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>
31 #include "bytestream.h"
34 #include "dirac_arith.h"
35 #include "mpeg12data.h"
36 #include "libavcodec/mpegvideo.h"
37 #include "mpegvideoencdsp.h"
38 #include "dirac_dwt.h"
44 * The spec limits the number of wavelet decompositions to 4 for both
45 * level 1 (VC-2) and 128 (long-gop default).
46 * 5 decompositions is the maximum before >16-bit buffers are needed.
47 * Schroedinger allows this for DD 9,7 and 13,7 wavelets only, limiting
48 * the others to 4 decompositions (or 3 for the fidelity filter).
50 * We use this instead of MAX_DECOMPOSITIONS to save some memory.
52 #define MAX_DWT_LEVELS 5
55 * The spec limits this to 3 for frame coding, but in practice can be as high as 6
57 #define MAX_REFERENCE_FRAMES 8
58 #define MAX_DELAY 5 /* limit for main profile for frame coding (TODO: field coding) */
59 #define MAX_FRAMES (MAX_REFERENCE_FRAMES + MAX_DELAY + 1)
60 #define MAX_QUANT 68 /* max quant for VC-2 */
61 #define MAX_BLOCKSIZE 32 /* maximum xblen/yblen we support */
64 * DiracBlock->ref flags, if set then the block does MC from the given ref
66 #define DIRAC_REF_MASK_REF1 1
67 #define DIRAC_REF_MASK_REF2 2
68 #define DIRAC_REF_MASK_GLOBAL 4
71 * Value of Picture.reference when Picture is not a reference picture, but
72 * is held for delayed output.
74 #define DELAYED_PIC_REF 4
76 #define CALC_PADDING(size, depth) \
77 (((size + (1 << depth) - 1) >> depth) << depth)
79 #define DIVRNDUP(a, b) (((a) + (b) - 1) / (b))
83 int interpolated[3]; /* 1 if hpel[] is valid */
85 uint8_t *hpel_base[3][4];
92 } u; /* anonymous unions aren't in C99 :( */
96 typedef struct SubBand {
104 struct SubBand *parent;
108 const uint8_t *coeff_data;
111 typedef struct Plane {
120 IDWTELEM *idwt_buf_base;
126 /* block separation (block n+1 starts after this many pixels in block n) */
129 /* amount of overspill on each edge (half of the overlap between blocks) */
133 SubBand band[MAX_DWT_LEVELS][4];
136 typedef struct DiracContext {
137 AVCodecContext *avctx;
138 MpegvideoEncDSPContext mpvencdsp;
139 VideoDSPContext vdsp;
140 DiracDSPContext diracdsp;
142 dirac_source_params source;
143 int seen_sequence_header;
144 int frame_number; /* number of the next frame to display */
149 int zero_res; /* zero residue flag */
150 int is_arith; /* whether coeffs use arith or golomb coding */
151 int low_delay; /* use the low delay syntax */
152 int globalmc_flag; /* use global motion compensation */
153 int num_refs; /* number of reference pictures */
155 /* wavelet decoding */
156 unsigned wavelet_depth; /* depth of the IDWT */
157 unsigned wavelet_idx;
160 * schroedinger older than 1.0.8 doesn't store
161 * quant delta if only one codebook exists in a band
163 unsigned old_delta_quant;
164 unsigned codeblock_mode;
169 } codeblock[MAX_DWT_LEVELS+1];
172 unsigned num_x; /* number of horizontal slices */
173 unsigned num_y; /* number of vertical slices */
174 AVRational bytes; /* average bytes per slice */
175 uint8_t quant[MAX_DWT_LEVELS][4]; /* [DIRAC_STD] E.1 */
179 int pan_tilt[2]; /* pan/tilt vector */
180 int zrs[2][2]; /* zoom/rotate/shear matrix */
181 int perspective[2]; /* perspective vector */
183 unsigned perspective_exp;
186 /* motion compensation */
187 uint8_t mv_precision; /* [DIRAC_STD] REFS_WT_PRECISION */
188 int16_t weight[2]; /* [DIRAC_STD] REF1_WT and REF2_WT */
189 unsigned weight_log2denom; /* [DIRAC_STD] REFS_WT_PRECISION */
191 int blwidth; /* number of blocks (horizontally) */
192 int blheight; /* number of blocks (vertically) */
193 int sbwidth; /* number of superblocks (horizontally) */
194 int sbheight; /* number of superblocks (vertically) */
197 DiracBlock *blmotion;
199 uint8_t *edge_emu_buffer[4];
200 uint8_t *edge_emu_buffer_base;
202 uint16_t *mctmp; /* buffer holding the MC data multiplied by OBMC weights */
206 DECLARE_ALIGNED(16, uint8_t, obmc_weight)[3][MAX_BLOCKSIZE*MAX_BLOCKSIZE];
208 void (*put_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
209 void (*avg_pixels_tab[4])(uint8_t *dst, const uint8_t *src[5], int stride, int h);
210 void (*add_obmc)(uint16_t *dst, const uint8_t *src, int stride, const uint8_t *obmc_weight, int yblen);
211 dirac_weight_func weight_func;
212 dirac_biweight_func biweight_func;
214 DiracFrame *current_picture;
215 DiracFrame *ref_pics[2];
217 DiracFrame *ref_frames[MAX_REFERENCE_FRAMES+1];
218 DiracFrame *delay_frames[MAX_DELAY+1];
219 DiracFrame all_frames[MAX_FRAMES];
223 * Dirac Specification ->
224 * Parse code values. 9.6.1 Table 9.1
226 enum dirac_parse_code {
227 pc_seq_header = 0x00,
241 static const uint8_t default_qmat[][4][4] = {
242 { { 5, 3, 3, 0}, { 0, 4, 4, 1}, { 0, 5, 5, 2}, { 0, 6, 6, 3} },
243 { { 4, 2, 2, 0}, { 0, 4, 4, 2}, { 0, 5, 5, 3}, { 0, 7, 7, 5} },
244 { { 5, 3, 3, 0}, { 0, 4, 4, 1}, { 0, 5, 5, 2}, { 0, 6, 6, 3} },
245 { { 8, 4, 4, 0}, { 0, 4, 4, 0}, { 0, 4, 4, 0}, { 0, 4, 4, 0} },
246 { { 8, 4, 4, 0}, { 0, 4, 4, 0}, { 0, 4, 4, 0}, { 0, 4, 4, 0} },
247 { { 0, 4, 4, 8}, { 0, 8, 8, 12}, { 0, 13, 13, 17}, { 0, 17, 17, 21} },
248 { { 3, 1, 1, 0}, { 0, 4, 4, 2}, { 0, 6, 6, 5}, { 0, 9, 9, 7} },
251 static const int qscale_tab[MAX_QUANT+1] = {
252 4, 5, 6, 7, 8, 10, 11, 13,
253 16, 19, 23, 27, 32, 38, 45, 54,
254 64, 76, 91, 108, 128, 152, 181, 215,
255 256, 304, 362, 431, 512, 609, 724, 861,
256 1024, 1218, 1448, 1722, 2048, 2435, 2896, 3444,
257 4096, 4871, 5793, 6889, 8192, 9742, 11585, 13777,
258 16384, 19484, 23170, 27554, 32768, 38968, 46341, 55109,
262 static const int qoffset_intra_tab[MAX_QUANT+1] = {
263 1, 2, 3, 4, 4, 5, 6, 7,
264 8, 10, 12, 14, 16, 19, 23, 27,
265 32, 38, 46, 54, 64, 76, 91, 108,
266 128, 152, 181, 216, 256, 305, 362, 431,
267 512, 609, 724, 861, 1024, 1218, 1448, 1722,
268 2048, 2436, 2897, 3445, 4096, 4871, 5793, 6889,
269 8192, 9742, 11585, 13777, 16384, 19484, 23171, 27555,
273 static const int qoffset_inter_tab[MAX_QUANT+1] = {
274 1, 2, 2, 3, 3, 4, 4, 5,
275 6, 7, 9, 10, 12, 14, 17, 20,
276 24, 29, 34, 41, 48, 57, 68, 81,
277 96, 114, 136, 162, 192, 228, 272, 323,
278 384, 457, 543, 646, 768, 913, 1086, 1292,
279 1536, 1827, 2172, 2583, 3072, 3653, 4344, 5166,
280 6144, 7307, 8689, 10333, 12288, 14613, 17378, 20666,
284 /* magic number division by 3 from schroedinger */
285 static inline int divide3(int x)
287 return ((x+1)*21845 + 10922) >> 16;
290 static DiracFrame *remove_frame(DiracFrame *framelist[], int picnum)
292 DiracFrame *remove_pic = NULL;
293 int i, remove_idx = -1;
295 for (i = 0; framelist[i]; i++)
296 if (framelist[i]->avframe->display_picture_number == picnum) {
297 remove_pic = framelist[i];
302 for (i = remove_idx; framelist[i]; i++)
303 framelist[i] = framelist[i+1];
308 static int add_frame(DiracFrame *framelist[], int maxframes, DiracFrame *frame)
311 for (i = 0; i < maxframes; i++)
313 framelist[i] = frame;
319 static int alloc_sequence_buffers(DiracContext *s)
321 int sbwidth = DIVRNDUP(s->source.width, 4);
322 int sbheight = DIVRNDUP(s->source.height, 4);
323 int i, w, h, top_padding;
325 /* todo: think more about this / use or set Plane here */
326 for (i = 0; i < 3; i++) {
327 int max_xblen = MAX_BLOCKSIZE >> (i ? s->chroma_x_shift : 0);
328 int max_yblen = MAX_BLOCKSIZE >> (i ? s->chroma_y_shift : 0);
329 w = s->source.width >> (i ? s->chroma_x_shift : 0);
330 h = s->source.height >> (i ? s->chroma_y_shift : 0);
332 /* we allocate the max we support here since num decompositions can
333 * change from frame to frame. Stride is aligned to 16 for SIMD, and
334 * 1<<MAX_DWT_LEVELS top padding to avoid if(y>0) in arith decoding
335 * MAX_BLOCKSIZE padding for MC: blocks can spill up to half of that
337 top_padding = FFMAX(1<<MAX_DWT_LEVELS, max_yblen/2);
338 w = FFALIGN(CALC_PADDING(w, MAX_DWT_LEVELS), 8); /* FIXME: Should this be 16 for SSE??? */
339 h = top_padding + CALC_PADDING(h, MAX_DWT_LEVELS) + max_yblen/2;
341 s->plane[i].idwt_buf_base = av_mallocz_array((w+max_xblen), h * sizeof(IDWTELEM));
342 s->plane[i].idwt_tmp = av_malloc_array((w+16), sizeof(IDWTELEM));
343 s->plane[i].idwt_buf = s->plane[i].idwt_buf_base + top_padding*w;
344 if (!s->plane[i].idwt_buf_base || !s->plane[i].idwt_tmp)
345 return AVERROR(ENOMEM);
348 /* fixme: allocate using real stride here */
349 s->sbsplit = av_malloc_array(sbwidth, sbheight);
350 s->blmotion = av_malloc_array(sbwidth, sbheight * 16 * sizeof(*s->blmotion));
352 if (!s->sbsplit || !s->blmotion)
353 return AVERROR(ENOMEM);
357 static int alloc_buffers(DiracContext *s, int stride)
359 int w = s->source.width;
360 int h = s->source.height;
362 av_assert0(stride >= w);
365 if (s->buffer_stride >= stride)
367 s->buffer_stride = 0;
369 av_freep(&s->edge_emu_buffer_base);
370 memset(s->edge_emu_buffer, 0, sizeof(s->edge_emu_buffer));
372 av_freep(&s->mcscratch);
374 s->edge_emu_buffer_base = av_malloc_array(stride, MAX_BLOCKSIZE);
376 s->mctmp = av_malloc_array((stride+MAX_BLOCKSIZE), (h+MAX_BLOCKSIZE) * sizeof(*s->mctmp));
377 s->mcscratch = av_malloc_array(stride, MAX_BLOCKSIZE);
379 if (!s->edge_emu_buffer_base || !s->mctmp || !s->mcscratch)
380 return AVERROR(ENOMEM);
382 s->buffer_stride = stride;
386 static void free_sequence_buffers(DiracContext *s)
390 for (i = 0; i < MAX_FRAMES; i++) {
391 if (s->all_frames[i].avframe->data[0]) {
392 av_frame_unref(s->all_frames[i].avframe);
393 memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));
396 for (j = 0; j < 3; j++)
397 for (k = 1; k < 4; k++)
398 av_freep(&s->all_frames[i].hpel_base[j][k]);
401 memset(s->ref_frames, 0, sizeof(s->ref_frames));
402 memset(s->delay_frames, 0, sizeof(s->delay_frames));
404 for (i = 0; i < 3; i++) {
405 av_freep(&s->plane[i].idwt_buf_base);
406 av_freep(&s->plane[i].idwt_tmp);
409 s->buffer_stride = 0;
410 av_freep(&s->sbsplit);
411 av_freep(&s->blmotion);
412 av_freep(&s->edge_emu_buffer_base);
415 av_freep(&s->mcscratch);
418 static av_cold int dirac_decode_init(AVCodecContext *avctx)
420 DiracContext *s = avctx->priv_data;
424 s->frame_number = -1;
426 ff_diracdsp_init(&s->diracdsp);
427 ff_mpegvideoencdsp_init(&s->mpvencdsp, avctx);
428 ff_videodsp_init(&s->vdsp, 8);
430 for (i = 0; i < MAX_FRAMES; i++) {
431 s->all_frames[i].avframe = av_frame_alloc();
432 if (!s->all_frames[i].avframe) {
434 av_frame_free(&s->all_frames[--i].avframe);
435 return AVERROR(ENOMEM);
442 static void dirac_decode_flush(AVCodecContext *avctx)
444 DiracContext *s = avctx->priv_data;
445 free_sequence_buffers(s);
446 s->seen_sequence_header = 0;
447 s->frame_number = -1;
450 static av_cold int dirac_decode_end(AVCodecContext *avctx)
452 DiracContext *s = avctx->priv_data;
455 dirac_decode_flush(avctx);
456 for (i = 0; i < MAX_FRAMES; i++)
457 av_frame_free(&s->all_frames[i].avframe);
462 #define SIGN_CTX(x) (CTX_SIGN_ZERO + ((x) > 0) - ((x) < 0))
464 static inline void coeff_unpack_arith(DiracArith *c, int qfactor, int qoffset,
465 SubBand *b, IDWTELEM *buf, int x, int y)
469 int pred_ctx = CTX_ZPZN_F1;
471 /* Check if the parent subband has a 0 in the corresponding position */
473 pred_ctx += !!b->parent->ibuf[b->parent->stride * (y>>1) + (x>>1)] << 1;
475 if (b->orientation == subband_hl)
476 sign_pred = buf[-b->stride];
478 /* Determine if the pixel has only zeros in its neighbourhood */
480 pred_ctx += !(buf[-1] | buf[-b->stride] | buf[-1-b->stride]);
481 if (b->orientation == subband_lh)
484 pred_ctx += !buf[-b->stride];
487 coeff = dirac_get_arith_uint(c, pred_ctx, CTX_COEFF_DATA);
489 coeff = (coeff * qfactor + qoffset + 2) >> 2;
490 sign = dirac_get_arith_bit(c, SIGN_CTX(sign_pred));
491 coeff = (coeff ^ -sign) + sign;
496 static inline int coeff_unpack_golomb(GetBitContext *gb, int qfactor, int qoffset)
500 coeff = svq3_get_ue_golomb(gb);
502 coeff = (coeff * qfactor + qoffset + 2) >> 2;
503 sign = get_bits1(gb);
504 coeff = (coeff ^ -sign) + sign;
510 * Decode the coeffs in the rectangle defined by left, right, top, bottom
511 * [DIRAC_STD] 13.4.3.2 Codeblock unpacking loop. codeblock()
513 static inline void codeblock(DiracContext *s, SubBand *b,
514 GetBitContext *gb, DiracArith *c,
515 int left, int right, int top, int bottom,
516 int blockcnt_one, int is_arith)
518 int x, y, zero_block;
519 int qoffset, qfactor;
522 /* check for any coded coefficients in this codeblock */
525 zero_block = dirac_get_arith_bit(c, CTX_ZERO_BLOCK);
527 zero_block = get_bits1(gb);
533 if (s->codeblock_mode && !(s->old_delta_quant && blockcnt_one)) {
534 int quant = b->quant;
536 quant += dirac_get_arith_int(c, CTX_DELTA_Q_F, CTX_DELTA_Q_DATA);
538 quant += dirac_get_se_golomb(gb);
540 av_log(s->avctx, AV_LOG_ERROR, "Invalid quant\n");
546 b->quant = FFMIN(b->quant, MAX_QUANT);
548 qfactor = qscale_tab[b->quant];
549 /* TODO: context pointer? */
551 qoffset = qoffset_intra_tab[b->quant];
553 qoffset = qoffset_inter_tab[b->quant];
555 buf = b->ibuf + top * b->stride;
556 for (y = top; y < bottom; y++) {
557 for (x = left; x < right; x++) {
558 /* [DIRAC_STD] 13.4.4 Subband coefficients. coeff_unpack() */
560 coeff_unpack_arith(c, qfactor, qoffset, b, buf+x, x, y);
562 buf[x] = coeff_unpack_golomb(gb, qfactor, qoffset);
569 * Dirac Specification ->
570 * 13.3 intra_dc_prediction(band)
572 static inline void intra_dc_prediction(SubBand *b)
574 IDWTELEM *buf = b->ibuf;
577 for (x = 1; x < b->width; x++)
581 for (y = 1; y < b->height; y++) {
582 buf[0] += buf[-b->stride];
584 for (x = 1; x < b->width; x++) {
585 int pred = buf[x - 1] + buf[x - b->stride] + buf[x - b->stride-1];
586 buf[x] += divide3(pred);
593 * Dirac Specification ->
594 * 13.4.2 Non-skipped subbands. subband_coeffs()
596 static av_always_inline void decode_subband_internal(DiracContext *s, SubBand *b, int is_arith)
598 int cb_x, cb_y, left, right, top, bottom;
601 int cb_width = s->codeblock[b->level + (b->orientation != subband_ll)].width;
602 int cb_height = s->codeblock[b->level + (b->orientation != subband_ll)].height;
603 int blockcnt_one = (cb_width + cb_height) == 2;
608 init_get_bits8(&gb, b->coeff_data, b->length);
611 ff_dirac_init_arith_decoder(&c, &gb, b->length);
614 for (cb_y = 0; cb_y < cb_height; cb_y++) {
615 bottom = (b->height * (cb_y+1LL)) / cb_height;
617 for (cb_x = 0; cb_x < cb_width; cb_x++) {
618 right = (b->width * (cb_x+1LL)) / cb_width;
619 codeblock(s, b, &gb, &c, left, right, top, bottom, blockcnt_one, is_arith);
625 if (b->orientation == subband_ll && s->num_refs == 0)
626 intra_dc_prediction(b);
629 static int decode_subband_arith(AVCodecContext *avctx, void *b)
631 DiracContext *s = avctx->priv_data;
632 decode_subband_internal(s, b, 1);
636 static int decode_subband_golomb(AVCodecContext *avctx, void *arg)
638 DiracContext *s = avctx->priv_data;
640 decode_subband_internal(s, *b, 0);
645 * Dirac Specification ->
646 * [DIRAC_STD] 13.4.1 core_transform_data()
648 static void decode_component(DiracContext *s, int comp)
650 AVCodecContext *avctx = s->avctx;
651 SubBand *bands[3*MAX_DWT_LEVELS+1];
652 enum dirac_subband orientation;
653 int level, num_bands = 0;
655 /* Unpack all subbands at all levels. */
656 for (level = 0; level < s->wavelet_depth; level++) {
657 for (orientation = !!level; orientation < 4; orientation++) {
658 SubBand *b = &s->plane[comp].band[level][orientation];
659 bands[num_bands++] = b;
661 align_get_bits(&s->gb);
662 /* [DIRAC_STD] 13.4.2 subband() */
663 b->length = svq3_get_ue_golomb(&s->gb);
665 b->quant = svq3_get_ue_golomb(&s->gb);
666 align_get_bits(&s->gb);
667 b->coeff_data = s->gb.buffer + get_bits_count(&s->gb)/8;
668 b->length = FFMIN(b->length, FFMAX(get_bits_left(&s->gb)/8, 0));
669 skip_bits_long(&s->gb, b->length*8);
672 /* arithmetic coding has inter-level dependencies, so we can only execute one level at a time */
674 avctx->execute(avctx, decode_subband_arith, &s->plane[comp].band[level][!!level],
675 NULL, 4-!!level, sizeof(SubBand));
677 /* golomb coding has no inter-level dependencies, so we can execute all subbands in parallel */
679 avctx->execute(avctx, decode_subband_golomb, bands, NULL, num_bands, sizeof(SubBand*));
682 /* [DIRAC_STD] 13.5.5.2 Luma slice subband data. luma_slice_band(level,orient,sx,sy) --> if b2 == NULL */
683 /* [DIRAC_STD] 13.5.5.3 Chroma slice subband data. chroma_slice_band(level,orient,sx,sy) --> if b2 != NULL */
684 static void lowdelay_subband(DiracContext *s, GetBitContext *gb, int quant,
685 int slice_x, int slice_y, int bits_end,
686 SubBand *b1, SubBand *b2)
688 int left = b1->width * slice_x / s->lowdelay.num_x;
689 int right = b1->width *(slice_x+1) / s->lowdelay.num_x;
690 int top = b1->height * slice_y / s->lowdelay.num_y;
691 int bottom = b1->height *(slice_y+1) / s->lowdelay.num_y;
693 int qfactor = qscale_tab[FFMIN(quant, MAX_QUANT)];
694 int qoffset = qoffset_intra_tab[FFMIN(quant, MAX_QUANT)];
696 IDWTELEM *buf1 = b1->ibuf + top * b1->stride;
697 IDWTELEM *buf2 = b2 ? b2->ibuf + top * b2->stride : NULL;
699 /* we have to constantly check for overread since the spec explicitly
700 requires this, with the meaning that all remaining coeffs are set to 0 */
701 if (get_bits_count(gb) >= bits_end)
704 for (y = top; y < bottom; y++) {
705 for (x = left; x < right; x++) {
706 buf1[x] = coeff_unpack_golomb(gb, qfactor, qoffset);
707 if (get_bits_count(gb) >= bits_end)
710 buf2[x] = coeff_unpack_golomb(gb, qfactor, qoffset);
711 if (get_bits_count(gb) >= bits_end)
721 struct lowdelay_slice {
730 * Dirac Specification ->
731 * 13.5.2 Slices. slice(sx,sy)
733 static int decode_lowdelay_slice(AVCodecContext *avctx, void *arg)
735 DiracContext *s = avctx->priv_data;
736 struct lowdelay_slice *slice = arg;
737 GetBitContext *gb = &slice->gb;
738 enum dirac_subband orientation;
739 int level, quant, chroma_bits, chroma_end;
741 int quant_base = get_bits(gb, 7); /*[DIRAC_STD] qindex */
742 int length_bits = av_log2(8 * slice->bytes)+1;
743 int luma_bits = get_bits_long(gb, length_bits);
744 int luma_end = get_bits_count(gb) + FFMIN(luma_bits, get_bits_left(gb));
746 /* [DIRAC_STD] 13.5.5.2 luma_slice_band */
747 for (level = 0; level < s->wavelet_depth; level++)
748 for (orientation = !!level; orientation < 4; orientation++) {
749 quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);
750 lowdelay_subband(s, gb, quant, slice->slice_x, slice->slice_y, luma_end,
751 &s->plane[0].band[level][orientation], NULL);
754 /* consume any unused bits from luma */
755 skip_bits_long(gb, get_bits_count(gb) - luma_end);
757 chroma_bits = 8*slice->bytes - 7 - length_bits - luma_bits;
758 chroma_end = get_bits_count(gb) + FFMIN(chroma_bits, get_bits_left(gb));
759 /* [DIRAC_STD] 13.5.5.3 chroma_slice_band */
760 for (level = 0; level < s->wavelet_depth; level++)
761 for (orientation = !!level; orientation < 4; orientation++) {
762 quant = FFMAX(quant_base - s->lowdelay.quant[level][orientation], 0);
763 lowdelay_subband(s, gb, quant, slice->slice_x, slice->slice_y, chroma_end,
764 &s->plane[1].band[level][orientation],
765 &s->plane[2].band[level][orientation]);
772 * Dirac Specification ->
773 * 13.5.1 low_delay_transform_data()
775 static int decode_lowdelay(DiracContext *s)
777 AVCodecContext *avctx = s->avctx;
778 int slice_x, slice_y, bytes, bufsize;
780 struct lowdelay_slice *slices;
783 slices = av_mallocz_array(s->lowdelay.num_x, s->lowdelay.num_y * sizeof(struct lowdelay_slice));
785 return AVERROR(ENOMEM);
787 align_get_bits(&s->gb);
788 /*[DIRAC_STD] 13.5.2 Slices. slice(sx,sy) */
789 buf = s->gb.buffer + get_bits_count(&s->gb)/8;
790 bufsize = get_bits_left(&s->gb);
792 for (slice_y = 0; bufsize > 0 && slice_y < s->lowdelay.num_y; slice_y++)
793 for (slice_x = 0; bufsize > 0 && slice_x < s->lowdelay.num_x; slice_x++) {
794 bytes = (slice_num+1) * s->lowdelay.bytes.num / s->lowdelay.bytes.den
795 - slice_num * s->lowdelay.bytes.num / s->lowdelay.bytes.den;
797 slices[slice_num].bytes = bytes;
798 slices[slice_num].slice_x = slice_x;
799 slices[slice_num].slice_y = slice_y;
800 init_get_bits(&slices[slice_num].gb, buf, bufsize);
807 avctx->execute(avctx, decode_lowdelay_slice, slices, NULL, slice_num,
808 sizeof(struct lowdelay_slice)); /* [DIRAC_STD] 13.5.2 Slices */
809 intra_dc_prediction(&s->plane[0].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
810 intra_dc_prediction(&s->plane[1].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
811 intra_dc_prediction(&s->plane[2].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
816 static void init_planes(DiracContext *s)
818 int i, w, h, level, orientation;
820 for (i = 0; i < 3; i++) {
821 Plane *p = &s->plane[i];
823 p->width = s->source.width >> (i ? s->chroma_x_shift : 0);
824 p->height = s->source.height >> (i ? s->chroma_y_shift : 0);
825 p->idwt_width = w = CALC_PADDING(p->width , s->wavelet_depth);
826 p->idwt_height = h = CALC_PADDING(p->height, s->wavelet_depth);
827 p->idwt_stride = FFALIGN(p->idwt_width, 8);
829 for (level = s->wavelet_depth-1; level >= 0; level--) {
832 for (orientation = !!level; orientation < 4; orientation++) {
833 SubBand *b = &p->band[level][orientation];
835 b->ibuf = p->idwt_buf;
837 b->stride = p->idwt_stride << (s->wavelet_depth - level);
840 b->orientation = orientation;
845 b->ibuf += b->stride>>1;
848 b->parent = &p->band[level-1][orientation];
853 p->xblen = s->plane[0].xblen >> s->chroma_x_shift;
854 p->yblen = s->plane[0].yblen >> s->chroma_y_shift;
855 p->xbsep = s->plane[0].xbsep >> s->chroma_x_shift;
856 p->ybsep = s->plane[0].ybsep >> s->chroma_y_shift;
859 p->xoffset = (p->xblen - p->xbsep)/2;
860 p->yoffset = (p->yblen - p->ybsep)/2;
865 * Unpack the motion compensation parameters
866 * Dirac Specification ->
867 * 11.2 Picture prediction data. picture_prediction()
869 static int dirac_unpack_prediction_parameters(DiracContext *s)
871 static const uint8_t default_blen[] = { 4, 12, 16, 24 };
872 static const uint8_t default_bsep[] = { 4, 8, 12, 16 };
874 GetBitContext *gb = &s->gb;
878 /* [DIRAC_STD] 11.2.2 Block parameters. block_parameters() */
879 /* Luma and Chroma are equal. 11.2.3 */
880 idx = svq3_get_ue_golomb(gb); /* [DIRAC_STD] index */
883 av_log(s->avctx, AV_LOG_ERROR, "Block prediction index too high\n");
888 s->plane[0].xblen = svq3_get_ue_golomb(gb);
889 s->plane[0].yblen = svq3_get_ue_golomb(gb);
890 s->plane[0].xbsep = svq3_get_ue_golomb(gb);
891 s->plane[0].ybsep = svq3_get_ue_golomb(gb);
893 /*[DIRAC_STD] preset_block_params(index). Table 11.1 */
894 s->plane[0].xblen = default_blen[idx-1];
895 s->plane[0].yblen = default_blen[idx-1];
896 s->plane[0].xbsep = default_bsep[idx-1];
897 s->plane[0].ybsep = default_bsep[idx-1];
899 /*[DIRAC_STD] 11.2.4 motion_data_dimensions()
900 Calculated in function dirac_unpack_block_motion_data */
902 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) {
903 av_log(s->avctx, AV_LOG_ERROR, "Block separation too small\n");
906 if (s->plane[0].xbsep > s->plane[0].xblen || s->plane[0].ybsep > s->plane[0].yblen) {
907 av_log(s->avctx, AV_LOG_ERROR, "Block separation greater than size\n");
910 if (FFMAX(s->plane[0].xblen, s->plane[0].yblen) > MAX_BLOCKSIZE) {
911 av_log(s->avctx, AV_LOG_ERROR, "Unsupported large block size\n");
915 /*[DIRAC_STD] 11.2.5 Motion vector precision. motion_vector_precision()
916 Read motion vector precision */
917 s->mv_precision = svq3_get_ue_golomb(gb);
918 if (s->mv_precision > 3) {
919 av_log(s->avctx, AV_LOG_ERROR, "MV precision finer than eighth-pel\n");
923 /*[DIRAC_STD] 11.2.6 Global motion. global_motion()
924 Read the global motion compensation parameters */
925 s->globalmc_flag = get_bits1(gb);
926 if (s->globalmc_flag) {
927 memset(s->globalmc, 0, sizeof(s->globalmc));
928 /* [DIRAC_STD] pan_tilt(gparams) */
929 for (ref = 0; ref < s->num_refs; ref++) {
931 s->globalmc[ref].pan_tilt[0] = dirac_get_se_golomb(gb);
932 s->globalmc[ref].pan_tilt[1] = dirac_get_se_golomb(gb);
934 /* [DIRAC_STD] zoom_rotate_shear(gparams)
935 zoom/rotation/shear parameters */
937 s->globalmc[ref].zrs_exp = svq3_get_ue_golomb(gb);
938 s->globalmc[ref].zrs[0][0] = dirac_get_se_golomb(gb);
939 s->globalmc[ref].zrs[0][1] = dirac_get_se_golomb(gb);
940 s->globalmc[ref].zrs[1][0] = dirac_get_se_golomb(gb);
941 s->globalmc[ref].zrs[1][1] = dirac_get_se_golomb(gb);
943 s->globalmc[ref].zrs[0][0] = 1;
944 s->globalmc[ref].zrs[1][1] = 1;
946 /* [DIRAC_STD] perspective(gparams) */
948 s->globalmc[ref].perspective_exp = svq3_get_ue_golomb(gb);
949 s->globalmc[ref].perspective[0] = dirac_get_se_golomb(gb);
950 s->globalmc[ref].perspective[1] = dirac_get_se_golomb(gb);
955 /*[DIRAC_STD] 11.2.7 Picture prediction mode. prediction_mode()
956 Picture prediction mode, not currently used. */
957 if (svq3_get_ue_golomb(gb)) {
958 av_log(s->avctx, AV_LOG_ERROR, "Unknown picture prediction mode\n");
962 /* [DIRAC_STD] 11.2.8 Reference picture weight. reference_picture_weights()
963 just data read, weight calculation will be done later on. */
964 s->weight_log2denom = 1;
969 s->weight_log2denom = svq3_get_ue_golomb(gb);
970 s->weight[0] = dirac_get_se_golomb(gb);
971 if (s->num_refs == 2)
972 s->weight[1] = dirac_get_se_golomb(gb);
978 * Dirac Specification ->
979 * 11.3 Wavelet transform data. wavelet_transform()
981 static int dirac_unpack_idwt_params(DiracContext *s)
983 GetBitContext *gb = &s->gb;
987 #define CHECKEDREAD(dst, cond, errmsg) \
988 tmp = svq3_get_ue_golomb(gb); \
990 av_log(s->avctx, AV_LOG_ERROR, errmsg); \
997 s->zero_res = s->num_refs ? get_bits1(gb) : 0;
1001 /*[DIRAC_STD] 11.3.1 Transform parameters. transform_parameters() */
1002 CHECKEDREAD(s->wavelet_idx, tmp > 6, "wavelet_idx is too big\n")
1004 CHECKEDREAD(s->wavelet_depth, tmp > MAX_DWT_LEVELS || tmp < 1, "invalid number of DWT decompositions\n")
1006 if (!s->low_delay) {
1007 /* Codeblock parameters (core syntax only) */
1008 if (get_bits1(gb)) {
1009 for (i = 0; i <= s->wavelet_depth; i++) {
1010 CHECKEDREAD(s->codeblock[i].width , tmp < 1 || tmp > (s->avctx->width >>s->wavelet_depth-i), "codeblock width invalid\n")
1011 CHECKEDREAD(s->codeblock[i].height, tmp < 1 || tmp > (s->avctx->height>>s->wavelet_depth-i), "codeblock height invalid\n")
1014 CHECKEDREAD(s->codeblock_mode, tmp > 1, "unknown codeblock mode\n")
1016 for (i = 0; i <= s->wavelet_depth; i++)
1017 s->codeblock[i].width = s->codeblock[i].height = 1;
1019 /* Slice parameters + quantization matrix*/
1020 /*[DIRAC_STD] 11.3.4 Slice coding Parameters (low delay syntax only). slice_parameters() */
1021 s->lowdelay.num_x = svq3_get_ue_golomb(gb);
1022 s->lowdelay.num_y = svq3_get_ue_golomb(gb);
1023 s->lowdelay.bytes.num = svq3_get_ue_golomb(gb);
1024 s->lowdelay.bytes.den = svq3_get_ue_golomb(gb);
1026 if (s->lowdelay.bytes.den <= 0) {
1027 av_log(s->avctx,AV_LOG_ERROR,"Invalid lowdelay.bytes.den\n");
1028 return AVERROR_INVALIDDATA;
1031 /* [DIRAC_STD] 11.3.5 Quantisation matrices (low-delay syntax). quant_matrix() */
1032 if (get_bits1(gb)) {
1033 av_log(s->avctx,AV_LOG_DEBUG,"Low Delay: Has Custom Quantization Matrix!\n");
1034 /* custom quantization matrix */
1035 s->lowdelay.quant[0][0] = svq3_get_ue_golomb(gb);
1036 for (level = 0; level < s->wavelet_depth; level++) {
1037 s->lowdelay.quant[level][1] = svq3_get_ue_golomb(gb);
1038 s->lowdelay.quant[level][2] = svq3_get_ue_golomb(gb);
1039 s->lowdelay.quant[level][3] = svq3_get_ue_golomb(gb);
1042 if (s->wavelet_depth > 4) {
1043 av_log(s->avctx,AV_LOG_ERROR,"Mandatory custom low delay matrix missing for depth %d\n", s->wavelet_depth);
1044 return AVERROR_INVALIDDATA;
1046 /* default quantization matrix */
1047 for (level = 0; level < s->wavelet_depth; level++)
1048 for (i = 0; i < 4; i++) {
1049 s->lowdelay.quant[level][i] = default_qmat[s->wavelet_idx][level][i];
1050 /* haar with no shift differs for different depths */
1051 if (s->wavelet_idx == 3)
1052 s->lowdelay.quant[level][i] += 4*(s->wavelet_depth-1 - level);
1059 static inline int pred_sbsplit(uint8_t *sbsplit, int stride, int x, int y)
1061 static const uint8_t avgsplit[7] = { 0, 0, 1, 1, 1, 2, 2 };
1068 return sbsplit[-stride];
1070 return avgsplit[sbsplit[-1] + sbsplit[-stride] + sbsplit[-stride-1]];
1073 static inline int pred_block_mode(DiracBlock *block, int stride, int x, int y, int refmask)
1080 return block[-1].ref & refmask;
1082 return block[-stride].ref & refmask;
1084 /* return the majority */
1085 pred = (block[-1].ref & refmask) + (block[-stride].ref & refmask) + (block[-stride-1].ref & refmask);
1086 return (pred >> 1) & refmask;
1089 static inline void pred_block_dc(DiracBlock *block, int stride, int x, int y)
1093 memset(block->u.dc, 0, sizeof(block->u.dc));
1095 if (x && !(block[-1].ref & 3)) {
1096 for (i = 0; i < 3; i++)
1097 block->u.dc[i] += block[-1].u.dc[i];
1101 if (y && !(block[-stride].ref & 3)) {
1102 for (i = 0; i < 3; i++)
1103 block->u.dc[i] += block[-stride].u.dc[i];
1107 if (x && y && !(block[-1-stride].ref & 3)) {
1108 for (i = 0; i < 3; i++)
1109 block->u.dc[i] += block[-1-stride].u.dc[i];
1114 for (i = 0; i < 3; i++)
1115 block->u.dc[i] = (block->u.dc[i]+1)>>1;
1116 } else if (n == 3) {
1117 for (i = 0; i < 3; i++)
1118 block->u.dc[i] = divide3(block->u.dc[i]);
1122 static inline void pred_mv(DiracBlock *block, int stride, int x, int y, int ref)
1125 int refmask = ref+1;
1126 int mask = refmask | DIRAC_REF_MASK_GLOBAL; /* exclude gmc blocks */
1129 if (x && (block[-1].ref & mask) == refmask)
1130 pred[n++] = block[-1].u.mv[ref];
1132 if (y && (block[-stride].ref & mask) == refmask)
1133 pred[n++] = block[-stride].u.mv[ref];
1135 if (x && y && (block[-stride-1].ref & mask) == refmask)
1136 pred[n++] = block[-stride-1].u.mv[ref];
1140 block->u.mv[ref][0] = 0;
1141 block->u.mv[ref][1] = 0;
1144 block->u.mv[ref][0] = pred[0][0];
1145 block->u.mv[ref][1] = pred[0][1];
1148 block->u.mv[ref][0] = (pred[0][0] + pred[1][0] + 1) >> 1;
1149 block->u.mv[ref][1] = (pred[0][1] + pred[1][1] + 1) >> 1;
1152 block->u.mv[ref][0] = mid_pred(pred[0][0], pred[1][0], pred[2][0]);
1153 block->u.mv[ref][1] = mid_pred(pred[0][1], pred[1][1], pred[2][1]);
1158 static void global_mv(DiracContext *s, DiracBlock *block, int x, int y, int ref)
1160 int ez = s->globalmc[ref].zrs_exp;
1161 int ep = s->globalmc[ref].perspective_exp;
1162 int (*A)[2] = s->globalmc[ref].zrs;
1163 int *b = s->globalmc[ref].pan_tilt;
1164 int *c = s->globalmc[ref].perspective;
1166 int m = (1<<ep) - (c[0]*x + c[1]*y);
1167 int mx = m * ((A[0][0] * x + A[0][1]*y) + (1<<ez) * b[0]);
1168 int my = m * ((A[1][0] * x + A[1][1]*y) + (1<<ez) * b[1]);
1170 block->u.mv[ref][0] = (mx + (1<<(ez+ep))) >> (ez+ep);
1171 block->u.mv[ref][1] = (my + (1<<(ez+ep))) >> (ez+ep);
1174 static void decode_block_params(DiracContext *s, DiracArith arith[8], DiracBlock *block,
1175 int stride, int x, int y)
1179 block->ref = pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF1);
1180 block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF1);
1182 if (s->num_refs == 2) {
1183 block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF2);
1184 block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF2) << 1;
1188 pred_block_dc(block, stride, x, y);
1189 for (i = 0; i < 3; i++)
1190 block->u.dc[i] += dirac_get_arith_int(arith+1+i, CTX_DC_F1, CTX_DC_DATA);
1194 if (s->globalmc_flag) {
1195 block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_GLOBAL);
1196 block->ref ^= dirac_get_arith_bit(arith, CTX_GLOBAL_BLOCK) << 2;
1199 for (i = 0; i < s->num_refs; i++)
1200 if (block->ref & (i+1)) {
1201 if (block->ref & DIRAC_REF_MASK_GLOBAL) {
1202 global_mv(s, block, x, y, i);
1204 pred_mv(block, stride, x, y, i);
1205 block->u.mv[i][0] += dirac_get_arith_int(arith + 4 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
1206 block->u.mv[i][1] += dirac_get_arith_int(arith + 5 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
1212 * Copies the current block to the other blocks covered by the current superblock split mode
1214 static void propagate_block_data(DiracBlock *block, int stride, int size)
1217 DiracBlock *dst = block;
1219 for (x = 1; x < size; x++)
1222 for (y = 1; y < size; y++) {
1224 for (x = 0; x < size; x++)
1230 * Dirac Specification ->
1231 * 12. Block motion data syntax
1233 static int dirac_unpack_block_motion_data(DiracContext *s)
1235 GetBitContext *gb = &s->gb;
1236 uint8_t *sbsplit = s->sbsplit;
1238 DiracArith arith[8];
1242 /* [DIRAC_STD] 11.2.4 and 12.2.1 Number of blocks and superblocks */
1243 s->sbwidth = DIVRNDUP(s->source.width, 4*s->plane[0].xbsep);
1244 s->sbheight = DIVRNDUP(s->source.height, 4*s->plane[0].ybsep);
1245 s->blwidth = 4 * s->sbwidth;
1246 s->blheight = 4 * s->sbheight;
1248 /* [DIRAC_STD] 12.3.1 Superblock splitting modes. superblock_split_modes()
1249 decode superblock split modes */
1250 ff_dirac_init_arith_decoder(arith, gb, svq3_get_ue_golomb(gb)); /* svq3_get_ue_golomb(gb) is the length */
1251 for (y = 0; y < s->sbheight; y++) {
1252 for (x = 0; x < s->sbwidth; x++) {
1253 unsigned int split = dirac_get_arith_uint(arith, CTX_SB_F1, CTX_SB_DATA);
1256 sbsplit[x] = (split + pred_sbsplit(sbsplit+x, s->sbwidth, x, y)) % 3;
1258 sbsplit += s->sbwidth;
1261 /* setup arith decoding */
1262 ff_dirac_init_arith_decoder(arith, gb, svq3_get_ue_golomb(gb));
1263 for (i = 0; i < s->num_refs; i++) {
1264 ff_dirac_init_arith_decoder(arith + 4 + 2 * i, gb, svq3_get_ue_golomb(gb));
1265 ff_dirac_init_arith_decoder(arith + 5 + 2 * i, gb, svq3_get_ue_golomb(gb));
1267 for (i = 0; i < 3; i++)
1268 ff_dirac_init_arith_decoder(arith+1+i, gb, svq3_get_ue_golomb(gb));
1270 for (y = 0; y < s->sbheight; y++)
1271 for (x = 0; x < s->sbwidth; x++) {
1272 int blkcnt = 1 << s->sbsplit[y * s->sbwidth + x];
1273 int step = 4 >> s->sbsplit[y * s->sbwidth + x];
1275 for (q = 0; q < blkcnt; q++)
1276 for (p = 0; p < blkcnt; p++) {
1277 int bx = 4 * x + p*step;
1278 int by = 4 * y + q*step;
1279 DiracBlock *block = &s->blmotion[by*s->blwidth + bx];
1280 decode_block_params(s, arith, block, s->blwidth, bx, by);
1281 propagate_block_data(block, s->blwidth, step);
1288 static int weight(int i, int blen, int offset)
1290 #define ROLLOFF(i) offset == 1 ? ((i) ? 5 : 3) : \
1291 (1 + (6*(i) + offset - 1) / (2*offset - 1))
1295 else if (i > blen-1 - 2*offset)
1296 return ROLLOFF(blen-1 - i);
1300 static void init_obmc_weight_row(Plane *p, uint8_t *obmc_weight, int stride,
1301 int left, int right, int wy)
1304 for (x = 0; left && x < p->xblen >> 1; x++)
1305 obmc_weight[x] = wy*8;
1306 for (; x < p->xblen >> right; x++)
1307 obmc_weight[x] = wy*weight(x, p->xblen, p->xoffset);
1308 for (; x < p->xblen; x++)
1309 obmc_weight[x] = wy*8;
1310 for (; x < stride; x++)
1314 static void init_obmc_weight(Plane *p, uint8_t *obmc_weight, int stride,
1315 int left, int right, int top, int bottom)
1318 for (y = 0; top && y < p->yblen >> 1; y++) {
1319 init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
1320 obmc_weight += stride;
1322 for (; y < p->yblen >> bottom; y++) {
1323 int wy = weight(y, p->yblen, p->yoffset);
1324 init_obmc_weight_row(p, obmc_weight, stride, left, right, wy);
1325 obmc_weight += stride;
1327 for (; y < p->yblen; y++) {
1328 init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
1329 obmc_weight += stride;
1333 static void init_obmc_weights(DiracContext *s, Plane *p, int by)
1336 int bottom = by == s->blheight-1;
1338 /* don't bother re-initing for rows 2 to blheight-2, the weights don't change */
1339 if (top || bottom || by == 1) {
1340 init_obmc_weight(p, s->obmc_weight[0], MAX_BLOCKSIZE, 1, 0, top, bottom);
1341 init_obmc_weight(p, s->obmc_weight[1], MAX_BLOCKSIZE, 0, 0, top, bottom);
1342 init_obmc_weight(p, s->obmc_weight[2], MAX_BLOCKSIZE, 0, 1, top, bottom);
1346 static const uint8_t epel_weights[4][4][4] = {
1366 * For block x,y, determine which of the hpel planes to do bilinear
1367 * interpolation from and set src[] to the location in each hpel plane
1370 * @return the index of the put_dirac_pixels_tab function to use
1371 * 0 for 1 plane (fpel,hpel), 1 for 2 planes (qpel), 2 for 4 planes (qpel), and 3 for epel
1373 static int mc_subpel(DiracContext *s, DiracBlock *block, const uint8_t *src[5],
1374 int x, int y, int ref, int plane)
1376 Plane *p = &s->plane[plane];
1377 uint8_t **ref_hpel = s->ref_pics[ref]->hpel[plane];
1378 int motion_x = block->u.mv[ref][0];
1379 int motion_y = block->u.mv[ref][1];
1380 int mx, my, i, epel, nplanes = 0;
1383 motion_x >>= s->chroma_x_shift;
1384 motion_y >>= s->chroma_y_shift;
1387 mx = motion_x & ~(-1U << s->mv_precision);
1388 my = motion_y & ~(-1U << s->mv_precision);
1389 motion_x >>= s->mv_precision;
1390 motion_y >>= s->mv_precision;
1391 /* normalize subpel coordinates to epel */
1392 /* TODO: template this function? */
1393 mx <<= 3 - s->mv_precision;
1394 my <<= 3 - s->mv_precision;
1403 src[0] = ref_hpel[(my>>1)+(mx>>2)] + y*p->stride + x;
1407 for (i = 0; i < 4; i++)
1408 src[i] = ref_hpel[i] + y*p->stride + x;
1410 /* if we're interpolating in the right/bottom halves, adjust the planes as needed
1411 we increment x/y because the edge changes for half of the pixels */
1418 src[0] += p->stride;
1419 src[1] += p->stride;
1427 /* check if we really only need 2 planes since either mx or my is
1428 a hpel position. (epel weights of 0 handle this there) */
1430 /* mx == 0: average [0] and [2]
1431 mx == 4: average [1] and [3] */
1432 src[!mx] = src[2 + !!mx];
1434 } else if (!(my&3)) {
1435 src[0] = src[(my>>1) ];
1436 src[1] = src[(my>>1)+1];
1440 /* adjust the ordering if needed so the weights work */
1442 FFSWAP(const uint8_t *, src[0], src[1]);
1443 FFSWAP(const uint8_t *, src[2], src[3]);
1446 FFSWAP(const uint8_t *, src[0], src[2]);
1447 FFSWAP(const uint8_t *, src[1], src[3]);
1449 src[4] = epel_weights[my&3][mx&3];
1453 /* fixme: v/h _edge_pos */
1454 if (x + p->xblen > p->width +EDGE_WIDTH/2 ||
1455 y + p->yblen > p->height+EDGE_WIDTH/2 ||
1457 for (i = 0; i < nplanes; i++) {
1458 s->vdsp.emulated_edge_mc(s->edge_emu_buffer[i], src[i],
1459 p->stride, p->stride,
1460 p->xblen, p->yblen, x, y,
1461 p->width+EDGE_WIDTH/2, p->height+EDGE_WIDTH/2);
1462 src[i] = s->edge_emu_buffer[i];
1465 return (nplanes>>1) + epel;
1468 static void add_dc(uint16_t *dst, int dc, int stride,
1469 uint8_t *obmc_weight, int xblen, int yblen)
1474 for (y = 0; y < yblen; y++) {
1475 for (x = 0; x < xblen; x += 2) {
1476 dst[x ] += dc * obmc_weight[x ];
1477 dst[x+1] += dc * obmc_weight[x+1];
1480 obmc_weight += MAX_BLOCKSIZE;
1484 static void block_mc(DiracContext *s, DiracBlock *block,
1485 uint16_t *mctmp, uint8_t *obmc_weight,
1486 int plane, int dstx, int dsty)
1488 Plane *p = &s->plane[plane];
1489 const uint8_t *src[5];
1492 switch (block->ref&3) {
1494 add_dc(mctmp, block->u.dc[plane], p->stride, obmc_weight, p->xblen, p->yblen);
1498 idx = mc_subpel(s, block, src, dstx, dsty, (block->ref&3)-1, plane);
1499 s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1501 s->weight_func(s->mcscratch, p->stride, s->weight_log2denom,
1502 s->weight[0] + s->weight[1], p->yblen);
1505 idx = mc_subpel(s, block, src, dstx, dsty, 0, plane);
1506 s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1507 idx = mc_subpel(s, block, src, dstx, dsty, 1, plane);
1508 if (s->biweight_func) {
1509 /* fixme: +32 is a quick hack */
1510 s->put_pixels_tab[idx](s->mcscratch + 32, src, p->stride, p->yblen);
1511 s->biweight_func(s->mcscratch, s->mcscratch+32, p->stride, s->weight_log2denom,
1512 s->weight[0], s->weight[1], p->yblen);
1514 s->avg_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1517 s->add_obmc(mctmp, s->mcscratch, p->stride, obmc_weight, p->yblen);
1520 static void mc_row(DiracContext *s, DiracBlock *block, uint16_t *mctmp, int plane, int dsty)
1522 Plane *p = &s->plane[plane];
1523 int x, dstx = p->xbsep - p->xoffset;
1525 block_mc(s, block, mctmp, s->obmc_weight[0], plane, -p->xoffset, dsty);
1528 for (x = 1; x < s->blwidth-1; x++) {
1529 block_mc(s, block+x, mctmp, s->obmc_weight[1], plane, dstx, dsty);
1533 block_mc(s, block+x, mctmp, s->obmc_weight[2], plane, dstx, dsty);
1536 static void select_dsp_funcs(DiracContext *s, int width, int height, int xblen, int yblen)
1544 memcpy(s->put_pixels_tab, s->diracdsp.put_dirac_pixels_tab[idx], sizeof(s->put_pixels_tab));
1545 memcpy(s->avg_pixels_tab, s->diracdsp.avg_dirac_pixels_tab[idx], sizeof(s->avg_pixels_tab));
1546 s->add_obmc = s->diracdsp.add_dirac_obmc[idx];
1547 if (s->weight_log2denom > 1 || s->weight[0] != 1 || s->weight[1] != 1) {
1548 s->weight_func = s->diracdsp.weight_dirac_pixels_tab[idx];
1549 s->biweight_func = s->diracdsp.biweight_dirac_pixels_tab[idx];
1551 s->weight_func = NULL;
1552 s->biweight_func = NULL;
1556 static void interpolate_refplane(DiracContext *s, DiracFrame *ref, int plane, int width, int height)
1558 /* chroma allocates an edge of 8 when subsampled
1559 which for 4:2:2 means an h edge of 16 and v edge of 8
1560 just use 8 for everything for the moment */
1561 int i, edge = EDGE_WIDTH/2;
1563 ref->hpel[plane][0] = ref->avframe->data[plane];
1564 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 */
1566 /* no need for hpel if we only have fpel vectors */
1567 if (!s->mv_precision)
1570 for (i = 1; i < 4; i++) {
1571 if (!ref->hpel_base[plane][i])
1572 ref->hpel_base[plane][i] = av_malloc((height+2*edge) * ref->avframe->linesize[plane] + 32);
1573 /* we need to be 16-byte aligned even for chroma */
1574 ref->hpel[plane][i] = ref->hpel_base[plane][i] + edge*ref->avframe->linesize[plane] + 16;
1577 if (!ref->interpolated[plane]) {
1578 s->diracdsp.dirac_hpel_filter(ref->hpel[plane][1], ref->hpel[plane][2],
1579 ref->hpel[plane][3], ref->hpel[plane][0],
1580 ref->avframe->linesize[plane], width, height);
1581 s->mpvencdsp.draw_edges(ref->hpel[plane][1], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1582 s->mpvencdsp.draw_edges(ref->hpel[plane][2], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1583 s->mpvencdsp.draw_edges(ref->hpel[plane][3], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1585 ref->interpolated[plane] = 1;
1589 * Dirac Specification ->
1590 * 13.0 Transform data syntax. transform_data()
1592 static int dirac_decode_frame_internal(DiracContext *s)
1595 int y, i, comp, dsty;
1599 /* [DIRAC_STD] 13.5.1 low_delay_transform_data() */
1600 for (comp = 0; comp < 3; comp++) {
1601 Plane *p = &s->plane[comp];
1602 memset(p->idwt_buf, 0, p->idwt_stride * p->idwt_height * sizeof(IDWTELEM));
1605 if ((ret = decode_lowdelay(s)) < 0)
1610 for (comp = 0; comp < 3; comp++) {
1611 Plane *p = &s->plane[comp];
1612 uint8_t *frame = s->current_picture->avframe->data[comp];
1614 /* FIXME: small resolutions */
1615 for (i = 0; i < 4; i++)
1616 s->edge_emu_buffer[i] = s->edge_emu_buffer_base + i*FFALIGN(p->width, 16);
1618 if (!s->zero_res && !s->low_delay)
1620 memset(p->idwt_buf, 0, p->idwt_stride * p->idwt_height * sizeof(IDWTELEM));
1621 decode_component(s, comp); /* [DIRAC_STD] 13.4.1 core_transform_data() */
1623 if (ff_spatial_idwt_init2(&d, p->idwt_buf, p->idwt_width, p->idwt_height, p->idwt_stride,
1624 s->wavelet_idx+2, s->wavelet_depth, p->idwt_tmp))
1627 if (!s->num_refs) { /* intra */
1628 for (y = 0; y < p->height; y += 16) {
1629 ff_spatial_idwt_slice2(&d, y+16); /* decode */
1630 s->diracdsp.put_signed_rect_clamped(frame + y*p->stride, p->stride,
1631 p->idwt_buf + y*p->idwt_stride, p->idwt_stride, p->width, 16);
1633 } else { /* inter */
1634 int rowheight = p->ybsep*p->stride;
1636 select_dsp_funcs(s, p->width, p->height, p->xblen, p->yblen);
1638 for (i = 0; i < s->num_refs; i++)
1639 interpolate_refplane(s, s->ref_pics[i], comp, p->width, p->height);
1641 memset(s->mctmp, 0, 4*p->yoffset*p->stride);
1644 for (y = 0; y < s->blheight; y++) {
1646 start = FFMAX(dsty, 0);
1647 uint16_t *mctmp = s->mctmp + y*rowheight;
1648 DiracBlock *blocks = s->blmotion + y*s->blwidth;
1650 init_obmc_weights(s, p, y);
1652 if (y == s->blheight-1 || start+p->ybsep > p->height)
1653 h = p->height - start;
1655 h = p->ybsep - (start - dsty);
1659 memset(mctmp+2*p->yoffset*p->stride, 0, 2*rowheight);
1660 mc_row(s, blocks, mctmp, comp, dsty);
1662 mctmp += (start - dsty)*p->stride + p->xoffset;
1663 ff_spatial_idwt_slice2(&d, start + h); /* decode */
1664 s->diracdsp.add_rect_clamped(frame + start*p->stride, mctmp, p->stride,
1665 p->idwt_buf + start*p->idwt_stride, p->idwt_stride, p->width, h);
1676 static int get_buffer_with_edge(AVCodecContext *avctx, AVFrame *f, int flags)
1679 int chroma_x_shift, chroma_y_shift;
1680 avcodec_get_chroma_sub_sample(avctx->pix_fmt, &chroma_x_shift, &chroma_y_shift);
1682 f->width = avctx->width + 2 * EDGE_WIDTH;
1683 f->height = avctx->height + 2 * EDGE_WIDTH + 2;
1684 ret = ff_get_buffer(avctx, f, flags);
1688 for (i = 0; f->data[i]; i++) {
1689 int offset = (EDGE_WIDTH >> (i && i<3 ? chroma_y_shift : 0)) *
1690 f->linesize[i] + 32;
1691 f->data[i] += offset;
1693 f->width = avctx->width;
1694 f->height = avctx->height;
1700 * Dirac Specification ->
1701 * 11.1.1 Picture Header. picture_header()
1703 static int dirac_decode_picture_header(DiracContext *s)
1706 int i, j, refnum, refdist;
1707 GetBitContext *gb = &s->gb;
1709 /* [DIRAC_STD] 11.1.1 Picture Header. picture_header() PICTURE_NUM */
1710 picnum = s->current_picture->avframe->display_picture_number = get_bits_long(gb, 32);
1713 av_log(s->avctx,AV_LOG_DEBUG,"PICTURE_NUM: %d\n",picnum);
1715 /* if this is the first keyframe after a sequence header, start our
1716 reordering from here */
1717 if (s->frame_number < 0)
1718 s->frame_number = picnum;
1720 s->ref_pics[0] = s->ref_pics[1] = NULL;
1721 for (i = 0; i < s->num_refs; i++) {
1722 refnum = picnum + dirac_get_se_golomb(gb);
1725 /* find the closest reference to the one we want */
1726 /* Jordi: this is needed if the referenced picture hasn't yet arrived */
1727 for (j = 0; j < MAX_REFERENCE_FRAMES && refdist; j++)
1728 if (s->ref_frames[j]
1729 && FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum) < refdist) {
1730 s->ref_pics[i] = s->ref_frames[j];
1731 refdist = FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum);
1734 if (!s->ref_pics[i] || refdist)
1735 av_log(s->avctx, AV_LOG_DEBUG, "Reference not found\n");
1737 /* if there were no references at all, allocate one */
1738 if (!s->ref_pics[i])
1739 for (j = 0; j < MAX_FRAMES; j++)
1740 if (!s->all_frames[j].avframe->data[0]) {
1741 s->ref_pics[i] = &s->all_frames[j];
1742 get_buffer_with_edge(s->avctx, s->ref_pics[i]->avframe, AV_GET_BUFFER_FLAG_REF);
1747 /* retire the reference frames that are not used anymore */
1748 if (s->current_picture->avframe->reference) {
1749 retire = picnum + dirac_get_se_golomb(gb);
1750 if (retire != picnum) {
1751 DiracFrame *retire_pic = remove_frame(s->ref_frames, retire);
1754 retire_pic->avframe->reference &= DELAYED_PIC_REF;
1756 av_log(s->avctx, AV_LOG_DEBUG, "Frame to retire not found\n");
1759 /* if reference array is full, remove the oldest as per the spec */
1760 while (add_frame(s->ref_frames, MAX_REFERENCE_FRAMES, s->current_picture)) {
1761 av_log(s->avctx, AV_LOG_ERROR, "Reference frame overflow\n");
1762 remove_frame(s->ref_frames, s->ref_frames[0]->avframe->display_picture_number)->avframe->reference &= DELAYED_PIC_REF;
1767 if (dirac_unpack_prediction_parameters(s)) /* [DIRAC_STD] 11.2 Picture Prediction Data. picture_prediction() */
1769 if (dirac_unpack_block_motion_data(s)) /* [DIRAC_STD] 12. Block motion data syntax */
1772 if (dirac_unpack_idwt_params(s)) /* [DIRAC_STD] 11.3 Wavelet transform data */
1779 static int get_delayed_pic(DiracContext *s, AVFrame *picture, int *got_frame)
1781 DiracFrame *out = s->delay_frames[0];
1785 /* find frame with lowest picture number */
1786 for (i = 1; s->delay_frames[i]; i++)
1787 if (s->delay_frames[i]->avframe->display_picture_number < out->avframe->display_picture_number) {
1788 out = s->delay_frames[i];
1792 for (i = out_idx; s->delay_frames[i]; i++)
1793 s->delay_frames[i] = s->delay_frames[i+1];
1796 out->avframe->reference ^= DELAYED_PIC_REF;
1798 if((ret = av_frame_ref(picture, out->avframe)) < 0)
1806 * Dirac Specification ->
1807 * 9.6 Parse Info Header Syntax. parse_info()
1808 * 4 byte start code + byte parse code + 4 byte size + 4 byte previous size
1810 #define DATA_UNIT_HEADER_SIZE 13
1812 /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3
1813 inside the function parse_sequence() */
1814 static int dirac_decode_data_unit(AVCodecContext *avctx, const uint8_t *buf, int size)
1816 DiracContext *s = avctx->priv_data;
1817 DiracFrame *pic = NULL;
1818 int ret, i, parse_code = buf[4];
1821 if (size < DATA_UNIT_HEADER_SIZE)
1824 init_get_bits(&s->gb, &buf[13], 8*(size - DATA_UNIT_HEADER_SIZE));
1826 if (parse_code == pc_seq_header) {
1827 if (s->seen_sequence_header)
1830 /* [DIRAC_STD] 10. Sequence header */
1831 if (avpriv_dirac_parse_sequence_header(avctx, &s->gb, &s->source))
1834 avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_x_shift, &s->chroma_y_shift);
1836 if (alloc_sequence_buffers(s))
1839 s->seen_sequence_header = 1;
1840 } else if (parse_code == pc_eos) { /* [DIRAC_STD] End of Sequence */
1841 free_sequence_buffers(s);
1842 s->seen_sequence_header = 0;
1843 } else if (parse_code == pc_aux_data) {
1844 if (buf[13] == 1) { /* encoder implementation/version */
1846 /* versions older than 1.0.8 don't store quant delta for
1847 subbands with only one codeblock */
1848 if (sscanf(buf+14, "Schroedinger %d.%d.%d", ver, ver+1, ver+2) == 3)
1849 if (ver[0] == 1 && ver[1] == 0 && ver[2] <= 7)
1850 s->old_delta_quant = 1;
1852 } else if (parse_code & 0x8) { /* picture data unit */
1853 if (!s->seen_sequence_header) {
1854 av_log(avctx, AV_LOG_DEBUG, "Dropping frame without sequence header\n");
1858 /* find an unused frame */
1859 for (i = 0; i < MAX_FRAMES; i++)
1860 if (s->all_frames[i].avframe->data[0] == NULL)
1861 pic = &s->all_frames[i];
1863 av_log(avctx, AV_LOG_ERROR, "framelist full\n");
1867 av_frame_unref(pic->avframe);
1869 /* [DIRAC_STD] Defined in 9.6.1 ... */
1870 tmp = parse_code & 0x03; /* [DIRAC_STD] num_refs() */
1872 av_log(avctx, AV_LOG_ERROR, "num_refs of 3\n");
1876 s->is_arith = (parse_code & 0x48) == 0x08; /* [DIRAC_STD] using_ac() */
1877 s->low_delay = (parse_code & 0x88) == 0x88; /* [DIRAC_STD] is_low_delay() */
1878 pic->avframe->reference = (parse_code & 0x0C) == 0x0C; /* [DIRAC_STD] is_reference() */
1879 pic->avframe->key_frame = s->num_refs == 0; /* [DIRAC_STD] is_intra() */
1880 pic->avframe->pict_type = s->num_refs + 1; /* Definition of AVPictureType in avutil.h */
1882 if ((ret = get_buffer_with_edge(avctx, pic->avframe, (parse_code & 0x0C) == 0x0C ? AV_GET_BUFFER_FLAG_REF : 0)) < 0)
1884 s->current_picture = pic;
1885 s->plane[0].stride = pic->avframe->linesize[0];
1886 s->plane[1].stride = pic->avframe->linesize[1];
1887 s->plane[2].stride = pic->avframe->linesize[2];
1889 if (alloc_buffers(s, FFMAX3(FFABS(s->plane[0].stride), FFABS(s->plane[1].stride), FFABS(s->plane[2].stride))) < 0)
1890 return AVERROR(ENOMEM);
1892 /* [DIRAC_STD] 11.1 Picture parse. picture_parse() */
1893 if (dirac_decode_picture_header(s))
1896 /* [DIRAC_STD] 13.0 Transform data syntax. transform_data() */
1897 if (dirac_decode_frame_internal(s))
1903 static int dirac_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *pkt)
1905 DiracContext *s = avctx->priv_data;
1906 AVFrame *picture = data;
1907 uint8_t *buf = pkt->data;
1908 int buf_size = pkt->size;
1909 int i, data_unit_size, buf_idx = 0;
1912 /* release unused frames */
1913 for (i = 0; i < MAX_FRAMES; i++)
1914 if (s->all_frames[i].avframe->data[0] && !s->all_frames[i].avframe->reference) {
1915 av_frame_unref(s->all_frames[i].avframe);
1916 memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));
1919 s->current_picture = NULL;
1922 /* end of stream, so flush delayed pics */
1924 return get_delayed_pic(s, (AVFrame *)data, got_frame);
1927 /*[DIRAC_STD] Here starts the code from parse_info() defined in 9.6
1928 [DIRAC_STD] PARSE_INFO_PREFIX = "BBCD" as defined in ISO/IEC 646
1929 BBCD start code search */
1930 for (; buf_idx + DATA_UNIT_HEADER_SIZE < buf_size; buf_idx++) {
1931 if (buf[buf_idx ] == 'B' && buf[buf_idx+1] == 'B' &&
1932 buf[buf_idx+2] == 'C' && buf[buf_idx+3] == 'D')
1935 /* BBCD found or end of data */
1936 if (buf_idx + DATA_UNIT_HEADER_SIZE >= buf_size)
1939 data_unit_size = AV_RB32(buf+buf_idx+5);
1940 if (buf_idx + data_unit_size > buf_size || !data_unit_size) {
1941 if(buf_idx + data_unit_size > buf_size)
1942 av_log(s->avctx, AV_LOG_ERROR,
1943 "Data unit with size %d is larger than input buffer, discarding\n",
1948 /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3 inside the function parse_sequence() */
1949 if (dirac_decode_data_unit(avctx, buf+buf_idx, data_unit_size))
1951 av_log(s->avctx, AV_LOG_ERROR,"Error in dirac_decode_data_unit\n");
1954 buf_idx += data_unit_size;
1957 if (!s->current_picture)
1960 if (s->current_picture->avframe->display_picture_number > s->frame_number) {
1961 DiracFrame *delayed_frame = remove_frame(s->delay_frames, s->frame_number);
1963 s->current_picture->avframe->reference |= DELAYED_PIC_REF;
1965 if (add_frame(s->delay_frames, MAX_DELAY, s->current_picture)) {
1966 int min_num = s->delay_frames[0]->avframe->display_picture_number;
1967 /* Too many delayed frames, so we display the frame with the lowest pts */
1968 av_log(avctx, AV_LOG_ERROR, "Delay frame overflow\n");
1970 for (i = 1; s->delay_frames[i]; i++)
1971 if (s->delay_frames[i]->avframe->display_picture_number < min_num)
1972 min_num = s->delay_frames[i]->avframe->display_picture_number;
1974 delayed_frame = remove_frame(s->delay_frames, min_num);
1975 add_frame(s->delay_frames, MAX_DELAY, s->current_picture);
1978 if (delayed_frame) {
1979 delayed_frame->avframe->reference ^= DELAYED_PIC_REF;
1980 if((ret=av_frame_ref(data, delayed_frame->avframe)) < 0)
1984 } else if (s->current_picture->avframe->display_picture_number == s->frame_number) {
1985 /* The right frame at the right time :-) */
1986 if((ret=av_frame_ref(data, s->current_picture->avframe)) < 0)
1992 s->frame_number = picture->display_picture_number + 1;
1997 AVCodec ff_dirac_decoder = {
1999 .long_name = NULL_IF_CONFIG_SMALL("BBC Dirac VC-2"),
2000 .type = AVMEDIA_TYPE_VIDEO,
2001 .id = AV_CODEC_ID_DIRAC,
2002 .priv_data_size = sizeof(DiracContext),
2003 .init = dirac_decode_init,
2004 .close = dirac_decode_end,
2005 .decode = dirac_decode_frame,
2006 .capabilities = CODEC_CAP_DELAY,
2007 .flush = dirac_decode_flush,