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