1/*
2 * H.26L/H.264/AVC/JVT/14496-10/... motion vector prediction
3 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22/**
23 * @file
24 * H.264 / AVC / MPEG-4 part10 motion vector prediction.
25 * @author Michael Niedermayer <michaelni@gmx.at>
26 */
27
28#ifndef AVCODEC_H264_MVPRED_H
29#define AVCODEC_H264_MVPRED_H
30
31#include "h264dec.h"
32#include "mpegutils.h"
33#include "libavutil/avassert.h"
34#include "libavutil/mem_internal.h"
35
36
37static av_always_inline int fetch_diagonal_mv(const H264Context *h, H264SliceContext *sl,
38                                              const int16_t **C,
39                                              int i, int list, int part_width)
40{
41    const int topright_ref = sl->ref_cache[list][i - 8 + part_width];
42
43    /* there is no consistent mapping of mvs to neighboring locations that will
44     * make mbaff happy, so we can't move all this logic to fill_caches */
45    if (FRAME_MBAFF(h)) {
46#define SET_DIAG_MV(MV_OP, REF_OP, XY, Y4)                              \
47        const int xy = XY, y4 = Y4;                                     \
48        const int mb_type = mb_types[xy + (y4 >> 2) * h->mb_stride];    \
49        if (!USES_LIST(mb_type, list))                                  \
50            return LIST_NOT_USED;                                       \
51        mv = h->cur_pic_ptr->motion_val[list][h->mb2b_xy[xy] + 3 + y4 * h->b_stride]; \
52        sl->mv_cache[list][scan8[0] - 2][0] = mv[0];                     \
53        sl->mv_cache[list][scan8[0] - 2][1] = mv[1] MV_OP;               \
54        return h->cur_pic_ptr->ref_index[list][4 * xy + 1 + (y4 & ~1)] REF_OP;
55
56        if (topright_ref == PART_NOT_AVAILABLE
57            && i >= scan8[0] + 8 && (i & 7) == 4
58            && sl->ref_cache[list][scan8[0] - 1] != PART_NOT_AVAILABLE) {
59            const uint32_t *mb_types = h->cur_pic_ptr->mb_type;
60            const int16_t *mv;
61            AV_ZERO32(sl->mv_cache[list][scan8[0] - 2]);
62            *C = sl->mv_cache[list][scan8[0] - 2];
63
64            if (!MB_FIELD(sl) && IS_INTERLACED(sl->left_type[0])) {
65                SET_DIAG_MV(* 2, >> 1, sl->left_mb_xy[0] + h->mb_stride,
66                            (sl->mb_y & 1) * 2 + (i >> 5));
67            }
68            if (MB_FIELD(sl) && !IS_INTERLACED(sl->left_type[0])) {
69                // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
70                SET_DIAG_MV(/ 2, *2, sl->left_mb_xy[i >= 36], ((i >> 2)) & 3);
71            }
72        }
73#undef SET_DIAG_MV
74    }
75
76    if (topright_ref != PART_NOT_AVAILABLE) {
77        *C = sl->mv_cache[list][i - 8 + part_width];
78        return topright_ref;
79    } else {
80        ff_tlog(h->avctx, "topright MV not available\n");
81
82        *C = sl->mv_cache[list][i - 8 - 1];
83        return sl->ref_cache[list][i - 8 - 1];
84    }
85}
86
87/**
88 * Get the predicted MV.
89 * @param n the block index
90 * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
91 * @param mx the x component of the predicted motion vector
92 * @param my the y component of the predicted motion vector
93 */
94static av_always_inline void pred_motion(const H264Context *const h,
95                                         H264SliceContext *sl,
96                                         int n,
97                                         int part_width, int list, int ref,
98                                         int *const mx, int *const my)
99{
100    const int index8       = scan8[n];
101    const int top_ref      = sl->ref_cache[list][index8 - 8];
102    const int left_ref     = sl->ref_cache[list][index8 - 1];
103    const int16_t *const A = sl->mv_cache[list][index8 - 1];
104    const int16_t *const B = sl->mv_cache[list][index8 - 8];
105    const int16_t *C;
106    int diagonal_ref, match_count;
107
108    av_assert2(part_width == 1 || part_width == 2 || part_width == 4);
109
110/* mv_cache
111 * B . . A T T T T
112 * U . . L . . , .
113 * U . . L . . . .
114 * U . . L . . , .
115 * . . . L . . . .
116 */
117
118    diagonal_ref = fetch_diagonal_mv(h, sl, &C, index8, list, part_width);
119    match_count  = (diagonal_ref == ref) + (top_ref == ref) + (left_ref == ref);
120    ff_tlog(h->avctx, "pred_motion match_count=%d\n", match_count);
121    if (match_count > 1) { //most common
122        *mx = mid_pred(A[0], B[0], C[0]);
123        *my = mid_pred(A[1], B[1], C[1]);
124    } else if (match_count == 1) {
125        if (left_ref == ref) {
126            *mx = A[0];
127            *my = A[1];
128        } else if (top_ref == ref) {
129            *mx = B[0];
130            *my = B[1];
131        } else {
132            *mx = C[0];
133            *my = C[1];
134        }
135    } else {
136        if (top_ref      == PART_NOT_AVAILABLE &&
137            diagonal_ref == PART_NOT_AVAILABLE &&
138            left_ref     != PART_NOT_AVAILABLE) {
139            *mx = A[0];
140            *my = A[1];
141        } else {
142            *mx = mid_pred(A[0], B[0], C[0]);
143            *my = mid_pred(A[1], B[1], C[1]);
144        }
145    }
146
147    ff_tlog(h->avctx,
148            "pred_motion (%2d %2d %2d) (%2d %2d %2d) (%2d %2d %2d) -> (%2d %2d %2d) at %2d %2d %d list %d\n",
149            top_ref, B[0], B[1], diagonal_ref, C[0], C[1], left_ref,
150            A[0], A[1], ref, *mx, *my, sl->mb_x, sl->mb_y, n, list);
151}
152
153/**
154 * Get the directionally predicted 16x8 MV.
155 * @param n the block index
156 * @param mx the x component of the predicted motion vector
157 * @param my the y component of the predicted motion vector
158 */
159static av_always_inline void pred_16x8_motion(const H264Context *const h,
160                                              H264SliceContext *sl,
161                                              int n, int list, int ref,
162                                              int *const mx, int *const my)
163{
164    if (n == 0) {
165        const int top_ref      = sl->ref_cache[list][scan8[0] - 8];
166        const int16_t *const B = sl->mv_cache[list][scan8[0] - 8];
167
168        ff_tlog(h->avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n",
169                top_ref, B[0], B[1], sl->mb_x, sl->mb_y, n, list);
170
171        if (top_ref == ref) {
172            *mx = B[0];
173            *my = B[1];
174            return;
175        }
176    } else {
177        const int left_ref     = sl->ref_cache[list][scan8[8] - 1];
178        const int16_t *const A = sl->mv_cache[list][scan8[8] - 1];
179
180        ff_tlog(h->avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n",
181                left_ref, A[0], A[1], sl->mb_x, sl->mb_y, n, list);
182
183        if (left_ref == ref) {
184            *mx = A[0];
185            *my = A[1];
186            return;
187        }
188    }
189
190    //RARE
191    pred_motion(h, sl, n, 4, list, ref, mx, my);
192}
193
194/**
195 * Get the directionally predicted 8x16 MV.
196 * @param n the block index
197 * @param mx the x component of the predicted motion vector
198 * @param my the y component of the predicted motion vector
199 */
200static av_always_inline void pred_8x16_motion(const H264Context *const h,
201                                              H264SliceContext *sl,
202                                              int n, int list, int ref,
203                                              int *const mx, int *const my)
204{
205    if (n == 0) {
206        const int left_ref     = sl->ref_cache[list][scan8[0] - 1];
207        const int16_t *const A = sl->mv_cache[list][scan8[0] - 1];
208
209        ff_tlog(h->avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n",
210                left_ref, A[0], A[1], sl->mb_x, sl->mb_y, n, list);
211
212        if (left_ref == ref) {
213            *mx = A[0];
214            *my = A[1];
215            return;
216        }
217    } else {
218        const int16_t *C;
219        int diagonal_ref;
220
221        diagonal_ref = fetch_diagonal_mv(h, sl, &C, scan8[4], list, 2);
222
223        ff_tlog(h->avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n",
224                diagonal_ref, C[0], C[1], sl->mb_x, sl->mb_y, n, list);
225
226        if (diagonal_ref == ref) {
227            *mx = C[0];
228            *my = C[1];
229            return;
230        }
231    }
232
233    //RARE
234    pred_motion(h, sl, n, 2, list, ref, mx, my);
235}
236
237#define FIX_MV_MBAFF(type, refn, mvn, idx)      \
238    if (FRAME_MBAFF(h)) {                       \
239        if (MB_FIELD(sl)) {                     \
240            if (!IS_INTERLACED(type)) {         \
241                refn <<= 1;                     \
242                AV_COPY32(mvbuf[idx], mvn);     \
243                mvbuf[idx][1] /= 2;             \
244                mvn = mvbuf[idx];               \
245            }                                   \
246        } else {                                \
247            if (IS_INTERLACED(type)) {          \
248                refn >>= 1;                     \
249                AV_COPY32(mvbuf[idx], mvn);     \
250                mvbuf[idx][1] *= 2;             \
251                mvn = mvbuf[idx];               \
252            }                                   \
253        }                                       \
254    }
255
256static av_always_inline void pred_pskip_motion(const H264Context *const h,
257                                               H264SliceContext *sl)
258{
259    DECLARE_ALIGNED(4, static const int16_t, zeromv)[2] = { 0 };
260    DECLARE_ALIGNED(4, int16_t, mvbuf)[3][2];
261    int8_t *ref     = h->cur_pic.ref_index[0];
262    int16_t(*mv)[2] = h->cur_pic.motion_val[0];
263    int top_ref, left_ref, diagonal_ref, match_count, mx, my;
264    const int16_t *A, *B, *C;
265    int b_stride = h->b_stride;
266
267    fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, 0, 1);
268
269    /* To avoid doing an entire fill_decode_caches, we inline the relevant
270     * parts here.
271     * FIXME: this is a partial duplicate of the logic in fill_decode_caches,
272     * but it's faster this way.  Is there a way to avoid this duplication?
273     */
274    if (USES_LIST(sl->left_type[LTOP], 0)) {
275        left_ref = ref[4 * sl->left_mb_xy[LTOP] + 1 + (sl->left_block[0] & ~1)];
276        A = mv[h->mb2b_xy[sl->left_mb_xy[LTOP]] + 3 + b_stride * sl->left_block[0]];
277        FIX_MV_MBAFF(sl->left_type[LTOP], left_ref, A, 0);
278        if (!(left_ref | AV_RN32A(A)))
279            goto zeromv;
280    } else if (sl->left_type[LTOP]) {
281        left_ref = LIST_NOT_USED;
282        A        = zeromv;
283    } else {
284        goto zeromv;
285    }
286
287    if (USES_LIST(sl->top_type, 0)) {
288        top_ref = ref[4 * sl->top_mb_xy + 2];
289        B       = mv[h->mb2b_xy[sl->top_mb_xy] + 3 * b_stride];
290        FIX_MV_MBAFF(sl->top_type, top_ref, B, 1);
291        if (!(top_ref | AV_RN32A(B)))
292            goto zeromv;
293    } else if (sl->top_type) {
294        top_ref = LIST_NOT_USED;
295        B       = zeromv;
296    } else {
297        goto zeromv;
298    }
299
300    ff_tlog(h->avctx, "pred_pskip: (%d) (%d) at %2d %2d\n",
301            top_ref, left_ref, sl->mb_x, sl->mb_y);
302
303    if (USES_LIST(sl->topright_type, 0)) {
304        diagonal_ref = ref[4 * sl->topright_mb_xy + 2];
305        C = mv[h->mb2b_xy[sl->topright_mb_xy] + 3 * b_stride];
306        FIX_MV_MBAFF(sl->topright_type, diagonal_ref, C, 2);
307    } else if (sl->topright_type) {
308        diagonal_ref = LIST_NOT_USED;
309        C = zeromv;
310    } else {
311        if (USES_LIST(sl->topleft_type, 0)) {
312            diagonal_ref = ref[4 * sl->topleft_mb_xy + 1 +
313                               (sl->topleft_partition & 2)];
314            C = mv[h->mb2b_xy[sl->topleft_mb_xy] + 3 + b_stride +
315                   (sl->topleft_partition & 2 * b_stride)];
316            FIX_MV_MBAFF(sl->topleft_type, diagonal_ref, C, 2);
317        } else if (sl->topleft_type) {
318            diagonal_ref = LIST_NOT_USED;
319            C            = zeromv;
320        } else {
321            diagonal_ref = PART_NOT_AVAILABLE;
322            C            = zeromv;
323        }
324    }
325
326    match_count = !diagonal_ref + !top_ref + !left_ref;
327    ff_tlog(h->avctx, "pred_pskip_motion match_count=%d\n", match_count);
328    if (match_count > 1) {
329        mx = mid_pred(A[0], B[0], C[0]);
330        my = mid_pred(A[1], B[1], C[1]);
331    } else if (match_count == 1) {
332        if (!left_ref) {
333            mx = A[0];
334            my = A[1];
335        } else if (!top_ref) {
336            mx = B[0];
337            my = B[1];
338        } else {
339            mx = C[0];
340            my = C[1];
341        }
342    } else {
343        mx = mid_pred(A[0], B[0], C[0]);
344        my = mid_pred(A[1], B[1], C[1]);
345    }
346
347    fill_rectangle(sl->mv_cache[0][scan8[0]], 4, 4, 8, pack16to32(mx, my), 4);
348    return;
349
350zeromv:
351    fill_rectangle(sl->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
352    return;
353}
354
355static void fill_decode_neighbors(const H264Context *h, H264SliceContext *sl, int mb_type)
356{
357    const int mb_xy = sl->mb_xy;
358    int topleft_xy, top_xy, topright_xy, left_xy[LEFT_MBS];
359    static const uint8_t left_block_options[4][32] = {
360        { 0, 1, 2, 3, 7, 10, 8, 11, 3 + 0 * 4, 3 + 1 * 4, 3 + 2 * 4, 3 + 3 * 4, 1 + 4 * 4, 1 + 8 * 4, 1 + 5 * 4, 1 + 9 * 4 },
361        { 2, 2, 3, 3, 8, 11, 8, 11, 3 + 2 * 4, 3 + 2 * 4, 3 + 3 * 4, 3 + 3 * 4, 1 + 5 * 4, 1 + 9 * 4, 1 + 5 * 4, 1 + 9 * 4 },
362        { 0, 0, 1, 1, 7, 10, 7, 10, 3 + 0 * 4, 3 + 0 * 4, 3 + 1 * 4, 3 + 1 * 4, 1 + 4 * 4, 1 + 8 * 4, 1 + 4 * 4, 1 + 8 * 4 },
363        { 0, 2, 0, 2, 7, 10, 7, 10, 3 + 0 * 4, 3 + 2 * 4, 3 + 0 * 4, 3 + 2 * 4, 1 + 4 * 4, 1 + 8 * 4, 1 + 4 * 4, 1 + 8 * 4 }
364    };
365
366    sl->topleft_partition = -1;
367
368    top_xy = mb_xy - (h->mb_stride << MB_FIELD(sl));
369
370    /* Wow, what a mess, why didn't they simplify the interlacing & intra
371     * stuff, I can't imagine that these complex rules are worth it. */
372
373    topleft_xy    = top_xy - 1;
374    topright_xy   = top_xy + 1;
375    left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
376    sl->left_block = left_block_options[0];
377    if (FRAME_MBAFF(h)) {
378        const int left_mb_field_flag = IS_INTERLACED(h->cur_pic.mb_type[mb_xy - 1]);
379        const int curr_mb_field_flag = IS_INTERLACED(mb_type);
380        if (sl->mb_y & 1) {
381            if (left_mb_field_flag != curr_mb_field_flag) {
382                left_xy[LBOT] = left_xy[LTOP] = mb_xy - h->mb_stride - 1;
383                if (curr_mb_field_flag) {
384                    left_xy[LBOT] += h->mb_stride;
385                    sl->left_block  = left_block_options[3];
386                } else {
387                    topleft_xy += h->mb_stride;
388                    /* take top left mv from the middle of the mb, as opposed
389                     * to all other modes which use the bottom right partition */
390                    sl->topleft_partition = 0;
391                    sl->left_block        = left_block_options[1];
392                }
393            }
394        } else {
395            if (curr_mb_field_flag) {
396                topleft_xy  += h->mb_stride & (((h->cur_pic.mb_type[top_xy - 1] >> 7) & 1) - 1);
397                topright_xy += h->mb_stride & (((h->cur_pic.mb_type[top_xy + 1] >> 7) & 1) - 1);
398                top_xy      += h->mb_stride & (((h->cur_pic.mb_type[top_xy]     >> 7) & 1) - 1);
399            }
400            if (left_mb_field_flag != curr_mb_field_flag) {
401                if (curr_mb_field_flag) {
402                    left_xy[LBOT] += h->mb_stride;
403                    sl->left_block  = left_block_options[3];
404                } else {
405                    sl->left_block = left_block_options[2];
406                }
407            }
408        }
409    }
410
411    sl->topleft_mb_xy    = topleft_xy;
412    sl->top_mb_xy        = top_xy;
413    sl->topright_mb_xy   = topright_xy;
414    sl->left_mb_xy[LTOP] = left_xy[LTOP];
415    sl->left_mb_xy[LBOT] = left_xy[LBOT];
416    //FIXME do we need all in the context?
417
418    sl->topleft_type    = h->cur_pic.mb_type[topleft_xy];
419    sl->top_type        = h->cur_pic.mb_type[top_xy];
420    sl->topright_type   = h->cur_pic.mb_type[topright_xy];
421    sl->left_type[LTOP] = h->cur_pic.mb_type[left_xy[LTOP]];
422    sl->left_type[LBOT] = h->cur_pic.mb_type[left_xy[LBOT]];
423
424    if (FMO) {
425        if (h->slice_table[topleft_xy] != sl->slice_num)
426            sl->topleft_type = 0;
427        if (h->slice_table[top_xy] != sl->slice_num)
428            sl->top_type = 0;
429        if (h->slice_table[left_xy[LTOP]] != sl->slice_num)
430            sl->left_type[LTOP] = sl->left_type[LBOT] = 0;
431    } else {
432        if (h->slice_table[topleft_xy] != sl->slice_num) {
433            sl->topleft_type = 0;
434            if (h->slice_table[top_xy] != sl->slice_num)
435                sl->top_type = 0;
436            if (h->slice_table[left_xy[LTOP]] != sl->slice_num)
437                sl->left_type[LTOP] = sl->left_type[LBOT] = 0;
438        }
439    }
440    if (h->slice_table[topright_xy] != sl->slice_num)
441        sl->topright_type = 0;
442}
443
444static void fill_decode_caches(const H264Context *h, H264SliceContext *sl, int mb_type)
445{
446    int topleft_xy, top_xy, topright_xy, left_xy[LEFT_MBS];
447    int topleft_type, top_type, topright_type, left_type[LEFT_MBS];
448    const uint8_t *left_block = sl->left_block;
449    int i;
450    uint8_t *nnz;
451    uint8_t *nnz_cache;
452
453    topleft_xy      = sl->topleft_mb_xy;
454    top_xy          = sl->top_mb_xy;
455    topright_xy     = sl->topright_mb_xy;
456    left_xy[LTOP]   = sl->left_mb_xy[LTOP];
457    left_xy[LBOT]   = sl->left_mb_xy[LBOT];
458    topleft_type    = sl->topleft_type;
459    top_type        = sl->top_type;
460    topright_type   = sl->topright_type;
461    left_type[LTOP] = sl->left_type[LTOP];
462    left_type[LBOT] = sl->left_type[LBOT];
463
464    if (!IS_SKIP(mb_type)) {
465        if (IS_INTRA(mb_type)) {
466            int type_mask = h->ps.pps->constrained_intra_pred ? IS_INTRA(-1) : -1;
467            sl->topleft_samples_available     =
468                sl->top_samples_available     =
469                    sl->left_samples_available = 0xFFFF;
470            sl->topright_samples_available     = 0xEEEA;
471
472            if (!(top_type & type_mask)) {
473                sl->topleft_samples_available  = 0xB3FF;
474                sl->top_samples_available      = 0x33FF;
475                sl->topright_samples_available = 0x26EA;
476            }
477            if (IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[LTOP])) {
478                if (IS_INTERLACED(mb_type)) {
479                    if (!(left_type[LTOP] & type_mask)) {
480                        sl->topleft_samples_available &= 0xDFFF;
481                        sl->left_samples_available    &= 0x5FFF;
482                    }
483                    if (!(left_type[LBOT] & type_mask)) {
484                        sl->topleft_samples_available &= 0xFF5F;
485                        sl->left_samples_available    &= 0xFF5F;
486                    }
487                } else {
488                    int left_typei = h->cur_pic.mb_type[left_xy[LTOP] + h->mb_stride];
489
490                    av_assert2(left_xy[LTOP] == left_xy[LBOT]);
491                    if (!((left_typei & type_mask) && (left_type[LTOP] & type_mask))) {
492                        sl->topleft_samples_available &= 0xDF5F;
493                        sl->left_samples_available    &= 0x5F5F;
494                    }
495                }
496            } else {
497                if (!(left_type[LTOP] & type_mask)) {
498                    sl->topleft_samples_available &= 0xDF5F;
499                    sl->left_samples_available    &= 0x5F5F;
500                }
501            }
502
503            if (!(topleft_type & type_mask))
504                sl->topleft_samples_available &= 0x7FFF;
505
506            if (!(topright_type & type_mask))
507                sl->topright_samples_available &= 0xFBFF;
508
509            if (IS_INTRA4x4(mb_type)) {
510                if (IS_INTRA4x4(top_type)) {
511                    AV_COPY32(sl->intra4x4_pred_mode_cache + 4 + 8 * 0, sl->intra4x4_pred_mode + h->mb2br_xy[top_xy]);
512                } else {
513                    sl->intra4x4_pred_mode_cache[4 + 8 * 0] =
514                    sl->intra4x4_pred_mode_cache[5 + 8 * 0] =
515                    sl->intra4x4_pred_mode_cache[6 + 8 * 0] =
516                    sl->intra4x4_pred_mode_cache[7 + 8 * 0] = 2 - 3 * !(top_type & type_mask);
517                }
518                for (i = 0; i < 2; i++) {
519                    if (IS_INTRA4x4(left_type[LEFT(i)])) {
520                        int8_t *mode = sl->intra4x4_pred_mode + h->mb2br_xy[left_xy[LEFT(i)]];
521                        sl->intra4x4_pred_mode_cache[3 + 8 * 1 + 2 * 8 * i] = mode[6 - left_block[0 + 2 * i]];
522                        sl->intra4x4_pred_mode_cache[3 + 8 * 2 + 2 * 8 * i] = mode[6 - left_block[1 + 2 * i]];
523                    } else {
524                        sl->intra4x4_pred_mode_cache[3 + 8 * 1 + 2 * 8 * i] =
525                        sl->intra4x4_pred_mode_cache[3 + 8 * 2 + 2 * 8 * i] = 2 - 3 * !(left_type[LEFT(i)] & type_mask);
526                    }
527                }
528            }
529        }
530
531        /*
532         * 0 . T T. T T T T
533         * 1 L . .L . . . .
534         * 2 L . .L . . . .
535         * 3 . T TL . . . .
536         * 4 L . .L . . . .
537         * 5 L . .. . . . .
538         */
539        /* FIXME: constraint_intra_pred & partitioning & nnz
540         * (let us hope this is just a typo in the spec) */
541        nnz_cache = sl->non_zero_count_cache;
542        if (top_type) {
543            nnz = h->non_zero_count[top_xy];
544            AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[4 * 3]);
545            if (!h->chroma_y_shift) {
546                AV_COPY32(&nnz_cache[4 + 8 *  5], &nnz[4 *  7]);
547                AV_COPY32(&nnz_cache[4 + 8 * 10], &nnz[4 * 11]);
548            } else {
549                AV_COPY32(&nnz_cache[4 + 8 *  5], &nnz[4 * 5]);
550                AV_COPY32(&nnz_cache[4 + 8 * 10], &nnz[4 * 9]);
551            }
552        } else {
553            uint32_t top_empty = CABAC(h) && !IS_INTRA(mb_type) ? 0 : 0x40404040;
554            AV_WN32A(&nnz_cache[4 + 8 *  0], top_empty);
555            AV_WN32A(&nnz_cache[4 + 8 *  5], top_empty);
556            AV_WN32A(&nnz_cache[4 + 8 * 10], top_empty);
557        }
558
559        for (i = 0; i < 2; i++) {
560            if (left_type[LEFT(i)]) {
561                nnz = h->non_zero_count[left_xy[LEFT(i)]];
562                nnz_cache[3 + 8 * 1 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i]];
563                nnz_cache[3 + 8 * 2 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i]];
564                if (CHROMA444(h)) {
565                    nnz_cache[3 + 8 *  6 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] + 4 * 4];
566                    nnz_cache[3 + 8 *  7 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] + 4 * 4];
567                    nnz_cache[3 + 8 * 11 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] + 8 * 4];
568                    nnz_cache[3 + 8 * 12 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] + 8 * 4];
569                } else if (CHROMA422(h)) {
570                    nnz_cache[3 + 8 *  6 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] - 2 + 4 * 4];
571                    nnz_cache[3 + 8 *  7 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] - 2 + 4 * 4];
572                    nnz_cache[3 + 8 * 11 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] - 2 + 8 * 4];
573                    nnz_cache[3 + 8 * 12 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] - 2 + 8 * 4];
574                } else {
575                    nnz_cache[3 + 8 *  6 + 8 * i] = nnz[left_block[8 + 4 + 2 * i]];
576                    nnz_cache[3 + 8 * 11 + 8 * i] = nnz[left_block[8 + 5 + 2 * i]];
577                }
578            } else {
579                nnz_cache[3 + 8 *  1 + 2 * 8 * i] =
580                nnz_cache[3 + 8 *  2 + 2 * 8 * i] =
581                nnz_cache[3 + 8 *  6 + 2 * 8 * i] =
582                nnz_cache[3 + 8 *  7 + 2 * 8 * i] =
583                nnz_cache[3 + 8 * 11 + 2 * 8 * i] =
584                nnz_cache[3 + 8 * 12 + 2 * 8 * i] = CABAC(h) && !IS_INTRA(mb_type) ? 0 : 64;
585            }
586        }
587
588        if (CABAC(h)) {
589            // top_cbp
590            if (top_type)
591                sl->top_cbp = h->cbp_table[top_xy];
592            else
593                sl->top_cbp = IS_INTRA(mb_type) ? 0x7CF : 0x00F;
594            // left_cbp
595            if (left_type[LTOP]) {
596                sl->left_cbp =   (h->cbp_table[left_xy[LTOP]] & 0x7F0) |
597                               ((h->cbp_table[left_xy[LTOP]] >> (left_block[0] & (~1))) & 2) |
598                              (((h->cbp_table[left_xy[LBOT]] >> (left_block[2] & (~1))) & 2) << 2);
599            } else {
600                sl->left_cbp = IS_INTRA(mb_type) ? 0x7CF : 0x00F;
601            }
602        }
603    }
604
605    if (IS_INTER(mb_type) || (IS_DIRECT(mb_type) && sl->direct_spatial_mv_pred)) {
606        int list;
607        int b_stride = h->b_stride;
608        for (list = 0; list < sl->list_count; list++) {
609            int8_t *ref_cache = &sl->ref_cache[list][scan8[0]];
610            int8_t *ref       = h->cur_pic.ref_index[list];
611            int16_t(*mv_cache)[2] = &sl->mv_cache[list][scan8[0]];
612            int16_t(*mv)[2]       = h->cur_pic.motion_val[list];
613            if (!USES_LIST(mb_type, list))
614                continue;
615            av_assert2(!(IS_DIRECT(mb_type) && !sl->direct_spatial_mv_pred));
616
617            if (USES_LIST(top_type, list)) {
618                const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
619                AV_COPY128(mv_cache[0 - 1 * 8], mv[b_xy + 0]);
620                ref_cache[0 - 1 * 8] =
621                ref_cache[1 - 1 * 8] = ref[4 * top_xy + 2];
622                ref_cache[2 - 1 * 8] =
623                ref_cache[3 - 1 * 8] = ref[4 * top_xy + 3];
624            } else {
625                AV_ZERO128(mv_cache[0 - 1 * 8]);
626                AV_WN32A(&ref_cache[0 - 1 * 8],
627                         ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE) & 0xFF) * 0x01010101u);
628            }
629
630            if (mb_type & (MB_TYPE_16x8 | MB_TYPE_8x8)) {
631                for (i = 0; i < 2; i++) {
632                    int cache_idx = -1 + i * 2 * 8;
633                    if (USES_LIST(left_type[LEFT(i)], list)) {
634                        const int b_xy  = h->mb2b_xy[left_xy[LEFT(i)]] + 3;
635                        const int b8_xy = 4 * left_xy[LEFT(i)] + 1;
636                        AV_COPY32(mv_cache[cache_idx],
637                                  mv[b_xy + b_stride * left_block[0 + i * 2]]);
638                        AV_COPY32(mv_cache[cache_idx + 8],
639                                  mv[b_xy + b_stride * left_block[1 + i * 2]]);
640                        ref_cache[cache_idx]     = ref[b8_xy + (left_block[0 + i * 2] & ~1)];
641                        ref_cache[cache_idx + 8] = ref[b8_xy + (left_block[1 + i * 2] & ~1)];
642                    } else {
643                        AV_ZERO32(mv_cache[cache_idx]);
644                        AV_ZERO32(mv_cache[cache_idx + 8]);
645                        ref_cache[cache_idx]     =
646                        ref_cache[cache_idx + 8] = (left_type[LEFT(i)]) ? LIST_NOT_USED
647                                                                        : PART_NOT_AVAILABLE;
648                    }
649                }
650            } else {
651                if (USES_LIST(left_type[LTOP], list)) {
652                    const int b_xy  = h->mb2b_xy[left_xy[LTOP]] + 3;
653                    const int b8_xy = 4 * left_xy[LTOP] + 1;
654                    AV_COPY32(mv_cache[-1], mv[b_xy + b_stride * left_block[0]]);
655                    ref_cache[-1] = ref[b8_xy + (left_block[0] & ~1)];
656                } else {
657                    AV_ZERO32(mv_cache[-1]);
658                    ref_cache[-1] = left_type[LTOP] ? LIST_NOT_USED
659                                                    : PART_NOT_AVAILABLE;
660                }
661            }
662
663            if (USES_LIST(topright_type, list)) {
664                const int b_xy = h->mb2b_xy[topright_xy] + 3 * b_stride;
665                AV_COPY32(mv_cache[4 - 1 * 8], mv[b_xy]);
666                ref_cache[4 - 1 * 8] = ref[4 * topright_xy + 2];
667            } else {
668                AV_ZERO32(mv_cache[4 - 1 * 8]);
669                ref_cache[4 - 1 * 8] = topright_type ? LIST_NOT_USED
670                                                     : PART_NOT_AVAILABLE;
671            }
672            if(ref_cache[2 - 1*8] < 0 || ref_cache[4 - 1 * 8] < 0) {
673                if (USES_LIST(topleft_type, list)) {
674                    const int b_xy  = h->mb2b_xy[topleft_xy] + 3 + b_stride +
675                                      (sl->topleft_partition & 2 * b_stride);
676                    const int b8_xy = 4 * topleft_xy + 1 + (sl->topleft_partition & 2);
677                    AV_COPY32(mv_cache[-1 - 1 * 8], mv[b_xy]);
678                    ref_cache[-1 - 1 * 8] = ref[b8_xy];
679                } else {
680                    AV_ZERO32(mv_cache[-1 - 1 * 8]);
681                    ref_cache[-1 - 1 * 8] = topleft_type ? LIST_NOT_USED
682                                                         : PART_NOT_AVAILABLE;
683                }
684            }
685
686            if ((mb_type & (MB_TYPE_SKIP | MB_TYPE_DIRECT2)) && !FRAME_MBAFF(h))
687                continue;
688
689            if (!(mb_type & (MB_TYPE_SKIP | MB_TYPE_DIRECT2))) {
690                uint8_t(*mvd_cache)[2]   = &sl->mvd_cache[list][scan8[0]];
691                uint8_t(*mvd)[2]         = sl->mvd_table[list];
692                ref_cache[2 + 8 * 0] =
693                ref_cache[2 + 8 * 2] = PART_NOT_AVAILABLE;
694                AV_ZERO32(mv_cache[2 + 8 * 0]);
695                AV_ZERO32(mv_cache[2 + 8 * 2]);
696
697                if (CABAC(h)) {
698                    if (USES_LIST(top_type, list)) {
699                        const int b_xy = h->mb2br_xy[top_xy];
700                        AV_COPY64(mvd_cache[0 - 1 * 8], mvd[b_xy + 0]);
701                    } else {
702                        AV_ZERO64(mvd_cache[0 - 1 * 8]);
703                    }
704                    if (USES_LIST(left_type[LTOP], list)) {
705                        const int b_xy = h->mb2br_xy[left_xy[LTOP]] + 6;
706                        AV_COPY16(mvd_cache[-1 + 0 * 8], mvd[b_xy - left_block[0]]);
707                        AV_COPY16(mvd_cache[-1 + 1 * 8], mvd[b_xy - left_block[1]]);
708                    } else {
709                        AV_ZERO16(mvd_cache[-1 + 0 * 8]);
710                        AV_ZERO16(mvd_cache[-1 + 1 * 8]);
711                    }
712                    if (USES_LIST(left_type[LBOT], list)) {
713                        const int b_xy = h->mb2br_xy[left_xy[LBOT]] + 6;
714                        AV_COPY16(mvd_cache[-1 + 2 * 8], mvd[b_xy - left_block[2]]);
715                        AV_COPY16(mvd_cache[-1 + 3 * 8], mvd[b_xy - left_block[3]]);
716                    } else {
717                        AV_ZERO16(mvd_cache[-1 + 2 * 8]);
718                        AV_ZERO16(mvd_cache[-1 + 3 * 8]);
719                    }
720                    AV_ZERO16(mvd_cache[2 + 8 * 0]);
721                    AV_ZERO16(mvd_cache[2 + 8 * 2]);
722                    if (sl->slice_type_nos == AV_PICTURE_TYPE_B) {
723                        uint8_t *direct_cache = &sl->direct_cache[scan8[0]];
724                        uint8_t *direct_table = h->direct_table;
725                        fill_rectangle(direct_cache, 4, 4, 8, MB_TYPE_16x16 >> 1, 1);
726
727                        if (IS_DIRECT(top_type)) {
728                            AV_WN32A(&direct_cache[-1 * 8],
729                                     0x01010101u * (MB_TYPE_DIRECT2 >> 1));
730                        } else if (IS_8X8(top_type)) {
731                            int b8_xy = 4 * top_xy;
732                            direct_cache[0 - 1 * 8] = direct_table[b8_xy + 2];
733                            direct_cache[2 - 1 * 8] = direct_table[b8_xy + 3];
734                        } else {
735                            AV_WN32A(&direct_cache[-1 * 8],
736                                     0x01010101 * (MB_TYPE_16x16 >> 1));
737                        }
738
739                        if (IS_DIRECT(left_type[LTOP]))
740                            direct_cache[-1 + 0 * 8] = MB_TYPE_DIRECT2 >> 1;
741                        else if (IS_8X8(left_type[LTOP]))
742                            direct_cache[-1 + 0 * 8] = direct_table[4 * left_xy[LTOP] + 1 + (left_block[0] & ~1)];
743                        else
744                            direct_cache[-1 + 0 * 8] = MB_TYPE_16x16 >> 1;
745
746                        if (IS_DIRECT(left_type[LBOT]))
747                            direct_cache[-1 + 2 * 8] = MB_TYPE_DIRECT2 >> 1;
748                        else if (IS_8X8(left_type[LBOT]))
749                            direct_cache[-1 + 2 * 8] = direct_table[4 * left_xy[LBOT] + 1 + (left_block[2] & ~1)];
750                        else
751                            direct_cache[-1 + 2 * 8] = MB_TYPE_16x16 >> 1;
752                    }
753                }
754            }
755
756#define MAP_MVS                                                         \
757    MAP_F2F(scan8[0] - 1 - 1 * 8, topleft_type)                         \
758    MAP_F2F(scan8[0] + 0 - 1 * 8, top_type)                             \
759    MAP_F2F(scan8[0] + 1 - 1 * 8, top_type)                             \
760    MAP_F2F(scan8[0] + 2 - 1 * 8, top_type)                             \
761    MAP_F2F(scan8[0] + 3 - 1 * 8, top_type)                             \
762    MAP_F2F(scan8[0] + 4 - 1 * 8, topright_type)                        \
763    MAP_F2F(scan8[0] - 1 + 0 * 8, left_type[LTOP])                      \
764    MAP_F2F(scan8[0] - 1 + 1 * 8, left_type[LTOP])                      \
765    MAP_F2F(scan8[0] - 1 + 2 * 8, left_type[LBOT])                      \
766    MAP_F2F(scan8[0] - 1 + 3 * 8, left_type[LBOT])
767
768            if (FRAME_MBAFF(h)) {
769                if (MB_FIELD(sl)) {
770
771#define MAP_F2F(idx, mb_type)                                           \
772    if (!IS_INTERLACED(mb_type) && sl->ref_cache[list][idx] >= 0) {     \
773        sl->ref_cache[list][idx]     *= 2;                              \
774        sl->mv_cache[list][idx][1]   /= 2;                              \
775        sl->mvd_cache[list][idx][1] >>= 1;                              \
776    }
777
778                    MAP_MVS
779                } else {
780
781#undef MAP_F2F
782#define MAP_F2F(idx, mb_type)                                           \
783    if (IS_INTERLACED(mb_type) && sl->ref_cache[list][idx] >= 0) {      \
784        sl->ref_cache[list][idx]    >>= 1;                              \
785        sl->mv_cache[list][idx][1]   *= 2;                              \
786        sl->mvd_cache[list][idx][1] <<= 1;                              \
787    }
788
789                    MAP_MVS
790#undef MAP_F2F
791                }
792            }
793        }
794    }
795
796    sl->neighbor_transform_size = !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[LTOP]);
797}
798
799/**
800 * decodes a P_SKIP or B_SKIP macroblock
801 */
802static void av_unused decode_mb_skip(const H264Context *h, H264SliceContext *sl)
803{
804    const int mb_xy = sl->mb_xy;
805    int mb_type     = 0;
806
807    memset(h->non_zero_count[mb_xy], 0, 48);
808
809    if (MB_FIELD(sl))
810        mb_type |= MB_TYPE_INTERLACED;
811
812    if (sl->slice_type_nos == AV_PICTURE_TYPE_B) {
813        // just for fill_caches. pred_direct_motion will set the real mb_type
814        mb_type |= MB_TYPE_L0L1 | MB_TYPE_DIRECT2 | MB_TYPE_SKIP;
815        if (sl->direct_spatial_mv_pred) {
816            fill_decode_neighbors(h, sl, mb_type);
817            fill_decode_caches(h, sl, mb_type); //FIXME check what is needed and what not ...
818        }
819        ff_h264_pred_direct_motion(h, sl, &mb_type);
820        mb_type |= MB_TYPE_SKIP;
821    } else {
822        mb_type |= MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P1L0 | MB_TYPE_SKIP;
823
824        fill_decode_neighbors(h, sl, mb_type);
825        pred_pskip_motion(h, sl);
826    }
827
828    write_back_motion(h, sl, mb_type);
829    h->cur_pic.mb_type[mb_xy]      = mb_type;
830    h->cur_pic.qscale_table[mb_xy] = sl->qscale;
831    h->slice_table[mb_xy]          = sl->slice_num;
832    sl->prev_mb_skipped            = 1;
833}
834
835#endif /* AVCODEC_H264_MVPRED_H */
836