path: root/libavcodec/h264_mvpred.h (plain)
blob: bf395e3fe296f651eb6b83951ddea34405c31f28

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