path: root/libavcodec/diracdec.c (plain)
blob: 202ae94922785f3ab4004204983391068e2040d1

1	/*
2	* Copyright (C) 2007 Marco Gerards <marco@gnu.org>
3	* Copyright (C) 2009 David Conrad
4	* Copyright (C) 2011 Jordi Ortiz
5	*
6	* This file is part of FFmpeg.
7	*
8	* FFmpeg is free software; you can redistribute it and/or
9	* modify it under the terms of the GNU Lesser General Public
10	* License as published by the Free Software Foundation; either
11	* version 2.1 of the License, or (at your option) any later version.
12	*
13	* FFmpeg is distributed in the hope that it will be useful,
14	* but WITHOUT ANY WARRANTY; without even the implied warranty of
15	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16	* Lesser General Public License for more details.
17	*
18	* You should have received a copy of the GNU Lesser General Public
19	* License along with FFmpeg; if not, write to the Free Software
20	* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21	*/
22
23	/**
24	* @file
25	* Dirac Decoder
26	* @author Marco Gerards <marco@gnu.org>, David Conrad, Jordi Ortiz <nenjordi@gmail.com>
27	*/
28
29	#include "libavutil/thread.h"
30	#include "avcodec.h"
31	#include "get_bits.h"
32	#include "bytestream.h"
33	#include "internal.h"
34	#include "golomb.h"
35	#include "dirac_arith.h"
36	#include "dirac_vlc.h"
37	#include "mpeg12data.h"
38	#include "libavcodec/mpegvideo.h"
39	#include "mpegvideoencdsp.h"
40	#include "dirac_dwt.h"
41	#include "dirac.h"
42	#include "diractab.h"
43	#include "diracdsp.h"
44	#include "videodsp.h"
45
46	/**
47	* The spec limits this to 3 for frame coding, but in practice can be as high as 6
48	*/
49	#define MAX_REFERENCE_FRAMES 8
50	#define MAX_DELAY 5 /* limit for main profile for frame coding (TODO: field coding) */
51	#define MAX_FRAMES (MAX_REFERENCE_FRAMES + MAX_DELAY + 1)
52	#define MAX_QUANT 255 /* max quant for VC-2 */
53	#define MAX_BLOCKSIZE 32 /* maximum xblen/yblen we support */
54
55	/**
56	* DiracBlock->ref flags, if set then the block does MC from the given ref
57	*/
58	#define DIRAC_REF_MASK_REF1 1
59	#define DIRAC_REF_MASK_REF2 2
60	#define DIRAC_REF_MASK_GLOBAL 4
61
62	/**
63	* Value of Picture.reference when Picture is not a reference picture, but
64	* is held for delayed output.
65	*/
66	#define DELAYED_PIC_REF 4
67
68	#define CALC_PADDING(size, depth) \
69	(((size + (1 << depth) - 1) >> depth) << depth)
70
71	#define DIVRNDUP(a, b) (((a) + (b) - 1) / (b))
72
73	typedef struct {
74	AVFrame *avframe;
75	int interpolated[3]; /* 1 if hpel[] is valid */
76	uint8_t *hpel[3][4];
77	uint8_t *hpel_base[3][4];
78	int reference;
79	} DiracFrame;
80
81	typedef struct {
82	union {
83	int16_t mv[2][2];
84	int16_t dc[3];
85	} u; /* anonymous unions aren't in C99 :( */
86	uint8_t ref;
87	} DiracBlock;
88
89	typedef struct SubBand {
90	int level;
91	int orientation;
92	int stride; /* in bytes */
93	int width;
94	int height;
95	int pshift;
96	int quant;
97	uint8_t *ibuf;
98	struct SubBand *parent;
99
100	/* for low delay */
101	unsigned length;
102	const uint8_t *coeff_data;
103	} SubBand;
104
105	typedef struct Plane {
106	DWTPlane idwt;
107
108	int width;
109	int height;
110	ptrdiff_t stride;
111
112	/* block length */
113	uint8_t xblen;
114	uint8_t yblen;
115	/* block separation (block n+1 starts after this many pixels in block n) */
116	uint8_t xbsep;
117	uint8_t ybsep;
118	/* amount of overspill on each edge (half of the overlap between blocks) */
119	uint8_t xoffset;
120	uint8_t yoffset;
121
122	SubBand band[MAX_DWT_LEVELS][4];
123	} Plane;
124
125	/* Used by Low Delay and High Quality profiles */
126	typedef struct DiracSlice {
127	GetBitContext gb;
128	int slice_x;
129	int slice_y;
130	int bytes;
131	} DiracSlice;
132
133	typedef struct DiracContext {
134	AVCodecContext *avctx;
135	MpegvideoEncDSPContext mpvencdsp;
136	VideoDSPContext vdsp;
137	DiracDSPContext diracdsp;
138	DiracGolombLUT *reader_ctx;
139	DiracVersionInfo version;
140	GetBitContext gb;
141	AVDiracSeqHeader seq;
142	int seen_sequence_header;
143	int frame_number; /* number of the next frame to display */
144	Plane plane[3];
145	int chroma_x_shift;
146	int chroma_y_shift;
147
148	int bit_depth; /* bit depth */
149	int pshift; /* pixel shift = bit_depth > 8 */
150
151	int zero_res; /* zero residue flag */
152	int is_arith; /* whether coeffs use arith or golomb coding */
153	int core_syntax; /* use core syntax only */
154	int low_delay; /* use the low delay syntax */
155	int hq_picture; /* high quality picture, enables low_delay */
156	int ld_picture; /* use low delay picture, turns on low_delay */
157	int dc_prediction; /* has dc prediction */
158	int globalmc_flag; /* use global motion compensation */
159	int num_refs; /* number of reference pictures */
160
161	/* wavelet decoding */
162	unsigned wavelet_depth; /* depth of the IDWT */
163	unsigned wavelet_idx;
164
165	/**
166	* schroedinger older than 1.0.8 doesn't store
167	* quant delta if only one codebook exists in a band
168	*/
169	unsigned old_delta_quant;
170	unsigned codeblock_mode;
171
172	unsigned num_x; /* number of horizontal slices */
173	unsigned num_y; /* number of vertical slices */
174
175	uint8_t *thread_buf; /* Per-thread buffer for coefficient storage */
176	int threads_num_buf; /* Current # of buffers allocated */
177	int thread_buf_size; /* Each thread has a buffer this size */
178
179	DiracSlice *slice_params_buf;
180	int slice_params_num_buf;
181
182	struct {
183	unsigned width;
184	unsigned height;
185	} codeblock[MAX_DWT_LEVELS+1];
186
187	struct {
188	AVRational bytes; /* average bytes per slice */
189	uint8_t quant[MAX_DWT_LEVELS][4]; /* [DIRAC_STD] E.1 */
190	} lowdelay;
191
192	struct {
193	unsigned prefix_bytes;
194	uint64_t size_scaler;
195	} highquality;
196
197	struct {
198	int pan_tilt[2]; /* pan/tilt vector */
199	int zrs[2][2]; /* zoom/rotate/shear matrix */
200	int perspective[2]; /* perspective vector */
201	unsigned zrs_exp;
202	unsigned perspective_exp;
203	} globalmc[2];
204
205	/* motion compensation */
206	uint8_t mv_precision; /* [DIRAC_STD] REFS_WT_PRECISION */
207	int16_t weight[2]; /* [DIRAC_STD] REF1_WT and REF2_WT */
208	unsigned weight_log2denom; /* [DIRAC_STD] REFS_WT_PRECISION */
209
210	int blwidth; /* number of blocks (horizontally) */
211	int blheight; /* number of blocks (vertically) */
212	int sbwidth; /* number of superblocks (horizontally) */
213	int sbheight; /* number of superblocks (vertically) */
214
215	uint8_t *sbsplit;
216	DiracBlock *blmotion;
217
218	uint8_t *edge_emu_buffer[4];
219	uint8_t *edge_emu_buffer_base;
220
221	uint16_t *mctmp; /* buffer holding the MC data multiplied by OBMC weights */
222	uint8_t *mcscratch;
223	int buffer_stride;
224
225	DECLARE_ALIGNED(16, uint8_t, obmc_weight)[3][MAX_BLOCKSIZE*MAX_BLOCKSIZE];
226
227	void (*put_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
228	void (*avg_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
229	void (*add_obmc)(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
230	dirac_weight_func weight_func;
231	dirac_biweight_func biweight_func;
232
233	DiracFrame *current_picture;
234	DiracFrame *ref_pics[2];
235
236	DiracFrame *ref_frames[MAX_REFERENCE_FRAMES+1];
237	DiracFrame *delay_frames[MAX_DELAY+1];
238	DiracFrame all_frames[MAX_FRAMES];
239	} DiracContext;
240
241	enum dirac_subband {
242	subband_ll = 0,
243	subband_hl = 1,
244	subband_lh = 2,
245	subband_hh = 3,
246	subband_nb,
247	};
248
249	/* magic number division by 3 from schroedinger */
250	static inline int divide3(int x)
251	{
252	return ((x+1)*21845 + 10922) >> 16;
253	}
254
255	static DiracFrame *remove_frame(DiracFrame *framelist[], int picnum)
256	{
257	DiracFrame *remove_pic = NULL;
258	int i, remove_idx = -1;
259
260	for (i = 0; framelist[i]; i++)
261	if (framelist[i]->avframe->display_picture_number == picnum) {
262	remove_pic = framelist[i];
263	remove_idx = i;
264	}
265
266	if (remove_pic)
267	for (i = remove_idx; framelist[i]; i++)
268	framelist[i] = framelist[i+1];
269
270	return remove_pic;
271	}
272
273	static int add_frame(DiracFrame *framelist[], int maxframes, DiracFrame *frame)
274	{
275	int i;
276	for (i = 0; i < maxframes; i++)
277	if (!framelist[i]) {
278	framelist[i] = frame;
279	return 0;
280	}
281	return -1;
282	}
283
284	static int alloc_sequence_buffers(DiracContext *s)
285	{
286	int sbwidth = DIVRNDUP(s->seq.width, 4);
287	int sbheight = DIVRNDUP(s->seq.height, 4);
288	int i, w, h, top_padding;
289
290	/* todo: think more about this / use or set Plane here */
291	for (i = 0; i < 3; i++) {
292	int max_xblen = MAX_BLOCKSIZE >> (i ? s->chroma_x_shift : 0);
293	int max_yblen = MAX_BLOCKSIZE >> (i ? s->chroma_y_shift : 0);
294	w = s->seq.width >> (i ? s->chroma_x_shift : 0);
295	h = s->seq.height >> (i ? s->chroma_y_shift : 0);
296
297	/* we allocate the max we support here since num decompositions can
298	* change from frame to frame. Stride is aligned to 16 for SIMD, and
299	* 1<<MAX_DWT_LEVELS top padding to avoid if(y>0) in arith decoding
300	* MAX_BLOCKSIZE padding for MC: blocks can spill up to half of that
301	* on each side */
302	top_padding = FFMAX(1<<MAX_DWT_LEVELS, max_yblen/2);
303	w = FFALIGN(CALC_PADDING(w, MAX_DWT_LEVELS), 8); /* FIXME: Should this be 16 for SSE??? */
304	h = top_padding + CALC_PADDING(h, MAX_DWT_LEVELS) + max_yblen/2;
305
306	s->plane[i].idwt.buf_base = av_mallocz_array((w+max_xblen), h * (2 << s->pshift));
307	s->plane[i].idwt.tmp = av_malloc_array((w+16), 2 << s->pshift);
308	s->plane[i].idwt.buf = s->plane[i].idwt.buf_base + (top_padding*w)*(2 << s->pshift);
309	if (!s->plane[i].idwt.buf_base || !s->plane[i].idwt.tmp)
310	return AVERROR(ENOMEM);
311	}
312
313	/* fixme: allocate using real stride here */
314	s->sbsplit = av_malloc_array(sbwidth, sbheight);
315	s->blmotion = av_malloc_array(sbwidth, sbheight * 16 * sizeof(*s->blmotion));
316
317	if (!s->sbsplit || !s->blmotion)
318	return AVERROR(ENOMEM);
319	return 0;
320	}
321
322	static int alloc_buffers(DiracContext *s, int stride)
323	{
324	int w = s->seq.width;
325	int h = s->seq.height;
326
327	av_assert0(stride >= w);
328	stride += 64;
329
330	if (s->buffer_stride >= stride)
331	return 0;
332	s->buffer_stride = 0;
333
334	av_freep(&s->edge_emu_buffer_base);
335	memset(s->edge_emu_buffer, 0, sizeof(s->edge_emu_buffer));
336	av_freep(&s->mctmp);
337	av_freep(&s->mcscratch);
338
339	s->edge_emu_buffer_base = av_malloc_array(stride, MAX_BLOCKSIZE);
340
341	s->mctmp = av_malloc_array((stride+MAX_BLOCKSIZE), (h+MAX_BLOCKSIZE) * sizeof(*s->mctmp));
342	s->mcscratch = av_malloc_array(stride, MAX_BLOCKSIZE);
343
344	if (!s->edge_emu_buffer_base || !s->mctmp || !s->mcscratch)
345	return AVERROR(ENOMEM);
346
347	s->buffer_stride = stride;
348	return 0;
349	}
350
351	static void free_sequence_buffers(DiracContext *s)
352	{
353	int i, j, k;
354
355	for (i = 0; i < MAX_FRAMES; i++) {
356	if (s->all_frames[i].avframe->data[0]) {
357	av_frame_unref(s->all_frames[i].avframe);
358	memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));
359	}
360
361	for (j = 0; j < 3; j++)
362	for (k = 1; k < 4; k++)
363	av_freep(&s->all_frames[i].hpel_base[j][k]);
364	}
365
366	memset(s->ref_frames, 0, sizeof(s->ref_frames));
367	memset(s->delay_frames, 0, sizeof(s->delay_frames));
368
369	for (i = 0; i < 3; i++) {
370	av_freep(&s->plane[i].idwt.buf_base);
371	av_freep(&s->plane[i].idwt.tmp);
372	}
373
374	s->buffer_stride = 0;
375	av_freep(&s->sbsplit);
376	av_freep(&s->blmotion);
377	av_freep(&s->edge_emu_buffer_base);
378
379	av_freep(&s->mctmp);
380	av_freep(&s->mcscratch);
381	}
382
383	static AVOnce dirac_arith_init = AV_ONCE_INIT;
384
385	static av_cold int dirac_decode_init(AVCodecContext *avctx)
386	{
387	DiracContext *s = avctx->priv_data;
388	int i, ret;
389
390	s->avctx = avctx;
391	s->frame_number = -1;
392
393	s->thread_buf = NULL;
394	s->threads_num_buf = -1;
395	s->thread_buf_size = -1;
396
397	ff_dirac_golomb_reader_init(&s->reader_ctx);
398	ff_diracdsp_init(&s->diracdsp);
399	ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
400	ff_videodsp_init(&s->vdsp, 8);
401
402	for (i = 0; i < MAX_FRAMES; i++) {
403	s->all_frames[i].avframe = av_frame_alloc();
404	if (!s->all_frames[i].avframe) {
405	while (i > 0)
406	av_frame_free(&s->all_frames[--i].avframe);
407	return AVERROR(ENOMEM);
408	}
409	}
410	ret = ff_thread_once(&dirac_arith_init, ff_dirac_init_arith_tables);
411	if (ret != 0)
412	return AVERROR_UNKNOWN;
413
414	return 0;
415	}
416
417	static void dirac_decode_flush(AVCodecContext *avctx)
418	{
419	DiracContext *s = avctx->priv_data;
420	free_sequence_buffers(s);
421	s->seen_sequence_header = 0;
422	s->frame_number = -1;
423	}
424
425	static av_cold int dirac_decode_end(AVCodecContext *avctx)
426	{
427	DiracContext *s = avctx->priv_data;
428	int i;
429
430	ff_dirac_golomb_reader_end(&s->reader_ctx);
431
432	dirac_decode_flush(avctx);
433	for (i = 0; i < MAX_FRAMES; i++)
434	av_frame_free(&s->all_frames[i].avframe);
435
436	av_freep(&s->thread_buf);
437	av_freep(&s->slice_params_buf);
438
439	return 0;
440	}
441
442	static inline int coeff_unpack_golomb(GetBitContext *gb, int qfactor, int qoffset)
443	{
444	int coeff = dirac_get_se_golomb(gb);
445	const int sign = FFSIGN(coeff);
446	if (coeff)
447	coeff = sign*((sign * coeff * qfactor + qoffset) >> 2);
448	return coeff;
449	}
450
451	#define SIGN_CTX(x) (CTX_SIGN_ZERO + ((x) > 0) - ((x) < 0))
452
453	#define UNPACK_ARITH(n, type) \
454	static inline void coeff_unpack_arith_##n(DiracArith *c, int qfactor, int qoffset, \
455	SubBand *b, type *buf, int x, int y) \
456	{ \
457	int coeff, sign, sign_pred = 0, pred_ctx = CTX_ZPZN_F1; \
458	const int mstride = -(b->stride >> (1+b->pshift)); \
459	if (b->parent) { \
460	const type *pbuf = (type *)b->parent->ibuf; \
461	const int stride = b->parent->stride >> (1+b->parent->pshift); \
462	pred_ctx += !!pbuf[stride * (y>>1) + (x>>1)] << 1; \
463	} \
464	if (b->orientation == subband_hl) \
465	sign_pred = buf[mstride]; \
466	if (x) { \
467	pred_ctx += !(buf[-1] | buf[mstride] | buf[-1 + mstride]); \
468	if (b->orientation == subband_lh) \
469	sign_pred = buf[-1]; \
470	} else { \
471	pred_ctx += !buf[mstride]; \
472	} \
473	coeff = dirac_get_arith_uint(c, pred_ctx, CTX_COEFF_DATA); \
474	if (coeff) { \
475	coeff = (coeff * qfactor + qoffset) >> 2; \
476	sign = dirac_get_arith_bit(c, SIGN_CTX(sign_pred)); \
477	coeff = (coeff ^ -sign) + sign; \
478	} \
479	*buf = coeff; \
480	} \
481
482	UNPACK_ARITH(8, int16_t)
483	UNPACK_ARITH(10, int32_t)
484
485	/**
486	* Decode the coeffs in the rectangle defined by left, right, top, bottom
487	* [DIRAC_STD] 13.4.3.2 Codeblock unpacking loop. codeblock()
488	*/
489	static inline void codeblock(DiracContext *s, SubBand *b,
490	GetBitContext *gb, DiracArith *c,
491	int left, int right, int top, int bottom,
492	int blockcnt_one, int is_arith)
493	{
494	int x, y, zero_block;
495	int qoffset, qfactor;
496	uint8_t *buf;
497
498	/* check for any coded coefficients in this codeblock */
499	if (!blockcnt_one) {
500	if (is_arith)
501	zero_block = dirac_get_arith_bit(c, CTX_ZERO_BLOCK);
502	else
503	zero_block = get_bits1(gb);
504
505	if (zero_block)
506	return;
507	}
508
509	if (s->codeblock_mode && !(s->old_delta_quant && blockcnt_one)) {
510	int quant = b->quant;
511	if (is_arith)
512	quant += dirac_get_arith_int(c, CTX_DELTA_Q_F, CTX_DELTA_Q_DATA);
513	else
514	quant += dirac_get_se_golomb(gb);
515	if (quant < 0) {
516	av_log(s->avctx, AV_LOG_ERROR, "Invalid quant\n");
517	return;
518	}
519	b->quant = quant;
520	}
521
522	if (b->quant > (DIRAC_MAX_QUANT_INDEX - 1)) {
523	av_log(s->avctx, AV_LOG_ERROR, "Unsupported quant %d\n", b->quant);
524	b->quant = 0;
525	return;
526	}
527
528	qfactor = ff_dirac_qscale_tab[b->quant];
529	/* TODO: context pointer? */
530	if (!s->num_refs)
531	qoffset = ff_dirac_qoffset_intra_tab[b->quant] + 2;
532	else
533	qoffset = ff_dirac_qoffset_inter_tab[b->quant] + 2;
534
535	buf = b->ibuf + top * b->stride;
536	if (is_arith) {
537	for (y = top; y < bottom; y++) {
538	for (x = left; x < right; x++) {
539	if (b->pshift) {
540	coeff_unpack_arith_10(c, qfactor, qoffset, b, (int32_t*)(buf)+x, x, y);
541	} else {
542	coeff_unpack_arith_8(c, qfactor, qoffset, b, (int16_t*)(buf)+x, x, y);
543	}
544	}
545	buf += b->stride;
546	}
547	} else {
548	for (y = top; y < bottom; y++) {
549	for (x = left; x < right; x++) {
550	int val = coeff_unpack_golomb(gb, qfactor, qoffset);
551	if (b->pshift) {
552	AV_WN32(&buf[4*x], val);
553	} else {
554	AV_WN16(&buf[2*x], val);
555	}
556	}
557	buf += b->stride;
558	}
559	}
560	}
561
562	/**
563	* Dirac Specification ->
564	* 13.3 intra_dc_prediction(band)
565	*/
566	#define INTRA_DC_PRED(n, type) \
567	static inline void intra_dc_prediction_##n(SubBand *b) \
568	{ \
569	type *buf = (type*)b->ibuf; \
570	int x, y; \
571	\
572	for (x = 1; x < b->width; x++) \
573	buf[x] += buf[x-1]; \
574	buf += (b->stride >> (1+b->pshift)); \
575	\
576	for (y = 1; y < b->height; y++) { \
577	buf[0] += buf[-(b->stride >> (1+b->pshift))]; \
578	\
579	for (x = 1; x < b->width; x++) { \
580	int pred = buf[x - 1] + buf[x - (b->stride >> (1+b->pshift))] + buf[x - (b->stride >> (1+b->pshift))-1]; \
581	buf[x] += divide3(pred); \
582	} \
583	buf += (b->stride >> (1+b->pshift)); \
584	} \
585	} \
586
587	INTRA_DC_PRED(8, int16_t)
588	INTRA_DC_PRED(10, int32_t)
589
590	/**
591	* Dirac Specification ->
592	* 13.4.2 Non-skipped subbands. subband_coeffs()
593	*/
594	static av_always_inline void decode_subband_internal(DiracContext *s, SubBand *b, int is_arith)
595	{
596	int cb_x, cb_y, left, right, top, bottom;
597	DiracArith c;
598	GetBitContext gb;
599	int cb_width = s->codeblock[b->level + (b->orientation != subband_ll)].width;
600	int cb_height = s->codeblock[b->level + (b->orientation != subband_ll)].height;
601	int blockcnt_one = (cb_width + cb_height) == 2;
602
603	if (!b->length)
604	return;
605
606	init_get_bits8(&gb, b->coeff_data, b->length);
607
608	if (is_arith)
609	ff_dirac_init_arith_decoder(&c, &gb, b->length);
610
611	top = 0;
612	for (cb_y = 0; cb_y < cb_height; cb_y++) {
613	bottom = (b->height * (cb_y+1LL)) / cb_height;
614	left = 0;
615	for (cb_x = 0; cb_x < cb_width; cb_x++) {
616	right = (b->width * (cb_x+1LL)) / cb_width;
617	codeblock(s, b, &gb, &c, left, right, top, bottom, blockcnt_one, is_arith);
618	left = right;
619	}
620	top = bottom;
621	}
622
623	if (b->orientation == subband_ll && s->num_refs == 0) {
624	if (s->pshift) {
625	intra_dc_prediction_10(b);
626	} else {
627	intra_dc_prediction_8(b);
628	}
629	}
630	}
631
632	static int decode_subband_arith(AVCodecContext *avctx, void *b)
633	{
634	DiracContext *s = avctx->priv_data;
635	decode_subband_internal(s, b, 1);
636	return 0;
637	}
638
639	static int decode_subband_golomb(AVCodecContext *avctx, void *arg)
640	{
641	DiracContext *s = avctx->priv_data;
642	SubBand **b = arg;
643	decode_subband_internal(s, *b, 0);
644	return 0;
645	}
646
647	/**
648	* Dirac Specification ->
649	* [DIRAC_STD] 13.4.1 core_transform_data()
650	*/
651	static void decode_component(DiracContext *s, int comp)
652	{
653	AVCodecContext *avctx = s->avctx;
654	SubBand *bands[3*MAX_DWT_LEVELS+1];
655	enum dirac_subband orientation;
656	int level, num_bands = 0;
657
658	/* Unpack all subbands at all levels. */
659	for (level = 0; level < s->wavelet_depth; level++) {
660	for (orientation = !!level; orientation < 4; orientation++) {
661	SubBand *b = &s->plane[comp].band[level][orientation];
662	bands[num_bands++] = b;
663
664	align_get_bits(&s->gb);
665	/* [DIRAC_STD] 13.4.2 subband() */
666	b->length = get_interleaved_ue_golomb(&s->gb);
667	if (b->length) {
668	b->quant = get_interleaved_ue_golomb(&s->gb);
669	align_get_bits(&s->gb);
670	b->coeff_data = s->gb.buffer + get_bits_count(&s->gb)/8;
671	b->length = FFMIN(b->length, FFMAX(get_bits_left(&s->gb)/8, 0));
672	skip_bits_long(&s->gb, b->length*8);
673	}
674	}
675	/* arithmetic coding has inter-level dependencies, so we can only execute one level at a time */
676	if (s->is_arith)
677	avctx->execute(avctx, decode_subband_arith, &s->plane[comp].band[level][!!level],
678	NULL, 4-!!level, sizeof(SubBand));
679	}
680	/* golomb coding has no inter-level dependencies, so we can execute all subbands in parallel */
681	if (!s->is_arith)
682	avctx->execute(avctx, decode_subband_golomb, bands, NULL, num_bands, sizeof(SubBand*));
683	}
684
685	#define PARSE_VALUES(type, x, gb, ebits, buf1, buf2) \
686	type *buf = (type *)buf1; \
687	buf[x] = coeff_unpack_golomb(gb, qfactor, qoffset); \
688	if (get_bits_count(gb) >= ebits) \
689	return; \
690	if (buf2) { \
691	buf = (type *)buf2; \
692	buf[x] = coeff_unpack_golomb(gb, qfactor, qoffset); \
693	if (get_bits_count(gb) >= ebits) \
694	return; \
695	} \
696
697	static void decode_subband(DiracContext *s, GetBitContext *gb, int quant,
698	int slice_x, int slice_y, int bits_end,
699	SubBand *b1, SubBand *b2)
700	{
701	int left = b1->width * slice_x / s->num_x;
702	int right = b1->width *(slice_x+1) / s->num_x;
703	int top = b1->height * slice_y / s->num_y;
704	int bottom = b1->height *(slice_y+1) / s->num_y;
705
706	int qfactor, qoffset;
707
708	uint8_t *buf1 = b1->ibuf + top * b1->stride;
709	uint8_t *buf2 = b2 ? b2->ibuf + top * b2->stride: NULL;
710	int x, y;
711
712	if (quant > (DIRAC_MAX_QUANT_INDEX - 1)) {
713	av_log(s->avctx, AV_LOG_ERROR, "Unsupported quant %d\n", quant);
714	return;
715	}
716	qfactor = ff_dirac_qscale_tab[quant];
717	qoffset = ff_dirac_qoffset_intra_tab[quant] + 2;
718	/* we have to constantly check for overread since the spec explicitly
719	requires this, with the meaning that all remaining coeffs are set to 0 */
720	if (get_bits_count(gb) >= bits_end)
721	return;
722
723	if (s->pshift) {
724	for (y = top; y < bottom; y++) {
725	for (x = left; x < right; x++) {
726	PARSE_VALUES(int32_t, x, gb, bits_end, buf1, buf2);
727	}
728	buf1 += b1->stride;
729	if (buf2)
730	buf2 += b2->stride;
731	}
732	}
733	else {
734	for (y = top; y < bottom; y++) {
735	for (x = left; x < right; x++) {
736	PARSE_VALUES(int16_t, x, gb, bits_end, buf1, buf2);
737	}
738	buf1 += b1->stride;
739	if (buf2)
740	buf2 += b2->stride;
741	}
742	}
743	}
744
745	/**
746	* Dirac Specification ->
747	* 13.5.2 Slices. slice(sx,sy)
748	*/
749	static int decode_lowdelay_slice(AVCodecContext *avctx, void *arg)
750	{
751	DiracContext *s = avctx->priv_data;
752	DiracSlice *slice = arg;
753	GetBitContext *gb = &slice->gb;
754	enum dirac_subband orientation;
755	int level, quant, chroma_bits, chroma_end;
756
757	int quant_base = get_bits(gb, 7); /*[DIRAC_STD] qindex */
758	int length_bits = av_log2(8 * slice->bytes)+1;
759	int luma_bits = get_bits_long(gb, length_bits);
760	int luma_end = get_bits_count(gb) + FFMIN(luma_bits, get_bits_left(gb));
761
762	/* [DIRAC_STD] 13.5.5.2 luma_slice_band */
763	for (level = 0; level < s->wavelet_depth; level++)
764	for (orientation = !!level; orientation < 4; orientation++) {
765	quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);
766	decode_subband(s, gb, quant, slice->slice_x, slice->slice_y, luma_end,
767	&s->plane[0].band[level][orientation], NULL);
768	}
769
770	/* consume any unused bits from luma */
771	skip_bits_long(gb, get_bits_count(gb) - luma_end);
772
773	chroma_bits = 8*slice->bytes - 7 - length_bits - luma_bits;
774	chroma_end = get_bits_count(gb) + FFMIN(chroma_bits, get_bits_left(gb));
775	/* [DIRAC_STD] 13.5.5.3 chroma_slice_band */
776	for (level = 0; level < s->wavelet_depth; level++)
777	for (orientation = !!level; orientation < 4; orientation++) {
778	quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);
779	decode_subband(s, gb, quant, slice->slice_x, slice->slice_y, chroma_end,
780	&s->plane[1].band[level][orientation],
781	&s->plane[2].band[level][orientation]);
782	}
783
784	return 0;
785	}
786
787	typedef struct SliceCoeffs {
788	int left;
789	int top;
790	int tot_h;
791	int tot_v;
792	int tot;
793	} SliceCoeffs;
794
795	static int subband_coeffs(DiracContext *s, int x, int y, int p,
796	SliceCoeffs c[MAX_DWT_LEVELS])
797	{
798	int level, coef = 0;
799	for (level = 0; level < s->wavelet_depth; level++) {
800	SliceCoeffs *o = &c[level];
801	SubBand *b = &s->plane[p].band[level][3]; /* orientation doens't matter */
802	o->top = b->height * y / s->num_y;
803	o->left = b->width * x / s->num_x;
804	o->tot_h = ((b->width * (x + 1)) / s->num_x) - o->left;
805	o->tot_v = ((b->height * (y + 1)) / s->num_y) - o->top;
806	o->tot = o->tot_h*o->tot_v;
807	coef += o->tot * (4 - !!level);
808	}
809	return coef;
810	}
811
812	/**
813	* VC-2 Specification ->
814	* 13.5.3 hq_slice(sx,sy)
815	*/
816	static int decode_hq_slice(DiracContext *s, DiracSlice *slice, uint8_t *tmp_buf)
817	{
818	int i, level, orientation, quant_idx;
819	int qfactor[MAX_DWT_LEVELS][4], qoffset[MAX_DWT_LEVELS][4];
820	GetBitContext *gb = &slice->gb;
821	SliceCoeffs coeffs_num[MAX_DWT_LEVELS];
822
823	skip_bits_long(gb, 8*s->highquality.prefix_bytes);
824	quant_idx = get_bits(gb, 8);
825
826	if (quant_idx > DIRAC_MAX_QUANT_INDEX) {
827	av_log(s->avctx, AV_LOG_ERROR, "Invalid quantization index - %i\n", quant_idx);
828	return AVERROR_INVALIDDATA;
829	}
830
831	/* Slice quantization (slice_quantizers() in the specs) */
832	for (level = 0; level < s->wavelet_depth; level++) {
833	for (orientation = !!level; orientation < 4; orientation++) {
834	const int quant = FFMAX(quant_idx - s->lowdelay.quant[level][orientation], 0);
835	qfactor[level][orientation] = ff_dirac_qscale_tab[quant];
836	qoffset[level][orientation] = ff_dirac_qoffset_intra_tab[quant] + 2;
837	}
838	}
839
840	/* Luma + 2 Chroma planes */
841	for (i = 0; i < 3; i++) {
842	int coef_num, coef_par, off = 0;
843	int64_t length = s->highquality.size_scaler*get_bits(gb, 8);
844	int64_t bits_end = get_bits_count(gb) + 8*length;
845	const uint8_t *addr = align_get_bits(gb);
846
847	if (length*8 > get_bits_left(gb)) {
848	av_log(s->avctx, AV_LOG_ERROR, "end too far away\n");
849	return AVERROR_INVALIDDATA;
850	}
851
852	coef_num = subband_coeffs(s, slice->slice_x, slice->slice_y, i, coeffs_num);
853
854	if (s->pshift)
855	coef_par = ff_dirac_golomb_read_32bit(s->reader_ctx, addr,
856	length, tmp_buf, coef_num);
857	else
858	coef_par = ff_dirac_golomb_read_16bit(s->reader_ctx, addr,
859	length, tmp_buf, coef_num);
860
861	if (coef_num > coef_par) {
862	const int start_b = coef_par * (1 << (s->pshift + 1));
863	const int end_b = coef_num * (1 << (s->pshift + 1));
864	memset(&tmp_buf[start_b], 0, end_b - start_b);
865	}
866
867	for (level = 0; level < s->wavelet_depth; level++) {
868	const SliceCoeffs *c = &coeffs_num[level];
869	for (orientation = !!level; orientation < 4; orientation++) {
870	const SubBand *b1 = &s->plane[i].band[level][orientation];
871	uint8_t *buf = b1->ibuf + c->top * b1->stride + (c->left << (s->pshift + 1));
872
873	/* Change to c->tot_h <= 4 for AVX2 dequantization */
874	const int qfunc = s->pshift + 2*(c->tot_h <= 2);
875	s->diracdsp.dequant_subband[qfunc](&tmp_buf[off], buf, b1->stride,
876	qfactor[level][orientation],
877	qoffset[level][orientation],
878	c->tot_v, c->tot_h);
879
880	off += c->tot << (s->pshift + 1);
881	}
882	}
883
884	skip_bits_long(gb, bits_end - get_bits_count(gb));
885	}
886
887	return 0;
888	}
889
890	static int decode_hq_slice_row(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
891	{
892	int i;
893	DiracContext *s = avctx->priv_data;
894	DiracSlice *slices = ((DiracSlice *)arg) + s->num_x*jobnr;
895	uint8_t *thread_buf = &s->thread_buf[s->thread_buf_size*threadnr];
896	for (i = 0; i < s->num_x; i++)
897	decode_hq_slice(s, &slices[i], thread_buf);
898	return 0;
899	}
900
901	/**
902	* Dirac Specification ->
903	* 13.5.1 low_delay_transform_data()
904	*/
905	static int decode_lowdelay(DiracContext *s)
906	{
907	AVCodecContext *avctx = s->avctx;
908	int slice_x, slice_y, bufsize;
909	int64_t coef_buf_size, bytes = 0;
910	const uint8_t *buf;
911	DiracSlice *slices;
912	SliceCoeffs tmp[MAX_DWT_LEVELS];
913	int slice_num = 0;
914
915	if (s->slice_params_num_buf != (s->num_x * s->num_y)) {
916	s->slice_params_buf = av_realloc_f(s->slice_params_buf, s->num_x * s->num_y, sizeof(DiracSlice));
917	if (!s->slice_params_buf) {
918	av_log(s->avctx, AV_LOG_ERROR, "slice params buffer allocation failure\n");
919	s->slice_params_num_buf = 0;
920	return AVERROR(ENOMEM);
921	}
922	s->slice_params_num_buf = s->num_x * s->num_y;
923	}
924	slices = s->slice_params_buf;
925
926	/* 8 becacuse that's how much the golomb reader could overread junk data
927	* from another plane/slice at most, and 512 because SIMD */
928	coef_buf_size = subband_coeffs(s, s->num_x - 1, s->num_y - 1, 0, tmp) + 8;
929	coef_buf_size = (coef_buf_size << (1 + s->pshift)) + 512;
930
931	if (s->threads_num_buf != avctx->thread_count ||
932	s->thread_buf_size != coef_buf_size) {
933	s->threads_num_buf = avctx->thread_count;
934	s->thread_buf_size = coef_buf_size;
935	s->thread_buf = av_realloc_f(s->thread_buf, avctx->thread_count, s->thread_buf_size);
936	if (!s->thread_buf) {
937	av_log(s->avctx, AV_LOG_ERROR, "thread buffer allocation failure\n");
938	return AVERROR(ENOMEM);
939	}
940	}
941
942	align_get_bits(&s->gb);
943	/*[DIRAC_STD] 13.5.2 Slices. slice(sx,sy) */
944	buf = s->gb.buffer + get_bits_count(&s->gb)/8;
945	bufsize = get_bits_left(&s->gb);
946
947	if (s->hq_picture) {
948	int i;
949
950	for (slice_y = 0; bufsize > 0 && slice_y < s->num_y; slice_y++) {
951	for (slice_x = 0; bufsize > 0 && slice_x < s->num_x; slice_x++) {
952	bytes = s->highquality.prefix_bytes + 1;
953	for (i = 0; i < 3; i++) {
954	if (bytes <= bufsize/8)
955	bytes += buf[bytes] * s->highquality.size_scaler + 1;
956	}
957	if (bytes >= INT_MAX || bytes*8 > bufsize) {
958	av_log(s->avctx, AV_LOG_ERROR, "too many bytes\n");
959	return AVERROR_INVALIDDATA;
960	}
961
962	slices[slice_num].bytes = bytes;
963	slices[slice_num].slice_x = slice_x;
964	slices[slice_num].slice_y = slice_y;
965	init_get_bits(&slices[slice_num].gb, buf, bufsize);
966	slice_num++;
967
968	buf += bytes;
969	if (bufsize/8 >= bytes)
970	bufsize -= bytes*8;
971	else
972	bufsize = 0;
973	}
974	}
975
976	if (s->num_x*s->num_y != slice_num) {
977	av_log(s->avctx, AV_LOG_ERROR, "too few slices\n");
978	return AVERROR_INVALIDDATA;
979	}
980
981	avctx->execute2(avctx, decode_hq_slice_row, slices, NULL, s->num_y);
982	} else {
983	for (slice_y = 0; bufsize > 0 && slice_y < s->num_y; slice_y++) {
984	for (slice_x = 0; bufsize > 0 && slice_x < s->num_x; slice_x++) {
985	bytes = (slice_num+1) * (int64_t)s->lowdelay.bytes.num / s->lowdelay.bytes.den
986	- slice_num * (int64_t)s->lowdelay.bytes.num / s->lowdelay.bytes.den;
987	slices[slice_num].bytes = bytes;
988	slices[slice_num].slice_x = slice_x;
989	slices[slice_num].slice_y = slice_y;
990	init_get_bits(&slices[slice_num].gb, buf, bufsize);
991	slice_num++;
992
993	buf += bytes;
994	if (bufsize/8 >= bytes)
995	bufsize -= bytes*8;
996	else
997	bufsize = 0;
998	}
999	}
1000	avctx->execute(avctx, decode_lowdelay_slice, slices, NULL, slice_num,
1001	sizeof(DiracSlice)); /* [DIRAC_STD] 13.5.2 Slices */
1002	}
1003
1004	if (s->dc_prediction) {
1005	if (s->pshift) {
1006	intra_dc_prediction_10(&s->plane[0].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
1007	intra_dc_prediction_10(&s->plane[1].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
1008	intra_dc_prediction_10(&s->plane[2].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
1009	} else {
1010	intra_dc_prediction_8(&s->plane[0].band[0][0]);
1011	intra_dc_prediction_8(&s->plane[1].band[0][0]);
1012	intra_dc_prediction_8(&s->plane[2].band[0][0]);
1013	}
1014	}
1015
1016	return 0;
1017	}
1018
1019	static void init_planes(DiracContext *s)
1020	{
1021	int i, w, h, level, orientation;
1022
1023	for (i = 0; i < 3; i++) {
1024	Plane *p = &s->plane[i];
1025
1026	p->width = s->seq.width >> (i ? s->chroma_x_shift : 0);
1027	p->height = s->seq.height >> (i ? s->chroma_y_shift : 0);
1028	p->idwt.width = w = CALC_PADDING(p->width , s->wavelet_depth);
1029	p->idwt.height = h = CALC_PADDING(p->height, s->wavelet_depth);
1030	p->idwt.stride = FFALIGN(p->idwt.width, 8) << (1 + s->pshift);
1031
1032	for (level = s->wavelet_depth-1; level >= 0; level--) {
1033	w = w>>1;
1034	h = h>>1;
1035	for (orientation = !!level; orientation < 4; orientation++) {
1036	SubBand *b = &p->band[level][orientation];
1037
1038	b->pshift = s->pshift;
1039	b->ibuf = p->idwt.buf;
1040	b->level = level;
1041	b->stride = p->idwt.stride << (s->wavelet_depth - level);
1042	b->width = w;
1043	b->height = h;
1044	b->orientation = orientation;
1045
1046	if (orientation & 1)
1047	b->ibuf += w << (1+b->pshift);
1048	if (orientation > 1)
1049	b->ibuf += (b->stride>>1);
1050
1051	if (level)
1052	b->parent = &p->band[level-1][orientation];
1053	}
1054	}
1055
1056	if (i > 0) {
1057	p->xblen = s->plane[0].xblen >> s->chroma_x_shift;
1058	p->yblen = s->plane[0].yblen >> s->chroma_y_shift;
1059	p->xbsep = s->plane[0].xbsep >> s->chroma_x_shift;
1060	p->ybsep = s->plane[0].ybsep >> s->chroma_y_shift;
1061	}
1062
1063	p->xoffset = (p->xblen - p->xbsep)/2;
1064	p->yoffset = (p->yblen - p->ybsep)/2;
1065	}
1066	}
1067
1068	/**
1069	* Unpack the motion compensation parameters
1070	* Dirac Specification ->
1071	* 11.2 Picture prediction data. picture_prediction()
1072	*/
1073	static int dirac_unpack_prediction_parameters(DiracContext *s)
1074	{
1075	static const uint8_t default_blen[] = { 4, 12, 16, 24 };
1076
1077	GetBitContext *gb = &s->gb;
1078	unsigned idx, ref;
1079
1080	align_get_bits(gb);
1081	/* [DIRAC_STD] 11.2.2 Block parameters. block_parameters() */
1082	/* Luma and Chroma are equal. 11.2.3 */
1083	idx = get_interleaved_ue_golomb(gb); /* [DIRAC_STD] index */
1084
1085	if (idx > 4) {
1086	av_log(s->avctx, AV_LOG_ERROR, "Block prediction index too high\n");
1087	return AVERROR_INVALIDDATA;
1088	}
1089
1090	if (idx == 0) {
1091	s->plane[0].xblen = get_interleaved_ue_golomb(gb);
1092	s->plane[0].yblen = get_interleaved_ue_golomb(gb);
1093	s->plane[0].xbsep = get_interleaved_ue_golomb(gb);
1094	s->plane[0].ybsep = get_interleaved_ue_golomb(gb);
1095	} else {
1096	/*[DIRAC_STD] preset_block_params(index). Table 11.1 */
1097	s->plane[0].xblen = default_blen[idx-1];
1098	s->plane[0].yblen = default_blen[idx-1];
1099	s->plane[0].xbsep = 4 * idx;
1100	s->plane[0].ybsep = 4 * idx;
1101	}
1102	/*[DIRAC_STD] 11.2.4 motion_data_dimensions()
1103	Calculated in function dirac_unpack_block_motion_data */
1104
1105	if (s->plane[0].xblen % (1 << s->chroma_x_shift) != 0 ||
1106	s->plane[0].yblen % (1 << s->chroma_y_shift) != 0 ||
1107	!s->plane[0].xblen || !s->plane[0].yblen) {
1108	av_log(s->avctx, AV_LOG_ERROR,
1109	"invalid x/y block length (%d/%d) for x/y chroma shift (%d/%d)\n",
1110	s->plane[0].xblen, s->plane[0].yblen, s->chroma_x_shift, s->chroma_y_shift);
1111	return AVERROR_INVALIDDATA;
1112	}
1113	if (!s->plane[0].xbsep || !s->plane[0].ybsep || s->plane[0].xbsep < s->plane[0].xblen/2 || s->plane[0].ybsep < s->plane[0].yblen/2) {
1114	av_log(s->avctx, AV_LOG_ERROR, "Block separation too small\n");
1115	return AVERROR_INVALIDDATA;
1116	}
1117	if (s->plane[0].xbsep > s->plane[0].xblen || s->plane[0].ybsep > s->plane[0].yblen) {
1118	av_log(s->avctx, AV_LOG_ERROR, "Block separation greater than size\n");
1119	return AVERROR_INVALIDDATA;
1120	}
1121	if (FFMAX(s->plane[0].xblen, s->plane[0].yblen) > MAX_BLOCKSIZE) {
1122	av_log(s->avctx, AV_LOG_ERROR, "Unsupported large block size\n");
1123	return AVERROR_PATCHWELCOME;
1124	}
1125
1126	/*[DIRAC_STD] 11.2.5 Motion vector precision. motion_vector_precision()
1127	Read motion vector precision */
1128	s->mv_precision = get_interleaved_ue_golomb(gb);
1129	if (s->mv_precision > 3) {
1130	av_log(s->avctx, AV_LOG_ERROR, "MV precision finer than eighth-pel\n");
1131	return AVERROR_INVALIDDATA;
1132	}
1133
1134	/*[DIRAC_STD] 11.2.6 Global motion. global_motion()
1135	Read the global motion compensation parameters */
1136	s->globalmc_flag = get_bits1(gb);
1137	if (s->globalmc_flag) {
1138	memset(s->globalmc, 0, sizeof(s->globalmc));
1139	/* [DIRAC_STD] pan_tilt(gparams) */
1140	for (ref = 0; ref < s->num_refs; ref++) {
1141	if (get_bits1(gb)) {
1142	s->globalmc[ref].pan_tilt[0] = dirac_get_se_golomb(gb);
1143	s->globalmc[ref].pan_tilt[1] = dirac_get_se_golomb(gb);
1144	}
1145	/* [DIRAC_STD] zoom_rotate_shear(gparams)
1146	zoom/rotation/shear parameters */
1147	if (get_bits1(gb)) {
1148	s->globalmc[ref].zrs_exp = get_interleaved_ue_golomb(gb);
1149	s->globalmc[ref].zrs[0][0] = dirac_get_se_golomb(gb);
1150	s->globalmc[ref].zrs[0][1] = dirac_get_se_golomb(gb);
1151	s->globalmc[ref].zrs[1][0] = dirac_get_se_golomb(gb);
1152	s->globalmc[ref].zrs[1][1] = dirac_get_se_golomb(gb);
1153	} else {
1154	s->globalmc[ref].zrs[0][0] = 1;
1155	s->globalmc[ref].zrs[1][1] = 1;
1156	}
1157	/* [DIRAC_STD] perspective(gparams) */
1158	if (get_bits1(gb)) {
1159	s->globalmc[ref].perspective_exp = get_interleaved_ue_golomb(gb);
1160	s->globalmc[ref].perspective[0] = dirac_get_se_golomb(gb);
1161	s->globalmc[ref].perspective[1] = dirac_get_se_golomb(gb);
1162	}
1163	}
1164	}
1165
1166	/*[DIRAC_STD] 11.2.7 Picture prediction mode. prediction_mode()
1167	Picture prediction mode, not currently used. */
1168	if (get_interleaved_ue_golomb(gb)) {
1169	av_log(s->avctx, AV_LOG_ERROR, "Unknown picture prediction mode\n");
1170	return AVERROR_INVALIDDATA;
1171	}
1172
1173	/* [DIRAC_STD] 11.2.8 Reference picture weight. reference_picture_weights()
1174	just data read, weight calculation will be done later on. */
1175	s->weight_log2denom = 1;
1176	s->weight[0] = 1;
1177	s->weight[1] = 1;
1178
1179	if (get_bits1(gb)) {
1180	s->weight_log2denom = get_interleaved_ue_golomb(gb);
1181	s->weight[0] = dirac_get_se_golomb(gb);
1182	if (s->num_refs == 2)
1183	s->weight[1] = dirac_get_se_golomb(gb);
1184	}
1185	return 0;
1186	}
1187
1188	/**
1189	* Dirac Specification ->
1190	* 11.3 Wavelet transform data. wavelet_transform()
1191	*/
1192	static int dirac_unpack_idwt_params(DiracContext *s)
1193	{
1194	GetBitContext *gb = &s->gb;
1195	int i, level;
1196	unsigned tmp;
1197
1198	#define CHECKEDREAD(dst, cond, errmsg) \
1199	tmp = get_interleaved_ue_golomb(gb); \
1200	if (cond) { \
1201	av_log(s->avctx, AV_LOG_ERROR, errmsg); \
1202	return AVERROR_INVALIDDATA; \
1203	}\
1204	dst = tmp;
1205
1206	align_get_bits(gb);
1207
1208	s->zero_res = s->num_refs ? get_bits1(gb) : 0;
1209	if (s->zero_res)
1210	return 0;
1211
1212	/*[DIRAC_STD] 11.3.1 Transform parameters. transform_parameters() */
1213	CHECKEDREAD(s->wavelet_idx, tmp > 6, "wavelet_idx is too big\n")
1214
1215	CHECKEDREAD(s->wavelet_depth, tmp > MAX_DWT_LEVELS || tmp < 1, "invalid number of DWT decompositions\n")
1216
1217	if (!s->low_delay) {
1218	/* Codeblock parameters (core syntax only) */
1219	if (get_bits1(gb)) {
1220	for (i = 0; i <= s->wavelet_depth; i++) {
1221	CHECKEDREAD(s->codeblock[i].width , tmp < 1 || tmp > (s->avctx->width >>s->wavelet_depth-i), "codeblock width invalid\n")
1222	CHECKEDREAD(s->codeblock[i].height, tmp < 1 || tmp > (s->avctx->height>>s->wavelet_depth-i), "codeblock height invalid\n")
1223	}
1224
1225	CHECKEDREAD(s->codeblock_mode, tmp > 1, "unknown codeblock mode\n")
1226	}
1227	else {
1228	for (i = 0; i <= s->wavelet_depth; i++)
1229	s->codeblock[i].width = s->codeblock[i].height = 1;
1230	}
1231	}
1232	else {
1233	s->num_x = get_interleaved_ue_golomb(gb);
1234	s->num_y = get_interleaved_ue_golomb(gb);
1235	if (s->num_x * s->num_y == 0 || s->num_x * (uint64_t)s->num_y > INT_MAX) {
1236	av_log(s->avctx,AV_LOG_ERROR,"Invalid numx/y\n");
1237	s->num_x = s->num_y = 0;
1238	return AVERROR_INVALIDDATA;
1239	}
1240	if (s->ld_picture) {
1241	s->lowdelay.bytes.num = get_interleaved_ue_golomb(gb);
1242	s->lowdelay.bytes.den = get_interleaved_ue_golomb(gb);
1243	if (s->lowdelay.bytes.den <= 0) {
1244	av_log(s->avctx,AV_LOG_ERROR,"Invalid lowdelay.bytes.den\n");
1245	return AVERROR_INVALIDDATA;
1246	}
1247	} else if (s->hq_picture) {
1248	s->highquality.prefix_bytes = get_interleaved_ue_golomb(gb);
1249	s->highquality.size_scaler = get_interleaved_ue_golomb(gb);
1250	if (s->highquality.prefix_bytes >= INT_MAX / 8) {
1251	av_log(s->avctx,AV_LOG_ERROR,"too many prefix bytes\n");
1252	return AVERROR_INVALIDDATA;
1253	}
1254	}
1255
1256	/* [DIRAC_STD] 11.3.5 Quantisation matrices (low-delay syntax). quant_matrix() */
1257	if (get_bits1(gb)) {
1258	av_log(s->avctx,AV_LOG_DEBUG,"Low Delay: Has Custom Quantization Matrix!\n");
1259	/* custom quantization matrix */
1260	s->lowdelay.quant[0][0] = get_interleaved_ue_golomb(gb);
1261	for (level = 0; level < s->wavelet_depth; level++) {
1262	s->lowdelay.quant[level][1] = get_interleaved_ue_golomb(gb);
1263	s->lowdelay.quant[level][2] = get_interleaved_ue_golomb(gb);
1264	s->lowdelay.quant[level][3] = get_interleaved_ue_golomb(gb);
1265	}
1266	} else {
1267	if (s->wavelet_depth > 4) {
1268	av_log(s->avctx,AV_LOG_ERROR,"Mandatory custom low delay matrix missing for depth %d\n", s->wavelet_depth);
1269	return AVERROR_INVALIDDATA;
1270	}
1271	/* default quantization matrix */
1272	for (level = 0; level < s->wavelet_depth; level++)
1273	for (i = 0; i < 4; i++) {
1274	s->lowdelay.quant[level][i] = ff_dirac_default_qmat[s->wavelet_idx][level][i];
1275	/* haar with no shift differs for different depths */
1276	if (s->wavelet_idx == 3)
1277	s->lowdelay.quant[level][i] += 4*(s->wavelet_depth-1 - level);
1278	}
1279	}
1280	}
1281	return 0;
1282	}
1283
1284	static inline int pred_sbsplit(uint8_t *sbsplit, int stride, int x, int y)
1285	{
1286	static const uint8_t avgsplit[7] = { 0, 0, 1, 1, 1, 2, 2 };
1287
1288	if (!(x|y))
1289	return 0;
1290	else if (!y)
1291	return sbsplit[-1];
1292	else if (!x)
1293	return sbsplit[-stride];
1294
1295	return avgsplit[sbsplit[-1] + sbsplit[-stride] + sbsplit[-stride-1]];
1296	}
1297
1298	static inline int pred_block_mode(DiracBlock *block, int stride, int x, int y, int refmask)
1299	{
1300	int pred;
1301
1302	if (!(x|y))
1303	return 0;
1304	else if (!y)
1305	return block[-1].ref & refmask;
1306	else if (!x)
1307	return block[-stride].ref & refmask;
1308
1309	/* return the majority */
1310	pred = (block[-1].ref & refmask) + (block[-stride].ref & refmask) + (block[-stride-1].ref & refmask);
1311	return (pred >> 1) & refmask;
1312	}
1313
1314	static inline void pred_block_dc(DiracBlock *block, int stride, int x, int y)
1315	{
1316	int i, n = 0;
1317
1318	memset(block->u.dc, 0, sizeof(block->u.dc));
1319
1320	if (x && !(block[-1].ref & 3)) {
1321	for (i = 0; i < 3; i++)
1322	block->u.dc[i] += block[-1].u.dc[i];
1323	n++;
1324	}
1325
1326	if (y && !(block[-stride].ref & 3)) {
1327	for (i = 0; i < 3; i++)
1328	block->u.dc[i] += block[-stride].u.dc[i];
1329	n++;
1330	}
1331
1332	if (x && y && !(block[-1-stride].ref & 3)) {
1333	for (i = 0; i < 3; i++)
1334	block->u.dc[i] += block[-1-stride].u.dc[i];
1335	n++;
1336	}
1337
1338	if (n == 2) {
1339	for (i = 0; i < 3; i++)
1340	block->u.dc[i] = (block->u.dc[i]+1)>>1;
1341	} else if (n == 3) {
1342	for (i = 0; i < 3; i++)
1343	block->u.dc[i] = divide3(block->u.dc[i]);
1344	}
1345	}
1346
1347	static inline void pred_mv(DiracBlock *block, int stride, int x, int y, int ref)
1348	{
1349	int16_t *pred[3];
1350	int refmask = ref+1;
1351	int mask = refmask | DIRAC_REF_MASK_GLOBAL; /* exclude gmc blocks */
1352	int n = 0;
1353
1354	if (x && (block[-1].ref & mask) == refmask)
1355	pred[n++] = block[-1].u.mv[ref];
1356
1357	if (y && (block[-stride].ref & mask) == refmask)
1358	pred[n++] = block[-stride].u.mv[ref];
1359
1360	if (x && y && (block[-stride-1].ref & mask) == refmask)
1361	pred[n++] = block[-stride-1].u.mv[ref];
1362
1363	switch (n) {
1364	case 0:
1365	block->u.mv[ref][0] = 0;
1366	block->u.mv[ref][1] = 0;
1367	break;
1368	case 1:
1369	block->u.mv[ref][0] = pred[0][0];
1370	block->u.mv[ref][1] = pred[0][1];
1371	break;
1372	case 2:
1373	block->u.mv[ref][0] = (pred[0][0] + pred[1][0] + 1) >> 1;
1374	block->u.mv[ref][1] = (pred[0][1] + pred[1][1] + 1) >> 1;
1375	break;
1376	case 3:
1377	block->u.mv[ref][0] = mid_pred(pred[0][0], pred[1][0], pred[2][0]);
1378	block->u.mv[ref][1] = mid_pred(pred[0][1], pred[1][1], pred[2][1]);
1379	break;
1380	}
1381	}
1382
1383	static void global_mv(DiracContext *s, DiracBlock *block, int x, int y, int ref)
1384	{
1385	int ez = s->globalmc[ref].zrs_exp;
1386	int ep = s->globalmc[ref].perspective_exp;
1387	int (*A)[2] = s->globalmc[ref].zrs;
1388	int *b = s->globalmc[ref].pan_tilt;
1389	int *c = s->globalmc[ref].perspective;
1390
1391	int m = (1<<ep) - (c[0]*x + c[1]*y);
1392	int mx = m * ((A[0][0] * x + A[0][1]*y) + (1<<ez) * b[0]);
1393	int my = m * ((A[1][0] * x + A[1][1]*y) + (1<<ez) * b[1]);
1394
1395	block->u.mv[ref][0] = (mx + (1<<(ez+ep))) >> (ez+ep);
1396	block->u.mv[ref][1] = (my + (1<<(ez+ep))) >> (ez+ep);
1397	}
1398
1399	static void decode_block_params(DiracContext *s, DiracArith arith[8], DiracBlock *block,
1400	int stride, int x, int y)
1401	{
1402	int i;
1403
1404	block->ref = pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF1);
1405	block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF1);
1406
1407	if (s->num_refs == 2) {
1408	block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF2);
1409	block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF2) << 1;
1410	}
1411
1412	if (!block->ref) {
1413	pred_block_dc(block, stride, x, y);
1414	for (i = 0; i < 3; i++)
1415	block->u.dc[i] += dirac_get_arith_int(arith+1+i, CTX_DC_F1, CTX_DC_DATA);
1416	return;
1417	}
1418
1419	if (s->globalmc_flag) {
1420	block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_GLOBAL);
1421	block->ref ^= dirac_get_arith_bit(arith, CTX_GLOBAL_BLOCK) << 2;
1422	}
1423
1424	for (i = 0; i < s->num_refs; i++)
1425	if (block->ref & (i+1)) {
1426	if (block->ref & DIRAC_REF_MASK_GLOBAL) {
1427	global_mv(s, block, x, y, i);
1428	} else {
1429	pred_mv(block, stride, x, y, i);
1430	block->u.mv[i][0] += dirac_get_arith_int(arith + 4 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
1431	block->u.mv[i][1] += dirac_get_arith_int(arith + 5 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
1432	}
1433	}
1434	}
1435
1436	/**
1437	* Copies the current block to the other blocks covered by the current superblock split mode
1438	*/
1439	static void propagate_block_data(DiracBlock *block, int stride, int size)
1440	{
1441	int x, y;
1442	DiracBlock *dst = block;
1443
1444	for (x = 1; x < size; x++)
1445	dst[x] = *block;
1446
1447	for (y = 1; y < size; y++) {
1448	dst += stride;
1449	for (x = 0; x < size; x++)
1450	dst[x] = *block;
1451	}
1452	}
1453
1454	/**
1455	* Dirac Specification ->
1456	* 12. Block motion data syntax
1457	*/
1458	static int dirac_unpack_block_motion_data(DiracContext *s)
1459	{
1460	GetBitContext *gb = &s->gb;
1461	uint8_t *sbsplit = s->sbsplit;
1462	int i, x, y, q, p;
1463	DiracArith arith[8];
1464
1465	align_get_bits(gb);
1466
1467	/* [DIRAC_STD] 11.2.4 and 12.2.1 Number of blocks and superblocks */
1468	s->sbwidth = DIVRNDUP(s->seq.width, 4*s->plane[0].xbsep);
1469	s->sbheight = DIVRNDUP(s->seq.height, 4*s->plane[0].ybsep);
1470	s->blwidth = 4 * s->sbwidth;
1471	s->blheight = 4 * s->sbheight;
1472
1473	/* [DIRAC_STD] 12.3.1 Superblock splitting modes. superblock_split_modes()
1474	decode superblock split modes */
1475	ff_dirac_init_arith_decoder(arith, gb, get_interleaved_ue_golomb(gb)); /* get_interleaved_ue_golomb(gb) is the length */
1476	for (y = 0; y < s->sbheight; y++) {
1477	for (x = 0; x < s->sbwidth; x++) {
1478	unsigned int split = dirac_get_arith_uint(arith, CTX_SB_F1, CTX_SB_DATA);
1479	if (split > 2)
1480	return AVERROR_INVALIDDATA;
1481	sbsplit[x] = (split + pred_sbsplit(sbsplit+x, s->sbwidth, x, y)) % 3;
1482	}
1483	sbsplit += s->sbwidth;
1484	}
1485
1486	/* setup arith decoding */
1487	ff_dirac_init_arith_decoder(arith, gb, get_interleaved_ue_golomb(gb));
1488	for (i = 0; i < s->num_refs; i++) {
1489	ff_dirac_init_arith_decoder(arith + 4 + 2 * i, gb, get_interleaved_ue_golomb(gb));
1490	ff_dirac_init_arith_decoder(arith + 5 + 2 * i, gb, get_interleaved_ue_golomb(gb));
1491	}
1492	for (i = 0; i < 3; i++)
1493	ff_dirac_init_arith_decoder(arith+1+i, gb, get_interleaved_ue_golomb(gb));
1494
1495	for (y = 0; y < s->sbheight; y++)
1496	for (x = 0; x < s->sbwidth; x++) {
1497	int blkcnt = 1 << s->sbsplit[y * s->sbwidth + x];
1498	int step = 4 >> s->sbsplit[y * s->sbwidth + x];
1499
1500	for (q = 0; q < blkcnt; q++)
1501	for (p = 0; p < blkcnt; p++) {
1502	int bx = 4 * x + p*step;
1503	int by = 4 * y + q*step;
1504	DiracBlock *block = &s->blmotion[by*s->blwidth + bx];
1505	decode_block_params(s, arith, block, s->blwidth, bx, by);
1506	propagate_block_data(block, s->blwidth, step);
1507	}
1508	}
1509
1510	return 0;
1511	}
1512
1513	static int weight(int i, int blen, int offset)
1514	{
1515	#define ROLLOFF(i) offset == 1 ? ((i) ? 5 : 3) : \
1516	(1 + (6*(i) + offset - 1) / (2*offset - 1))
1517
1518	if (i < 2*offset)
1519	return ROLLOFF(i);
1520	else if (i > blen-1 - 2*offset)
1521	return ROLLOFF(blen-1 - i);
1522	return 8;
1523	}
1524
1525	static void init_obmc_weight_row(Plane *p, uint8_t *obmc_weight, int stride,
1526	int left, int right, int wy)
1527	{
1528	int x;
1529	for (x = 0; left && x < p->xblen >> 1; x++)
1530	obmc_weight[x] = wy*8;
1531	for (; x < p->xblen >> right; x++)
1532	obmc_weight[x] = wy*weight(x, p->xblen, p->xoffset);
1533	for (; x < p->xblen; x++)
1534	obmc_weight[x] = wy*8;
1535	for (; x < stride; x++)
1536	obmc_weight[x] = 0;
1537	}
1538
1539	static void init_obmc_weight(Plane *p, uint8_t *obmc_weight, int stride,
1540	int left, int right, int top, int bottom)
1541	{
1542	int y;
1543	for (y = 0; top && y < p->yblen >> 1; y++) {
1544	init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
1545	obmc_weight += stride;
1546	}
1547	for (; y < p->yblen >> bottom; y++) {
1548	int wy = weight(y, p->yblen, p->yoffset);
1549	init_obmc_weight_row(p, obmc_weight, stride, left, right, wy);
1550	obmc_weight += stride;
1551	}
1552	for (; y < p->yblen; y++) {
1553	init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
1554	obmc_weight += stride;
1555	}
1556	}
1557
1558	static void init_obmc_weights(DiracContext *s, Plane *p, int by)
1559	{
1560	int top = !by;
1561	int bottom = by == s->blheight-1;
1562
1563	/* don't bother re-initing for rows 2 to blheight-2, the weights don't change */
1564	if (top || bottom || by == 1) {
1565	init_obmc_weight(p, s->obmc_weight[0], MAX_BLOCKSIZE, 1, 0, top, bottom);
1566	init_obmc_weight(p, s->obmc_weight[1], MAX_BLOCKSIZE, 0, 0, top, bottom);
1567	init_obmc_weight(p, s->obmc_weight[2], MAX_BLOCKSIZE, 0, 1, top, bottom);
1568	}
1569	}
1570
1571	static const uint8_t epel_weights[4][4][4] = {
1572	{{ 16, 0, 0, 0 },
1573	{ 12, 4, 0, 0 },
1574	{ 8, 8, 0, 0 },
1575	{ 4, 12, 0, 0 }},
1576	{{ 12, 0, 4, 0 },
1577	{ 9, 3, 3, 1 },
1578	{ 6, 6, 2, 2 },
1579	{ 3, 9, 1, 3 }},
1580	{{ 8, 0, 8, 0 },
1581	{ 6, 2, 6, 2 },
1582	{ 4, 4, 4, 4 },
1583	{ 2, 6, 2, 6 }},
1584	{{ 4, 0, 12, 0 },
1585	{ 3, 1, 9, 3 },
1586	{ 2, 2, 6, 6 },
1587	{ 1, 3, 3, 9 }}
1588	};
1589
1590	/**
1591	* For block x,y, determine which of the hpel planes to do bilinear
1592	* interpolation from and set src[] to the location in each hpel plane
1593	* to MC from.
1594	*
1595	* @return the index of the put_dirac_pixels_tab function to use
1596	* 0 for 1 plane (fpel,hpel), 1 for 2 planes (qpel), 2 for 4 planes (qpel), and 3 for epel
1597	*/
1598	static int mc_subpel(DiracContext *s, DiracBlock *block, const uint8_t *src[5],
1599	int x, int y, int ref, int plane)
1600	{
1601	Plane *p = &s->plane[plane];
1602	uint8_t **ref_hpel = s->ref_pics[ref]->hpel[plane];
1603	int motion_x = block->u.mv[ref][0];
1604	int motion_y = block->u.mv[ref][1];
1605	int mx, my, i, epel, nplanes = 0;
1606
1607	if (plane) {
1608	motion_x >>= s->chroma_x_shift;
1609	motion_y >>= s->chroma_y_shift;
1610	}
1611
1612	mx = motion_x & ~(-1U << s->mv_precision);
1613	my = motion_y & ~(-1U << s->mv_precision);
1614	motion_x >>= s->mv_precision;
1615	motion_y >>= s->mv_precision;
1616	/* normalize subpel coordinates to epel */
1617	/* TODO: template this function? */
1618	mx <<= 3 - s->mv_precision;
1619	my <<= 3 - s->mv_precision;
1620
1621	x += motion_x;
1622	y += motion_y;
1623	epel = (mx|my)&1;
1624
1625	/* hpel position */
1626	if (!((mx|my)&3)) {
1627	nplanes = 1;
1628	src[0] = ref_hpel[(my>>1)+(mx>>2)] + y*p->stride + x;
1629	} else {
1630	/* qpel or epel */
1631	nplanes = 4;
1632	for (i = 0; i < 4; i++)
1633	src[i] = ref_hpel[i] + y*p->stride + x;
1634
1635	/* if we're interpolating in the right/bottom halves, adjust the planes as needed
1636	we increment x/y because the edge changes for half of the pixels */
1637	if (mx > 4) {
1638	src[0] += 1;
1639	src[2] += 1;
1640	x++;
1641	}
1642	if (my > 4) {
1643	src[0] += p->stride;
1644	src[1] += p->stride;
1645	y++;
1646	}
1647
1648	/* hpel planes are:
1649	[0]: F [1]: H
1650	[2]: V [3]: C */
1651	if (!epel) {
1652	/* check if we really only need 2 planes since either mx or my is
1653	a hpel position. (epel weights of 0 handle this there) */
1654	if (!(mx&3)) {
1655	/* mx == 0: average [0] and [2]
1656	mx == 4: average [1] and [3] */
1657	src[!mx] = src[2 + !!mx];
1658	nplanes = 2;
1659	} else if (!(my&3)) {
1660	src[0] = src[(my>>1) ];
1661	src[1] = src[(my>>1)+1];
1662	nplanes = 2;
1663	}
1664	} else {
1665	/* adjust the ordering if needed so the weights work */
1666	if (mx > 4) {
1667	FFSWAP(const uint8_t *, src[0], src[1]);
1668	FFSWAP(const uint8_t *, src[2], src[3]);
1669	}
1670	if (my > 4) {
1671	FFSWAP(const uint8_t *, src[0], src[2]);
1672	FFSWAP(const uint8_t *, src[1], src[3]);
1673	}
1674	src[4] = epel_weights[my&3][mx&3];
1675	}
1676	}
1677
1678	/* fixme: v/h _edge_pos */
1679	if (x + p->xblen > p->width +EDGE_WIDTH/2 ||
1680	y + p->yblen > p->height+EDGE_WIDTH/2 ||
1681	x < 0 || y < 0) {
1682	for (i = 0; i < nplanes; i++) {
1683	s->vdsp.emulated_edge_mc(s->edge_emu_buffer[i], src[i],
1684	p->stride, p->stride,
1685	p->xblen, p->yblen, x, y,
1686	p->width+EDGE_WIDTH/2, p->height+EDGE_WIDTH/2);
1687	src[i] = s->edge_emu_buffer[i];
1688	}
1689	}
1690	return (nplanes>>1) + epel;
1691	}
1692
1693	static void add_dc(uint16_t *dst, int dc, int stride,
1694	uint8_t *obmc_weight, int xblen, int yblen)
1695	{
1696	int x, y;
1697	dc += 128;
1698
1699	for (y = 0; y < yblen; y++) {
1700	for (x = 0; x < xblen; x += 2) {
1701	dst[x ] += dc * obmc_weight[x ];
1702	dst[x+1] += dc * obmc_weight[x+1];
1703	}
1704	dst += stride;
1705	obmc_weight += MAX_BLOCKSIZE;
1706	}
1707	}
1708
1709	static void block_mc(DiracContext *s, DiracBlock *block,
1710	uint16_t *mctmp, uint8_t *obmc_weight,
1711	int plane, int dstx, int dsty)
1712	{
1713	Plane *p = &s->plane[plane];
1714	const uint8_t *src[5];
1715	int idx;
1716
1717	switch (block->ref&3) {
1718	case 0: /* DC */
1719	add_dc(mctmp, block->u.dc[plane], p->stride, obmc_weight, p->xblen, p->yblen);
1720	return;
1721	case 1:
1722	case 2:
1723	idx = mc_subpel(s, block, src, dstx, dsty, (block->ref&3)-1, plane);
1724	s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1725	if (s->weight_func)
1726	s->weight_func(s->mcscratch, p->stride, s->weight_log2denom,
1727	s->weight[0] + s->weight[1], p->yblen);
1728	break;
1729	case 3:
1730	idx = mc_subpel(s, block, src, dstx, dsty, 0, plane);
1731	s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1732	idx = mc_subpel(s, block, src, dstx, dsty, 1, plane);
1733	if (s->biweight_func) {
1734	/* fixme: +32 is a quick hack */
1735	s->put_pixels_tab[idx](s->mcscratch + 32, src, p->stride, p->yblen);
1736	s->biweight_func(s->mcscratch, s->mcscratch+32, p->stride, s->weight_log2denom,
1737	s->weight[0], s->weight[1], p->yblen);
1738	} else
1739	s->avg_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1740	break;
1741	}
1742	s->add_obmc(mctmp, s->mcscratch, p->stride, obmc_weight, p->yblen);
1743	}
1744
1745	static void mc_row(DiracContext *s, DiracBlock *block, uint16_t *mctmp, int plane, int dsty)
1746	{
1747	Plane *p = &s->plane[plane];
1748	int x, dstx = p->xbsep - p->xoffset;
1749
1750	block_mc(s, block, mctmp, s->obmc_weight[0], plane, -p->xoffset, dsty);
1751	mctmp += p->xbsep;
1752
1753	for (x = 1; x < s->blwidth-1; x++) {
1754	block_mc(s, block+x, mctmp, s->obmc_weight[1], plane, dstx, dsty);
1755	dstx += p->xbsep;
1756	mctmp += p->xbsep;
1757	}
1758	block_mc(s, block+x, mctmp, s->obmc_weight[2], plane, dstx, dsty);
1759	}
1760
1761	static void select_dsp_funcs(DiracContext *s, int width, int height, int xblen, int yblen)
1762	{
1763	int idx = 0;
1764	if (xblen > 8)
1765	idx = 1;
1766	if (xblen > 16)
1767	idx = 2;
1768
1769	memcpy(s->put_pixels_tab, s->diracdsp.put_dirac_pixels_tab[idx], sizeof(s->put_pixels_tab));
1770	memcpy(s->avg_pixels_tab, s->diracdsp.avg_dirac_pixels_tab[idx], sizeof(s->avg_pixels_tab));
1771	s->add_obmc = s->diracdsp.add_dirac_obmc[idx];
1772	if (s->weight_log2denom > 1 || s->weight[0] != 1 || s->weight[1] != 1) {
1773	s->weight_func = s->diracdsp.weight_dirac_pixels_tab[idx];
1774	s->biweight_func = s->diracdsp.biweight_dirac_pixels_tab[idx];
1775	} else {
1776	s->weight_func = NULL;
1777	s->biweight_func = NULL;
1778	}
1779	}
1780
1781	static int interpolate_refplane(DiracContext *s, DiracFrame *ref, int plane, int width, int height)
1782	{
1783	/* chroma allocates an edge of 8 when subsampled
1784	which for 4:2:2 means an h edge of 16 and v edge of 8
1785	just use 8 for everything for the moment */
1786	int i, edge = EDGE_WIDTH/2;
1787
1788	ref->hpel[plane][0] = ref->avframe->data[plane];
1789	s->mpvencdsp.draw_edges(ref->hpel[plane][0], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM); /* EDGE_TOP | EDGE_BOTTOM values just copied to make it build, this needs to be ensured */
1790
1791	/* no need for hpel if we only have fpel vectors */
1792	if (!s->mv_precision)
1793	return 0;
1794
1795	for (i = 1; i < 4; i++) {
1796	if (!ref->hpel_base[plane][i])
1797	ref->hpel_base[plane][i] = av_malloc((height+2*edge) * ref->avframe->linesize[plane] + 32);
1798	if (!ref->hpel_base[plane][i]) {
1799	return AVERROR(ENOMEM);
1800	}
1801	/* we need to be 16-byte aligned even for chroma */
1802	ref->hpel[plane][i] = ref->hpel_base[plane][i] + edge*ref->avframe->linesize[plane] + 16;
1803	}
1804
1805	if (!ref->interpolated[plane]) {
1806	s->diracdsp.dirac_hpel_filter(ref->hpel[plane][1], ref->hpel[plane][2],
1807	ref->hpel[plane][3], ref->hpel[plane][0],
1808	ref->avframe->linesize[plane], width, height);
1809	s->mpvencdsp.draw_edges(ref->hpel[plane][1], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1810	s->mpvencdsp.draw_edges(ref->hpel[plane][2], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1811	s->mpvencdsp.draw_edges(ref->hpel[plane][3], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1812	}
1813	ref->interpolated[plane] = 1;
1814
1815	return 0;
1816	}
1817
1818	/**
1819	* Dirac Specification ->
1820	* 13.0 Transform data syntax. transform_data()
1821	*/
1822	static int dirac_decode_frame_internal(DiracContext *s)
1823	{
1824	DWTContext d;
1825	int y, i, comp, dsty;
1826	int ret;
1827
1828	if (s->low_delay) {
1829	/* [DIRAC_STD] 13.5.1 low_delay_transform_data() */
1830	if (!s->hq_picture) {
1831	for (comp = 0; comp < 3; comp++) {
1832	Plane *p = &s->plane[comp];
1833	memset(p->idwt.buf, 0, p->idwt.stride * p->idwt.height);
1834	}
1835	}
1836	if (!s->zero_res) {
1837	if ((ret = decode_lowdelay(s)) < 0)
1838	return ret;
1839	}
1840	}
1841
1842	for (comp = 0; comp < 3; comp++) {
1843	Plane *p = &s->plane[comp];
1844	uint8_t *frame = s->current_picture->avframe->data[comp];
1845
1846	/* FIXME: small resolutions */
1847	for (i = 0; i < 4; i++)
1848	s->edge_emu_buffer[i] = s->edge_emu_buffer_base + i*FFALIGN(p->width, 16);
1849
1850	if (!s->zero_res && !s->low_delay)
1851	{
1852	memset(p->idwt.buf, 0, p->idwt.stride * p->idwt.height);
1853	decode_component(s, comp); /* [DIRAC_STD] 13.4.1 core_transform_data() */
1854	}
1855	ret = ff_spatial_idwt_init(&d, &p->idwt, s->wavelet_idx+2,
1856	s->wavelet_depth, s->bit_depth);
1857	if (ret < 0)
1858	return ret;
1859
1860	if (!s->num_refs) { /* intra */
1861	for (y = 0; y < p->height; y += 16) {
1862	int idx = (s->bit_depth - 8) >> 1;
1863	ff_spatial_idwt_slice2(&d, y+16); /* decode */
1864	s->diracdsp.put_signed_rect_clamped[idx](frame + y*p->stride,
1865	p->stride,
1866	p->idwt.buf + y*p->idwt.stride,
1867	p->idwt.stride, p->width, 16);
1868	}
1869	} else { /* inter */
1870	int rowheight = p->ybsep*p->stride;
1871
1872	select_dsp_funcs(s, p->width, p->height, p->xblen, p->yblen);
1873
1874	for (i = 0; i < s->num_refs; i++) {
1875	int ret = interpolate_refplane(s, s->ref_pics[i], comp, p->width, p->height);
1876	if (ret < 0)
1877	return ret;
1878	}
1879
1880	memset(s->mctmp, 0, 4*p->yoffset*p->stride);
1881
1882	dsty = -p->yoffset;
1883	for (y = 0; y < s->blheight; y++) {
1884	int h = 0,
1885	start = FFMAX(dsty, 0);
1886	uint16_t *mctmp = s->mctmp + y*rowheight;
1887	DiracBlock *blocks = s->blmotion + y*s->blwidth;
1888
1889	init_obmc_weights(s, p, y);
1890
1891	if (y == s->blheight-1 || start+p->ybsep > p->height)
1892	h = p->height - start;
1893	else
1894	h = p->ybsep - (start - dsty);
1895	if (h < 0)
1896	break;
1897
1898	memset(mctmp+2*p->yoffset*p->stride, 0, 2*rowheight);
1899	mc_row(s, blocks, mctmp, comp, dsty);
1900
1901	mctmp += (start - dsty)*p->stride + p->xoffset;
1902	ff_spatial_idwt_slice2(&d, start + h); /* decode */
1903	/* NOTE: add_rect_clamped hasn't been templated hence the shifts.
1904	* idwt.stride is passed as pixels, not in bytes as in the rest of the decoder */
1905	s->diracdsp.add_rect_clamped(frame + start*p->stride, mctmp, p->stride,
1906	(int16_t*)(p->idwt.buf) + start*(p->idwt.stride >> 1), (p->idwt.stride >> 1), p->width, h);
1907
1908	dsty += p->ybsep;
1909	}
1910	}
1911	}
1912
1913
1914	return 0;
1915	}
1916
1917	static int get_buffer_with_edge(AVCodecContext *avctx, AVFrame *f, int flags)
1918	{
1919	int ret, i;
1920	int chroma_x_shift, chroma_y_shift;
1921	avcodec_get_chroma_sub_sample(avctx->pix_fmt, &chroma_x_shift, &chroma_y_shift);
1922
1923	f->width = avctx->width + 2 * EDGE_WIDTH;
1924	f->height = avctx->height + 2 * EDGE_WIDTH + 2;
1925	ret = ff_get_buffer(avctx, f, flags);
1926	if (ret < 0)
1927	return ret;
1928
1929	for (i = 0; f->data[i]; i++) {
1930	int offset = (EDGE_WIDTH >> (i && i<3 ? chroma_y_shift : 0)) *
1931	f->linesize[i] + 32;
1932	f->data[i] += offset;
1933	}
1934	f->width = avctx->width;
1935	f->height = avctx->height;
1936
1937	return 0;
1938	}
1939
1940	/**
1941	* Dirac Specification ->
1942	* 11.1.1 Picture Header. picture_header()
1943	*/
1944	static int dirac_decode_picture_header(DiracContext *s)
1945	{
1946	unsigned retire, picnum;
1947	int i, j, ret;
1948	int64_t refdist, refnum;
1949	GetBitContext *gb = &s->gb;
1950
1951	/* [DIRAC_STD] 11.1.1 Picture Header. picture_header() PICTURE_NUM */
1952	picnum = s->current_picture->avframe->display_picture_number = get_bits_long(gb, 32);
1953
1954
1955	av_log(s->avctx,AV_LOG_DEBUG,"PICTURE_NUM: %d\n",picnum);
1956
1957	/* if this is the first keyframe after a sequence header, start our
1958	reordering from here */
1959	if (s->frame_number < 0)
1960	s->frame_number = picnum;
1961
1962	s->ref_pics[0] = s->ref_pics[1] = NULL;
1963	for (i = 0; i < s->num_refs; i++) {
1964	refnum = (picnum + dirac_get_se_golomb(gb)) & 0xFFFFFFFF;
1965	refdist = INT64_MAX;
1966
1967	/* find the closest reference to the one we want */
1968	/* Jordi: this is needed if the referenced picture hasn't yet arrived */
1969	for (j = 0; j < MAX_REFERENCE_FRAMES && refdist; j++)
1970	if (s->ref_frames[j]
1971	&& FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum) < refdist) {
1972	s->ref_pics[i] = s->ref_frames[j];
1973	refdist = FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum);
1974	}
1975
1976	if (!s->ref_pics[i] || refdist)
1977	av_log(s->avctx, AV_LOG_DEBUG, "Reference not found\n");
1978
1979	/* if there were no references at all, allocate one */
1980	if (!s->ref_pics[i])
1981	for (j = 0; j < MAX_FRAMES; j++)
1982	if (!s->all_frames[j].avframe->data[0]) {
1983	s->ref_pics[i] = &s->all_frames[j];
1984	ret = get_buffer_with_edge(s->avctx, s->ref_pics[i]->avframe, AV_GET_BUFFER_FLAG_REF);
1985	if (ret < 0)
1986	return ret;
1987	break;
1988	}
1989
1990	if (!s->ref_pics[i]) {
1991	av_log(s->avctx, AV_LOG_ERROR, "Reference could not be allocated\n");
1992	return AVERROR_INVALIDDATA;
1993	}
1994
1995	}
1996
1997	/* retire the reference frames that are not used anymore */
1998	if (s->current_picture->reference) {
1999	retire = (picnum + dirac_get_se_golomb(gb)) & 0xFFFFFFFF;
2000	if (retire != picnum) {
2001	DiracFrame *retire_pic = remove_frame(s->ref_frames, retire);
2002
2003	if (retire_pic)
2004	retire_pic->reference &= DELAYED_PIC_REF;
2005	else
2006	av_log(s->avctx, AV_LOG_DEBUG, "Frame to retire not found\n");
2007	}
2008
2009	/* if reference array is full, remove the oldest as per the spec */
2010	while (add_frame(s->ref_frames, MAX_REFERENCE_FRAMES, s->current_picture)) {
2011	av_log(s->avctx, AV_LOG_ERROR, "Reference frame overflow\n");
2012	remove_frame(s->ref_frames, s->ref_frames[0]->avframe->display_picture_number)->reference &= DELAYED_PIC_REF;
2013	}
2014	}
2015
2016	if (s->num_refs) {
2017	ret = dirac_unpack_prediction_parameters(s); /* [DIRAC_STD] 11.2 Picture Prediction Data. picture_prediction() */
2018	if (ret < 0)
2019	return ret;
2020	ret = dirac_unpack_block_motion_data(s); /* [DIRAC_STD] 12. Block motion data syntax */
2021	if (ret < 0)
2022	return ret;
2023	}
2024	ret = dirac_unpack_idwt_params(s); /* [DIRAC_STD] 11.3 Wavelet transform data */
2025	if (ret < 0)
2026	return ret;
2027
2028	init_planes(s);
2029	return 0;
2030	}
2031
2032	static int get_delayed_pic(DiracContext *s, AVFrame *picture, int *got_frame)
2033	{
2034	DiracFrame *out = s->delay_frames[0];
2035	int i, out_idx = 0;
2036	int ret;
2037
2038	/* find frame with lowest picture number */
2039	for (i = 1; s->delay_frames[i]; i++)
2040	if (s->delay_frames[i]->avframe->display_picture_number < out->avframe->display_picture_number) {
2041	out = s->delay_frames[i];
2042	out_idx = i;
2043	}
2044
2045	for (i = out_idx; s->delay_frames[i]; i++)
2046	s->delay_frames[i] = s->delay_frames[i+1];
2047
2048	if (out) {
2049	out->reference ^= DELAYED_PIC_REF;
2050	*got_frame = 1;
2051	if((ret = av_frame_ref(picture, out->avframe)) < 0)
2052	return ret;
2053	}
2054
2055	return 0;
2056	}
2057
2058	/**
2059	* Dirac Specification ->
2060	* 9.6 Parse Info Header Syntax. parse_info()
2061	* 4 byte start code + byte parse code + 4 byte size + 4 byte previous size
2062	*/
2063	#define DATA_UNIT_HEADER_SIZE 13
2064
2065	/* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3
2066	inside the function parse_sequence() */
2067	static int dirac_decode_data_unit(AVCodecContext *avctx, const uint8_t *buf, int size)
2068	{
2069	DiracContext *s = avctx->priv_data;
2070	DiracFrame *pic = NULL;
2071	AVDiracSeqHeader *dsh;
2072	int ret, i;
2073	uint8_t parse_code;
2074	unsigned tmp;
2075
2076	if (size < DATA_UNIT_HEADER_SIZE)
2077	return AVERROR_INVALIDDATA;
2078
2079	parse_code = buf[4];
2080
2081	init_get_bits(&s->gb, &buf[13], 8*(size - DATA_UNIT_HEADER_SIZE));
2082
2083	if (parse_code == DIRAC_PCODE_SEQ_HEADER) {
2084	if (s->seen_sequence_header)
2085	return 0;
2086
2087	/* [DIRAC_STD] 10. Sequence header */
2088	ret = av_dirac_parse_sequence_header(&dsh, buf + DATA_UNIT_HEADER_SIZE, size - DATA_UNIT_HEADER_SIZE, avctx);
2089	if (ret < 0) {
2090	av_log(avctx, AV_LOG_ERROR, "error parsing sequence header");
2091	return ret;
2092	}
2093
2094	ret = ff_set_dimensions(avctx, dsh->width, dsh->height);
2095	if (ret < 0) {
2096	av_freep(&dsh);
2097	return ret;
2098	}
2099
2100	ff_set_sar(avctx, dsh->sample_aspect_ratio);
2101	avctx->pix_fmt = dsh->pix_fmt;
2102	avctx->color_range = dsh->color_range;
2103	avctx->color_trc = dsh->color_trc;
2104	avctx->color_primaries = dsh->color_primaries;
2105	avctx->colorspace = dsh->colorspace;
2106	avctx->profile = dsh->profile;
2107	avctx->level = dsh->level;
2108	avctx->framerate = dsh->framerate;
2109	s->bit_depth = dsh->bit_depth;
2110	s->version.major = dsh->version.major;
2111	s->version.minor = dsh->version.minor;
2112	s->seq = *dsh;
2113	av_freep(&dsh);
2114
2115	s->pshift = s->bit_depth > 8;
2116
2117	avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_x_shift, &s->chroma_y_shift);
2118
2119	ret = alloc_sequence_buffers(s);
2120	if (ret < 0)
2121	return ret;
2122
2123	s->seen_sequence_header = 1;
2124	} else if (parse_code == DIRAC_PCODE_END_SEQ) { /* [DIRAC_STD] End of Sequence */
2125	free_sequence_buffers(s);
2126	s->seen_sequence_header = 0;
2127	} else if (parse_code == DIRAC_PCODE_AUX) {
2128	if (buf[13] == 1) { /* encoder implementation/version */
2129	int ver[3];
2130	/* versions older than 1.0.8 don't store quant delta for
2131	subbands with only one codeblock */
2132	if (sscanf(buf+14, "Schroedinger %d.%d.%d", ver, ver+1, ver+2) == 3)
2133	if (ver[0] == 1 && ver[1] == 0 && ver[2] <= 7)
2134	s->old_delta_quant = 1;
2135	}
2136	} else if (parse_code & 0x8) { /* picture data unit */
2137	if (!s->seen_sequence_header) {
2138	av_log(avctx, AV_LOG_DEBUG, "Dropping frame without sequence header\n");
2139	return AVERROR_INVALIDDATA;
2140	}
2141
2142	/* find an unused frame */
2143	for (i = 0; i < MAX_FRAMES; i++)
2144	if (s->all_frames[i].avframe->data[0] == NULL)
2145	pic = &s->all_frames[i];
2146	if (!pic) {
2147	av_log(avctx, AV_LOG_ERROR, "framelist full\n");
2148	return AVERROR_INVALIDDATA;
2149	}
2150
2151	av_frame_unref(pic->avframe);
2152
2153	/* [DIRAC_STD] Defined in 9.6.1 ... */
2154	tmp = parse_code & 0x03; /* [DIRAC_STD] num_refs() */
2155	if (tmp > 2) {
2156	av_log(avctx, AV_LOG_ERROR, "num_refs of 3\n");
2157	return AVERROR_INVALIDDATA;
2158	}
2159	s->num_refs = tmp;
2160	s->is_arith = (parse_code & 0x48) == 0x08; /* [DIRAC_STD] using_ac() */
2161	s->low_delay = (parse_code & 0x88) == 0x88; /* [DIRAC_STD] is_low_delay() */
2162	s->core_syntax = (parse_code & 0x88) == 0x08; /* [DIRAC_STD] is_core_syntax() */
2163	s->ld_picture = (parse_code & 0xF8) == 0xC8; /* [DIRAC_STD] is_ld_picture() */
2164	s->hq_picture = (parse_code & 0xF8) == 0xE8; /* [DIRAC_STD] is_hq_picture() */
2165	s->dc_prediction = (parse_code & 0x28) == 0x08; /* [DIRAC_STD] using_dc_prediction() */
2166	pic->reference = (parse_code & 0x0C) == 0x0C; /* [DIRAC_STD] is_reference() */
2167	pic->avframe->key_frame = s->num_refs == 0; /* [DIRAC_STD] is_intra() */
2168	pic->avframe->pict_type = s->num_refs + 1; /* Definition of AVPictureType in avutil.h */
2169
2170	/* VC-2 Low Delay has a different parse code than the Dirac Low Delay */
2171	if (s->version.minor == 2 && parse_code == 0x88)
2172	s->ld_picture = 1;
2173
2174	if (s->low_delay && !(s->ld_picture || s->hq_picture) ) {
2175	av_log(avctx, AV_LOG_ERROR, "Invalid low delay flag\n");
2176	return AVERROR_INVALIDDATA;
2177	}
2178
2179	if ((ret = get_buffer_with_edge(avctx, pic->avframe, (parse_code & 0x0C) == 0x0C ? AV_GET_BUFFER_FLAG_REF : 0)) < 0)
2180	return ret;
2181	s->current_picture = pic;
2182	s->plane[0].stride = pic->avframe->linesize[0];
2183	s->plane[1].stride = pic->avframe->linesize[1];
2184	s->plane[2].stride = pic->avframe->linesize[2];
2185
2186	if (alloc_buffers(s, FFMAX3(FFABS(s->plane[0].stride), FFABS(s->plane[1].stride), FFABS(s->plane[2].stride))) < 0)
2187	return AVERROR(ENOMEM);
2188
2189	/* [DIRAC_STD] 11.1 Picture parse. picture_parse() */
2190	ret = dirac_decode_picture_header(s);
2191	if (ret < 0)
2192	return ret;
2193
2194	/* [DIRAC_STD] 13.0 Transform data syntax. transform_data() */
2195	ret = dirac_decode_frame_internal(s);
2196	if (ret < 0)
2197	return ret;
2198	}
2199	return 0;
2200	}
2201
2202	static int dirac_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *pkt)
2203	{
2204	DiracContext *s = avctx->priv_data;
2205	AVFrame *picture = data;
2206	uint8_t *buf = pkt->data;
2207	int buf_size = pkt->size;
2208	int i, buf_idx = 0;
2209	int ret;
2210	unsigned data_unit_size;
2211
2212	/* release unused frames */
2213	for (i = 0; i < MAX_FRAMES; i++)
2214	if (s->all_frames[i].avframe->data[0] && !s->all_frames[i].reference) {
2215	av_frame_unref(s->all_frames[i].avframe);
2216	memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));
2217	}
2218
2219	s->current_picture = NULL;
2220	*got_frame = 0;
2221
2222	/* end of stream, so flush delayed pics */
2223	if (buf_size == 0)
2224	return get_delayed_pic(s, (AVFrame *)data, got_frame);
2225
2226	for (;;) {
2227	/*[DIRAC_STD] Here starts the code from parse_info() defined in 9.6
2228	[DIRAC_STD] PARSE_INFO_PREFIX = "BBCD" as defined in ISO/IEC 646
2229	BBCD start code search */
2230	for (; buf_idx + DATA_UNIT_HEADER_SIZE < buf_size; buf_idx++) {
2231	if (buf[buf_idx ] == 'B' && buf[buf_idx+1] == 'B' &&
2232	buf[buf_idx+2] == 'C' && buf[buf_idx+3] == 'D')
2233	break;
2234	}
2235	/* BBCD found or end of data */
2236	if (buf_idx + DATA_UNIT_HEADER_SIZE >= buf_size)
2237	break;
2238
2239	data_unit_size = AV_RB32(buf+buf_idx+5);
2240	if (data_unit_size > buf_size - buf_idx || !data_unit_size) {
2241	if(data_unit_size > buf_size - buf_idx)
2242	av_log(s->avctx, AV_LOG_ERROR,
2243	"Data unit with size %d is larger than input buffer, discarding\n",
2244	data_unit_size);
2245	buf_idx += 4;
2246	continue;
2247	}
2248	/* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3 inside the function parse_sequence() */
2249	ret = dirac_decode_data_unit(avctx, buf+buf_idx, data_unit_size);
2250	if (ret < 0)
2251	{
2252	av_log(s->avctx, AV_LOG_ERROR,"Error in dirac_decode_data_unit\n");
2253	return ret;
2254	}
2255	buf_idx += data_unit_size;
2256	}
2257
2258	if (!s->current_picture)
2259	return buf_size;
2260
2261	if (s->current_picture->avframe->display_picture_number > s->frame_number) {
2262	DiracFrame *delayed_frame = remove_frame(s->delay_frames, s->frame_number);
2263
2264	s->current_picture->reference |= DELAYED_PIC_REF;
2265
2266	if (add_frame(s->delay_frames, MAX_DELAY, s->current_picture)) {
2267	int min_num = s->delay_frames[0]->avframe->display_picture_number;
2268	/* Too many delayed frames, so we display the frame with the lowest pts */
2269	av_log(avctx, AV_LOG_ERROR, "Delay frame overflow\n");
2270
2271	for (i = 1; s->delay_frames[i]; i++)
2272	if (s->delay_frames[i]->avframe->display_picture_number < min_num)
2273	min_num = s->delay_frames[i]->avframe->display_picture_number;
2274
2275	delayed_frame = remove_frame(s->delay_frames, min_num);
2276	add_frame(s->delay_frames, MAX_DELAY, s->current_picture);
2277	}
2278
2279	if (delayed_frame) {
2280	delayed_frame->reference ^= DELAYED_PIC_REF;
2281	if((ret=av_frame_ref(data, delayed_frame->avframe)) < 0)
2282	return ret;
2283	*got_frame = 1;
2284	}
2285	} else if (s->current_picture->avframe->display_picture_number == s->frame_number) {
2286	/* The right frame at the right time :-) */
2287	if((ret=av_frame_ref(data, s->current_picture->avframe)) < 0)
2288	return ret;
2289	*got_frame = 1;
2290	}
2291
2292	if (*got_frame)
2293	s->frame_number = picture->display_picture_number + 1;
2294
2295	return buf_idx;
2296	}
2297
2298	AVCodec ff_dirac_decoder = {
2299	.name = "dirac",
2300	.long_name = NULL_IF_CONFIG_SMALL("BBC Dirac VC-2"),
2301	.type = AVMEDIA_TYPE_VIDEO,
2302	.id = AV_CODEC_ID_DIRAC,
2303	.priv_data_size = sizeof(DiracContext),
2304	.init = dirac_decode_init,
2305	.close = dirac_decode_end,
2306	.decode = dirac_decode_frame,
2307	.capabilities = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_DR1,
2308	.caps_internal = FF_CODEC_CAP_INIT_THREADSAFE,
2309	.flush = dirac_decode_flush,
2310	};
2311