Statistics
| Branch: | Revision:

ffmpeg / libavcodec / motion_est.c @ bb21f176

History | View | Annotate | Download (76.4 KB)

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

    
226
#include "motion_est_template.c"
227

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

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

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

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

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

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

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

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

    
282
    // 8x8 fullpel search would need a 4x4 chroma compare, which we dont have yet, and even if we had the motion estimation code doesnt expect it
283
    if(s->codec_id != CODEC_ID_SNOW){
284
        if((c->avctx->me_cmp&FF_CMP_CHROMA)/* && !s->dsp.me_cmp[2]*/){
285
            s->dsp.me_cmp[2]= zero_cmp;
286
        }
287
        if((c->avctx->me_sub_cmp&FF_CMP_CHROMA) && !s->dsp.me_sub_cmp[2]){
288
            s->dsp.me_sub_cmp[2]= zero_cmp;
289
        }
290
        c->hpel_put[2][0]= c->hpel_put[2][1]=
291
        c->hpel_put[2][2]= c->hpel_put[2][3]= zero_hpel;
292
    }
293

    
294
    if(s->codec_id == CODEC_ID_H261){
295
        c->sub_motion_search= no_sub_motion_search;
296
    }
297

    
298
    c->temp= c->scratchpad;
299
}
300

    
301
#if 0
302
static int pix_dev(uint8_t * pix, int line_size, int mean)
303
{
304
    int s, i, j;
305

306
    s = 0;
307
    for (i = 0; i < 16; i++) {
308
        for (j = 0; j < 16; j += 8) {
309
            s += FFABS(pix[0]-mean);
310
            s += FFABS(pix[1]-mean);
311
            s += FFABS(pix[2]-mean);
312
            s += FFABS(pix[3]-mean);
313
            s += FFABS(pix[4]-mean);
314
            s += FFABS(pix[5]-mean);
315
            s += FFABS(pix[6]-mean);
316
            s += FFABS(pix[7]-mean);
317
            pix += 8;
318
        }
319
        pix += line_size - 16;
320
    }
321
    return s;
322
}
323
#endif
324

    
325
static inline void no_motion_search(MpegEncContext * s,
326
                                    int *mx_ptr, int *my_ptr)
327
{
328
    *mx_ptr = 16 * s->mb_x;
329
    *my_ptr = 16 * s->mb_y;
330
}
331

    
332
#if 0  /* the use of these functions is inside #if 0 */
333
static int full_motion_search(MpegEncContext * s,
334
                              int *mx_ptr, int *my_ptr, int range,
335
                              int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
336
{
337
    int x1, y1, x2, y2, xx, yy, x, y;
338
    int mx, my, dmin, d;
339
    uint8_t *pix;
340

341
    xx = 16 * s->mb_x;
342
    yy = 16 * s->mb_y;
343
    x1 = xx - range + 1;        /* we loose one pixel to avoid boundary pb with half pixel pred */
344
    if (x1 < xmin)
345
        x1 = xmin;
346
    x2 = xx + range - 1;
347
    if (x2 > xmax)
348
        x2 = xmax;
349
    y1 = yy - range + 1;
350
    if (y1 < ymin)
351
        y1 = ymin;
352
    y2 = yy + range - 1;
353
    if (y2 > ymax)
354
        y2 = ymax;
355
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
356
    dmin = 0x7fffffff;
357
    mx = 0;
358
    my = 0;
359
    for (y = y1; y <= y2; y++) {
360
        for (x = x1; x <= x2; x++) {
361
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x,
362
                             s->linesize, 16);
363
            if (d < dmin ||
364
                (d == dmin &&
365
                 (abs(x - xx) + abs(y - yy)) <
366
                 (abs(mx - xx) + abs(my - yy)))) {
367
                dmin = d;
368
                mx = x;
369
                my = y;
370
            }
371
        }
372
    }
373

374
    *mx_ptr = mx;
375
    *my_ptr = my;
376

377
#if 0
378
    if (*mx_ptr < -(2 * range) || *mx_ptr >= (2 * range) ||
379
        *my_ptr < -(2 * range) || *my_ptr >= (2 * range)) {
380
        av_log(NULL, AV_LOG_ERROR, "error %d %d\n", *mx_ptr, *my_ptr);
381
    }
382
#endif
383
    return dmin;
384
}
385

    
386

    
387
static int log_motion_search(MpegEncContext * s,
388
                             int *mx_ptr, int *my_ptr, int range,
389
                             int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
390
{
391
    int x1, y1, x2, y2, xx, yy, x, y;
392
    int mx, my, dmin, d;
393
    uint8_t *pix;
394

    
395
    xx = s->mb_x << 4;
396
    yy = s->mb_y << 4;
397

    
398
    /* Left limit */
399
    x1 = xx - range;
400
    if (x1 < xmin)
401
        x1 = xmin;
402

    
403
    /* Right limit */
404
    x2 = xx + range;
405
    if (x2 > xmax)
406
        x2 = xmax;
407

    
408
    /* Upper limit */
409
    y1 = yy - range;
410
    if (y1 < ymin)
411
        y1 = ymin;
412

    
413
    /* Lower limit */
414
    y2 = yy + range;
415
    if (y2 > ymax)
416
        y2 = ymax;
417

    
418
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
419
    dmin = 0x7fffffff;
420
    mx = 0;
421
    my = 0;
422

    
423
    do {
424
        for (y = y1; y <= y2; y += range) {
425
            for (x = x1; x <= x2; x += range) {
426
                d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
427
                if (d < dmin || (d == dmin && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
428
                    dmin = d;
429
                    mx = x;
430
                    my = y;
431
                }
432
            }
433
        }
434

    
435
        range = range >> 1;
436

    
437
        x1 = mx - range;
438
        if (x1 < xmin)
439
            x1 = xmin;
440

    
441
        x2 = mx + range;
442
        if (x2 > xmax)
443
            x2 = xmax;
444

    
445
        y1 = my - range;
446
        if (y1 < ymin)
447
            y1 = ymin;
448

    
449
        y2 = my + range;
450
        if (y2 > ymax)
451
            y2 = ymax;
452

    
453
    } while (range >= 1);
454

    
455
#ifdef DEBUG
456
    av_log(s->avctx, AV_LOG_DEBUG, "log       - MX: %d\tMY: %d\n", mx, my);
457
#endif
458
    *mx_ptr = mx;
459
    *my_ptr = my;
460
    return dmin;
461
}
462

    
463
static int phods_motion_search(MpegEncContext * s,
464
                               int *mx_ptr, int *my_ptr, int range,
465
                               int xmin, int ymin, int xmax, int ymax, uint8_t *ref_picture)
466
{
467
    int x1, y1, x2, y2, xx, yy, x, y, lastx, d;
468
    int mx, my, dminx, dminy;
469
    uint8_t *pix;
470

    
471
    xx = s->mb_x << 4;
472
    yy = s->mb_y << 4;
473

    
474
    /* Left limit */
475
    x1 = xx - range;
476
    if (x1 < xmin)
477
        x1 = xmin;
478

    
479
    /* Right limit */
480
    x2 = xx + range;
481
    if (x2 > xmax)
482
        x2 = xmax;
483

    
484
    /* Upper limit */
485
    y1 = yy - range;
486
    if (y1 < ymin)
487
        y1 = ymin;
488

    
489
    /* Lower limit */
490
    y2 = yy + range;
491
    if (y2 > ymax)
492
        y2 = ymax;
493

    
494
    pix = s->new_picture.data[0] + (yy * s->linesize) + xx;
495
    mx = 0;
496
    my = 0;
497

    
498
    x = xx;
499
    y = yy;
500
    do {
501
        dminx = 0x7fffffff;
502
        dminy = 0x7fffffff;
503

    
504
        lastx = x;
505
        for (x = x1; x <= x2; x += range) {
506
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
507
            if (d < dminx || (d == dminx && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
508
                dminx = d;
509
                mx = x;
510
            }
511
        }
512

    
513
        x = lastx;
514
        for (y = y1; y <= y2; y += range) {
515
            d = s->dsp.pix_abs[0][0](NULL, pix, ref_picture + (y * s->linesize) + x, s->linesize, 16);
516
            if (d < dminy || (d == dminy && (abs(x - xx) + abs(y - yy)) < (abs(mx - xx) + abs(my - yy)))) {
517
                dminy = d;
518
                my = y;
519
            }
520
        }
521

    
522
        range = range >> 1;
523

    
524
        x = mx;
525
        y = my;
526
        x1 = mx - range;
527
        if (x1 < xmin)
528
            x1 = xmin;
529

    
530
        x2 = mx + range;
531
        if (x2 > xmax)
532
            x2 = xmax;
533

    
534
        y1 = my - range;
535
        if (y1 < ymin)
536
            y1 = ymin;
537

    
538
        y2 = my + range;
539
        if (y2 > ymax)
540
            y2 = ymax;
541

    
542
    } while (range >= 1);
543

    
544
#ifdef DEBUG
545
    av_log(s->avctx, AV_LOG_DEBUG, "phods     - MX: %d\tMY: %d\n", mx, my);
546
#endif
547

    
548
    /* half pixel search */
549
    *mx_ptr = mx;
550
    *my_ptr = my;
551
    return dminy;
552
}
553
#endif /* 0 */
554

    
555
#define Z_THRESHOLD 256
556

    
557
#define CHECK_SAD_HALF_MV(suffix, x, y) \
558
{\
559
    d= s->dsp.pix_abs[size][(x?1:0)+(y?2:0)](NULL, pix, ptr+((x)>>1), stride, h);\
560
    d += (mv_penalty[pen_x + x] + mv_penalty[pen_y + y])*penalty_factor;\
561
    COPY3_IF_LT(dminh, d, dx, x, dy, y)\
562
}
563

    
564
static inline int sad_hpel_motion_search(MpegEncContext * s,
565
                                  int *mx_ptr, int *my_ptr, int dmin,
566
                                  int src_index, int ref_index,
567
                                  int size, int h)
568
{
569
    MotionEstContext * const c= &s->me;
570
    const int penalty_factor= c->sub_penalty_factor;
571
    int mx, my, dminh;
572
    uint8_t *pix, *ptr;
573
    int stride= c->stride;
574
    const int flags= c->sub_flags;
575
    LOAD_COMMON
576

    
577
    assert(flags == 0);
578

    
579
    if(c->skip){
580
//    printf("S");
581
        *mx_ptr = 0;
582
        *my_ptr = 0;
583
        return dmin;
584
    }
585
//    printf("N");
586

    
587
    pix = c->src[src_index][0];
588

    
589
    mx = *mx_ptr;
590
    my = *my_ptr;
591
    ptr = c->ref[ref_index][0] + (my * stride) + mx;
592

    
593
    dminh = dmin;
594

    
595
    if (mx > xmin && mx < xmax &&
596
        my > ymin && my < ymax) {
597
        int dx=0, dy=0;
598
        int d, pen_x, pen_y;
599
        const int index= (my<<ME_MAP_SHIFT) + mx;
600
        const int t= score_map[(index-(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
601
        const int l= score_map[(index- 1               )&(ME_MAP_SIZE-1)];
602
        const int r= score_map[(index+ 1               )&(ME_MAP_SIZE-1)];
603
        const int b= score_map[(index+(1<<ME_MAP_SHIFT))&(ME_MAP_SIZE-1)];
604
        mx<<=1;
605
        my<<=1;
606

    
607

    
608
        pen_x= pred_x + mx;
609
        pen_y= pred_y + my;
610

    
611
        ptr-= stride;
612
        if(t<=b){
613
            CHECK_SAD_HALF_MV(y2 , 0, -1)
614
            if(l<=r){
615
                CHECK_SAD_HALF_MV(xy2, -1, -1)
616
                if(t+r<=b+l){
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
            }else{
625
                CHECK_SAD_HALF_MV(xy2, +1, -1)
626
                if(t+l<=b+r){
627
                    CHECK_SAD_HALF_MV(xy2, -1, -1)
628
                    ptr+= stride;
629
                }else{
630
                    ptr+= stride;
631
                    CHECK_SAD_HALF_MV(xy2, +1, +1)
632
                }
633
                CHECK_SAD_HALF_MV(x2 , +1,  0)
634
            }
635
        }else{
636
            if(l<=r){
637
                if(t+l<=b+r){
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
            }else{
647
                if(t+r<=b+l){
648
                    CHECK_SAD_HALF_MV(xy2, +1, -1)
649
                    ptr+= stride;
650
                }else{
651
                    ptr+= stride;
652
                    CHECK_SAD_HALF_MV(xy2, -1, +1)
653
                }
654
                CHECK_SAD_HALF_MV(x2 , +1,  0)
655
                CHECK_SAD_HALF_MV(xy2, +1, +1)
656
            }
657
            CHECK_SAD_HALF_MV(y2 ,  0, +1)
658
        }
659
        mx+=dx;
660
        my+=dy;
661

    
662
    }else{
663
        mx<<=1;
664
        my<<=1;
665
    }
666

    
667
    *mx_ptr = mx;
668
    *my_ptr = my;
669
    return dminh;
670
}
671

    
672
static inline void set_p_mv_tables(MpegEncContext * s, int mx, int my, int mv4)
673
{
674
    const int xy= s->mb_x + s->mb_y*s->mb_stride;
675

    
676
    s->p_mv_table[xy][0] = mx;
677
    s->p_mv_table[xy][1] = my;
678

    
679
    /* has already been set to the 4 MV if 4MV is done */
680
    if(mv4){
681
        int mot_xy= s->block_index[0];
682

    
683
        s->current_picture.motion_val[0][mot_xy  ][0]= mx;
684
        s->current_picture.motion_val[0][mot_xy  ][1]= my;
685
        s->current_picture.motion_val[0][mot_xy+1][0]= mx;
686
        s->current_picture.motion_val[0][mot_xy+1][1]= my;
687

    
688
        mot_xy += s->b8_stride;
689
        s->current_picture.motion_val[0][mot_xy  ][0]= mx;
690
        s->current_picture.motion_val[0][mot_xy  ][1]= my;
691
        s->current_picture.motion_val[0][mot_xy+1][0]= mx;
692
        s->current_picture.motion_val[0][mot_xy+1][1]= my;
693
    }
694
}
695

    
696
/**
697
 * get fullpel ME search limits.
698
 */
699
static inline void get_limits(MpegEncContext *s, int x, int y)
700
{
701
    MotionEstContext * const c= &s->me;
702
    int range= c->avctx->me_range >> (1 + !!(c->flags&FLAG_QPEL));
703
/*
704
    if(c->avctx->me_range) c->range= c->avctx->me_range >> 1;
705
    else                   c->range= 16;
706
*/
707
    if (s->unrestricted_mv) {
708
        c->xmin = - x - 16;
709
        c->ymin = - y - 16;
710
        c->xmax = - x + s->mb_width *16;
711
        c->ymax = - y + s->mb_height*16;
712
    } else if (s->out_format == FMT_H261){
713
        // Search range of H261 is different from other codec standards
714
        c->xmin = (x > 15) ? - 15 : 0;
715
        c->ymin = (y > 15) ? - 15 : 0;
716
        c->xmax = (x < s->mb_width * 16 - 16) ? 15 : 0;
717
        c->ymax = (y < s->mb_height * 16 - 16) ? 15 : 0;
718
    } else {
719
        c->xmin = - x;
720
        c->ymin = - y;
721
        c->xmax = - x + s->mb_width *16 - 16;
722
        c->ymax = - y + s->mb_height*16 - 16;
723
    }
724
    if(range){
725
        c->xmin = FFMAX(c->xmin,-range);
726
        c->xmax = FFMIN(c->xmax, range);
727
        c->ymin = FFMAX(c->ymin,-range);
728
        c->ymax = FFMIN(c->ymax, range);
729
    }
730
}
731

    
732
static inline void init_mv4_ref(MotionEstContext *c){
733
    const int stride= c->stride;
734

    
735
    c->ref[1][0] = c->ref[0][0] + 8;
736
    c->ref[2][0] = c->ref[0][0] + 8*stride;
737
    c->ref[3][0] = c->ref[2][0] + 8;
738
    c->src[1][0] = c->src[0][0] + 8;
739
    c->src[2][0] = c->src[0][0] + 8*stride;
740
    c->src[3][0] = c->src[2][0] + 8;
741
}
742

    
743
static inline int h263_mv4_search(MpegEncContext *s, int mx, int my, int shift)
744
{
745
    MotionEstContext * const c= &s->me;
746
    const int size= 1;
747
    const int h=8;
748
    int block;
749
    int P[10][2];
750
    int dmin_sum=0, mx4_sum=0, my4_sum=0;
751
    int same=1;
752
    const int stride= c->stride;
753
    uint8_t *mv_penalty= c->current_mv_penalty;
754

    
755
    init_mv4_ref(c);
756

    
757
    for(block=0; block<4; block++){
758
        int mx4, my4;
759
        int pred_x4, pred_y4;
760
        int dmin4;
761
        static const int off[4]= {2, 1, 1, -1};
762
        const int mot_stride = s->b8_stride;
763
        const int mot_xy = s->block_index[block];
764

    
765
        P_LEFT[0] = s->current_picture.motion_val[0][mot_xy - 1][0];
766
        P_LEFT[1] = s->current_picture.motion_val[0][mot_xy - 1][1];
767

    
768
        if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
769

    
770
        /* special case for first line */
771
        if (s->first_slice_line && block<2) {
772
            c->pred_x= pred_x4= P_LEFT[0];
773
            c->pred_y= pred_y4= P_LEFT[1];
774
        } else {
775
            P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][0];
776
            P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride             ][1];
777
            P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][0];
778
            P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + off[block]][1];
779
            if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
780
            if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
781
            if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
782
            if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
783

    
784
            P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
785
            P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
786

    
787
            c->pred_x= pred_x4 = P_MEDIAN[0];
788
            c->pred_y= pred_y4 = P_MEDIAN[1];
789
        }
790
        P_MV1[0]= mx;
791
        P_MV1[1]= my;
792

    
793
        dmin4 = epzs_motion_search4(s, &mx4, &my4, P, block, block, s->p_mv_table, (1<<16)>>shift);
794

    
795
        dmin4= c->sub_motion_search(s, &mx4, &my4, dmin4, block, block, size, h);
796

    
797
        if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
798
            int dxy;
799
            const int offset= ((block&1) + (block>>1)*stride)*8;
800
            uint8_t *dest_y = c->scratchpad + offset;
801
            if(s->quarter_sample){
802
                uint8_t *ref= c->ref[block][0] + (mx4>>2) + (my4>>2)*stride;
803
                dxy = ((my4 & 3) << 2) | (mx4 & 3);
804

    
805
                if(s->no_rounding)
806
                    s->dsp.put_no_rnd_qpel_pixels_tab[1][dxy](dest_y   , ref    , stride);
807
                else
808
                    s->dsp.put_qpel_pixels_tab       [1][dxy](dest_y   , ref    , stride);
809
            }else{
810
                uint8_t *ref= c->ref[block][0] + (mx4>>1) + (my4>>1)*stride;
811
                dxy = ((my4 & 1) << 1) | (mx4 & 1);
812

    
813
                if(s->no_rounding)
814
                    s->dsp.put_no_rnd_pixels_tab[1][dxy](dest_y    , ref    , stride, h);
815
                else
816
                    s->dsp.put_pixels_tab       [1][dxy](dest_y    , ref    , stride, h);
817
            }
818
            dmin_sum+= (mv_penalty[mx4-pred_x4] + mv_penalty[my4-pred_y4])*c->mb_penalty_factor;
819
        }else
820
            dmin_sum+= dmin4;
821

    
822
        if(s->quarter_sample){
823
            mx4_sum+= mx4/2;
824
            my4_sum+= my4/2;
825
        }else{
826
            mx4_sum+= mx4;
827
            my4_sum+= my4;
828
        }
829

    
830
        s->current_picture.motion_val[0][ s->block_index[block] ][0]= mx4;
831
        s->current_picture.motion_val[0][ s->block_index[block] ][1]= my4;
832

    
833
        if(mx4 != mx || my4 != my) same=0;
834
    }
835

    
836
    if(same)
837
        return INT_MAX;
838

    
839
    if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
840
        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);
841
    }
842

    
843
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
844
        int dxy;
845
        int mx, my;
846
        int offset;
847

    
848
        mx= ff_h263_round_chroma(mx4_sum);
849
        my= ff_h263_round_chroma(my4_sum);
850
        dxy = ((my & 1) << 1) | (mx & 1);
851

    
852
        offset= (s->mb_x*8 + (mx>>1)) + (s->mb_y*8 + (my>>1))*s->uvlinesize;
853

    
854
        if(s->no_rounding){
855
            s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
856
            s->dsp.put_no_rnd_pixels_tab[1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
857
        }else{
858
            s->dsp.put_pixels_tab       [1][dxy](c->scratchpad    , s->last_picture.data[1] + offset, s->uvlinesize, 8);
859
            s->dsp.put_pixels_tab       [1][dxy](c->scratchpad+8  , s->last_picture.data[2] + offset, s->uvlinesize, 8);
860
        }
861

    
862
        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);
863
        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);
864
    }
865

    
866
    c->pred_x= mx;
867
    c->pred_y= my;
868

    
869
    switch(c->avctx->mb_cmp&0xFF){
870
    /*case FF_CMP_SSE:
871
        return dmin_sum+ 32*s->qscale*s->qscale;*/
872
    case FF_CMP_RD:
873
        return dmin_sum;
874
    default:
875
        return dmin_sum+ 11*c->mb_penalty_factor;
876
    }
877
}
878

    
879
static inline void init_interlaced_ref(MpegEncContext *s, int ref_index){
880
    MotionEstContext * const c= &s->me;
881

    
882
    c->ref[1+ref_index][0] = c->ref[0+ref_index][0] + s->linesize;
883
    c->src[1][0] = c->src[0][0] + s->linesize;
884
    if(c->flags & FLAG_CHROMA){
885
        c->ref[1+ref_index][1] = c->ref[0+ref_index][1] + s->uvlinesize;
886
        c->ref[1+ref_index][2] = c->ref[0+ref_index][2] + s->uvlinesize;
887
        c->src[1][1] = c->src[0][1] + s->uvlinesize;
888
        c->src[1][2] = c->src[0][2] + s->uvlinesize;
889
    }
890
}
891

    
892
static int interlaced_search(MpegEncContext *s, int ref_index,
893
                             int16_t (*mv_tables[2][2])[2], uint8_t *field_select_tables[2], int mx, int my, int user_field_select)
894
{
895
    MotionEstContext * const c= &s->me;
896
    const int size=0;
897
    const int h=8;
898
    int block;
899
    int P[10][2];
900
    uint8_t * const mv_penalty= c->current_mv_penalty;
901
    int same=1;
902
    const int stride= 2*s->linesize;
903
    int dmin_sum= 0;
904
    const int mot_stride= s->mb_stride;
905
    const int xy= s->mb_x + s->mb_y*mot_stride;
906

    
907
    c->ymin>>=1;
908
    c->ymax>>=1;
909
    c->stride<<=1;
910
    c->uvstride<<=1;
911
    init_interlaced_ref(s, ref_index);
912

    
913
    for(block=0; block<2; block++){
914
        int field_select;
915
        int best_dmin= INT_MAX;
916
        int best_field= -1;
917

    
918
        for(field_select=0; field_select<2; field_select++){
919
            int dmin, mx_i, my_i;
920
            int16_t (*mv_table)[2]= mv_tables[block][field_select];
921

    
922
            if(user_field_select){
923
                assert(field_select==0 || field_select==1);
924
                assert(field_select_tables[block][xy]==0 || field_select_tables[block][xy]==1);
925
                if(field_select_tables[block][xy] != field_select)
926
                    continue;
927
            }
928

    
929
            P_LEFT[0] = mv_table[xy - 1][0];
930
            P_LEFT[1] = mv_table[xy - 1][1];
931
            if(P_LEFT[0]       > (c->xmax<<1)) P_LEFT[0]       = (c->xmax<<1);
932

    
933
            c->pred_x= P_LEFT[0];
934
            c->pred_y= P_LEFT[1];
935

    
936
            if(!s->first_slice_line){
937
                P_TOP[0]      = mv_table[xy - mot_stride][0];
938
                P_TOP[1]      = mv_table[xy - mot_stride][1];
939
                P_TOPRIGHT[0] = mv_table[xy - mot_stride + 1][0];
940
                P_TOPRIGHT[1] = mv_table[xy - mot_stride + 1][1];
941
                if(P_TOP[1]      > (c->ymax<<1)) P_TOP[1]     = (c->ymax<<1);
942
                if(P_TOPRIGHT[0] < (c->xmin<<1)) P_TOPRIGHT[0]= (c->xmin<<1);
943
                if(P_TOPRIGHT[0] > (c->xmax<<1)) P_TOPRIGHT[0]= (c->xmax<<1);
944
                if(P_TOPRIGHT[1] > (c->ymax<<1)) P_TOPRIGHT[1]= (c->ymax<<1);
945

    
946
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
947
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
948
            }
949
            P_MV1[0]= mx; //FIXME not correct if block != field_select
950
            P_MV1[1]= my / 2;
951

    
952
            dmin = epzs_motion_search2(s, &mx_i, &my_i, P, block, field_select+ref_index, mv_table, (1<<16)>>1);
953

    
954
            dmin= c->sub_motion_search(s, &mx_i, &my_i, dmin, block, field_select+ref_index, size, h);
955

    
956
            mv_table[xy][0]= mx_i;
957
            mv_table[xy][1]= my_i;
958

    
959
            if(s->dsp.me_sub_cmp[0] != s->dsp.mb_cmp[0]){
960
                int dxy;
961

    
962
                //FIXME chroma ME
963
                uint8_t *ref= c->ref[field_select+ref_index][0] + (mx_i>>1) + (my_i>>1)*stride;
964
                dxy = ((my_i & 1) << 1) | (mx_i & 1);
965

    
966
                if(s->no_rounding){
967
                    s->dsp.put_no_rnd_pixels_tab[size][dxy](c->scratchpad, ref    , stride, h);
968
                }else{
969
                    s->dsp.put_pixels_tab       [size][dxy](c->scratchpad, ref    , stride, h);
970
                }
971
                dmin= s->dsp.mb_cmp[size](s, c->src[block][0], c->scratchpad, stride, h);
972
                dmin+= (mv_penalty[mx_i-c->pred_x] + mv_penalty[my_i-c->pred_y] + 1)*c->mb_penalty_factor;
973
            }else
974
                dmin+= c->mb_penalty_factor; //field_select bits
975

    
976
            dmin += field_select != block; //slightly prefer same field
977

    
978
            if(dmin < best_dmin){
979
                best_dmin= dmin;
980
                best_field= field_select;
981
            }
982
        }
983
        {
984
            int16_t (*mv_table)[2]= mv_tables[block][best_field];
985

    
986
            if(mv_table[xy][0] != mx) same=0; //FIXME check if these checks work and are any good at all
987
            if(mv_table[xy][1]&1) same=0;
988
            if(mv_table[xy][1]*2 != my) same=0;
989
            if(best_field != block) same=0;
990
        }
991

    
992
        field_select_tables[block][xy]= best_field;
993
        dmin_sum += best_dmin;
994
    }
995

    
996
    c->ymin<<=1;
997
    c->ymax<<=1;
998
    c->stride>>=1;
999
    c->uvstride>>=1;
1000

    
1001
    if(same)
1002
        return INT_MAX;
1003

    
1004
    switch(c->avctx->mb_cmp&0xFF){
1005
    /*case FF_CMP_SSE:
1006
        return dmin_sum+ 32*s->qscale*s->qscale;*/
1007
    case FF_CMP_RD:
1008
        return dmin_sum;
1009
    default:
1010
        return dmin_sum+ 11*c->mb_penalty_factor;
1011
    }
1012
}
1013

    
1014
static void clip_input_mv(MpegEncContext * s, int16_t *mv, int interlaced){
1015
    int ymax= s->me.ymax>>interlaced;
1016
    int ymin= s->me.ymin>>interlaced;
1017

    
1018
    if(mv[0] < s->me.xmin) mv[0] = s->me.xmin;
1019
    if(mv[0] > s->me.xmax) mv[0] = s->me.xmax;
1020
    if(mv[1] <       ymin) mv[1] =       ymin;
1021
    if(mv[1] >       ymax) mv[1] =       ymax;
1022
}
1023

    
1024
static inline int check_input_motion(MpegEncContext * s, int mb_x, int mb_y, int p_type){
1025
    MotionEstContext * const c= &s->me;
1026
    Picture *p= s->current_picture_ptr;
1027
    int mb_xy= mb_x + mb_y*s->mb_stride;
1028
    int xy= 2*mb_x + 2*mb_y*s->b8_stride;
1029
    int mb_type= s->current_picture.mb_type[mb_xy];
1030
    int flags= c->flags;
1031
    int shift= (flags&FLAG_QPEL) + 1;
1032
    int mask= (1<<shift)-1;
1033
    int x, y, i;
1034
    int d=0;
1035
    me_cmp_func cmpf= s->dsp.sse[0];
1036
    me_cmp_func chroma_cmpf= s->dsp.sse[1];
1037

    
1038
    if(p_type && USES_LIST(mb_type, 1)){
1039
        av_log(c->avctx, AV_LOG_ERROR, "backward motion vector in P frame\n");
1040
        return INT_MAX/2;
1041
    }
1042
    assert(IS_INTRA(mb_type) || USES_LIST(mb_type,0) || USES_LIST(mb_type,1));
1043

    
1044
    for(i=0; i<4; i++){
1045
        int xy= s->block_index[i];
1046
        clip_input_mv(s, p->motion_val[0][xy], !!IS_INTERLACED(mb_type));
1047
        clip_input_mv(s, p->motion_val[1][xy], !!IS_INTERLACED(mb_type));
1048
    }
1049

    
1050
    if(IS_INTERLACED(mb_type)){
1051
        int xy2= xy  + s->b8_stride;
1052
        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1053
        c->stride<<=1;
1054
        c->uvstride<<=1;
1055

    
1056
        if(!(s->flags & CODEC_FLAG_INTERLACED_ME)){
1057
            av_log(c->avctx, AV_LOG_ERROR, "Interlaced macroblock selected but interlaced motion estimation disabled\n");
1058
            return INT_MAX/2;
1059
        }
1060

    
1061
        if(USES_LIST(mb_type, 0)){
1062
            int field_select0= p->ref_index[0][xy ];
1063
            int field_select1= p->ref_index[0][xy2];
1064
            assert(field_select0==0 ||field_select0==1);
1065
            assert(field_select1==0 ||field_select1==1);
1066
            init_interlaced_ref(s, 0);
1067

    
1068
            if(p_type){
1069
                s->p_field_select_table[0][mb_xy]= field_select0;
1070
                s->p_field_select_table[1][mb_xy]= field_select1;
1071
                *(uint32_t*)s->p_field_mv_table[0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1072
                *(uint32_t*)s->p_field_mv_table[1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1073
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER_I;
1074
            }else{
1075
                s->b_field_select_table[0][0][mb_xy]= field_select0;
1076
                s->b_field_select_table[0][1][mb_xy]= field_select1;
1077
                *(uint32_t*)s->b_field_mv_table[0][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[0][xy ];
1078
                *(uint32_t*)s->b_field_mv_table[0][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[0][xy2];
1079
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_FORWARD_I;
1080
            }
1081

    
1082
            x= p->motion_val[0][xy ][0];
1083
            y= p->motion_val[0][xy ][1];
1084
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0, 0, cmpf, chroma_cmpf, flags);
1085
            x= p->motion_val[0][xy2][0];
1086
            y= p->motion_val[0][xy2][1];
1087
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1, 1, cmpf, chroma_cmpf, flags);
1088
        }
1089
        if(USES_LIST(mb_type, 1)){
1090
            int field_select0= p->ref_index[1][xy ];
1091
            int field_select1= p->ref_index[1][xy2];
1092
            assert(field_select0==0 ||field_select0==1);
1093
            assert(field_select1==0 ||field_select1==1);
1094
            init_interlaced_ref(s, 2);
1095

    
1096
            s->b_field_select_table[1][0][mb_xy]= field_select0;
1097
            s->b_field_select_table[1][1][mb_xy]= field_select1;
1098
            *(uint32_t*)s->b_field_mv_table[1][0][field_select0][mb_xy]= *(uint32_t*)p->motion_val[1][xy ];
1099
            *(uint32_t*)s->b_field_mv_table[1][1][field_select1][mb_xy]= *(uint32_t*)p->motion_val[1][xy2];
1100
            if(USES_LIST(mb_type, 0)){
1101
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BIDIR_I;
1102
            }else{
1103
                s->mb_type[mb_xy]= CANDIDATE_MB_TYPE_BACKWARD_I;
1104
            }
1105

    
1106
            x= p->motion_val[1][xy ][0];
1107
            y= p->motion_val[1][xy ][1];
1108
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select0+2, 0, cmpf, chroma_cmpf, flags);
1109
            x= p->motion_val[1][xy2][0];
1110
            y= p->motion_val[1][xy2][1];
1111
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 8, field_select1+2, 1, cmpf, chroma_cmpf, flags);
1112
            //FIXME bidir scores
1113
        }
1114
        c->stride>>=1;
1115
        c->uvstride>>=1;
1116
    }else if(IS_8X8(mb_type)){
1117
        if(!(s->flags & CODEC_FLAG_4MV)){
1118
            av_log(c->avctx, AV_LOG_ERROR, "4MV macroblock selected but 4MV encoding disabled\n");
1119
            return INT_MAX/2;
1120
        }
1121
        cmpf= s->dsp.sse[1];
1122
        chroma_cmpf= s->dsp.sse[1];
1123
        init_mv4_ref(c);
1124
        for(i=0; i<4; i++){
1125
            xy= s->block_index[i];
1126
            x= p->motion_val[0][xy][0];
1127
            y= p->motion_val[0][xy][1];
1128
            d+= cmp(s, x>>shift, y>>shift, x&mask, y&mask, 1, 8, i, i, cmpf, chroma_cmpf, flags);
1129
        }
1130
        s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER4V;
1131
    }else{
1132
        if(USES_LIST(mb_type, 0)){
1133
            if(p_type){
1134
                *(uint32_t*)s->p_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1135
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTER;
1136
            }else if(USES_LIST(mb_type, 1)){
1137
                *(uint32_t*)s->b_bidir_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1138
                *(uint32_t*)s->b_bidir_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1139
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BIDIR;
1140
            }else{
1141
                *(uint32_t*)s->b_forw_mv_table[mb_xy]= *(uint32_t*)p->motion_val[0][xy];
1142
                s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_FORWARD;
1143
            }
1144
            x= p->motion_val[0][xy][0];
1145
            y= p->motion_val[0][xy][1];
1146
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 0, 0, cmpf, chroma_cmpf, flags);
1147
        }else if(USES_LIST(mb_type, 1)){
1148
            *(uint32_t*)s->b_back_mv_table[mb_xy]= *(uint32_t*)p->motion_val[1][xy];
1149
            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_BACKWARD;
1150

    
1151
            x= p->motion_val[1][xy][0];
1152
            y= p->motion_val[1][xy][1];
1153
            d = cmp(s, x>>shift, y>>shift, x&mask, y&mask, 0, 16, 2, 0, cmpf, chroma_cmpf, flags);
1154
        }else
1155
            s->mb_type[mb_xy]=CANDIDATE_MB_TYPE_INTRA;
1156
    }
1157
    return d;
1158
}
1159

    
1160
void ff_estimate_p_frame_motion(MpegEncContext * s,
1161
                                int mb_x, int mb_y)
1162
{
1163
    MotionEstContext * const c= &s->me;
1164
    uint8_t *pix, *ppix;
1165
    int sum, mx, my, dmin;
1166
    int varc;            ///< the variance of the block (sum of squared (p[y][x]-average))
1167
    int vard;            ///< sum of squared differences with the estimated motion vector
1168
    int P[10][2];
1169
    const int shift= 1+s->quarter_sample;
1170
    int mb_type=0;
1171
    Picture * const pic= &s->current_picture;
1172

    
1173
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1174

    
1175
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1176
    assert(s->linesize == c->stride);
1177
    assert(s->uvlinesize == c->uvstride);
1178

    
1179
    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1180
    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1181
    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1182
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1183

    
1184
    get_limits(s, 16*mb_x, 16*mb_y);
1185
    c->skip=0;
1186

    
1187
    /* intra / predictive decision */
1188
    pix = c->src[0][0];
1189
    sum = s->dsp.pix_sum(pix, s->linesize);
1190
    varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1191

    
1192
    pic->mb_mean[s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1193
    pic->mb_var [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1194
    c->mb_var_sum_temp += (varc+128)>>8;
1195

    
1196
    if(c->avctx->me_threshold){
1197
        vard= check_input_motion(s, mb_x, mb_y, 1);
1198

    
1199
        if((vard+128)>>8 < c->avctx->me_threshold){
1200
            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1201
            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1202
            pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1203
            c->mc_mb_var_sum_temp += (vard+128)>>8;
1204
            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1205
            return;
1206
        }
1207
        if((vard+128)>>8 < c->avctx->mb_threshold)
1208
            mb_type= s->mb_type[mb_x + mb_y*s->mb_stride];
1209
    }
1210

    
1211
    switch(s->me_method) {
1212
    case ME_ZERO:
1213
    default:
1214
        no_motion_search(s, &mx, &my);
1215
        mx-= mb_x*16;
1216
        my-= mb_y*16;
1217
        dmin = 0;
1218
        break;
1219
#if 0
1220
    case ME_FULL:
1221
        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1222
        mx-= mb_x*16;
1223
        my-= mb_y*16;
1224
        break;
1225
    case ME_LOG:
1226
        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1227
        mx-= mb_x*16;
1228
        my-= mb_y*16;
1229
        break;
1230
    case ME_PHODS:
1231
        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1232
        mx-= mb_x*16;
1233
        my-= mb_y*16;
1234
        break;
1235
#endif
1236
    case ME_X1:
1237
    case ME_EPZS:
1238
       {
1239
            const int mot_stride = s->b8_stride;
1240
            const int mot_xy = s->block_index[0];
1241

    
1242
            P_LEFT[0]       = s->current_picture.motion_val[0][mot_xy - 1][0];
1243
            P_LEFT[1]       = s->current_picture.motion_val[0][mot_xy - 1][1];
1244

    
1245
            if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
1246

    
1247
            if(!s->first_slice_line) {
1248
                P_TOP[0]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][0];
1249
                P_TOP[1]      = s->current_picture.motion_val[0][mot_xy - mot_stride    ][1];
1250
                P_TOPRIGHT[0] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][0];
1251
                P_TOPRIGHT[1] = s->current_picture.motion_val[0][mot_xy - mot_stride + 2][1];
1252
                if(P_TOP[1]      > (c->ymax<<shift)) P_TOP[1]     = (c->ymax<<shift);
1253
                if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1254
                if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1255

    
1256
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1257
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1258

    
1259
                if(s->out_format == FMT_H263){
1260
                    c->pred_x = P_MEDIAN[0];
1261
                    c->pred_y = P_MEDIAN[1];
1262
                }else { /* mpeg1 at least */
1263
                    c->pred_x= P_LEFT[0];
1264
                    c->pred_y= P_LEFT[1];
1265
                }
1266
            }else{
1267
                c->pred_x= P_LEFT[0];
1268
                c->pred_y= P_LEFT[1];
1269
            }
1270

    
1271
        }
1272
        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1273

    
1274
        break;
1275
    }
1276

    
1277
    /* At this point (mx,my) are full-pell and the relative displacement */
1278
    ppix = c->ref[0][0] + (my * s->linesize) + mx;
1279

    
1280
    vard = s->dsp.sse[0](NULL, pix, ppix, s->linesize, 16);
1281

    
1282
    pic->mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1283
//    pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1284
    c->mc_mb_var_sum_temp += (vard+128)>>8;
1285

    
1286
#if 0
1287
    printf("varc=%4d avg_var=%4d (sum=%4d) vard=%4d mx=%2d my=%2d\n",
1288
           varc, s->avg_mb_var, sum, vard, mx - xx, my - yy);
1289
#endif
1290
    if(mb_type){
1291
        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1292
        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1293
        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1294

    
1295
        if(mb_type == CANDIDATE_MB_TYPE_INTER){
1296
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1297
            set_p_mv_tables(s, mx, my, 1);
1298
        }else{
1299
            mx <<=shift;
1300
            my <<=shift;
1301
        }
1302
        if(mb_type == CANDIDATE_MB_TYPE_INTER4V){
1303
            h263_mv4_search(s, mx, my, shift);
1304

    
1305
            set_p_mv_tables(s, mx, my, 0);
1306
        }
1307
        if(mb_type == CANDIDATE_MB_TYPE_INTER_I){
1308
            interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 1);
1309
        }
1310
    }else if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1311
        int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1312
        int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1313
        c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1314

    
1315
        if (vard*2 + 200*256 > varc)
1316
            mb_type|= CANDIDATE_MB_TYPE_INTRA;
1317
        if (varc*2 + 200*256 > vard || s->qscale > 24){
1318
//        if (varc*2 + 200*256 + 50*(s->lambda2>>FF_LAMBDA_SHIFT) > vard){
1319
            mb_type|= CANDIDATE_MB_TYPE_INTER;
1320
            c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1321
            if(s->flags&CODEC_FLAG_MV0)
1322
                if(mx || my)
1323
                    mb_type |= CANDIDATE_MB_TYPE_SKIPPED; //FIXME check difference
1324
        }else{
1325
            mx <<=shift;
1326
            my <<=shift;
1327
        }
1328
        if((s->flags&CODEC_FLAG_4MV)
1329
           && !c->skip && varc>50<<8 && vard>10<<8){
1330
            if(h263_mv4_search(s, mx, my, shift) < INT_MAX)
1331
                mb_type|=CANDIDATE_MB_TYPE_INTER4V;
1332

    
1333
            set_p_mv_tables(s, mx, my, 0);
1334
        }else
1335
            set_p_mv_tables(s, mx, my, 1);
1336
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
1337
           && !c->skip){ //FIXME varc/d checks
1338
            if(interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0) < INT_MAX)
1339
                mb_type |= CANDIDATE_MB_TYPE_INTER_I;
1340
        }
1341
    }else{
1342
        int intra_score, i;
1343
        mb_type= CANDIDATE_MB_TYPE_INTER;
1344

    
1345
        dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1346
        if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1347
            dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1348

    
1349
        if((s->flags&CODEC_FLAG_4MV)
1350
           && !c->skip && varc>50<<8 && vard>10<<8){
1351
            int dmin4= h263_mv4_search(s, mx, my, shift);
1352
            if(dmin4 < dmin){
1353
                mb_type= CANDIDATE_MB_TYPE_INTER4V;
1354
                dmin=dmin4;
1355
            }
1356
        }
1357
        if((s->flags&CODEC_FLAG_INTERLACED_ME)
1358
           && !c->skip){ //FIXME varc/d checks
1359
            int dmin_i= interlaced_search(s, 0, s->p_field_mv_table, s->p_field_select_table, mx, my, 0);
1360
            if(dmin_i < dmin){
1361
                mb_type = CANDIDATE_MB_TYPE_INTER_I;
1362
                dmin= dmin_i;
1363
            }
1364
        }
1365

    
1366
//        pic->mb_cmp_score[s->mb_stride * mb_y + mb_x] = dmin;
1367
        set_p_mv_tables(s, mx, my, mb_type!=CANDIDATE_MB_TYPE_INTER4V);
1368

    
1369
        /* get intra luma score */
1370
        if((c->avctx->mb_cmp&0xFF)==FF_CMP_SSE){
1371
            intra_score= varc - 500;
1372
        }else{
1373
            int mean= (sum+128)>>8;
1374
            mean*= 0x01010101;
1375

    
1376
            for(i=0; i<16; i++){
1377
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 0]) = mean;
1378
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 4]) = mean;
1379
                *(uint32_t*)(&c->scratchpad[i*s->linesize+ 8]) = mean;
1380
                *(uint32_t*)(&c->scratchpad[i*s->linesize+12]) = mean;
1381
            }
1382

    
1383
            intra_score= s->dsp.mb_cmp[0](s, c->scratchpad, pix, s->linesize, 16);
1384
        }
1385
#if 0 //FIXME
1386
        /* get chroma score */
1387
        if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1388
            for(i=1; i<3; i++){
1389
                uint8_t *dest_c;
1390
                int mean;
1391

1392
                if(s->out_format == FMT_H263){
1393
                    mean= (s->dc_val[i][mb_x + mb_y*s->b8_stride] + 4)>>3; //FIXME not exact but simple ;)
1394
                }else{
1395
                    mean= (s->last_dc[i] + 4)>>3;
1396
                }
1397
                dest_c = s->new_picture.data[i] + (mb_y * 8  * (s->uvlinesize)) + mb_x * 8;
1398

1399
                mean*= 0x01010101;
1400
                for(i=0; i<8; i++){
1401
                    *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 0]) = mean;
1402
                    *(uint32_t*)(&c->scratchpad[i*s->uvlinesize+ 4]) = mean;
1403
                }
1404

1405
                intra_score+= s->dsp.mb_cmp[1](s, c->scratchpad, dest_c, s->uvlinesize);
1406
            }
1407
        }
1408
#endif
1409
        intra_score += c->mb_penalty_factor*16;
1410

    
1411
        if(intra_score < dmin){
1412
            mb_type= CANDIDATE_MB_TYPE_INTRA;
1413
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_INTRA; //FIXME cleanup
1414
        }else
1415
            s->current_picture.mb_type[mb_y*s->mb_stride + mb_x]= 0;
1416

    
1417
        {
1418
            int p_score= FFMIN(vard, varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*100);
1419
            int i_score= varc-500+(s->lambda2>>FF_LAMBDA_SHIFT)*20;
1420
            c->scene_change_score+= ff_sqrt(p_score) - ff_sqrt(i_score);
1421
        }
1422
    }
1423

    
1424
    s->mb_type[mb_y*s->mb_stride + mb_x]= mb_type;
1425
}
1426

    
1427
int ff_pre_estimate_p_frame_motion(MpegEncContext * s,
1428
                                    int mb_x, int mb_y)
1429
{
1430
    MotionEstContext * const c= &s->me;
1431
    int mx, my, dmin;
1432
    int P[10][2];
1433
    const int shift= 1+s->quarter_sample;
1434
    const int xy= mb_x + mb_y*s->mb_stride;
1435
    init_ref(c, s->new_picture.data, s->last_picture.data, NULL, 16*mb_x, 16*mb_y, 0);
1436

    
1437
    assert(s->quarter_sample==0 || s->quarter_sample==1);
1438

    
1439
    c->pre_penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_pre_cmp);
1440
    c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1441

    
1442
    get_limits(s, 16*mb_x, 16*mb_y);
1443
    c->skip=0;
1444

    
1445
    P_LEFT[0]       = s->p_mv_table[xy + 1][0];
1446
    P_LEFT[1]       = s->p_mv_table[xy + 1][1];
1447

    
1448
    if(P_LEFT[0]       < (c->xmin<<shift)) P_LEFT[0]       = (c->xmin<<shift);
1449

    
1450
    /* special case for first line */
1451
    if (s->first_slice_line) {
1452
        c->pred_x= P_LEFT[0];
1453
        c->pred_y= P_LEFT[1];
1454
        P_TOP[0]= P_TOPRIGHT[0]= P_MEDIAN[0]=
1455
        P_TOP[1]= P_TOPRIGHT[1]= P_MEDIAN[1]= 0; //FIXME
1456
    } else {
1457
        P_TOP[0]      = s->p_mv_table[xy + s->mb_stride    ][0];
1458
        P_TOP[1]      = s->p_mv_table[xy + s->mb_stride    ][1];
1459
        P_TOPRIGHT[0] = s->p_mv_table[xy + s->mb_stride - 1][0];
1460
        P_TOPRIGHT[1] = s->p_mv_table[xy + s->mb_stride - 1][1];
1461
        if(P_TOP[1]      < (c->ymin<<shift)) P_TOP[1]     = (c->ymin<<shift);
1462
        if(P_TOPRIGHT[0] > (c->xmax<<shift)) P_TOPRIGHT[0]= (c->xmax<<shift);
1463
        if(P_TOPRIGHT[1] < (c->ymin<<shift)) P_TOPRIGHT[1]= (c->ymin<<shift);
1464

    
1465
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1466
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1467

    
1468
        c->pred_x = P_MEDIAN[0];
1469
        c->pred_y = P_MEDIAN[1];
1470
    }
1471

    
1472
    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, s->p_mv_table, (1<<16)>>shift, 0, 16);
1473

    
1474
    s->p_mv_table[xy][0] = mx<<shift;
1475
    s->p_mv_table[xy][1] = my<<shift;
1476

    
1477
    return dmin;
1478
}
1479

    
1480
static int ff_estimate_motion_b(MpegEncContext * s,
1481
                       int mb_x, int mb_y, int16_t (*mv_table)[2], int ref_index, int f_code)
1482
{
1483
    MotionEstContext * const c= &s->me;
1484
    int mx, my, dmin;
1485
    int P[10][2];
1486
    const int shift= 1+s->quarter_sample;
1487
    const int mot_stride = s->mb_stride;
1488
    const int mot_xy = mb_y*mot_stride + mb_x;
1489
    uint8_t * const mv_penalty= c->mv_penalty[f_code] + MAX_MV;
1490
    int mv_scale;
1491

    
1492
    c->penalty_factor    = get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_cmp);
1493
    c->sub_penalty_factor= get_penalty_factor(s->lambda, s->lambda2, c->avctx->me_sub_cmp);
1494
    c->mb_penalty_factor = get_penalty_factor(s->lambda, s->lambda2, c->avctx->mb_cmp);
1495
    c->current_mv_penalty= mv_penalty;
1496

    
1497
    get_limits(s, 16*mb_x, 16*mb_y);
1498

    
1499
    switch(s->me_method) {
1500
    case ME_ZERO:
1501
    default:
1502
        no_motion_search(s, &mx, &my);
1503
        dmin = 0;
1504
        mx-= mb_x*16;
1505
        my-= mb_y*16;
1506
        break;
1507
#if 0
1508
    case ME_FULL:
1509
        dmin = full_motion_search(s, &mx, &my, range, ref_picture);
1510
        mx-= mb_x*16;
1511
        my-= mb_y*16;
1512
        break;
1513
    case ME_LOG:
1514
        dmin = log_motion_search(s, &mx, &my, range / 2, ref_picture);
1515
        mx-= mb_x*16;
1516
        my-= mb_y*16;
1517
        break;
1518
    case ME_PHODS:
1519
        dmin = phods_motion_search(s, &mx, &my, range / 2, ref_picture);
1520
        mx-= mb_x*16;
1521
        my-= mb_y*16;
1522
        break;
1523
#endif
1524
    case ME_X1:
1525
    case ME_EPZS:
1526
       {
1527
            P_LEFT[0]        = mv_table[mot_xy - 1][0];
1528
            P_LEFT[1]        = mv_table[mot_xy - 1][1];
1529

    
1530
            if(P_LEFT[0]       > (c->xmax<<shift)) P_LEFT[0]       = (c->xmax<<shift);
1531

    
1532
            /* special case for first line */
1533
            if (!s->first_slice_line) {
1534
                P_TOP[0] = mv_table[mot_xy - mot_stride             ][0];
1535
                P_TOP[1] = mv_table[mot_xy - mot_stride             ][1];
1536
                P_TOPRIGHT[0] = mv_table[mot_xy - mot_stride + 1         ][0];
1537
                P_TOPRIGHT[1] = mv_table[mot_xy - mot_stride + 1         ][1];
1538
                if(P_TOP[1] > (c->ymax<<shift)) P_TOP[1]= (c->ymax<<shift);
1539
                if(P_TOPRIGHT[0] < (c->xmin<<shift)) P_TOPRIGHT[0]= (c->xmin<<shift);
1540
                if(P_TOPRIGHT[1] > (c->ymax<<shift)) P_TOPRIGHT[1]= (c->ymax<<shift);
1541

    
1542
                P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1543
                P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1544
            }
1545
            c->pred_x= P_LEFT[0];
1546
            c->pred_y= P_LEFT[1];
1547
        }
1548

    
1549
        if(mv_table == s->b_forw_mv_table){
1550
            mv_scale= (s->pb_time<<16) / (s->pp_time<<shift);
1551
        }else{
1552
            mv_scale= ((s->pb_time - s->pp_time)<<16) / (s->pp_time<<shift);
1553
        }
1554

    
1555
        dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, ref_index, s->p_mv_table, mv_scale, 0, 16);
1556

    
1557
        break;
1558
    }
1559

    
1560
    dmin= c->sub_motion_search(s, &mx, &my, dmin, 0, ref_index, 0, 16);
1561

    
1562
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1563
        dmin= ff_get_mb_score(s, mx, my, 0, ref_index, 0, 16, 1);
1564

    
1565
//printf("%d %d %d %d//", s->mb_x, s->mb_y, mx, my);
1566
//    s->mb_type[mb_y*s->mb_width + mb_x]= mb_type;
1567
    mv_table[mot_xy][0]= mx;
1568
    mv_table[mot_xy][1]= my;
1569

    
1570
    return dmin;
1571
}
1572

    
1573
static inline int check_bidir_mv(MpegEncContext * s,
1574
                   int motion_fx, int motion_fy,
1575
                   int motion_bx, int motion_by,
1576
                   int pred_fx, int pred_fy,
1577
                   int pred_bx, int pred_by,
1578
                   int size, int h)
1579
{
1580
    //FIXME optimize?
1581
    //FIXME better f_code prediction (max mv & distance)
1582
    //FIXME pointers
1583
    MotionEstContext * const c= &s->me;
1584
    uint8_t * const mv_penalty_f= c->mv_penalty[s->f_code] + MAX_MV; // f_code of the prev frame
1585
    uint8_t * const mv_penalty_b= c->mv_penalty[s->b_code] + MAX_MV; // f_code of the prev frame
1586
    int stride= c->stride;
1587
    uint8_t *dest_y = c->scratchpad;
1588
    uint8_t *ptr;
1589
    int dxy;
1590
    int src_x, src_y;
1591
    int fbmin;
1592
    uint8_t **src_data= c->src[0];
1593
    uint8_t **ref_data= c->ref[0];
1594
    uint8_t **ref2_data= c->ref[2];
1595

    
1596
    if(s->quarter_sample){
1597
        dxy = ((motion_fy & 3) << 2) | (motion_fx & 3);
1598
        src_x = motion_fx >> 2;
1599
        src_y = motion_fy >> 2;
1600

    
1601
        ptr = ref_data[0] + (src_y * stride) + src_x;
1602
        s->dsp.put_qpel_pixels_tab[0][dxy](dest_y    , ptr    , stride);
1603

    
1604
        dxy = ((motion_by & 3) << 2) | (motion_bx & 3);
1605
        src_x = motion_bx >> 2;
1606
        src_y = motion_by >> 2;
1607

    
1608
        ptr = ref2_data[0] + (src_y * stride) + src_x;
1609
        s->dsp.avg_qpel_pixels_tab[size][dxy](dest_y    , ptr    , stride);
1610
    }else{
1611
        dxy = ((motion_fy & 1) << 1) | (motion_fx & 1);
1612
        src_x = motion_fx >> 1;
1613
        src_y = motion_fy >> 1;
1614

    
1615
        ptr = ref_data[0] + (src_y * stride) + src_x;
1616
        s->dsp.put_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
1617

    
1618
        dxy = ((motion_by & 1) << 1) | (motion_bx & 1);
1619
        src_x = motion_bx >> 1;
1620
        src_y = motion_by >> 1;
1621

    
1622
        ptr = ref2_data[0] + (src_y * stride) + src_x;
1623
        s->dsp.avg_pixels_tab[size][dxy](dest_y    , ptr    , stride, h);
1624
    }
1625

    
1626
    fbmin = (mv_penalty_f[motion_fx-pred_fx] + mv_penalty_f[motion_fy-pred_fy])*c->mb_penalty_factor
1627
           +(mv_penalty_b[motion_bx-pred_bx] + mv_penalty_b[motion_by-pred_by])*c->mb_penalty_factor
1628
           + s->dsp.mb_cmp[size](s, src_data[0], dest_y, stride, h); //FIXME new_pic
1629

    
1630
    if(c->avctx->mb_cmp&FF_CMP_CHROMA){
1631
    }
1632
    //FIXME CHROMA !!!
1633

    
1634
    return fbmin;
1635
}
1636

    
1637
/* refine the bidir vectors in hq mode and return the score in both lq & hq mode*/
1638
static inline int bidir_refine(MpegEncContext * s, int mb_x, int mb_y)
1639
{
1640
    MotionEstContext * const c= &s->me;
1641
    const int mot_stride = s->mb_stride;
1642
    const int xy = mb_y *mot_stride + mb_x;
1643
    int fbmin;
1644
    int pred_fx= s->b_bidir_forw_mv_table[xy-1][0];
1645
    int pred_fy= s->b_bidir_forw_mv_table[xy-1][1];
1646
    int pred_bx= s->b_bidir_back_mv_table[xy-1][0];
1647
    int pred_by= s->b_bidir_back_mv_table[xy-1][1];
1648
    int motion_fx= s->b_bidir_forw_mv_table[xy][0]= s->b_forw_mv_table[xy][0];
1649
    int motion_fy= s->b_bidir_forw_mv_table[xy][1]= s->b_forw_mv_table[xy][1];
1650
    int motion_bx= s->b_bidir_back_mv_table[xy][0]= s->b_back_mv_table[xy][0];
1651
    int motion_by= s->b_bidir_back_mv_table[xy][1]= s->b_back_mv_table[xy][1];
1652
    const int flags= c->sub_flags;
1653
    const int qpel= flags&FLAG_QPEL;
1654
    const int shift= 1+qpel;
1655
    const int xmin= c->xmin<<shift;
1656
    const int ymin= c->ymin<<shift;
1657
    const int xmax= c->xmax<<shift;
1658
    const int ymax= c->ymax<<shift;
1659
    uint8_t map[8][8][8][8];
1660

    
1661
    memset(map,0,sizeof(map));
1662
#define BIDIR_MAP(fx,fy,bx,by) \
1663
    map[(motion_fx+fx)&7][(motion_fy+fy)&7][(motion_bx+bx)&7][(motion_by+by)&7]
1664
    BIDIR_MAP(0,0,0,0) = 1;
1665

    
1666
    fbmin= check_bidir_mv(s, motion_fx, motion_fy,
1667
                          motion_bx, motion_by,
1668
                          pred_fx, pred_fy,
1669
                          pred_bx, pred_by,
1670
                          0, 16);
1671

    
1672
    if(s->avctx->bidir_refine){
1673
        int score, end;
1674
#define CHECK_BIDIR(fx,fy,bx,by)\
1675
    if( !BIDIR_MAP(fx,fy,bx,by)\
1676
       &&(fx<=0 || motion_fx+fx<=xmax) && (fy<=0 || motion_fy+fy<=ymax) && (bx<=0 || motion_bx+bx<=xmax) && (by<=0 || motion_by+by<=ymax)\
1677
       &&(fx>=0 || motion_fx+fx>=xmin) && (fy>=0 || motion_fy+fy>=ymin) && (bx>=0 || motion_bx+bx>=xmin) && (by>=0 || motion_by+by>=ymin)){\
1678
        BIDIR_MAP(fx,fy,bx,by) = 1;\
1679
        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);\
1680
        if(score < fbmin){\
1681
            fbmin= score;\
1682
            motion_fx+=fx;\
1683
            motion_fy+=fy;\
1684
            motion_bx+=bx;\
1685
            motion_by+=by;\
1686
            end=0;\
1687
        }\
1688
    }
1689
#define CHECK_BIDIR2(a,b,c,d)\
1690
CHECK_BIDIR(a,b,c,d)\
1691
CHECK_BIDIR(-(a),-(b),-(c),-(d))
1692

    
1693
#define CHECK_BIDIRR(a,b,c,d)\
1694
CHECK_BIDIR2(a,b,c,d)\
1695
CHECK_BIDIR2(b,c,d,a)\
1696
CHECK_BIDIR2(c,d,a,b)\
1697
CHECK_BIDIR2(d,a,b,c)
1698

    
1699
        do{
1700
            end=1;
1701

    
1702
            CHECK_BIDIRR( 0, 0, 0, 1)
1703
            if(s->avctx->bidir_refine > 1){
1704
                CHECK_BIDIRR( 0, 0, 1, 1)
1705
                CHECK_BIDIR2( 0, 1, 0, 1)
1706
                CHECK_BIDIR2( 1, 0, 1, 0)
1707
                CHECK_BIDIRR( 0, 0,-1, 1)
1708
                CHECK_BIDIR2( 0,-1, 0, 1)
1709
                CHECK_BIDIR2(-1, 0, 1, 0)
1710
                if(s->avctx->bidir_refine > 2){
1711
                    CHECK_BIDIRR( 0, 1, 1, 1)
1712
                    CHECK_BIDIRR( 0,-1, 1, 1)
1713
                    CHECK_BIDIRR( 0, 1,-1, 1)
1714
                    CHECK_BIDIRR( 0, 1, 1,-1)
1715
                    if(s->avctx->bidir_refine > 3){
1716
                        CHECK_BIDIR2( 1, 1, 1, 1)
1717
                        CHECK_BIDIRR( 1, 1, 1,-1)
1718
                        CHECK_BIDIR2( 1, 1,-1,-1)
1719
                        CHECK_BIDIR2( 1,-1,-1, 1)
1720
                        CHECK_BIDIR2( 1,-1, 1,-1)
1721
                    }
1722
                }
1723
            }
1724
        }while(!end);
1725
    }
1726

    
1727
    s->b_bidir_forw_mv_table[xy][0]= motion_fx;
1728
    s->b_bidir_forw_mv_table[xy][1]= motion_fy;
1729
    s->b_bidir_back_mv_table[xy][0]= motion_bx;
1730
    s->b_bidir_back_mv_table[xy][1]= motion_by;
1731

    
1732
    return fbmin;
1733
}
1734

    
1735
static inline int direct_search(MpegEncContext * s, int mb_x, int mb_y)
1736
{
1737
    MotionEstContext * const c= &s->me;
1738
    int P[10][2];
1739
    const int mot_stride = s->mb_stride;
1740
    const int mot_xy = mb_y*mot_stride + mb_x;
1741
    const int shift= 1+s->quarter_sample;
1742
    int dmin, i;
1743
    const int time_pp= s->pp_time;
1744
    const int time_pb= s->pb_time;
1745
    int mx, my, xmin, xmax, ymin, ymax;
1746
    int16_t (*mv_table)[2]= s->b_direct_mv_table;
1747

    
1748
    c->current_mv_penalty= c->mv_penalty[1] + MAX_MV;
1749
    ymin= xmin=(-32)>>shift;
1750
    ymax= xmax=   31>>shift;
1751

    
1752
    if(IS_8X8(s->next_picture.mb_type[mot_xy])){
1753
        s->mv_type= MV_TYPE_8X8;
1754
    }else{
1755
        s->mv_type= MV_TYPE_16X16;
1756
    }
1757

    
1758
    for(i=0; i<4; i++){
1759
        int index= s->block_index[i];
1760
        int min, max;
1761

    
1762
        c->co_located_mv[i][0]= s->next_picture.motion_val[0][index][0];
1763
        c->co_located_mv[i][1]= s->next_picture.motion_val[0][index][1];
1764
        c->direct_basis_mv[i][0]= c->co_located_mv[i][0]*time_pb/time_pp + ((i& 1)<<(shift+3));
1765
        c->direct_basis_mv[i][1]= c->co_located_mv[i][1]*time_pb/time_pp + ((i>>1)<<(shift+3));
1766
//        c->direct_basis_mv[1][i][0]= c->co_located_mv[i][0]*(time_pb - time_pp)/time_pp + ((i &1)<<(shift+3);
1767
//        c->direct_basis_mv[1][i][1]= c->co_located_mv[i][1]*(time_pb - time_pp)/time_pp + ((i>>1)<<(shift+3);
1768

    
1769
        max= FFMAX(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1770
        min= FFMIN(c->direct_basis_mv[i][0], c->direct_basis_mv[i][0] - c->co_located_mv[i][0])>>shift;
1771
        max+= 16*mb_x + 1; // +-1 is for the simpler rounding
1772
        min+= 16*mb_x - 1;
1773
        xmax= FFMIN(xmax, s->width - max);
1774
        xmin= FFMAX(xmin, - 16     - min);
1775

    
1776
        max= FFMAX(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1777
        min= FFMIN(c->direct_basis_mv[i][1], c->direct_basis_mv[i][1] - c->co_located_mv[i][1])>>shift;
1778
        max+= 16*mb_y + 1; // +-1 is for the simpler rounding
1779
        min+= 16*mb_y - 1;
1780
        ymax= FFMIN(ymax, s->height - max);
1781
        ymin= FFMAX(ymin, - 16      - min);
1782

    
1783
        if(s->mv_type == MV_TYPE_16X16) break;
1784
    }
1785

    
1786
    assert(xmax <= 15 && ymax <= 15 && xmin >= -16 && ymin >= -16);
1787

    
1788
    if(xmax < 0 || xmin >0 || ymax < 0 || ymin > 0){
1789
        s->b_direct_mv_table[mot_xy][0]= 0;
1790
        s->b_direct_mv_table[mot_xy][1]= 0;
1791

    
1792
        return 256*256*256*64;
1793
    }
1794

    
1795
    c->xmin= xmin;
1796
    c->ymin= ymin;
1797
    c->xmax= xmax;
1798
    c->ymax= ymax;
1799
    c->flags     |= FLAG_DIRECT;
1800
    c->sub_flags |= FLAG_DIRECT;
1801
    c->pred_x=0;
1802
    c->pred_y=0;
1803

    
1804
    P_LEFT[0]        = av_clip(mv_table[mot_xy - 1][0], xmin<<shift, xmax<<shift);
1805
    P_LEFT[1]        = av_clip(mv_table[mot_xy - 1][1], ymin<<shift, ymax<<shift);
1806

    
1807
    /* special case for first line */
1808
    if (!s->first_slice_line) { //FIXME maybe allow this over thread boundary as its clipped
1809
        P_TOP[0]      = av_clip(mv_table[mot_xy - mot_stride             ][0], xmin<<shift, xmax<<shift);
1810
        P_TOP[1]      = av_clip(mv_table[mot_xy - mot_stride             ][1], ymin<<shift, ymax<<shift);
1811
        P_TOPRIGHT[0] = av_clip(mv_table[mot_xy - mot_stride + 1         ][0], xmin<<shift, xmax<<shift);
1812
        P_TOPRIGHT[1] = av_clip(mv_table[mot_xy - mot_stride + 1         ][1], ymin<<shift, ymax<<shift);
1813

    
1814
        P_MEDIAN[0]= mid_pred(P_LEFT[0], P_TOP[0], P_TOPRIGHT[0]);
1815
        P_MEDIAN[1]= mid_pred(P_LEFT[1], P_TOP[1], P_TOPRIGHT[1]);
1816
    }
1817

    
1818
    dmin = ff_epzs_motion_search(s, &mx, &my, P, 0, 0, mv_table, 1<<(16-shift), 0, 16);
1819
    if(c->sub_flags&FLAG_QPEL)
1820
        dmin = qpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1821
    else
1822
        dmin = hpel_motion_search(s, &mx, &my, dmin, 0, 0, 0, 16);
1823

    
1824
    if(c->avctx->me_sub_cmp != c->avctx->mb_cmp && !c->skip)
1825
        dmin= ff_get_mb_score(s, mx, my, 0, 0, 0, 16, 1);
1826

    
1827
    get_limits(s, 16*mb_x, 16*mb_y); //restore c->?min/max, maybe not needed
1828

    
1829
    mv_table[mot_xy][0]= mx;
1830
    mv_table[mot_xy][1]= my;
1831
    c->flags     &= ~FLAG_DIRECT;
1832
    c->sub_flags &= ~FLAG_DIRECT;
1833

    
1834
    return dmin;
1835
}
1836

    
1837
void ff_estimate_b_frame_motion(MpegEncContext * s,
1838
                             int mb_x, int mb_y)
1839
{
1840
    MotionEstContext * const c= &s->me;
1841
    const int penalty_factor= c->mb_penalty_factor;
1842
    int fmin, bmin, dmin, fbmin, bimin, fimin;
1843
    int type=0;
1844
    const int xy = mb_y*s->mb_stride + mb_x;
1845
    init_ref(c, s->new_picture.data, s->last_picture.data, s->next_picture.data, 16*mb_x, 16*mb_y, 2);
1846

    
1847
    get_limits(s, 16*mb_x, 16*mb_y);
1848

    
1849
    c->skip=0;
1850

    
1851
    if(s->codec_id == CODEC_ID_MPEG4 && s->next_picture.mbskip_table[xy]){
1852
        int score= direct_search(s, mb_x, mb_y); //FIXME just check 0,0
1853

    
1854
        score= ((unsigned)(score*score + 128*256))>>16;
1855
        c->mc_mb_var_sum_temp += score;
1856
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1857
        s->mb_type[mb_y*s->mb_stride + mb_x]= CANDIDATE_MB_TYPE_DIRECT0;
1858

    
1859
        return;
1860
    }
1861

    
1862
    if(c->avctx->me_threshold){
1863
        int vard= check_input_motion(s, mb_x, mb_y, 0);
1864

    
1865
        if((vard+128)>>8 < c->avctx->me_threshold){
1866
//            pix = c->src[0][0];
1867
//            sum = s->dsp.pix_sum(pix, s->linesize);
1868
//            varc = s->dsp.pix_norm1(pix, s->linesize) - (((unsigned)(sum*sum))>>8) + 500;
1869

    
1870
//            pic->mb_var   [s->mb_stride * mb_y + mb_x] = (varc+128)>>8;
1871
             s->current_picture.mc_mb_var[s->mb_stride * mb_y + mb_x] = (vard+128)>>8;
1872
/*            pic->mb_mean  [s->mb_stride * mb_y + mb_x] = (sum+128)>>8;
1873
            c->mb_var_sum_temp    += (varc+128)>>8;*/
1874
            c->mc_mb_var_sum_temp += (vard+128)>>8;
1875
/*            if (vard <= 64<<8 || vard < varc) {
1876
                c->scene_change_score+= ff_sqrt(vard) - ff_sqrt(varc);
1877
            }else{
1878
                c->scene_change_score+= s->qscale * s->avctx->scenechange_factor;
1879
            }*/
1880
            return;
1881
        }
1882
        if((vard+128)>>8 < c->avctx->mb_threshold){
1883
            type= s->mb_type[mb_y*s->mb_stride + mb_x];
1884
            if(type == CANDIDATE_MB_TYPE_DIRECT){
1885
                direct_search(s, mb_x, mb_y);
1886
            }
1887
            if(type == CANDIDATE_MB_TYPE_FORWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1888
                c->skip=0;
1889
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code);
1890
            }
1891
            if(type == CANDIDATE_MB_TYPE_BACKWARD || type == CANDIDATE_MB_TYPE_BIDIR){
1892
                c->skip=0;
1893
                ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code);
1894
            }
1895
            if(type == CANDIDATE_MB_TYPE_FORWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1896
                c->skip=0;
1897
                c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1898
                interlaced_search(s, 0,
1899
                                        s->b_field_mv_table[0], s->b_field_select_table[0],
1900
                                        s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 1);
1901
            }
1902
            if(type == CANDIDATE_MB_TYPE_BACKWARD_I || type == CANDIDATE_MB_TYPE_BIDIR_I){
1903
                c->skip=0;
1904
                c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1905
                interlaced_search(s, 2,
1906
                                        s->b_field_mv_table[1], s->b_field_select_table[1],
1907
                                        s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 1);
1908
            }
1909
            return;
1910
        }
1911
    }
1912

    
1913
    if (s->codec_id == CODEC_ID_MPEG4)
1914
        dmin= direct_search(s, mb_x, mb_y);
1915
    else
1916
        dmin= INT_MAX;
1917
//FIXME penalty stuff for non mpeg4
1918
    c->skip=0;
1919
    fmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_forw_mv_table, 0, s->f_code) + 3*penalty_factor;
1920

    
1921
    c->skip=0;
1922
    bmin= ff_estimate_motion_b(s, mb_x, mb_y, s->b_back_mv_table, 2, s->b_code) + 2*penalty_factor;
1923
//printf(" %d %d ", s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1]);
1924

    
1925
    c->skip=0;
1926
    fbmin= bidir_refine(s, mb_x, mb_y) + penalty_factor;
1927
//printf("%d %d %d %d\n", dmin, fmin, bmin, fbmin);
1928

    
1929
    if(s->flags & CODEC_FLAG_INTERLACED_ME){
1930
//FIXME mb type penalty
1931
        c->skip=0;
1932
        c->current_mv_penalty= c->mv_penalty[s->f_code] + MAX_MV;
1933
        fimin= interlaced_search(s, 0,
1934
                                 s->b_field_mv_table[0], s->b_field_select_table[0],
1935
                                 s->b_forw_mv_table[xy][0], s->b_forw_mv_table[xy][1], 0);
1936
        c->current_mv_penalty= c->mv_penalty[s->b_code] + MAX_MV;
1937
        bimin= interlaced_search(s, 2,
1938
                                 s->b_field_mv_table[1], s->b_field_select_table[1],
1939
                                 s->b_back_mv_table[xy][0], s->b_back_mv_table[xy][1], 0);
1940
    }else
1941
        fimin= bimin= INT_MAX;
1942

    
1943
    {
1944
        int score= fmin;
1945
        type = CANDIDATE_MB_TYPE_FORWARD;
1946

    
1947
        if (dmin <= score){
1948
            score = dmin;
1949
            type = CANDIDATE_MB_TYPE_DIRECT;
1950
        }
1951
        if(bmin<score){
1952
            score=bmin;
1953
            type= CANDIDATE_MB_TYPE_BACKWARD;
1954
        }
1955
        if(fbmin<score){
1956
            score=fbmin;
1957
            type= CANDIDATE_MB_TYPE_BIDIR;
1958
        }
1959
        if(fimin<score){
1960
            score=fimin;
1961
            type= CANDIDATE_MB_TYPE_FORWARD_I;
1962
        }
1963
        if(bimin<score){
1964
            score=bimin;
1965
            type= CANDIDATE_MB_TYPE_BACKWARD_I;
1966
        }
1967

    
1968
        score= ((unsigned)(score*score + 128*256))>>16;
1969
        c->mc_mb_var_sum_temp += score;
1970
        s->current_picture.mc_mb_var[mb_y*s->mb_stride + mb_x] = score; //FIXME use SSE
1971
    }
1972

    
1973
    if(c->avctx->mb_decision > FF_MB_DECISION_SIMPLE){
1974
        type= CANDIDATE_MB_TYPE_FORWARD | CANDIDATE_MB_TYPE_BACKWARD | CANDIDATE_MB_TYPE_BIDIR | CANDIDATE_MB_TYPE_DIRECT;
1975
        if(fimin < INT_MAX)
1976
            type |= CANDIDATE_MB_TYPE_FORWARD_I;
1977
        if(bimin < INT_MAX)
1978
            type |= CANDIDATE_MB_TYPE_BACKWARD_I;
1979
        if(fimin < INT_MAX && bimin < INT_MAX){
1980
            type |= CANDIDATE_MB_TYPE_BIDIR_I;
1981
        }
1982
         //FIXME something smarter
1983
        if(dmin>256*256*16) type&= ~CANDIDATE_MB_TYPE_DIRECT; //dont try direct mode if its invalid for this MB
1984
        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])
1985
            type |= CANDIDATE_MB_TYPE_DIRECT0;
1986
#if 0
1987
        if(s->out_format == FMT_MPEG1)
1988
            type |= CANDIDATE_MB_TYPE_INTRA;
1989
#endif
1990
    }
1991

    
1992
    s->mb_type[mb_y*s->mb_stride + mb_x]= type;
1993
}
1994

    
1995
/* find best f_code for ME which do unlimited searches */
1996
int ff_get_best_fcode(MpegEncContext * s, int16_t (*mv_table)[2], int type)
1997
{
1998
    if(s->me_method>=ME_EPZS){
1999
        int score[8];
2000
        int i, y, range= s->avctx->me_range ? s->avctx->me_range : (INT_MAX/2);
2001
        uint8_t * fcode_tab= s->fcode_tab;
2002
        int best_fcode=-1;
2003
        int best_score=-10000000;
2004

    
2005
        if(s->msmpeg4_version)
2006
            range= FFMIN(range, 16);
2007
        else if(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL)
2008
            range= FFMIN(range, 256);
2009

    
2010
        for(i=0; i<8; i++) score[i]= s->mb_num*(8-i);
2011

    
2012
        for(y=0; y<s->mb_height; y++){
2013
            int x;
2014
            int xy= y*s->mb_stride;
2015
            for(x=0; x<s->mb_width; x++){
2016
                if(s->mb_type[xy] & type){
2017
                    int mx= mv_table[xy][0];
2018
                    int my= mv_table[xy][1];
2019
                    int fcode= FFMAX(fcode_tab[mx + MAX_MV],
2020
                                     fcode_tab[my + MAX_MV]);
2021
                    int j;
2022

    
2023
                        if(mx >= range || mx < -range ||
2024
                           my >= range || my < -range)
2025
                            continue;
2026

    
2027
                    for(j=0; j<fcode && j<8; j++){
2028
                        if(s->pict_type==B_TYPE || s->current_picture.mc_mb_var[xy] < s->current_picture.mb_var[xy])
2029
                            score[j]-= 170;
2030
                    }
2031
                }
2032
                xy++;
2033
            }
2034
        }
2035

    
2036
        for(i=1; i<8; i++){
2037
            if(score[i] > best_score){
2038
                best_score= score[i];
2039
                best_fcode= i;
2040
            }
2041
//            printf("%d %d\n", i, score[i]);
2042
        }
2043

    
2044
//    printf("fcode: %d type: %d\n", i, s->pict_type);
2045
        return best_fcode;
2046
/*        for(i=0; i<=MAX_FCODE; i++){
2047
            printf("%d ", mv_num[i]);
2048
        }
2049
        printf("\n");*/
2050
    }else{
2051
        return 1;
2052
    }
2053
}
2054

    
2055
void ff_fix_long_p_mvs(MpegEncContext * s)
2056
{
2057
    MotionEstContext * const c= &s->me;
2058
    const int f_code= s->f_code;
2059
    int y, range;
2060
    assert(s->pict_type==P_TYPE);
2061

    
2062
    range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2063

    
2064
    assert(range <= 16 || !s->msmpeg4_version);
2065
    assert(range <=256 || !(s->codec_id == CODEC_ID_MPEG2VIDEO && s->avctx->strict_std_compliance >= FF_COMPLIANCE_NORMAL));
2066

    
2067
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2068

    
2069
//printf("%d no:%d %d//\n", clip, noclip, f_code);
2070
    if(s->flags&CODEC_FLAG_4MV){
2071
        const int wrap= s->b8_stride;
2072

    
2073
        /* clip / convert to intra 8x8 type MVs */
2074
        for(y=0; y<s->mb_height; y++){
2075
            int xy= y*2*wrap;
2076
            int i= y*s->mb_stride;
2077
            int x;
2078

    
2079
            for(x=0; x<s->mb_width; x++){
2080
                if(s->mb_type[i]&CANDIDATE_MB_TYPE_INTER4V){
2081
                    int block;
2082
                    for(block=0; block<4; block++){
2083
                        int off= (block& 1) + (block>>1)*wrap;
2084
                        int mx= s->current_picture.motion_val[0][ xy + off ][0];
2085
                        int my= s->current_picture.motion_val[0][ xy + off ][1];
2086

    
2087
                        if(   mx >=range || mx <-range
2088
                           || my >=range || my <-range){
2089
                            s->mb_type[i] &= ~CANDIDATE_MB_TYPE_INTER4V;
2090
                            s->mb_type[i] |= CANDIDATE_MB_TYPE_INTRA;
2091
                            s->current_picture.mb_type[i]= CANDIDATE_MB_TYPE_INTRA;
2092
                        }
2093
                    }
2094
                }
2095
                xy+=2;
2096
                i++;
2097
            }
2098
        }
2099
    }
2100
}
2101

    
2102
/**
2103
 *
2104
 * @param truncate 1 for truncation, 0 for using intra
2105
 */
2106
void ff_fix_long_mvs(MpegEncContext * s, uint8_t *field_select_table, int field_select,
2107
                     int16_t (*mv_table)[2], int f_code, int type, int truncate)
2108
{
2109
    MotionEstContext * const c= &s->me;
2110
    int y, h_range, v_range;
2111

    
2112
    // RAL: 8 in MPEG-1, 16 in MPEG-4
2113
    int range = (((s->out_format == FMT_MPEG1 || s->msmpeg4_version) ? 8 : 16) << f_code);
2114

    
2115
    if(c->avctx->me_range && range > c->avctx->me_range) range= c->avctx->me_range;
2116

    
2117
    h_range= range;
2118
    v_range= field_select_table ? range>>1 : range;
2119

    
2120
    /* clip / convert to intra 16x16 type MVs */
2121
    for(y=0; y<s->mb_height; y++){
2122
        int x;
2123
        int xy= y*s->mb_stride;
2124
        for(x=0; x<s->mb_width; x++){
2125
            if (s->mb_type[xy] & type){    // RAL: "type" test added...
2126
                if(field_select_table==NULL || field_select_table[xy] == field_select){
2127
                    if(   mv_table[xy][0] >=h_range || mv_table[xy][0] <-h_range
2128
                       || mv_table[xy][1] >=v_range || mv_table[xy][1] <-v_range){
2129

    
2130
                        if(truncate){
2131
                            if     (mv_table[xy][0] > h_range-1) mv_table[xy][0]=  h_range-1;
2132
                            else if(mv_table[xy][0] < -h_range ) mv_table[xy][0]= -h_range;
2133
                            if     (mv_table[xy][1] > v_range-1) mv_table[xy][1]=  v_range-1;
2134
                            else if(mv_table[xy][1] < -v_range ) mv_table[xy][1]= -v_range;
2135
                        }else{
2136
                            s->mb_type[xy] &= ~type;
2137
                            s->mb_type[xy] |= CANDIDATE_MB_TYPE_INTRA;
2138
                            mv_table[xy][0]=
2139
                            mv_table[xy][1]= 0;
2140
                        }
2141
                    }
2142
                }
2143
            }
2144
            xy++;
2145
        }
2146
    }
2147
}