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 "internal.h"
32 #include "avcodec.h"
33 #include "h264dec.h"
34 #include "mpegutils.h"
35 #include "libavutil/avassert.h"
36 #include "libavutil/mem_internal.h"
37
38
fetch_diagonal_mv(const H264Context * h,H264SliceContext * sl,const int16_t ** C,int i,int list,int part_width)39 static av_always_inline int fetch_diagonal_mv(const H264Context *h, H264SliceContext *sl,
40 const int16_t **C,
41 int i, int list, int part_width)
42 {
43 const int topright_ref = sl->ref_cache[list][i - 8 + part_width];
44
45 /* there is no consistent mapping of mvs to neighboring locations that will
46 * make mbaff happy, so we can't move all this logic to fill_caches */
47 if (FRAME_MBAFF(h)) {
48 #define SET_DIAG_MV(MV_OP, REF_OP, XY, Y4) \
49 const int xy = XY, y4 = Y4; \
50 const int mb_type = mb_types[xy + (y4 >> 2) * h->mb_stride]; \
51 if (!USES_LIST(mb_type, list)) \
52 return LIST_NOT_USED; \
53 mv = h->cur_pic_ptr->motion_val[list][h->mb2b_xy[xy] + 3 + y4 * h->b_stride]; \
54 sl->mv_cache[list][scan8[0] - 2][0] = mv[0]; \
55 sl->mv_cache[list][scan8[0] - 2][1] = mv[1] MV_OP; \
56 return h->cur_pic_ptr->ref_index[list][4 * xy + 1 + (y4 & ~1)] REF_OP;
57
58 if (topright_ref == PART_NOT_AVAILABLE
59 && i >= scan8[0] + 8 && (i & 7) == 4
60 && sl->ref_cache[list][scan8[0] - 1] != PART_NOT_AVAILABLE) {
61 const uint32_t *mb_types = h->cur_pic_ptr->mb_type;
62 const int16_t *mv;
63 AV_ZERO32(sl->mv_cache[list][scan8[0] - 2]);
64 *C = sl->mv_cache[list][scan8[0] - 2];
65
66 if (!MB_FIELD(sl) && IS_INTERLACED(sl->left_type[0])) {
67 SET_DIAG_MV(* 2, >> 1, sl->left_mb_xy[0] + h->mb_stride,
68 (sl->mb_y & 1) * 2 + (i >> 5));
69 }
70 if (MB_FIELD(sl) && !IS_INTERLACED(sl->left_type[0])) {
71 // left shift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's OK.
72 SET_DIAG_MV(/ 2, *2, sl->left_mb_xy[i >= 36], ((i >> 2)) & 3);
73 }
74 }
75 #undef SET_DIAG_MV
76 }
77
78 if (topright_ref != PART_NOT_AVAILABLE) {
79 *C = sl->mv_cache[list][i - 8 + part_width];
80 return topright_ref;
81 } else {
82 ff_tlog(h->avctx, "topright MV not available\n");
83
84 *C = sl->mv_cache[list][i - 8 - 1];
85 return sl->ref_cache[list][i - 8 - 1];
86 }
87 }
88
89 /**
90 * Get the predicted MV.
91 * @param n the block index
92 * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
93 * @param mx the x component of the predicted motion vector
94 * @param my the y component of the predicted motion vector
95 */
pred_motion(const H264Context * const h,H264SliceContext * sl,int n,int part_width,int list,int ref,int * const mx,int * const my)96 static av_always_inline void pred_motion(const H264Context *const h,
97 H264SliceContext *sl,
98 int n,
99 int part_width, int list, int ref,
100 int *const mx, int *const my)
101 {
102 const int index8 = scan8[n];
103 const int top_ref = sl->ref_cache[list][index8 - 8];
104 const int left_ref = sl->ref_cache[list][index8 - 1];
105 const int16_t *const A = sl->mv_cache[list][index8 - 1];
106 const int16_t *const B = sl->mv_cache[list][index8 - 8];
107 const int16_t *C;
108 int diagonal_ref, match_count;
109
110 av_assert2(part_width == 1 || part_width == 2 || part_width == 4);
111
112 /* mv_cache
113 * B . . A T T T T
114 * U . . L . . , .
115 * U . . L . . . .
116 * U . . L . . , .
117 * . . . L . . . .
118 */
119
120 diagonal_ref = fetch_diagonal_mv(h, sl, &C, index8, list, part_width);
121 match_count = (diagonal_ref == ref) + (top_ref == ref) + (left_ref == ref);
122 ff_tlog(h->avctx, "pred_motion match_count=%d\n", match_count);
123 if (match_count > 1) { //most common
124 *mx = mid_pred(A[0], B[0], C[0]);
125 *my = mid_pred(A[1], B[1], C[1]);
126 } else if (match_count == 1) {
127 if (left_ref == ref) {
128 *mx = A[0];
129 *my = A[1];
130 } else if (top_ref == ref) {
131 *mx = B[0];
132 *my = B[1];
133 } else {
134 *mx = C[0];
135 *my = C[1];
136 }
137 } else {
138 if (top_ref == PART_NOT_AVAILABLE &&
139 diagonal_ref == PART_NOT_AVAILABLE &&
140 left_ref != PART_NOT_AVAILABLE) {
141 *mx = A[0];
142 *my = A[1];
143 } else {
144 *mx = mid_pred(A[0], B[0], C[0]);
145 *my = mid_pred(A[1], B[1], C[1]);
146 }
147 }
148
149 ff_tlog(h->avctx,
150 "pred_motion (%2d %2d %2d) (%2d %2d %2d) (%2d %2d %2d) -> (%2d %2d %2d) at %2d %2d %d list %d\n",
151 top_ref, B[0], B[1], diagonal_ref, C[0], C[1], left_ref,
152 A[0], A[1], ref, *mx, *my, sl->mb_x, sl->mb_y, n, list);
153 }
154
155 /**
156 * Get the directionally predicted 16x8 MV.
157 * @param n the block index
158 * @param mx the x component of the predicted motion vector
159 * @param my the y component of the predicted motion vector
160 */
pred_16x8_motion(const H264Context * const h,H264SliceContext * sl,int n,int list,int ref,int * const mx,int * const my)161 static av_always_inline void pred_16x8_motion(const H264Context *const h,
162 H264SliceContext *sl,
163 int n, int list, int ref,
164 int *const mx, int *const my)
165 {
166 if (n == 0) {
167 const int top_ref = sl->ref_cache[list][scan8[0] - 8];
168 const int16_t *const B = sl->mv_cache[list][scan8[0] - 8];
169
170 ff_tlog(h->avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n",
171 top_ref, B[0], B[1], sl->mb_x, sl->mb_y, n, list);
172
173 if (top_ref == ref) {
174 *mx = B[0];
175 *my = B[1];
176 return;
177 }
178 } else {
179 const int left_ref = sl->ref_cache[list][scan8[8] - 1];
180 const int16_t *const A = sl->mv_cache[list][scan8[8] - 1];
181
182 ff_tlog(h->avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n",
183 left_ref, A[0], A[1], sl->mb_x, sl->mb_y, n, list);
184
185 if (left_ref == ref) {
186 *mx = A[0];
187 *my = A[1];
188 return;
189 }
190 }
191
192 //RARE
193 pred_motion(h, sl, n, 4, list, ref, mx, my);
194 }
195
196 /**
197 * Get the directionally predicted 8x16 MV.
198 * @param n the block index
199 * @param mx the x component of the predicted motion vector
200 * @param my the y component of the predicted motion vector
201 */
pred_8x16_motion(const H264Context * const h,H264SliceContext * sl,int n,int list,int ref,int * const mx,int * const my)202 static av_always_inline void pred_8x16_motion(const H264Context *const h,
203 H264SliceContext *sl,
204 int n, int list, int ref,
205 int *const mx, int *const my)
206 {
207 if (n == 0) {
208 const int left_ref = sl->ref_cache[list][scan8[0] - 1];
209 const int16_t *const A = sl->mv_cache[list][scan8[0] - 1];
210
211 ff_tlog(h->avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n",
212 left_ref, A[0], A[1], sl->mb_x, sl->mb_y, n, list);
213
214 if (left_ref == ref) {
215 *mx = A[0];
216 *my = A[1];
217 return;
218 }
219 } else {
220 const int16_t *C;
221 int diagonal_ref;
222
223 diagonal_ref = fetch_diagonal_mv(h, sl, &C, scan8[4], list, 2);
224
225 ff_tlog(h->avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n",
226 diagonal_ref, C[0], C[1], sl->mb_x, sl->mb_y, n, list);
227
228 if (diagonal_ref == ref) {
229 *mx = C[0];
230 *my = C[1];
231 return;
232 }
233 }
234
235 //RARE
236 pred_motion(h, sl, n, 2, list, ref, mx, my);
237 }
238
239 #define FIX_MV_MBAFF(type, refn, mvn, idx) \
240 if (FRAME_MBAFF(h)) { \
241 if (MB_FIELD(sl)) { \
242 if (!IS_INTERLACED(type)) { \
243 refn <<= 1; \
244 AV_COPY32(mvbuf[idx], mvn); \
245 mvbuf[idx][1] /= 2; \
246 mvn = mvbuf[idx]; \
247 } \
248 } else { \
249 if (IS_INTERLACED(type)) { \
250 refn >>= 1; \
251 AV_COPY32(mvbuf[idx], mvn); \
252 mvbuf[idx][1] *= 2; \
253 mvn = mvbuf[idx]; \
254 } \
255 } \
256 }
257
pred_pskip_motion(const H264Context * const h,H264SliceContext * sl)258 static av_always_inline void pred_pskip_motion(const H264Context *const h,
259 H264SliceContext *sl)
260 {
261 DECLARE_ALIGNED(4, static const int16_t, zeromv)[2] = { 0 };
262 DECLARE_ALIGNED(4, int16_t, mvbuf)[3][2];
263 int8_t *ref = h->cur_pic.ref_index[0];
264 int16_t(*mv)[2] = h->cur_pic.motion_val[0];
265 int top_ref, left_ref, diagonal_ref, match_count, mx, my;
266 const int16_t *A, *B, *C;
267 int b_stride = h->b_stride;
268
269 fill_rectangle(&sl->ref_cache[0][scan8[0]], 4, 4, 8, 0, 1);
270
271 /* To avoid doing an entire fill_decode_caches, we inline the relevant
272 * parts here.
273 * FIXME: this is a partial duplicate of the logic in fill_decode_caches,
274 * but it's faster this way. Is there a way to avoid this duplication?
275 */
276 if (USES_LIST(sl->left_type[LTOP], 0)) {
277 left_ref = ref[4 * sl->left_mb_xy[LTOP] + 1 + (sl->left_block[0] & ~1)];
278 A = mv[h->mb2b_xy[sl->left_mb_xy[LTOP]] + 3 + b_stride * sl->left_block[0]];
279 FIX_MV_MBAFF(sl->left_type[LTOP], left_ref, A, 0);
280 if (!(left_ref | AV_RN32A(A)))
281 goto zeromv;
282 } else if (sl->left_type[LTOP]) {
283 left_ref = LIST_NOT_USED;
284 A = zeromv;
285 } else {
286 goto zeromv;
287 }
288
289 if (USES_LIST(sl->top_type, 0)) {
290 top_ref = ref[4 * sl->top_mb_xy + 2];
291 B = mv[h->mb2b_xy[sl->top_mb_xy] + 3 * b_stride];
292 FIX_MV_MBAFF(sl->top_type, top_ref, B, 1);
293 if (!(top_ref | AV_RN32A(B)))
294 goto zeromv;
295 } else if (sl->top_type) {
296 top_ref = LIST_NOT_USED;
297 B = zeromv;
298 } else {
299 goto zeromv;
300 }
301
302 ff_tlog(h->avctx, "pred_pskip: (%d) (%d) at %2d %2d\n",
303 top_ref, left_ref, sl->mb_x, sl->mb_y);
304
305 if (USES_LIST(sl->topright_type, 0)) {
306 diagonal_ref = ref[4 * sl->topright_mb_xy + 2];
307 C = mv[h->mb2b_xy[sl->topright_mb_xy] + 3 * b_stride];
308 FIX_MV_MBAFF(sl->topright_type, diagonal_ref, C, 2);
309 } else if (sl->topright_type) {
310 diagonal_ref = LIST_NOT_USED;
311 C = zeromv;
312 } else {
313 if (USES_LIST(sl->topleft_type, 0)) {
314 diagonal_ref = ref[4 * sl->topleft_mb_xy + 1 +
315 (sl->topleft_partition & 2)];
316 C = mv[h->mb2b_xy[sl->topleft_mb_xy] + 3 + b_stride +
317 (sl->topleft_partition & 2 * b_stride)];
318 FIX_MV_MBAFF(sl->topleft_type, diagonal_ref, C, 2);
319 } else if (sl->topleft_type) {
320 diagonal_ref = LIST_NOT_USED;
321 C = zeromv;
322 } else {
323 diagonal_ref = PART_NOT_AVAILABLE;
324 C = zeromv;
325 }
326 }
327
328 match_count = !diagonal_ref + !top_ref + !left_ref;
329 ff_tlog(h->avctx, "pred_pskip_motion match_count=%d\n", match_count);
330 if (match_count > 1) {
331 mx = mid_pred(A[0], B[0], C[0]);
332 my = mid_pred(A[1], B[1], C[1]);
333 } else if (match_count == 1) {
334 if (!left_ref) {
335 mx = A[0];
336 my = A[1];
337 } else if (!top_ref) {
338 mx = B[0];
339 my = B[1];
340 } else {
341 mx = C[0];
342 my = C[1];
343 }
344 } else {
345 mx = mid_pred(A[0], B[0], C[0]);
346 my = mid_pred(A[1], B[1], C[1]);
347 }
348
349 fill_rectangle(sl->mv_cache[0][scan8[0]], 4, 4, 8, pack16to32(mx, my), 4);
350 return;
351
352 zeromv:
353 fill_rectangle(sl->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
354 return;
355 }
356
fill_decode_neighbors(const H264Context * h,H264SliceContext * sl,int mb_type)357 static void fill_decode_neighbors(const H264Context *h, H264SliceContext *sl, int mb_type)
358 {
359 const int mb_xy = sl->mb_xy;
360 int topleft_xy, top_xy, topright_xy, left_xy[LEFT_MBS];
361 static const uint8_t left_block_options[4][32] = {
362 { 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 },
363 { 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 },
364 { 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 },
365 { 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 }
366 };
367
368 sl->topleft_partition = -1;
369
370 top_xy = mb_xy - (h->mb_stride << MB_FIELD(sl));
371
372 /* Wow, what a mess, why didn't they simplify the interlacing & intra
373 * stuff, I can't imagine that these complex rules are worth it. */
374
375 topleft_xy = top_xy - 1;
376 topright_xy = top_xy + 1;
377 left_xy[LBOT] = left_xy[LTOP] = mb_xy - 1;
378 sl->left_block = left_block_options[0];
379 if (FRAME_MBAFF(h)) {
380 const int left_mb_field_flag = IS_INTERLACED(h->cur_pic.mb_type[mb_xy - 1]);
381 const int curr_mb_field_flag = IS_INTERLACED(mb_type);
382 if (sl->mb_y & 1) {
383 if (left_mb_field_flag != curr_mb_field_flag) {
384 left_xy[LBOT] = left_xy[LTOP] = mb_xy - h->mb_stride - 1;
385 if (curr_mb_field_flag) {
386 left_xy[LBOT] += h->mb_stride;
387 sl->left_block = left_block_options[3];
388 } else {
389 topleft_xy += h->mb_stride;
390 /* take top left mv from the middle of the mb, as opposed
391 * to all other modes which use the bottom right partition */
392 sl->topleft_partition = 0;
393 sl->left_block = left_block_options[1];
394 }
395 }
396 } else {
397 if (curr_mb_field_flag) {
398 topleft_xy += h->mb_stride & (((h->cur_pic.mb_type[top_xy - 1] >> 7) & 1) - 1);
399 topright_xy += h->mb_stride & (((h->cur_pic.mb_type[top_xy + 1] >> 7) & 1) - 1);
400 top_xy += h->mb_stride & (((h->cur_pic.mb_type[top_xy] >> 7) & 1) - 1);
401 }
402 if (left_mb_field_flag != curr_mb_field_flag) {
403 if (curr_mb_field_flag) {
404 left_xy[LBOT] += h->mb_stride;
405 sl->left_block = left_block_options[3];
406 } else {
407 sl->left_block = left_block_options[2];
408 }
409 }
410 }
411 }
412
413 sl->topleft_mb_xy = topleft_xy;
414 sl->top_mb_xy = top_xy;
415 sl->topright_mb_xy = topright_xy;
416 sl->left_mb_xy[LTOP] = left_xy[LTOP];
417 sl->left_mb_xy[LBOT] = left_xy[LBOT];
418 //FIXME do we need all in the context?
419
420 sl->topleft_type = h->cur_pic.mb_type[topleft_xy];
421 sl->top_type = h->cur_pic.mb_type[top_xy];
422 sl->topright_type = h->cur_pic.mb_type[topright_xy];
423 sl->left_type[LTOP] = h->cur_pic.mb_type[left_xy[LTOP]];
424 sl->left_type[LBOT] = h->cur_pic.mb_type[left_xy[LBOT]];
425
426 if (FMO) {
427 if (h->slice_table[topleft_xy] != sl->slice_num)
428 sl->topleft_type = 0;
429 if (h->slice_table[top_xy] != sl->slice_num)
430 sl->top_type = 0;
431 if (h->slice_table[left_xy[LTOP]] != sl->slice_num)
432 sl->left_type[LTOP] = sl->left_type[LBOT] = 0;
433 } else {
434 if (h->slice_table[topleft_xy] != sl->slice_num) {
435 sl->topleft_type = 0;
436 if (h->slice_table[top_xy] != sl->slice_num)
437 sl->top_type = 0;
438 if (h->slice_table[left_xy[LTOP]] != sl->slice_num)
439 sl->left_type[LTOP] = sl->left_type[LBOT] = 0;
440 }
441 }
442 if (h->slice_table[topright_xy] != sl->slice_num)
443 sl->topright_type = 0;
444 }
445
fill_decode_caches(const H264Context * h,H264SliceContext * sl,int mb_type)446 static void fill_decode_caches(const H264Context *h, H264SliceContext *sl, int mb_type)
447 {
448 int topleft_xy, top_xy, topright_xy, left_xy[LEFT_MBS];
449 int topleft_type, top_type, topright_type, left_type[LEFT_MBS];
450 const uint8_t *left_block = sl->left_block;
451 int i;
452 uint8_t *nnz;
453 uint8_t *nnz_cache;
454
455 topleft_xy = sl->topleft_mb_xy;
456 top_xy = sl->top_mb_xy;
457 topright_xy = sl->topright_mb_xy;
458 left_xy[LTOP] = sl->left_mb_xy[LTOP];
459 left_xy[LBOT] = sl->left_mb_xy[LBOT];
460 topleft_type = sl->topleft_type;
461 top_type = sl->top_type;
462 topright_type = sl->topright_type;
463 left_type[LTOP] = sl->left_type[LTOP];
464 left_type[LBOT] = sl->left_type[LBOT];
465
466 if (!IS_SKIP(mb_type)) {
467 if (IS_INTRA(mb_type)) {
468 int type_mask = h->ps.pps->constrained_intra_pred ? IS_INTRA(-1) : -1;
469 sl->topleft_samples_available =
470 sl->top_samples_available =
471 sl->left_samples_available = 0xFFFF;
472 sl->topright_samples_available = 0xEEEA;
473
474 if (!(top_type & type_mask)) {
475 sl->topleft_samples_available = 0xB3FF;
476 sl->top_samples_available = 0x33FF;
477 sl->topright_samples_available = 0x26EA;
478 }
479 if (IS_INTERLACED(mb_type) != IS_INTERLACED(left_type[LTOP])) {
480 if (IS_INTERLACED(mb_type)) {
481 if (!(left_type[LTOP] & type_mask)) {
482 sl->topleft_samples_available &= 0xDFFF;
483 sl->left_samples_available &= 0x5FFF;
484 }
485 if (!(left_type[LBOT] & type_mask)) {
486 sl->topleft_samples_available &= 0xFF5F;
487 sl->left_samples_available &= 0xFF5F;
488 }
489 } else {
490 int left_typei = h->cur_pic.mb_type[left_xy[LTOP] + h->mb_stride];
491
492 av_assert2(left_xy[LTOP] == left_xy[LBOT]);
493 if (!((left_typei & type_mask) && (left_type[LTOP] & type_mask))) {
494 sl->topleft_samples_available &= 0xDF5F;
495 sl->left_samples_available &= 0x5F5F;
496 }
497 }
498 } else {
499 if (!(left_type[LTOP] & type_mask)) {
500 sl->topleft_samples_available &= 0xDF5F;
501 sl->left_samples_available &= 0x5F5F;
502 }
503 }
504
505 if (!(topleft_type & type_mask))
506 sl->topleft_samples_available &= 0x7FFF;
507
508 if (!(topright_type & type_mask))
509 sl->topright_samples_available &= 0xFBFF;
510
511 if (IS_INTRA4x4(mb_type)) {
512 if (IS_INTRA4x4(top_type)) {
513 AV_COPY32(sl->intra4x4_pred_mode_cache + 4 + 8 * 0, sl->intra4x4_pred_mode + h->mb2br_xy[top_xy]);
514 } else {
515 sl->intra4x4_pred_mode_cache[4 + 8 * 0] =
516 sl->intra4x4_pred_mode_cache[5 + 8 * 0] =
517 sl->intra4x4_pred_mode_cache[6 + 8 * 0] =
518 sl->intra4x4_pred_mode_cache[7 + 8 * 0] = 2 - 3 * !(top_type & type_mask);
519 }
520 for (i = 0; i < 2; i++) {
521 if (IS_INTRA4x4(left_type[LEFT(i)])) {
522 int8_t *mode = sl->intra4x4_pred_mode + h->mb2br_xy[left_xy[LEFT(i)]];
523 sl->intra4x4_pred_mode_cache[3 + 8 * 1 + 2 * 8 * i] = mode[6 - left_block[0 + 2 * i]];
524 sl->intra4x4_pred_mode_cache[3 + 8 * 2 + 2 * 8 * i] = mode[6 - left_block[1 + 2 * i]];
525 } else {
526 sl->intra4x4_pred_mode_cache[3 + 8 * 1 + 2 * 8 * i] =
527 sl->intra4x4_pred_mode_cache[3 + 8 * 2 + 2 * 8 * i] = 2 - 3 * !(left_type[LEFT(i)] & type_mask);
528 }
529 }
530 }
531 }
532
533 /*
534 * 0 . T T. T T T T
535 * 1 L . .L . . . .
536 * 2 L . .L . . . .
537 * 3 . T TL . . . .
538 * 4 L . .L . . . .
539 * 5 L . .. . . . .
540 */
541 /* FIXME: constraint_intra_pred & partitioning & nnz
542 * (let us hope this is just a typo in the spec) */
543 nnz_cache = sl->non_zero_count_cache;
544 if (top_type) {
545 nnz = h->non_zero_count[top_xy];
546 AV_COPY32(&nnz_cache[4 + 8 * 0], &nnz[4 * 3]);
547 if (!h->chroma_y_shift) {
548 AV_COPY32(&nnz_cache[4 + 8 * 5], &nnz[4 * 7]);
549 AV_COPY32(&nnz_cache[4 + 8 * 10], &nnz[4 * 11]);
550 } else {
551 AV_COPY32(&nnz_cache[4 + 8 * 5], &nnz[4 * 5]);
552 AV_COPY32(&nnz_cache[4 + 8 * 10], &nnz[4 * 9]);
553 }
554 } else {
555 uint32_t top_empty = CABAC(h) && !IS_INTRA(mb_type) ? 0 : 0x40404040;
556 AV_WN32A(&nnz_cache[4 + 8 * 0], top_empty);
557 AV_WN32A(&nnz_cache[4 + 8 * 5], top_empty);
558 AV_WN32A(&nnz_cache[4 + 8 * 10], top_empty);
559 }
560
561 for (i = 0; i < 2; i++) {
562 if (left_type[LEFT(i)]) {
563 nnz = h->non_zero_count[left_xy[LEFT(i)]];
564 nnz_cache[3 + 8 * 1 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i]];
565 nnz_cache[3 + 8 * 2 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i]];
566 if (CHROMA444(h)) {
567 nnz_cache[3 + 8 * 6 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] + 4 * 4];
568 nnz_cache[3 + 8 * 7 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] + 4 * 4];
569 nnz_cache[3 + 8 * 11 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] + 8 * 4];
570 nnz_cache[3 + 8 * 12 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] + 8 * 4];
571 } else if (CHROMA422(h)) {
572 nnz_cache[3 + 8 * 6 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] - 2 + 4 * 4];
573 nnz_cache[3 + 8 * 7 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] - 2 + 4 * 4];
574 nnz_cache[3 + 8 * 11 + 2 * 8 * i] = nnz[left_block[8 + 0 + 2 * i] - 2 + 8 * 4];
575 nnz_cache[3 + 8 * 12 + 2 * 8 * i] = nnz[left_block[8 + 1 + 2 * i] - 2 + 8 * 4];
576 } else {
577 nnz_cache[3 + 8 * 6 + 8 * i] = nnz[left_block[8 + 4 + 2 * i]];
578 nnz_cache[3 + 8 * 11 + 8 * i] = nnz[left_block[8 + 5 + 2 * i]];
579 }
580 } else {
581 nnz_cache[3 + 8 * 1 + 2 * 8 * i] =
582 nnz_cache[3 + 8 * 2 + 2 * 8 * i] =
583 nnz_cache[3 + 8 * 6 + 2 * 8 * i] =
584 nnz_cache[3 + 8 * 7 + 2 * 8 * i] =
585 nnz_cache[3 + 8 * 11 + 2 * 8 * i] =
586 nnz_cache[3 + 8 * 12 + 2 * 8 * i] = CABAC(h) && !IS_INTRA(mb_type) ? 0 : 64;
587 }
588 }
589
590 if (CABAC(h)) {
591 // top_cbp
592 if (top_type)
593 sl->top_cbp = h->cbp_table[top_xy];
594 else
595 sl->top_cbp = IS_INTRA(mb_type) ? 0x7CF : 0x00F;
596 // left_cbp
597 if (left_type[LTOP]) {
598 sl->left_cbp = (h->cbp_table[left_xy[LTOP]] & 0x7F0) |
599 ((h->cbp_table[left_xy[LTOP]] >> (left_block[0] & (~1))) & 2) |
600 (((h->cbp_table[left_xy[LBOT]] >> (left_block[2] & (~1))) & 2) << 2);
601 } else {
602 sl->left_cbp = IS_INTRA(mb_type) ? 0x7CF : 0x00F;
603 }
604 }
605 }
606
607 if (IS_INTER(mb_type) || (IS_DIRECT(mb_type) && sl->direct_spatial_mv_pred)) {
608 int list;
609 int b_stride = h->b_stride;
610 for (list = 0; list < sl->list_count; list++) {
611 int8_t *ref_cache = &sl->ref_cache[list][scan8[0]];
612 int8_t *ref = h->cur_pic.ref_index[list];
613 int16_t(*mv_cache)[2] = &sl->mv_cache[list][scan8[0]];
614 int16_t(*mv)[2] = h->cur_pic.motion_val[list];
615 if (!USES_LIST(mb_type, list))
616 continue;
617 av_assert2(!(IS_DIRECT(mb_type) && !sl->direct_spatial_mv_pred));
618
619 if (USES_LIST(top_type, list)) {
620 const int b_xy = h->mb2b_xy[top_xy] + 3 * b_stride;
621 AV_COPY128(mv_cache[0 - 1 * 8], mv[b_xy + 0]);
622 ref_cache[0 - 1 * 8] =
623 ref_cache[1 - 1 * 8] = ref[4 * top_xy + 2];
624 ref_cache[2 - 1 * 8] =
625 ref_cache[3 - 1 * 8] = ref[4 * top_xy + 3];
626 } else {
627 AV_ZERO128(mv_cache[0 - 1 * 8]);
628 AV_WN32A(&ref_cache[0 - 1 * 8],
629 ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE) & 0xFF) * 0x01010101u);
630 }
631
632 if (mb_type & (MB_TYPE_16x8 | MB_TYPE_8x8)) {
633 for (i = 0; i < 2; i++) {
634 int cache_idx = -1 + i * 2 * 8;
635 if (USES_LIST(left_type[LEFT(i)], list)) {
636 const int b_xy = h->mb2b_xy[left_xy[LEFT(i)]] + 3;
637 const int b8_xy = 4 * left_xy[LEFT(i)] + 1;
638 AV_COPY32(mv_cache[cache_idx],
639 mv[b_xy + b_stride * left_block[0 + i * 2]]);
640 AV_COPY32(mv_cache[cache_idx + 8],
641 mv[b_xy + b_stride * left_block[1 + i * 2]]);
642 ref_cache[cache_idx] = ref[b8_xy + (left_block[0 + i * 2] & ~1)];
643 ref_cache[cache_idx + 8] = ref[b8_xy + (left_block[1 + i * 2] & ~1)];
644 } else {
645 AV_ZERO32(mv_cache[cache_idx]);
646 AV_ZERO32(mv_cache[cache_idx + 8]);
647 ref_cache[cache_idx] =
648 ref_cache[cache_idx + 8] = (left_type[LEFT(i)]) ? LIST_NOT_USED
649 : PART_NOT_AVAILABLE;
650 }
651 }
652 } else {
653 if (USES_LIST(left_type[LTOP], list)) {
654 const int b_xy = h->mb2b_xy[left_xy[LTOP]] + 3;
655 const int b8_xy = 4 * left_xy[LTOP] + 1;
656 AV_COPY32(mv_cache[-1], mv[b_xy + b_stride * left_block[0]]);
657 ref_cache[-1] = ref[b8_xy + (left_block[0] & ~1)];
658 } else {
659 AV_ZERO32(mv_cache[-1]);
660 ref_cache[-1] = left_type[LTOP] ? LIST_NOT_USED
661 : PART_NOT_AVAILABLE;
662 }
663 }
664
665 if (USES_LIST(topright_type, list)) {
666 const int b_xy = h->mb2b_xy[topright_xy] + 3 * b_stride;
667 AV_COPY32(mv_cache[4 - 1 * 8], mv[b_xy]);
668 ref_cache[4 - 1 * 8] = ref[4 * topright_xy + 2];
669 } else {
670 AV_ZERO32(mv_cache[4 - 1 * 8]);
671 ref_cache[4 - 1 * 8] = topright_type ? LIST_NOT_USED
672 : PART_NOT_AVAILABLE;
673 }
674 if(ref_cache[2 - 1*8] < 0 || ref_cache[4 - 1 * 8] < 0) {
675 if (USES_LIST(topleft_type, list)) {
676 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + b_stride +
677 (sl->topleft_partition & 2 * b_stride);
678 const int b8_xy = 4 * topleft_xy + 1 + (sl->topleft_partition & 2);
679 AV_COPY32(mv_cache[-1 - 1 * 8], mv[b_xy]);
680 ref_cache[-1 - 1 * 8] = ref[b8_xy];
681 } else {
682 AV_ZERO32(mv_cache[-1 - 1 * 8]);
683 ref_cache[-1 - 1 * 8] = topleft_type ? LIST_NOT_USED
684 : PART_NOT_AVAILABLE;
685 }
686 }
687
688 if ((mb_type & (MB_TYPE_SKIP | MB_TYPE_DIRECT2)) && !FRAME_MBAFF(h))
689 continue;
690
691 if (!(mb_type & (MB_TYPE_SKIP | MB_TYPE_DIRECT2))) {
692 uint8_t(*mvd_cache)[2] = &sl->mvd_cache[list][scan8[0]];
693 uint8_t(*mvd)[2] = sl->mvd_table[list];
694 ref_cache[2 + 8 * 0] =
695 ref_cache[2 + 8 * 2] = PART_NOT_AVAILABLE;
696 AV_ZERO32(mv_cache[2 + 8 * 0]);
697 AV_ZERO32(mv_cache[2 + 8 * 2]);
698
699 if (CABAC(h)) {
700 if (USES_LIST(top_type, list)) {
701 const int b_xy = h->mb2br_xy[top_xy];
702 AV_COPY64(mvd_cache[0 - 1 * 8], mvd[b_xy + 0]);
703 } else {
704 AV_ZERO64(mvd_cache[0 - 1 * 8]);
705 }
706 if (USES_LIST(left_type[LTOP], list)) {
707 const int b_xy = h->mb2br_xy[left_xy[LTOP]] + 6;
708 AV_COPY16(mvd_cache[-1 + 0 * 8], mvd[b_xy - left_block[0]]);
709 AV_COPY16(mvd_cache[-1 + 1 * 8], mvd[b_xy - left_block[1]]);
710 } else {
711 AV_ZERO16(mvd_cache[-1 + 0 * 8]);
712 AV_ZERO16(mvd_cache[-1 + 1 * 8]);
713 }
714 if (USES_LIST(left_type[LBOT], list)) {
715 const int b_xy = h->mb2br_xy[left_xy[LBOT]] + 6;
716 AV_COPY16(mvd_cache[-1 + 2 * 8], mvd[b_xy - left_block[2]]);
717 AV_COPY16(mvd_cache[-1 + 3 * 8], mvd[b_xy - left_block[3]]);
718 } else {
719 AV_ZERO16(mvd_cache[-1 + 2 * 8]);
720 AV_ZERO16(mvd_cache[-1 + 3 * 8]);
721 }
722 AV_ZERO16(mvd_cache[2 + 8 * 0]);
723 AV_ZERO16(mvd_cache[2 + 8 * 2]);
724 if (sl->slice_type_nos == AV_PICTURE_TYPE_B) {
725 uint8_t *direct_cache = &sl->direct_cache[scan8[0]];
726 uint8_t *direct_table = h->direct_table;
727 fill_rectangle(direct_cache, 4, 4, 8, MB_TYPE_16x16 >> 1, 1);
728
729 if (IS_DIRECT(top_type)) {
730 AV_WN32A(&direct_cache[-1 * 8],
731 0x01010101u * (MB_TYPE_DIRECT2 >> 1));
732 } else if (IS_8X8(top_type)) {
733 int b8_xy = 4 * top_xy;
734 direct_cache[0 - 1 * 8] = direct_table[b8_xy + 2];
735 direct_cache[2 - 1 * 8] = direct_table[b8_xy + 3];
736 } else {
737 AV_WN32A(&direct_cache[-1 * 8],
738 0x01010101 * (MB_TYPE_16x16 >> 1));
739 }
740
741 if (IS_DIRECT(left_type[LTOP]))
742 direct_cache[-1 + 0 * 8] = MB_TYPE_DIRECT2 >> 1;
743 else if (IS_8X8(left_type[LTOP]))
744 direct_cache[-1 + 0 * 8] = direct_table[4 * left_xy[LTOP] + 1 + (left_block[0] & ~1)];
745 else
746 direct_cache[-1 + 0 * 8] = MB_TYPE_16x16 >> 1;
747
748 if (IS_DIRECT(left_type[LBOT]))
749 direct_cache[-1 + 2 * 8] = MB_TYPE_DIRECT2 >> 1;
750 else if (IS_8X8(left_type[LBOT]))
751 direct_cache[-1 + 2 * 8] = direct_table[4 * left_xy[LBOT] + 1 + (left_block[2] & ~1)];
752 else
753 direct_cache[-1 + 2 * 8] = MB_TYPE_16x16 >> 1;
754 }
755 }
756 }
757
758 #define MAP_MVS \
759 MAP_F2F(scan8[0] - 1 - 1 * 8, topleft_type) \
760 MAP_F2F(scan8[0] + 0 - 1 * 8, top_type) \
761 MAP_F2F(scan8[0] + 1 - 1 * 8, top_type) \
762 MAP_F2F(scan8[0] + 2 - 1 * 8, top_type) \
763 MAP_F2F(scan8[0] + 3 - 1 * 8, top_type) \
764 MAP_F2F(scan8[0] + 4 - 1 * 8, topright_type) \
765 MAP_F2F(scan8[0] - 1 + 0 * 8, left_type[LTOP]) \
766 MAP_F2F(scan8[0] - 1 + 1 * 8, left_type[LTOP]) \
767 MAP_F2F(scan8[0] - 1 + 2 * 8, left_type[LBOT]) \
768 MAP_F2F(scan8[0] - 1 + 3 * 8, left_type[LBOT])
769
770 if (FRAME_MBAFF(h)) {
771 if (MB_FIELD(sl)) {
772
773 #define MAP_F2F(idx, mb_type) \
774 if (!IS_INTERLACED(mb_type) && sl->ref_cache[list][idx] >= 0) { \
775 sl->ref_cache[list][idx] *= 2; \
776 sl->mv_cache[list][idx][1] /= 2; \
777 sl->mvd_cache[list][idx][1] >>= 1; \
778 }
779
780 MAP_MVS
781 } else {
782
783 #undef MAP_F2F
784 #define MAP_F2F(idx, mb_type) \
785 if (IS_INTERLACED(mb_type) && sl->ref_cache[list][idx] >= 0) { \
786 sl->ref_cache[list][idx] >>= 1; \
787 sl->mv_cache[list][idx][1] *= 2; \
788 sl->mvd_cache[list][idx][1] <<= 1; \
789 }
790
791 MAP_MVS
792 #undef MAP_F2F
793 }
794 }
795 }
796 }
797
798 sl->neighbor_transform_size = !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[LTOP]);
799 }
800
801 /**
802 * decodes a P_SKIP or B_SKIP macroblock
803 */
decode_mb_skip(const H264Context * h,H264SliceContext * sl)804 static void av_unused decode_mb_skip(const H264Context *h, H264SliceContext *sl)
805 {
806 const int mb_xy = sl->mb_xy;
807 int mb_type = 0;
808
809 memset(h->non_zero_count[mb_xy], 0, 48);
810
811 if (MB_FIELD(sl))
812 mb_type |= MB_TYPE_INTERLACED;
813
814 if (sl->slice_type_nos == AV_PICTURE_TYPE_B) {
815 // just for fill_caches. pred_direct_motion will set the real mb_type
816 mb_type |= MB_TYPE_L0L1 | MB_TYPE_DIRECT2 | MB_TYPE_SKIP;
817 if (sl->direct_spatial_mv_pred) {
818 fill_decode_neighbors(h, sl, mb_type);
819 fill_decode_caches(h, sl, mb_type); //FIXME check what is needed and what not ...
820 }
821 ff_h264_pred_direct_motion(h, sl, &mb_type);
822 mb_type |= MB_TYPE_SKIP;
823 } else {
824 mb_type |= MB_TYPE_16x16 | MB_TYPE_P0L0 | MB_TYPE_P1L0 | MB_TYPE_SKIP;
825
826 fill_decode_neighbors(h, sl, mb_type);
827 pred_pskip_motion(h, sl);
828 }
829
830 write_back_motion(h, sl, mb_type);
831 h->cur_pic.mb_type[mb_xy] = mb_type;
832 h->cur_pic.qscale_table[mb_xy] = sl->qscale;
833 h->slice_table[mb_xy] = sl->slice_num;
834 sl->prev_mb_skipped = 1;
835 }
836
837 #endif /* AVCODEC_H264_MVPRED_H */
838