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);
804 if (bufsize/8 >= bytes)
810 avctx->execute(avctx, decode_lowdelay_slice, slices, NULL, slice_num,
811 sizeof(struct lowdelay_slice)); /* [DIRAC_STD] 13.5.2 Slices */
812 intra_dc_prediction(&s->plane[0].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
813 intra_dc_prediction(&s->plane[1].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
814 intra_dc_prediction(&s->plane[2].band[0][0]); /* [DIRAC_STD] 13.3 intra_dc_prediction() */
819 static void init_planes(DiracContext *s)
821 int i, w, h, level, orientation;
823 for (i = 0; i < 3; i++) {
824 Plane *p = &s->plane[i];
826 p->width = s->source.width >> (i ? s->chroma_x_shift : 0);
827 p->height = s->source.height >> (i ? s->chroma_y_shift : 0);
828 p->idwt_width = w = CALC_PADDING(p->width , s->wavelet_depth);
829 p->idwt_height = h = CALC_PADDING(p->height, s->wavelet_depth);
830 p->idwt_stride = FFALIGN(p->idwt_width, 8);
832 for (level = s->wavelet_depth-1; level >= 0; level--) {
835 for (orientation = !!level; orientation < 4; orientation++) {
836 SubBand *b = &p->band[level][orientation];
838 b->ibuf = p->idwt_buf;
840 b->stride = p->idwt_stride << (s->wavelet_depth - level);
843 b->orientation = orientation;
848 b->ibuf += b->stride>>1;
851 b->parent = &p->band[level-1][orientation];
856 p->xblen = s->plane[0].xblen >> s->chroma_x_shift;
857 p->yblen = s->plane[0].yblen >> s->chroma_y_shift;
858 p->xbsep = s->plane[0].xbsep >> s->chroma_x_shift;
859 p->ybsep = s->plane[0].ybsep >> s->chroma_y_shift;
862 p->xoffset = (p->xblen - p->xbsep)/2;
863 p->yoffset = (p->yblen - p->ybsep)/2;
868 * Unpack the motion compensation parameters
869 * Dirac Specification ->
870 * 11.2 Picture prediction data. picture_prediction()
872 static int dirac_unpack_prediction_parameters(DiracContext *s)
874 static const uint8_t default_blen[] = { 4, 12, 16, 24 };
875 static const uint8_t default_bsep[] = { 4, 8, 12, 16 };
877 GetBitContext *gb = &s->gb;
881 /* [DIRAC_STD] 11.2.2 Block parameters. block_parameters() */
882 /* Luma and Chroma are equal. 11.2.3 */
883 idx = svq3_get_ue_golomb(gb); /* [DIRAC_STD] index */
886 av_log(s->avctx, AV_LOG_ERROR, "Block prediction index too high\n");
891 s->plane[0].xblen = svq3_get_ue_golomb(gb);
892 s->plane[0].yblen = svq3_get_ue_golomb(gb);
893 s->plane[0].xbsep = svq3_get_ue_golomb(gb);
894 s->plane[0].ybsep = svq3_get_ue_golomb(gb);
896 /*[DIRAC_STD] preset_block_params(index). Table 11.1 */
897 s->plane[0].xblen = default_blen[idx-1];
898 s->plane[0].yblen = default_blen[idx-1];
899 s->plane[0].xbsep = default_bsep[idx-1];
900 s->plane[0].ybsep = default_bsep[idx-1];
902 /*[DIRAC_STD] 11.2.4 motion_data_dimensions()
903 Calculated in function dirac_unpack_block_motion_data */
905 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) {
906 av_log(s->avctx, AV_LOG_ERROR, "Block separation too small\n");
909 if (s->plane[0].xbsep > s->plane[0].xblen || s->plane[0].ybsep > s->plane[0].yblen) {
910 av_log(s->avctx, AV_LOG_ERROR, "Block separation greater than size\n");
913 if (FFMAX(s->plane[0].xblen, s->plane[0].yblen) > MAX_BLOCKSIZE) {
914 av_log(s->avctx, AV_LOG_ERROR, "Unsupported large block size\n");
918 /*[DIRAC_STD] 11.2.5 Motion vector precision. motion_vector_precision()
919 Read motion vector precision */
920 s->mv_precision = svq3_get_ue_golomb(gb);
921 if (s->mv_precision > 3) {
922 av_log(s->avctx, AV_LOG_ERROR, "MV precision finer than eighth-pel\n");
926 /*[DIRAC_STD] 11.2.6 Global motion. global_motion()
927 Read the global motion compensation parameters */
928 s->globalmc_flag = get_bits1(gb);
929 if (s->globalmc_flag) {
930 memset(s->globalmc, 0, sizeof(s->globalmc));
931 /* [DIRAC_STD] pan_tilt(gparams) */
932 for (ref = 0; ref < s->num_refs; ref++) {
934 s->globalmc[ref].pan_tilt[0] = dirac_get_se_golomb(gb);
935 s->globalmc[ref].pan_tilt[1] = dirac_get_se_golomb(gb);
937 /* [DIRAC_STD] zoom_rotate_shear(gparams)
938 zoom/rotation/shear parameters */
940 s->globalmc[ref].zrs_exp = svq3_get_ue_golomb(gb);
941 s->globalmc[ref].zrs[0][0] = dirac_get_se_golomb(gb);
942 s->globalmc[ref].zrs[0][1] = dirac_get_se_golomb(gb);
943 s->globalmc[ref].zrs[1][0] = dirac_get_se_golomb(gb);
944 s->globalmc[ref].zrs[1][1] = dirac_get_se_golomb(gb);
946 s->globalmc[ref].zrs[0][0] = 1;
947 s->globalmc[ref].zrs[1][1] = 1;
949 /* [DIRAC_STD] perspective(gparams) */
951 s->globalmc[ref].perspective_exp = svq3_get_ue_golomb(gb);
952 s->globalmc[ref].perspective[0] = dirac_get_se_golomb(gb);
953 s->globalmc[ref].perspective[1] = dirac_get_se_golomb(gb);
958 /*[DIRAC_STD] 11.2.7 Picture prediction mode. prediction_mode()
959 Picture prediction mode, not currently used. */
960 if (svq3_get_ue_golomb(gb)) {
961 av_log(s->avctx, AV_LOG_ERROR, "Unknown picture prediction mode\n");
965 /* [DIRAC_STD] 11.2.8 Reference picture weight. reference_picture_weights()
966 just data read, weight calculation will be done later on. */
967 s->weight_log2denom = 1;
972 s->weight_log2denom = svq3_get_ue_golomb(gb);
973 s->weight[0] = dirac_get_se_golomb(gb);
974 if (s->num_refs == 2)
975 s->weight[1] = dirac_get_se_golomb(gb);
981 * Dirac Specification ->
982 * 11.3 Wavelet transform data. wavelet_transform()
984 static int dirac_unpack_idwt_params(DiracContext *s)
986 GetBitContext *gb = &s->gb;
990 #define CHECKEDREAD(dst, cond, errmsg) \
991 tmp = svq3_get_ue_golomb(gb); \
993 av_log(s->avctx, AV_LOG_ERROR, errmsg); \
1000 s->zero_res = s->num_refs ? get_bits1(gb) : 0;
1004 /*[DIRAC_STD] 11.3.1 Transform parameters. transform_parameters() */
1005 CHECKEDREAD(s->wavelet_idx, tmp > 6, "wavelet_idx is too big\n")
1007 CHECKEDREAD(s->wavelet_depth, tmp > MAX_DWT_LEVELS || tmp < 1, "invalid number of DWT decompositions\n")
1009 if (!s->low_delay) {
1010 /* Codeblock parameters (core syntax only) */
1011 if (get_bits1(gb)) {
1012 for (i = 0; i <= s->wavelet_depth; i++) {
1013 CHECKEDREAD(s->codeblock[i].width , tmp < 1 || tmp > (s->avctx->width >>s->wavelet_depth-i), "codeblock width invalid\n")
1014 CHECKEDREAD(s->codeblock[i].height, tmp < 1 || tmp > (s->avctx->height>>s->wavelet_depth-i), "codeblock height invalid\n")
1017 CHECKEDREAD(s->codeblock_mode, tmp > 1, "unknown codeblock mode\n")
1019 for (i = 0; i <= s->wavelet_depth; i++)
1020 s->codeblock[i].width = s->codeblock[i].height = 1;
1022 /* Slice parameters + quantization matrix*/
1023 /*[DIRAC_STD] 11.3.4 Slice coding Parameters (low delay syntax only). slice_parameters() */
1024 s->lowdelay.num_x = svq3_get_ue_golomb(gb);
1025 s->lowdelay.num_y = svq3_get_ue_golomb(gb);
1026 s->lowdelay.bytes.num = svq3_get_ue_golomb(gb);
1027 s->lowdelay.bytes.den = svq3_get_ue_golomb(gb);
1029 if (s->lowdelay.bytes.den <= 0) {
1030 av_log(s->avctx,AV_LOG_ERROR,"Invalid lowdelay.bytes.den\n");
1031 return AVERROR_INVALIDDATA;
1034 /* [DIRAC_STD] 11.3.5 Quantisation matrices (low-delay syntax). quant_matrix() */
1035 if (get_bits1(gb)) {
1036 av_log(s->avctx,AV_LOG_DEBUG,"Low Delay: Has Custom Quantization Matrix!\n");
1037 /* custom quantization matrix */
1038 s->lowdelay.quant[0][0] = svq3_get_ue_golomb(gb);
1039 for (level = 0; level < s->wavelet_depth; level++) {
1040 s->lowdelay.quant[level][1] = svq3_get_ue_golomb(gb);
1041 s->lowdelay.quant[level][2] = svq3_get_ue_golomb(gb);
1042 s->lowdelay.quant[level][3] = svq3_get_ue_golomb(gb);
1045 if (s->wavelet_depth > 4) {
1046 av_log(s->avctx,AV_LOG_ERROR,"Mandatory custom low delay matrix missing for depth %d\n", s->wavelet_depth);
1047 return AVERROR_INVALIDDATA;
1049 /* default quantization matrix */
1050 for (level = 0; level < s->wavelet_depth; level++)
1051 for (i = 0; i < 4; i++) {
1052 s->lowdelay.quant[level][i] = default_qmat[s->wavelet_idx][level][i];
1053 /* haar with no shift differs for different depths */
1054 if (s->wavelet_idx == 3)
1055 s->lowdelay.quant[level][i] += 4*(s->wavelet_depth-1 - level);
1062 static inline int pred_sbsplit(uint8_t *sbsplit, int stride, int x, int y)
1064 static const uint8_t avgsplit[7] = { 0, 0, 1, 1, 1, 2, 2 };
1071 return sbsplit[-stride];
1073 return avgsplit[sbsplit[-1] + sbsplit[-stride] + sbsplit[-stride-1]];
1076 static inline int pred_block_mode(DiracBlock *block, int stride, int x, int y, int refmask)
1083 return block[-1].ref & refmask;
1085 return block[-stride].ref & refmask;
1087 /* return the majority */
1088 pred = (block[-1].ref & refmask) + (block[-stride].ref & refmask) + (block[-stride-1].ref & refmask);
1089 return (pred >> 1) & refmask;
1092 static inline void pred_block_dc(DiracBlock *block, int stride, int x, int y)
1096 memset(block->u.dc, 0, sizeof(block->u.dc));
1098 if (x && !(block[-1].ref & 3)) {
1099 for (i = 0; i < 3; i++)
1100 block->u.dc[i] += block[-1].u.dc[i];
1104 if (y && !(block[-stride].ref & 3)) {
1105 for (i = 0; i < 3; i++)
1106 block->u.dc[i] += block[-stride].u.dc[i];
1110 if (x && y && !(block[-1-stride].ref & 3)) {
1111 for (i = 0; i < 3; i++)
1112 block->u.dc[i] += block[-1-stride].u.dc[i];
1117 for (i = 0; i < 3; i++)
1118 block->u.dc[i] = (block->u.dc[i]+1)>>1;
1119 } else if (n == 3) {
1120 for (i = 0; i < 3; i++)
1121 block->u.dc[i] = divide3(block->u.dc[i]);
1125 static inline void pred_mv(DiracBlock *block, int stride, int x, int y, int ref)
1128 int refmask = ref+1;
1129 int mask = refmask | DIRAC_REF_MASK_GLOBAL; /* exclude gmc blocks */
1132 if (x && (block[-1].ref & mask) == refmask)
1133 pred[n++] = block[-1].u.mv[ref];
1135 if (y && (block[-stride].ref & mask) == refmask)
1136 pred[n++] = block[-stride].u.mv[ref];
1138 if (x && y && (block[-stride-1].ref & mask) == refmask)
1139 pred[n++] = block[-stride-1].u.mv[ref];
1143 block->u.mv[ref][0] = 0;
1144 block->u.mv[ref][1] = 0;
1147 block->u.mv[ref][0] = pred[0][0];
1148 block->u.mv[ref][1] = pred[0][1];
1151 block->u.mv[ref][0] = (pred[0][0] + pred[1][0] + 1) >> 1;
1152 block->u.mv[ref][1] = (pred[0][1] + pred[1][1] + 1) >> 1;
1155 block->u.mv[ref][0] = mid_pred(pred[0][0], pred[1][0], pred[2][0]);
1156 block->u.mv[ref][1] = mid_pred(pred[0][1], pred[1][1], pred[2][1]);
1161 static void global_mv(DiracContext *s, DiracBlock *block, int x, int y, int ref)
1163 int ez = s->globalmc[ref].zrs_exp;
1164 int ep = s->globalmc[ref].perspective_exp;
1165 int (*A)[2] = s->globalmc[ref].zrs;
1166 int *b = s->globalmc[ref].pan_tilt;
1167 int *c = s->globalmc[ref].perspective;
1169 int m = (1<<ep) - (c[0]*x + c[1]*y);
1170 int mx = m * ((A[0][0] * x + A[0][1]*y) + (1<<ez) * b[0]);
1171 int my = m * ((A[1][0] * x + A[1][1]*y) + (1<<ez) * b[1]);
1173 block->u.mv[ref][0] = (mx + (1<<(ez+ep))) >> (ez+ep);
1174 block->u.mv[ref][1] = (my + (1<<(ez+ep))) >> (ez+ep);
1177 static void decode_block_params(DiracContext *s, DiracArith arith[8], DiracBlock *block,
1178 int stride, int x, int y)
1182 block->ref = pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF1);
1183 block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF1);
1185 if (s->num_refs == 2) {
1186 block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_REF2);
1187 block->ref ^= dirac_get_arith_bit(arith, CTX_PMODE_REF2) << 1;
1191 pred_block_dc(block, stride, x, y);
1192 for (i = 0; i < 3; i++)
1193 block->u.dc[i] += dirac_get_arith_int(arith+1+i, CTX_DC_F1, CTX_DC_DATA);
1197 if (s->globalmc_flag) {
1198 block->ref |= pred_block_mode(block, stride, x, y, DIRAC_REF_MASK_GLOBAL);
1199 block->ref ^= dirac_get_arith_bit(arith, CTX_GLOBAL_BLOCK) << 2;
1202 for (i = 0; i < s->num_refs; i++)
1203 if (block->ref & (i+1)) {
1204 if (block->ref & DIRAC_REF_MASK_GLOBAL) {
1205 global_mv(s, block, x, y, i);
1207 pred_mv(block, stride, x, y, i);
1208 block->u.mv[i][0] += dirac_get_arith_int(arith + 4 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
1209 block->u.mv[i][1] += dirac_get_arith_int(arith + 5 + 2 * i, CTX_MV_F1, CTX_MV_DATA);
1215 * Copies the current block to the other blocks covered by the current superblock split mode
1217 static void propagate_block_data(DiracBlock *block, int stride, int size)
1220 DiracBlock *dst = block;
1222 for (x = 1; x < size; x++)
1225 for (y = 1; y < size; y++) {
1227 for (x = 0; x < size; x++)
1233 * Dirac Specification ->
1234 * 12. Block motion data syntax
1236 static int dirac_unpack_block_motion_data(DiracContext *s)
1238 GetBitContext *gb = &s->gb;
1239 uint8_t *sbsplit = s->sbsplit;
1241 DiracArith arith[8];
1245 /* [DIRAC_STD] 11.2.4 and 12.2.1 Number of blocks and superblocks */
1246 s->sbwidth = DIVRNDUP(s->source.width, 4*s->plane[0].xbsep);
1247 s->sbheight = DIVRNDUP(s->source.height, 4*s->plane[0].ybsep);
1248 s->blwidth = 4 * s->sbwidth;
1249 s->blheight = 4 * s->sbheight;
1251 /* [DIRAC_STD] 12.3.1 Superblock splitting modes. superblock_split_modes()
1252 decode superblock split modes */
1253 ff_dirac_init_arith_decoder(arith, gb, svq3_get_ue_golomb(gb)); /* svq3_get_ue_golomb(gb) is the length */
1254 for (y = 0; y < s->sbheight; y++) {
1255 for (x = 0; x < s->sbwidth; x++) {
1256 unsigned int split = dirac_get_arith_uint(arith, CTX_SB_F1, CTX_SB_DATA);
1259 sbsplit[x] = (split + pred_sbsplit(sbsplit+x, s->sbwidth, x, y)) % 3;
1261 sbsplit += s->sbwidth;
1264 /* setup arith decoding */
1265 ff_dirac_init_arith_decoder(arith, gb, svq3_get_ue_golomb(gb));
1266 for (i = 0; i < s->num_refs; i++) {
1267 ff_dirac_init_arith_decoder(arith + 4 + 2 * i, gb, svq3_get_ue_golomb(gb));
1268 ff_dirac_init_arith_decoder(arith + 5 + 2 * i, gb, svq3_get_ue_golomb(gb));
1270 for (i = 0; i < 3; i++)
1271 ff_dirac_init_arith_decoder(arith+1+i, gb, svq3_get_ue_golomb(gb));
1273 for (y = 0; y < s->sbheight; y++)
1274 for (x = 0; x < s->sbwidth; x++) {
1275 int blkcnt = 1 << s->sbsplit[y * s->sbwidth + x];
1276 int step = 4 >> s->sbsplit[y * s->sbwidth + x];
1278 for (q = 0; q < blkcnt; q++)
1279 for (p = 0; p < blkcnt; p++) {
1280 int bx = 4 * x + p*step;
1281 int by = 4 * y + q*step;
1282 DiracBlock *block = &s->blmotion[by*s->blwidth + bx];
1283 decode_block_params(s, arith, block, s->blwidth, bx, by);
1284 propagate_block_data(block, s->blwidth, step);
1291 static int weight(int i, int blen, int offset)
1293 #define ROLLOFF(i) offset == 1 ? ((i) ? 5 : 3) : \
1294 (1 + (6*(i) + offset - 1) / (2*offset - 1))
1298 else if (i > blen-1 - 2*offset)
1299 return ROLLOFF(blen-1 - i);
1303 static void init_obmc_weight_row(Plane *p, uint8_t *obmc_weight, int stride,
1304 int left, int right, int wy)
1307 for (x = 0; left && x < p->xblen >> 1; x++)
1308 obmc_weight[x] = wy*8;
1309 for (; x < p->xblen >> right; x++)
1310 obmc_weight[x] = wy*weight(x, p->xblen, p->xoffset);
1311 for (; x < p->xblen; x++)
1312 obmc_weight[x] = wy*8;
1313 for (; x < stride; x++)
1317 static void init_obmc_weight(Plane *p, uint8_t *obmc_weight, int stride,
1318 int left, int right, int top, int bottom)
1321 for (y = 0; top && y < p->yblen >> 1; y++) {
1322 init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
1323 obmc_weight += stride;
1325 for (; y < p->yblen >> bottom; y++) {
1326 int wy = weight(y, p->yblen, p->yoffset);
1327 init_obmc_weight_row(p, obmc_weight, stride, left, right, wy);
1328 obmc_weight += stride;
1330 for (; y < p->yblen; y++) {
1331 init_obmc_weight_row(p, obmc_weight, stride, left, right, 8);
1332 obmc_weight += stride;
1336 static void init_obmc_weights(DiracContext *s, Plane *p, int by)
1339 int bottom = by == s->blheight-1;
1341 /* don't bother re-initing for rows 2 to blheight-2, the weights don't change */
1342 if (top || bottom || by == 1) {
1343 init_obmc_weight(p, s->obmc_weight[0], MAX_BLOCKSIZE, 1, 0, top, bottom);
1344 init_obmc_weight(p, s->obmc_weight[1], MAX_BLOCKSIZE, 0, 0, top, bottom);
1345 init_obmc_weight(p, s->obmc_weight[2], MAX_BLOCKSIZE, 0, 1, top, bottom);
1349 static const uint8_t epel_weights[4][4][4] = {
1369 * For block x,y, determine which of the hpel planes to do bilinear
1370 * interpolation from and set src[] to the location in each hpel plane
1373 * @return the index of the put_dirac_pixels_tab function to use
1374 * 0 for 1 plane (fpel,hpel), 1 for 2 planes (qpel), 2 for 4 planes (qpel), and 3 for epel
1376 static int mc_subpel(DiracContext *s, DiracBlock *block, const uint8_t *src[5],
1377 int x, int y, int ref, int plane)
1379 Plane *p = &s->plane[plane];
1380 uint8_t **ref_hpel = s->ref_pics[ref]->hpel[plane];
1381 int motion_x = block->u.mv[ref][0];
1382 int motion_y = block->u.mv[ref][1];
1383 int mx, my, i, epel, nplanes = 0;
1386 motion_x >>= s->chroma_x_shift;
1387 motion_y >>= s->chroma_y_shift;
1390 mx = motion_x & ~(-1U << s->mv_precision);
1391 my = motion_y & ~(-1U << s->mv_precision);
1392 motion_x >>= s->mv_precision;
1393 motion_y >>= s->mv_precision;
1394 /* normalize subpel coordinates to epel */
1395 /* TODO: template this function? */
1396 mx <<= 3 - s->mv_precision;
1397 my <<= 3 - s->mv_precision;
1406 src[0] = ref_hpel[(my>>1)+(mx>>2)] + y*p->stride + x;
1410 for (i = 0; i < 4; i++)
1411 src[i] = ref_hpel[i] + y*p->stride + x;
1413 /* if we're interpolating in the right/bottom halves, adjust the planes as needed
1414 we increment x/y because the edge changes for half of the pixels */
1421 src[0] += p->stride;
1422 src[1] += p->stride;
1430 /* check if we really only need 2 planes since either mx or my is
1431 a hpel position. (epel weights of 0 handle this there) */
1433 /* mx == 0: average [0] and [2]
1434 mx == 4: average [1] and [3] */
1435 src[!mx] = src[2 + !!mx];
1437 } else if (!(my&3)) {
1438 src[0] = src[(my>>1) ];
1439 src[1] = src[(my>>1)+1];
1443 /* adjust the ordering if needed so the weights work */
1445 FFSWAP(const uint8_t *, src[0], src[1]);
1446 FFSWAP(const uint8_t *, src[2], src[3]);
1449 FFSWAP(const uint8_t *, src[0], src[2]);
1450 FFSWAP(const uint8_t *, src[1], src[3]);
1452 src[4] = epel_weights[my&3][mx&3];
1456 /* fixme: v/h _edge_pos */
1457 if (x + p->xblen > p->width +EDGE_WIDTH/2 ||
1458 y + p->yblen > p->height+EDGE_WIDTH/2 ||
1460 for (i = 0; i < nplanes; i++) {
1461 s->vdsp.emulated_edge_mc(s->edge_emu_buffer[i], src[i],
1462 p->stride, p->stride,
1463 p->xblen, p->yblen, x, y,
1464 p->width+EDGE_WIDTH/2, p->height+EDGE_WIDTH/2);
1465 src[i] = s->edge_emu_buffer[i];
1468 return (nplanes>>1) + epel;
1471 static void add_dc(uint16_t *dst, int dc, int stride,
1472 uint8_t *obmc_weight, int xblen, int yblen)
1477 for (y = 0; y < yblen; y++) {
1478 for (x = 0; x < xblen; x += 2) {
1479 dst[x ] += dc * obmc_weight[x ];
1480 dst[x+1] += dc * obmc_weight[x+1];
1483 obmc_weight += MAX_BLOCKSIZE;
1487 static void block_mc(DiracContext *s, DiracBlock *block,
1488 uint16_t *mctmp, uint8_t *obmc_weight,
1489 int plane, int dstx, int dsty)
1491 Plane *p = &s->plane[plane];
1492 const uint8_t *src[5];
1495 switch (block->ref&3) {
1497 add_dc(mctmp, block->u.dc[plane], p->stride, obmc_weight, p->xblen, p->yblen);
1501 idx = mc_subpel(s, block, src, dstx, dsty, (block->ref&3)-1, plane);
1502 s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1504 s->weight_func(s->mcscratch, p->stride, s->weight_log2denom,
1505 s->weight[0] + s->weight[1], p->yblen);
1508 idx = mc_subpel(s, block, src, dstx, dsty, 0, plane);
1509 s->put_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1510 idx = mc_subpel(s, block, src, dstx, dsty, 1, plane);
1511 if (s->biweight_func) {
1512 /* fixme: +32 is a quick hack */
1513 s->put_pixels_tab[idx](s->mcscratch + 32, src, p->stride, p->yblen);
1514 s->biweight_func(s->mcscratch, s->mcscratch+32, p->stride, s->weight_log2denom,
1515 s->weight[0], s->weight[1], p->yblen);
1517 s->avg_pixels_tab[idx](s->mcscratch, src, p->stride, p->yblen);
1520 s->add_obmc(mctmp, s->mcscratch, p->stride, obmc_weight, p->yblen);
1523 static void mc_row(DiracContext *s, DiracBlock *block, uint16_t *mctmp, int plane, int dsty)
1525 Plane *p = &s->plane[plane];
1526 int x, dstx = p->xbsep - p->xoffset;
1528 block_mc(s, block, mctmp, s->obmc_weight[0], plane, -p->xoffset, dsty);
1531 for (x = 1; x < s->blwidth-1; x++) {
1532 block_mc(s, block+x, mctmp, s->obmc_weight[1], plane, dstx, dsty);
1536 block_mc(s, block+x, mctmp, s->obmc_weight[2], plane, dstx, dsty);
1539 static void select_dsp_funcs(DiracContext *s, int width, int height, int xblen, int yblen)
1547 memcpy(s->put_pixels_tab, s->diracdsp.put_dirac_pixels_tab[idx], sizeof(s->put_pixels_tab));
1548 memcpy(s->avg_pixels_tab, s->diracdsp.avg_dirac_pixels_tab[idx], sizeof(s->avg_pixels_tab));
1549 s->add_obmc = s->diracdsp.add_dirac_obmc[idx];
1550 if (s->weight_log2denom > 1 || s->weight[0] != 1 || s->weight[1] != 1) {
1551 s->weight_func = s->diracdsp.weight_dirac_pixels_tab[idx];
1552 s->biweight_func = s->diracdsp.biweight_dirac_pixels_tab[idx];
1554 s->weight_func = NULL;
1555 s->biweight_func = NULL;
1559 static void interpolate_refplane(DiracContext *s, DiracFrame *ref, int plane, int width, int height)
1561 /* chroma allocates an edge of 8 when subsampled
1562 which for 4:2:2 means an h edge of 16 and v edge of 8
1563 just use 8 for everything for the moment */
1564 int i, edge = EDGE_WIDTH/2;
1566 ref->hpel[plane][0] = ref->avframe->data[plane];
1567 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 */
1569 /* no need for hpel if we only have fpel vectors */
1570 if (!s->mv_precision)
1573 for (i = 1; i < 4; i++) {
1574 if (!ref->hpel_base[plane][i])
1575 ref->hpel_base[plane][i] = av_malloc((height+2*edge) * ref->avframe->linesize[plane] + 32);
1576 /* we need to be 16-byte aligned even for chroma */
1577 ref->hpel[plane][i] = ref->hpel_base[plane][i] + edge*ref->avframe->linesize[plane] + 16;
1580 if (!ref->interpolated[plane]) {
1581 s->diracdsp.dirac_hpel_filter(ref->hpel[plane][1], ref->hpel[plane][2],
1582 ref->hpel[plane][3], ref->hpel[plane][0],
1583 ref->avframe->linesize[plane], width, height);
1584 s->mpvencdsp.draw_edges(ref->hpel[plane][1], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1585 s->mpvencdsp.draw_edges(ref->hpel[plane][2], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1586 s->mpvencdsp.draw_edges(ref->hpel[plane][3], ref->avframe->linesize[plane], width, height, edge, edge, EDGE_TOP | EDGE_BOTTOM);
1588 ref->interpolated[plane] = 1;
1592 * Dirac Specification ->
1593 * 13.0 Transform data syntax. transform_data()
1595 static int dirac_decode_frame_internal(DiracContext *s)
1598 int y, i, comp, dsty;
1602 /* [DIRAC_STD] 13.5.1 low_delay_transform_data() */
1603 for (comp = 0; comp < 3; comp++) {
1604 Plane *p = &s->plane[comp];
1605 memset(p->idwt_buf, 0, p->idwt_stride * p->idwt_height * sizeof(IDWTELEM));
1608 if ((ret = decode_lowdelay(s)) < 0)
1613 for (comp = 0; comp < 3; comp++) {
1614 Plane *p = &s->plane[comp];
1615 uint8_t *frame = s->current_picture->avframe->data[comp];
1617 /* FIXME: small resolutions */
1618 for (i = 0; i < 4; i++)
1619 s->edge_emu_buffer[i] = s->edge_emu_buffer_base + i*FFALIGN(p->width, 16);
1621 if (!s->zero_res && !s->low_delay)
1623 memset(p->idwt_buf, 0, p->idwt_stride * p->idwt_height * sizeof(IDWTELEM));
1624 decode_component(s, comp); /* [DIRAC_STD] 13.4.1 core_transform_data() */
1626 if (ff_spatial_idwt_init2(&d, p->idwt_buf, p->idwt_width, p->idwt_height, p->idwt_stride,
1627 s->wavelet_idx+2, s->wavelet_depth, p->idwt_tmp))
1630 if (!s->num_refs) { /* intra */
1631 for (y = 0; y < p->height; y += 16) {
1632 ff_spatial_idwt_slice2(&d, y+16); /* decode */
1633 s->diracdsp.put_signed_rect_clamped(frame + y*p->stride, p->stride,
1634 p->idwt_buf + y*p->idwt_stride, p->idwt_stride, p->width, 16);
1636 } else { /* inter */
1637 int rowheight = p->ybsep*p->stride;
1639 select_dsp_funcs(s, p->width, p->height, p->xblen, p->yblen);
1641 for (i = 0; i < s->num_refs; i++)
1642 interpolate_refplane(s, s->ref_pics[i], comp, p->width, p->height);
1644 memset(s->mctmp, 0, 4*p->yoffset*p->stride);
1647 for (y = 0; y < s->blheight; y++) {
1649 start = FFMAX(dsty, 0);
1650 uint16_t *mctmp = s->mctmp + y*rowheight;
1651 DiracBlock *blocks = s->blmotion + y*s->blwidth;
1653 init_obmc_weights(s, p, y);
1655 if (y == s->blheight-1 || start+p->ybsep > p->height)
1656 h = p->height - start;
1658 h = p->ybsep - (start - dsty);
1662 memset(mctmp+2*p->yoffset*p->stride, 0, 2*rowheight);
1663 mc_row(s, blocks, mctmp, comp, dsty);
1665 mctmp += (start - dsty)*p->stride + p->xoffset;
1666 ff_spatial_idwt_slice2(&d, start + h); /* decode */
1667 s->diracdsp.add_rect_clamped(frame + start*p->stride, mctmp, p->stride,
1668 p->idwt_buf + start*p->idwt_stride, p->idwt_stride, p->width, h);
1679 static int get_buffer_with_edge(AVCodecContext *avctx, AVFrame *f, int flags)
1682 int chroma_x_shift, chroma_y_shift;
1683 avcodec_get_chroma_sub_sample(avctx->pix_fmt, &chroma_x_shift, &chroma_y_shift);
1685 f->width = avctx->width + 2 * EDGE_WIDTH;
1686 f->height = avctx->height + 2 * EDGE_WIDTH + 2;
1687 ret = ff_get_buffer(avctx, f, flags);
1691 for (i = 0; f->data[i]; i++) {
1692 int offset = (EDGE_WIDTH >> (i && i<3 ? chroma_y_shift : 0)) *
1693 f->linesize[i] + 32;
1694 f->data[i] += offset;
1696 f->width = avctx->width;
1697 f->height = avctx->height;
1703 * Dirac Specification ->
1704 * 11.1.1 Picture Header. picture_header()
1706 static int dirac_decode_picture_header(DiracContext *s)
1709 int i, j, refnum, refdist;
1710 GetBitContext *gb = &s->gb;
1712 /* [DIRAC_STD] 11.1.1 Picture Header. picture_header() PICTURE_NUM */
1713 picnum = s->current_picture->avframe->display_picture_number = get_bits_long(gb, 32);
1716 av_log(s->avctx,AV_LOG_DEBUG,"PICTURE_NUM: %d\n",picnum);
1718 /* if this is the first keyframe after a sequence header, start our
1719 reordering from here */
1720 if (s->frame_number < 0)
1721 s->frame_number = picnum;
1723 s->ref_pics[0] = s->ref_pics[1] = NULL;
1724 for (i = 0; i < s->num_refs; i++) {
1725 refnum = picnum + dirac_get_se_golomb(gb);
1728 /* find the closest reference to the one we want */
1729 /* Jordi: this is needed if the referenced picture hasn't yet arrived */
1730 for (j = 0; j < MAX_REFERENCE_FRAMES && refdist; j++)
1731 if (s->ref_frames[j]
1732 && FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum) < refdist) {
1733 s->ref_pics[i] = s->ref_frames[j];
1734 refdist = FFABS(s->ref_frames[j]->avframe->display_picture_number - refnum);
1737 if (!s->ref_pics[i] || refdist)
1738 av_log(s->avctx, AV_LOG_DEBUG, "Reference not found\n");
1740 /* if there were no references at all, allocate one */
1741 if (!s->ref_pics[i])
1742 for (j = 0; j < MAX_FRAMES; j++)
1743 if (!s->all_frames[j].avframe->data[0]) {
1744 s->ref_pics[i] = &s->all_frames[j];
1745 get_buffer_with_edge(s->avctx, s->ref_pics[i]->avframe, AV_GET_BUFFER_FLAG_REF);
1749 if (!s->ref_pics[i]) {
1750 av_log(s->avctx, AV_LOG_ERROR, "Reference could not be allocated\n");
1756 /* retire the reference frames that are not used anymore */
1757 if (s->current_picture->avframe->reference) {
1758 retire = picnum + dirac_get_se_golomb(gb);
1759 if (retire != picnum) {
1760 DiracFrame *retire_pic = remove_frame(s->ref_frames, retire);
1763 retire_pic->avframe->reference &= DELAYED_PIC_REF;
1765 av_log(s->avctx, AV_LOG_DEBUG, "Frame to retire not found\n");
1768 /* if reference array is full, remove the oldest as per the spec */
1769 while (add_frame(s->ref_frames, MAX_REFERENCE_FRAMES, s->current_picture)) {
1770 av_log(s->avctx, AV_LOG_ERROR, "Reference frame overflow\n");
1771 remove_frame(s->ref_frames, s->ref_frames[0]->avframe->display_picture_number)->avframe->reference &= DELAYED_PIC_REF;
1776 if (dirac_unpack_prediction_parameters(s)) /* [DIRAC_STD] 11.2 Picture Prediction Data. picture_prediction() */
1778 if (dirac_unpack_block_motion_data(s)) /* [DIRAC_STD] 12. Block motion data syntax */
1781 if (dirac_unpack_idwt_params(s)) /* [DIRAC_STD] 11.3 Wavelet transform data */
1788 static int get_delayed_pic(DiracContext *s, AVFrame *picture, int *got_frame)
1790 DiracFrame *out = s->delay_frames[0];
1794 /* find frame with lowest picture number */
1795 for (i = 1; s->delay_frames[i]; i++)
1796 if (s->delay_frames[i]->avframe->display_picture_number < out->avframe->display_picture_number) {
1797 out = s->delay_frames[i];
1801 for (i = out_idx; s->delay_frames[i]; i++)
1802 s->delay_frames[i] = s->delay_frames[i+1];
1805 out->avframe->reference ^= DELAYED_PIC_REF;
1807 if((ret = av_frame_ref(picture, out->avframe)) < 0)
1815 * Dirac Specification ->
1816 * 9.6 Parse Info Header Syntax. parse_info()
1817 * 4 byte start code + byte parse code + 4 byte size + 4 byte previous size
1819 #define DATA_UNIT_HEADER_SIZE 13
1821 /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3
1822 inside the function parse_sequence() */
1823 static int dirac_decode_data_unit(AVCodecContext *avctx, const uint8_t *buf, int size)
1825 DiracContext *s = avctx->priv_data;
1826 DiracFrame *pic = NULL;
1827 int ret, i, parse_code;
1830 if (size < DATA_UNIT_HEADER_SIZE)
1833 parse_code = buf[4];
1835 init_get_bits(&s->gb, &buf[13], 8*(size - DATA_UNIT_HEADER_SIZE));
1837 if (parse_code == pc_seq_header) {
1838 if (s->seen_sequence_header)
1841 /* [DIRAC_STD] 10. Sequence header */
1842 if (avpriv_dirac_parse_sequence_header(avctx, &s->gb, &s->source))
1845 avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_x_shift, &s->chroma_y_shift);
1847 if (alloc_sequence_buffers(s))
1850 s->seen_sequence_header = 1;
1851 } else if (parse_code == pc_eos) { /* [DIRAC_STD] End of Sequence */
1852 free_sequence_buffers(s);
1853 s->seen_sequence_header = 0;
1854 } else if (parse_code == pc_aux_data) {
1855 if (buf[13] == 1) { /* encoder implementation/version */
1857 /* versions older than 1.0.8 don't store quant delta for
1858 subbands with only one codeblock */
1859 if (sscanf(buf+14, "Schroedinger %d.%d.%d", ver, ver+1, ver+2) == 3)
1860 if (ver[0] == 1 && ver[1] == 0 && ver[2] <= 7)
1861 s->old_delta_quant = 1;
1863 } else if (parse_code & 0x8) { /* picture data unit */
1864 if (!s->seen_sequence_header) {
1865 av_log(avctx, AV_LOG_DEBUG, "Dropping frame without sequence header\n");
1869 /* find an unused frame */
1870 for (i = 0; i < MAX_FRAMES; i++)
1871 if (s->all_frames[i].avframe->data[0] == NULL)
1872 pic = &s->all_frames[i];
1874 av_log(avctx, AV_LOG_ERROR, "framelist full\n");
1878 av_frame_unref(pic->avframe);
1880 /* [DIRAC_STD] Defined in 9.6.1 ... */
1881 tmp = parse_code & 0x03; /* [DIRAC_STD] num_refs() */
1883 av_log(avctx, AV_LOG_ERROR, "num_refs of 3\n");
1887 s->is_arith = (parse_code & 0x48) == 0x08; /* [DIRAC_STD] using_ac() */
1888 s->low_delay = (parse_code & 0x88) == 0x88; /* [DIRAC_STD] is_low_delay() */
1889 pic->avframe->reference = (parse_code & 0x0C) == 0x0C; /* [DIRAC_STD] is_reference() */
1890 pic->avframe->key_frame = s->num_refs == 0; /* [DIRAC_STD] is_intra() */
1891 pic->avframe->pict_type = s->num_refs + 1; /* Definition of AVPictureType in avutil.h */
1893 if ((ret = get_buffer_with_edge(avctx, pic->avframe, (parse_code & 0x0C) == 0x0C ? AV_GET_BUFFER_FLAG_REF : 0)) < 0)
1895 s->current_picture = pic;
1896 s->plane[0].stride = pic->avframe->linesize[0];
1897 s->plane[1].stride = pic->avframe->linesize[1];
1898 s->plane[2].stride = pic->avframe->linesize[2];
1900 if (alloc_buffers(s, FFMAX3(FFABS(s->plane[0].stride), FFABS(s->plane[1].stride), FFABS(s->plane[2].stride))) < 0)
1901 return AVERROR(ENOMEM);
1903 /* [DIRAC_STD] 11.1 Picture parse. picture_parse() */
1904 if (dirac_decode_picture_header(s))
1907 /* [DIRAC_STD] 13.0 Transform data syntax. transform_data() */
1908 if (dirac_decode_frame_internal(s))
1914 static int dirac_decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPacket *pkt)
1916 DiracContext *s = avctx->priv_data;
1917 AVFrame *picture = data;
1918 uint8_t *buf = pkt->data;
1919 int buf_size = pkt->size;
1920 int i, data_unit_size, buf_idx = 0;
1923 /* release unused frames */
1924 for (i = 0; i < MAX_FRAMES; i++)
1925 if (s->all_frames[i].avframe->data[0] && !s->all_frames[i].avframe->reference) {
1926 av_frame_unref(s->all_frames[i].avframe);
1927 memset(s->all_frames[i].interpolated, 0, sizeof(s->all_frames[i].interpolated));
1930 s->current_picture = NULL;
1933 /* end of stream, so flush delayed pics */
1935 return get_delayed_pic(s, (AVFrame *)data, got_frame);
1938 /*[DIRAC_STD] Here starts the code from parse_info() defined in 9.6
1939 [DIRAC_STD] PARSE_INFO_PREFIX = "BBCD" as defined in ISO/IEC 646
1940 BBCD start code search */
1941 for (; buf_idx + DATA_UNIT_HEADER_SIZE < buf_size; buf_idx++) {
1942 if (buf[buf_idx ] == 'B' && buf[buf_idx+1] == 'B' &&
1943 buf[buf_idx+2] == 'C' && buf[buf_idx+3] == 'D')
1946 /* BBCD found or end of data */
1947 if (buf_idx + DATA_UNIT_HEADER_SIZE >= buf_size)
1950 data_unit_size = AV_RB32(buf+buf_idx+5);
1951 if (data_unit_size > buf_size - buf_idx || !data_unit_size) {
1952 if(data_unit_size > buf_size - buf_idx)
1953 av_log(s->avctx, AV_LOG_ERROR,
1954 "Data unit with size %d is larger than input buffer, discarding\n",
1959 /* [DIRAC_STD] dirac_decode_data_unit makes reference to the while defined in 9.3 inside the function parse_sequence() */
1960 if (dirac_decode_data_unit(avctx, buf+buf_idx, data_unit_size))
1962 av_log(s->avctx, AV_LOG_ERROR,"Error in dirac_decode_data_unit\n");
1965 buf_idx += data_unit_size;
1968 if (!s->current_picture)
1971 if (s->current_picture->avframe->display_picture_number > s->frame_number) {
1972 DiracFrame *delayed_frame = remove_frame(s->delay_frames, s->frame_number);
1974 s->current_picture->avframe->reference |= DELAYED_PIC_REF;
1976 if (add_frame(s->delay_frames, MAX_DELAY, s->current_picture)) {
1977 int min_num = s->delay_frames[0]->avframe->display_picture_number;
1978 /* Too many delayed frames, so we display the frame with the lowest pts */
1979 av_log(avctx, AV_LOG_ERROR, "Delay frame overflow\n");
1981 for (i = 1; s->delay_frames[i]; i++)
1982 if (s->delay_frames[i]->avframe->display_picture_number < min_num)
1983 min_num = s->delay_frames[i]->avframe->display_picture_number;
1985 delayed_frame = remove_frame(s->delay_frames, min_num);
1986 add_frame(s->delay_frames, MAX_DELAY, s->current_picture);
1989 if (delayed_frame) {
1990 delayed_frame->avframe->reference ^= DELAYED_PIC_REF;
1991 if((ret=av_frame_ref(data, delayed_frame->avframe)) < 0)
1995 } else if (s->current_picture->avframe->display_picture_number == s->frame_number) {
1996 /* The right frame at the right time :-) */
1997 if((ret=av_frame_ref(data, s->current_picture->avframe)) < 0)
2003 s->frame_number = picture->display_picture_number + 1;
2008 AVCodec ff_dirac_decoder = {
2010 .long_name = NULL_IF_CONFIG_SMALL("BBC Dirac VC-2"),
2011 .type = AVMEDIA_TYPE_VIDEO,
2012 .id = AV_CODEC_ID_DIRAC,
2013 .priv_data_size = sizeof(DiracContext),
2014 .init = dirac_decode_init,
2015 .close = dirac_decode_end,
2016 .decode = dirac_decode_frame,
2017 .capabilities = CODEC_CAP_DELAY,
2018 .flush = dirac_decode_flush,