FFmpeg
 All Data Structures Namespaces 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 "mpegutils.h"
35 #include "libavutil/avassert.h"
36 
37 
38 static av_always_inline int fetch_diagonal_mv(H264Context *h, const int16_t **C,
39  int i, int list, int part_width)
40 {
41  const int topright_ref = h->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  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 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  && h->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(h->mv_cache[list][scan8[0] - 2]);
62  *C = h->mv_cache[list][scan8[0] - 2];
63 
64  if (!MB_FIELD(h) && IS_INTERLACED(h->left_type[0])) {
65  SET_DIAG_MV(* 2, >> 1, h->left_mb_xy[0] + h->mb_stride,
66  (h->mb_y & 1) * 2 + (i >> 5));
67  }
68  if (MB_FIELD(h) && !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(h->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->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->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->mb_x, h->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->avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n",
166  top_ref, B[0], B[1], h->mb_x, h->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->avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n",
178  left_ref, A[0], A[1], h->mb_x, h->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->avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n",
206  left_ref, A[0], A[1], h->mb_x, h->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->avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n",
220  diagonal_ref, C[0], C[1], h->mb_x, h->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(h)) { \
235  if (MB_FIELD(h)) { \
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  int8_t *ref = h->cur_pic.ref_index[0];
257  int16_t(*mv)[2] = h->cur_pic.motion_val[0];
258  int top_ref, left_ref, diagonal_ref, match_count, mx, my;
259  const int16_t *A, *B, *C;
260  int b_stride = h->b_stride;
261 
262  fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, 0, 1);
263 
264  /* To avoid doing an entire fill_decode_caches, we inline the relevant
265  * parts here.
266  * FIXME: this is a partial duplicate of the logic in fill_decode_caches,
267  * but it's faster this way. Is there a way to avoid this duplication?
268  */
269  if (USES_LIST(h->left_type[LTOP], 0)) {
270  left_ref = ref[4 * h->left_mb_xy[LTOP] + 1 + (h->left_block[0] & ~1)];
271  A = mv[h->mb2b_xy[h->left_mb_xy[LTOP]] + 3 + b_stride * h->left_block[0]];
272  FIX_MV_MBAFF(h->left_type[LTOP], left_ref, A, 0);
273  if (!(left_ref | AV_RN32A(A)))
274  goto zeromv;
275  } else if (h->left_type[LTOP]) {
276  left_ref = LIST_NOT_USED;
277  A = zeromv;
278  } else {
279  goto zeromv;
280  }
281 
282  if (USES_LIST(h->top_type, 0)) {
283  top_ref = ref[4 * h->top_mb_xy + 2];
284  B = mv[h->mb2b_xy[h->top_mb_xy] + 3 * b_stride];
285  FIX_MV_MBAFF(h->top_type, top_ref, B, 1);
286  if (!(top_ref | AV_RN32A(B)))
287  goto zeromv;
288  } else if (h->top_type) {
289  top_ref = LIST_NOT_USED;
290  B = zeromv;
291  } else {
292  goto zeromv;
293  }
294 
295  tprintf(h->avctx, "pred_pskip: (%d) (%d) at %2d %2d\n",
296  top_ref, left_ref, h->mb_x, h->mb_y);
297 
298  if (USES_LIST(h->topright_type, 0)) {
299  diagonal_ref = ref[4 * h->topright_mb_xy + 2];
300  C = mv[h->mb2b_xy[h->topright_mb_xy] + 3 * b_stride];
301  FIX_MV_MBAFF(h->topright_type, diagonal_ref, C, 2);
302  } else if (h->topright_type) {
303  diagonal_ref = LIST_NOT_USED;
304  C = zeromv;
305  } else {
306  if (USES_LIST(h->topleft_type, 0)) {
307  diagonal_ref = ref[4 * h->topleft_mb_xy + 1 +
308  (h->topleft_partition & 2)];
309  C = mv[h->mb2b_xy[h->topleft_mb_xy] + 3 + b_stride +
310  (h->topleft_partition & 2 * b_stride)];
311  FIX_MV_MBAFF(h->topleft_type, diagonal_ref, C, 2);
312  } else if (h->topleft_type) {
313  diagonal_ref = LIST_NOT_USED;
314  C = zeromv;
315  } else {
316  diagonal_ref = PART_NOT_AVAILABLE;
317  C = zeromv;
318  }
319  }
320 
321  match_count = !diagonal_ref + !top_ref + !left_ref;
322  tprintf(h->avctx, "pred_pskip_motion match_count=%d\n", match_count);
323  if (match_count > 1) {
324  mx = mid_pred(A[0], B[0], C[0]);
325  my = mid_pred(A[1], B[1], C[1]);
326  } else if (match_count == 1) {
327  if (!left_ref) {
328  mx = A[0];
329  my = A[1];
330  } else if (!top_ref) {
331  mx = B[0];
332  my = B[1];
333  } else {
334  mx = C[0];
335  my = C[1];
336  }
337  } else {
338  mx = mid_pred(A[0], B[0], C[0]);
339  my = mid_pred(A[1], B[1], C[1]);
340  }
341 
342  fill_rectangle(h->mv_cache[0][scan8[0]], 4, 4, 8, pack16to32(mx, my), 4);
343  return;
344 
345 zeromv:
346  fill_rectangle(h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
347  return;
348 }
349 
350 static void fill_decode_neighbors(H264Context *h, int mb_type)
351 {
352  const int mb_xy = h->mb_xy;
353  int topleft_xy, top_xy, topright_xy, left_xy[LEFT_MBS];
354  static const uint8_t left_block_options[4][32] = {
355  { 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 },
356  { 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 },
357  { 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 },
358  { 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 }
359  };
360 
361  h->topleft_partition = -1;
362 
363  top_xy = mb_xy - (h->mb_stride << MB_FIELD(h));
364 
365  /* Wow, what a mess, why didn't they simplify the interlacing & intra
366  * stuff, I can't imagine that these complex rules are worth it. */
367 
368  topleft_xy = top_xy - 1;
369  topright_xy = top_xy + 1;
370  left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
371  h->left_block = left_block_options[0];
372  if (FRAME_MBAFF(h)) {
373  const int left_mb_field_flag = IS_INTERLACED(h->cur_pic.mb_type[mb_xy - 1]);
374  const int curr_mb_field_flag = IS_INTERLACED(mb_type);
375  if (h->mb_y & 1) {
376  if (left_mb_field_flag != curr_mb_field_flag) {
377  left_xy[LBOT] = left_xy[LTOP] = mb_xy - h->mb_stride - 1;
378  if (curr_mb_field_flag) {
379  left_xy[LBOT] += h->mb_stride;
380  h->left_block = left_block_options[3];
381  } else {
382  topleft_xy += h->mb_stride;
383  /* take top left mv from the middle of the mb, as opposed
384  * to all other modes which use the bottom right partition */
385  h->topleft_partition = 0;
386  h->left_block = left_block_options[1];
387  }
388  }
389  } else {
390  if (curr_mb_field_flag) {
391  topleft_xy += h->mb_stride & (((h->cur_pic.mb_type[top_xy - 1] >> 7) & 1) - 1);
392  topright_xy += h->mb_stride & (((h->cur_pic.mb_type[top_xy + 1] >> 7) & 1) - 1);
393  top_xy += h->mb_stride & (((h->cur_pic.mb_type[top_xy] >> 7) & 1) - 1);
394  }
395  if (left_mb_field_flag != curr_mb_field_flag) {
396  if (curr_mb_field_flag) {
397  left_xy[LBOT] += h->mb_stride;
398  h->left_block = left_block_options[3];
399  } else {
400  h->left_block = left_block_options[2];
401  }
402  }
403  }
404  }
405 
406  h->topleft_mb_xy = topleft_xy;
407  h->top_mb_xy = top_xy;
408  h->topright_mb_xy = topright_xy;
409  h->left_mb_xy[LTOP] = left_xy[LTOP];
410  h->left_mb_xy[LBOT] = left_xy[LBOT];
411  //FIXME do we need all in the context?
412 
413  h->topleft_type = h->cur_pic.mb_type[topleft_xy];
414  h->top_type = h->cur_pic.mb_type[top_xy];
415  h->topright_type = h->cur_pic.mb_type[topright_xy];
416  h->left_type[LTOP] = h->cur_pic.mb_type[left_xy[LTOP]];
417  h->left_type[LBOT] = h->cur_pic.mb_type[left_xy[LBOT]];
418 
419  if (FMO) {
420  if (h->slice_table[topleft_xy] != h->slice_num)
421  h->topleft_type = 0;
422  if (h->slice_table[top_xy] != h->slice_num)
423  h->top_type = 0;
424  if (h->slice_table[left_xy[LTOP]] != h->slice_num)
425  h->left_type[LTOP] = h->left_type[LBOT] = 0;
426  } else {
427  if (h->slice_table[topleft_xy] != h->slice_num) {
428  h->topleft_type = 0;
429  if (h->slice_table[top_xy] != h->slice_num)
430  h->top_type = 0;
431  if (h->slice_table[left_xy[LTOP]] != h->slice_num)
432  h->left_type[LTOP] = h->left_type[LBOT] = 0;
433  }
434  }
435  if (h->slice_table[topright_xy] != h->slice_num)
436  h->topright_type = 0;
437 }
438 
439 static void fill_decode_caches(H264Context *h, int mb_type)
440 {
441  int topleft_xy, top_xy, topright_xy, left_xy[LEFT_MBS];
442  int topleft_type, top_type, topright_type, left_type[LEFT_MBS];
443  const uint8_t *left_block = h->left_block;
444  int i;
445  uint8_t *nnz;
446  uint8_t *nnz_cache;
447 
448  topleft_xy = h->topleft_mb_xy;
449  top_xy = h->top_mb_xy;
450  topright_xy = h->topright_mb_xy;
451  left_xy[LTOP] = h->left_mb_xy[LTOP];
452  left_xy[LBOT] = h->left_mb_xy[LBOT];
453  topleft_type = h->topleft_type;
454  top_type = h->top_type;
455  topright_type = h->topright_type;
456  left_type[LTOP] = h->left_type[LTOP];
457  left_type[LBOT] = h->left_type[LBOT];
458 
459  if (!IS_SKIP(mb_type)) {
460  if (IS_INTRA(mb_type)) {
461  int type_mask = h->pps.constrained_intra_pred ? IS_INTRA(-1) : -1;
464  h->left_samples_available = 0xFFFF;
465  h->topright_samples_available = 0xEEEA;
466 
467  if (!(top_type & type_mask)) {
468  h->topleft_samples_available = 0xB3FF;
469  h->top_samples_available = 0x33FF;
470  h->topright_samples_available = 0x26EA;
471  }
472  if (IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[LTOP])) {
473  if (IS_INTERLACED(mb_type)) {
474  if (!(left_type[LTOP] & type_mask)) {
475  h->topleft_samples_available &= 0xDFFF;
476  h->left_samples_available &= 0x5FFF;
477  }
478  if (!(left_type[LBOT] & type_mask)) {
479  h->topleft_samples_available &= 0xFF5F;
480  h->left_samples_available &= 0xFF5F;
481  }
482  } else {
483  int left_typei = h->cur_pic.mb_type[left_xy[LTOP] + h->mb_stride];
484 
485  av_assert2(left_xy[LTOP] == left_xy[LBOT]);
486  if (!((left_typei & type_mask) && (left_type[LTOP] & type_mask))) {
487  h->topleft_samples_available &= 0xDF5F;
488  h->left_samples_available &= 0x5F5F;
489  }
490  }
491  } else {
492  if (!(left_type[LTOP] & type_mask)) {
493  h->topleft_samples_available &= 0xDF5F;
494  h->left_samples_available &= 0x5F5F;
495  }
496  }
497 
498  if (!(topleft_type & type_mask))
499  h->topleft_samples_available &= 0x7FFF;
500 
501  if (!(topright_type & type_mask))
502  h->topright_samples_available &= 0xFBFF;
503 
504  if (IS_INTRA4x4(mb_type)) {
505  if (IS_INTRA4x4(top_type)) {
506  AV_COPY32(h->intra4x4_pred_mode_cache + 4 + 8 * 0, h->intra4x4_pred_mode + h->mb2br_xy[top_xy]);
507  } else {
508  h->intra4x4_pred_mode_cache[4 + 8 * 0] =
509  h->intra4x4_pred_mode_cache[5 + 8 * 0] =
510  h->intra4x4_pred_mode_cache[6 + 8 * 0] =
511  h->intra4x4_pred_mode_cache[7 + 8 * 0] = 2 - 3 * !(top_type & type_mask);
512  }
513  for (i = 0; i < 2; i++) {
514  if (IS_INTRA4x4(left_type[LEFT(i)])) {
515  int8_t *mode = h->intra4x4_pred_mode + h->mb2br_xy[left_xy[LEFT(i)]];
516  h->intra4x4_pred_mode_cache[3 + 8 * 1 + 2 * 8 * i] = mode[6 - left_block[0 + 2 * i]];
517  h->intra4x4_pred_mode_cache[3 + 8 * 2 + 2 * 8 * i] = mode[6 - left_block[1 + 2 * i]];
518  } else {
519  h->intra4x4_pred_mode_cache[3 + 8 * 1 + 2 * 8 * i] =
520  h->intra4x4_pred_mode_cache[3 + 8 * 2 + 2 * 8 * i] = 2 - 3 * !(left_type[LEFT(i)] & type_mask);
521  }
522  }
523  }
524  }
525 
526  /*
527  * 0 . T T. T T T T
528  * 1 L . .L . . . .
529  * 2 L . .L . . . .
530  * 3 . T TL . . . .
531  * 4 L . .L . . . .
532  * 5 L . .. . . . .
533  */
534  /* FIXME: constraint_intra_pred & partitioning & nnz
535  * (let us hope this is just a typo in the spec) */
536  nnz_cache = h->non_zero_count_cache;
537  if (top_type) {
538  nnz = h->non_zero_count[top_xy];
539  AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[4 * 3]);
540  if (!h->chroma_y_shift) {
541  AV_COPY32(&nnz_cache[4 + 8 * 5], &nnz[4 * 7]);
542  AV_COPY32(&nnz_cache[4 + 8 * 10], &nnz[4 * 11]);
543  } else {
544  AV_COPY32(&nnz_cache[4 + 8 * 5], &nnz[4 * 5]);
545  AV_COPY32(&nnz_cache[4 + 8 * 10], &nnz[4 * 9]);
546  }
547  } else {
548  uint32_t top_empty = CABAC(h) && !IS_INTRA(mb_type) ? 0 : 0x40404040;
549  AV_WN32A(&nnz_cache[4 + 8 * 0], top_empty);
550  AV_WN32A(&nnz_cache[4 + 8 * 5], top_empty);
551  AV_WN32A(&nnz_cache[4 + 8 * 10], top_empty);
552  }
553 
554  for (i = 0; i < 2; i++) {
555  if (left_type[LEFT(i)]) {
556  nnz = h->non_zero_count[left_xy[LEFT(i)]];
557  nnz_cache[3 + 8 * 1 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i]];
558  nnz_cache[3 + 8 * 2 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i]];
559  if (CHROMA444(h)) {
560  nnz_cache[3 + 8 * 6 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] + 4 * 4];
561  nnz_cache[3 + 8 * 7 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] + 4 * 4];
562  nnz_cache[3 + 8 * 11 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] + 8 * 4];
563  nnz_cache[3 + 8 * 12 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] + 8 * 4];
564  } else if (CHROMA422(h)) {
565  nnz_cache[3 + 8 * 6 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] - 2 + 4 * 4];
566  nnz_cache[3 + 8 * 7 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] - 2 + 4 * 4];
567  nnz_cache[3 + 8 * 11 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] - 2 + 8 * 4];
568  nnz_cache[3 + 8 * 12 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] - 2 + 8 * 4];
569  } else {
570  nnz_cache[3 + 8 * 6 + 8 * i] = nnz[left_block[8 + 4 + 2 * i]];
571  nnz_cache[3 + 8 * 11 + 8 * i] = nnz[left_block[8 + 5 + 2 * i]];
572  }
573  } else {
574  nnz_cache[3 + 8 * 1 + 2 * 8 * i] =
575  nnz_cache[3 + 8 * 2 + 2 * 8 * i] =
576  nnz_cache[3 + 8 * 6 + 2 * 8 * i] =
577  nnz_cache[3 + 8 * 7 + 2 * 8 * i] =
578  nnz_cache[3 + 8 * 11 + 2 * 8 * i] =
579  nnz_cache[3 + 8 * 12 + 2 * 8 * i] = CABAC(h) && !IS_INTRA(mb_type) ? 0 : 64;
580  }
581  }
582 
583  if (CABAC(h)) {
584  // top_cbp
585  if (top_type)
586  h->top_cbp = h->cbp_table[top_xy];
587  else
588  h->top_cbp = IS_INTRA(mb_type) ? 0x7CF : 0x00F;
589  // left_cbp
590  if (left_type[LTOP]) {
591  h->left_cbp = (h->cbp_table[left_xy[LTOP]] & 0x7F0) |
592  ((h->cbp_table[left_xy[LTOP]] >> (left_block[0] & (~1))) & 2) |
593  (((h->cbp_table[left_xy[LBOT]] >> (left_block[2] & (~1))) & 2) << 2);
594  } else {
595  h->left_cbp = IS_INTRA(mb_type) ? 0x7CF : 0x00F;
596  }
597  }
598  }
599 
600  if (IS_INTER(mb_type) || (IS_DIRECT(mb_type) && h->direct_spatial_mv_pred)) {
601  int list;
602  int b_stride = h->b_stride;
603  for (list = 0; list < h->list_count; list++) {
604  int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
605  int8_t *ref = h->cur_pic.ref_index[list];
606  int16_t(*mv_cache)[2] = &h->mv_cache[list][scan8[0]];
607  int16_t(*mv)[2] = h->cur_pic.motion_val[list];
608  if (!USES_LIST(mb_type, list))
609  continue;
610  av_assert2(!(IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred));
611 
612  if (USES_LIST(top_type, list)) {
613  const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
614  AV_COPY128(mv_cache[0 - 1 * 8], mv[b_xy + 0]);
615  ref_cache[0 - 1 * 8] =
616  ref_cache[1 - 1 * 8] = ref[4 * top_xy + 2];
617  ref_cache[2 - 1 * 8] =
618  ref_cache[3 - 1 * 8] = ref[4 * top_xy + 3];
619  } else {
620  AV_ZERO128(mv_cache[0 - 1 * 8]);
621  AV_WN32A(&ref_cache[0 - 1 * 8],
622  ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE) & 0xFF) * 0x01010101u);
623  }
624 
625  if (mb_type & (MB_TYPE_16x8 | MB_TYPE_8x8)) {
626  for (i = 0; i < 2; i++) {
627  int cache_idx = -1 + i * 2 * 8;
628  if (USES_LIST(left_type[LEFT(i)], list)) {
629  const int b_xy = h->mb2b_xy[left_xy[LEFT(i)]] + 3;
630  const int b8_xy = 4 * left_xy[LEFT(i)] + 1;
631  AV_COPY32(mv_cache[cache_idx],
632  mv[b_xy + b_stride * left_block[0 + i * 2]]);
633  AV_COPY32(mv_cache[cache_idx + 8],
634  mv[b_xy + b_stride * left_block[1 + i * 2]]);
635  ref_cache[cache_idx] = ref[b8_xy + (left_block[0 + i * 2] & ~1)];
636  ref_cache[cache_idx + 8] = ref[b8_xy + (left_block[1 + i * 2] & ~1)];
637  } else {
638  AV_ZERO32(mv_cache[cache_idx]);
639  AV_ZERO32(mv_cache[cache_idx + 8]);
640  ref_cache[cache_idx] =
641  ref_cache[cache_idx + 8] = (left_type[LEFT(i)]) ? LIST_NOT_USED
643  }
644  }
645  } else {
646  if (USES_LIST(left_type[LTOP], list)) {
647  const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
648  const int b8_xy = 4 * left_xy[LTOP] + 1;
649  AV_COPY32(mv_cache[-1], mv[b_xy + b_stride * left_block[0]]);
650  ref_cache[-1] = ref[b8_xy + (left_block[0] & ~1)];
651  } else {
652  AV_ZERO32(mv_cache[-1]);
653  ref_cache[-1] = left_type[LTOP] ? LIST_NOT_USED
655  }
656  }
657 
658  if (USES_LIST(topright_type, list)) {
659  const int b_xy = h->mb2b_xy[topright_xy] + 3 * b_stride;
660  AV_COPY32(mv_cache[4 - 1 * 8], mv[b_xy]);
661  ref_cache[4 - 1 * 8] = ref[4 * topright_xy + 2];
662  } else {
663  AV_ZERO32(mv_cache[4 - 1 * 8]);
664  ref_cache[4 - 1 * 8] = topright_type ? LIST_NOT_USED
666  }
667  if(ref_cache[2 - 1*8] < 0 || ref_cache[4 - 1 * 8] < 0) {
668  if (USES_LIST(topleft_type, list)) {
669  const int b_xy = h->mb2b_xy[topleft_xy] + 3 + b_stride +
670  (h->topleft_partition & 2 * b_stride);
671  const int b8_xy = 4 * topleft_xy + 1 + (h->topleft_partition & 2);
672  AV_COPY32(mv_cache[-1 - 1 * 8], mv[b_xy]);
673  ref_cache[-1 - 1 * 8] = ref[b8_xy];
674  } else {
675  AV_ZERO32(mv_cache[-1 - 1 * 8]);
676  ref_cache[-1 - 1 * 8] = topleft_type ? LIST_NOT_USED
678  }
679  }
680 
681  if ((mb_type & (MB_TYPE_SKIP | MB_TYPE_DIRECT2)) && !FRAME_MBAFF(h))
682  continue;
683 
684  if (!(mb_type & (MB_TYPE_SKIP | MB_TYPE_DIRECT2))) {
685  uint8_t(*mvd_cache)[2] = &h->mvd_cache[list][scan8[0]];
686  uint8_t(*mvd)[2] = h->mvd_table[list];
687  ref_cache[2 + 8 * 0] =
688  ref_cache[2 + 8 * 2] = PART_NOT_AVAILABLE;
689  AV_ZERO32(mv_cache[2 + 8 * 0]);
690  AV_ZERO32(mv_cache[2 + 8 * 2]);
691 
692  if (CABAC(h)) {
693  if (USES_LIST(top_type, list)) {
694  const int b_xy = h->mb2br_xy[top_xy];
695  AV_COPY64(mvd_cache[0 - 1 * 8], mvd[b_xy + 0]);
696  } else {
697  AV_ZERO64(mvd_cache[0 - 1 * 8]);
698  }
699  if (USES_LIST(left_type[LTOP], list)) {
700  const int b_xy = h->mb2br_xy[left_xy[LTOP]] + 6;
701  AV_COPY16(mvd_cache[-1 + 0 * 8], mvd[b_xy - left_block[0]]);
702  AV_COPY16(mvd_cache[-1 + 1 * 8], mvd[b_xy - left_block[1]]);
703  } else {
704  AV_ZERO16(mvd_cache[-1 + 0 * 8]);
705  AV_ZERO16(mvd_cache[-1 + 1 * 8]);
706  }
707  if (USES_LIST(left_type[LBOT], list)) {
708  const int b_xy = h->mb2br_xy[left_xy[LBOT]] + 6;
709  AV_COPY16(mvd_cache[-1 + 2 * 8], mvd[b_xy - left_block[2]]);
710  AV_COPY16(mvd_cache[-1 + 3 * 8], mvd[b_xy - left_block[3]]);
711  } else {
712  AV_ZERO16(mvd_cache[-1 + 2 * 8]);
713  AV_ZERO16(mvd_cache[-1 + 3 * 8]);
714  }
715  AV_ZERO16(mvd_cache[2 + 8 * 0]);
716  AV_ZERO16(mvd_cache[2 + 8 * 2]);
717  if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
718  uint8_t *direct_cache = &h->direct_cache[scan8[0]];
719  uint8_t *direct_table = h->direct_table;
720  fill_rectangle(direct_cache, 4, 4, 8, MB_TYPE_16x16 >> 1, 1);
721 
722  if (IS_DIRECT(top_type)) {
723  AV_WN32A(&direct_cache[-1 * 8],
724  0x01010101u * (MB_TYPE_DIRECT2 >> 1));
725  } else if (IS_8X8(top_type)) {
726  int b8_xy = 4 * top_xy;
727  direct_cache[0 - 1 * 8] = direct_table[b8_xy + 2];
728  direct_cache[2 - 1 * 8] = direct_table[b8_xy + 3];
729  } else {
730  AV_WN32A(&direct_cache[-1 * 8],
731  0x01010101 * (MB_TYPE_16x16 >> 1));
732  }
733 
734  if (IS_DIRECT(left_type[LTOP]))
735  direct_cache[-1 + 0 * 8] = MB_TYPE_DIRECT2 >> 1;
736  else if (IS_8X8(left_type[LTOP]))
737  direct_cache[-1 + 0 * 8] = direct_table[4 * left_xy[LTOP] + 1 + (left_block[0] & ~1)];
738  else
739  direct_cache[-1 + 0 * 8] = MB_TYPE_16x16 >> 1;
740 
741  if (IS_DIRECT(left_type[LBOT]))
742  direct_cache[-1 + 2 * 8] = MB_TYPE_DIRECT2 >> 1;
743  else if (IS_8X8(left_type[LBOT]))
744  direct_cache[-1 + 2 * 8] = direct_table[4 * left_xy[LBOT] + 1 + (left_block[2] & ~1)];
745  else
746  direct_cache[-1 + 2 * 8] = MB_TYPE_16x16 >> 1;
747  }
748  }
749  }
750 
751 #define MAP_MVS \
752  MAP_F2F(scan8[0] - 1 - 1 * 8, topleft_type) \
753  MAP_F2F(scan8[0] + 0 - 1 * 8, top_type) \
754  MAP_F2F(scan8[0] + 1 - 1 * 8, top_type) \
755  MAP_F2F(scan8[0] + 2 - 1 * 8, top_type) \
756  MAP_F2F(scan8[0] + 3 - 1 * 8, top_type) \
757  MAP_F2F(scan8[0] + 4 - 1 * 8, topright_type) \
758  MAP_F2F(scan8[0] - 1 + 0 * 8, left_type[LTOP]) \
759  MAP_F2F(scan8[0] - 1 + 1 * 8, left_type[LTOP]) \
760  MAP_F2F(scan8[0] - 1 + 2 * 8, left_type[LBOT]) \
761  MAP_F2F(scan8[0] - 1 + 3 * 8, left_type[LBOT])
762 
763  if (FRAME_MBAFF(h)) {
764  if (MB_FIELD(h)) {
765 
766 #define MAP_F2F(idx, mb_type) \
767  if (!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0) { \
768  h->ref_cache[list][idx] <<= 1; \
769  h->mv_cache[list][idx][1] /= 2; \
770  h->mvd_cache[list][idx][1] >>= 1; \
771  }
772 
773  MAP_MVS
774  } else {
775 
776 #undef MAP_F2F
777 #define MAP_F2F(idx, mb_type) \
778  if (IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0) { \
779  h->ref_cache[list][idx] >>= 1; \
780  h->mv_cache[list][idx][1] <<= 1; \
781  h->mvd_cache[list][idx][1] <<= 1; \
782  }
783 
784  MAP_MVS
785 #undef MAP_F2F
786  }
787  }
788  }
789  }
790 
791  h->neighbor_transform_size = !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[LTOP]);
792 }
793 
794 /**
795  * decodes a P_SKIP or B_SKIP macroblock
796  */
798 {
799  const int mb_xy = h->mb_xy;
800  int mb_type = 0;
801 
802  memset(h->non_zero_count[mb_xy], 0, 48);
803 
804  if (MB_FIELD(h))
805  mb_type |= MB_TYPE_INTERLACED;
806 
807  if (h->slice_type_nos == AV_PICTURE_TYPE_B) {
808  // just for fill_caches. pred_direct_motion will set the real mb_type
810  if (h->direct_spatial_mv_pred) {
811  fill_decode_neighbors(h, mb_type);
812  fill_decode_caches(h, mb_type); //FIXME check what is needed and what not ...
813  }
814  ff_h264_pred_direct_motion(h, &mb_type);
815  mb_type |= MB_TYPE_SKIP;
816  } else {
818 
819  fill_decode_neighbors(h, mb_type);
821  }
822 
823  write_back_motion(h, mb_type);
824  h->cur_pic.mb_type[mb_xy] = mb_type;
825  h->cur_pic.qscale_table[mb_xy] = h->qscale;
826  h->slice_table[mb_xy] = h->slice_num;
827  h->prev_mb_skipped = 1;
828 }
829 
830 #endif /* AVCODEC_H264_MVPRED_H */