2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
12 #include <stdlib.h> // qsort()
14 #include "./vp9_rtcd.h"
15 #include "./vpx_dsp_rtcd.h"
16 #include "./vpx_scale_rtcd.h"
18 #include "vpx_dsp/bitreader_buffer.h"
19 #include "vpx_dsp/bitreader.h"
20 #include "vpx_dsp/vpx_dsp_common.h"
21 #include "vpx_mem/vpx_mem.h"
22 #include "vpx_ports/mem.h"
23 #include "vpx_ports/mem_ops.h"
24 #include "vpx_scale/vpx_scale.h"
25 #include "vpx_util/vpx_thread.h"
27 #include "vp9/common/vp9_alloccommon.h"
28 #include "vp9/common/vp9_common.h"
29 #include "vp9/common/vp9_entropy.h"
30 #include "vp9/common/vp9_entropymode.h"
31 #include "vp9/common/vp9_idct.h"
32 #include "vp9/common/vp9_thread_common.h"
33 #include "vp9/common/vp9_pred_common.h"
34 #include "vp9/common/vp9_quant_common.h"
35 #include "vp9/common/vp9_reconintra.h"
36 #include "vp9/common/vp9_reconinter.h"
37 #include "vp9/common/vp9_seg_common.h"
38 #include "vp9/common/vp9_tile_common.h"
40 #include "vp9/decoder/vp9_decodeframe.h"
41 #include "vp9/decoder/vp9_detokenize.h"
42 #include "vp9/decoder/vp9_decodemv.h"
43 #include "vp9/decoder/vp9_decoder.h"
44 #include "vp9/decoder/vp9_dsubexp.h"
46 #define MAX_VP9_HEADER_SIZE 80
48 static int is_compound_reference_allowed(const VP9_COMMON *cm) {
50 for (i = 1; i < REFS_PER_FRAME; ++i)
51 if (cm->ref_frame_sign_bias[i + 1] != cm->ref_frame_sign_bias[1])
57 static void setup_compound_reference_mode(VP9_COMMON *cm) {
58 if (cm->ref_frame_sign_bias[LAST_FRAME] ==
59 cm->ref_frame_sign_bias[GOLDEN_FRAME]) {
60 cm->comp_fixed_ref = ALTREF_FRAME;
61 cm->comp_var_ref[0] = LAST_FRAME;
62 cm->comp_var_ref[1] = GOLDEN_FRAME;
63 } else if (cm->ref_frame_sign_bias[LAST_FRAME] ==
64 cm->ref_frame_sign_bias[ALTREF_FRAME]) {
65 cm->comp_fixed_ref = GOLDEN_FRAME;
66 cm->comp_var_ref[0] = LAST_FRAME;
67 cm->comp_var_ref[1] = ALTREF_FRAME;
69 cm->comp_fixed_ref = LAST_FRAME;
70 cm->comp_var_ref[0] = GOLDEN_FRAME;
71 cm->comp_var_ref[1] = ALTREF_FRAME;
75 static int read_is_valid(const uint8_t *start, size_t len, const uint8_t *end) {
76 return len != 0 && len <= (size_t)(end - start);
79 static int decode_unsigned_max(struct vpx_read_bit_buffer *rb, int max) {
80 const int data = vpx_rb_read_literal(rb, get_unsigned_bits(max));
81 return data > max ? max : data;
84 static TX_MODE read_tx_mode(vpx_reader *r) {
85 TX_MODE tx_mode = vpx_read_literal(r, 2);
86 if (tx_mode == ALLOW_32X32)
87 tx_mode += vpx_read_bit(r);
91 static void read_tx_mode_probs(struct tx_probs *tx_probs, vpx_reader *r) {
94 for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
95 for (j = 0; j < TX_SIZES - 3; ++j)
96 vp9_diff_update_prob(r, &tx_probs->p8x8[i][j]);
98 for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
99 for (j = 0; j < TX_SIZES - 2; ++j)
100 vp9_diff_update_prob(r, &tx_probs->p16x16[i][j]);
102 for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
103 for (j = 0; j < TX_SIZES - 1; ++j)
104 vp9_diff_update_prob(r, &tx_probs->p32x32[i][j]);
107 static void read_switchable_interp_probs(FRAME_CONTEXT *fc, vpx_reader *r) {
109 for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
110 for (i = 0; i < SWITCHABLE_FILTERS - 1; ++i)
111 vp9_diff_update_prob(r, &fc->switchable_interp_prob[j][i]);
114 static void read_inter_mode_probs(FRAME_CONTEXT *fc, vpx_reader *r) {
116 for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
117 for (j = 0; j < INTER_MODES - 1; ++j)
118 vp9_diff_update_prob(r, &fc->inter_mode_probs[i][j]);
121 static REFERENCE_MODE read_frame_reference_mode(const VP9_COMMON *cm,
123 if (is_compound_reference_allowed(cm)) {
124 return vpx_read_bit(r) ? (vpx_read_bit(r) ? REFERENCE_MODE_SELECT
125 : COMPOUND_REFERENCE)
128 return SINGLE_REFERENCE;
132 static void read_frame_reference_mode_probs(VP9_COMMON *cm, vpx_reader *r) {
133 FRAME_CONTEXT *const fc = cm->fc;
136 if (cm->reference_mode == REFERENCE_MODE_SELECT)
137 for (i = 0; i < COMP_INTER_CONTEXTS; ++i)
138 vp9_diff_update_prob(r, &fc->comp_inter_prob[i]);
140 if (cm->reference_mode != COMPOUND_REFERENCE)
141 for (i = 0; i < REF_CONTEXTS; ++i) {
142 vp9_diff_update_prob(r, &fc->single_ref_prob[i][0]);
143 vp9_diff_update_prob(r, &fc->single_ref_prob[i][1]);
146 if (cm->reference_mode != SINGLE_REFERENCE)
147 for (i = 0; i < REF_CONTEXTS; ++i)
148 vp9_diff_update_prob(r, &fc->comp_ref_prob[i]);
151 static void update_mv_probs(vpx_prob *p, int n, vpx_reader *r) {
153 for (i = 0; i < n; ++i)
154 if (vpx_read(r, MV_UPDATE_PROB))
155 p[i] = (vpx_read_literal(r, 7) << 1) | 1;
158 static void read_mv_probs(nmv_context *ctx, int allow_hp, vpx_reader *r) {
161 update_mv_probs(ctx->joints, MV_JOINTS - 1, r);
163 for (i = 0; i < 2; ++i) {
164 nmv_component *const comp_ctx = &ctx->comps[i];
165 update_mv_probs(&comp_ctx->sign, 1, r);
166 update_mv_probs(comp_ctx->classes, MV_CLASSES - 1, r);
167 update_mv_probs(comp_ctx->class0, CLASS0_SIZE - 1, r);
168 update_mv_probs(comp_ctx->bits, MV_OFFSET_BITS, r);
171 for (i = 0; i < 2; ++i) {
172 nmv_component *const comp_ctx = &ctx->comps[i];
173 for (j = 0; j < CLASS0_SIZE; ++j)
174 update_mv_probs(comp_ctx->class0_fp[j], MV_FP_SIZE - 1, r);
175 update_mv_probs(comp_ctx->fp, 3, r);
179 for (i = 0; i < 2; ++i) {
180 nmv_component *const comp_ctx = &ctx->comps[i];
181 update_mv_probs(&comp_ctx->class0_hp, 1, r);
182 update_mv_probs(&comp_ctx->hp, 1, r);
187 static void inverse_transform_block_inter(MACROBLOCKD* xd, int plane,
188 const TX_SIZE tx_size,
189 uint8_t *dst, int stride,
191 struct macroblockd_plane *const pd = &xd->plane[plane];
193 tran_low_t *const dqcoeff = pd->dqcoeff;
194 #if CONFIG_VP9_HIGHBITDEPTH
195 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
197 vp9_highbd_iwht4x4_add(dqcoeff, dst, stride, eob, xd->bd);
201 vp9_highbd_idct4x4_add(dqcoeff, dst, stride, eob, xd->bd);
204 vp9_highbd_idct8x8_add(dqcoeff, dst, stride, eob, xd->bd);
207 vp9_highbd_idct16x16_add(dqcoeff, dst, stride, eob, xd->bd);
210 vp9_highbd_idct32x32_add(dqcoeff, dst, stride, eob, xd->bd);
213 assert(0 && "Invalid transform size");
218 vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
222 vp9_idct4x4_add(dqcoeff, dst, stride, eob);
225 vp9_idct8x8_add(dqcoeff, dst, stride, eob);
228 vp9_idct16x16_add(dqcoeff, dst, stride, eob);
231 vp9_idct32x32_add(dqcoeff, dst, stride, eob);
234 assert(0 && "Invalid transform size");
241 vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
245 vp9_idct4x4_add(dqcoeff, dst, stride, eob);
248 vp9_idct8x8_add(dqcoeff, dst, stride, eob);
251 vp9_idct16x16_add(dqcoeff, dst, stride, eob);
254 vp9_idct32x32_add(dqcoeff, dst, stride, eob);
257 assert(0 && "Invalid transform size");
261 #endif // CONFIG_VP9_HIGHBITDEPTH
266 if (tx_size <= TX_16X16 && eob <= 10)
267 memset(dqcoeff, 0, 4 * (4 << tx_size) * sizeof(dqcoeff[0]));
268 else if (tx_size == TX_32X32 && eob <= 34)
269 memset(dqcoeff, 0, 256 * sizeof(dqcoeff[0]));
271 memset(dqcoeff, 0, (16 << (tx_size << 1)) * sizeof(dqcoeff[0]));
276 static void inverse_transform_block_intra(MACROBLOCKD* xd, int plane,
277 const TX_TYPE tx_type,
278 const TX_SIZE tx_size,
279 uint8_t *dst, int stride,
281 struct macroblockd_plane *const pd = &xd->plane[plane];
283 tran_low_t *const dqcoeff = pd->dqcoeff;
284 #if CONFIG_VP9_HIGHBITDEPTH
285 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
287 vp9_highbd_iwht4x4_add(dqcoeff, dst, stride, eob, xd->bd);
291 vp9_highbd_iht4x4_add(tx_type, dqcoeff, dst, stride, eob, xd->bd);
294 vp9_highbd_iht8x8_add(tx_type, dqcoeff, dst, stride, eob, xd->bd);
297 vp9_highbd_iht16x16_add(tx_type, dqcoeff, dst, stride, eob, xd->bd);
300 vp9_highbd_idct32x32_add(dqcoeff, dst, stride, eob, xd->bd);
303 assert(0 && "Invalid transform size");
308 vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
312 vp9_iht4x4_add(tx_type, dqcoeff, dst, stride, eob);
315 vp9_iht8x8_add(tx_type, dqcoeff, dst, stride, eob);
318 vp9_iht16x16_add(tx_type, dqcoeff, dst, stride, eob);
321 vp9_idct32x32_add(dqcoeff, dst, stride, eob);
324 assert(0 && "Invalid transform size");
331 vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
335 vp9_iht4x4_add(tx_type, dqcoeff, dst, stride, eob);
338 vp9_iht8x8_add(tx_type, dqcoeff, dst, stride, eob);
341 vp9_iht16x16_add(tx_type, dqcoeff, dst, stride, eob);
344 vp9_idct32x32_add(dqcoeff, dst, stride, eob);
347 assert(0 && "Invalid transform size");
351 #endif // CONFIG_VP9_HIGHBITDEPTH
356 if (tx_type == DCT_DCT && tx_size <= TX_16X16 && eob <= 10)
357 memset(dqcoeff, 0, 4 * (4 << tx_size) * sizeof(dqcoeff[0]));
358 else if (tx_size == TX_32X32 && eob <= 34)
359 memset(dqcoeff, 0, 256 * sizeof(dqcoeff[0]));
361 memset(dqcoeff, 0, (16 << (tx_size << 1)) * sizeof(dqcoeff[0]));
366 static void predict_and_reconstruct_intra_block(MACROBLOCKD *const xd,
368 MB_MODE_INFO *const mbmi,
372 struct macroblockd_plane *const pd = &xd->plane[plane];
373 PREDICTION_MODE mode = (plane == 0) ? mbmi->mode : mbmi->uv_mode;
375 dst = &pd->dst.buf[4 * row * pd->dst.stride + 4 * col];
377 if (mbmi->sb_type < BLOCK_8X8)
379 mode = xd->mi[0]->bmi[(row << 1) + col].as_mode;
381 vp9_predict_intra_block(xd, pd->n4_wl, tx_size, mode,
382 dst, pd->dst.stride, dst, pd->dst.stride,
386 const TX_TYPE tx_type = (plane || xd->lossless) ?
387 DCT_DCT : intra_mode_to_tx_type_lookup[mode];
388 const scan_order *sc = (plane || xd->lossless) ?
389 &vp9_default_scan_orders[tx_size] : &vp9_scan_orders[tx_size][tx_type];
390 const int eob = vp9_decode_block_tokens(xd, plane, sc, col, row, tx_size,
391 r, mbmi->segment_id);
392 inverse_transform_block_intra(xd, plane, tx_type, tx_size,
393 dst, pd->dst.stride, eob);
397 static int reconstruct_inter_block(MACROBLOCKD *const xd, vpx_reader *r,
398 MB_MODE_INFO *const mbmi, int plane,
399 int row, int col, TX_SIZE tx_size) {
400 struct macroblockd_plane *const pd = &xd->plane[plane];
401 const scan_order *sc = &vp9_default_scan_orders[tx_size];
402 const int eob = vp9_decode_block_tokens(xd, plane, sc, col, row, tx_size, r,
405 inverse_transform_block_inter(xd, plane, tx_size,
406 &pd->dst.buf[4 * row * pd->dst.stride + 4 * col],
407 pd->dst.stride, eob);
411 static void build_mc_border(const uint8_t *src, int src_stride,
412 uint8_t *dst, int dst_stride,
413 int x, int y, int b_w, int b_h, int w, int h) {
414 // Get a pointer to the start of the real data for this row.
415 const uint8_t *ref_row = src - x - y * src_stride;
418 ref_row += (h - 1) * src_stride;
420 ref_row += y * src_stride;
424 int left = x < 0 ? -x : 0;
435 copy = b_w - left - right;
438 memset(dst, ref_row[0], left);
441 memcpy(dst + left, ref_row + x + left, copy);
444 memset(dst + left + copy, ref_row[w - 1], right);
450 ref_row += src_stride;
454 #if CONFIG_VP9_HIGHBITDEPTH
455 static void high_build_mc_border(const uint8_t *src8, int src_stride,
456 uint16_t *dst, int dst_stride,
457 int x, int y, int b_w, int b_h,
459 // Get a pointer to the start of the real data for this row.
460 const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
461 const uint16_t *ref_row = src - x - y * src_stride;
464 ref_row += (h - 1) * src_stride;
466 ref_row += y * src_stride;
470 int left = x < 0 ? -x : 0;
481 copy = b_w - left - right;
484 vpx_memset16(dst, ref_row[0], left);
487 memcpy(dst + left, ref_row + x + left, copy * sizeof(uint16_t));
490 vpx_memset16(dst + left + copy, ref_row[w - 1], right);
496 ref_row += src_stride;
499 #endif // CONFIG_VP9_HIGHBITDEPTH
501 #if CONFIG_VP9_HIGHBITDEPTH
502 static void extend_and_predict(const uint8_t *buf_ptr1, int pre_buf_stride,
503 int x0, int y0, int b_w, int b_h,
504 int frame_width, int frame_height,
506 uint8_t *const dst, int dst_buf_stride,
507 int subpel_x, int subpel_y,
508 const InterpKernel *kernel,
509 const struct scale_factors *sf,
511 int w, int h, int ref, int xs, int ys) {
512 DECLARE_ALIGNED(16, uint16_t, mc_buf_high[80 * 2 * 80 * 2]);
513 const uint8_t *buf_ptr;
515 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
516 high_build_mc_border(buf_ptr1, pre_buf_stride, mc_buf_high, b_w,
517 x0, y0, b_w, b_h, frame_width, frame_height);
518 buf_ptr = CONVERT_TO_BYTEPTR(mc_buf_high) + border_offset;
520 build_mc_border(buf_ptr1, pre_buf_stride, (uint8_t *)mc_buf_high, b_w,
521 x0, y0, b_w, b_h, frame_width, frame_height);
522 buf_ptr = ((uint8_t *)mc_buf_high) + border_offset;
525 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
526 high_inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
527 subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
529 inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
530 subpel_y, sf, w, h, ref, kernel, xs, ys);
534 static void extend_and_predict(const uint8_t *buf_ptr1, int pre_buf_stride,
535 int x0, int y0, int b_w, int b_h,
536 int frame_width, int frame_height,
538 uint8_t *const dst, int dst_buf_stride,
539 int subpel_x, int subpel_y,
540 const InterpKernel *kernel,
541 const struct scale_factors *sf,
542 int w, int h, int ref, int xs, int ys) {
543 DECLARE_ALIGNED(16, uint8_t, mc_buf[80 * 2 * 80 * 2]);
544 const uint8_t *buf_ptr;
546 build_mc_border(buf_ptr1, pre_buf_stride, mc_buf, b_w,
547 x0, y0, b_w, b_h, frame_width, frame_height);
548 buf_ptr = mc_buf + border_offset;
550 inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
551 subpel_y, sf, w, h, ref, kernel, xs, ys);
553 #endif // CONFIG_VP9_HIGHBITDEPTH
555 static void dec_build_inter_predictors(VP9Decoder *const pbi, MACROBLOCKD *xd,
556 int plane, int bw, int bh, int x,
557 int y, int w, int h, int mi_x, int mi_y,
558 const InterpKernel *kernel,
559 const struct scale_factors *sf,
560 struct buf_2d *pre_buf,
561 struct buf_2d *dst_buf, const MV* mv,
562 RefCntBuffer *ref_frame_buf,
563 int is_scaled, int ref) {
564 struct macroblockd_plane *const pd = &xd->plane[plane];
565 uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
567 int xs, ys, x0, y0, x0_16, y0_16, frame_width, frame_height,
568 buf_stride, subpel_x, subpel_y;
569 uint8_t *ref_frame, *buf_ptr;
571 // Get reference frame pointer, width and height.
573 frame_width = ref_frame_buf->buf.y_crop_width;
574 frame_height = ref_frame_buf->buf.y_crop_height;
575 ref_frame = ref_frame_buf->buf.y_buffer;
577 frame_width = ref_frame_buf->buf.uv_crop_width;
578 frame_height = ref_frame_buf->buf.uv_crop_height;
579 ref_frame = plane == 1 ? ref_frame_buf->buf.u_buffer
580 : ref_frame_buf->buf.v_buffer;
584 const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, mv, bw, bh,
587 // Co-ordinate of containing block to pixel precision.
588 int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
589 int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
591 // Co-ordinate of the block to 1/16th pixel precision.
592 x0_16 = (x_start + x) << SUBPEL_BITS;
593 y0_16 = (y_start + y) << SUBPEL_BITS;
595 // Co-ordinate of current block in reference frame
596 // to 1/16th pixel precision.
597 x0_16 = sf->scale_value_x(x0_16, sf);
598 y0_16 = sf->scale_value_y(y0_16, sf);
600 // Map the top left corner of the block into the reference frame.
601 x0 = sf->scale_value_x(x_start + x, sf);
602 y0 = sf->scale_value_y(y_start + y, sf);
604 // Scale the MV and incorporate the sub-pixel offset of the block
605 // in the reference frame.
606 scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
610 // Co-ordinate of containing block to pixel precision.
611 x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
612 y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
614 // Co-ordinate of the block to 1/16th pixel precision.
615 x0_16 = x0 << SUBPEL_BITS;
616 y0_16 = y0 << SUBPEL_BITS;
618 scaled_mv.row = mv->row * (1 << (1 - pd->subsampling_y));
619 scaled_mv.col = mv->col * (1 << (1 - pd->subsampling_x));
622 subpel_x = scaled_mv.col & SUBPEL_MASK;
623 subpel_y = scaled_mv.row & SUBPEL_MASK;
625 // Calculate the top left corner of the best matching block in the
627 x0 += scaled_mv.col >> SUBPEL_BITS;
628 y0 += scaled_mv.row >> SUBPEL_BITS;
629 x0_16 += scaled_mv.col;
630 y0_16 += scaled_mv.row;
632 // Get reference block pointer.
633 buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
634 buf_stride = pre_buf->stride;
636 // Do border extension if there is motion or the
637 // width/height is not a multiple of 8 pixels.
638 if (is_scaled || scaled_mv.col || scaled_mv.row ||
639 (frame_width & 0x7) || (frame_height & 0x7)) {
640 int y1 = ((y0_16 + (h - 1) * ys) >> SUBPEL_BITS) + 1;
642 // Get reference block bottom right horizontal coordinate.
643 int x1 = ((x0_16 + (w - 1) * xs) >> SUBPEL_BITS) + 1;
644 int x_pad = 0, y_pad = 0;
646 if (subpel_x || (sf->x_step_q4 != SUBPEL_SHIFTS)) {
647 x0 -= VP9_INTERP_EXTEND - 1;
648 x1 += VP9_INTERP_EXTEND;
652 if (subpel_y || (sf->y_step_q4 != SUBPEL_SHIFTS)) {
653 y0 -= VP9_INTERP_EXTEND - 1;
654 y1 += VP9_INTERP_EXTEND;
658 // Wait until reference block is ready. Pad 7 more pixels as last 7
659 // pixels of each superblock row can be changed by next superblock row.
660 if (pbi->frame_parallel_decode)
661 vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
662 VPXMAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
664 // Skip border extension if block is inside the frame.
665 if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width - 1 ||
666 y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
667 // Extend the border.
668 const uint8_t *const buf_ptr1 = ref_frame + y0 * buf_stride + x0;
669 const int b_w = x1 - x0 + 1;
670 const int b_h = y1 - y0 + 1;
671 const int border_offset = y_pad * 3 * b_w + x_pad * 3;
673 extend_and_predict(buf_ptr1, buf_stride, x0, y0, b_w, b_h,
674 frame_width, frame_height, border_offset,
675 dst, dst_buf->stride,
678 #if CONFIG_VP9_HIGHBITDEPTH
685 // Wait until reference block is ready. Pad 7 more pixels as last 7
686 // pixels of each superblock row can be changed by next superblock row.
687 if (pbi->frame_parallel_decode) {
688 const int y1 = (y0_16 + (h - 1) * ys) >> SUBPEL_BITS;
689 vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
690 VPXMAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
693 #if CONFIG_VP9_HIGHBITDEPTH
694 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
695 high_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
696 subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
698 inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
699 subpel_y, sf, w, h, ref, kernel, xs, ys);
702 inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
703 subpel_y, sf, w, h, ref, kernel, xs, ys);
704 #endif // CONFIG_VP9_HIGHBITDEPTH
707 static void dec_build_inter_predictors_sb(VP9Decoder *const pbi,
709 int mi_row, int mi_col) {
711 const int mi_x = mi_col * MI_SIZE;
712 const int mi_y = mi_row * MI_SIZE;
713 const MODE_INFO *mi = xd->mi[0];
714 const InterpKernel *kernel = vp9_filter_kernels[mi->mbmi.interp_filter];
715 const BLOCK_SIZE sb_type = mi->mbmi.sb_type;
716 const int is_compound = has_second_ref(&mi->mbmi);
718 for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
719 struct macroblockd_plane *const pd = &xd->plane[plane];
720 struct buf_2d *const dst_buf = &pd->dst;
721 const int num_4x4_w = pd->n4_w;
722 const int num_4x4_h = pd->n4_h;
724 const int n4w_x4 = 4 * num_4x4_w;
725 const int n4h_x4 = 4 * num_4x4_h;
728 for (ref = 0; ref < 1 + is_compound; ++ref) {
729 const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
730 struct buf_2d *const pre_buf = &pd->pre[ref];
731 const int idx = xd->block_refs[ref]->idx;
732 BufferPool *const pool = pbi->common.buffer_pool;
733 RefCntBuffer *const ref_frame_buf = &pool->frame_bufs[idx];
734 const int is_scaled = vp9_is_scaled(sf);
736 if (sb_type < BLOCK_8X8) {
738 for (y = 0; y < num_4x4_h; ++y) {
739 for (x = 0; x < num_4x4_w; ++x) {
740 const MV mv = average_split_mvs(pd, mi, ref, i++);
741 dec_build_inter_predictors(pbi, xd, plane, n4w_x4, n4h_x4,
742 4 * x, 4 * y, 4, 4, mi_x, mi_y, kernel,
743 sf, pre_buf, dst_buf, &mv,
744 ref_frame_buf, is_scaled, ref);
748 const MV mv = mi->mbmi.mv[ref].as_mv;
749 dec_build_inter_predictors(pbi, xd, plane, n4w_x4, n4h_x4,
750 0, 0, n4w_x4, n4h_x4, mi_x, mi_y, kernel,
751 sf, pre_buf, dst_buf, &mv, ref_frame_buf,
758 static INLINE TX_SIZE dec_get_uv_tx_size(const MB_MODE_INFO *mbmi,
759 int n4_wl, int n4_hl) {
760 // get minimum log2 num4x4s dimension
761 const int x = VPXMIN(n4_wl, n4_hl);
762 return VPXMIN(mbmi->tx_size, x);
765 static INLINE void dec_reset_skip_context(MACROBLOCKD *xd) {
767 for (i = 0; i < MAX_MB_PLANE; i++) {
768 struct macroblockd_plane *const pd = &xd->plane[i];
769 memset(pd->above_context, 0, sizeof(ENTROPY_CONTEXT) * pd->n4_w);
770 memset(pd->left_context, 0, sizeof(ENTROPY_CONTEXT) * pd->n4_h);
774 static void set_plane_n4(MACROBLOCKD *const xd, int bw, int bh, int bwl,
777 for (i = 0; i < MAX_MB_PLANE; i++) {
778 xd->plane[i].n4_w = (bw << 1) >> xd->plane[i].subsampling_x;
779 xd->plane[i].n4_h = (bh << 1) >> xd->plane[i].subsampling_y;
780 xd->plane[i].n4_wl = bwl - xd->plane[i].subsampling_x;
781 xd->plane[i].n4_hl = bhl - xd->plane[i].subsampling_y;
785 static MB_MODE_INFO *set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd,
786 BLOCK_SIZE bsize, int mi_row, int mi_col,
787 int bw, int bh, int x_mis, int y_mis,
789 const int offset = mi_row * cm->mi_stride + mi_col;
791 const TileInfo *const tile = &xd->tile;
793 xd->mi = cm->mi_grid_visible + offset;
794 xd->mi[0] = &cm->mi[offset];
795 // TODO(slavarnway): Generate sb_type based on bwl and bhl, instead of
796 // passing bsize from decode_partition().
797 xd->mi[0]->mbmi.sb_type = bsize;
798 for (y = 0; y < y_mis; ++y)
799 for (x = !y; x < x_mis; ++x) {
800 xd->mi[y * cm->mi_stride + x] = xd->mi[0];
803 set_plane_n4(xd, bw, bh, bwl, bhl);
805 set_skip_context(xd, mi_row, mi_col);
807 // Distance of Mb to the various image edges. These are specified to 8th pel
808 // as they are always compared to values that are in 1/8th pel units
809 set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
811 vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
812 return &xd->mi[0]->mbmi;
815 static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd,
816 int mi_row, int mi_col,
817 vpx_reader *r, BLOCK_SIZE bsize,
819 VP9_COMMON *const cm = &pbi->common;
820 const int less8x8 = bsize < BLOCK_8X8;
821 const int bw = 1 << (bwl - 1);
822 const int bh = 1 << (bhl - 1);
823 const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
824 const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
826 MB_MODE_INFO *mbmi = set_offsets(cm, xd, bsize, mi_row, mi_col,
827 bw, bh, x_mis, y_mis, bwl, bhl);
829 if (bsize >= BLOCK_8X8 && (cm->subsampling_x || cm->subsampling_y)) {
830 const BLOCK_SIZE uv_subsize =
831 ss_size_lookup[bsize][cm->subsampling_x][cm->subsampling_y];
832 if (uv_subsize == BLOCK_INVALID)
833 vpx_internal_error(xd->error_info,
834 VPX_CODEC_CORRUPT_FRAME, "Invalid block size.");
837 vpx_read_mode_info(pbi, xd, mi_row, mi_col, r, x_mis, y_mis);
840 dec_reset_skip_context(xd);
843 if (!is_inter_block(mbmi)) {
845 for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
846 const struct macroblockd_plane *const pd = &xd->plane[plane];
847 const TX_SIZE tx_size =
848 plane ? dec_get_uv_tx_size(mbmi, pd->n4_wl, pd->n4_hl)
850 const int num_4x4_w = pd->n4_w;
851 const int num_4x4_h = pd->n4_h;
852 const int step = (1 << tx_size);
854 const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ?
855 0 : xd->mb_to_right_edge >> (5 + pd->subsampling_x));
856 const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ?
857 0 : xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
859 for (row = 0; row < max_blocks_high; row += step)
860 for (col = 0; col < max_blocks_wide; col += step)
861 predict_and_reconstruct_intra_block(xd, r, mbmi, plane,
866 dec_build_inter_predictors_sb(pbi, xd, mi_row, mi_col);
873 for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
874 const struct macroblockd_plane *const pd = &xd->plane[plane];
875 const TX_SIZE tx_size =
876 plane ? dec_get_uv_tx_size(mbmi, pd->n4_wl, pd->n4_hl)
878 const int num_4x4_w = pd->n4_w;
879 const int num_4x4_h = pd->n4_h;
880 const int step = (1 << tx_size);
882 const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ?
883 0 : xd->mb_to_right_edge >> (5 + pd->subsampling_x));
884 const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ?
885 0 : xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
887 for (row = 0; row < max_blocks_high; row += step)
888 for (col = 0; col < max_blocks_wide; col += step)
889 eobtotal += reconstruct_inter_block(xd, r, mbmi, plane, row, col,
893 if (!less8x8 && eobtotal == 0)
894 mbmi->skip = 1; // skip loopfilter
898 xd->corrupted |= vpx_reader_has_error(r);
900 if (cm->lf.filter_level) {
901 vp9_build_mask(cm, mbmi, mi_row, mi_col, bw, bh);
905 static INLINE int dec_partition_plane_context(const MACROBLOCKD *xd,
906 int mi_row, int mi_col,
908 const PARTITION_CONTEXT *above_ctx = xd->above_seg_context + mi_col;
909 const PARTITION_CONTEXT *left_ctx = xd->left_seg_context + (mi_row & MI_MASK);
910 int above = (*above_ctx >> bsl) & 1 , left = (*left_ctx >> bsl) & 1;
914 return (left * 2 + above) + bsl * PARTITION_PLOFFSET;
917 static INLINE void dec_update_partition_context(MACROBLOCKD *xd,
918 int mi_row, int mi_col,
921 PARTITION_CONTEXT *const above_ctx = xd->above_seg_context + mi_col;
922 PARTITION_CONTEXT *const left_ctx = xd->left_seg_context + (mi_row & MI_MASK);
924 // update the partition context at the end notes. set partition bits
925 // of block sizes larger than the current one to be one, and partition
926 // bits of smaller block sizes to be zero.
927 memset(above_ctx, partition_context_lookup[subsize].above, bw);
928 memset(left_ctx, partition_context_lookup[subsize].left, bw);
931 static PARTITION_TYPE read_partition(MACROBLOCKD *xd, int mi_row, int mi_col,
933 int has_rows, int has_cols, int bsl) {
934 const int ctx = dec_partition_plane_context(xd, mi_row, mi_col, bsl);
935 const vpx_prob *const probs = get_partition_probs(xd, ctx);
936 FRAME_COUNTS *counts = xd->counts;
939 if (has_rows && has_cols)
940 p = (PARTITION_TYPE)vpx_read_tree(r, vp9_partition_tree, probs);
941 else if (!has_rows && has_cols)
942 p = vpx_read(r, probs[1]) ? PARTITION_SPLIT : PARTITION_HORZ;
943 else if (has_rows && !has_cols)
944 p = vpx_read(r, probs[2]) ? PARTITION_SPLIT : PARTITION_VERT;
949 ++counts->partition[ctx][p];
954 // TODO(slavarnway): eliminate bsize and subsize in future commits
955 static void decode_partition(VP9Decoder *const pbi, MACROBLOCKD *const xd,
956 int mi_row, int mi_col,
957 vpx_reader* r, BLOCK_SIZE bsize, int n4x4_l2) {
958 VP9_COMMON *const cm = &pbi->common;
959 const int n8x8_l2 = n4x4_l2 - 1;
960 const int num_8x8_wh = 1 << n8x8_l2;
961 const int hbs = num_8x8_wh >> 1;
962 PARTITION_TYPE partition;
964 const int has_rows = (mi_row + hbs) < cm->mi_rows;
965 const int has_cols = (mi_col + hbs) < cm->mi_cols;
967 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
970 partition = read_partition(xd, mi_row, mi_col, r, has_rows, has_cols,
972 subsize = subsize_lookup[partition][bsize]; // get_subsize(bsize, partition);
974 // calculate bmode block dimensions (log 2)
975 xd->bmode_blocks_wl = 1 >> !!(partition & PARTITION_VERT);
976 xd->bmode_blocks_hl = 1 >> !!(partition & PARTITION_HORZ);
977 decode_block(pbi, xd, mi_row, mi_col, r, subsize, 1, 1);
981 decode_block(pbi, xd, mi_row, mi_col, r, subsize, n4x4_l2, n4x4_l2);
984 decode_block(pbi, xd, mi_row, mi_col, r, subsize, n4x4_l2, n8x8_l2);
986 decode_block(pbi, xd, mi_row + hbs, mi_col, r, subsize, n4x4_l2,
990 decode_block(pbi, xd, mi_row, mi_col, r, subsize, n8x8_l2, n4x4_l2);
992 decode_block(pbi, xd, mi_row, mi_col + hbs, r, subsize, n8x8_l2,
995 case PARTITION_SPLIT:
996 decode_partition(pbi, xd, mi_row, mi_col, r, subsize, n8x8_l2);
997 decode_partition(pbi, xd, mi_row, mi_col + hbs, r, subsize, n8x8_l2);
998 decode_partition(pbi, xd, mi_row + hbs, mi_col, r, subsize, n8x8_l2);
999 decode_partition(pbi, xd, mi_row + hbs, mi_col + hbs, r, subsize,
1003 assert(0 && "Invalid partition type");
1007 // update partition context
1008 if (bsize >= BLOCK_8X8 &&
1009 (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
1010 dec_update_partition_context(xd, mi_row, mi_col, subsize, num_8x8_wh);
1013 static void setup_token_decoder(const uint8_t *data,
1014 const uint8_t *data_end,
1016 struct vpx_internal_error_info *error_info,
1018 vpx_decrypt_cb decrypt_cb,
1019 void *decrypt_state) {
1020 // Validate the calculated partition length. If the buffer
1021 // described by the partition can't be fully read, then restrict
1022 // it to the portion that can be (for EC mode) or throw an error.
1023 if (!read_is_valid(data, read_size, data_end))
1024 vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
1025 "Truncated packet or corrupt tile length");
1027 if (vpx_reader_init(r, data, read_size, decrypt_cb, decrypt_state))
1028 vpx_internal_error(error_info, VPX_CODEC_MEM_ERROR,
1029 "Failed to allocate bool decoder %d", 1);
1032 static void read_coef_probs_common(vp9_coeff_probs_model *coef_probs,
1036 if (vpx_read_bit(r))
1037 for (i = 0; i < PLANE_TYPES; ++i)
1038 for (j = 0; j < REF_TYPES; ++j)
1039 for (k = 0; k < COEF_BANDS; ++k)
1040 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
1041 for (m = 0; m < UNCONSTRAINED_NODES; ++m)
1042 vp9_diff_update_prob(r, &coef_probs[i][j][k][l][m]);
1045 static void read_coef_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode,
1047 const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
1049 for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
1050 read_coef_probs_common(fc->coef_probs[tx_size], r);
1053 static void setup_segmentation(struct segmentation *seg,
1054 struct vpx_read_bit_buffer *rb) {
1057 seg->update_map = 0;
1058 seg->update_data = 0;
1060 seg->enabled = vpx_rb_read_bit(rb);
1064 // Segmentation map update
1065 seg->update_map = vpx_rb_read_bit(rb);
1066 if (seg->update_map) {
1067 for (i = 0; i < SEG_TREE_PROBS; i++)
1068 seg->tree_probs[i] = vpx_rb_read_bit(rb) ? vpx_rb_read_literal(rb, 8)
1071 seg->temporal_update = vpx_rb_read_bit(rb);
1072 if (seg->temporal_update) {
1073 for (i = 0; i < PREDICTION_PROBS; i++)
1074 seg->pred_probs[i] = vpx_rb_read_bit(rb) ? vpx_rb_read_literal(rb, 8)
1077 for (i = 0; i < PREDICTION_PROBS; i++)
1078 seg->pred_probs[i] = MAX_PROB;
1082 // Segmentation data update
1083 seg->update_data = vpx_rb_read_bit(rb);
1084 if (seg->update_data) {
1085 seg->abs_delta = vpx_rb_read_bit(rb);
1087 vp9_clearall_segfeatures(seg);
1089 for (i = 0; i < MAX_SEGMENTS; i++) {
1090 for (j = 0; j < SEG_LVL_MAX; j++) {
1092 const int feature_enabled = vpx_rb_read_bit(rb);
1093 if (feature_enabled) {
1094 vp9_enable_segfeature(seg, i, j);
1095 data = decode_unsigned_max(rb, vp9_seg_feature_data_max(j));
1096 if (vp9_is_segfeature_signed(j))
1097 data = vpx_rb_read_bit(rb) ? -data : data;
1099 vp9_set_segdata(seg, i, j, data);
1105 static void setup_loopfilter(struct loopfilter *lf,
1106 struct vpx_read_bit_buffer *rb) {
1107 lf->filter_level = vpx_rb_read_literal(rb, 6);
1108 lf->sharpness_level = vpx_rb_read_literal(rb, 3);
1110 // Read in loop filter deltas applied at the MB level based on mode or ref
1112 lf->mode_ref_delta_update = 0;
1114 lf->mode_ref_delta_enabled = vpx_rb_read_bit(rb);
1115 if (lf->mode_ref_delta_enabled) {
1116 lf->mode_ref_delta_update = vpx_rb_read_bit(rb);
1117 if (lf->mode_ref_delta_update) {
1120 for (i = 0; i < MAX_REF_LF_DELTAS; i++)
1121 if (vpx_rb_read_bit(rb))
1122 lf->ref_deltas[i] = vpx_rb_read_signed_literal(rb, 6);
1124 for (i = 0; i < MAX_MODE_LF_DELTAS; i++)
1125 if (vpx_rb_read_bit(rb))
1126 lf->mode_deltas[i] = vpx_rb_read_signed_literal(rb, 6);
1131 static INLINE int read_delta_q(struct vpx_read_bit_buffer *rb) {
1132 return vpx_rb_read_bit(rb) ? vpx_rb_read_signed_literal(rb, 4) : 0;
1135 static void setup_quantization(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1136 struct vpx_read_bit_buffer *rb) {
1137 cm->base_qindex = vpx_rb_read_literal(rb, QINDEX_BITS);
1138 cm->y_dc_delta_q = read_delta_q(rb);
1139 cm->uv_dc_delta_q = read_delta_q(rb);
1140 cm->uv_ac_delta_q = read_delta_q(rb);
1141 cm->dequant_bit_depth = cm->bit_depth;
1142 xd->lossless = cm->base_qindex == 0 &&
1143 cm->y_dc_delta_q == 0 &&
1144 cm->uv_dc_delta_q == 0 &&
1145 cm->uv_ac_delta_q == 0;
1147 #if CONFIG_VP9_HIGHBITDEPTH
1148 xd->bd = (int)cm->bit_depth;
1152 static void setup_segmentation_dequant(VP9_COMMON *const cm) {
1153 // Build y/uv dequant values based on segmentation.
1154 if (cm->seg.enabled) {
1156 for (i = 0; i < MAX_SEGMENTS; ++i) {
1157 const int qindex = vp9_get_qindex(&cm->seg, i, cm->base_qindex);
1158 cm->y_dequant[i][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q,
1160 cm->y_dequant[i][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
1161 cm->uv_dequant[i][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
1163 cm->uv_dequant[i][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
1167 const int qindex = cm->base_qindex;
1168 // When segmentation is disabled, only the first value is used. The
1169 // remaining are don't cares.
1170 cm->y_dequant[0][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
1171 cm->y_dequant[0][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
1172 cm->uv_dequant[0][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
1174 cm->uv_dequant[0][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
1179 static INTERP_FILTER read_interp_filter(struct vpx_read_bit_buffer *rb) {
1180 const INTERP_FILTER literal_to_filter[] = { EIGHTTAP_SMOOTH,
1184 return vpx_rb_read_bit(rb) ? SWITCHABLE
1185 : literal_to_filter[vpx_rb_read_literal(rb, 2)];
1188 static void setup_render_size(VP9_COMMON *cm, struct vpx_read_bit_buffer *rb) {
1189 cm->render_width = cm->width;
1190 cm->render_height = cm->height;
1191 if (vpx_rb_read_bit(rb))
1192 vp9_read_frame_size(rb, &cm->render_width, &cm->render_height);
1195 static void resize_mv_buffer(VP9_COMMON *cm) {
1196 vpx_free(cm->cur_frame->mvs);
1197 cm->cur_frame->mi_rows = cm->mi_rows;
1198 cm->cur_frame->mi_cols = cm->mi_cols;
1199 cm->cur_frame->mvs = (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
1200 sizeof(*cm->cur_frame->mvs));
1203 static void resize_context_buffers(VP9_COMMON *cm, int width, int height) {
1204 #if CONFIG_SIZE_LIMIT
1205 if (width > DECODE_WIDTH_LIMIT || height > DECODE_HEIGHT_LIMIT)
1206 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1207 "Dimensions of %dx%d beyond allowed size of %dx%d.",
1208 width, height, DECODE_WIDTH_LIMIT, DECODE_HEIGHT_LIMIT);
1210 if (cm->width != width || cm->height != height) {
1211 const int new_mi_rows =
1212 ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2) >> MI_SIZE_LOG2;
1213 const int new_mi_cols =
1214 ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2) >> MI_SIZE_LOG2;
1216 // Allocations in vp9_alloc_context_buffers() depend on individual
1217 // dimensions as well as the overall size.
1218 if (new_mi_cols > cm->mi_cols || new_mi_rows > cm->mi_rows) {
1219 if (vp9_alloc_context_buffers(cm, width, height))
1220 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1221 "Failed to allocate context buffers");
1223 vp9_set_mb_mi(cm, width, height);
1225 vp9_init_context_buffers(cm);
1227 cm->height = height;
1229 if (cm->cur_frame->mvs == NULL || cm->mi_rows > cm->cur_frame->mi_rows ||
1230 cm->mi_cols > cm->cur_frame->mi_cols) {
1231 resize_mv_buffer(cm);
1235 static void setup_frame_size(VP9_COMMON *cm, struct vpx_read_bit_buffer *rb) {
1237 BufferPool *const pool = cm->buffer_pool;
1238 vp9_read_frame_size(rb, &width, &height);
1239 resize_context_buffers(cm, width, height);
1240 setup_render_size(cm, rb);
1242 lock_buffer_pool(pool);
1243 if (vpx_realloc_frame_buffer(
1244 get_frame_new_buffer(cm), cm->width, cm->height,
1245 cm->subsampling_x, cm->subsampling_y,
1246 #if CONFIG_VP9_HIGHBITDEPTH
1247 cm->use_highbitdepth,
1249 VP9_DEC_BORDER_IN_PIXELS,
1251 &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
1253 unlock_buffer_pool(pool);
1254 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1255 "Failed to allocate frame buffer");
1257 unlock_buffer_pool(pool);
1259 pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
1260 pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
1261 pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
1262 pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
1263 pool->frame_bufs[cm->new_fb_idx].buf.color_range = cm->color_range;
1264 pool->frame_bufs[cm->new_fb_idx].buf.render_width = cm->render_width;
1265 pool->frame_bufs[cm->new_fb_idx].buf.render_height = cm->render_height;
1268 static INLINE int valid_ref_frame_img_fmt(vpx_bit_depth_t ref_bit_depth,
1269 int ref_xss, int ref_yss,
1270 vpx_bit_depth_t this_bit_depth,
1271 int this_xss, int this_yss) {
1272 return ref_bit_depth == this_bit_depth && ref_xss == this_xss &&
1273 ref_yss == this_yss;
1276 static void setup_frame_size_with_refs(VP9_COMMON *cm,
1277 struct vpx_read_bit_buffer *rb) {
1280 int has_valid_ref_frame = 0;
1281 BufferPool *const pool = cm->buffer_pool;
1282 for (i = 0; i < REFS_PER_FRAME; ++i) {
1283 if (vpx_rb_read_bit(rb)) {
1284 YV12_BUFFER_CONFIG *const buf = cm->frame_refs[i].buf;
1285 width = buf->y_crop_width;
1286 height = buf->y_crop_height;
1293 vp9_read_frame_size(rb, &width, &height);
1295 if (width <= 0 || height <= 0)
1296 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1297 "Invalid frame size");
1299 // Check to make sure at least one of frames that this frame references
1300 // has valid dimensions.
1301 for (i = 0; i < REFS_PER_FRAME; ++i) {
1302 RefBuffer *const ref_frame = &cm->frame_refs[i];
1303 has_valid_ref_frame |= valid_ref_frame_size(ref_frame->buf->y_crop_width,
1304 ref_frame->buf->y_crop_height,
1307 if (!has_valid_ref_frame)
1308 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1309 "Referenced frame has invalid size");
1310 for (i = 0; i < REFS_PER_FRAME; ++i) {
1311 RefBuffer *const ref_frame = &cm->frame_refs[i];
1312 if (!valid_ref_frame_img_fmt(
1313 ref_frame->buf->bit_depth,
1314 ref_frame->buf->subsampling_x,
1315 ref_frame->buf->subsampling_y,
1319 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1320 "Referenced frame has incompatible color format");
1323 resize_context_buffers(cm, width, height);
1324 setup_render_size(cm, rb);
1326 lock_buffer_pool(pool);
1327 if (vpx_realloc_frame_buffer(
1328 get_frame_new_buffer(cm), cm->width, cm->height,
1329 cm->subsampling_x, cm->subsampling_y,
1330 #if CONFIG_VP9_HIGHBITDEPTH
1331 cm->use_highbitdepth,
1333 VP9_DEC_BORDER_IN_PIXELS,
1335 &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
1337 unlock_buffer_pool(pool);
1338 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1339 "Failed to allocate frame buffer");
1341 unlock_buffer_pool(pool);
1343 pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
1344 pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
1345 pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
1346 pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
1347 pool->frame_bufs[cm->new_fb_idx].buf.color_range = cm->color_range;
1348 pool->frame_bufs[cm->new_fb_idx].buf.render_width = cm->render_width;
1349 pool->frame_bufs[cm->new_fb_idx].buf.render_height = cm->render_height;
1352 static void setup_tile_info(VP9_COMMON *cm, struct vpx_read_bit_buffer *rb) {
1353 int min_log2_tile_cols, max_log2_tile_cols, max_ones;
1354 vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
1357 max_ones = max_log2_tile_cols - min_log2_tile_cols;
1358 cm->log2_tile_cols = min_log2_tile_cols;
1359 while (max_ones-- && vpx_rb_read_bit(rb))
1360 cm->log2_tile_cols++;
1362 if (cm->log2_tile_cols > 6)
1363 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1364 "Invalid number of tile columns");
1367 cm->log2_tile_rows = vpx_rb_read_bit(rb);
1368 if (cm->log2_tile_rows)
1369 cm->log2_tile_rows += vpx_rb_read_bit(rb);
1372 // Reads the next tile returning its size and adjusting '*data' accordingly
1373 // based on 'is_last'.
1374 static void get_tile_buffer(const uint8_t *const data_end,
1376 struct vpx_internal_error_info *error_info,
1377 const uint8_t **data,
1378 vpx_decrypt_cb decrypt_cb, void *decrypt_state,
1383 if (!read_is_valid(*data, 4, data_end))
1384 vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
1385 "Truncated packet or corrupt tile length");
1389 decrypt_cb(decrypt_state, *data, be_data, 4);
1390 size = mem_get_be32(be_data);
1392 size = mem_get_be32(*data);
1396 if (size > (size_t)(data_end - *data))
1397 vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
1398 "Truncated packet or corrupt tile size");
1400 size = data_end - *data;
1409 static void get_tile_buffers(VP9Decoder *pbi,
1410 const uint8_t *data, const uint8_t *data_end,
1411 int tile_cols, int tile_rows,
1412 TileBuffer (*tile_buffers)[1 << 6]) {
1415 for (r = 0; r < tile_rows; ++r) {
1416 for (c = 0; c < tile_cols; ++c) {
1417 const int is_last = (r == tile_rows - 1) && (c == tile_cols - 1);
1418 TileBuffer *const buf = &tile_buffers[r][c];
1420 get_tile_buffer(data_end, is_last, &pbi->common.error, &data,
1421 pbi->decrypt_cb, pbi->decrypt_state, buf);
1426 static const uint8_t *decode_tiles(VP9Decoder *pbi,
1427 const uint8_t *data,
1428 const uint8_t *data_end) {
1429 VP9_COMMON *const cm = &pbi->common;
1430 const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
1431 const int aligned_cols = mi_cols_aligned_to_sb(cm->mi_cols);
1432 const int tile_cols = 1 << cm->log2_tile_cols;
1433 const int tile_rows = 1 << cm->log2_tile_rows;
1434 TileBuffer tile_buffers[4][1 << 6];
1435 int tile_row, tile_col;
1437 TileData *tile_data = NULL;
1439 if (cm->lf.filter_level && !cm->skip_loop_filter &&
1440 pbi->lf_worker.data1 == NULL) {
1441 CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
1442 vpx_memalign(32, sizeof(LFWorkerData)));
1443 pbi->lf_worker.hook = (VPxWorkerHook)vp9_loop_filter_worker;
1444 if (pbi->max_threads > 1 && !winterface->reset(&pbi->lf_worker)) {
1445 vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1446 "Loop filter thread creation failed");
1450 if (cm->lf.filter_level && !cm->skip_loop_filter) {
1451 LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
1452 // Be sure to sync as we might be resuming after a failed frame decode.
1453 winterface->sync(&pbi->lf_worker);
1454 vp9_loop_filter_data_reset(lf_data, get_frame_new_buffer(cm), cm,
1458 assert(tile_rows <= 4);
1459 assert(tile_cols <= (1 << 6));
1461 // Note: this memset assumes above_context[0], [1] and [2]
1462 // are allocated as part of the same buffer.
1463 memset(cm->above_context, 0,
1464 sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_cols);
1466 memset(cm->above_seg_context, 0,
1467 sizeof(*cm->above_seg_context) * aligned_cols);
1471 get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
1473 if (pbi->tile_data == NULL ||
1474 (tile_cols * tile_rows) != pbi->total_tiles) {
1475 vpx_free(pbi->tile_data);
1479 vpx_memalign(32, tile_cols * tile_rows * (sizeof(*pbi->tile_data))));
1480 pbi->total_tiles = tile_rows * tile_cols;
1483 // Load all tile information into tile_data.
1484 for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
1485 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
1486 const TileBuffer *const buf = &tile_buffers[tile_row][tile_col];
1487 tile_data = pbi->tile_data + tile_cols * tile_row + tile_col;
1489 tile_data->xd = pbi->mb;
1490 tile_data->xd.corrupted = 0;
1491 tile_data->xd.counts = cm->frame_parallel_decoding_mode ?
1493 vp9_zero(tile_data->dqcoeff);
1494 vp9_tile_init(&tile_data->xd.tile, tile_data->cm, tile_row, tile_col);
1495 setup_token_decoder(buf->data, data_end, buf->size, &cm->error,
1496 &tile_data->bit_reader, pbi->decrypt_cb,
1497 pbi->decrypt_state);
1498 vp9_init_macroblockd(cm, &tile_data->xd, tile_data->dqcoeff);
1502 for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
1504 vp9_tile_set_row(&tile, cm, tile_row);
1505 for (mi_row = tile.mi_row_start; mi_row < tile.mi_row_end;
1506 mi_row += MI_BLOCK_SIZE) {
1507 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
1508 const int col = pbi->inv_tile_order ?
1509 tile_cols - tile_col - 1 : tile_col;
1510 tile_data = pbi->tile_data + tile_cols * tile_row + col;
1511 vp9_tile_set_col(&tile, tile_data->cm, col);
1512 vp9_zero(tile_data->xd.left_context);
1513 vp9_zero(tile_data->xd.left_seg_context);
1514 for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
1515 mi_col += MI_BLOCK_SIZE) {
1516 decode_partition(pbi, &tile_data->xd, mi_row,
1517 mi_col, &tile_data->bit_reader, BLOCK_64X64, 4);
1519 pbi->mb.corrupted |= tile_data->xd.corrupted;
1520 if (pbi->mb.corrupted)
1521 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1522 "Failed to decode tile data");
1524 // Loopfilter one row.
1525 if (cm->lf.filter_level && !cm->skip_loop_filter) {
1526 const int lf_start = mi_row - MI_BLOCK_SIZE;
1527 LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
1529 // delay the loopfilter by 1 macroblock row.
1530 if (lf_start < 0) continue;
1532 // decoding has completed: finish up the loop filter in this thread.
1533 if (mi_row + MI_BLOCK_SIZE >= cm->mi_rows) continue;
1535 winterface->sync(&pbi->lf_worker);
1536 lf_data->start = lf_start;
1537 lf_data->stop = mi_row;
1538 if (pbi->max_threads > 1) {
1539 winterface->launch(&pbi->lf_worker);
1541 winterface->execute(&pbi->lf_worker);
1544 // After loopfiltering, the last 7 row pixels in each superblock row may
1545 // still be changed by the longest loopfilter of the next superblock
1547 if (pbi->frame_parallel_decode)
1548 vp9_frameworker_broadcast(pbi->cur_buf,
1549 mi_row << MI_BLOCK_SIZE_LOG2);
1553 // Loopfilter remaining rows in the frame.
1554 if (cm->lf.filter_level && !cm->skip_loop_filter) {
1555 LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
1556 winterface->sync(&pbi->lf_worker);
1557 lf_data->start = lf_data->stop;
1558 lf_data->stop = cm->mi_rows;
1559 winterface->execute(&pbi->lf_worker);
1562 // Get last tile data.
1563 tile_data = pbi->tile_data + tile_cols * tile_rows - 1;
1565 if (pbi->frame_parallel_decode)
1566 vp9_frameworker_broadcast(pbi->cur_buf, INT_MAX);
1567 return vpx_reader_find_end(&tile_data->bit_reader);
1570 // On entry 'tile_data->data_end' points to the end of the input frame, on exit
1571 // it is updated to reflect the bitreader position of the final tile column if
1572 // present in the tile buffer group or NULL otherwise.
1573 static int tile_worker_hook(TileWorkerData *const tile_data,
1574 VP9Decoder *const pbi) {
1575 TileInfo *volatile tile = &tile_data->xd.tile;
1576 const int final_col = (1 << pbi->common.log2_tile_cols) - 1;
1577 const uint8_t *volatile bit_reader_end = NULL;
1578 volatile int n = tile_data->buf_start;
1579 tile_data->error_info.setjmp = 1;
1581 if (setjmp(tile_data->error_info.jmp)) {
1582 tile_data->error_info.setjmp = 0;
1583 tile_data->xd.corrupted = 1;
1584 tile_data->data_end = NULL;
1588 tile_data->xd.error_info = &tile_data->error_info;
1589 tile_data->xd.corrupted = 0;
1593 const TileBuffer *const buf = pbi->tile_buffers + n;
1594 vp9_zero(tile_data->dqcoeff);
1595 vp9_tile_init(tile, &pbi->common, 0, buf->col);
1596 setup_token_decoder(buf->data, tile_data->data_end, buf->size,
1597 &tile_data->error_info, &tile_data->bit_reader,
1598 pbi->decrypt_cb, pbi->decrypt_state);
1599 vp9_init_macroblockd(&pbi->common, &tile_data->xd, tile_data->dqcoeff);
1601 for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
1602 mi_row += MI_BLOCK_SIZE) {
1603 vp9_zero(tile_data->xd.left_context);
1604 vp9_zero(tile_data->xd.left_seg_context);
1605 for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
1606 mi_col += MI_BLOCK_SIZE) {
1607 decode_partition(pbi, &tile_data->xd, mi_row, mi_col,
1608 &tile_data->bit_reader, BLOCK_64X64, 4);
1612 if (buf->col == final_col) {
1613 bit_reader_end = vpx_reader_find_end(&tile_data->bit_reader);
1615 } while (!tile_data->xd.corrupted && ++n <= tile_data->buf_end);
1617 tile_data->data_end = bit_reader_end;
1618 return !tile_data->xd.corrupted;
1621 // sorts in descending order
1622 static int compare_tile_buffers(const void *a, const void *b) {
1623 const TileBuffer *const buf1 = (const TileBuffer*)a;
1624 const TileBuffer *const buf2 = (const TileBuffer*)b;
1625 return (int)(buf2->size - buf1->size);
1628 static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
1629 const uint8_t *data,
1630 const uint8_t *data_end) {
1631 VP9_COMMON *const cm = &pbi->common;
1632 const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
1633 const uint8_t *bit_reader_end = NULL;
1634 const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
1635 const int tile_cols = 1 << cm->log2_tile_cols;
1636 const int tile_rows = 1 << cm->log2_tile_rows;
1637 const int num_workers = VPXMIN(pbi->max_threads, tile_cols);
1640 assert(tile_cols <= (1 << 6));
1641 assert(tile_rows == 1);
1644 if (pbi->num_tile_workers == 0) {
1645 const int num_threads = pbi->max_threads;
1646 CHECK_MEM_ERROR(cm, pbi->tile_workers,
1647 vpx_malloc(num_threads * sizeof(*pbi->tile_workers)));
1648 // Ensure tile data offsets will be properly aligned. This may fail on
1649 // platforms without DECLARE_ALIGNED().
1650 assert((sizeof(*pbi->tile_worker_data) % 16) == 0);
1651 CHECK_MEM_ERROR(cm, pbi->tile_worker_data,
1652 vpx_memalign(32, num_threads *
1653 sizeof(*pbi->tile_worker_data)));
1654 for (n = 0; n < num_threads; ++n) {
1655 VPxWorker *const worker = &pbi->tile_workers[n];
1656 ++pbi->num_tile_workers;
1658 winterface->init(worker);
1659 if (n < num_threads - 1 && !winterface->reset(worker)) {
1660 vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1661 "Tile decoder thread creation failed");
1666 // Reset tile decoding hook
1667 for (n = 0; n < num_workers; ++n) {
1668 VPxWorker *const worker = &pbi->tile_workers[n];
1669 TileWorkerData *const tile_data = &pbi->tile_worker_data[n];
1670 winterface->sync(worker);
1671 tile_data->xd = pbi->mb;
1672 tile_data->xd.counts =
1673 cm->frame_parallel_decoding_mode ? NULL : &tile_data->counts;
1674 worker->hook = (VPxWorkerHook)tile_worker_hook;
1675 worker->data1 = tile_data;
1676 worker->data2 = pbi;
1679 // Note: this memset assumes above_context[0], [1] and [2]
1680 // are allocated as part of the same buffer.
1681 memset(cm->above_context, 0,
1682 sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_mi_cols);
1683 memset(cm->above_seg_context, 0,
1684 sizeof(*cm->above_seg_context) * aligned_mi_cols);
1688 // Load tile data into tile_buffers
1689 get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows,
1690 &pbi->tile_buffers);
1692 // Sort the buffers based on size in descending order.
1693 qsort(pbi->tile_buffers, tile_cols, sizeof(pbi->tile_buffers[0]),
1694 compare_tile_buffers);
1696 if (num_workers == tile_cols) {
1697 // Rearrange the tile buffers such that the largest, and
1698 // presumably the most difficult, tile will be decoded in the main thread.
1699 // This should help minimize the number of instances where the main thread
1700 // is waiting for a worker to complete.
1701 const TileBuffer largest = pbi->tile_buffers[0];
1702 memmove(pbi->tile_buffers, pbi->tile_buffers + 1,
1703 (tile_cols - 1) * sizeof(pbi->tile_buffers[0]));
1704 pbi->tile_buffers[tile_cols - 1] = largest;
1706 int start = 0, end = tile_cols - 2;
1709 // Interleave the tiles to distribute the load between threads, assuming a
1710 // larger tile implies it is more difficult to decode.
1711 while (start < end) {
1712 tmp = pbi->tile_buffers[start];
1713 pbi->tile_buffers[start] = pbi->tile_buffers[end];
1714 pbi->tile_buffers[end] = tmp;
1720 // Initialize thread frame counts.
1721 if (!cm->frame_parallel_decoding_mode) {
1722 for (n = 0; n < num_workers; ++n) {
1723 TileWorkerData *const tile_data =
1724 (TileWorkerData*)pbi->tile_workers[n].data1;
1725 vp9_zero(tile_data->counts);
1730 const int base = tile_cols / num_workers;
1731 const int remain = tile_cols % num_workers;
1734 for (n = 0; n < num_workers; ++n) {
1735 const int count = base + (remain + n) / num_workers;
1736 VPxWorker *const worker = &pbi->tile_workers[n];
1737 TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
1739 tile_data->buf_start = buf_start;
1740 tile_data->buf_end = buf_start + count - 1;
1741 tile_data->data_end = data_end;
1744 worker->had_error = 0;
1745 if (n == num_workers - 1) {
1746 assert(tile_data->buf_end == tile_cols - 1);
1747 winterface->execute(worker);
1749 winterface->launch(worker);
1753 for (; n > 0; --n) {
1754 VPxWorker *const worker = &pbi->tile_workers[n - 1];
1755 TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
1756 // TODO(jzern): The tile may have specific error data associated with
1757 // its vpx_internal_error_info which could be propagated to the main info
1758 // in cm. Additionally once the threads have been synced and an error is
1759 // detected, there's no point in continuing to decode tiles.
1760 pbi->mb.corrupted |= !winterface->sync(worker);
1761 if (!bit_reader_end) bit_reader_end = tile_data->data_end;
1765 // Accumulate thread frame counts.
1766 if (!cm->frame_parallel_decoding_mode) {
1767 for (n = 0; n < num_workers; ++n) {
1768 TileWorkerData *const tile_data =
1769 (TileWorkerData*)pbi->tile_workers[n].data1;
1770 vp9_accumulate_frame_counts(&cm->counts, &tile_data->counts, 1);
1774 assert(bit_reader_end || pbi->mb.corrupted);
1775 return bit_reader_end;
1778 static void error_handler(void *data) {
1779 VP9_COMMON *const cm = (VP9_COMMON *)data;
1780 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME, "Truncated packet");
1783 static void read_bitdepth_colorspace_sampling(
1784 VP9_COMMON *cm, struct vpx_read_bit_buffer *rb) {
1785 if (cm->profile >= PROFILE_2) {
1786 cm->bit_depth = vpx_rb_read_bit(rb) ? VPX_BITS_12 : VPX_BITS_10;
1787 #if CONFIG_VP9_HIGHBITDEPTH
1788 cm->use_highbitdepth = 1;
1791 cm->bit_depth = VPX_BITS_8;
1792 #if CONFIG_VP9_HIGHBITDEPTH
1793 cm->use_highbitdepth = 0;
1796 cm->color_space = vpx_rb_read_literal(rb, 3);
1797 if (cm->color_space != VPX_CS_SRGB) {
1798 cm->color_range = (vpx_color_range_t)vpx_rb_read_bit(rb);
1799 if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
1800 cm->subsampling_x = vpx_rb_read_bit(rb);
1801 cm->subsampling_y = vpx_rb_read_bit(rb);
1802 if (cm->subsampling_x == 1 && cm->subsampling_y == 1)
1803 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1804 "4:2:0 color not supported in profile 1 or 3");
1805 if (vpx_rb_read_bit(rb))
1806 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1807 "Reserved bit set");
1809 cm->subsampling_y = cm->subsampling_x = 1;
1812 cm->color_range = VPX_CR_FULL_RANGE;
1813 if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
1814 // Note if colorspace is SRGB then 4:4:4 chroma sampling is assumed.
1815 // 4:2:2 or 4:4:0 chroma sampling is not allowed.
1816 cm->subsampling_y = cm->subsampling_x = 0;
1817 if (vpx_rb_read_bit(rb))
1818 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1819 "Reserved bit set");
1821 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1822 "4:4:4 color not supported in profile 0 or 2");
1827 static size_t read_uncompressed_header(VP9Decoder *pbi,
1828 struct vpx_read_bit_buffer *rb) {
1829 VP9_COMMON *const cm = &pbi->common;
1830 BufferPool *const pool = cm->buffer_pool;
1831 RefCntBuffer *const frame_bufs = pool->frame_bufs;
1832 int i, mask, ref_index = 0;
1835 cm->last_frame_type = cm->frame_type;
1836 cm->last_intra_only = cm->intra_only;
1838 if (vpx_rb_read_literal(rb, 2) != VP9_FRAME_MARKER)
1839 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1840 "Invalid frame marker");
1842 cm->profile = vp9_read_profile(rb);
1843 #if CONFIG_VP9_HIGHBITDEPTH
1844 if (cm->profile >= MAX_PROFILES)
1845 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1846 "Unsupported bitstream profile");
1848 if (cm->profile >= PROFILE_2)
1849 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1850 "Unsupported bitstream profile");
1853 cm->show_existing_frame = vpx_rb_read_bit(rb);
1854 if (cm->show_existing_frame) {
1855 // Show an existing frame directly.
1856 const int frame_to_show = cm->ref_frame_map[vpx_rb_read_literal(rb, 3)];
1857 lock_buffer_pool(pool);
1858 if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1) {
1859 unlock_buffer_pool(pool);
1860 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1861 "Buffer %d does not contain a decoded frame",
1865 ref_cnt_fb(frame_bufs, &cm->new_fb_idx, frame_to_show);
1866 unlock_buffer_pool(pool);
1867 pbi->refresh_frame_flags = 0;
1868 cm->lf.filter_level = 0;
1871 if (pbi->frame_parallel_decode) {
1872 for (i = 0; i < REF_FRAMES; ++i)
1873 cm->next_ref_frame_map[i] = cm->ref_frame_map[i];
1878 cm->frame_type = (FRAME_TYPE) vpx_rb_read_bit(rb);
1879 cm->show_frame = vpx_rb_read_bit(rb);
1880 cm->error_resilient_mode = vpx_rb_read_bit(rb);
1882 if (cm->frame_type == KEY_FRAME) {
1883 if (!vp9_read_sync_code(rb))
1884 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1885 "Invalid frame sync code");
1887 read_bitdepth_colorspace_sampling(cm, rb);
1888 pbi->refresh_frame_flags = (1 << REF_FRAMES) - 1;
1890 for (i = 0; i < REFS_PER_FRAME; ++i) {
1891 cm->frame_refs[i].idx = INVALID_IDX;
1892 cm->frame_refs[i].buf = NULL;
1895 setup_frame_size(cm, rb);
1896 if (pbi->need_resync) {
1897 memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
1898 pbi->need_resync = 0;
1901 cm->intra_only = cm->show_frame ? 0 : vpx_rb_read_bit(rb);
1903 cm->reset_frame_context = cm->error_resilient_mode ?
1904 0 : vpx_rb_read_literal(rb, 2);
1906 if (cm->intra_only) {
1907 if (!vp9_read_sync_code(rb))
1908 vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1909 "Invalid frame sync code");
1910 if (cm->profile > PROFILE_0) {
1911 read_bitdepth_colorspace_sampling(cm, rb);
1913 // NOTE: The intra-only frame header does not include the specification
1914 // of either the color format or color sub-sampling in profile 0. VP9
1915 // specifies that the default color format should be YUV 4:2:0 in this
1916 // case (normative).
1917 cm->color_space = VPX_CS_BT_601;
1918 cm->color_range = VPX_CR_STUDIO_RANGE;
1919 cm->subsampling_y = cm->subsampling_x = 1;
1920 cm->bit_depth = VPX_BITS_8;
1921 #if CONFIG_VP9_HIGHBITDEPTH
1922 cm->use_highbitdepth = 0;
1926 pbi->refresh_frame_flags = vpx_rb_read_literal(rb, REF_FRAMES);
1927 setup_frame_size(cm, rb);
1928 if (pbi->need_resync) {
1929 memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
1930 pbi->need_resync = 0;
1932 } else if (pbi->need_resync != 1) { /* Skip if need resync */
1933 pbi->refresh_frame_flags = vpx_rb_read_literal(rb, REF_FRAMES);
1934 for (i = 0; i < REFS_PER_FRAME; ++i) {
1935 const int ref = vpx_rb_read_literal(rb, REF_FRAMES_LOG2);
1936 const int idx = cm->ref_frame_map[ref];
1937 RefBuffer *const ref_frame = &cm->frame_refs[i];
1938 ref_frame->idx = idx;
1939 ref_frame->buf = &frame_bufs[idx].buf;
1940 cm->ref_frame_sign_bias[LAST_FRAME + i] = vpx_rb_read_bit(rb);
1943 setup_frame_size_with_refs(cm, rb);
1945 cm->allow_high_precision_mv = vpx_rb_read_bit(rb);
1946 cm->interp_filter = read_interp_filter(rb);
1948 for (i = 0; i < REFS_PER_FRAME; ++i) {
1949 RefBuffer *const ref_buf = &cm->frame_refs[i];
1950 #if CONFIG_VP9_HIGHBITDEPTH
1951 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
1952 ref_buf->buf->y_crop_width,
1953 ref_buf->buf->y_crop_height,
1954 cm->width, cm->height,
1955 cm->use_highbitdepth);
1957 vp9_setup_scale_factors_for_frame(&ref_buf->sf,
1958 ref_buf->buf->y_crop_width,
1959 ref_buf->buf->y_crop_height,
1960 cm->width, cm->height);
1965 #if CONFIG_VP9_HIGHBITDEPTH
1966 get_frame_new_buffer(cm)->bit_depth = cm->bit_depth;
1968 get_frame_new_buffer(cm)->color_space = cm->color_space;
1969 get_frame_new_buffer(cm)->color_range = cm->color_range;
1970 get_frame_new_buffer(cm)->render_width = cm->render_width;
1971 get_frame_new_buffer(cm)->render_height = cm->render_height;
1973 if (pbi->need_resync) {
1974 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1975 "Keyframe / intra-only frame required to reset decoder"
1979 if (!cm->error_resilient_mode) {
1980 cm->refresh_frame_context = vpx_rb_read_bit(rb);
1981 cm->frame_parallel_decoding_mode = vpx_rb_read_bit(rb);
1983 cm->refresh_frame_context = 0;
1984 cm->frame_parallel_decoding_mode = 1;
1987 // This flag will be overridden by the call to vp9_setup_past_independence
1988 // below, forcing the use of context 0 for those frame types.
1989 cm->frame_context_idx = vpx_rb_read_literal(rb, FRAME_CONTEXTS_LOG2);
1991 // Generate next_ref_frame_map.
1992 lock_buffer_pool(pool);
1993 for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
1995 cm->next_ref_frame_map[ref_index] = cm->new_fb_idx;
1996 ++frame_bufs[cm->new_fb_idx].ref_count;
1998 cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
2000 // Current thread holds the reference frame.
2001 if (cm->ref_frame_map[ref_index] >= 0)
2002 ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
2006 for (; ref_index < REF_FRAMES; ++ref_index) {
2007 cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
2008 // Current thread holds the reference frame.
2009 if (cm->ref_frame_map[ref_index] >= 0)
2010 ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
2012 unlock_buffer_pool(pool);
2013 pbi->hold_ref_buf = 1;
2015 if (frame_is_intra_only(cm) || cm->error_resilient_mode)
2016 vp9_setup_past_independence(cm);
2018 setup_loopfilter(&cm->lf, rb);
2019 setup_quantization(cm, &pbi->mb, rb);
2020 setup_segmentation(&cm->seg, rb);
2021 setup_segmentation_dequant(cm);
2023 setup_tile_info(cm, rb);
2024 sz = vpx_rb_read_literal(rb, 16);
2027 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
2028 "Invalid header size");
2033 static int read_compressed_header(VP9Decoder *pbi, const uint8_t *data,
2034 size_t partition_size) {
2035 VP9_COMMON *const cm = &pbi->common;
2036 MACROBLOCKD *const xd = &pbi->mb;
2037 FRAME_CONTEXT *const fc = cm->fc;
2041 if (vpx_reader_init(&r, data, partition_size, pbi->decrypt_cb,
2042 pbi->decrypt_state))
2043 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
2044 "Failed to allocate bool decoder 0");
2046 cm->tx_mode = xd->lossless ? ONLY_4X4 : read_tx_mode(&r);
2047 if (cm->tx_mode == TX_MODE_SELECT)
2048 read_tx_mode_probs(&fc->tx_probs, &r);
2049 read_coef_probs(fc, cm->tx_mode, &r);
2051 for (k = 0; k < SKIP_CONTEXTS; ++k)
2052 vp9_diff_update_prob(&r, &fc->skip_probs[k]);
2054 if (!frame_is_intra_only(cm)) {
2055 nmv_context *const nmvc = &fc->nmvc;
2058 read_inter_mode_probs(fc, &r);
2060 if (cm->interp_filter == SWITCHABLE)
2061 read_switchable_interp_probs(fc, &r);
2063 for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
2064 vp9_diff_update_prob(&r, &fc->intra_inter_prob[i]);
2066 cm->reference_mode = read_frame_reference_mode(cm, &r);
2067 if (cm->reference_mode != SINGLE_REFERENCE)
2068 setup_compound_reference_mode(cm);
2069 read_frame_reference_mode_probs(cm, &r);
2071 for (j = 0; j < BLOCK_SIZE_GROUPS; j++)
2072 for (i = 0; i < INTRA_MODES - 1; ++i)
2073 vp9_diff_update_prob(&r, &fc->y_mode_prob[j][i]);
2075 for (j = 0; j < PARTITION_CONTEXTS; ++j)
2076 for (i = 0; i < PARTITION_TYPES - 1; ++i)
2077 vp9_diff_update_prob(&r, &fc->partition_prob[j][i]);
2079 read_mv_probs(nmvc, cm->allow_high_precision_mv, &r);
2082 return vpx_reader_has_error(&r);
2086 #define debug_check_frame_counts(cm) (void)0
2088 // Counts should only be incremented when frame_parallel_decoding_mode and
2089 // error_resilient_mode are disabled.
2090 static void debug_check_frame_counts(const VP9_COMMON *const cm) {
2091 FRAME_COUNTS zero_counts;
2092 vp9_zero(zero_counts);
2093 assert(cm->frame_parallel_decoding_mode || cm->error_resilient_mode);
2094 assert(!memcmp(cm->counts.y_mode, zero_counts.y_mode,
2095 sizeof(cm->counts.y_mode)));
2096 assert(!memcmp(cm->counts.uv_mode, zero_counts.uv_mode,
2097 sizeof(cm->counts.uv_mode)));
2098 assert(!memcmp(cm->counts.partition, zero_counts.partition,
2099 sizeof(cm->counts.partition)));
2100 assert(!memcmp(cm->counts.coef, zero_counts.coef,
2101 sizeof(cm->counts.coef)));
2102 assert(!memcmp(cm->counts.eob_branch, zero_counts.eob_branch,
2103 sizeof(cm->counts.eob_branch)));
2104 assert(!memcmp(cm->counts.switchable_interp, zero_counts.switchable_interp,
2105 sizeof(cm->counts.switchable_interp)));
2106 assert(!memcmp(cm->counts.inter_mode, zero_counts.inter_mode,
2107 sizeof(cm->counts.inter_mode)));
2108 assert(!memcmp(cm->counts.intra_inter, zero_counts.intra_inter,
2109 sizeof(cm->counts.intra_inter)));
2110 assert(!memcmp(cm->counts.comp_inter, zero_counts.comp_inter,
2111 sizeof(cm->counts.comp_inter)));
2112 assert(!memcmp(cm->counts.single_ref, zero_counts.single_ref,
2113 sizeof(cm->counts.single_ref)));
2114 assert(!memcmp(cm->counts.comp_ref, zero_counts.comp_ref,
2115 sizeof(cm->counts.comp_ref)));
2116 assert(!memcmp(&cm->counts.tx, &zero_counts.tx, sizeof(cm->counts.tx)));
2117 assert(!memcmp(cm->counts.skip, zero_counts.skip, sizeof(cm->counts.skip)));
2118 assert(!memcmp(&cm->counts.mv, &zero_counts.mv, sizeof(cm->counts.mv)));
2122 static struct vpx_read_bit_buffer *init_read_bit_buffer(
2124 struct vpx_read_bit_buffer *rb,
2125 const uint8_t *data,
2126 const uint8_t *data_end,
2127 uint8_t clear_data[MAX_VP9_HEADER_SIZE]) {
2129 rb->error_handler = error_handler;
2130 rb->error_handler_data = &pbi->common;
2131 if (pbi->decrypt_cb) {
2132 const int n = (int)VPXMIN(MAX_VP9_HEADER_SIZE, data_end - data);
2133 pbi->decrypt_cb(pbi->decrypt_state, data, clear_data, n);
2134 rb->bit_buffer = clear_data;
2135 rb->bit_buffer_end = clear_data + n;
2137 rb->bit_buffer = data;
2138 rb->bit_buffer_end = data_end;
2143 //------------------------------------------------------------------------------
2145 int vp9_read_sync_code(struct vpx_read_bit_buffer *const rb) {
2146 return vpx_rb_read_literal(rb, 8) == VP9_SYNC_CODE_0 &&
2147 vpx_rb_read_literal(rb, 8) == VP9_SYNC_CODE_1 &&
2148 vpx_rb_read_literal(rb, 8) == VP9_SYNC_CODE_2;
2151 void vp9_read_frame_size(struct vpx_read_bit_buffer *rb,
2152 int *width, int *height) {
2153 *width = vpx_rb_read_literal(rb, 16) + 1;
2154 *height = vpx_rb_read_literal(rb, 16) + 1;
2157 BITSTREAM_PROFILE vp9_read_profile(struct vpx_read_bit_buffer *rb) {
2158 int profile = vpx_rb_read_bit(rb);
2159 profile |= vpx_rb_read_bit(rb) << 1;
2161 profile += vpx_rb_read_bit(rb);
2162 return (BITSTREAM_PROFILE) profile;
2165 void vp9_decode_frame(VP9Decoder *pbi,
2166 const uint8_t *data, const uint8_t *data_end,
2167 const uint8_t **p_data_end) {
2168 VP9_COMMON *const cm = &pbi->common;
2169 MACROBLOCKD *const xd = &pbi->mb;
2170 struct vpx_read_bit_buffer rb;
2171 int context_updated = 0;
2172 uint8_t clear_data[MAX_VP9_HEADER_SIZE];
2173 const size_t first_partition_size = read_uncompressed_header(pbi,
2174 init_read_bit_buffer(pbi, &rb, data, data_end, clear_data));
2175 const int tile_rows = 1 << cm->log2_tile_rows;
2176 const int tile_cols = 1 << cm->log2_tile_cols;
2177 YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
2178 xd->cur_buf = new_fb;
2180 if (!first_partition_size) {
2181 // showing a frame directly
2182 *p_data_end = data + (cm->profile <= PROFILE_2 ? 1 : 2);
2186 data += vpx_rb_bytes_read(&rb);
2187 if (!read_is_valid(data, first_partition_size, data_end))
2188 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
2189 "Truncated packet or corrupt header length");
2191 cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
2192 cm->width == cm->last_width &&
2193 cm->height == cm->last_height &&
2194 !cm->last_intra_only &&
2195 cm->last_show_frame &&
2196 (cm->last_frame_type != KEY_FRAME);
2198 vp9_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);
2200 *cm->fc = cm->frame_contexts[cm->frame_context_idx];
2201 if (!cm->fc->initialized)
2202 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
2203 "Uninitialized entropy context.");
2205 vp9_zero(cm->counts);
2208 new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
2209 if (new_fb->corrupted)
2210 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
2211 "Decode failed. Frame data header is corrupted.");
2213 if (cm->lf.filter_level && !cm->skip_loop_filter) {
2214 vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
2217 // If encoded in frame parallel mode, frame context is ready after decoding
2218 // the frame header.
2219 if (pbi->frame_parallel_decode && cm->frame_parallel_decoding_mode) {
2220 VPxWorker *const worker = pbi->frame_worker_owner;
2221 FrameWorkerData *const frame_worker_data = worker->data1;
2222 if (cm->refresh_frame_context) {
2223 context_updated = 1;
2224 cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
2226 vp9_frameworker_lock_stats(worker);
2227 pbi->cur_buf->row = -1;
2228 pbi->cur_buf->col = -1;
2229 frame_worker_data->frame_context_ready = 1;
2230 // Signal the main thread that context is ready.
2231 vp9_frameworker_signal_stats(worker);
2232 vp9_frameworker_unlock_stats(worker);
2235 if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1) {
2236 // Multi-threaded tile decoder
2237 *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
2238 if (!xd->corrupted) {
2239 if (!cm->skip_loop_filter) {
2240 // If multiple threads are used to decode tiles, then we use those
2241 // threads to do parallel loopfiltering.
2242 vp9_loop_filter_frame_mt(new_fb, cm, pbi->mb.plane,
2243 cm->lf.filter_level, 0, 0, pbi->tile_workers,
2244 pbi->num_tile_workers, &pbi->lf_row_sync);
2247 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
2248 "Decode failed. Frame data is corrupted.");
2251 *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
2254 if (!xd->corrupted) {
2255 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
2256 vp9_adapt_coef_probs(cm);
2258 if (!frame_is_intra_only(cm)) {
2259 vp9_adapt_mode_probs(cm);
2260 vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
2263 debug_check_frame_counts(cm);
2266 vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
2267 "Decode failed. Frame data is corrupted.");
2270 // Non frame parallel update frame context here.
2271 if (cm->refresh_frame_context && !context_updated)
2272 cm->frame_contexts[cm->frame_context_idx] = *cm->fc;