2 * Copyright (C) 2006 Aurelien Jacobs <aurel@gnuage.org>
4 * This file is part of Libav.
6 * Libav is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * Libav is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with Libav; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * VP5 and VP6 compatible video decoder (common features)
27 #include "bytestream.h"
34 void ff_vp56_init_dequant(VP56Context *s, int quantizer)
36 s->quantizer = quantizer;
37 s->dequant_dc = vp56_dc_dequant[quantizer] << 2;
38 s->dequant_ac = vp56_ac_dequant[quantizer] << 2;
39 memset(s->qscale_table, quantizer, s->mb_width);
42 static int vp56_get_vectors_predictors(VP56Context *s, int row, int col,
46 VP56mv vect[2] = {{0,0}, {0,0}};
50 for (pos=0; pos<12; pos++) {
51 mvp.x = col + vp56_candidate_predictor_pos[pos][0];
52 mvp.y = row + vp56_candidate_predictor_pos[pos][1];
53 if (mvp.x < 0 || mvp.x >= s->mb_width ||
54 mvp.y < 0 || mvp.y >= s->mb_height)
56 offset = mvp.x + s->mb_width*mvp.y;
58 if (vp56_reference_frame[s->macroblocks[offset].type] != ref_frame)
60 if ((s->macroblocks[offset].mv.x == vect[0].x &&
61 s->macroblocks[offset].mv.y == vect[0].y) ||
62 (s->macroblocks[offset].mv.x == 0 &&
63 s->macroblocks[offset].mv.y == 0))
66 vect[nb_pred++] = s->macroblocks[offset].mv;
71 s->vector_candidate_pos = pos;
74 s->vector_candidate[0] = vect[0];
75 s->vector_candidate[1] = vect[1];
80 static void vp56_parse_mb_type_models(VP56Context *s)
82 VP56RangeCoder *c = &s->c;
83 VP56Model *model = s->modelp;
86 for (ctx=0; ctx<3; ctx++) {
87 if (vp56_rac_get_prob(c, 174)) {
88 int idx = vp56_rac_gets(c, 4);
89 memcpy(model->mb_types_stats[ctx],
90 vp56_pre_def_mb_type_stats[idx][ctx],
91 sizeof(model->mb_types_stats[ctx]));
93 if (vp56_rac_get_prob(c, 254)) {
94 for (type=0; type<10; type++) {
96 if (vp56_rac_get_prob(c, 205)) {
97 int delta, sign = vp56_rac_get(c);
99 delta = vp56_rac_get_tree(c, vp56_pmbtm_tree,
100 vp56_mb_type_model_model);
102 delta = 4 * vp56_rac_gets(c, 7);
103 model->mb_types_stats[ctx][type][i] += (delta ^ -sign) + sign;
110 /* compute MB type probability tables based on previous MB type */
111 for (ctx=0; ctx<3; ctx++) {
114 for (type=0; type<10; type++)
115 p[type] = 100 * model->mb_types_stats[ctx][type][1];
117 for (type=0; type<10; type++) {
118 int p02, p34, p0234, p17, p56, p89, p5689, p156789;
120 /* conservative MB type probability */
121 model->mb_type[ctx][type][0] = 255 - (255 * model->mb_types_stats[ctx][type][0]) / (1 + model->mb_types_stats[ctx][type][0] + model->mb_types_stats[ctx][type][1]);
123 p[type] = 0; /* same MB type => weight is null */
125 /* binary tree parsing probabilities */
133 p156789 = p17 + p5689;
135 model->mb_type[ctx][type][1] = 1 + 255 * p0234/(1+p0234+p156789);
136 model->mb_type[ctx][type][2] = 1 + 255 * p02 / (1+p0234);
137 model->mb_type[ctx][type][3] = 1 + 255 * p17 / (1+p156789);
138 model->mb_type[ctx][type][4] = 1 + 255 * p[0] / (1+p02);
139 model->mb_type[ctx][type][5] = 1 + 255 * p[3] / (1+p34);
140 model->mb_type[ctx][type][6] = 1 + 255 * p[1] / (1+p17);
141 model->mb_type[ctx][type][7] = 1 + 255 * p56 / (1+p5689);
142 model->mb_type[ctx][type][8] = 1 + 255 * p[5] / (1+p56);
143 model->mb_type[ctx][type][9] = 1 + 255 * p[8] / (1+p89);
145 /* restore initial value */
146 p[type] = 100 * model->mb_types_stats[ctx][type][1];
151 static VP56mb vp56_parse_mb_type(VP56Context *s,
152 VP56mb prev_type, int ctx)
154 uint8_t *mb_type_model = s->modelp->mb_type[ctx][prev_type];
155 VP56RangeCoder *c = &s->c;
157 if (vp56_rac_get_prob(c, mb_type_model[0]))
160 return vp56_rac_get_tree(c, vp56_pmbt_tree, mb_type_model);
163 static void vp56_decode_4mv(VP56Context *s, int row, int col)
169 /* parse each block type */
170 for (b=0; b<4; b++) {
171 type[b] = vp56_rac_gets(&s->c, 2);
173 type[b]++; /* only returns 0, 2, 3 or 4 (all INTER_PF) */
177 for (b=0; b<4; b++) {
179 case VP56_MB_INTER_NOVEC_PF:
180 s->mv[b] = (VP56mv) {0,0};
182 case VP56_MB_INTER_DELTA_PF:
183 s->parse_vector_adjustment(s, &s->mv[b]);
185 case VP56_MB_INTER_V1_PF:
186 s->mv[b] = s->vector_candidate[0];
188 case VP56_MB_INTER_V2_PF:
189 s->mv[b] = s->vector_candidate[1];
196 /* this is the one selected for the whole MB for prediction */
197 s->macroblocks[row * s->mb_width + col].mv = s->mv[3];
199 /* chroma vectors are average luma vectors */
200 if (s->avctx->codec->id == AV_CODEC_ID_VP5) {
201 s->mv[4].x = s->mv[5].x = RSHIFT(mv.x,2);
202 s->mv[4].y = s->mv[5].y = RSHIFT(mv.y,2);
204 s->mv[4] = s->mv[5] = (VP56mv) {mv.x/4, mv.y/4};
208 static VP56mb vp56_decode_mv(VP56Context *s, int row, int col)
210 VP56mv *mv, vect = {0,0};
213 ctx = vp56_get_vectors_predictors(s, row, col, VP56_FRAME_PREVIOUS);
214 s->mb_type = vp56_parse_mb_type(s, s->mb_type, ctx);
215 s->macroblocks[row * s->mb_width + col].type = s->mb_type;
217 switch (s->mb_type) {
218 case VP56_MB_INTER_V1_PF:
219 mv = &s->vector_candidate[0];
222 case VP56_MB_INTER_V2_PF:
223 mv = &s->vector_candidate[1];
226 case VP56_MB_INTER_V1_GF:
227 vp56_get_vectors_predictors(s, row, col, VP56_FRAME_GOLDEN);
228 mv = &s->vector_candidate[0];
231 case VP56_MB_INTER_V2_GF:
232 vp56_get_vectors_predictors(s, row, col, VP56_FRAME_GOLDEN);
233 mv = &s->vector_candidate[1];
236 case VP56_MB_INTER_DELTA_PF:
237 s->parse_vector_adjustment(s, &vect);
241 case VP56_MB_INTER_DELTA_GF:
242 vp56_get_vectors_predictors(s, row, col, VP56_FRAME_GOLDEN);
243 s->parse_vector_adjustment(s, &vect);
247 case VP56_MB_INTER_4V:
248 vp56_decode_4mv(s, row, col);
256 s->macroblocks[row*s->mb_width + col].mv = *mv;
258 /* same vector for all blocks */
265 static void vp56_add_predictors_dc(VP56Context *s, VP56Frame ref_frame)
267 int idx = s->scantable.permutated[0];
270 for (b=0; b<6; b++) {
271 VP56RefDc *ab = &s->above_blocks[s->above_block_idx[b]];
272 VP56RefDc *lb = &s->left_block[ff_vp56_b6to4[b]];
277 if (ref_frame == lb->ref_frame) {
281 if (ref_frame == ab->ref_frame) {
285 if (s->avctx->codec->id == AV_CODEC_ID_VP5)
287 if (count < 2 && ref_frame == ab[-1+2*i].ref_frame) {
288 dc += ab[-1+2*i].dc_coeff;
292 dc = s->prev_dc[ff_vp56_b2p[b]][ref_frame];
296 s->block_coeff[b][idx] += dc;
297 s->prev_dc[ff_vp56_b2p[b]][ref_frame] = s->block_coeff[b][idx];
298 ab->dc_coeff = s->block_coeff[b][idx];
299 ab->ref_frame = ref_frame;
300 lb->dc_coeff = s->block_coeff[b][idx];
301 lb->ref_frame = ref_frame;
302 s->block_coeff[b][idx] *= s->dequant_dc;
306 static void vp56_deblock_filter(VP56Context *s, uint8_t *yuv,
307 int stride, int dx, int dy)
309 int t = vp56_filter_threshold[s->quantizer];
310 if (dx) s->vp56dsp.edge_filter_hor(yuv + 10-dx , stride, t);
311 if (dy) s->vp56dsp.edge_filter_ver(yuv + stride*(10-dy), stride, t);
314 static void vp56_mc(VP56Context *s, int b, int plane, uint8_t *src,
315 int stride, int x, int y)
317 uint8_t *dst=s->framep[VP56_FRAME_CURRENT]->data[plane]+s->block_offset[b];
320 int overlap_offset = 0;
321 int mask = s->vp56_coord_div[b] - 1;
322 int deblock_filtering = s->deblock_filtering;
326 if (s->avctx->skip_loop_filter >= AVDISCARD_ALL ||
327 (s->avctx->skip_loop_filter >= AVDISCARD_NONKEY
328 && !s->framep[VP56_FRAME_CURRENT]->key_frame))
329 deblock_filtering = 0;
331 dx = s->mv[b].x / s->vp56_coord_div[b];
332 dy = s->mv[b].y / s->vp56_coord_div[b];
341 if (x<0 || x+12>=s->plane_width[plane] ||
342 y<0 || y+12>=s->plane_height[plane]) {
343 s->dsp.emulated_edge_mc(s->edge_emu_buffer,
344 src + s->block_offset[b] + (dy-2)*stride + (dx-2),
345 stride, 12, 12, x, y,
346 s->plane_width[plane],
347 s->plane_height[plane]);
348 src_block = s->edge_emu_buffer;
349 src_offset = 2 + 2*stride;
350 } else if (deblock_filtering) {
351 /* only need a 12x12 block, but there is no such dsp function, */
352 /* so copy a 16x12 block */
353 s->dsp.put_pixels_tab[0][0](s->edge_emu_buffer,
354 src + s->block_offset[b] + (dy-2)*stride + (dx-2),
356 src_block = s->edge_emu_buffer;
357 src_offset = 2 + 2*stride;
360 src_offset = s->block_offset[b] + dy*stride + dx;
363 if (deblock_filtering)
364 vp56_deblock_filter(s, src_block, stride, dx&7, dy&7);
366 if (s->mv[b].x & mask)
367 overlap_offset += (s->mv[b].x > 0) ? 1 : -1;
368 if (s->mv[b].y & mask)
369 overlap_offset += (s->mv[b].y > 0) ? stride : -stride;
371 if (overlap_offset) {
373 s->filter(s, dst, src_block, src_offset, src_offset+overlap_offset,
374 stride, s->mv[b], mask, s->filter_selection, b<4);
376 s->dsp.put_no_rnd_pixels_l2[1](dst, src_block+src_offset,
377 src_block+src_offset+overlap_offset,
380 s->dsp.put_pixels_tab[1][0](dst, src_block+src_offset, stride, 8);
384 static void vp56_decode_mb(VP56Context *s, int row, int col, int is_alpha)
386 AVFrame *frame_current, *frame_ref;
389 int b, ab, b_max, plane, off;
391 if (s->framep[VP56_FRAME_CURRENT]->key_frame)
392 mb_type = VP56_MB_INTRA;
394 mb_type = vp56_decode_mv(s, row, col);
395 ref_frame = vp56_reference_frame[mb_type];
397 s->dsp.clear_blocks(*s->block_coeff);
401 vp56_add_predictors_dc(s, ref_frame);
403 frame_current = s->framep[VP56_FRAME_CURRENT];
404 frame_ref = s->framep[ref_frame];
405 if (mb_type != VP56_MB_INTRA && !frame_ref->data[0])
409 b_max = 6 - 2*is_alpha;
413 for (b=0; b<b_max; b++) {
414 plane = ff_vp56_b2p[b+ab];
415 s->vp3dsp.idct_put(frame_current->data[plane] + s->block_offset[b],
416 s->stride[plane], s->block_coeff[b]);
420 case VP56_MB_INTER_NOVEC_PF:
421 case VP56_MB_INTER_NOVEC_GF:
422 for (b=0; b<b_max; b++) {
423 plane = ff_vp56_b2p[b+ab];
424 off = s->block_offset[b];
425 s->dsp.put_pixels_tab[1][0](frame_current->data[plane] + off,
426 frame_ref->data[plane] + off,
427 s->stride[plane], 8);
428 s->vp3dsp.idct_add(frame_current->data[plane] + off,
429 s->stride[plane], s->block_coeff[b]);
433 case VP56_MB_INTER_DELTA_PF:
434 case VP56_MB_INTER_V1_PF:
435 case VP56_MB_INTER_V2_PF:
436 case VP56_MB_INTER_DELTA_GF:
437 case VP56_MB_INTER_4V:
438 case VP56_MB_INTER_V1_GF:
439 case VP56_MB_INTER_V2_GF:
440 for (b=0; b<b_max; b++) {
441 int x_off = b==1 || b==3 ? 8 : 0;
442 int y_off = b==2 || b==3 ? 8 : 0;
443 plane = ff_vp56_b2p[b+ab];
444 vp56_mc(s, b, plane, frame_ref->data[plane], s->stride[plane],
445 16*col+x_off, 16*row+y_off);
446 s->vp3dsp.idct_add(frame_current->data[plane] + s->block_offset[b],
447 s->stride[plane], s->block_coeff[b]);
453 static int vp56_size_changed(AVCodecContext *avctx)
455 VP56Context *s = avctx->priv_data;
456 int stride = s->framep[VP56_FRAME_CURRENT]->linesize[0];
459 s->plane_width[0] = s->plane_width[3] = avctx->coded_width;
460 s->plane_width[1] = s->plane_width[2] = avctx->coded_width/2;
461 s->plane_height[0] = s->plane_height[3] = avctx->coded_height;
462 s->plane_height[1] = s->plane_height[2] = avctx->coded_height/2;
465 s->stride[i] = s->flip * s->framep[VP56_FRAME_CURRENT]->linesize[i];
467 s->mb_width = (avctx->coded_width +15) / 16;
468 s->mb_height = (avctx->coded_height+15) / 16;
470 if (s->mb_width > 1000 || s->mb_height > 1000) {
471 avcodec_set_dimensions(avctx, 0, 0);
472 av_log(avctx, AV_LOG_ERROR, "picture too big\n");
476 s->qscale_table = av_realloc(s->qscale_table, s->mb_width);
477 s->above_blocks = av_realloc(s->above_blocks,
478 (4*s->mb_width+6) * sizeof(*s->above_blocks));
479 s->macroblocks = av_realloc(s->macroblocks,
480 s->mb_width*s->mb_height*sizeof(*s->macroblocks));
481 av_free(s->edge_emu_buffer_alloc);
482 s->edge_emu_buffer_alloc = av_malloc(16*stride);
483 s->edge_emu_buffer = s->edge_emu_buffer_alloc;
485 s->edge_emu_buffer += 15 * stride;
490 int ff_vp56_decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
493 const uint8_t *buf = avpkt->data;
494 VP56Context *s = avctx->priv_data;
495 AVFrame *const p = s->framep[VP56_FRAME_CURRENT];
496 int remaining_buf_size = avpkt->size;
497 int is_alpha, av_uninit(alpha_offset);
500 if (remaining_buf_size < 3)
502 alpha_offset = bytestream_get_be24(&buf);
503 remaining_buf_size -= 3;
504 if (remaining_buf_size < alpha_offset)
508 for (is_alpha=0; is_alpha < 1+s->has_alpha; is_alpha++) {
509 int mb_row, mb_col, mb_row_flip, mb_offset = 0;
510 int block, y, uv, stride_y, stride_uv;
511 int golden_frame = 0;
514 s->modelp = &s->models[is_alpha];
516 res = s->parse_header(s, buf, remaining_buf_size, &golden_frame);
519 for (i = 0; i < 4; i++) {
520 if (s->frames[i].data[0])
521 avctx->release_buffer(avctx, &s->frames[i]);
526 if (res == VP56_SIZE_CHANGE) {
528 for (i = 0; i < 4; i++) {
529 if (s->frames[i].data[0])
530 avctx->release_buffer(avctx, &s->frames[i]);
533 avcodec_set_dimensions(avctx, 0, 0);
540 if (ff_get_buffer(avctx, p) < 0) {
541 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
545 if (res == VP56_SIZE_CHANGE)
546 if (vp56_size_changed(avctx)) {
547 avctx->release_buffer(avctx, p);
553 p->pict_type = AV_PICTURE_TYPE_I;
554 s->default_models_init(s);
555 for (block=0; block<s->mb_height*s->mb_width; block++)
556 s->macroblocks[block].type = VP56_MB_INTRA;
558 p->pict_type = AV_PICTURE_TYPE_P;
559 vp56_parse_mb_type_models(s);
560 s->parse_vector_models(s);
561 s->mb_type = VP56_MB_INTER_NOVEC_PF;
564 if (s->parse_coeff_models(s))
567 memset(s->prev_dc, 0, sizeof(s->prev_dc));
568 s->prev_dc[1][VP56_FRAME_CURRENT] = 128;
569 s->prev_dc[2][VP56_FRAME_CURRENT] = 128;
571 for (block=0; block < 4*s->mb_width+6; block++) {
572 s->above_blocks[block].ref_frame = VP56_FRAME_NONE;
573 s->above_blocks[block].dc_coeff = 0;
574 s->above_blocks[block].not_null_dc = 0;
576 s->above_blocks[2*s->mb_width + 2].ref_frame = VP56_FRAME_CURRENT;
577 s->above_blocks[3*s->mb_width + 4].ref_frame = VP56_FRAME_CURRENT;
579 stride_y = p->linesize[0];
580 stride_uv = p->linesize[1];
585 /* main macroblocks loop */
586 for (mb_row=0; mb_row<s->mb_height; mb_row++) {
588 mb_row_flip = s->mb_height - mb_row - 1;
590 mb_row_flip = mb_row;
592 for (block=0; block<4; block++) {
593 s->left_block[block].ref_frame = VP56_FRAME_NONE;
594 s->left_block[block].dc_coeff = 0;
595 s->left_block[block].not_null_dc = 0;
597 memset(s->coeff_ctx, 0, sizeof(s->coeff_ctx));
598 memset(s->coeff_ctx_last, 24, sizeof(s->coeff_ctx_last));
600 s->above_block_idx[0] = 1;
601 s->above_block_idx[1] = 2;
602 s->above_block_idx[2] = 1;
603 s->above_block_idx[3] = 2;
604 s->above_block_idx[4] = 2*s->mb_width + 2 + 1;
605 s->above_block_idx[5] = 3*s->mb_width + 4 + 1;
607 s->block_offset[s->frbi] = (mb_row_flip*16 + mb_offset) * stride_y;
608 s->block_offset[s->srbi] = s->block_offset[s->frbi] + 8*stride_y;
609 s->block_offset[1] = s->block_offset[0] + 8;
610 s->block_offset[3] = s->block_offset[2] + 8;
611 s->block_offset[4] = (mb_row_flip*8 + mb_offset) * stride_uv;
612 s->block_offset[5] = s->block_offset[4];
614 for (mb_col=0; mb_col<s->mb_width; mb_col++) {
615 vp56_decode_mb(s, mb_row, mb_col, is_alpha);
617 for (y=0; y<4; y++) {
618 s->above_block_idx[y] += 2;
619 s->block_offset[y] += 16;
622 for (uv=4; uv<6; uv++) {
623 s->above_block_idx[uv] += 1;
624 s->block_offset[uv] += 8;
630 if (p->key_frame || golden_frame) {
631 if (s->framep[VP56_FRAME_GOLDEN]->data[0] &&
632 s->framep[VP56_FRAME_GOLDEN] != s->framep[VP56_FRAME_GOLDEN2])
633 avctx->release_buffer(avctx, s->framep[VP56_FRAME_GOLDEN]);
634 s->framep[VP56_FRAME_GOLDEN] = p;
638 FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN],
639 s->framep[VP56_FRAME_GOLDEN2]);
641 remaining_buf_size -= alpha_offset;
645 if (s->framep[VP56_FRAME_PREVIOUS] == s->framep[VP56_FRAME_GOLDEN] ||
646 s->framep[VP56_FRAME_PREVIOUS] == s->framep[VP56_FRAME_GOLDEN2]) {
647 if (s->framep[VP56_FRAME_UNUSED] != s->framep[VP56_FRAME_GOLDEN] &&
648 s->framep[VP56_FRAME_UNUSED] != s->framep[VP56_FRAME_GOLDEN2])
649 FFSWAP(AVFrame *, s->framep[VP56_FRAME_PREVIOUS],
650 s->framep[VP56_FRAME_UNUSED]);
652 FFSWAP(AVFrame *, s->framep[VP56_FRAME_PREVIOUS],
653 s->framep[VP56_FRAME_UNUSED2]);
654 } else if (s->framep[VP56_FRAME_PREVIOUS]->data[0])
655 avctx->release_buffer(avctx, s->framep[VP56_FRAME_PREVIOUS]);
656 FFSWAP(AVFrame *, s->framep[VP56_FRAME_CURRENT],
657 s->framep[VP56_FRAME_PREVIOUS]);
660 p->qscale_table = s->qscale_table;
661 p->qscale_type = FF_QSCALE_TYPE_VP56;
662 *(AVFrame*)data = *p;
668 av_cold void ff_vp56_init(AVCodecContext *avctx, int flip, int has_alpha)
670 VP56Context *s = avctx->priv_data;
674 avctx->pix_fmt = has_alpha ? AV_PIX_FMT_YUVA420P : AV_PIX_FMT_YUV420P;
676 ff_dsputil_init(&s->dsp, avctx);
677 ff_vp3dsp_init(&s->vp3dsp, avctx->flags);
678 ff_vp56dsp_init(&s->vp56dsp, avctx->codec->id);
679 ff_init_scantable_permutation(s->dsp.idct_permutation, s->vp3dsp.idct_perm);
680 ff_init_scantable(s->dsp.idct_permutation, &s->scantable,ff_zigzag_direct);
683 s->framep[i] = &s->frames[i];
684 s->framep[VP56_FRAME_UNUSED] = s->framep[VP56_FRAME_GOLDEN];
685 s->framep[VP56_FRAME_UNUSED2] = s->framep[VP56_FRAME_GOLDEN2];
686 s->edge_emu_buffer_alloc = NULL;
688 s->above_blocks = NULL;
689 s->macroblocks = NULL;
691 s->deblock_filtering = 1;
695 s->has_alpha = has_alpha;
707 av_cold int ff_vp56_free(AVCodecContext *avctx)
709 VP56Context *s = avctx->priv_data;
711 av_freep(&s->qscale_table);
712 av_freep(&s->above_blocks);
713 av_freep(&s->macroblocks);
714 av_freep(&s->edge_emu_buffer_alloc);
715 if (s->framep[VP56_FRAME_GOLDEN]->data[0])
716 avctx->release_buffer(avctx, s->framep[VP56_FRAME_GOLDEN]);
717 if (s->framep[VP56_FRAME_GOLDEN2]->data[0])
718 avctx->release_buffer(avctx, s->framep[VP56_FRAME_GOLDEN2]);
719 if (s->framep[VP56_FRAME_PREVIOUS]->data[0])
720 avctx->release_buffer(avctx, s->framep[VP56_FRAME_PREVIOUS]);