h264_direct.c 29.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
/*
 * H.26L/H.264/AVC/JVT/14496-10/... direct mb/block decoding
 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
23
 * @file
24
 * H.264 / AVC / MPEG-4 part10 direct mb/block decoding.
25 26 27 28 29
 * @author Michael Niedermayer <michaelni@gmx.at>
 */

#include "internal.h"
#include "avcodec.h"
30
#include "h264dec.h"
31
#include "h264_ps.h"
32
#include "mpegutils.h"
33
#include "rectangle.h"
34
#include "thread.h"
35 36 37

#include <assert.h>

38
static int get_scale_factor(H264SliceContext *sl,
39
                            int poc, int poc1, int i)
40
{
41
    int poc0 = sl->ref_list[0][i].poc;
42
    int td = av_clip_int8(poc1 - poc0);
43
    if (td == 0 || sl->ref_list[0][i].parent->long_ref) {
44
        return 256;
45
    } else {
46
        int tb = av_clip_int8(poc - poc0);
47
        int tx = (16384 + (FFABS(td) >> 1)) / td;
48
        return av_clip_intp2((tb * tx + 32) >> 6, 10);
49 50 51
    }
}

52
void ff_h264_direct_dist_scale_factor(const H264Context *const h,
53
                                      H264SliceContext *sl)
54
{
55 56
    const int poc  = FIELD_PICTURE(h) ? h->cur_pic_ptr->field_poc[h->picture_structure == PICT_BOTTOM_FIELD]
                                      : h->cur_pic_ptr->poc;
57
    const int poc1 = sl->ref_list[1][0].poc;
58 59
    int i, field;

60
    if (FRAME_MBAFF(h))
61
        for (field = 0; field < 2; field++) {
62
            const int poc  = h->cur_pic_ptr->field_poc[field];
63
            const int poc1 = sl->ref_list[1][0].parent->field_poc[field];
64
            for (i = 0; i < 2 * sl->ref_count[0]; i++)
65
                sl->dist_scale_factor_field[field][i ^ field] =
66
                    get_scale_factor(sl, poc, poc1, i + 16);
67 68
        }

69
    for (i = 0; i < sl->ref_count[0]; i++)
70
        sl->dist_scale_factor[i] = get_scale_factor(sl, poc, poc1, i);
71 72
}

73
static void fill_colmap(const H264Context *h, H264SliceContext *sl,
74
                        int map[2][16 + 32], int list,
75 76
                        int field, int colfield, int mbafi)
{
77
    H264Picture *const ref1 = sl->ref_list[1][0].parent;
78
    int j, old_ref, rfield;
79
    int start  = mbafi ? 16                       : 0;
80
    int end    = mbafi ? 16 + 2 * sl->ref_count[0] : sl->ref_count[0];
81
    int interl = mbafi || h->picture_structure != PICT_FRAME;
82 83 84 85

    /* bogus; fills in for missing frames */
    memset(map[list], 0, sizeof(map[list]));

86 87
    for (rfield = 0; rfield < 2; rfield++) {
        for (old_ref = 0; old_ref < ref1->ref_count[colfield][list]; old_ref++) {
88 89
            int poc = ref1->ref_poc[colfield][list][old_ref];

90
            if (!interl)
91
                poc |= 3;
92 93 94 95 96
            // FIXME: store all MBAFF references so this is not needed
            else if (interl && (poc & 3) == 3)
                poc = (poc & ~3) + rfield + 1;

            for (j = start; j < end; j++) {
97
                if (4 * sl->ref_list[0][j].parent->frame_num +
98
                    (sl->ref_list[0][j].reference & 3) == poc) {
99
                    int cur_ref = mbafi ? (j - 16) ^ field : j;
100
                    if (ref1->mbaff)
101 102
                        map[list][2 * old_ref + (rfield ^ field) + 16] = cur_ref;
                    if (rfield == field || !interl)
103 104 105 106 107 108 109 110
                        map[list][old_ref] = cur_ref;
                    break;
                }
            }
        }
    }
}

111
void ff_h264_direct_ref_list_init(const H264Context *const h, H264SliceContext *sl)
112
{
113
    H264Ref *const ref1 = &sl->ref_list[1][0];
114
    H264Picture *const cur = h->cur_pic_ptr;
115
    int list, j, field;
116 117
    int sidx     = (h->picture_structure & 1) ^ 1;
    int ref1sidx = (ref1->reference      & 1) ^ 1;
118

119
    for (list = 0; list < sl->list_count; list++) {
120 121
        cur->ref_count[sidx][list] = sl->ref_count[list];
        for (j = 0; j < sl->ref_count[list]; j++)
122
            cur->ref_poc[sidx][list][j] = 4 * sl->ref_list[list][j].parent->frame_num +
123
                                          (sl->ref_list[list][j].reference & 3);
124 125
    }

126
    if (h->picture_structure == PICT_FRAME) {
127
        memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
128
        memcpy(cur->ref_poc[1],   cur->ref_poc[0],   sizeof(cur->ref_poc[0]));
129 130
    }

131
    cur->mbaff = FRAME_MBAFF(h);
132

133
    sl->col_fieldoff = 0;
134 135 136 137

    if (sl->list_count != 2 || !sl->ref_count[1])
        return;

138 139
    if (h->picture_structure == PICT_FRAME) {
        int cur_poc  = h->cur_pic_ptr->poc;
140
        int *col_poc = sl->ref_list[1][0].parent->field_poc;
141 142 143 144
        if (col_poc[0] == INT_MAX && col_poc[1] == INT_MAX) {
            av_log(h->avctx, AV_LOG_ERROR, "co located POCs unavailable\n");
            sl->col_parity = 1;
        } else
145 146
        sl->col_parity = (FFABS(col_poc[0] - cur_poc) >=
                          FFABS(col_poc[1] - cur_poc));
147
        ref1sidx =
148
        sidx     = sl->col_parity;
149
    // FL -> FL & differ parity
150
    } else if (!(h->picture_structure & sl->ref_list[1][0].reference) &&
151
               !sl->ref_list[1][0].parent->mbaff) {
152
        sl->col_fieldoff = 2 * sl->ref_list[1][0].reference - 3;
153 154
    }

155
    if (sl->slice_type_nos != AV_PICTURE_TYPE_B || sl->direct_spatial_mv_pred)
156 157
        return;

158
    for (list = 0; list < 2; list++) {
159
        fill_colmap(h, sl, sl->map_col_to_list0, list, sidx, ref1sidx, 0);
160
        if (FRAME_MBAFF(h))
161
            for (field = 0; field < 2; field++)
162
                fill_colmap(h, sl, sl->map_col_to_list0_field[field], list, field,
163
                            field, 1);
164 165 166
    }
}

167
static void await_reference_mb_row(const H264Context *const h, H264Ref *ref,
168
                                   int mb_y)
169
{
170
    int ref_field         = ref->reference - 1;
171
    int ref_field_picture = ref->parent->field_picture;
172
    int ref_height        = 16 * h->mb_height >> ref_field_picture;
173

174
    if (!HAVE_THREADS || !(h->avctx->active_thread_type & FF_THREAD_FRAME))
175 176
        return;

177 178
    /* FIXME: It can be safe to access mb stuff
     * even if pixels aren't deblocked yet. */
179

180
    ff_thread_await_progress(&ref->parent->tf,
181 182
                             FFMIN(16 * mb_y >> ref_field_picture,
                                   ref_height - 1),
183 184 185
                             ref_field_picture && ref_field);
}

186
static void pred_spatial_direct_motion(const H264Context *const h, H264SliceContext *sl,
187
                                       int *mb_type)
188
{
189
    int b8_stride = 2;
190
    int b4_stride = h->b_stride;
191
    int mb_xy = sl->mb_xy, mb_y = sl->mb_y;
192 193 194 195
    int mb_type_col[2];
    const int16_t (*l1mv0)[2], (*l1mv1)[2];
    const int8_t *l1ref0, *l1ref1;
    const int is_b8x8 = IS_8X8(*mb_type);
196
    unsigned int sub_mb_type = MB_TYPE_L0L1;
197
    int i8, i4;
198 199 200
    int ref[2];
    int mv[2];
    int list;
201

202
    assert(sl->ref_list[1][0].reference & 3);
203

204
    await_reference_mb_row(h, &sl->ref_list[1][0],
205
                           sl->mb_y + !!IS_INTERLACED(*mb_type));
206

207 208
#define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16 | MB_TYPE_INTRA4x4 | \
                                MB_TYPE_INTRA16x16 | MB_TYPE_INTRA_PCM)
209 210

    /* ref = min(neighbors) */
211
    for (list = 0; list < 2; list++) {
212 213 214 215
        int left_ref     = sl->ref_cache[list][scan8[0] - 1];
        int top_ref      = sl->ref_cache[list][scan8[0] - 8];
        int refc         = sl->ref_cache[list][scan8[0] - 8 + 4];
        const int16_t *C = sl->mv_cache[list][scan8[0]  - 8 + 4];
216
        if (refc == PART_NOT_AVAILABLE) {
217 218
            refc = sl->ref_cache[list][scan8[0] - 8 - 1];
            C    = sl->mv_cache[list][scan8[0]  - 8 - 1];
219
        }
220 221 222 223 224 225
        ref[list] = FFMIN3((unsigned)left_ref,
                           (unsigned)top_ref,
                           (unsigned)refc);
        if (ref[list] >= 0) {
            /* This is just pred_motion() but with the cases removed that
             * cannot happen for direct blocks. */
226 227
            const int16_t *const A = sl->mv_cache[list][scan8[0] - 1];
            const int16_t *const B = sl->mv_cache[list][scan8[0] - 8];
228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243

            int match_count = (left_ref == ref[list]) +
                              (top_ref  == ref[list]) +
                              (refc     == ref[list]);

            if (match_count > 1) { // most common
                mv[list] = pack16to32(mid_pred(A[0], B[0], C[0]),
                                      mid_pred(A[1], B[1], C[1]));
            } else {
                assert(match_count == 1);
                if (left_ref == ref[list])
                    mv[list] = AV_RN32A(A);
                else if (top_ref == ref[list])
                    mv[list] = AV_RN32A(B);
                else
                    mv[list] = AV_RN32A(C);
244
            }
245
            av_assert2(ref[list] < (sl->ref_count[list] << !!FRAME_MBAFF(h)));
246 247 248
        } else {
            int mask = ~(MB_TYPE_L0 << (2 * list));
            mv[list]  = 0;
249
            ref[list] = -1;
250
            if (!is_b8x8)
251 252 253 254
                *mb_type &= mask;
            sub_mb_type &= mask;
        }
    }
255
    if (ref[0] < 0 && ref[1] < 0) {
256
        ref[0] = ref[1] = 0;
257
        if (!is_b8x8)
258 259 260 261
            *mb_type |= MB_TYPE_L0L1;
        sub_mb_type |= MB_TYPE_L0L1;
    }

262
    if (!(is_b8x8 | mv[0] | mv[1])) {
263 264 265 266
        fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
        fill_rectangle(&sl->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
        fill_rectangle(&sl->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
        fill_rectangle(&sl->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
267 268 269
        *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 |
                                 MB_TYPE_P1L0 | MB_TYPE_P1L1)) |
                   MB_TYPE_16x16 | MB_TYPE_DIRECT2;
270 271 272
        return;
    }

273
    if (IS_INTERLACED(sl->ref_list[1][0].parent->mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL
274
        if (!IS_INTERLACED(*mb_type)) {                    //     AFR/FR    -> AFL/FL
275 276 277
            mb_y  = (sl->mb_y & ~1) + sl->col_parity;
            mb_xy = sl->mb_x +
                    ((sl->mb_y & ~1) + sl->col_parity) * h->mb_stride;
278
            b8_stride = 0;
279
        } else {
280 281
            mb_y  += sl->col_fieldoff;
            mb_xy += h->mb_stride * sl->col_fieldoff; // non-zero for FL -> FL & differ parity
282 283
        }
        goto single_col;
284 285
    } else {                                             // AFL/AFR/FR/FL -> AFR/FR
        if (IS_INTERLACED(*mb_type)) {                   // AFL       /FL -> AFR/FR
286 287
            mb_y           =  sl->mb_y & ~1;
            mb_xy          = (sl->mb_y & ~1) * h->mb_stride + sl->mb_x;
288 289
            mb_type_col[0] = sl->ref_list[1][0].parent->mb_type[mb_xy];
            mb_type_col[1] = sl->ref_list[1][0].parent->mb_type[mb_xy + h->mb_stride];
290 291 292 293
            b8_stride      = 2 + 4 * h->mb_stride;
            b4_stride     *= 6;
            if (IS_INTERLACED(mb_type_col[0]) !=
                IS_INTERLACED(mb_type_col[1])) {
294 295 296
                mb_type_col[0] &= ~MB_TYPE_INTERLACED;
                mb_type_col[1] &= ~MB_TYPE_INTERLACED;
            }
297

298 299 300 301 302 303 304
            sub_mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */
            if ((mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) &&
                (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) &&
                !is_b8x8) {
                *mb_type |= MB_TYPE_16x8 | MB_TYPE_DIRECT2;  /* B_16x8 */
            } else {
                *mb_type |= MB_TYPE_8x8;
305
            }
306
        } else {                                         //     AFR/FR    -> AFR/FR
307 308
single_col:
            mb_type_col[0] =
309
            mb_type_col[1] = sl->ref_list[1][0].parent->mb_type[mb_xy];
310

311 312 313 314 315 316 317 318
            sub_mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_SUB_8x8 */
            if (!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)) {
                *mb_type |= MB_TYPE_16x16 | MB_TYPE_DIRECT2; /* B_16x16 */
            } else if (!is_b8x8 &&
                       (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16))) {
                *mb_type |= MB_TYPE_DIRECT2 |
                            (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16));
            } else {
319
                if (!h->ps.sps->direct_8x8_inference_flag) {
320 321 322
                    /* FIXME: Save sub mb types from previous frames (or derive
                     * from MVs) so we know exactly what block size to use. */
                    sub_mb_type += (MB_TYPE_8x8 - MB_TYPE_16x16); /* B_SUB_4x4 */
323
                }
324
                *mb_type |= MB_TYPE_8x8;
325 326 327 328
            }
        }
    }

329
    await_reference_mb_row(h, &sl->ref_list[1][0], mb_y);
330

331 332
    l1mv0  = (void*)&sl->ref_list[1][0].parent->motion_val[0][h->mb2b_xy[mb_xy]];
    l1mv1  = (void*)&sl->ref_list[1][0].parent->motion_val[1][h->mb2b_xy[mb_xy]];
333 334
    l1ref0 = &sl->ref_list[1][0].parent->ref_index[0][4 * mb_xy];
    l1ref1 = &sl->ref_list[1][0].parent->ref_index[1][4 * mb_xy];
335
    if (!b8_stride) {
336
        if (sl->mb_y & 1) {
337 338
            l1ref0 += 2;
            l1ref1 += 2;
339 340
            l1mv0  += 2 * b4_stride;
            l1mv1  += 2 * b4_stride;
341 342 343
        }
    }

344 345 346 347 348 349 350 351 352
    if (IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])) {
        int n = 0;
        for (i8 = 0; i8 < 4; i8++) {
            int x8  = i8 & 1;
            int y8  = i8 >> 1;
            int xy8 = x8     + y8 * b8_stride;
            int xy4 = x8 * 3 + y8 * b4_stride;
            int a, b;

353
            if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8]))
354
                continue;
355
            sl->sub_mb_type[i8] = sub_mb_type;
356

357
            fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8,
358
                           (uint8_t)ref[0], 1);
359
            fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8,
360
                           (uint8_t)ref[1], 1);
361
            if (!IS_INTRA(mb_type_col[y8]) && !sl->ref_list[1][0].parent->long_ref &&
362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378
                ((l1ref0[xy8] == 0 &&
                  FFABS(l1mv0[xy4][0]) <= 1 &&
                  FFABS(l1mv0[xy4][1]) <= 1) ||
                 (l1ref0[xy8] < 0 &&
                  l1ref1[xy8] == 0 &&
                  FFABS(l1mv1[xy4][0]) <= 1 &&
                  FFABS(l1mv1[xy4][1]) <= 1))) {
                a =
                b = 0;
                if (ref[0] > 0)
                    a = mv[0];
                if (ref[1] > 0)
                    b = mv[1];
                n++;
            } else {
                a = mv[0];
                b = mv[1];
379
            }
380 381
            fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, a, 4);
            fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, b, 4);
382 383 384 385 386 387 388
        }
        if (!is_b8x8 && !(n & 3))
            *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 |
                                     MB_TYPE_P1L0 | MB_TYPE_P1L1)) |
                       MB_TYPE_16x16 | MB_TYPE_DIRECT2;
    } else if (IS_16X16(*mb_type)) {
        int a, b;
389

390 391
        fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
        fill_rectangle(&sl->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
392
        if (!IS_INTRA(mb_type_col[0]) && !sl->ref_list[1][0].parent->long_ref &&
393 394 395 396 397 398
            ((l1ref0[0] == 0 &&
              FFABS(l1mv0[0][0]) <= 1 &&
              FFABS(l1mv0[0][1]) <= 1) ||
             (l1ref0[0] < 0 && !l1ref1[0] &&
              FFABS(l1mv1[0][0]) <= 1 &&
              FFABS(l1mv1[0][1]) <= 1 &&
399
              h->sei.unregistered.x264_build > 33U))) {
400 401 402 403 404 405 406 407 408
            a = b = 0;
            if (ref[0] > 0)
                a = mv[0];
            if (ref[1] > 0)
                b = mv[1];
        } else {
            a = mv[0];
            b = mv[1];
        }
409 410
        fill_rectangle(&sl->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
        fill_rectangle(&sl->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
411 412 413 414 415 416
    } else {
        int n = 0;
        for (i8 = 0; i8 < 4; i8++) {
            const int x8 = i8 & 1;
            const int y8 = i8 >> 1;

417
            if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8]))
418
                continue;
419
            sl->sub_mb_type[i8] = sub_mb_type;
420

421 422 423
            fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, mv[0], 4);
            fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, mv[1], 4);
            fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8,
424
                           (uint8_t)ref[0], 1);
425
            fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8,
426 427 428 429
                           (uint8_t)ref[1], 1);

            assert(b8_stride == 2);
            /* col_zero_flag */
430
            if (!IS_INTRA(mb_type_col[0]) && !sl->ref_list[1][0].parent->long_ref &&
431 432 433
                (l1ref0[i8] == 0 ||
                 (l1ref0[i8] < 0 &&
                  l1ref1[i8] == 0 &&
434
                  h->sei.unregistered.x264_build > 33U))) {
435 436 437 438 439
                const int16_t (*l1mv)[2] = l1ref0[i8] == 0 ? l1mv0 : l1mv1;
                if (IS_SUB_8X8(sub_mb_type)) {
                    const int16_t *mv_col = l1mv[x8 * 3 + y8 * 3 * b4_stride];
                    if (FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1) {
                        if (ref[0] == 0)
440
                            fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2,
441 442
                                           8, 0, 4);
                        if (ref[1] == 0)
443
                            fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2,
444 445 446 447 448 449 450 451 452 453
                                           8, 0, 4);
                        n += 4;
                    }
                } else {
                    int m = 0;
                    for (i4 = 0; i4 < 4; i4++) {
                        const int16_t *mv_col = l1mv[x8 * 2 + (i4 & 1) +
                                                     (y8 * 2 + (i4 >> 1)) * b4_stride];
                        if (FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1) {
                            if (ref[0] == 0)
454
                                AV_ZERO32(sl->mv_cache[0][scan8[i8 * 4 + i4]]);
455
                            if (ref[1] == 0)
456
                                AV_ZERO32(sl->mv_cache[1][scan8[i8 * 4 + i4]]);
457
                            m++;
458 459
                        }
                    }
460
                    if (!(m & 3))
461
                        sl->sub_mb_type[i8] += MB_TYPE_16x16 - MB_TYPE_8x8;
462
                    n += m;
463 464 465
                }
            }
        }
466 467 468 469 470
        if (!is_b8x8 && !(n & 15))
            *mb_type = (*mb_type & ~(MB_TYPE_8x8 | MB_TYPE_16x8 | MB_TYPE_8x16 |
                                     MB_TYPE_P1L0 | MB_TYPE_P1L1)) |
                       MB_TYPE_16x16 | MB_TYPE_DIRECT2;
    }
471 472
}

473
static void pred_temp_direct_motion(const H264Context *const h, H264SliceContext *sl,
474
                                    int *mb_type)
475
{
476
    int b8_stride = 2;
477
    int b4_stride = h->b_stride;
478
    int mb_xy = sl->mb_xy, mb_y = sl->mb_y;
479 480 481 482 483 484 485
    int mb_type_col[2];
    const int16_t (*l1mv0)[2], (*l1mv1)[2];
    const int8_t *l1ref0, *l1ref1;
    const int is_b8x8 = IS_8X8(*mb_type);
    unsigned int sub_mb_type;
    int i8, i4;

486
    assert(sl->ref_list[1][0].reference & 3);
487

488
    await_reference_mb_row(h, &sl->ref_list[1][0],
489
                           sl->mb_y + !!IS_INTERLACED(*mb_type));
490

491
    if (IS_INTERLACED(sl->ref_list[1][0].parent->mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL
492
        if (!IS_INTERLACED(*mb_type)) {                    //     AFR/FR    -> AFL/FL
493 494 495
            mb_y  = (sl->mb_y & ~1) + sl->col_parity;
            mb_xy = sl->mb_x +
                    ((sl->mb_y & ~1) + sl->col_parity) * h->mb_stride;
496
            b8_stride = 0;
497
        } else {
498 499
            mb_y  += sl->col_fieldoff;
            mb_xy += h->mb_stride * sl->col_fieldoff; // non-zero for FL -> FL & differ parity
500 501
        }
        goto single_col;
502 503
    } else {                                        // AFL/AFR/FR/FL -> AFR/FR
        if (IS_INTERLACED(*mb_type)) {              // AFL       /FL -> AFR/FR
504 505
            mb_y           = sl->mb_y & ~1;
            mb_xy          = sl->mb_x + (sl->mb_y & ~1) * h->mb_stride;
506 507
            mb_type_col[0] = sl->ref_list[1][0].parent->mb_type[mb_xy];
            mb_type_col[1] = sl->ref_list[1][0].parent->mb_type[mb_xy + h->mb_stride];
508 509 510 511
            b8_stride      = 2 + 4 * h->mb_stride;
            b4_stride     *= 6;
            if (IS_INTERLACED(mb_type_col[0]) !=
                IS_INTERLACED(mb_type_col[1])) {
512 513 514
                mb_type_col[0] &= ~MB_TYPE_INTERLACED;
                mb_type_col[1] &= ~MB_TYPE_INTERLACED;
            }
515

516 517
            sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |
                          MB_TYPE_DIRECT2;                  /* B_SUB_8x8 */
518

519 520 521 522 523 524 525
            if ((mb_type_col[0] & MB_TYPE_16x16_OR_INTRA) &&
                (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA) &&
                !is_b8x8) {
                *mb_type |= MB_TYPE_16x8 | MB_TYPE_L0L1 |
                            MB_TYPE_DIRECT2;                /* B_16x8 */
            } else {
                *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1;
526
            }
527
        } else {                                    //     AFR/FR    -> AFR/FR
528
single_col:
529
            mb_type_col[0]     =
530
                mb_type_col[1] = sl->ref_list[1][0].parent->mb_type[mb_xy];
531 532 533 534 535 536 537 538 539 540 541

            sub_mb_type = MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |
                          MB_TYPE_DIRECT2;                  /* B_SUB_8x8 */
            if (!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)) {
                *mb_type |= MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |
                            MB_TYPE_DIRECT2;                /* B_16x16 */
            } else if (!is_b8x8 &&
                       (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16))) {
                *mb_type |= MB_TYPE_L0L1 | MB_TYPE_DIRECT2 |
                            (mb_type_col[0] & (MB_TYPE_16x8 | MB_TYPE_8x16));
            } else {
542
                if (!h->ps.sps->direct_8x8_inference_flag) {
543 544 545 546
                    /* FIXME: save sub mb types from previous frames (or derive
                     * from MVs) so we know exactly what block size to use */
                    sub_mb_type = MB_TYPE_8x8 | MB_TYPE_P0L0 | MB_TYPE_P0L1 |
                                  MB_TYPE_DIRECT2;          /* B_SUB_4x4 */
547
                }
548
                *mb_type |= MB_TYPE_8x8 | MB_TYPE_L0L1;
549 550 551 552
            }
        }
    }

553
    await_reference_mb_row(h, &sl->ref_list[1][0], mb_y);
554

555 556
    l1mv0  = (void*)&sl->ref_list[1][0].parent->motion_val[0][h->mb2b_xy[mb_xy]];
    l1mv1  = (void*)&sl->ref_list[1][0].parent->motion_val[1][h->mb2b_xy[mb_xy]];
557 558
    l1ref0 = &sl->ref_list[1][0].parent->ref_index[0][4 * mb_xy];
    l1ref1 = &sl->ref_list[1][0].parent->ref_index[1][4 * mb_xy];
559
    if (!b8_stride) {
560
        if (sl->mb_y & 1) {
561 562
            l1ref0 += 2;
            l1ref1 += 2;
563 564
            l1mv0  += 2 * b4_stride;
            l1mv1  += 2 * b4_stride;
565 566 567 568
        }
    }

    {
569 570
        const int *map_col_to_list0[2] = { sl->map_col_to_list0[0],
                                           sl->map_col_to_list0[1] };
571
        const int *dist_scale_factor = sl->dist_scale_factor;
572
        int ref_offset;
573

574
        if (FRAME_MBAFF(h) && IS_INTERLACED(*mb_type)) {
575 576 577
            map_col_to_list0[0] = sl->map_col_to_list0_field[sl->mb_y & 1][0];
            map_col_to_list0[1] = sl->map_col_to_list0_field[sl->mb_y & 1][1];
            dist_scale_factor   = sl->dist_scale_factor_field[sl->mb_y & 1];
578
        }
579
        ref_offset = (sl->ref_list[1][0].parent->mbaff << 4) & (mb_type_col[0] >> 3);
580

581 582
        if (IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])) {
            int y_shift = 2 * !IS_INTERLACED(*mb_type);
583
            assert(h->ps.sps->direct_8x8_inference_flag);
584

585 586 587
            for (i8 = 0; i8 < 4; i8++) {
                const int x8 = i8 & 1;
                const int y8 = i8 >> 1;
588
                int ref0, scale;
589
                const int16_t (*l1mv)[2] = l1mv0;
590

591
                if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8]))
592
                    continue;
593
                sl->sub_mb_type[i8] = sub_mb_type;
594

595
                fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 1);
596
                if (IS_INTRA(mb_type_col[y8])) {
597 598 599
                    fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 1);
                    fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 4);
                    fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 4);
600 601 602
                    continue;
                }

603 604
                ref0 = l1ref0[x8 + y8 * b8_stride];
                if (ref0 >= 0)
605
                    ref0 = map_col_to_list0[0][ref0 + ref_offset];
606 607 608 609
                else {
                    ref0 = map_col_to_list0[1][l1ref1[x8 + y8 * b8_stride] +
                                               ref_offset];
                    l1mv = l1mv1;
610 611
                }
                scale = dist_scale_factor[ref0];
612
                fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8,
613
                               ref0, 1);
614 615

                {
616 617 618 619
                    const int16_t *mv_col = l1mv[x8 * 3 + y8 * b4_stride];
                    int my_col            = (mv_col[1] << y_shift) / 2;
                    int mx                = (scale * mv_col[0] + 128) >> 8;
                    int my                = (scale * my_col    + 128) >> 8;
620
                    fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8,
621
                                   pack16to32(mx, my), 4);
622
                    fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8,
623
                                   pack16to32(mx - mv_col[0], my - my_col), 4);
624 625 626 627 628 629 630
                }
            }
            return;
        }

        /* one-to-one mv scaling */

631
        if (IS_16X16(*mb_type)) {
632 633
            int ref, mv0, mv1;

634
            fill_rectangle(&sl->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
635 636 637
            if (IS_INTRA(mb_type_col[0])) {
                ref = mv0 = mv1 = 0;
            } else {
638 639 640 641 642 643 644
                const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
                                                : map_col_to_list0[1][l1ref1[0] + ref_offset];
                const int scale = dist_scale_factor[ref0];
                const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
                int mv_l0[2];
                mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
                mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
645 646 647
                ref      = ref0;
                mv0      = pack16to32(mv_l0[0], mv_l0[1]);
                mv1      = pack16to32(mv_l0[0] - mv_col[0], mv_l0[1] - mv_col[1]);
648
            }
649 650 651
            fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
            fill_rectangle(&sl->mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
            fill_rectangle(&sl->mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
652 653 654 655
        } else {
            for (i8 = 0; i8 < 4; i8++) {
                const int x8 = i8 & 1;
                const int y8 = i8 >> 1;
656
                int ref0, scale;
657
                const int16_t (*l1mv)[2] = l1mv0;
658

659
                if (is_b8x8 && !IS_DIRECT(sl->sub_mb_type[i8]))
660
                    continue;
661
                sl->sub_mb_type[i8] = sub_mb_type;
662
                fill_rectangle(&sl->ref_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 1);
663
                if (IS_INTRA(mb_type_col[0])) {
664 665 666
                    fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 1);
                    fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8, 0, 4);
                    fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8, 0, 4);
667 668 669
                    continue;
                }

670 671
                assert(b8_stride == 2);
                ref0 = l1ref0[i8];
672
                if (ref0 >= 0)
673
                    ref0 = map_col_to_list0[0][ref0 + ref_offset];
674
                else {
675
                    ref0 = map_col_to_list0[1][l1ref1[i8] + ref_offset];
676
                    l1mv = l1mv1;
677 678 679
                }
                scale = dist_scale_factor[ref0];

680
                fill_rectangle(&sl->ref_cache[0][scan8[i8 * 4]], 2, 2, 8,
681 682 683 684 685
                               ref0, 1);
                if (IS_SUB_8X8(sub_mb_type)) {
                    const int16_t *mv_col = l1mv[x8 * 3 + y8 * 3 * b4_stride];
                    int mx                = (scale * mv_col[0] + 128) >> 8;
                    int my                = (scale * mv_col[1] + 128) >> 8;
686
                    fill_rectangle(&sl->mv_cache[0][scan8[i8 * 4]], 2, 2, 8,
687
                                   pack16to32(mx, my), 4);
688
                    fill_rectangle(&sl->mv_cache[1][scan8[i8 * 4]], 2, 2, 8,
689 690 691 692 693
                                   pack16to32(mx - mv_col[0], my - mv_col[1]), 4);
                } else {
                    for (i4 = 0; i4 < 4; i4++) {
                        const int16_t *mv_col = l1mv[x8 * 2 + (i4 & 1) +
                                                     (y8 * 2 + (i4 >> 1)) * b4_stride];
694
                        int16_t *mv_l0 = sl->mv_cache[0][scan8[i8 * 4 + i4]];
695 696
                        mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
                        mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
697
                        AV_WN32A(sl->mv_cache[1][scan8[i8 * 4 + i4]],
698 699 700
                                 pack16to32(mv_l0[0] - mv_col[0],
                                            mv_l0[1] - mv_col[1]));
                    }
701 702 703 704 705
                }
            }
        }
    }
}
706

707
void ff_h264_pred_direct_motion(const H264Context *const h, H264SliceContext *sl,
708
                                int *mb_type)
709
{
710
    if (sl->direct_spatial_mv_pred)
711
        pred_spatial_direct_motion(h, sl, mb_type);
712
    else
713
        pred_temp_direct_motion(h, sl, mb_type);
714
}