2 * Copyright (C) 2007 Marco Gerards <marco@gnu.org>
3 * Copyright (C) 2009 David Conrad
4 * Copyright (C) 2011 Jordi Ortiz
6 * This file is part of FFmpeg.
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.
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.
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
26 * @author Marco Gerards <marco@gnu.org>, David Conrad, Jordi Ortiz <nenjordi@gmail.com>
29 #include "libavutil/pixdesc.h"
30 #include "libavutil/thread.h"
33 #include "bytestream.h"
36 #include "dirac_arith.h"
37 #include "dirac_vlc.h"
38 #include "mpeg12data.h"
39 #include "libavcodec/mpegvideo.h"
40 #include "mpegvideoencdsp.h"
41 #include "dirac_dwt.h"
48 * The spec limits this to 3 for frame coding, but in practice can be as high as 6
50 #define MAX_REFERENCE_FRAMES 8
51 #define MAX_DELAY 5 /* limit for main profile for frame coding (TODO: field coding) */
52 #define MAX_FRAMES (MAX_REFERENCE_FRAMES + MAX_DELAY + 1)
53 #define MAX_QUANT 255 /* max quant for VC-2 */
54 #define MAX_BLOCKSIZE 32 /* maximum xblen/yblen we support */
57 * DiracBlock->ref flags, if set then the block does MC from the given ref
59 #define DIRAC_REF_MASK_REF1 1
60 #define DIRAC_REF_MASK_REF2 2
61 #define DIRAC_REF_MASK_GLOBAL 4
64 * Value of Picture.reference when Picture is not a reference picture, but
65 * is held for delayed output.
67 #define DELAYED_PIC_REF 4
69 #define CALC_PADDING(size, depth) \
70 (((size + (1 << depth) - 1) >> depth) << depth)
72 #define DIVRNDUP(a, b) (((a) + (b) - 1) / (b))
76 int interpolated[3]; /* 1 if hpel[] is valid */
78 uint8_t *hpel_base[3][4];
86 } u; /* anonymous unions aren't in C99 :( */
90 typedef struct SubBand {
93 int stride; /* in bytes */
99 struct SubBand *parent;
103 const uint8_t *coeff_data;
106 typedef struct Plane {
116 /* block separation (block n+1 starts after this many pixels in block n) */
119 /* amount of overspill on each edge (half of the overlap between blocks) */
123 SubBand band[MAX_DWT_LEVELS][4];
126 /* Used by Low Delay and High Quality profiles */
127 typedef struct DiracSlice {
134 typedef struct DiracContext {
135 AVCodecContext *avctx;
136 MpegvideoEncDSPContext mpvencdsp;
137 VideoDSPContext vdsp;
138 DiracDSPContext diracdsp;
139 DiracGolombLUT *reader_ctx;
140 DiracVersionInfo version;
142 AVDiracSeqHeader seq;
143 int seen_sequence_header;
144 int64_t frame_number; /* number of the next frame to display */
149 int bit_depth; /* bit depth */
150 int pshift; /* pixel shift = bit_depth > 8 */
152 int zero_res; /* zero residue flag */
153 int is_arith; /* whether coeffs use arith or golomb coding */
154 int core_syntax; /* use core syntax only */
155 int low_delay; /* use the low delay syntax */
156 int hq_picture; /* high quality picture, enables low_delay */
157 int ld_picture; /* use low delay picture, turns on low_delay */
158 int dc_prediction; /* has dc prediction */
159 int globalmc_flag; /* use global motion compensation */
160 int num_refs; /* number of reference pictures */
162 /* wavelet decoding */
163 unsigned wavelet_depth; /* depth of the IDWT */
164 unsigned wavelet_idx;
167 * schroedinger older than 1.0.8 doesn't store
168 * quant delta if only one codebook exists in a band
170 unsigned old_delta_quant;
171 unsigned codeblock_mode;
173 unsigned num_x; /* number of horizontal slices */
174 unsigned num_y; /* number of vertical slices */
176 uint8_t *thread_buf; /* Per-thread buffer for coefficient storage */
177 int threads_num_buf; /* Current # of buffers allocated */
178 int thread_buf_size; /* Each thread has a buffer this size */
180 DiracSlice *slice_params_buf;
181 int slice_params_num_buf;
186 } codeblock[MAX_DWT_LEVELS+1];
189 AVRational bytes; /* average bytes per slice */
190 uint8_t quant[MAX_DWT_LEVELS][4]; /* [DIRAC_STD] E.1 */
194 unsigned prefix_bytes;
195 uint64_t size_scaler;
199 int pan_tilt[2]; /* pan/tilt vector */
200 int zrs[2][2]; /* zoom/rotate/shear matrix */
201 int perspective[2]; /* perspective vector */
203 unsigned perspective_exp;
206 /* motion compensation */
207 uint8_t mv_precision; /* [DIRAC_STD] REFS_WT_PRECISION */
208 int16_t weight[2]; /* [DIRAC_STD] REF1_WT and REF2_WT */
209 unsigned weight_log2denom; /* [DIRAC_STD] REFS_WT_PRECISION */
211 int blwidth; /* number of blocks (horizontally) */
212 int blheight; /* number of blocks (vertically) */
213 int sbwidth; /* number of superblocks (horizontally) */
214 int sbheight; /* number of superblocks (vertically) */
217 DiracBlock *blmotion;
219 uint8_t *edge_emu_buffer[4];
220 uint8_t *edge_emu_buffer_base;
222 uint16_t *mctmp; /* buffer holding the MC data multiplied by OBMC weights */
226 DECLARE_ALIGNED(16, uint8_t, obmc_weight)[3][MAX_BLOCKSIZE*MAX_BLOCKSIZE];
228 void (*put_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
229 void (*avg_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
230 void (*add_obmc)(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
231 dirac_weight_func weight_func;
232 dirac_biweight_func biweight_func;
234 DiracFrame *current_picture;
235 DiracFrame *ref_pics[2];
237 DiracFrame *ref_frames[MAX_REFERENCE_FRAMES+1];
238 DiracFrame *delay_frames[MAX_DELAY+1];
239 DiracFrame all_frames[MAX_FRAMES];
250 /* magic number division by 3 from schroedinger */
251 static inline int divide3(int x)
253 return (int)((x+1U)*21845 + 10922) >> 16;
256 static DiracFrame *remove_frame(DiracFrame *framelist[], int picnum)
258 DiracFrame *remove_pic = NULL;
259 int i, remove_idx = -1;
261 for (i = 0; framelist[i]; i++)
262 if (framelist[i]->avframe->display_picture_number == picnum) {
263 remove_pic = framelist[i];
268 for (i = remove_idx; framelist[i]; i++)
269 framelist[i] = framelist[i+1];
274 static int add_frame(DiracFrame *framelist[], int maxframes, DiracFrame *frame)
277 for (i = 0; i < maxframes; i++)
279 framelist[i] = frame;
285 static int alloc_sequence_buffers(DiracContext *s)
287 int sbwidth = DIVRNDUP(s->seq.width, 4);
288 int sbheight = DIVRNDUP(s->seq.height, 4);
289 int i, w, h, top_padding;
291 /* todo: think more about this / use or set Plane here */
292 for (i = 0; i < 3; i++) {
293 int max_xblen = MAX_BLOCKSIZE >> (i ? s->chroma_x_shift : 0);
294 int max_yblen = MAX_BLOCKSIZE >> (i ? s->chroma_y_shift : 0);
295 w = s->seq.width >> (i ? s->chroma_x_shift : 0);
296 h = s->seq.height >> (i ? s->chroma_y_shift : 0);
298 /* we allocate the max we support here since num decompositions can
299 * change from frame to frame. Stride is aligned to 16 for SIMD, and
300 * 1<<MAX_DWT_LEVELS top padding to avoid if(y>0) in arith decoding
301 * MAX_BLOCKSIZE padding for MC: blocks can spill up to half of that
303 top_padding = FFMAX(1<<MAX_DWT_LEVELS, max_yblen/2);
304 w = FFALIGN(CALC_PADDING(w, MAX_DWT_LEVELS), 8); /* FIXME: Should this be 16 for SSE??? */
305 h = top_padding + CALC_PADDING(h, MAX_DWT_LEVELS) + max_yblen/2;
307 s->plane[i].idwt.buf_base = av_mallocz_array((w+max_xblen), h * (2 << s->pshift));
308 s->plane[i].idwt.tmp = av_malloc_array((w+16), 2 << s->pshift);
309 s->plane[i].idwt.buf = s->plane[i].idwt.buf_base + (top_padding*w)*(2 << s->pshift);
310 if (!s->plane[i].idwt.buf_base || !s->plane[i].idwt.tmp)
311 return AVERROR(ENOMEM);
314 /* fixme: allocate using real stride here */
315 s->sbsplit = av_malloc_array(sbwidth, sbheight);
316 s->blmotion = av_malloc_array(sbwidth, sbheight * 16 * sizeof(*s->blmotion));
318 if (!s->sbsplit || !s->blmotion)
319 return AVERROR(ENOMEM);
323 static int alloc_buffers(DiracContext *s, int stride)
325 int w = s->seq.width;
326 int h = s->seq.height;
328 av_assert0(stride >= w);
331 if (s->buffer_stride >= stride)
333 s->buffer_stride = 0;
335 av_freep(&s->edge_emu_buffer_base);
336 memset(s->edge_emu_buffer, 0, sizeof(s->edge_emu_buffer));
338 av_freep(&s->mcscratch);
340 s->edge_emu_buffer_base = av_malloc_array(stride, MAX_BLOCKSIZE);
342 s->mctmp = av_malloc_array((stride+MAX_BLOCKSIZE), (h+MAX_BLOCKSIZE) * sizeof(*s->mctmp));
343 s->mcscratch = av_malloc_array(stride, MAX_BLOCKSIZE);
345 if (!s->edge_emu_buffer_base || !s->mctmp || !s->mcscratch)
346 return AVERROR(ENOMEM);
348 s->buffer_stride = stride;
352 static void free_sequence_buffers(DiracContext *s)
356 for (i = 0; i < MAX_FRAMES; i++) {
357 if (s->all_frames[i].avframe->data[0]) {
358 av_frame_unref(s->all_frames[i].avframe);
359 memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));
362 for (j = 0; j < 3; j++)
363 for (k = 1; k < 4; k++)
364 av_freep(&s->all_frames[i].hpel_base[j][k]);
367 memset(s->ref_frames, 0, sizeof(s->ref_frames));
368 memset(s->delay_frames, 0, sizeof(s->delay_frames));
370 for (i = 0; i < 3; i++) {
371 av_freep(&s->plane[i].idwt.buf_base);
372 av_freep(&s->plane[i].idwt.tmp);
375 s->buffer_stride = 0;
376 av_freep(&s->sbsplit);
377 av_freep(&s->blmotion);
378 av_freep(&s->edge_emu_buffer_base);
381 av_freep(&s->mcscratch);
384 static AVOnce dirac_arith_init = AV_ONCE_INIT;
386 static av_cold int dirac_decode_init(AVCodecContext *avctx)
388 DiracContext *s = avctx->priv_data;
392 s->frame_number = -1;
394 s->thread_buf = NULL;
395 s->threads_num_buf = -1;
396 s->thread_buf_size = -1;
398 ff_dirac_golomb_reader_init(&s->reader_ctx);
399 ff_diracdsp_init(&s->diracdsp);
400 ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
401 ff_videodsp_init(&s->vdsp, 8);
403 for (i = 0; i < MAX_FRAMES; i++) {
404 s->all_frames[i].avframe = av_frame_alloc();
405 if (!s->all_frames[i].avframe) {
407 av_frame_free(&s->all_frames[--i].avframe);
408 return AVERROR(ENOMEM);
411 ret = ff_thread_once(&dirac_arith_init, ff_dirac_init_arith_tables);
413 return AVERROR_UNKNOWN;
418 static void dirac_decode_flush(AVCodecContext *avctx)
420 DiracContext *s = avctx->priv_data;
421 free_sequence_buffers(s);
422 s->seen_sequence_header = 0;
423 s->frame_number = -1;
426 static av_cold int dirac_decode_end(AVCodecContext *avctx)
428 DiracContext *s = avctx->priv_data;
431 ff_dirac_golomb_reader_end(&s->reader_ctx);
433 dirac_decode_flush(avctx);
434 for (i = 0; i < MAX_FRAMES; i++)
435 av_frame_free(&s->all_frames[i].avframe);
437 av_freep(&s->thread_buf);
438 av_freep(&s->slice_params_buf);
443 static inline int coeff_unpack_golomb(GetBitContext *gb, int qfactor, int qoffset)
445 int coeff = dirac_get_se_golomb(gb);
446 const unsigned sign = FFSIGN(coeff);
448 coeff = sign*((sign * coeff * qfactor + qoffset) >> 2);
452 #define SIGN_CTX(x) (CTX_SIGN_ZERO + ((x) > 0) - ((x) < 0))
454 #define UNPACK_ARITH(n, type) \
455 static inline void coeff_unpack_arith_##n(DiracArith *c, int qfactor, int qoffset, \
456 SubBand *b, type *buf, int x, int y) \
458 int sign, sign_pred = 0, pred_ctx = CTX_ZPZN_F1; \
460 const int mstride = -(b->stride >> (1+b->pshift)); \
462 const type *pbuf = (type *)b->parent->ibuf; \
463 const int stride = b->parent->stride >> (1+b->parent->pshift); \
464 pred_ctx += !!pbuf[stride * (y>>1) + (x>>1)] << 1; \
466 if (b->orientation == subband_hl) \
467 sign_pred = buf[mstride]; \
469 pred_ctx += !(buf[-1] | buf[mstride] | buf[-1 + mstride]); \
470 if (b->orientation == subband_lh) \
471 sign_pred = buf[-1]; \
473 pred_ctx += !buf[mstride]; \
475 coeff = dirac_get_arith_uint(c, pred_ctx, CTX_COEFF_DATA); \
477 coeff = (coeff * qfactor + qoffset) >> 2; \
478 sign = dirac_get_arith_bit(c, SIGN_CTX(sign_pred)); \
479 coeff = (coeff ^ -sign) + sign; \
484 UNPACK_ARITH(8, int16_t)
485 UNPACK_ARITH(10, int32_t)
488 * Decode the coeffs in the rectangle defined by left, right, top, bottom
489 * [DIRAC_STD] 13.4.3.2 Codeblock unpacking loop. codeblock()
491 static inline int codeblock(DiracContext *s, SubBand *b,
492 GetBitContext *gb, DiracArith *c,
493 int left, int right, int top, int bottom,
494 int blockcnt_one, int is_arith)
496 int x, y, zero_block;
497 int qoffset, qfactor;
500 /* check for any coded coefficients in this codeblock */
503 zero_block = dirac_get_arith_bit(c, CTX_ZERO_BLOCK);
505 zero_block = get_bits1(gb);
511 if (s->codeblock_mode && !(s->old_delta_quant && blockcnt_one)) {
514 quant = dirac_get_arith_int(c, CTX_DELTA_Q_F, CTX_DELTA_Q_DATA);
516 quant = dirac_get_se_golomb(gb);
517 if (quant > INT_MAX - b->quant || b->quant + quant < 0) {
518 av_log(s->avctx, AV_LOG_ERROR, "Invalid quant\n");
519 return AVERROR_INVALIDDATA;
524 if (b->quant > (DIRAC_MAX_QUANT_INDEX - 1)) {
525 av_log(s->avctx, AV_LOG_ERROR, "Unsupported quant %d\n", b->quant);
527 return AVERROR_INVALIDDATA;
530 qfactor = ff_dirac_qscale_tab[b->quant];
531 /* TODO: context pointer? */
533 qoffset = ff_dirac_qoffset_intra_tab[b->quant] + 2;
535 qoffset = ff_dirac_qoffset_inter_tab[b->quant] + 2;
537 buf = b->ibuf + top * b->stride;
539 for (y = top; y < bottom; y++) {
542 for (x = left; x < right; x++) {
544 coeff_unpack_arith_10(c, qfactor, qoffset, b, (int32_t*)(buf)+x, x, y);
546 coeff_unpack_arith_8(c, qfactor, qoffset, b, (int16_t*)(buf)+x, x, y);
552 for (y = top; y < bottom; y++) {
553 if (get_bits_left(gb) < 1)
554 return AVERROR_INVALIDDATA;
555 for (x = left; x < right; x++) {
556 int val = coeff_unpack_golomb(gb, qfactor, qoffset);
558 AV_WN32(&buf[4*x], val);
560 AV_WN16(&buf[2*x], val);
570 * Dirac Specification ->
571 * 13.3 intra_dc_prediction(band)
573 #define INTRA_DC_PRED(n, type) \
574 static inline void intra_dc_prediction_##n(SubBand *b) \
576 type *buf = (type*)b->ibuf; \
579 for (x = 1; x < b->width; x++) \
580 buf[x] += buf[x-1]; \
581 buf += (b->stride >> (1+b->pshift)); \
583 for (y = 1; y < b->height; y++) { \
584 buf[0] += buf[-(b->stride >> (1+b->pshift))]; \
586 for (x = 1; x < b->width; x++) { \
587 int pred = buf[x - 1] + buf[x - (b->stride >> (1+b->pshift))] + buf[x - (b->stride >> (1+b->pshift))-1]; \
588 buf[x] += divide3(pred); \
590 buf += (b->stride >> (1+b->pshift)); \
594 INTRA_DC_PRED(8, int16_t)
595 INTRA_DC_PRED(10, uint32_t)
598 * Dirac Specification ->
599 * 13.4.2 Non-skipped subbands. subband_coeffs()
601 static av_always_inline int decode_subband_internal(DiracContext *s, SubBand *b, int is_arith)
603 int cb_x, cb_y, left, right, top, bottom;
606 int cb_width = s->codeblock[b->level + (b->orientation != subband_ll)].width;
607 int cb_height = s->codeblock[b->level + (b->orientation != subband_ll)].height;
608 int blockcnt_one = (cb_width + cb_height) == 2;
614 init_get_bits8(&gb, b->coeff_data, b->length);
617 ff_dirac_init_arith_decoder(&c, &gb, b->length);
620 for (cb_y = 0; cb_y < cb_height; cb_y++) {
621 bottom = (b->height * (cb_y+1LL)) / cb_height;
623 for (cb_x = 0; cb_x < cb_width; cb_x++) {
624 right = (b->width * (cb_x+1LL)) / cb_width;
625 ret = codeblock(s, b, &gb, &c, left, right, top, bottom, blockcnt_one, is_arith);
633 if (b->orientation == subband_ll && s->num_refs == 0) {
635 intra_dc_prediction_10(b);
637 intra_dc_prediction_8(b);
643 static int decode_subband_arith(AVCodecContext *avctx, void *b)
645 DiracContext *s = avctx->priv_data;
646 return decode_subband_internal(s, b, 1);
649 static int decode_subband_golomb(AVCodecContext *avctx, void *arg)
651 DiracContext *s = avctx->priv_data;
653 return decode_subband_internal(s, *b, 0);
657 * Dirac Specification ->
658 * [DIRAC_STD] 13.4.1 core_transform_data()
660 static int decode_component(DiracContext *s, int comp)
662 AVCodecContext *avctx = s->avctx;
663 SubBand *bands[3*MAX_DWT_LEVELS+1];
664 enum dirac_subband orientation;
665 int level, num_bands = 0;
666 int ret[3*MAX_DWT_LEVELS+1];
668 int damaged_count = 0;
670 /* Unpack all subbands at all levels. */
671 for (level = 0; level < s->wavelet_depth; level++) {
672 for (orientation = !!level; orientation < 4; orientation++) {
673 SubBand *b = &s->plane[comp].band[level][orientation];
674 bands[num_bands++] = b;
676 align_get_bits(&s->gb);
677 /* [DIRAC_STD] 13.4.2 subband() */
678 b->length = get_interleaved_ue_golomb(&s->gb);
680 b->quant = get_interleaved_ue_golomb(&s->gb);
681 if (b->quant > (DIRAC_MAX_QUANT_INDEX - 1)) {
682 av_log(s->avctx, AV_LOG_ERROR, "Unsupported quant %d\n", b->quant);
684 return AVERROR_INVALIDDATA;
686 align_get_bits(&s->gb);
687 b->coeff_data = s->gb.buffer + get_bits_count(&s->gb)/8;
688 b->length = FFMIN(b->length, FFMAX(get_bits_left(&s->gb)/8, 0));
689 skip_bits_long(&s->gb, b->length*8);
692 /* arithmetic coding has inter-level dependencies, so we can only execute one level at a time */
694 avctx->execute(avctx, decode_subband_arith, &s->plane[comp].band[level][!!level],
695 ret + 3*level + !!level, 4-!!level, sizeof(SubBand));
697 /* golomb coding has no inter-level dependencies, so we can execute all subbands in parallel */
699 avctx->execute(avctx, decode_subband_golomb, bands, ret, num_bands, sizeof(SubBand*));
701 for (i = 0; i < s->wavelet_depth * 3 + 1; i++) {
705 if (damaged_count > (s->wavelet_depth * 3 + 1) /2)
706 return AVERROR_INVALIDDATA;
711 #define PARSE_VALUES(type, x, gb, ebits, buf1, buf2) \
712 type *buf = (type *)buf1; \
713 buf[x] = coeff_unpack_golomb(gb, qfactor, qoffset); \
714 if (get_bits_count(gb) >= ebits) \
717 buf = (type *)buf2; \
718 buf[x] = coeff_unpack_golomb(gb, qfactor, qoffset); \
719 if (get_bits_count(gb) >= ebits) \
723 static void decode_subband(DiracContext *s, GetBitContext *gb, int quant,
724 int slice_x, int slice_y, int bits_end,
725 SubBand *b1, SubBand *b2)
727 int left = b1->width * slice_x / s->num_x;
728 int right = b1->width *(slice_x+1) / s->num_x;
729 int top = b1->height * slice_y / s->num_y;
730 int bottom = b1->height *(slice_y+1) / s->num_y;
732 int qfactor, qoffset;
734 uint8_t *buf1 = b1->ibuf + top * b1->stride;
735 uint8_t *buf2 = b2 ? b2->ibuf + top * b2->stride: NULL;
738 if (quant > (DIRAC_MAX_QUANT_INDEX - 1)) {
739 av_log(s->avctx, AV_LOG_ERROR, "Unsupported quant %d\n", quant);
742 qfactor = ff_dirac_qscale_tab[quant];
743 qoffset = ff_dirac_qoffset_intra_tab[quant] + 2;
744 /* we have to constantly check for overread since the spec explicitly
745 requires this, with the meaning that all remaining coeffs are set to 0 */
746 if (get_bits_count(gb) >= bits_end)
750 for (y = top; y < bottom; y++) {
751 for (x = left; x < right; x++) {
752 PARSE_VALUES(int32_t, x, gb, bits_end, buf1, buf2);
760 for (y = top; y < bottom; y++) {
761 for (x = left; x < right; x++) {
762 PARSE_VALUES(int16_t, x, gb, bits_end, buf1, buf2);
772 * Dirac Specification ->
773 * 13.5.2 Slices. slice(sx,sy)
775 static int decode_lowdelay_slice(AVCodecContext *avctx, void *arg)
777 DiracContext *s = avctx->priv_data;
778 DiracSlice *slice = arg;
779 GetBitContext *gb = &slice->gb;
780 enum dirac_subband orientation;
781 int level, quant, chroma_bits, chroma_end;
783 int quant_base = get_bits(gb, 7); /*[DIRAC_STD] qindex */
784 int length_bits = av_log2(8 * slice->bytes)+1;
785 int luma_bits = get_bits_long(gb, length_bits);
786 int luma_end = get_bits_count(gb) + FFMIN(luma_bits, get_bits_left(gb));
788 /* [DIRAC_STD] 13.5.5.2 luma_slice_band */
789 for (level = 0; level < s->wavelet_depth; level++)
790 for (orientation = !!level; orientation < 4; orientation++) {
791 quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);
792 decode_subband(s, gb, quant, slice->slice_x, slice->slice_y, luma_end,
793 &s->plane[0].band[level][orientation], NULL);
796 /* consume any unused bits from luma */
797 skip_bits_long(gb, get_bits_count(gb) - luma_end);
799 chroma_bits = 8*slice->bytes - 7 - length_bits - luma_bits;
800 chroma_end = get_bits_count(gb) + FFMIN(chroma_bits, get_bits_left(gb));
801 /* [DIRAC_STD] 13.5.5.3 chroma_slice_band */
802 for (level = 0; level < s->wavelet_depth; level++)
803 for (orientation = !!level; orientation < 4; orientation++) {
804 quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);
805 decode_subband(s, gb, quant, slice->slice_x, slice->slice_y, chroma_end,
806 &s->plane[1].band[level][orientation],
807 &s->plane[2].band[level][orientation]);
813 typedef struct SliceCoeffs {
821 static int subband_coeffs(DiracContext *s, int x, int y, int p,
822 SliceCoeffs c[MAX_DWT_LEVELS])
825 for (level = 0; level < s->wavelet_depth; level++) {
826 SliceCoeffs *o = &c[level];
827 SubBand *b = &s->plane[p].band[level][3]; /* orientation doens't matter */
828 o->top = b->height * y / s->num_y;
829 o->left = b->width * x / s->num_x;
830 o->tot_h = ((b->width * (x + 1)) / s->num_x) - o->left;
831 o->tot_v = ((b->height * (y + 1)) / s->num_y) - o->top;
832 o->tot = o->tot_h*o->tot_v;
833 coef += o->tot * (4 - !!level);
839 * VC-2 Specification ->
840 * 13.5.3 hq_slice(sx,sy)
842 static int decode_hq_slice(DiracContext *s, DiracSlice *slice, uint8_t *tmp_buf)
844 int i, level, orientation, quant_idx;
845 int qfactor[MAX_DWT_LEVELS][4], qoffset[MAX_DWT_LEVELS][4];
846 GetBitContext *gb = &slice->gb;
847 SliceCoeffs coeffs_num[MAX_DWT_LEVELS];
849 skip_bits_long(gb, 8*s->highquality.prefix_bytes);
850 quant_idx = get_bits(gb, 8);
852 if (quant_idx > DIRAC_MAX_QUANT_INDEX - 1) {
853 av_log(s->avctx, AV_LOG_ERROR, "Invalid quantization index - %i\n", quant_idx);
854 return AVERROR_INVALIDDATA;
857 /* Slice quantization (slice_quantizers() in the specs) */
858 for (level = 0; level < s->wavelet_depth; level++) {
859 for (orientation = !!level; orientation < 4; orientation++) {
860 const int quant = FFMAX(quant_idx - s->lowdelay.quant[level][orientation], 0);
861 qfactor[level][orientation] = ff_dirac_qscale_tab[quant];
862 qoffset[level][orientation] = ff_dirac_qoffset_intra_tab[quant] + 2;
866 /* Luma + 2 Chroma planes */
867 for (i = 0; i < 3; i++) {
868 int coef_num, coef_par, off = 0;
869 int64_t length = s->highquality.size_scaler*get_bits(gb, 8);
870 int64_t bits_end = get_bits_count(gb) + 8*length;
871 const uint8_t *addr = align_get_bits(gb);
873 if (length*8 > get_bits_left(gb)) {
874 av_log(s->avctx, AV_LOG_ERROR, "end too far away\n");
875 return AVERROR_INVALIDDATA;
878 coef_num = subband_coeffs(s, slice->slice_x, slice->slice_y, i, coeffs_num);
881 coef_par = ff_dirac_golomb_read_32bit(s->reader_ctx, addr,
882 length, tmp_buf, coef_num);
884 coef_par = ff_dirac_golomb_read_16bit(s->reader_ctx, addr,
885 length, tmp_buf, coef_num);
887 if (coef_num > coef_par) {
888 const int start_b = coef_par * (1 << (s->pshift + 1));
889 const int end_b = coef_num * (1 << (s->pshift + 1));
890 memset(&tmp_buf[start_b], 0, end_b - start_b);
893 for (level = 0; level < s->wavelet_depth; level++) {
894 const SliceCoeffs *c = &coeffs_num[level];
895 for (orientation = !!level; orientation < 4; orientation++) {
896 const SubBand *b1 = &s->plane[i].band[level][orientation];
897 uint8_t *buf = b1->ibuf + c->top * b1->stride + (c->left << (s->pshift + 1));
899 /* Change to c->tot_h <= 4 for AVX2 dequantization */
900 const int qfunc = s->pshift + 2*(c->tot_h <= 2);
901 s->diracdsp.dequant_subband[qfunc](&tmp_buf[off], buf, b1->stride,
902 qfactor[level][orientation],
903 qoffset[level][orientation],
906 off += c->tot << (s->pshift + 1);
910 skip_bits_long(gb, bits_end - get_bits_count(gb));
916 static int decode_hq_slice_row(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
919 DiracContext *s = avctx->priv_data;
920 DiracSlice *slices = ((DiracSlice *)arg) + s->num_x*jobnr;
921 uint8_t *thread_buf = &s->thread_buf[s->thread_buf_size*threadnr];
922 for (i = 0; i < s->num_x; i++)
923 decode_hq_slice(s, &slices[i], thread_buf);
928 * Dirac Specification ->
929 * 13.5.1 low_delay_transform_data()
931 static int decode_lowdelay(DiracContext *s)
933 AVCodecContext *avctx = s->avctx;
934 int slice_x, slice_y, bufsize;
935 int64_t coef_buf_size, bytes = 0;
938 SliceCoeffs tmp[MAX_DWT_LEVELS];
941 if (s->slice_params_num_buf != (s->num_x * s->num_y)) {
942 s->slice_params_buf = av_realloc_f(s->slice_params_buf, s->num_x * s->num_y, sizeof(DiracSlice));
943 if (!s->slice_params_buf) {
944 av_log(s->avctx, AV_LOG_ERROR, "slice params buffer allocation failure\n");
945 s->slice_params_num_buf = 0;
946 return AVERROR(ENOMEM);
948 s->slice_params_num_buf = s->num_x * s->num_y;
950 slices = s->slice_params_buf;
952 /* 8 becacuse that's how much the golomb reader could overread junk data
953 * from another plane/slice at most, and 512 because SIMD */
954 coef_buf_size = subband_coeffs(s, s->num_x - 1, s->num_y - 1, 0, tmp) + 8;
955 coef_buf_size = (coef_buf_size << (1 + s->pshift)) + 512;
957 if (s->threads_num_buf != avctx->thread_count ||
958 s->thread_buf_size != coef_buf_size) {
959 s->threads_num_buf = avctx->thread_count;
960 s->thread_buf_size = coef_buf_size;
961 s->thread_buf = av_realloc_f(s->thread_buf, avctx->thread_count, s->thread_buf_size);
962 if (!s->thread_buf) {
963 av_log(s->avctx, AV_LOG_ERROR, "thread buffer allocation failure\n");
964 return AVERROR(ENOMEM);
968 align_get_bits(&s->gb);
969 /*[DIRAC_STD] 13.5.2 Slices. slice(sx,sy) */
970 buf = s->gb.buffer + get_bits_count(&s->gb)/8;
971 bufsize = get_bits_left(&s->gb);
976 for (slice_y = 0; bufsize > 0 && slice_y < s->num_y; slice_y++) {
977 for (slice_x = 0; bufsize > 0 && slice_x < s->num_x; slice_x++) {
978 bytes = s->highquality.prefix_bytes + 1;
979 for (i = 0; i < 3; i++) {
980 if (bytes <= bufsize/8)
981 bytes += buf[bytes] * s->highquality.size_scaler + 1;
983 if (bytes >= INT_MAX || bytes*8 > bufsize) {
984 av_log(s->avctx, AV_LOG_ERROR, "too many bytes\n");
985 return AVERROR_INVALIDDATA;
988 slices[slice_num].bytes = bytes;
989 slices[slice_num].slice_x = slice_x;
990 slices[slice_num].slice_y = slice_y;
991 init_get_bits(&slices[slice_num].gb, buf, bufsize);
995 if (bufsize/8 >= bytes)
1002 if (s->num_x*s->num_y != slice_num) {
1003 av_log(s->avctx, AV_LOG_ERROR, "too few slices\n");
1004 return AVERROR_INVALIDDATA;
1007 avctx->execute2(avctx, decode_hq_slice_row, slices, NULL, s->num_y);
1009 for (slice_y = 0; bufsize > 0 && slice_y < s->num_y; slice_y++) {
1010 for (slice_x = 0; bufsize > 0 && slice_x < s->num_x; slice_x++) {
1011 bytes = (slice_num+1) * (int64_t)s->lowdelay.bytes.num / s->lowdelay.bytes.den
1012 - slice_num * (int64_t)s->lowdelay.bytes.num / s->lowdelay.bytes.den;
1013 if (bytes >= INT_MAX || bytes*8 > bufsize) {
1014 av_log(s->avctx, AV_LOG_ERROR, "too many bytes\n");
1015 return AVERROR_INVALIDDATA;
1017 slices[slice_num].bytes = bytes;
1018 slices[slice_num].slice_x = slice_x;
1019 slices[slice_num].slice_y = slice_y;
1020 init_get_bits(&slices[slice_num].gb, buf, bufsize);
1024 if (bufsize/8 >= bytes)
1030 avctx->execute(avctx, decode_lowdelay_slice, slices, NULL, slice_num,
1031 sizeof(DiracSlice)); /* [DIRAC_STD] 13.5.2 Slices */
1034 if (s->dc_prediction) {
1036 intra_dc_prediction_10(&s->plane[0].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
1037 intra_dc_prediction_10(&s->plane[1].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
1038 intra_dc_prediction_10(&s->plane[2].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
1040 intra_dc_prediction_8(&s->plane[0].band[0][0]);
1041 intra_dc_prediction_8(&s->plane[1].band[0][0]);
1042 intra_dc_prediction_8(&s->plane[2].band[0][0]);
1049 static void init_planes(DiracContext *s)
1051 int i, w, h, level, orientation;
1053 for (i = 0; i < 3; i++) {
1054 Plane *p = &s->plane[i];
1056 p->width = s->seq.width >> (i ? s->chroma_x_shift : 0);
1057 p->height = s->seq.height >> (i ? s->chroma_y_shift : 0);
1058 p->idwt.width = w = CALC_PADDING(p->width , s->wavelet_depth);
1059 p->idwt.height = h = CALC_PADDING(p->height, s->wavelet_depth);
1060 p->idwt.stride = FFALIGN(p->idwt.width, 8) << (1 + s->pshift);
1062 for (level = s->wavelet_depth-1; level >= 0; level--) {
1065 for (orientation = !!level; orientation < 4; orientation++) {
1066 SubBand *b = &p->band[level][orientation];
1068 b->pshift = s->pshift;
1069 b->ibuf = p->idwt.buf;
1071 b->stride = p->idwt.stride << (s->wavelet_depth - level);
1074 b->orientation = orientation;
1076 if (orientation & 1)
1077 b->ibuf += w << (1+b->pshift);
1078 if (orientation > 1)
1079 b->ibuf += (b->stride>>1);
1082 b->parent = &p->band[level-1][orientation];
1087 p->xblen = s->plane[0].xblen >> s->chroma_x_shift;
1088 p->yblen = s->plane[0].yblen >> s->chroma_y_shift;
1089 p->xbsep = s->plane[0].xbsep >> s->chroma_x_shift;
1090 p->ybsep = s->plane[0].ybsep >> s->chroma_y_shift;
1093 p->xoffset = (p->xblen - p->xbsep)/2;
1094 p->yoffset = (p->yblen - p->ybsep)/2;
1099 * Unpack the motion compensation parameters
1100 * Dirac Specification ->
1101 * 11.2 Picture prediction data. picture_prediction()
1103 static int dirac_unpack_prediction_parameters(DiracContext *s)
1105 static const uint8_t default_blen[] = { 4, 12, 16, 24 };
1107 GetBitContext *gb = &s->gb;
1111 /* [DIRAC_STD] 11.2.2 Block parameters. block_parameters() */
1112 /* Luma and Chroma are equal. 11.2.3 */
1113 idx = get_interleaved_ue_golomb(gb); /* [DIRAC_STD] index */
1116 av_log(s->avctx, AV_LOG_ERROR, "Block prediction index too high\n");
1117 return AVERROR_INVALIDDATA;
1121 s->plane[0].xblen = get_interleaved_ue_golomb(gb);
1122 s->plane[0].yblen = get_interleaved_ue_golomb(gb);
1123 s->plane[0].xbsep = get_interleaved_ue_golomb(gb);
1124 s->plane[0].ybsep = get_interleaved_ue_golomb(gb);
1126 /*[DIRAC_STD] preset_block_params(index). Table 11.1 */
1127 s->plane[0].xblen = default_blen[idx-1];
1128 s->plane[0].yblen = default_blen[idx-1];
1129 s->plane[0].xbsep = 4 * idx;
1130 s->plane[0].ybsep = 4 * idx;
1132 /*[DIRAC_STD] 11.2.4 motion_data_dimensions()
1133 Calculated in function dirac_unpack_block_motion_data */
1135 if (s->plane[0].xblen % (1 << s->chroma_x_shift) != 0 ||
1136 s->plane[0].yblen % (1 << s->chroma_y_shift) != 0 ||
1137 !s->plane[0].xblen || !s->plane[0].yblen) {
1138 av_log(s->avctx, AV_LOG_ERROR,
1139 "invalid x/y block length (%d/%d) for x/y chroma shift (%d/%d)\n",
1140 s->plane[0].xblen, s->plane[0].yblen, s->chroma_x_shift, s->chroma_y_shift);
1141 return AVERROR_INVALIDDATA;
1143 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) {
1144 av_log(s->avctx, AV_LOG_ERROR, "Block separation too small\n");
1145 return AVERROR_INVALIDDATA;
1147 if (s->plane[0].xbsep > s->plane[0].xblen || s->plane[0].ybsep > s->plane[0].yblen) {
1148 av_log(s->avctx, AV_LOG_ERROR, "Block separation greater than size\n");
1149 return AVERROR_INVALIDDATA;
1151 if (FFMAX(s->plane[0].xblen, s->plane[0].yblen) > MAX_BLOCKSIZE) {
1152 av_log(s->avctx, AV_LOG_ERROR, "Unsupported large block size\n");
1153 return AVERROR_PATCHWELCOME;
1156 /*[DIRAC_STD] 11.2.5 Motion vector precision. motion_vector_precision()
1157 Read motion vector precision */
1158 s->mv_precision = get_interleaved_ue_golomb(gb);
1159 if (s->mv_precision > 3) {
1160 av_log(s->avctx, AV_LOG_ERROR, "MV precision finer than eighth-pel\n");
1161 return AVERROR_INVALIDDATA;
1164 /*[DIRAC_STD] 11.2.6 Global motion. global_motion()
1165 Read the global motion compensation parameters */
1166 s->globalmc_flag = get_bits1(gb);
1167 if (s->globalmc_flag) {
1168 memset(s->globalmc, 0, sizeof(s->globalmc));
1169 /* [DIRAC_STD] pan_tilt(gparams) */
1170 for (ref = 0; ref < s->num_refs; ref++) {
1171 if (get_bits1(gb)) {
1172 s->globalmc[ref].pan_tilt[0] = dirac_get_se_golomb(gb);
1173 s->globalmc[ref].pan_tilt[1] = dirac_get_se_golomb(gb);
1175 /* [DIRAC_STD] zoom_rotate_shear(gparams)
1176 zoom/rotation/shear parameters */
1177 if (get_bits1(gb)) {
1178 s->globalmc[ref].zrs_exp = get_interleaved_ue_golomb(gb);
1179 s->globalmc[ref].zrs[0][0] = dirac_get_se_golomb(gb);
1180 s->globalmc[ref].zrs[0][1] = dirac_get_se_golomb(gb);
1181 s->globalmc[ref].zrs[1][0] = dirac_get_se_golomb(gb);
1182 s->globalmc[ref].zrs[1][1] = dirac_get_se_golomb(gb);
1184 s->globalmc[ref].zrs[0][0] = 1;
1185 s->globalmc[ref].zrs[1][1] = 1;
1187 /* [DIRAC_STD] perspective(gparams) */
1188 if (get_bits1(gb)) {
1189 s->globalmc[ref].perspective_exp = get_interleaved_ue_golomb(gb);
1190 s->globalmc[ref].perspective[0] = dirac_get_se_golomb(gb);
1191 s->globalmc[ref].perspective[1] = dirac_get_se_golomb(gb);
1193 if (s->globalmc[ref].perspective_exp + (uint64_t)s->globalmc[ref].zrs_exp > 30) {
1194 return AVERROR_INVALIDDATA;
1200 /*[DIRAC_STD] 11.2.7 Picture prediction mode. prediction_mode()
1201 Picture prediction mode, not currently used. */
1202 if (get_interleaved_ue_golomb(gb)) {
1203 av_log(s->avctx, AV_LOG_ERROR, "Unknown picture prediction mode\n");
1204 return AVERROR_INVALIDDATA;
1207 /* [DIRAC_STD] 11.2.8 Reference picture weight. reference_picture_weights()
1208 just data read, weight calculation will be done later on. */
1209 s->weight_log2denom = 1;
1213 if (get_bits1(gb)) {
1214 s->weight_log2denom = get_interleaved_ue_golomb(gb);
1215 if (s->weight_log2denom < 1 || s->weight_log2denom > 8) {
1216 av_log(s->avctx, AV_LOG_ERROR, "weight_log2denom unsupported or invalid\n");
1217 s->weight_log2denom = 1;
1218 return AVERROR_INVALIDDATA;
1220 s->weight[0] = dirac_get_se_golomb(gb);
1221 if (s->num_refs == 2)
1222 s->weight[1] = dirac_get_se_golomb(gb);
1228 * Dirac Specification ->
1229 * 11.3 Wavelet transform data. wavelet_transform()
1231 static int dirac_unpack_idwt_params(DiracContext *s)
1233 GetBitContext *gb = &s->gb;
1237 #define CHECKEDREAD(dst, cond, errmsg) \
1238 tmp = get_interleaved_ue_golomb(gb); \
1240 av_log(s->avctx, AV_LOG_ERROR, errmsg); \
1241 return AVERROR_INVALIDDATA; \
1247 s->zero_res = s->num_refs ? get_bits1(gb) : 0;
1251 /*[DIRAC_STD] 11.3.1 Transform parameters. transform_parameters() */
1252 CHECKEDREAD(s->wavelet_idx, tmp > 6, "wavelet_idx is too big\n")
1254 CHECKEDREAD(s->wavelet_depth, tmp > MAX_DWT_LEVELS || tmp < 1, "invalid number of DWT decompositions\n")
1256 if (!s->low_delay) {
1257 /* Codeblock parameters (core syntax only) */
1258 if (get_bits1(gb)) {
1259 for (i = 0; i <= s->wavelet_depth; i++) {
1260 CHECKEDREAD(s->codeblock[i].width , tmp < 1 || tmp > (s->avctx->width >>s->wavelet_depth-i), "codeblock width invalid\n")
1261 CHECKEDREAD(s->codeblock[i].height, tmp < 1 || tmp > (s->avctx->height>>s->wavelet_depth-i), "codeblock height invalid\n")
1264 CHECKEDREAD(s->codeblock_mode, tmp > 1, "unknown codeblock mode\n")
1267 for (i = 0; i <= s->wavelet_depth; i++)
1268 s->codeblock[i].width = s->codeblock[i].height = 1;
1272 s->num_x = get_interleaved_ue_golomb(gb);
1273 s->num_y = get_interleaved_ue_golomb(gb);
1274 if (s->num_x * s->num_y == 0 || s->num_x * (uint64_t)s->num_y > INT_MAX ||
1275 s->num_x * (uint64_t)s->avctx->width > INT_MAX ||
1276 s->num_y * (uint64_t)s->avctx->height > INT_MAX
1278 av_log(s->avctx,AV_LOG_ERROR,"Invalid numx/y\n");
1279 s->num_x = s->num_y = 0;
1280 return AVERROR_INVALIDDATA;
1282 if (s->ld_picture) {
1283 s->lowdelay.bytes.num = get_interleaved_ue_golomb(gb);
1284 s->lowdelay.bytes.den = get_interleaved_ue_golomb(gb);
1285 if (s->lowdelay.bytes.den <= 0) {
1286 av_log(s->avctx,AV_LOG_ERROR,"Invalid lowdelay.bytes.den\n");
1287 return AVERROR_INVALIDDATA;
1289 } else if (s->hq_picture) {
1290 s->highquality.prefix_bytes = get_interleaved_ue_golomb(gb);
1291 s->highquality.size_scaler = get_interleaved_ue_golomb(gb);
1292 if (s->highquality.prefix_bytes >= INT_MAX / 8) {
1293 av_log(s->avctx,AV_LOG_ERROR,"too many prefix bytes\n");
1294 return AVERROR_INVALIDDATA;
1298 /* [DIRAC_STD] 11.3.5 Quantisation matrices (low-delay syntax). quant_matrix() */
1299 if (get_bits1(gb)) {
1300 av_log(s->avctx,AV_LOG_DEBUG,"Low Delay: Has Custom Quantization Matrix!\n");
1301 /* custom quantization matrix */
1302 for (level = 0; level < s->wavelet_depth; level++) {
1303 for (i = !!level; i < 4; i++) {
1304 s->lowdelay.quant[level][i] = get_interleaved_ue_golomb(gb);
1308 if (s->wavelet_depth > 4) {
1309 av_log(s->avctx,AV_LOG_ERROR,"Mandatory custom low delay matrix missing for depth %d\n", s->wavelet_depth);
1310 return AVERROR_INVALIDDATA;
1312 /* default quantization matrix */
1313 for (level = 0; level < s->wavelet_depth; level++)
1314 for (i = 0; i < 4; i++) {
1315 s->lowdelay.quant[level][i] = ff_dirac_default_qmat[s->wavelet_idx][level][i];
1316 /* haar with no shift differs for different depths */
1317 if (s->wavelet_idx == 3)
1318 s->lowdelay.quant[level][i] += 4*(s->wavelet_depth-1 - level);
1325 static inline int pred_sbsplit(uint8_t *sbsplit, int stride, int x, int y)
1327 static const uint8_t avgsplit[7] = { 0, 0, 1, 1, 1, 2, 2 };
1334 return sbsplit[-stride];
1336 return avgsplit[sbsplit[-1] + sbsplit[-stride] + sbsplit[-stride-1]];
1339 static inline int pred_block_mode(DiracBlock *block, int stride, int x, int y, int refmask)
1346 return block[-1].ref & refmask;
1348 return block[-stride].ref & refmask;
1350 /* return the majority */
1351 pred = (block[-1].ref & refmask) + (block[-stride].ref & refmask) + (block[-stride-1].ref & refmask);
1352 return (pred >> 1) & refmask;
1355 static inline void pred_block_dc(DiracBlock *block, int stride, int x, int y)
1359 memset(block->u.dc, 0, sizeof(block->u.dc));
1361 if (x && !(block[-1].ref & 3)) {
1362 for (i = 0; i < 3; i++)
1363 block->u.dc[i] += block[-1].u.dc[i];
1367 if (y && !(block[-stride].ref & 3)) {
1368 for (i = 0; i < 3; i++)
1369 block->u.dc[i] += block[-stride].u.dc[i];
1373 if (x && y && !(block[-1-stride].ref & 3)) {
1374 for (i = 0; i < 3; i++)
1375 block->u.dc[i] += block[-1-stride].u.dc[i];
1380 for (i = 0; i < 3; i++)
1381 block->u.dc[i] = (block->u.dc[i]+1)>>1;
1382 } else if (n == 3) {
1383 for (i = 0; i < 3; i++)
1384 block->u.dc[i] = divide3(block->u.dc[i]);
1388 static inline void pred_mv(DiracBlock *block, int stride, int x, int y, int ref)
1391 int refmask = ref+1;
1392 int mask = refmask | DIRAC_REF_MASK_GLOBAL; /* exclude gmc blocks */
1395 if (x && (block[-1].ref & mask) == refmask)
1396 pred[n++] = block[-1].u.mv[ref];
1398 if (y && (block[-stride].ref & mask) == refmask)
1399 pred[n++] = block[-stride].u.mv[ref];
1401 if (x && y && (block[-stride-1].ref & mask) == refmask)
1402 pred[n++] = block[-stride-1].u.mv[ref];
1406 block->u.mv[ref][0] = 0;
1407 block->u.mv[ref][1] = 0;
1410 block->u.mv[ref][0] = pred[0][0];
1411 block->u.mv[ref][1] = pred[0][1];
1414 block->u.mv[ref][0] = (pred[0][0] + pred[1][0] + 1) >> 1;
1415 block->u.mv[ref][1] = (pred[0][1] + pred[1][1] + 1) >> 1;
1418 block->u.mv[ref][0] = mid_pred(pred[0][0], pred[1][0], pred[2][0]);
1419 block->u.mv[ref][1] = mid_pred(pred[0][1], pred[1][1], pred[2][1]);
1424 static void global_mv(DiracContext *s, DiracBlock *block, int x, int y, int ref)
1426 int ez = s->globalmc[ref].zrs_exp;
1427 int ep = s->globalmc[ref].perspective_exp;
1428 int (*A)[2] = s->globalmc[ref].zrs;
1429 int *b = s->globalmc[ref].pan_tilt;
1430 int *c = s->globalmc[ref].perspective;
1432 int m = (1<<ep) - (c[0]*x + c[1]*y);
1433 int64_t mx = m * (int64_t)((A[0][0] * (int64_t)x + A[0][1]*(int64_t)y) + (1<<ez) * b[0]);
1434 int64_t my = m * (int64_t)((A[1][0] * (int64_t)x + A[1][1]*(int64_t)y) + (1<<ez) * b[1]);
1436 block->u.mv[ref][0] = (mx + (1<<(ez+ep))) >> (ez+ep);
1437 block->u.mv[ref][1] = (my + (1<<(ez+ep))) >> (ez+ep);
1440 static void decode_block_params(DiracContext *s, DiracArith arith[8], DiracBlock *block,
1441 int stride, int x, int y)
1445 block->ref = pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF1);
1446 block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF1);
1448 if (s->num_refs == 2) {
1449 block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF2);
1450 block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF2) << 1;
1454 pred_block_dc(block, stride, x, y);
1455 for (i = 0; i < 3; i++)
1456 block->u.dc[i] += (unsigned)dirac_get_arith_int(arith+1+i, CTX_DC_F1, CTX_DC_DATA);
1460 if (s->globalmc_flag) {
1461 block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_GLOBAL);
1462 block->ref ^= dirac_get_arith_bit(arith, CTX_GLOBAL_BLOCK) << 2;
1465 for (i = 0; i < s->num_refs; i++)
1466 if (block->ref & (i+1)) {
1467 if (block->ref & DIRAC_REF_MASK_GLOBAL) {
1468 global_mv(s, block, x, y, i);
1470 pred_mv(block, stride, x, y, i);
1471 block->u.mv[i][0] += (unsigned)dirac_get_arith_int(arith + 4 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
1472 block->u.mv[i][1] += (unsigned)dirac_get_arith_int(arith + 5 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
1478 * Copies the current block to the other blocks covered by the current superblock split mode
1480 static void propagate_block_data(DiracBlock *block, int stride, int size)
1483 DiracBlock *dst = block;
1485 for (x = 1; x < size; x++)
1488 for (y = 1; y < size; y++) {
1490 for (x = 0; x < size; x++)
1496 * Dirac Specification ->
1497 * 12. Block motion data syntax
1499 static int dirac_unpack_block_motion_data(DiracContext *s)
1501 GetBitContext *gb = &s->gb;
1502 uint8_t *sbsplit = s->sbsplit;
1504 DiracArith arith[8];
1508 /* [DIRAC_STD] 11.2.4 and 12.2.1 Number of blocks and superblocks */
1509 s->sbwidth = DIVRNDUP(s->seq.width, 4*s->plane[0].xbsep);
1510 s->sbheight = DIVRNDUP(s->seq.height, 4*s->plane[0].ybsep);
1511 s->blwidth = 4 * s->sbwidth;
1512 s->blheight = 4 * s->sbheight;
1514 /* [DIRAC_STD] 12.3.1 Superblock splitting modes. superblock_split_modes()
1515 decode superblock split modes */
1516 ff_dirac_init_arith_decoder(arith, gb, get_interleaved_ue_golomb(gb)); /* get_interleaved_ue_golomb(gb) is the length */
1517 for (y = 0; y < s->sbheight; y++) {
1518 for (x = 0; x < s->sbwidth; x++) {
1519 unsigned int split = dirac_get_arith_uint(arith, CTX_SB_F1, CTX_SB_DATA);
1521 return AVERROR_INVALIDDATA;
1522 sbsplit[x] = (split + pred_sbsplit(sbsplit+x, s->sbwidth, x, y)) % 3;
1524 sbsplit += s->sbwidth;
1527 /* setup arith decoding */
1528 ff_dirac_init_arith_decoder(arith, gb, get_interleaved_ue_golomb(gb));
1529 for (i = 0; i < s->num_refs; i++) {
1530 ff_dirac_init_arith_decoder(arith + 4 + 2 * i, gb, get_interleaved_ue_golomb(gb));
1531 ff_dirac_init_arith_decoder(arith + 5 + 2 * i, gb, get_interleaved_ue_golomb(gb));
1533 for (i = 0; i < 3; i++)
1534 ff_dirac_init_arith_decoder(arith+1+i, gb, get_interleaved_ue_golomb(gb));
1536 for (y = 0; y < s->sbheight; y++)
1537 for (x = 0; x < s->sbwidth; x++) {
1538 int blkcnt = 1 << s->sbsplit[y * s->sbwidth + x];
1539 int step = 4 >> s->sbsplit[y * s->sbwidth + x];
1541 for (q = 0; q < blkcnt; q++)
1542 for (p = 0; p < blkcnt; p++) {
1543 int bx = 4 * x + p*step;
1544 int by = 4 * y + q*step;
1545 DiracBlock *block = &s->blmotion[by*s->blwidth + bx];
1546 decode_block_params(s, arith, block, s->blwidth, bx, by);
1547 propagate_block_data(block, s->blwidth, step);
1554 static int weight(int i, int blen, int offset)
1556 #define ROLLOFF(i) offset == 1 ? ((i) ? 5 : 3) : \
1557 (1 + (6*(i) + offset - 1) / (2*offset - 1))
1561 else if (i > blen-1 - 2*offset)
1562 return ROLLOFF(blen-1 - i);
1566 static void init_obmc_weight_row(Plane *p, uint8_t *obmc_weight, int stride,
1567 int left, int right, int wy)
1570 for (x = 0; left && x < p->xblen >> 1; x++)
1571 obmc_weight[x] = wy*8;
1572 for (; x < p->xblen >> right; x++)
1573 obmc_weight[x] = wy*weight(x, p->xblen, p->xoffset);
1574 for (; x < p->xblen; x++)
1575 obmc_weight[x] = wy*8;
1576 for (; x < stride; x++)
1580 static void init_obmc_weight(Plane *p, uint8_t *obmc_weight, int stride,
1581 int left, int right, int top, int bottom)
1584 for (y = 0; top && y < p->yblen >> 1; y++) {
1585 init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
1586 obmc_weight += stride;
1588 for (; y < p->yblen >> bottom; y++) {
1589 int wy = weight(y, p->yblen, p->yoffset);
1590 init_obmc_weight_row(p, obmc_weight, stride, left, right, wy);
1591 obmc_weight += stride;
1593 for (; y < p->yblen; y++) {
1594 init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
1595 obmc_weight += stride;
1599 static void init_obmc_weights(DiracContext *s, Plane *p, int by)
1602 int bottom = by == s->blheight-1;
1604 /* don't bother re-initing for rows 2 to blheight-2, the weights don't change */
1605 if (top || bottom || by == 1) {
1606 init_obmc_weight(p, s->obmc_weight[0], MAX_BLOCKSIZE, 1, 0, top, bottom);
1607 init_obmc_weight(p, s->obmc_weight[1], MAX_BLOCKSIZE, 0, 0, top, bottom);
1608 init_obmc_weight(p, s->obmc_weight[2], MAX_BLOCKSIZE, 0, 1, top, bottom);
1612 static const uint8_t epel_weights[4][4][4] = {
1632 * For block x,y, determine which of the hpel planes to do bilinear
1633 * interpolation from and set src[] to the location in each hpel plane
1636 * @return the index of the put_dirac_pixels_tab function to use
1637 * 0 for 1 plane (fpel,hpel), 1 for 2 planes (qpel), 2 for 4 planes (qpel), and 3 for epel
1639 static int mc_subpel(DiracContext *s, DiracBlock *block, const uint8_t *src[5],
1640 int x, int y, int ref, int plane)
1642 Plane *p = &s->plane[plane];
1643 uint8_t **ref_hpel = s->ref_pics[ref]->hpel[plane];
1644 int motion_x = block->u.mv[ref][0];
1645 int motion_y = block->u.mv[ref][1];
1646 int mx, my, i, epel, nplanes = 0;
1649 motion_x >>= s->chroma_x_shift;
1650 motion_y >>= s->chroma_y_shift;
1653 mx = motion_x & ~(-1U << s->mv_precision);
1654 my = motion_y & ~(-1U << s->mv_precision);
1655 motion_x >>= s->mv_precision;
1656 motion_y >>= s->mv_precision;
1657 /* normalize subpel coordinates to epel */
1658 /* TODO: template this function? */
1659 mx <<= 3 - s->mv_precision;
1660 my <<= 3 - s->mv_precision;
1669 src[0] = ref_hpel[(my>>1)+(mx>>2)] + y*p->stride + x;
1673 for (i = 0; i < 4; i++)
1674 src[i] = ref_hpel[i] + y*p->stride + x;
1676 /* if we're interpolating in the right/bottom halves, adjust the planes as needed
1677 we increment x/y because the edge changes for half of the pixels */
1684 src[0] += p->stride;
1685 src[1] += p->stride;
1693 /* check if we really only need 2 planes since either mx or my is
1694 a hpel position. (epel weights of 0 handle this there) */
1696 /* mx == 0: average [0] and [2]
1697 mx == 4: average [1] and [3] */
1698 src[!mx] = src[2 + !!mx];
1700 } else if (!(my&3)) {
1701 src[0] = src[(my>>1) ];
1702 src[1] = src[(my>>1)+1];
1706 /* adjust the ordering if needed so the weights work */
1708 FFSWAP(const uint8_t *, src[0], src[1]);
1709 FFSWAP(const uint8_t *, src[2], src[3]);
1712 FFSWAP(const uint8_t *, src[0], src[2]);
1713 FFSWAP(const uint8_t *, src[1], src[3]);
1715 src[4] = epel_weights[my&3][mx&3];
1719 /* fixme: v/h _edge_pos */
1720 if (x + p->xblen > p->width +EDGE_WIDTH/2 ||
1721 y + p->yblen > p->height+EDGE_WIDTH/2 ||
1723 for (i = 0; i < nplanes; i++) {
1724 s->vdsp.emulated_edge_mc(s->edge_emu_buffer[i], src[i],
1725 p->stride, p->stride,
1726 p->xblen, p->yblen, x, y,
1727 p->width+EDGE_WIDTH/2, p->height+EDGE_WIDTH/2);
1728 src[i] = s->edge_emu_buffer[i];
1731 return (nplanes>>1) + epel;
1734 static void add_dc(uint16_t *dst, int dc, int stride,
1735 uint8_t *obmc_weight, int xblen, int yblen)
1740 for (y = 0; y < yblen; y++) {
1741 for (x = 0; x < xblen; x += 2) {
1742 dst[x ] += dc * obmc_weight[x ];
1743 dst[x+1] += dc * obmc_weight[x+1];
1746 obmc_weight += MAX_BLOCKSIZE;
1750 static void block_mc(DiracContext *s, DiracBlock *block,
1751 uint16_t *mctmp, uint8_t *obmc_weight,
1752 int plane, int dstx, int dsty)
1754 Plane *p = &s->plane[plane];
1755 const uint8_t *src[5];
1758 switch (block->ref&3) {
1760 add_dc(mctmp, block->u.dc[plane], p->stride, obmc_weight, p->xblen, p->yblen);
1764 idx = mc_subpel(s, block, src, dstx, dsty, (block->ref&3)-1, plane);
1765 s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1767 s->weight_func(s->mcscratch, p->stride, s->weight_log2denom,
1768 s->weight[0] + s->weight[1], p->yblen);
1771 idx = mc_subpel(s, block, src, dstx, dsty, 0, plane);
1772 s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1773 idx = mc_subpel(s, block, src, dstx, dsty, 1, plane);
1774 if (s->biweight_func) {
1775 /* fixme: +32 is a quick hack */
1776 s->put_pixels_tab[idx](s->mcscratch + 32, src, p->stride, p->yblen);
1777 s->biweight_func(s->mcscratch, s->mcscratch+32, p->stride, s->weight_log2denom,
1778 s->weight[0], s->weight[1], p->yblen);
1780 s->avg_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1783 s->add_obmc(mctmp, s->mcscratch, p->stride, obmc_weight, p->yblen);
1786 static void mc_row(DiracContext *s, DiracBlock *block, uint16_t *mctmp, int plane, int dsty)
1788 Plane *p = &s->plane[plane];
1789 int x, dstx = p->xbsep - p->xoffset;
1791 block_mc(s, block, mctmp, s->obmc_weight[0], plane, -p->xoffset, dsty);
1794 for (x = 1; x < s->blwidth-1; x++) {
1795 block_mc(s, block+x, mctmp, s->obmc_weight[1], plane, dstx, dsty);
1799 block_mc(s, block+x, mctmp, s->obmc_weight[2], plane, dstx, dsty);
1802 static void select_dsp_funcs(DiracContext *s, int width, int height, int xblen, int yblen)
1810 memcpy(s->put_pixels_tab, s->diracdsp.put_dirac_pixels_tab[idx], sizeof(s->put_pixels_tab));
1811 memcpy(s->avg_pixels_tab, s->diracdsp.avg_dirac_pixels_tab[idx], sizeof(s->avg_pixels_tab));
1812 s->add_obmc = s->diracdsp.add_dirac_obmc[idx];
1813 if (s->weight_log2denom > 1 || s->weight[0] != 1 || s->weight[1] != 1) {
1814 s->weight_func = s->diracdsp.weight_dirac_pixels_tab[idx];
1815 s->biweight_func = s->diracdsp.biweight_dirac_pixels_tab[idx];
1817 s->weight_func = NULL;
1818 s->biweight_func = NULL;
1822 static int interpolate_refplane(DiracContext *s, DiracFrame *ref, int plane, int width, int height)
1824 /* chroma allocates an edge of 8 when subsampled
1825 which for 4:2:2 means an h edge of 16 and v edge of 8
1826 just use 8 for everything for the moment */
1827 int i, edge = EDGE_WIDTH/2;
1829 ref->hpel[plane][0] = ref->avframe->data[plane];
1830 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 */
1832 /* no need for hpel if we only have fpel vectors */
1833 if (!s->mv_precision)
1836 for (i = 1; i < 4; i++) {
1837 if (!ref->hpel_base[plane][i])
1838 ref->hpel_base[plane][i] = av_malloc((height+2*edge) * ref->avframe->linesize[plane] + 32);
1839 if (!ref->hpel_base[plane][i]) {
1840 return AVERROR(ENOMEM);
1842 /* we need to be 16-byte aligned even for chroma */
1843 ref->hpel[plane][i] = ref->hpel_base[plane][i] + edge*ref->avframe->linesize[plane] + 16;
1846 if (!ref->interpolated[plane]) {
1847 s->diracdsp.dirac_hpel_filter(ref->hpel[plane][1], ref->hpel[plane][2],
1848 ref->hpel[plane][3], ref->hpel[plane][0],
1849 ref->avframe->linesize[plane], width, height);
1850 s->mpvencdsp.draw_edges(ref->hpel[plane][1], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1851 s->mpvencdsp.draw_edges(ref->hpel[plane][2], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1852 s->mpvencdsp.draw_edges(ref->hpel[plane][3], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1854 ref->interpolated[plane] = 1;
1860 * Dirac Specification ->
1861 * 13.0 Transform data syntax. transform_data()
1863 static int dirac_decode_frame_internal(DiracContext *s)
1866 int y, i, comp, dsty;
1870 /* [DIRAC_STD] 13.5.1 low_delay_transform_data() */
1871 if (!s->hq_picture) {
1872 for (comp = 0; comp < 3; comp++) {
1873 Plane *p = &s->plane[comp];
1874 memset(p->idwt.buf, 0, p->idwt.stride * p->idwt.height);
1878 if ((ret = decode_lowdelay(s)) < 0)
1883 for (comp = 0; comp < 3; comp++) {
1884 Plane *p = &s->plane[comp];
1885 uint8_t *frame = s->current_picture->avframe->data[comp];
1887 /* FIXME: small resolutions */
1888 for (i = 0; i < 4; i++)
1889 s->edge_emu_buffer[i] = s->edge_emu_buffer_base + i*FFALIGN(p->width, 16);
1891 if (!s->zero_res && !s->low_delay)
1893 memset(p->idwt.buf, 0, p->idwt.stride * p->idwt.height);
1894 ret = decode_component(s, comp); /* [DIRAC_STD] 13.4.1 core_transform_data() */
1898 ret = ff_spatial_idwt_init(&d, &p->idwt, s->wavelet_idx+2,
1899 s->wavelet_depth, s->bit_depth);
1903 if (!s->num_refs) { /* intra */
1904 for (y = 0; y < p->height; y += 16) {
1905 int idx = (s->bit_depth - 8) >> 1;
1906 ff_spatial_idwt_slice2(&d, y+16); /* decode */
1907 s->diracdsp.put_signed_rect_clamped[idx](frame + y*p->stride,
1909 p->idwt.buf + y*p->idwt.stride,
1910 p->idwt.stride, p->width, 16);
1912 } else { /* inter */
1913 int rowheight = p->ybsep*p->stride;
1915 select_dsp_funcs(s, p->width, p->height, p->xblen, p->yblen);
1917 for (i = 0; i < s->num_refs; i++) {
1918 int ret = interpolate_refplane(s, s->ref_pics[i], comp, p->width, p->height);
1923 memset(s->mctmp, 0, 4*p->yoffset*p->stride);
1926 for (y = 0; y < s->blheight; y++) {
1928 start = FFMAX(dsty, 0);
1929 uint16_t *mctmp = s->mctmp + y*rowheight;
1930 DiracBlock *blocks = s->blmotion + y*s->blwidth;
1932 init_obmc_weights(s, p, y);
1934 if (y == s->blheight-1 || start+p->ybsep > p->height)
1935 h = p->height - start;
1937 h = p->ybsep - (start - dsty);
1941 memset(mctmp+2*p->yoffset*p->stride, 0, 2*rowheight);
1942 mc_row(s, blocks, mctmp, comp, dsty);
1944 mctmp += (start - dsty)*p->stride + p->xoffset;
1945 ff_spatial_idwt_slice2(&d, start + h); /* decode */
1946 /* NOTE: add_rect_clamped hasn't been templated hence the shifts.
1947 * idwt.stride is passed as pixels, not in bytes as in the rest of the decoder */
1948 s->diracdsp.add_rect_clamped(frame + start*p->stride, mctmp, p->stride,
1949 (int16_t*)(p->idwt.buf) + start*(p->idwt.stride >> 1), (p->idwt.stride >> 1), p->width, h);
1960 static int get_buffer_with_edge(AVCodecContext *avctx, AVFrame *f, int flags)
1963 int chroma_x_shift, chroma_y_shift;
1964 ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt, &chroma_x_shift,
1969 f->width = avctx->width + 2 * EDGE_WIDTH;
1970 f->height = avctx->height + 2 * EDGE_WIDTH + 2;
1971 ret = ff_get_buffer(avctx, f, flags);
1975 for (i = 0; f->data[i]; i++) {
1976 int offset = (EDGE_WIDTH >> (i && i<3 ? chroma_y_shift : 0)) *
1977 f->linesize[i] + 32;
1978 f->data[i] += offset;
1980 f->width = avctx->width;
1981 f->height = avctx->height;
1987 * Dirac Specification ->
1988 * 11.1.1 Picture Header. picture_header()
1990 static int dirac_decode_picture_header(DiracContext *s)
1992 unsigned retire, picnum;
1994 int64_t refdist, refnum;
1995 GetBitContext *gb = &s->gb;
1997 /* [DIRAC_STD] 11.1.1 Picture Header. picture_header() PICTURE_NUM */
1998 picnum = s->current_picture->avframe->display_picture_number = get_bits_long(gb, 32);
2001 av_log(s->avctx,AV_LOG_DEBUG,"PICTURE_NUM: %d\n",picnum);
2003 /* if this is the first keyframe after a sequence header, start our
2004 reordering from here */
2005 if (s->frame_number < 0)
2006 s->frame_number = picnum;
2008 s->ref_pics[0] = s->ref_pics[1] = NULL;
2009 for (i = 0; i < s->num_refs; i++) {
2010 refnum = (picnum + dirac_get_se_golomb(gb)) & 0xFFFFFFFF;
2011 refdist = INT64_MAX;
2013 /* find the closest reference to the one we want */
2014 /* Jordi: this is needed if the referenced picture hasn't yet arrived */
2015 for (j = 0; j < MAX_REFERENCE_FRAMES && refdist; j++)
2016 if (s->ref_frames[j]
2017 && FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum) < refdist) {
2018 s->ref_pics[i] = s->ref_frames[j];
2019 refdist = FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum);
2022 if (!s->ref_pics[i] || refdist)
2023 av_log(s->avctx, AV_LOG_DEBUG, "Reference not found\n");
2025 /* if there were no references at all, allocate one */
2026 if (!s->ref_pics[i])
2027 for (j = 0; j < MAX_FRAMES; j++)
2028 if (!s->all_frames[j].avframe->data[0]) {
2029 s->ref_pics[i] = &s->all_frames[j];
2030 ret = get_buffer_with_edge(s->avctx, s->ref_pics[i]->avframe, AV_GET_BUFFER_FLAG_REF);
2036 if (!s->ref_pics[i]) {
2037 av_log(s->avctx, AV_LOG_ERROR, "Reference could not be allocated\n");
2038 return AVERROR_INVALIDDATA;
2043 /* retire the reference frames that are not used anymore */
2044 if (s->current_picture->reference) {
2045 retire = (picnum + dirac_get_se_golomb(gb)) & 0xFFFFFFFF;
2046 if (retire != picnum) {
2047 DiracFrame *retire_pic = remove_frame(s->ref_frames, retire);
2050 retire_pic->reference &= DELAYED_PIC_REF;
2052 av_log(s->avctx, AV_LOG_DEBUG, "Frame to retire not found\n");
2055 /* if reference array is full, remove the oldest as per the spec */
2056 while (add_frame(s->ref_frames, MAX_REFERENCE_FRAMES, s->current_picture)) {
2057 av_log(s->avctx, AV_LOG_ERROR, "Reference frame overflow\n");
2058 remove_frame(s->ref_frames, s->ref_frames[0]->avframe->display_picture_number)->reference &= DELAYED_PIC_REF;
2063 ret = dirac_unpack_prediction_parameters(s); /* [DIRAC_STD] 11.2 Picture Prediction Data. picture_prediction() */
2066 ret = dirac_unpack_block_motion_data(s); /* [DIRAC_STD] 12. Block motion data syntax */
2070 ret = dirac_unpack_idwt_params(s); /* [DIRAC_STD] 11.3 Wavelet transform data */
2078 static int get_delayed_pic(DiracContext *s, AVFrame *picture, int *got_frame)
2080 DiracFrame *out = s->delay_frames[0];
2084 /* find frame with lowest picture number */
2085 for (i = 1; s->delay_frames[i]; i++)
2086 if (s->delay_frames[i]->avframe->display_picture_number < out->avframe->display_picture_number) {
2087 out = s->delay_frames[i];
2091 for (i = out_idx; s->delay_frames[i]; i++)
2092 s->delay_frames[i] = s->delay_frames[i+1];
2095 out->reference ^= DELAYED_PIC_REF;
2096 if((ret = av_frame_ref(picture, out->avframe)) < 0)
2105 * Dirac Specification ->
2106 * 9.6 Parse Info Header Syntax. parse_info()
2107 * 4 byte start code + byte parse code + 4 byte size + 4 byte previous size
2109 #define DATA_UNIT_HEADER_SIZE 13
2111 /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3
2112 inside the function parse_sequence() */
2113 static int dirac_decode_data_unit(AVCodecContext *avctx, const uint8_t *buf, int size)
2115 DiracContext *s = avctx->priv_data;
2116 DiracFrame *pic = NULL;
2117 AVDiracSeqHeader *dsh;
2122 if (size < DATA_UNIT_HEADER_SIZE)
2123 return AVERROR_INVALIDDATA;
2125 parse_code = buf[4];
2127 init_get_bits(&s->gb, &buf[13], 8*(size - DATA_UNIT_HEADER_SIZE));
2129 if (parse_code == DIRAC_PCODE_SEQ_HEADER) {
2130 if (s->seen_sequence_header)
2133 /* [DIRAC_STD] 10. Sequence header */
2134 ret = av_dirac_parse_sequence_header(&dsh, buf + DATA_UNIT_HEADER_SIZE, size - DATA_UNIT_HEADER_SIZE, avctx);
2136 av_log(avctx, AV_LOG_ERROR, "error parsing sequence header");
2140 if (CALC_PADDING((int64_t)dsh->width, MAX_DWT_LEVELS) * CALC_PADDING((int64_t)dsh->height, MAX_DWT_LEVELS) > avctx->max_pixels)
2141 ret = AVERROR(ERANGE);
2143 ret = ff_set_dimensions(avctx, dsh->width, dsh->height);
2149 ff_set_sar(avctx, dsh->sample_aspect_ratio);
2150 avctx->pix_fmt = dsh->pix_fmt;
2151 avctx->color_range = dsh->color_range;
2152 avctx->color_trc = dsh->color_trc;
2153 avctx->color_primaries = dsh->color_primaries;
2154 avctx->colorspace = dsh->colorspace;
2155 avctx->profile = dsh->profile;
2156 avctx->level = dsh->level;
2157 avctx->framerate = dsh->framerate;
2158 s->bit_depth = dsh->bit_depth;
2159 s->version.major = dsh->version.major;
2160 s->version.minor = dsh->version.minor;
2164 s->pshift = s->bit_depth > 8;
2166 ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt,
2168 &s->chroma_y_shift);
2172 ret = alloc_sequence_buffers(s);
2176 s->seen_sequence_header = 1;
2177 } else if (parse_code == DIRAC_PCODE_END_SEQ) { /* [DIRAC_STD] End of Sequence */
2178 free_sequence_buffers(s);
2179 s->seen_sequence_header = 0;
2180 } else if (parse_code == DIRAC_PCODE_AUX) {
2181 if (buf[13] == 1) { /* encoder implementation/version */
2183 /* versions older than 1.0.8 don't store quant delta for
2184 subbands with only one codeblock */
2185 if (sscanf(buf+14, "Schroedinger %d.%d.%d", ver, ver+1, ver+2) == 3)
2186 if (ver[0] == 1 && ver[1] == 0 && ver[2] <= 7)
2187 s->old_delta_quant = 1;
2189 } else if (parse_code & 0x8) { /* picture data unit */
2190 if (!s->seen_sequence_header) {
2191 av_log(avctx, AV_LOG_DEBUG, "Dropping frame without sequence header\n");
2192 return AVERROR_INVALIDDATA;
2195 /* find an unused frame */
2196 for (i = 0; i < MAX_FRAMES; i++)
2197 if (s->all_frames[i].avframe->data[0] == NULL)
2198 pic = &s->all_frames[i];
2200 av_log(avctx, AV_LOG_ERROR, "framelist full\n");
2201 return AVERROR_INVALIDDATA;
2204 av_frame_unref(pic->avframe);
2206 /* [DIRAC_STD] Defined in 9.6.1 ... */
2207 tmp = parse_code & 0x03; /* [DIRAC_STD] num_refs() */
2209 av_log(avctx, AV_LOG_ERROR, "num_refs of 3\n");
2210 return AVERROR_INVALIDDATA;
2213 s->is_arith = (parse_code & 0x48) == 0x08; /* [DIRAC_STD] using_ac() */
2214 s->low_delay = (parse_code & 0x88) == 0x88; /* [DIRAC_STD] is_low_delay() */
2215 s->core_syntax = (parse_code & 0x88) == 0x08; /* [DIRAC_STD] is_core_syntax() */
2216 s->ld_picture = (parse_code & 0xF8) == 0xC8; /* [DIRAC_STD] is_ld_picture() */
2217 s->hq_picture = (parse_code & 0xF8) == 0xE8; /* [DIRAC_STD] is_hq_picture() */
2218 s->dc_prediction = (parse_code & 0x28) == 0x08; /* [DIRAC_STD] using_dc_prediction() */
2219 pic->reference = (parse_code & 0x0C) == 0x0C; /* [DIRAC_STD] is_reference() */
2220 pic->avframe->key_frame = s->num_refs == 0; /* [DIRAC_STD] is_intra() */
2221 pic->avframe->pict_type = s->num_refs + 1; /* Definition of AVPictureType in avutil.h */
2223 /* VC-2 Low Delay has a different parse code than the Dirac Low Delay */
2224 if (s->version.minor == 2 && parse_code == 0x88)
2227 if (s->low_delay && !(s->ld_picture || s->hq_picture) ) {
2228 av_log(avctx, AV_LOG_ERROR, "Invalid low delay flag\n");
2229 return AVERROR_INVALIDDATA;
2232 if ((ret = get_buffer_with_edge(avctx, pic->avframe, (parse_code & 0x0C) == 0x0C ? AV_GET_BUFFER_FLAG_REF : 0)) < 0)
2234 s->current_picture = pic;
2235 s->plane[0].stride = pic->avframe->linesize[0];
2236 s->plane[1].stride = pic->avframe->linesize[1];
2237 s->plane[2].stride = pic->avframe->linesize[2];
2239 if (alloc_buffers(s, FFMAX3(FFABS(s->plane[0].stride), FFABS(s->plane[1].stride), FFABS(s->plane[2].stride))) < 0)
2240 return AVERROR(ENOMEM);
2242 /* [DIRAC_STD] 11.1 Picture parse. picture_parse() */
2243 ret = dirac_decode_picture_header(s);
2247 /* [DIRAC_STD] 13.0 Transform data syntax. transform_data() */
2248 ret = dirac_decode_frame_internal(s);
2255 static int dirac_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *pkt)
2257 DiracContext *s = avctx->priv_data;
2258 AVFrame *picture = data;
2259 uint8_t *buf = pkt->data;
2260 int buf_size = pkt->size;
2263 unsigned data_unit_size;
2265 /* release unused frames */
2266 for (i = 0; i < MAX_FRAMES; i++)
2267 if (s->all_frames[i].avframe->data[0] && !s->all_frames[i].reference) {
2268 av_frame_unref(s->all_frames[i].avframe);
2269 memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));
2272 s->current_picture = NULL;
2275 /* end of stream, so flush delayed pics */
2277 return get_delayed_pic(s, (AVFrame *)data, got_frame);
2280 /*[DIRAC_STD] Here starts the code from parse_info() defined in 9.6
2281 [DIRAC_STD] PARSE_INFO_PREFIX = "BBCD" as defined in ISO/IEC 646
2282 BBCD start code search */
2283 for (; buf_idx + DATA_UNIT_HEADER_SIZE < buf_size; buf_idx++) {
2284 if (buf[buf_idx ] == 'B' && buf[buf_idx+1] == 'B' &&
2285 buf[buf_idx+2] == 'C' && buf[buf_idx+3] == 'D')
2288 /* BBCD found or end of data */
2289 if (buf_idx + DATA_UNIT_HEADER_SIZE >= buf_size)
2292 data_unit_size = AV_RB32(buf+buf_idx+5);
2293 if (data_unit_size > buf_size - buf_idx || !data_unit_size) {
2294 if(data_unit_size > buf_size - buf_idx)
2295 av_log(s->avctx, AV_LOG_ERROR,
2296 "Data unit with size %d is larger than input buffer, discarding\n",
2301 /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3 inside the function parse_sequence() */
2302 ret = dirac_decode_data_unit(avctx, buf+buf_idx, data_unit_size);
2305 av_log(s->avctx, AV_LOG_ERROR,"Error in dirac_decode_data_unit\n");
2308 buf_idx += data_unit_size;
2311 if (!s->current_picture)
2314 if (s->current_picture->avframe->display_picture_number > s->frame_number) {
2315 DiracFrame *delayed_frame = remove_frame(s->delay_frames, s->frame_number);
2317 s->current_picture->reference |= DELAYED_PIC_REF;
2319 if (add_frame(s->delay_frames, MAX_DELAY, s->current_picture)) {
2320 int min_num = s->delay_frames[0]->avframe->display_picture_number;
2321 /* Too many delayed frames, so we display the frame with the lowest pts */
2322 av_log(avctx, AV_LOG_ERROR, "Delay frame overflow\n");
2324 for (i = 1; s->delay_frames[i]; i++)
2325 if (s->delay_frames[i]->avframe->display_picture_number < min_num)
2326 min_num = s->delay_frames[i]->avframe->display_picture_number;
2328 delayed_frame = remove_frame(s->delay_frames, min_num);
2329 add_frame(s->delay_frames, MAX_DELAY, s->current_picture);
2332 if (delayed_frame) {
2333 delayed_frame->reference ^= DELAYED_PIC_REF;
2334 if((ret=av_frame_ref(data, delayed_frame->avframe)) < 0)
2338 } else if (s->current_picture->avframe->display_picture_number == s->frame_number) {
2339 /* The right frame at the right time :-) */
2340 if((ret=av_frame_ref(data, s->current_picture->avframe)) < 0)
2346 s->frame_number = picture->display_picture_number + 1LL;
2351 AVCodec ff_dirac_decoder = {
2353 .long_name = NULL_IF_CONFIG_SMALL("BBC Dirac VC-2"),
2354 .type = AVMEDIA_TYPE_VIDEO,
2355 .id = AV_CODEC_ID_DIRAC,
2356 .priv_data_size = sizeof(DiracContext),
2357 .init = dirac_decode_init,
2358 .close = dirac_decode_end,
2359 .decode = dirac_decode_frame,
2360 .capabilities = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_DR1,
2361 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE,
2362 .flush = dirac_decode_flush,