FFmpeg
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
h264_mvpred.h
Go to the documentation of this file.
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... motion vector predicion
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 / MPEG4 part10 motion vector predicion.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27 
28 #ifndef AVCODEC_H264_MVPRED_H
29 #define AVCODEC_H264_MVPRED_H
30 
31 #include "internal.h"
32 #include "avcodec.h"
33 #include "h264.h"
34 #include "libavutil/avassert.h"
35 
36 
37 static av_always_inline int fetch_diagonal_mv(H264Context *h, const int16_t **C,
38  int i, int list, int part_width)
39 {
40  const int topright_ref = h->ref_cache[list][i - 8 + part_width];
41  MpegEncContext *s = &h->s;
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) {
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) * s->mb_stride]; \
49  if (!USES_LIST(mb_type, list)) \
50  return LIST_NOT_USED; \
51  mv = s->current_picture_ptr->f.motion_val[list][h->mb2b_xy[xy] + 3 + y4 * h->b_stride]; \
52  h->mv_cache[list][scan8[0] - 2][0] = mv[0]; \
53  h->mv_cache[list][scan8[0] - 2][1] = mv[1] MV_OP; \
54  return s->current_picture_ptr->f.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  && h->ref_cache[list][scan8[0] - 1] != PART_NOT_AVAILABLE) {
59  const uint32_t *mb_types = s->current_picture_ptr->f.mb_type;
60  const int16_t *mv;
61  AV_ZERO32(h->mv_cache[list][scan8[0] - 2]);
62  *C = h->mv_cache[list][scan8[0] - 2];
63 
64  if (!MB_FIELD && IS_INTERLACED(h->left_type[0])) {
65  SET_DIAG_MV(* 2, >> 1, h->left_mb_xy[0] + s->mb_stride,
66  (s->mb_y & 1) * 2 + (i >> 5));
67  }
68  if (MB_FIELD && !IS_INTERLACED(h->left_type[0])) {
69  // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
70  SET_DIAG_MV(/ 2, << 1, h->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 = h->mv_cache[list][i - 8 + part_width];
78  return topright_ref;
79  } else {
80  tprintf(s->avctx, "topright MV not available\n");
81 
82  *C = h->mv_cache[list][i - 8 - 1];
83  return h->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  */
94 static av_always_inline void pred_motion(H264Context *const h, int n,
95  int part_width, int list, int ref,
96  int *const mx, int *const my)
97 {
98  const int index8 = scan8[n];
99  const int top_ref = h->ref_cache[list][index8 - 8];
100  const int left_ref = h->ref_cache[list][index8 - 1];
101  const int16_t *const A = h->mv_cache[list][index8 - 1];
102  const int16_t *const B = h->mv_cache[list][index8 - 8];
103  const int16_t *C;
104  int diagonal_ref, match_count;
105 
106  av_assert2(part_width == 1 || part_width == 2 || part_width == 4);
107 
108 /* mv_cache
109  * B . . A T T T T
110  * U . . L . . , .
111  * U . . L . . . .
112  * U . . L . . , .
113  * . . . L . . . .
114  */
115 
116  diagonal_ref = fetch_diagonal_mv(h, &C, index8, list, part_width);
117  match_count = (diagonal_ref == ref) + (top_ref == ref) + (left_ref == ref);
118  tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
119  if (match_count > 1) { //most common
120  *mx = mid_pred(A[0], B[0], C[0]);
121  *my = mid_pred(A[1], B[1], C[1]);
122  } else if (match_count == 1) {
123  if (left_ref == ref) {
124  *mx = A[0];
125  *my = A[1];
126  } else if (top_ref == ref) {
127  *mx = B[0];
128  *my = B[1];
129  } else {
130  *mx = C[0];
131  *my = C[1];
132  }
133  } else {
134  if (top_ref == PART_NOT_AVAILABLE &&
135  diagonal_ref == PART_NOT_AVAILABLE &&
136  left_ref != PART_NOT_AVAILABLE) {
137  *mx = A[0];
138  *my = A[1];
139  } else {
140  *mx = mid_pred(A[0], B[0], C[0]);
141  *my = mid_pred(A[1], B[1], C[1]);
142  }
143  }
144 
145  tprintf(h->s.avctx,
146  "pred_motion (%2d %2d %2d) (%2d %2d %2d) (%2d %2d %2d) -> (%2d %2d %2d) at %2d %2d %d list %d\n",
147  top_ref, B[0], B[1], diagonal_ref, C[0], C[1], left_ref,
148  A[0], A[1], ref, *mx, *my, h->s.mb_x, h->s.mb_y, n, list);
149 }
150 
151 /**
152  * Get the directionally predicted 16x8 MV.
153  * @param n the block index
154  * @param mx the x component of the predicted motion vector
155  * @param my the y component of the predicted motion vector
156  */
158  int n, int list, int ref,
159  int *const mx, int *const my)
160 {
161  if (n == 0) {
162  const int top_ref = h->ref_cache[list][scan8[0] - 8];
163  const int16_t *const B = h->mv_cache[list][scan8[0] - 8];
164 
165  tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n",
166  top_ref, B[0], B[1], h->s.mb_x, h->s.mb_y, n, list);
167 
168  if (top_ref == ref) {
169  *mx = B[0];
170  *my = B[1];
171  return;
172  }
173  } else {
174  const int left_ref = h->ref_cache[list][scan8[8] - 1];
175  const int16_t *const A = h->mv_cache[list][scan8[8] - 1];
176 
177  tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n",
178  left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
179 
180  if (left_ref == ref) {
181  *mx = A[0];
182  *my = A[1];
183  return;
184  }
185  }
186 
187  //RARE
188  pred_motion(h, n, 4, list, ref, mx, my);
189 }
190 
191 /**
192  * Get the directionally predicted 8x16 MV.
193  * @param n the block index
194  * @param mx the x component of the predicted motion vector
195  * @param my the y component of the predicted motion vector
196  */
198  int n, int list, int ref,
199  int *const mx, int *const my)
200 {
201  if (n == 0) {
202  const int left_ref = h->ref_cache[list][scan8[0] - 1];
203  const int16_t *const A = h->mv_cache[list][scan8[0] - 1];
204 
205  tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n",
206  left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
207 
208  if (left_ref == ref) {
209  *mx = A[0];
210  *my = A[1];
211  return;
212  }
213  } else {
214  const int16_t *C;
215  int diagonal_ref;
216 
217  diagonal_ref = fetch_diagonal_mv(h, &C, scan8[4], list, 2);
218 
219  tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n",
220  diagonal_ref, C[0], C[1], h->s.mb_x, h->s.mb_y, n, list);
221 
222  if (diagonal_ref == ref) {
223  *mx = C[0];
224  *my = C[1];
225  return;
226  }
227  }
228 
229  //RARE
230  pred_motion(h, n, 2, list, ref, mx, my);
231 }
232 
233 #define FIX_MV_MBAFF(type, refn, mvn, idx) \
234  if (FRAME_MBAFF) { \
235  if (MB_FIELD) { \
236  if (!IS_INTERLACED(type)) { \
237  refn <<= 1; \
238  AV_COPY32(mvbuf[idx], mvn); \
239  mvbuf[idx][1] /= 2; \
240  mvn = mvbuf[idx]; \
241  } \
242  } else { \
243  if (IS_INTERLACED(type)) { \
244  refn >>= 1; \
245  AV_COPY32(mvbuf[idx], mvn); \
246  mvbuf[idx][1] <<= 1; \
247  mvn = mvbuf[idx]; \
248  } \
249  } \
250  }
251 
253 {
254  DECLARE_ALIGNED(4, static const int16_t, zeromv)[2] = { 0 };
255  DECLARE_ALIGNED(4, int16_t, mvbuf)[3][2];
256  MpegEncContext *const s = &h->s;
257  int8_t *ref = s->current_picture.f.ref_index[0];
258  int16_t(*mv)[2] = s->current_picture.f.motion_val[0];
259  int top_ref, left_ref, diagonal_ref, match_count, mx, my;
260  const int16_t *A, *B, *C;
261  int b_stride = h->b_stride;
262 
263  fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, 0, 1);
264 
265  /* To avoid doing an entire fill_decode_caches, we inline the relevant
266  * parts here.
267  * FIXME: this is a partial duplicate of the logic in fill_decode_caches,
268  * but it's faster this way. Is there a way to avoid this duplication?
269  */
270  if (USES_LIST(h->left_type[LTOP], 0)) {
271  left_ref = ref[4 * h->left_mb_xy[LTOP] + 1 + (h->left_block[0] & ~1)];
272  A = mv[h->mb2b_xy[h->left_mb_xy[LTOP]] + 3 + b_stride * h->left_block[0]];
273  FIX_MV_MBAFF(h->left_type[LTOP], left_ref, A, 0);
274  if (!(left_ref | AV_RN32A(A)))
275  goto zeromv;
276  } else if (h->left_type[LTOP]) {
277  left_ref = LIST_NOT_USED;
278  A = zeromv;
279  } else {
280  goto zeromv;
281  }
282 
283  if (USES_LIST(h->top_type, 0)) {
284  top_ref = ref[4 * h->top_mb_xy + 2];
285  B = mv[h->mb2b_xy[h->top_mb_xy] + 3 * b_stride];
286  FIX_MV_MBAFF(h->top_type, top_ref, B, 1);
287  if (!(top_ref | AV_RN32A(B)))
288  goto zeromv;
289  } else if (h->top_type) {
290  top_ref = LIST_NOT_USED;
291  B = zeromv;
292  } else {
293  goto zeromv;
294  }
295 
296  tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n",
297  top_ref, left_ref, h->s.mb_x, h->s.mb_y);
298 
299  if (USES_LIST(h->topright_type, 0)) {
300  diagonal_ref = ref[4 * h->topright_mb_xy + 2];
301  C = mv[h->mb2b_xy[h->topright_mb_xy] + 3 * b_stride];
302  FIX_MV_MBAFF(h->topright_type, diagonal_ref, C, 2);
303  } else if (h->topright_type) {
304  diagonal_ref = LIST_NOT_USED;
305  C = zeromv;
306  } else {
307  if (USES_LIST(h->topleft_type, 0)) {
308  diagonal_ref = ref[4 * h->topleft_mb_xy + 1 +
309  (h->topleft_partition & 2)];
310  C = mv[h->mb2b_xy[h->topleft_mb_xy] + 3 + b_stride +
311  (h->topleft_partition & 2 * b_stride)];
312  FIX_MV_MBAFF(h->topleft_type, diagonal_ref, C, 2);
313  } else if (h->topleft_type) {
314  diagonal_ref = LIST_NOT_USED;
315  C = zeromv;
316  } else {
317  diagonal_ref = PART_NOT_AVAILABLE;
318  C = zeromv;
319  }
320  }
321 
322  match_count = !diagonal_ref + !top_ref + !left_ref;
323  tprintf(h->s.avctx, "pred_pskip_motion match_count=%d\n", match_count);
324  if (match_count > 1) {
325  mx = mid_pred(A[0], B[0], C[0]);
326  my = mid_pred(A[1], B[1], C[1]);
327  } else if (match_count == 1) {
328  if (!left_ref) {
329  mx = A[0];
330  my = A[1];
331  } else if (!top_ref) {
332  mx = B[0];
333  my = B[1];
334  } else {
335  mx = C[0];
336  my = C[1];
337  }
338  } else {
339  mx = mid_pred(A[0], B[0], C[0]);
340  my = mid_pred(A[1], B[1], C[1]);
341  }
342 
343  fill_rectangle(h->mv_cache[0][scan8[0]], 4, 4, 8, pack16to32(mx, my), 4);
344  return;
345 
346 zeromv:
347  fill_rectangle(h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
348  return;
349 }
350 
351 static void fill_decode_neighbors(H264Context *h, int mb_type)
352 {
353  MpegEncContext *const s = &h->s;
354  const int mb_xy = h->mb_xy;
355  int topleft_xy, top_xy, topright_xy, left_xy[LEFT_MBS];
356  static const uint8_t left_block_options[4][32] = {
357  { 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 },
358  { 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 },
359  { 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 },
360  { 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 }
361  };
362 
363  h->topleft_partition = -1;
364 
365  top_xy = mb_xy - (s->mb_stride << MB_FIELD);
366 
367  /* Wow, what a mess, why didn't they simplify the interlacing & intra
368  * stuff, I can't imagine that these complex rules are worth it. */
369 
370  topleft_xy = top_xy - 1;
371  topright_xy = top_xy + 1;
372  left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
373  h->left_block = left_block_options[0];
374  if (FRAME_MBAFF) {
375  const int left_mb_field_flag = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
376  const int curr_mb_field_flag = IS_INTERLACED(mb_type);
377  if (s->mb_y & 1) {
378  if (left_mb_field_flag != curr_mb_field_flag) {
379  left_xy[LBOT] = left_xy[LTOP] = mb_xy - s->mb_stride - 1;
380  if (curr_mb_field_flag) {
381  left_xy[LBOT] += s->mb_stride;
382  h->left_block = left_block_options[3];
383  } else {
384  topleft_xy += s->mb_stride;
385  /* take top left mv from the middle of the mb, as opposed
386  * to all other modes which use the bottom right partition */
387  h->topleft_partition = 0;
388  h->left_block = left_block_options[1];
389  }
390  }
391  } else {
392  if (curr_mb_field_flag) {
393  topleft_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy - 1] >> 7) & 1) - 1);
394  topright_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy + 1] >> 7) & 1) - 1);
395  top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
396  }
397  if (left_mb_field_flag != curr_mb_field_flag) {
398  if (curr_mb_field_flag) {
399  left_xy[LBOT] += s->mb_stride;
400  h->left_block = left_block_options[3];
401  } else {
402  h->left_block = left_block_options[2];
403  }
404  }
405  }
406  }
407 
408  h->topleft_mb_xy = topleft_xy;
409  h->top_mb_xy = top_xy;
410  h->topright_mb_xy = topright_xy;
411  h->left_mb_xy[LTOP] = left_xy[LTOP];
412  h->left_mb_xy[LBOT] = left_xy[LBOT];
413  //FIXME do we need all in the context?
414 
415  h->topleft_type = s->current_picture.f.mb_type[topleft_xy];
416  h->top_type = s->current_picture.f.mb_type[top_xy];
417  h->topright_type = s->current_picture.f.mb_type[topright_xy];
418  h->left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
419  h->left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
420 
421  if (FMO) {
422  if (h->slice_table[topleft_xy] != h->slice_num)
423  h->topleft_type = 0;
424  if (h->slice_table[top_xy] != h->slice_num)
425  h->top_type = 0;
426  if (h->slice_table[left_xy[LTOP]] != h->slice_num)
427  h->left_type[LTOP] = h->left_type[LBOT] = 0;
428  } else {
429  if (h->slice_table[topleft_xy] != h->slice_num) {
430  h->topleft_type = 0;
431  if (h->slice_table[top_xy] != h->slice_num)
432  h->top_type = 0;
433  if (h->slice_table[left_xy[LTOP]] != h->slice_num)
434  h->left_type[LTOP] = h->left_type[LBOT] = 0;
435  }
436  }
437  if (h->slice_table[topright_xy] != h->slice_num)
438  h->topright_type = 0;
439 }
440 
441 static void fill_decode_caches(H264Context *h, int mb_type)
442 {
443  MpegEncContext *const s = &h->s;
444  int topleft_xy, top_xy, topright_xy, left_xy[LEFT_MBS];
445  int topleft_type, top_type, topright_type, left_type[LEFT_MBS];
446  const uint8_t *left_block = h->left_block;
447  int i;
448  uint8_t *nnz;
449  uint8_t *nnz_cache;
450 
451  topleft_xy = h->topleft_mb_xy;
452  top_xy = h->top_mb_xy;
453  topright_xy = h->topright_mb_xy;
454  left_xy[LTOP] = h->left_mb_xy[LTOP];
455  left_xy[LBOT] = h->left_mb_xy[LBOT];
456  topleft_type = h->topleft_type;
457  top_type = h->top_type;
458  topright_type = h->topright_type;
459  left_type[LTOP] = h->left_type[LTOP];
460  left_type[LBOT] = h->left_type[LBOT];
461 
462  if (!IS_SKIP(mb_type)) {
463  if (IS_INTRA(mb_type)) {
464  int type_mask = h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
467  h->left_samples_available = 0xFFFF;
468  h->topright_samples_available = 0xEEEA;
469 
470  if (!(top_type & type_mask)) {
471  h->topleft_samples_available = 0xB3FF;
472  h->top_samples_available = 0x33FF;
473  h->topright_samples_available = 0x26EA;
474  }
475  if (IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[LTOP])) {
476  if (IS_INTERLACED(mb_type)) {
477  if (!(left_type[LTOP] & type_mask)) {
478  h->topleft_samples_available &= 0xDFFF;
479  h->left_samples_available &= 0x5FFF;
480  }
481  if (!(left_type[LBOT] & type_mask)) {
482  h->topleft_samples_available &= 0xFF5F;
483  h->left_samples_available &= 0xFF5F;
484  }
485  } else {
486  int left_typei = s->current_picture.f.mb_type[left_xy[LTOP] + s->mb_stride];
487 
488  av_assert2(left_xy[LTOP] == left_xy[LBOT]);
489  if (!((left_typei & type_mask) && (left_type[LTOP] & type_mask))) {
490  h->topleft_samples_available &= 0xDF5F;
491  h->left_samples_available &= 0x5F5F;
492  }
493  }
494  } else {
495  if (!(left_type[LTOP] & type_mask)) {
496  h->topleft_samples_available &= 0xDF5F;
497  h->left_samples_available &= 0x5F5F;
498  }
499  }
500 
501  if (!(topleft_type & type_mask))
502  h->topleft_samples_available &= 0x7FFF;
503 
504  if (!(topright_type & type_mask))
505  h->topright_samples_available &= 0xFBFF;
506 
507  if (IS_INTRA4x4(mb_type)) {
508  if (IS_INTRA4x4(top_type)) {
509  AV_COPY32(h->intra4x4_pred_mode_cache + 4 + 8 * 0, h->intra4x4_pred_mode + h->mb2br_xy[top_xy]);
510  } else {
511  h->intra4x4_pred_mode_cache[4 + 8 * 0] =
512  h->intra4x4_pred_mode_cache[5 + 8 * 0] =
513  h->intra4x4_pred_mode_cache[6 + 8 * 0] =
514  h->intra4x4_pred_mode_cache[7 + 8 * 0] = 2 - 3 * !(top_type & type_mask);
515  }
516  for (i = 0; i < 2; i++) {
517  if (IS_INTRA4x4(left_type[LEFT(i)])) {
518  int8_t *mode = h->intra4x4_pred_mode + h->mb2br_xy[left_xy[LEFT(i)]];
519  h->intra4x4_pred_mode_cache[3 + 8 * 1 + 2 * 8 * i] = mode[6 - left_block[0 + 2 * i]];
520  h->intra4x4_pred_mode_cache[3 + 8 * 2 + 2 * 8 * i] = mode[6 - left_block[1 + 2 * i]];
521  } else {
522  h->intra4x4_pred_mode_cache[3 + 8 * 1 + 2 * 8 * i] =
523  h->intra4x4_pred_mode_cache[3 + 8 * 2 + 2 * 8 * i] = 2 - 3 * !(left_type[LEFT(i)] & type_mask);
524  }
525  }
526  }
527  }
528 
529  /*
530  * 0 . T T. T T T T
531  * 1 L . .L . . . .
532  * 2 L . .L . . . .
533  * 3 . T TL . . . .
534  * 4 L . .L . . . .
535  * 5 L . .. . . . .
536  */
537  /* FIXME: constraint_intra_pred & partitioning & nnz
538  * (let us hope this is just a typo in the spec) */
539  nnz_cache = h->non_zero_count_cache;
540  if (top_type) {
541  nnz = h->non_zero_count[top_xy];
542  AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[4 * 3]);
543  if (!s->chroma_y_shift) {
544  AV_COPY32(&nnz_cache[4 + 8 * 5], &nnz[4 * 7]);
545  AV_COPY32(&nnz_cache[4 + 8 * 10], &nnz[4 * 11]);
546  } else {
547  AV_COPY32(&nnz_cache[4 + 8 * 5], &nnz[4 * 5]);
548  AV_COPY32(&nnz_cache[4 + 8 * 10], &nnz[4 * 9]);
549  }
550  } else {
551  uint32_t top_empty = CABAC && !IS_INTRA(mb_type) ? 0 : 0x40404040;
552  AV_WN32A(&nnz_cache[4 + 8 * 0], top_empty);
553  AV_WN32A(&nnz_cache[4 + 8 * 5], top_empty);
554  AV_WN32A(&nnz_cache[4 + 8 * 10], top_empty);
555  }
556 
557  for (i = 0; i < 2; i++) {
558  if (left_type[LEFT(i)]) {
559  nnz = h->non_zero_count[left_xy[LEFT(i)]];
560  nnz_cache[3 + 8 * 1 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i]];
561  nnz_cache[3 + 8 * 2 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i]];
562  if (CHROMA444) {
563  nnz_cache[3 + 8 * 6 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] + 4 * 4];
564  nnz_cache[3 + 8 * 7 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] + 4 * 4];
565  nnz_cache[3 + 8 * 11 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] + 8 * 4];
566  nnz_cache[3 + 8 * 12 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] + 8 * 4];
567  } else if (CHROMA422) {
568  nnz_cache[3 + 8 * 6 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] - 2 + 4 * 4];
569  nnz_cache[3 + 8 * 7 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] - 2 + 4 * 4];
570  nnz_cache[3 + 8 * 11 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] - 2 + 8 * 4];
571  nnz_cache[3 + 8 * 12 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] - 2 + 8 * 4];
572  } else {
573  nnz_cache[3 + 8 * 6 + 8 * i] = nnz[left_block[8 + 4 + 2 * i]];
574  nnz_cache[3 + 8 * 11 + 8 * i] = nnz[left_block[8 + 5 + 2 * i]];
575  }
576  } else {
577  nnz_cache[3 + 8 * 1 + 2 * 8 * i] =
578  nnz_cache[3 + 8 * 2 + 2 * 8 * i] =
579  nnz_cache[3 + 8 * 6 + 2 * 8 * i] =
580  nnz_cache[3 + 8 * 7 + 2 * 8 * i] =
581  nnz_cache[3 + 8 * 11 + 2 * 8 * i] =
582  nnz_cache[3 + 8 * 12 + 2 * 8 * i] = CABAC && !IS_INTRA(mb_type) ? 0 : 64;
583  }
584  }
585 
586  if (CABAC) {
587  // top_cbp
588  if (top_type)
589  h->top_cbp = h->cbp_table[top_xy];
590  else
591  h->top_cbp = IS_INTRA(mb_type) ? 0x7CF : 0x00F;
592  // left_cbp
593  if (left_type[LTOP]) {
594  h->left_cbp = (h->cbp_table[left_xy[LTOP]] & 0x7F0) |
595  ((h->cbp_table[left_xy[LTOP]] >> (left_block[0] & (~1))) & 2) |
596  (((h->cbp_table[left_xy[LBOT]] >> (left_block[2] & (~1))) & 2) << 2);
597  } else {
598  h->left_cbp = IS_INTRA(mb_type) ? 0x7CF : 0x00F;
599  }
600  }
601  }
602 
603  if (IS_INTER(mb_type) || (IS_DIRECT(mb_type) && h->direct_spatial_mv_pred)) {
604  int list;
605  int b_stride = h->b_stride;
606  for (list = 0; list < h->list_count; list++) {
607  int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
608  int8_t *ref = s->current_picture.f.ref_index[list];
609  int16_t(*mv_cache)[2] = &h->mv_cache[list][scan8[0]];
610  int16_t(*mv)[2] = s->current_picture.f.motion_val[list];
611  if (!USES_LIST(mb_type, list))
612  continue;
613  av_assert2(!(IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred));
614 
615  if (USES_LIST(top_type, list)) {
616  const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
617  AV_COPY128(mv_cache[0 - 1 * 8], mv[b_xy + 0]);
618  ref_cache[0 - 1 * 8] =
619  ref_cache[1 - 1 * 8] = ref[4 * top_xy + 2];
620  ref_cache[2 - 1 * 8] =
621  ref_cache[3 - 1 * 8] = ref[4 * top_xy + 3];
622  } else {
623  AV_ZERO128(mv_cache[0 - 1 * 8]);
624  AV_WN32A(&ref_cache[0 - 1 * 8],
625  ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE) & 0xFF) * 0x01010101u);
626  }
627 
628  if (mb_type & (MB_TYPE_16x8 | MB_TYPE_8x8)) {
629  for (i = 0; i < 2; i++) {
630  int cache_idx = -1 + i * 2 * 8;
631  if (USES_LIST(left_type[LEFT(i)], list)) {
632  const int b_xy = h->mb2b_xy[left_xy[LEFT(i)]] + 3;
633  const int b8_xy = 4 * left_xy[LEFT(i)] + 1;
634  AV_COPY32(mv_cache[cache_idx],
635  mv[b_xy + b_stride * left_block[0 + i * 2]]);
636  AV_COPY32(mv_cache[cache_idx + 8],
637  mv[b_xy + b_stride * left_block[1 + i * 2]]);
638  ref_cache[cache_idx] = ref[b8_xy + (left_block[0 + i * 2] & ~1)];
639  ref_cache[cache_idx + 8] = ref[b8_xy + (left_block[1 + i * 2] & ~1)];
640  } else {
641  AV_ZERO32(mv_cache[cache_idx]);
642  AV_ZERO32(mv_cache[cache_idx + 8]);
643  ref_cache[cache_idx] =
644  ref_cache[cache_idx + 8] = (left_type[LEFT(i)]) ? LIST_NOT_USED
646  }
647  }
648  } else {
649  if (USES_LIST(left_type[LTOP], list)) {
650  const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
651  const int b8_xy = 4 * left_xy[LTOP] + 1;
652  AV_COPY32(mv_cache[-1], mv[b_xy + b_stride * left_block[0]]);
653  ref_cache[-1] = ref[b8_xy + (left_block[0] & ~1)];
654  } else {
655  AV_ZERO32(mv_cache[-1]);
656  ref_cache[-1] = left_type[LTOP] ? LIST_NOT_USED
658  }
659  }
660 
661  if (USES_LIST(topright_type, list)) {
662  const int b_xy = h->mb2b_xy[topright_xy] + 3 * b_stride;
663  AV_COPY32(mv_cache[4 - 1 * 8], mv[b_xy]);
664  ref_cache[4 - 1 * 8] = ref[4 * topright_xy + 2];
665  } else {
666  AV_ZERO32(mv_cache[4 - 1 * 8]);
667  ref_cache[4 - 1 * 8] = topright_type ? LIST_NOT_USED
669  }
670  if(ref_cache[2 - 1*8] < 0 || ref_cache[4 - 1*8] < 0){
671  if (USES_LIST(topleft_type, list)) {
672  const int b_xy = h->mb2b_xy[topleft_xy] + 3 + b_stride +
673  (h->topleft_partition & 2 * b_stride);
674  const int b8_xy = 4 * topleft_xy + 1 + (h->topleft_partition & 2);
675  AV_COPY32(mv_cache[-1 - 1 * 8], mv[b_xy]);
676  ref_cache[-1 - 1 * 8] = ref[b8_xy];
677  } else {
678  AV_ZERO32(mv_cache[-1 - 1 * 8]);
679  ref_cache[-1 - 1 * 8] = topleft_type ? LIST_NOT_USED
681  }
682  }
683 
684  if ((mb_type & (MB_TYPE_SKIP | MB_TYPE_DIRECT2)) && !FRAME_MBAFF)
685  continue;
686 
687  if (!(mb_type & (MB_TYPE_SKIP | MB_TYPE_DIRECT2))) {
688  uint8_t(*mvd_cache)[2] = &h->mvd_cache[list][scan8[0]];
689  uint8_t(*mvd)[2] = h->mvd_table[list];
690  ref_cache[2 + 8 * 0] =
691  ref_cache[2 + 8 * 2] = PART_NOT_AVAILABLE;
692  AV_ZERO32(mv_cache[2 + 8 * 0]);
693  AV_ZERO32(mv_cache[2 + 8 * 2]);
694 
695  if (CABAC) {
696  if (USES_LIST(top_type, list)) {
697  const int b_xy = h->mb2br_xy[top_xy];
698  AV_COPY64(mvd_cache[0 - 1 * 8], mvd[b_xy + 0]);
699  } else {
700  AV_ZERO64(mvd_cache[0 - 1 * 8]);
701  }
702  if (USES_LIST(left_type[LTOP], list)) {
703  const int b_xy = h->mb2br_xy[left_xy[LTOP]] + 6;
704  AV_COPY16(mvd_cache[-1 + 0 * 8], mvd[b_xy - left_block[0]]);
705  AV_COPY16(mvd_cache[-1 + 1 * 8], mvd[b_xy - left_block[1]]);
706  } else {
707  AV_ZERO16(mvd_cache[-1 + 0 * 8]);
708  AV_ZERO16(mvd_cache[-1 + 1 * 8]);
709  }
710  if (USES_LIST(left_type[LBOT], list)) {
711  const int b_xy = h->mb2br_xy[left_xy[LBOT]] + 6;
712  AV_COPY16(mvd_cache[-1 + 2 * 8], mvd[b_xy - left_block[2]]);
713  AV_COPY16(mvd_cache[-1 + 3 * 8], mvd[b_xy - left_block[3]]);
714  } else {
715  AV_ZERO16(mvd_cache[-1 + 2 * 8]);
716  AV_ZERO16(mvd_cache[-1 + 3 * 8]);
717  }
718  AV_ZERO16(mvd_cache[2 + 8 * 0]);
719  AV_ZERO16(mvd_cache[2 + 8 * 2]);
720  if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
721  uint8_t *direct_cache = &h->direct_cache[scan8[0]];
722  uint8_t *direct_table = h->direct_table;
723  fill_rectangle(direct_cache, 4, 4, 8, MB_TYPE_16x16 >> 1, 1);
724 
725  if (IS_DIRECT(top_type)) {
726  AV_WN32A(&direct_cache[-1 * 8],
727  0x01010101u * (MB_TYPE_DIRECT2 >> 1));
728  } else if (IS_8X8(top_type)) {
729  int b8_xy = 4 * top_xy;
730  direct_cache[0 - 1 * 8] = direct_table[b8_xy + 2];
731  direct_cache[2 - 1 * 8] = direct_table[b8_xy + 3];
732  } else {
733  AV_WN32A(&direct_cache[-1 * 8],
734  0x01010101 * (MB_TYPE_16x16 >> 1));
735  }
736 
737  if (IS_DIRECT(left_type[LTOP]))
738  direct_cache[-1 + 0 * 8] = MB_TYPE_DIRECT2 >> 1;
739  else if (IS_8X8(left_type[LTOP]))
740  direct_cache[-1 + 0 * 8] = direct_table[4 * left_xy[LTOP] + 1 + (left_block[0] & ~1)];
741  else
742  direct_cache[-1 + 0 * 8] = MB_TYPE_16x16 >> 1;
743 
744  if (IS_DIRECT(left_type[LBOT]))
745  direct_cache[-1 + 2 * 8] = MB_TYPE_DIRECT2 >> 1;
746  else if (IS_8X8(left_type[LBOT]))
747  direct_cache[-1 + 2 * 8] = direct_table[4 * left_xy[LBOT] + 1 + (left_block[2] & ~1)];
748  else
749  direct_cache[-1 + 2 * 8] = MB_TYPE_16x16 >> 1;
750  }
751  }
752  }
753 
754 #define MAP_MVS \
755  MAP_F2F(scan8[0] - 1 - 1 * 8, topleft_type) \
756  MAP_F2F(scan8[0] + 0 - 1 * 8, top_type) \
757  MAP_F2F(scan8[0] + 1 - 1 * 8, top_type) \
758  MAP_F2F(scan8[0] + 2 - 1 * 8, top_type) \
759  MAP_F2F(scan8[0] + 3 - 1 * 8, top_type) \
760  MAP_F2F(scan8[0] + 4 - 1 * 8, topright_type) \
761  MAP_F2F(scan8[0] - 1 + 0 * 8, left_type[LTOP]) \
762  MAP_F2F(scan8[0] - 1 + 1 * 8, left_type[LTOP]) \
763  MAP_F2F(scan8[0] - 1 + 2 * 8, left_type[LBOT]) \
764  MAP_F2F(scan8[0] - 1 + 3 * 8, left_type[LBOT])
765 
766  if (FRAME_MBAFF) {
767  if (MB_FIELD) {
768 
769 #define MAP_F2F(idx, mb_type) \
770  if (!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0) { \
771  h->ref_cache[list][idx] <<= 1; \
772  h->mv_cache[list][idx][1] /= 2; \
773  h->mvd_cache[list][idx][1] >>= 1; \
774  }
775 
776  MAP_MVS
777  } else {
778 
779 #undef MAP_F2F
780 #define MAP_F2F(idx, mb_type) \
781  if (IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0) { \
782  h->ref_cache[list][idx] >>= 1; \
783  h->mv_cache[list][idx][1] <<= 1; \
784  h->mvd_cache[list][idx][1] <<= 1; \
785  }
786 
787  MAP_MVS
788 #undef MAP_F2F
789  }
790  }
791  }
792  }
793 
794  h->neighbor_transform_size = !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[LTOP]);
795 }
796 
797 /**
798  * decodes a P_SKIP or B_SKIP macroblock
799  */
801 {
802  MpegEncContext *const s = &h->s;
803  const int mb_xy = h->mb_xy;
804  int mb_type = 0;
805 
806  memset(h->non_zero_count[mb_xy], 0, 48);
807 
808  if (MB_FIELD)
809  mb_type |= MB_TYPE_INTERLACED;
810 
811  if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
812  // just for fill_caches. pred_direct_motion will set the real mb_type
814  if (h->direct_spatial_mv_pred) {
815  fill_decode_neighbors(h, mb_type);
816  fill_decode_caches(h, mb_type); //FIXME check what is needed and what not ...
817  }
818  ff_h264_pred_direct_motion(h, &mb_type);
819  mb_type |= MB_TYPE_SKIP;
820  } else {
822 
823  fill_decode_neighbors(h, mb_type);
825  }
826 
827  write_back_motion(h, mb_type);
828  s->current_picture.f.mb_type[mb_xy] = mb_type;
829  s->current_picture.f.qscale_table[mb_xy] = s->qscale;
830  h->slice_table[mb_xy] = h->slice_num;
831  h->prev_mb_skipped = 1;
832 }
833 
834 #endif /* AVCODEC_H264_MVPRED_H */