1865
                c->skip=0;
1866
                c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1867
                interlaced_search(s, 0,
1868
                                        s->b_field_mv_table[0], s->b_field_select_table[0],
1869
                                        s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1870
            }
1871
            if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1872
                c->skip=0;
1873
                c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1874
                interlaced_search(s, 2,
1875
                                        s->b_field_mv_table[1], s->b_field_select_table[1],
1876
                                        s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1877
            }
1878
            return;
1879
        }
1880
    }
1881

    
1882
    if (s->codec_id == CODEC_ID_MPEG4)
1883
        dmin= direct_search(s, mb_x, mb_y);
1884
    else
1885
        dmin= INT_MAX;
1886
//FIXME penalty stuff for non mpeg4
1887
    c->skip=0;
1888
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1889

    
1890
    c->skip=0;
1891
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1892
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1893

    
1894
    c->skip=0;
1895
    fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1896
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1897

    
1898
    if(s->flags & CODEC_FLAG_INTERLACED_ME){
1899
//FIXME mb type penalty
1900
        c->skip=0;
1901
        c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1902
        fimin= interlaced_search(s, 0,
1903
                                 s->b_field_mv_table[0], s->b_field_select_table[0],
1904
                                 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1905
        c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1906
        bimin= interlaced_search(s, 2,
1907
                                 s->b_field_mv_table[1], s->b_field_select_table[1],
1908
                                 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1909
    }else
1910
        fimin= bimin= INT_MAX;
1911

    
1912
    {
1913
        int score= fmin;
1914
        type = CANDIDATE_MB_TYPE_FORWARD;
1915

    
1916
        if (dmin <= score){
1917
            score = dmin;
1918
            type = CANDIDATE_MB_TYPE_DIRECT;
1919
        }
1920
        if(bmin<score){
1921
            score=bmin;
1922
            type= CANDIDATE_MB_TYPE_BACKWARD;
1923
        }
1924
        if(fbmin<score){
1925
            score=fbmin;
1926
            type= CANDIDATE_MB_TYPE_BIDIR;
1927
        }
1928
        if(fimin<score){
1929
            score=fimin;
1930
            type= CANDIDATE_MB_TYPE_FORWARD_I;
1931
        }
1932
        if(bimin<score){
1933
            score=bimin;
1934
            type= CANDIDATE_MB_TYPE_BACKWARD_I;
1935
        }
1936

    
1937
        score= ((unsigned)(score*score + 128*256))>>16;
1938
        c->mc_mb_var_sum_temp += score;
1939
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1940
    }
1941

    
1942
    if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1943
        type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1944
        if(fimin < INT_MAX)
1945
            type |= CANDIDATE_MB_TYPE_FORWARD_I;
1946
        if(bimin < INT_MAX)
1947
            type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1948
        if(fimin < INT_MAX && bimin < INT_MAX){
1949
            type |= CANDIDATE_MB_TYPE_BIDIR_I;
1950
        }
1951
         //FIXME something smarter
1952
        if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1953
#if 0
1954
        if(s->out_format == FMT_MPEG1)
1955
            type |= CANDIDATE_MB_TYPE_INTRA;
1956
#endif
1957
    }
1958

    
1959
    s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1960
}
1961

    
1962
/* find best f_code for ME which do unlimited searches */
1963
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1964
{
1965
    if(s->me_method>=ME_EPZS){
1966
        int score[8];
1967
        int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
1968
        uint8_t * fcode_tab= s->fcode_tab;
1969
        int best_fcode=-1;
1970
        int best_score=-10000000;
1971

    
1972
        if(s->msmpeg4_version)
1973
            range= FFMIN(range, 16);
1974
        else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
1975
            range= FFMIN(range, 256);
1976

    
1977
        for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
1978

    
1979
        for(y=0; y<s->mb_height; y++){
1980
            int x;
1981
            int xy= y*s->mb_stride;
1982
            for(x=0; x<s->mb_width; x++){
1983
                if(s->mb_type[xy] & type){
1984
                    int mx= mv_table[xy][0];
1985
                    int my= mv_table[xy][1];
1986
                    int fcode= FFMAX(fcode_tab[mx + MAX_MV],
1987
                                     fcode_tab[my + MAX_MV]);
1988
                    int j;
1989

    
1990
                        if(mx >= range || mx < -range ||
1991
                           my >= range || my < -range)
1992
                            continue;
1993

    
1994
                    for(j=0; j<fcode && j<8; j++){
1995
                        if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
1996
                            score[j]-= 170;
1997
                    }
1998
                }
1999
                xy++;
2000
            }
2001
        }
2002

    
2003
        for(i=1; i<8; i++){
2004
            if(score[i] > best_score){
2005
                best_score= score[i];
2006
                best_fcode= i;
2007
            }
2008
//            printf("%d %d\n", i, score[i]);
2009
        }
2010

    
2011
//    printf("fcode: %d type: %d\n", i, s->pict_type);
2012
        return best_fcode;
2013
/*        for(i=0; i<=MAX_FCODE; i++){
2014
            printf("%d ", mv_num[i]);
2015
        }
2016
        printf("\n");*/
2017
    }else{
2018
        return 1;
2019
    }
2020
}
2021

    
2022
void ff_fix_long_p_mvs(MpegEncContext * s)
2023
{
2024
    MotionEstContext * const c= &s->me;
2025
    const int f_code= s->f_code;
2026
    int y, range;
2027
    assert(s->pict_type==P_TYPE);
2028

    
2029
    range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2030

    
2031
    assert(range <= 16 || !s->msmpeg4_version);
2032
    assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
2033

    
2034
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2035

    
2036
//printf("%d no:%d %d//\n", clip, noclip, f_code);
2037
    if(s->flags&CODEC_FLAG_4MV){
2038
        const int wrap= s->b8_stride;
2039

    
2040
        /* clip / convert to intra 8x8 type MVs */
2041
        for(y=0; y<s->mb_height; y++){
2042
            int xy= y*2*wrap;
2043
            int i= y*s->mb_stride;
2044
            int x;
2045

    
2046
            for(x=0; x<s->mb_width; x++){
2047
                if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
2048
                    int block;
2049
                    for(block=0; block<4; block++){
2050
                        int off= (block& 1) + (block>>1)*wrap;
2051
                        int mx= s->current_picture.motion_val[0][ xy + off ][0];
2052
                        int my= s->current_picture.motion_val[0][ xy + off ][1];
2053

    
2054
                        if(   mx >=range || mx <-range
2055
                           || my >=range || my <-range){
2056
                            s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
2057
                            s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
2058
                            s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2059
                        }
2060
                    }
2061
                }
2062
                xy+=2;
2063
                i++;
2064
            }
2065
        }
2066
    }
2067
}
2068

    
2069
/**
2070
 *
2071
 * @param truncate 1 for truncation, 0 for using intra
2072
 */
2073
void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
2074
                     int16_t (*mv_table)[2], int f_code, int type, int truncate)
2075
{
2076
    MotionEstContext * const c= &s->me;
2077
    int y, h_range, v_range;
2078

    
2079
    // RAL: 8 in MPEG-1, 16 in MPEG-4
2080
    int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2081

    
2082
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2083

    
2084
    h_range= range;
2085
    v_range= field_select_table ? range>>1 : range;
2086

    
2087
    /* clip / convert to intra 16x16 type MVs */
2088
    for(y=0; y<s->mb_height; y++){
2089
        int x;
2090
        int xy= y*s->mb_stride;
2091
        for(x=0; x<s->mb_width; x++){
2092
            if (s->mb_type[xy] & type){    // RAL: "type" test added...
2093
                if(field_select_table==NULL || field_select_table[xy] == field_select){
2094
                    if(   mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2095
                       || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2096

    
2097
                        if(truncate){
2098
                            if     (mv_table[xy][0] > h_range-1) mv_table[xy][0]=  h_range-1;
2099
                            else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2100
                            if     (mv_table[xy][1] > v_range-1) mv_table[xy][1]=  v_range-1;
2101
                            else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2102
                        }else{
2103
                            s->mb_type[xy] &= ~type;
2104
                            s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
2105
                            mv_table[xy][0]=
2106
                            mv_table[xy][1]= 0;
2107
                        }
2108
                    }
2109
                }
2110
            }
2111
            xy++;
2112
        }
2113
    }
2114
}