1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libvpx/vp9/decoder/vp9_decodframe.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,1389 @@
1.4 +/*
1.5 + * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
1.6 + *
1.7 + * Use of this source code is governed by a BSD-style license
1.8 + * that can be found in the LICENSE file in the root of the source
1.9 + * tree. An additional intellectual property rights grant can be found
1.10 + * in the file PATENTS. All contributing project authors may
1.11 + * be found in the AUTHORS file in the root of the source tree.
1.12 + */
1.13 +
1.14 +#include <assert.h>
1.15 +
1.16 +#include "./vp9_rtcd.h"
1.17 +#include "vpx_mem/vpx_mem.h"
1.18 +#include "vpx_scale/vpx_scale.h"
1.19 +
1.20 +#include "vp9/common/vp9_alloccommon.h"
1.21 +#include "vp9/common/vp9_common.h"
1.22 +#include "vp9/common/vp9_entropy.h"
1.23 +#include "vp9/common/vp9_entropymode.h"
1.24 +#include "vp9/common/vp9_extend.h"
1.25 +#include "vp9/common/vp9_idct.h"
1.26 +#include "vp9/common/vp9_pred_common.h"
1.27 +#include "vp9/common/vp9_quant_common.h"
1.28 +#include "vp9/common/vp9_reconintra.h"
1.29 +#include "vp9/common/vp9_reconinter.h"
1.30 +#include "vp9/common/vp9_seg_common.h"
1.31 +#include "vp9/common/vp9_tile_common.h"
1.32 +
1.33 +#include "vp9/decoder/vp9_dboolhuff.h"
1.34 +#include "vp9/decoder/vp9_decodframe.h"
1.35 +#include "vp9/decoder/vp9_detokenize.h"
1.36 +#include "vp9/decoder/vp9_decodemv.h"
1.37 +#include "vp9/decoder/vp9_dsubexp.h"
1.38 +#include "vp9/decoder/vp9_onyxd_int.h"
1.39 +#include "vp9/decoder/vp9_read_bit_buffer.h"
1.40 +#include "vp9/decoder/vp9_thread.h"
1.41 +#include "vp9/decoder/vp9_treereader.h"
1.42 +
1.43 +typedef struct TileWorkerData {
1.44 + VP9_COMMON *cm;
1.45 + vp9_reader bit_reader;
1.46 + DECLARE_ALIGNED(16, MACROBLOCKD, xd);
1.47 + DECLARE_ALIGNED(16, unsigned char, token_cache[1024]);
1.48 + DECLARE_ALIGNED(16, int16_t, qcoeff[MAX_MB_PLANE][64 * 64]);
1.49 + DECLARE_ALIGNED(16, int16_t, dqcoeff[MAX_MB_PLANE][64 * 64]);
1.50 + DECLARE_ALIGNED(16, uint16_t, eobs[MAX_MB_PLANE][256]);
1.51 +} TileWorkerData;
1.52 +
1.53 +static int read_be32(const uint8_t *p) {
1.54 + return (p[0] << 24) | (p[1] << 16) | (p[2] << 8) | p[3];
1.55 +}
1.56 +
1.57 +static int is_compound_prediction_allowed(const VP9_COMMON *cm) {
1.58 + int i;
1.59 + for (i = 1; i < ALLOWED_REFS_PER_FRAME; ++i)
1.60 + if (cm->ref_frame_sign_bias[i + 1] != cm->ref_frame_sign_bias[1])
1.61 + return 1;
1.62 +
1.63 + return 0;
1.64 +}
1.65 +
1.66 +static void setup_compound_prediction(VP9_COMMON *cm) {
1.67 + if (cm->ref_frame_sign_bias[LAST_FRAME] ==
1.68 + cm->ref_frame_sign_bias[GOLDEN_FRAME]) {
1.69 + cm->comp_fixed_ref = ALTREF_FRAME;
1.70 + cm->comp_var_ref[0] = LAST_FRAME;
1.71 + cm->comp_var_ref[1] = GOLDEN_FRAME;
1.72 + } else if (cm->ref_frame_sign_bias[LAST_FRAME] ==
1.73 + cm->ref_frame_sign_bias[ALTREF_FRAME]) {
1.74 + cm->comp_fixed_ref = GOLDEN_FRAME;
1.75 + cm->comp_var_ref[0] = LAST_FRAME;
1.76 + cm->comp_var_ref[1] = ALTREF_FRAME;
1.77 + } else {
1.78 + cm->comp_fixed_ref = LAST_FRAME;
1.79 + cm->comp_var_ref[0] = GOLDEN_FRAME;
1.80 + cm->comp_var_ref[1] = ALTREF_FRAME;
1.81 + }
1.82 +}
1.83 +
1.84 +// len == 0 is not allowed
1.85 +static int read_is_valid(const uint8_t *start, size_t len, const uint8_t *end) {
1.86 + return start + len > start && start + len <= end;
1.87 +}
1.88 +
1.89 +static int decode_unsigned_max(struct vp9_read_bit_buffer *rb, int max) {
1.90 + const int data = vp9_rb_read_literal(rb, get_unsigned_bits(max));
1.91 + return data > max ? max : data;
1.92 +}
1.93 +
1.94 +static TX_MODE read_tx_mode(vp9_reader *r) {
1.95 + TX_MODE tx_mode = vp9_read_literal(r, 2);
1.96 + if (tx_mode == ALLOW_32X32)
1.97 + tx_mode += vp9_read_bit(r);
1.98 + return tx_mode;
1.99 +}
1.100 +
1.101 +static void read_tx_probs(struct tx_probs *tx_probs, vp9_reader *r) {
1.102 + int i, j;
1.103 +
1.104 + for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
1.105 + for (j = 0; j < TX_SIZES - 3; ++j)
1.106 + vp9_diff_update_prob(r, &tx_probs->p8x8[i][j]);
1.107 +
1.108 + for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
1.109 + for (j = 0; j < TX_SIZES - 2; ++j)
1.110 + vp9_diff_update_prob(r, &tx_probs->p16x16[i][j]);
1.111 +
1.112 + for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
1.113 + for (j = 0; j < TX_SIZES - 1; ++j)
1.114 + vp9_diff_update_prob(r, &tx_probs->p32x32[i][j]);
1.115 +}
1.116 +
1.117 +static void read_switchable_interp_probs(FRAME_CONTEXT *fc, vp9_reader *r) {
1.118 + int i, j;
1.119 + for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
1.120 + for (i = 0; i < SWITCHABLE_FILTERS - 1; ++i)
1.121 + vp9_diff_update_prob(r, &fc->switchable_interp_prob[j][i]);
1.122 +}
1.123 +
1.124 +static void read_inter_mode_probs(FRAME_CONTEXT *fc, vp9_reader *r) {
1.125 + int i, j;
1.126 + for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
1.127 + for (j = 0; j < INTER_MODES - 1; ++j)
1.128 + vp9_diff_update_prob(r, &fc->inter_mode_probs[i][j]);
1.129 +}
1.130 +
1.131 +static INLINE COMPPREDMODE_TYPE read_comp_pred_mode(vp9_reader *r) {
1.132 + COMPPREDMODE_TYPE mode = vp9_read_bit(r);
1.133 + if (mode)
1.134 + mode += vp9_read_bit(r);
1.135 + return mode;
1.136 +}
1.137 +
1.138 +static void read_comp_pred(VP9_COMMON *cm, vp9_reader *r) {
1.139 + int i;
1.140 +
1.141 + const int compound_allowed = is_compound_prediction_allowed(cm);
1.142 + cm->comp_pred_mode = compound_allowed ? read_comp_pred_mode(r)
1.143 + : SINGLE_PREDICTION_ONLY;
1.144 + if (compound_allowed)
1.145 + setup_compound_prediction(cm);
1.146 +
1.147 + if (cm->comp_pred_mode == HYBRID_PREDICTION)
1.148 + for (i = 0; i < COMP_INTER_CONTEXTS; i++)
1.149 + vp9_diff_update_prob(r, &cm->fc.comp_inter_prob[i]);
1.150 +
1.151 + if (cm->comp_pred_mode != COMP_PREDICTION_ONLY)
1.152 + for (i = 0; i < REF_CONTEXTS; i++) {
1.153 + vp9_diff_update_prob(r, &cm->fc.single_ref_prob[i][0]);
1.154 + vp9_diff_update_prob(r, &cm->fc.single_ref_prob[i][1]);
1.155 + }
1.156 +
1.157 + if (cm->comp_pred_mode != SINGLE_PREDICTION_ONLY)
1.158 + for (i = 0; i < REF_CONTEXTS; i++)
1.159 + vp9_diff_update_prob(r, &cm->fc.comp_ref_prob[i]);
1.160 +}
1.161 +
1.162 +static void update_mv_probs(vp9_prob *p, int n, vp9_reader *r) {
1.163 + int i;
1.164 + for (i = 0; i < n; ++i)
1.165 + if (vp9_read(r, NMV_UPDATE_PROB))
1.166 + p[i] = (vp9_read_literal(r, 7) << 1) | 1;
1.167 +}
1.168 +
1.169 +static void read_mv_probs(nmv_context *ctx, int allow_hp, vp9_reader *r) {
1.170 + int i, j;
1.171 +
1.172 + update_mv_probs(ctx->joints, MV_JOINTS - 1, r);
1.173 +
1.174 + for (i = 0; i < 2; ++i) {
1.175 + nmv_component *const comp_ctx = &ctx->comps[i];
1.176 + update_mv_probs(&comp_ctx->sign, 1, r);
1.177 + update_mv_probs(comp_ctx->classes, MV_CLASSES - 1, r);
1.178 + update_mv_probs(comp_ctx->class0, CLASS0_SIZE - 1, r);
1.179 + update_mv_probs(comp_ctx->bits, MV_OFFSET_BITS, r);
1.180 + }
1.181 +
1.182 + for (i = 0; i < 2; ++i) {
1.183 + nmv_component *const comp_ctx = &ctx->comps[i];
1.184 + for (j = 0; j < CLASS0_SIZE; ++j)
1.185 + update_mv_probs(comp_ctx->class0_fp[j], 3, r);
1.186 + update_mv_probs(comp_ctx->fp, 3, r);
1.187 + }
1.188 +
1.189 + if (allow_hp) {
1.190 + for (i = 0; i < 2; ++i) {
1.191 + nmv_component *const comp_ctx = &ctx->comps[i];
1.192 + update_mv_probs(&comp_ctx->class0_hp, 1, r);
1.193 + update_mv_probs(&comp_ctx->hp, 1, r);
1.194 + }
1.195 + }
1.196 +}
1.197 +
1.198 +static void setup_plane_dequants(VP9_COMMON *cm, MACROBLOCKD *xd, int q_index) {
1.199 + int i;
1.200 + xd->plane[0].dequant = cm->y_dequant[q_index];
1.201 +
1.202 + for (i = 1; i < MAX_MB_PLANE; i++)
1.203 + xd->plane[i].dequant = cm->uv_dequant[q_index];
1.204 +}
1.205 +
1.206 +// Allocate storage for each tile column.
1.207 +// TODO(jzern): when max_threads <= 1 the same storage could be used for each
1.208 +// tile.
1.209 +static void alloc_tile_storage(VP9D_COMP *pbi, int tile_rows, int tile_cols) {
1.210 + VP9_COMMON *const cm = &pbi->common;
1.211 + const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
1.212 + int i, tile_row, tile_col;
1.213 +
1.214 + CHECK_MEM_ERROR(cm, pbi->mi_streams,
1.215 + vpx_realloc(pbi->mi_streams, tile_rows * tile_cols *
1.216 + sizeof(*pbi->mi_streams)));
1.217 + for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
1.218 + for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
1.219 + TileInfo tile;
1.220 + vp9_tile_init(&tile, cm, tile_row, tile_col);
1.221 + pbi->mi_streams[tile_row * tile_cols + tile_col] =
1.222 + &cm->mi[tile.mi_row_start * cm->mode_info_stride
1.223 + + tile.mi_col_start];
1.224 + }
1.225 + }
1.226 +
1.227 + // 2 contexts per 'mi unit', so that we have one context per 4x4 txfm
1.228 + // block where mi unit size is 8x8.
1.229 + CHECK_MEM_ERROR(cm, pbi->above_context[0],
1.230 + vpx_realloc(pbi->above_context[0],
1.231 + sizeof(*pbi->above_context[0]) * MAX_MB_PLANE *
1.232 + 2 * aligned_mi_cols));
1.233 + for (i = 1; i < MAX_MB_PLANE; ++i) {
1.234 + pbi->above_context[i] = pbi->above_context[0] +
1.235 + i * sizeof(*pbi->above_context[0]) *
1.236 + 2 * aligned_mi_cols;
1.237 + }
1.238 +
1.239 + // This is sized based on the entire frame. Each tile operates within its
1.240 + // column bounds.
1.241 + CHECK_MEM_ERROR(cm, pbi->above_seg_context,
1.242 + vpx_realloc(pbi->above_seg_context,
1.243 + sizeof(*pbi->above_seg_context) *
1.244 + aligned_mi_cols));
1.245 +}
1.246 +
1.247 +static void inverse_transform_block(MACROBLOCKD* xd, int plane, int block,
1.248 + TX_SIZE tx_size, int x, int y) {
1.249 + struct macroblockd_plane *const pd = &xd->plane[plane];
1.250 + const int eob = pd->eobs[block];
1.251 + if (eob > 0) {
1.252 + TX_TYPE tx_type;
1.253 + const int plane_type = pd->plane_type;
1.254 + const int stride = pd->dst.stride;
1.255 + int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1.256 + uint8_t *const dst = &pd->dst.buf[4 * y * stride + 4 * x];
1.257 +
1.258 + switch (tx_size) {
1.259 + case TX_4X4:
1.260 + tx_type = get_tx_type_4x4(plane_type, xd, block);
1.261 + if (tx_type == DCT_DCT)
1.262 + xd->itxm_add(dqcoeff, dst, stride, eob);
1.263 + else
1.264 + vp9_iht4x4_16_add(dqcoeff, dst, stride, tx_type);
1.265 + break;
1.266 + case TX_8X8:
1.267 + tx_type = get_tx_type_8x8(plane_type, xd);
1.268 + vp9_iht8x8_add(tx_type, dqcoeff, dst, stride, eob);
1.269 + break;
1.270 + case TX_16X16:
1.271 + tx_type = get_tx_type_16x16(plane_type, xd);
1.272 + vp9_iht16x16_add(tx_type, dqcoeff, dst, stride, eob);
1.273 + break;
1.274 + case TX_32X32:
1.275 + tx_type = DCT_DCT;
1.276 + vp9_idct32x32_add(dqcoeff, dst, stride, eob);
1.277 + break;
1.278 + default:
1.279 + assert(!"Invalid transform size");
1.280 + }
1.281 +
1.282 + if (eob == 1) {
1.283 + vpx_memset(dqcoeff, 0, 2 * sizeof(dqcoeff[0]));
1.284 + } else {
1.285 + if (tx_type == DCT_DCT && tx_size <= TX_16X16 && eob <= 10)
1.286 + vpx_memset(dqcoeff, 0, 4 * (4 << tx_size) * sizeof(dqcoeff[0]));
1.287 + else if (tx_size == TX_32X32 && eob <= 34)
1.288 + vpx_memset(dqcoeff, 0, 256 * sizeof(dqcoeff[0]));
1.289 + else
1.290 + vpx_memset(dqcoeff, 0, (16 << (tx_size << 1)) * sizeof(dqcoeff[0]));
1.291 + }
1.292 + }
1.293 +}
1.294 +
1.295 +struct intra_args {
1.296 + VP9_COMMON *cm;
1.297 + MACROBLOCKD *xd;
1.298 + vp9_reader *r;
1.299 + uint8_t *token_cache;
1.300 +};
1.301 +
1.302 +static void predict_and_reconstruct_intra_block(int plane, int block,
1.303 + BLOCK_SIZE plane_bsize,
1.304 + TX_SIZE tx_size, void *arg) {
1.305 + struct intra_args *const args = arg;
1.306 + VP9_COMMON *const cm = args->cm;
1.307 + MACROBLOCKD *const xd = args->xd;
1.308 + struct macroblockd_plane *const pd = &xd->plane[plane];
1.309 + MODE_INFO *const mi = xd->mi_8x8[0];
1.310 + const MB_PREDICTION_MODE mode = (plane == 0)
1.311 + ? ((mi->mbmi.sb_type < BLOCK_8X8) ? mi->bmi[block].as_mode
1.312 + : mi->mbmi.mode)
1.313 + : mi->mbmi.uv_mode;
1.314 + int x, y;
1.315 + uint8_t *dst;
1.316 + txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
1.317 + dst = &pd->dst.buf[4 * y * pd->dst.stride + 4 * x];
1.318 +
1.319 + if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0)
1.320 + extend_for_intra(xd, plane_bsize, plane, block, tx_size);
1.321 +
1.322 + vp9_predict_intra_block(xd, block >> (tx_size << 1),
1.323 + b_width_log2(plane_bsize), tx_size, mode,
1.324 + dst, pd->dst.stride, dst, pd->dst.stride);
1.325 +
1.326 + if (!mi->mbmi.skip_coeff) {
1.327 + vp9_decode_block_tokens(cm, xd, plane, block, plane_bsize, x, y, tx_size,
1.328 + args->r, args->token_cache);
1.329 + inverse_transform_block(xd, plane, block, tx_size, x, y);
1.330 + }
1.331 +}
1.332 +
1.333 +struct inter_args {
1.334 + VP9_COMMON *cm;
1.335 + MACROBLOCKD *xd;
1.336 + vp9_reader *r;
1.337 + int *eobtotal;
1.338 + uint8_t *token_cache;
1.339 +};
1.340 +
1.341 +static void reconstruct_inter_block(int plane, int block,
1.342 + BLOCK_SIZE plane_bsize,
1.343 + TX_SIZE tx_size, void *arg) {
1.344 + struct inter_args *args = arg;
1.345 + VP9_COMMON *const cm = args->cm;
1.346 + MACROBLOCKD *const xd = args->xd;
1.347 + int x, y;
1.348 + txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
1.349 +
1.350 + *args->eobtotal += vp9_decode_block_tokens(cm, xd, plane, block,
1.351 + plane_bsize, x, y, tx_size,
1.352 + args->r, args->token_cache);
1.353 + inverse_transform_block(xd, plane, block, tx_size, x, y);
1.354 +}
1.355 +
1.356 +static void set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1.357 + const TileInfo *const tile,
1.358 + BLOCK_SIZE bsize, int mi_row, int mi_col) {
1.359 + const int bh = num_8x8_blocks_high_lookup[bsize];
1.360 + const int bw = num_8x8_blocks_wide_lookup[bsize];
1.361 + const int offset = mi_row * cm->mode_info_stride + mi_col;
1.362 + const int tile_offset = tile->mi_row_start * cm->mode_info_stride +
1.363 + tile->mi_col_start;
1.364 +
1.365 + xd->mi_8x8 = cm->mi_grid_visible + offset;
1.366 + xd->prev_mi_8x8 = cm->prev_mi_grid_visible + offset;
1.367 +
1.368 + // we are using the mode info context stream here
1.369 + xd->mi_8x8[0] = xd->mi_stream + offset - tile_offset;
1.370 + xd->mi_8x8[0]->mbmi.sb_type = bsize;
1.371 +
1.372 + // Special case: if prev_mi is NULL, the previous mode info context
1.373 + // cannot be used.
1.374 + xd->last_mi = cm->prev_mi ? xd->prev_mi_8x8[0] : NULL;
1.375 +
1.376 + set_skip_context(xd, xd->above_context, xd->left_context, mi_row, mi_col);
1.377 +
1.378 + // Distance of Mb to the various image edges. These are specified to 8th pel
1.379 + // as they are always compared to values that are in 1/8th pel units
1.380 + set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
1.381 +
1.382 + setup_dst_planes(xd, get_frame_new_buffer(cm), mi_row, mi_col);
1.383 +}
1.384 +
1.385 +static void set_ref(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1.386 + int idx, int mi_row, int mi_col) {
1.387 + MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
1.388 + const int ref = mbmi->ref_frame[idx] - LAST_FRAME;
1.389 + const YV12_BUFFER_CONFIG *cfg = get_frame_ref_buffer(cm, ref);
1.390 + const struct scale_factors_common *sfc = &cm->active_ref_scale_comm[ref];
1.391 + if (!vp9_is_valid_scale(sfc))
1.392 + vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1.393 + "Invalid scale factors");
1.394 +
1.395 + xd->scale_factor[idx].sfc = sfc;
1.396 + setup_pre_planes(xd, idx, cfg, mi_row, mi_col, &xd->scale_factor[idx]);
1.397 + xd->corrupted |= cfg->corrupted;
1.398 +}
1.399 +
1.400 +static void decode_modes_b(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1.401 + const TileInfo *const tile,
1.402 + int mi_row, int mi_col,
1.403 + vp9_reader *r, BLOCK_SIZE bsize,
1.404 + uint8_t *token_cache) {
1.405 + const int less8x8 = bsize < BLOCK_8X8;
1.406 + MB_MODE_INFO *mbmi;
1.407 +
1.408 + set_offsets(cm, xd, tile, bsize, mi_row, mi_col);
1.409 + vp9_read_mode_info(cm, xd, tile, mi_row, mi_col, r);
1.410 +
1.411 + if (less8x8)
1.412 + bsize = BLOCK_8X8;
1.413 +
1.414 + // Has to be called after set_offsets
1.415 + mbmi = &xd->mi_8x8[0]->mbmi;
1.416 +
1.417 + if (mbmi->skip_coeff) {
1.418 + reset_skip_context(xd, bsize);
1.419 + } else {
1.420 + if (cm->seg.enabled)
1.421 + setup_plane_dequants(cm, xd, vp9_get_qindex(&cm->seg, mbmi->segment_id,
1.422 + cm->base_qindex));
1.423 + }
1.424 +
1.425 + if (!is_inter_block(mbmi)) {
1.426 + struct intra_args arg = {
1.427 + cm, xd, r, token_cache
1.428 + };
1.429 + foreach_transformed_block(xd, bsize, predict_and_reconstruct_intra_block,
1.430 + &arg);
1.431 + } else {
1.432 + // Setup
1.433 + set_ref(cm, xd, 0, mi_row, mi_col);
1.434 + if (has_second_ref(mbmi))
1.435 + set_ref(cm, xd, 1, mi_row, mi_col);
1.436 +
1.437 + xd->subpix.filter_x = xd->subpix.filter_y =
1.438 + vp9_get_filter_kernel(mbmi->interp_filter);
1.439 +
1.440 + // Prediction
1.441 + vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
1.442 +
1.443 + // Reconstruction
1.444 + if (!mbmi->skip_coeff) {
1.445 + int eobtotal = 0;
1.446 + struct inter_args arg = {
1.447 + cm, xd, r, &eobtotal, token_cache
1.448 + };
1.449 + foreach_transformed_block(xd, bsize, reconstruct_inter_block, &arg);
1.450 + if (!less8x8 && eobtotal == 0)
1.451 + mbmi->skip_coeff = 1; // skip loopfilter
1.452 + }
1.453 + }
1.454 +
1.455 + xd->corrupted |= vp9_reader_has_error(r);
1.456 +}
1.457 +
1.458 +static PARTITION_TYPE read_partition(VP9_COMMON *cm, MACROBLOCKD *xd, int hbs,
1.459 + int mi_row, int mi_col, BLOCK_SIZE bsize,
1.460 + vp9_reader *r) {
1.461 + const int ctx = partition_plane_context(xd->above_seg_context,
1.462 + xd->left_seg_context,
1.463 + mi_row, mi_col, bsize);
1.464 + const vp9_prob *const probs = get_partition_probs(cm, ctx);
1.465 + const int has_rows = (mi_row + hbs) < cm->mi_rows;
1.466 + const int has_cols = (mi_col + hbs) < cm->mi_cols;
1.467 + PARTITION_TYPE p;
1.468 +
1.469 + if (has_rows && has_cols)
1.470 + p = treed_read(r, vp9_partition_tree, probs);
1.471 + else if (!has_rows && has_cols)
1.472 + p = vp9_read(r, probs[1]) ? PARTITION_SPLIT : PARTITION_HORZ;
1.473 + else if (has_rows && !has_cols)
1.474 + p = vp9_read(r, probs[2]) ? PARTITION_SPLIT : PARTITION_VERT;
1.475 + else
1.476 + p = PARTITION_SPLIT;
1.477 +
1.478 + if (!cm->frame_parallel_decoding_mode)
1.479 + ++cm->counts.partition[ctx][p];
1.480 +
1.481 + return p;
1.482 +}
1.483 +
1.484 +static void decode_modes_sb(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1.485 + const TileInfo *const tile,
1.486 + int mi_row, int mi_col,
1.487 + vp9_reader* r, BLOCK_SIZE bsize,
1.488 + uint8_t *token_cache) {
1.489 + const int hbs = num_8x8_blocks_wide_lookup[bsize] / 2;
1.490 + PARTITION_TYPE partition;
1.491 + BLOCK_SIZE subsize;
1.492 +
1.493 + if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1.494 + return;
1.495 +
1.496 + partition = read_partition(cm, xd, hbs, mi_row, mi_col, bsize, r);
1.497 + subsize = get_subsize(bsize, partition);
1.498 + if (subsize < BLOCK_8X8) {
1.499 + decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize, token_cache);
1.500 + } else {
1.501 + switch (partition) {
1.502 + case PARTITION_NONE:
1.503 + decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize, token_cache);
1.504 + break;
1.505 + case PARTITION_HORZ:
1.506 + decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize, token_cache);
1.507 + if (mi_row + hbs < cm->mi_rows)
1.508 + decode_modes_b(cm, xd, tile, mi_row + hbs, mi_col, r, subsize,
1.509 + token_cache);
1.510 + break;
1.511 + case PARTITION_VERT:
1.512 + decode_modes_b(cm, xd, tile, mi_row, mi_col, r, subsize, token_cache);
1.513 + if (mi_col + hbs < cm->mi_cols)
1.514 + decode_modes_b(cm, xd, tile, mi_row, mi_col + hbs, r, subsize,
1.515 + token_cache);
1.516 + break;
1.517 + case PARTITION_SPLIT:
1.518 + decode_modes_sb(cm, xd, tile, mi_row, mi_col, r, subsize,
1.519 + token_cache);
1.520 + decode_modes_sb(cm, xd, tile, mi_row, mi_col + hbs, r, subsize,
1.521 + token_cache);
1.522 + decode_modes_sb(cm, xd, tile, mi_row + hbs, mi_col, r, subsize,
1.523 + token_cache);
1.524 + decode_modes_sb(cm, xd, tile, mi_row + hbs, mi_col + hbs, r, subsize,
1.525 + token_cache);
1.526 + break;
1.527 + default:
1.528 + assert(!"Invalid partition type");
1.529 + }
1.530 + }
1.531 +
1.532 + // update partition context
1.533 + if (bsize >= BLOCK_8X8 &&
1.534 + (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
1.535 + update_partition_context(xd->above_seg_context, xd->left_seg_context,
1.536 + mi_row, mi_col, subsize, bsize);
1.537 +}
1.538 +
1.539 +static void setup_token_decoder(const uint8_t *data,
1.540 + const uint8_t *data_end,
1.541 + size_t read_size,
1.542 + struct vpx_internal_error_info *error_info,
1.543 + vp9_reader *r) {
1.544 + // Validate the calculated partition length. If the buffer
1.545 + // described by the partition can't be fully read, then restrict
1.546 + // it to the portion that can be (for EC mode) or throw an error.
1.547 + if (!read_is_valid(data, read_size, data_end))
1.548 + vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
1.549 + "Truncated packet or corrupt tile length");
1.550 +
1.551 + if (vp9_reader_init(r, data, read_size))
1.552 + vpx_internal_error(error_info, VPX_CODEC_MEM_ERROR,
1.553 + "Failed to allocate bool decoder %d", 1);
1.554 +}
1.555 +
1.556 +static void read_coef_probs_common(vp9_coeff_probs_model *coef_probs,
1.557 + vp9_reader *r) {
1.558 + int i, j, k, l, m;
1.559 +
1.560 + if (vp9_read_bit(r))
1.561 + for (i = 0; i < BLOCK_TYPES; i++)
1.562 + for (j = 0; j < REF_TYPES; j++)
1.563 + for (k = 0; k < COEF_BANDS; k++)
1.564 + for (l = 0; l < PREV_COEF_CONTEXTS; l++)
1.565 + if (k > 0 || l < 3)
1.566 + for (m = 0; m < UNCONSTRAINED_NODES; m++)
1.567 + vp9_diff_update_prob(r, &coef_probs[i][j][k][l][m]);
1.568 +}
1.569 +
1.570 +static void read_coef_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode,
1.571 + vp9_reader *r) {
1.572 + const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
1.573 + TX_SIZE tx_size;
1.574 + for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
1.575 + read_coef_probs_common(fc->coef_probs[tx_size], r);
1.576 +}
1.577 +
1.578 +static void setup_segmentation(struct segmentation *seg,
1.579 + struct vp9_read_bit_buffer *rb) {
1.580 + int i, j;
1.581 +
1.582 + seg->update_map = 0;
1.583 + seg->update_data = 0;
1.584 +
1.585 + seg->enabled = vp9_rb_read_bit(rb);
1.586 + if (!seg->enabled)
1.587 + return;
1.588 +
1.589 + // Segmentation map update
1.590 + seg->update_map = vp9_rb_read_bit(rb);
1.591 + if (seg->update_map) {
1.592 + for (i = 0; i < SEG_TREE_PROBS; i++)
1.593 + seg->tree_probs[i] = vp9_rb_read_bit(rb) ? vp9_rb_read_literal(rb, 8)
1.594 + : MAX_PROB;
1.595 +
1.596 + seg->temporal_update = vp9_rb_read_bit(rb);
1.597 + if (seg->temporal_update) {
1.598 + for (i = 0; i < PREDICTION_PROBS; i++)
1.599 + seg->pred_probs[i] = vp9_rb_read_bit(rb) ? vp9_rb_read_literal(rb, 8)
1.600 + : MAX_PROB;
1.601 + } else {
1.602 + for (i = 0; i < PREDICTION_PROBS; i++)
1.603 + seg->pred_probs[i] = MAX_PROB;
1.604 + }
1.605 + }
1.606 +
1.607 + // Segmentation data update
1.608 + seg->update_data = vp9_rb_read_bit(rb);
1.609 + if (seg->update_data) {
1.610 + seg->abs_delta = vp9_rb_read_bit(rb);
1.611 +
1.612 + vp9_clearall_segfeatures(seg);
1.613 +
1.614 + for (i = 0; i < MAX_SEGMENTS; i++) {
1.615 + for (j = 0; j < SEG_LVL_MAX; j++) {
1.616 + int data = 0;
1.617 + const int feature_enabled = vp9_rb_read_bit(rb);
1.618 + if (feature_enabled) {
1.619 + vp9_enable_segfeature(seg, i, j);
1.620 + data = decode_unsigned_max(rb, vp9_seg_feature_data_max(j));
1.621 + if (vp9_is_segfeature_signed(j))
1.622 + data = vp9_rb_read_bit(rb) ? -data : data;
1.623 + }
1.624 + vp9_set_segdata(seg, i, j, data);
1.625 + }
1.626 + }
1.627 + }
1.628 +}
1.629 +
1.630 +static void setup_loopfilter(struct loopfilter *lf,
1.631 + struct vp9_read_bit_buffer *rb) {
1.632 + lf->filter_level = vp9_rb_read_literal(rb, 6);
1.633 + lf->sharpness_level = vp9_rb_read_literal(rb, 3);
1.634 +
1.635 + // Read in loop filter deltas applied at the MB level based on mode or ref
1.636 + // frame.
1.637 + lf->mode_ref_delta_update = 0;
1.638 +
1.639 + lf->mode_ref_delta_enabled = vp9_rb_read_bit(rb);
1.640 + if (lf->mode_ref_delta_enabled) {
1.641 + lf->mode_ref_delta_update = vp9_rb_read_bit(rb);
1.642 + if (lf->mode_ref_delta_update) {
1.643 + int i;
1.644 +
1.645 + for (i = 0; i < MAX_REF_LF_DELTAS; i++)
1.646 + if (vp9_rb_read_bit(rb))
1.647 + lf->ref_deltas[i] = vp9_rb_read_signed_literal(rb, 6);
1.648 +
1.649 + for (i = 0; i < MAX_MODE_LF_DELTAS; i++)
1.650 + if (vp9_rb_read_bit(rb))
1.651 + lf->mode_deltas[i] = vp9_rb_read_signed_literal(rb, 6);
1.652 + }
1.653 + }
1.654 +}
1.655 +
1.656 +static int read_delta_q(struct vp9_read_bit_buffer *rb, int *delta_q) {
1.657 + const int old = *delta_q;
1.658 + *delta_q = vp9_rb_read_bit(rb) ? vp9_rb_read_signed_literal(rb, 4) : 0;
1.659 + return old != *delta_q;
1.660 +}
1.661 +
1.662 +static void setup_quantization(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1.663 + struct vp9_read_bit_buffer *rb) {
1.664 + int update = 0;
1.665 +
1.666 + cm->base_qindex = vp9_rb_read_literal(rb, QINDEX_BITS);
1.667 + update |= read_delta_q(rb, &cm->y_dc_delta_q);
1.668 + update |= read_delta_q(rb, &cm->uv_dc_delta_q);
1.669 + update |= read_delta_q(rb, &cm->uv_ac_delta_q);
1.670 + if (update)
1.671 + vp9_init_dequantizer(cm);
1.672 +
1.673 + xd->lossless = cm->base_qindex == 0 &&
1.674 + cm->y_dc_delta_q == 0 &&
1.675 + cm->uv_dc_delta_q == 0 &&
1.676 + cm->uv_ac_delta_q == 0;
1.677 +
1.678 + xd->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
1.679 +}
1.680 +
1.681 +static INTERPOLATION_TYPE read_interp_filter_type(
1.682 + struct vp9_read_bit_buffer *rb) {
1.683 + const INTERPOLATION_TYPE literal_to_type[] = { EIGHTTAP_SMOOTH,
1.684 + EIGHTTAP,
1.685 + EIGHTTAP_SHARP,
1.686 + BILINEAR };
1.687 + return vp9_rb_read_bit(rb) ? SWITCHABLE
1.688 + : literal_to_type[vp9_rb_read_literal(rb, 2)];
1.689 +}
1.690 +
1.691 +static void read_frame_size(struct vp9_read_bit_buffer *rb,
1.692 + int *width, int *height) {
1.693 + const int w = vp9_rb_read_literal(rb, 16) + 1;
1.694 + const int h = vp9_rb_read_literal(rb, 16) + 1;
1.695 + *width = w;
1.696 + *height = h;
1.697 +}
1.698 +
1.699 +static void setup_display_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
1.700 + cm->display_width = cm->width;
1.701 + cm->display_height = cm->height;
1.702 + if (vp9_rb_read_bit(rb))
1.703 + read_frame_size(rb, &cm->display_width, &cm->display_height);
1.704 +}
1.705 +
1.706 +static void apply_frame_size(VP9D_COMP *pbi, int width, int height) {
1.707 + VP9_COMMON *cm = &pbi->common;
1.708 +
1.709 + if (cm->width != width || cm->height != height) {
1.710 + // Change in frame size.
1.711 + if (cm->width == 0 || cm->height == 0) {
1.712 + // Assign new frame buffer on first call.
1.713 + cm->new_fb_idx = NUM_YV12_BUFFERS - 1;
1.714 + cm->fb_idx_ref_cnt[cm->new_fb_idx] = 1;
1.715 + }
1.716 +
1.717 + // TODO(agrange) Don't test width/height, check overall size.
1.718 + if (width > cm->width || height > cm->height) {
1.719 + // Rescale frame buffers only if they're not big enough already.
1.720 + if (vp9_resize_frame_buffers(cm, width, height))
1.721 + vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1.722 + "Failed to allocate frame buffers");
1.723 + }
1.724 +
1.725 + cm->width = width;
1.726 + cm->height = height;
1.727 +
1.728 + vp9_update_frame_size(cm);
1.729 + }
1.730 +
1.731 + vp9_realloc_frame_buffer(get_frame_new_buffer(cm), cm->width, cm->height,
1.732 + cm->subsampling_x, cm->subsampling_y,
1.733 + VP9BORDERINPIXELS);
1.734 +}
1.735 +
1.736 +static void setup_frame_size(VP9D_COMP *pbi,
1.737 + struct vp9_read_bit_buffer *rb) {
1.738 + int width, height;
1.739 + read_frame_size(rb, &width, &height);
1.740 + apply_frame_size(pbi, width, height);
1.741 + setup_display_size(&pbi->common, rb);
1.742 +}
1.743 +
1.744 +static void setup_frame_size_with_refs(VP9D_COMP *pbi,
1.745 + struct vp9_read_bit_buffer *rb) {
1.746 + VP9_COMMON *const cm = &pbi->common;
1.747 +
1.748 + int width, height;
1.749 + int found = 0, i;
1.750 + for (i = 0; i < ALLOWED_REFS_PER_FRAME; ++i) {
1.751 + if (vp9_rb_read_bit(rb)) {
1.752 + YV12_BUFFER_CONFIG *const cfg = get_frame_ref_buffer(cm, i);
1.753 + width = cfg->y_crop_width;
1.754 + height = cfg->y_crop_height;
1.755 + found = 1;
1.756 + break;
1.757 + }
1.758 + }
1.759 +
1.760 + if (!found)
1.761 + read_frame_size(rb, &width, &height);
1.762 +
1.763 + if (!width || !height)
1.764 + vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1.765 + "Referenced frame with invalid size");
1.766 +
1.767 + apply_frame_size(pbi, width, height);
1.768 + setup_display_size(cm, rb);
1.769 +}
1.770 +
1.771 +static void setup_tile_context(VP9D_COMP *const pbi, MACROBLOCKD *const xd,
1.772 + int tile_row, int tile_col) {
1.773 + int i;
1.774 + const int tile_cols = 1 << pbi->common.log2_tile_cols;
1.775 + xd->mi_stream = pbi->mi_streams[tile_row * tile_cols + tile_col];
1.776 +
1.777 + for (i = 0; i < MAX_MB_PLANE; ++i) {
1.778 + xd->above_context[i] = pbi->above_context[i];
1.779 + }
1.780 + // see note in alloc_tile_storage().
1.781 + xd->above_seg_context = pbi->above_seg_context;
1.782 +}
1.783 +
1.784 +static void decode_tile(VP9D_COMP *pbi, const TileInfo *const tile,
1.785 + vp9_reader *r) {
1.786 + const int num_threads = pbi->oxcf.max_threads;
1.787 + VP9_COMMON *const cm = &pbi->common;
1.788 + int mi_row, mi_col;
1.789 + MACROBLOCKD *xd = &pbi->mb;
1.790 +
1.791 + if (pbi->do_loopfilter_inline) {
1.792 + LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
1.793 + lf_data->frame_buffer = get_frame_new_buffer(cm);
1.794 + lf_data->cm = cm;
1.795 + lf_data->xd = pbi->mb;
1.796 + lf_data->stop = 0;
1.797 + lf_data->y_only = 0;
1.798 + vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
1.799 + }
1.800 +
1.801 + for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
1.802 + mi_row += MI_BLOCK_SIZE) {
1.803 + // For a SB there are 2 left contexts, each pertaining to a MB row within
1.804 + vp9_zero(xd->left_context);
1.805 + vp9_zero(xd->left_seg_context);
1.806 + for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
1.807 + mi_col += MI_BLOCK_SIZE) {
1.808 + decode_modes_sb(cm, xd, tile, mi_row, mi_col, r, BLOCK_64X64,
1.809 + pbi->token_cache);
1.810 + }
1.811 +
1.812 + if (pbi->do_loopfilter_inline) {
1.813 + const int lf_start = mi_row - MI_BLOCK_SIZE;
1.814 + LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
1.815 +
1.816 + // delay the loopfilter by 1 macroblock row.
1.817 + if (lf_start < 0) continue;
1.818 +
1.819 + // decoding has completed: finish up the loop filter in this thread.
1.820 + if (mi_row + MI_BLOCK_SIZE >= tile->mi_row_end) continue;
1.821 +
1.822 + vp9_worker_sync(&pbi->lf_worker);
1.823 + lf_data->start = lf_start;
1.824 + lf_data->stop = mi_row;
1.825 + if (num_threads > 1) {
1.826 + vp9_worker_launch(&pbi->lf_worker);
1.827 + } else {
1.828 + vp9_worker_execute(&pbi->lf_worker);
1.829 + }
1.830 + }
1.831 + }
1.832 +
1.833 + if (pbi->do_loopfilter_inline) {
1.834 + LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
1.835 +
1.836 + vp9_worker_sync(&pbi->lf_worker);
1.837 + lf_data->start = lf_data->stop;
1.838 + lf_data->stop = cm->mi_rows;
1.839 + vp9_worker_execute(&pbi->lf_worker);
1.840 + }
1.841 +}
1.842 +
1.843 +static void setup_tile_info(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
1.844 + int min_log2_tile_cols, max_log2_tile_cols, max_ones;
1.845 + vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
1.846 +
1.847 + // columns
1.848 + max_ones = max_log2_tile_cols - min_log2_tile_cols;
1.849 + cm->log2_tile_cols = min_log2_tile_cols;
1.850 + while (max_ones-- && vp9_rb_read_bit(rb))
1.851 + cm->log2_tile_cols++;
1.852 +
1.853 + // rows
1.854 + cm->log2_tile_rows = vp9_rb_read_bit(rb);
1.855 + if (cm->log2_tile_rows)
1.856 + cm->log2_tile_rows += vp9_rb_read_bit(rb);
1.857 +}
1.858 +
1.859 +// Reads the next tile returning its size and adjusting '*data' accordingly
1.860 +// based on 'is_last'.
1.861 +static size_t get_tile(const uint8_t *const data_end,
1.862 + int is_last,
1.863 + struct vpx_internal_error_info *error_info,
1.864 + const uint8_t **data) {
1.865 + size_t size;
1.866 +
1.867 + if (!is_last) {
1.868 + if (!read_is_valid(*data, 4, data_end))
1.869 + vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
1.870 + "Truncated packet or corrupt tile length");
1.871 +
1.872 + size = read_be32(*data);
1.873 + *data += 4;
1.874 +
1.875 + if (size > data_end - *data) {
1.876 + vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
1.877 + "Truncated packet or corrupt tile size");
1.878 + }
1.879 + } else {
1.880 + size = data_end - *data;
1.881 + }
1.882 + return size;
1.883 +}
1.884 +
1.885 +typedef struct TileBuffer {
1.886 + const uint8_t *data;
1.887 + size_t size;
1.888 +} TileBuffer;
1.889 +
1.890 +static const uint8_t *decode_tiles(VP9D_COMP *pbi, const uint8_t *data) {
1.891 + VP9_COMMON *const cm = &pbi->common;
1.892 + MACROBLOCKD *const xd = &pbi->mb;
1.893 + const int aligned_cols = mi_cols_aligned_to_sb(cm->mi_cols);
1.894 + const int tile_cols = 1 << cm->log2_tile_cols;
1.895 + const int tile_rows = 1 << cm->log2_tile_rows;
1.896 + TileBuffer tile_buffers[4][1 << 6];
1.897 + int tile_row, tile_col;
1.898 + const uint8_t *const data_end = pbi->source + pbi->source_sz;
1.899 + const uint8_t *end = NULL;
1.900 + vp9_reader r;
1.901 +
1.902 + assert(tile_rows <= 4);
1.903 + assert(tile_cols <= (1 << 6));
1.904 +
1.905 + // Note: this memset assumes above_context[0], [1] and [2]
1.906 + // are allocated as part of the same buffer.
1.907 + vpx_memset(pbi->above_context[0], 0,
1.908 + sizeof(*pbi->above_context[0]) * MAX_MB_PLANE * 2 * aligned_cols);
1.909 +
1.910 + vpx_memset(pbi->above_seg_context, 0,
1.911 + sizeof(*pbi->above_seg_context) * aligned_cols);
1.912 +
1.913 + // Load tile data into tile_buffers
1.914 + for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
1.915 + for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
1.916 + const int last_tile = tile_row == tile_rows - 1 &&
1.917 + tile_col == tile_cols - 1;
1.918 + const size_t size = get_tile(data_end, last_tile, &cm->error, &data);
1.919 + TileBuffer *const buf = &tile_buffers[tile_row][tile_col];
1.920 + buf->data = data;
1.921 + buf->size = size;
1.922 + data += size;
1.923 + }
1.924 + }
1.925 +
1.926 + // Decode tiles using data from tile_buffers
1.927 + for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
1.928 + for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
1.929 + const int col = pbi->oxcf.inv_tile_order ? tile_cols - tile_col - 1
1.930 + : tile_col;
1.931 + const int last_tile = tile_row == tile_rows - 1 &&
1.932 + col == tile_cols - 1;
1.933 + const TileBuffer *const buf = &tile_buffers[tile_row][col];
1.934 + TileInfo tile;
1.935 +
1.936 + vp9_tile_init(&tile, cm, tile_row, col);
1.937 + setup_token_decoder(buf->data, data_end, buf->size, &cm->error, &r);
1.938 + setup_tile_context(pbi, xd, tile_row, col);
1.939 + decode_tile(pbi, &tile, &r);
1.940 +
1.941 + if (last_tile)
1.942 + end = vp9_reader_find_end(&r);
1.943 + }
1.944 + }
1.945 +
1.946 + return end;
1.947 +}
1.948 +
1.949 +static void setup_tile_macroblockd(TileWorkerData *const tile_data) {
1.950 + MACROBLOCKD *xd = &tile_data->xd;
1.951 + struct macroblockd_plane *const pd = xd->plane;
1.952 + int i;
1.953 +
1.954 + for (i = 0; i < MAX_MB_PLANE; ++i) {
1.955 + pd[i].qcoeff = tile_data->qcoeff[i];
1.956 + pd[i].dqcoeff = tile_data->dqcoeff[i];
1.957 + pd[i].eobs = tile_data->eobs[i];
1.958 + vpx_memset(xd->plane[i].dqcoeff, 0, 64 * 64 * sizeof(int16_t));
1.959 + }
1.960 +}
1.961 +
1.962 +static int tile_worker_hook(void *arg1, void *arg2) {
1.963 + TileWorkerData *const tile_data = (TileWorkerData*)arg1;
1.964 + const TileInfo *const tile = (TileInfo*)arg2;
1.965 + int mi_row, mi_col;
1.966 +
1.967 + for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
1.968 + mi_row += MI_BLOCK_SIZE) {
1.969 + vp9_zero(tile_data->xd.left_context);
1.970 + vp9_zero(tile_data->xd.left_seg_context);
1.971 + for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
1.972 + mi_col += MI_BLOCK_SIZE) {
1.973 + decode_modes_sb(tile_data->cm, &tile_data->xd, tile,
1.974 + mi_row, mi_col, &tile_data->bit_reader, BLOCK_64X64,
1.975 + tile_data->token_cache);
1.976 + }
1.977 + }
1.978 + return !tile_data->xd.corrupted;
1.979 +}
1.980 +
1.981 +static const uint8_t *decode_tiles_mt(VP9D_COMP *pbi, const uint8_t *data) {
1.982 + VP9_COMMON *const cm = &pbi->common;
1.983 + const uint8_t *const data_end = pbi->source + pbi->source_sz;
1.984 + const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
1.985 + const int tile_cols = 1 << cm->log2_tile_cols;
1.986 + const int tile_rows = 1 << cm->log2_tile_rows;
1.987 + const int num_workers = MIN(pbi->oxcf.max_threads & ~1, tile_cols);
1.988 + int tile_col = 0;
1.989 +
1.990 + assert(tile_rows == 1);
1.991 + (void)tile_rows;
1.992 +
1.993 + if (num_workers > pbi->num_tile_workers) {
1.994 + int i;
1.995 + CHECK_MEM_ERROR(cm, pbi->tile_workers,
1.996 + vpx_realloc(pbi->tile_workers,
1.997 + num_workers * sizeof(*pbi->tile_workers)));
1.998 + for (i = pbi->num_tile_workers; i < num_workers; ++i) {
1.999 + VP9Worker *const worker = &pbi->tile_workers[i];
1.1000 + ++pbi->num_tile_workers;
1.1001 +
1.1002 + vp9_worker_init(worker);
1.1003 + worker->hook = (VP9WorkerHook)tile_worker_hook;
1.1004 + CHECK_MEM_ERROR(cm, worker->data1,
1.1005 + vpx_memalign(32, sizeof(TileWorkerData)));
1.1006 + CHECK_MEM_ERROR(cm, worker->data2, vpx_malloc(sizeof(TileInfo)));
1.1007 + if (i < num_workers - 1 && !vp9_worker_reset(worker)) {
1.1008 + vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1.1009 + "Tile decoder thread creation failed");
1.1010 + }
1.1011 + }
1.1012 + }
1.1013 +
1.1014 + // Note: this memset assumes above_context[0], [1] and [2]
1.1015 + // are allocated as part of the same buffer.
1.1016 + vpx_memset(pbi->above_context[0], 0,
1.1017 + sizeof(*pbi->above_context[0]) * MAX_MB_PLANE *
1.1018 + 2 * aligned_mi_cols);
1.1019 + vpx_memset(pbi->above_seg_context, 0,
1.1020 + sizeof(*pbi->above_seg_context) * aligned_mi_cols);
1.1021 +
1.1022 + while (tile_col < tile_cols) {
1.1023 + int i;
1.1024 + for (i = 0; i < num_workers && tile_col < tile_cols; ++i) {
1.1025 + VP9Worker *const worker = &pbi->tile_workers[i];
1.1026 + TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
1.1027 + TileInfo *const tile = (TileInfo*)worker->data2;
1.1028 + const size_t size =
1.1029 + get_tile(data_end, tile_col == tile_cols - 1, &cm->error, &data);
1.1030 +
1.1031 + tile_data->cm = cm;
1.1032 + tile_data->xd = pbi->mb;
1.1033 + tile_data->xd.corrupted = 0;
1.1034 + vp9_tile_init(tile, tile_data->cm, 0, tile_col);
1.1035 +
1.1036 + setup_token_decoder(data, data_end, size, &cm->error,
1.1037 + &tile_data->bit_reader);
1.1038 + setup_tile_context(pbi, &tile_data->xd, 0, tile_col);
1.1039 + setup_tile_macroblockd(tile_data);
1.1040 +
1.1041 + worker->had_error = 0;
1.1042 + if (i == num_workers - 1 || tile_col == tile_cols - 1) {
1.1043 + vp9_worker_execute(worker);
1.1044 + } else {
1.1045 + vp9_worker_launch(worker);
1.1046 + }
1.1047 +
1.1048 + data += size;
1.1049 + ++tile_col;
1.1050 + }
1.1051 +
1.1052 + for (; i > 0; --i) {
1.1053 + VP9Worker *const worker = &pbi->tile_workers[i - 1];
1.1054 + pbi->mb.corrupted |= !vp9_worker_sync(worker);
1.1055 + }
1.1056 + }
1.1057 +
1.1058 + {
1.1059 + const int final_worker = (tile_cols + num_workers - 1) % num_workers;
1.1060 + TileWorkerData *const tile_data =
1.1061 + (TileWorkerData*)pbi->tile_workers[final_worker].data1;
1.1062 + return vp9_reader_find_end(&tile_data->bit_reader);
1.1063 + }
1.1064 +}
1.1065 +
1.1066 +static void check_sync_code(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
1.1067 + if (vp9_rb_read_literal(rb, 8) != VP9_SYNC_CODE_0 ||
1.1068 + vp9_rb_read_literal(rb, 8) != VP9_SYNC_CODE_1 ||
1.1069 + vp9_rb_read_literal(rb, 8) != VP9_SYNC_CODE_2) {
1.1070 + vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1.1071 + "Invalid frame sync code");
1.1072 + }
1.1073 +}
1.1074 +
1.1075 +static void error_handler(void *data, size_t bit_offset) {
1.1076 + VP9_COMMON *const cm = (VP9_COMMON *)data;
1.1077 + vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME, "Truncated packet");
1.1078 +}
1.1079 +
1.1080 +#define RESERVED \
1.1081 + if (vp9_rb_read_bit(rb)) \
1.1082 + vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM, \
1.1083 + "Reserved bit must be unset")
1.1084 +
1.1085 +static size_t read_uncompressed_header(VP9D_COMP *pbi,
1.1086 + struct vp9_read_bit_buffer *rb) {
1.1087 + VP9_COMMON *const cm = &pbi->common;
1.1088 + size_t sz;
1.1089 + int i;
1.1090 +
1.1091 + cm->last_frame_type = cm->frame_type;
1.1092 +
1.1093 + if (vp9_rb_read_literal(rb, 2) != VP9_FRAME_MARKER)
1.1094 + vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1.1095 + "Invalid frame marker");
1.1096 +
1.1097 + cm->version = vp9_rb_read_bit(rb);
1.1098 + RESERVED;
1.1099 +
1.1100 + if (vp9_rb_read_bit(rb)) {
1.1101 + // show an existing frame directly
1.1102 + int frame_to_show = cm->ref_frame_map[vp9_rb_read_literal(rb, 3)];
1.1103 + ref_cnt_fb(cm->fb_idx_ref_cnt, &cm->new_fb_idx, frame_to_show);
1.1104 + pbi->refresh_frame_flags = 0;
1.1105 + cm->lf.filter_level = 0;
1.1106 + return 0;
1.1107 + }
1.1108 +
1.1109 + cm->frame_type = (FRAME_TYPE) vp9_rb_read_bit(rb);
1.1110 + cm->show_frame = vp9_rb_read_bit(rb);
1.1111 + cm->error_resilient_mode = vp9_rb_read_bit(rb);
1.1112 +
1.1113 + if (cm->frame_type == KEY_FRAME) {
1.1114 + check_sync_code(cm, rb);
1.1115 +
1.1116 + cm->color_space = vp9_rb_read_literal(rb, 3); // colorspace
1.1117 + if (cm->color_space != SRGB) {
1.1118 + vp9_rb_read_bit(rb); // [16,235] (including xvycc) vs [0,255] range
1.1119 + if (cm->version == 1) {
1.1120 + cm->subsampling_x = vp9_rb_read_bit(rb);
1.1121 + cm->subsampling_y = vp9_rb_read_bit(rb);
1.1122 + vp9_rb_read_bit(rb); // has extra plane
1.1123 + } else {
1.1124 + cm->subsampling_y = cm->subsampling_x = 1;
1.1125 + }
1.1126 + } else {
1.1127 + if (cm->version == 1) {
1.1128 + cm->subsampling_y = cm->subsampling_x = 0;
1.1129 + vp9_rb_read_bit(rb); // has extra plane
1.1130 + } else {
1.1131 + vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
1.1132 + "RGB not supported in profile 0");
1.1133 + }
1.1134 + }
1.1135 +
1.1136 + pbi->refresh_frame_flags = (1 << NUM_REF_FRAMES) - 1;
1.1137 +
1.1138 + for (i = 0; i < ALLOWED_REFS_PER_FRAME; ++i)
1.1139 + cm->active_ref_idx[i] = cm->new_fb_idx;
1.1140 +
1.1141 + setup_frame_size(pbi, rb);
1.1142 + } else {
1.1143 + cm->intra_only = cm->show_frame ? 0 : vp9_rb_read_bit(rb);
1.1144 +
1.1145 + cm->reset_frame_context = cm->error_resilient_mode ?
1.1146 + 0 : vp9_rb_read_literal(rb, 2);
1.1147 +
1.1148 + if (cm->intra_only) {
1.1149 + check_sync_code(cm, rb);
1.1150 +
1.1151 + pbi->refresh_frame_flags = vp9_rb_read_literal(rb, NUM_REF_FRAMES);
1.1152 + setup_frame_size(pbi, rb);
1.1153 + } else {
1.1154 + pbi->refresh_frame_flags = vp9_rb_read_literal(rb, NUM_REF_FRAMES);
1.1155 +
1.1156 + for (i = 0; i < ALLOWED_REFS_PER_FRAME; ++i) {
1.1157 + const int ref = vp9_rb_read_literal(rb, NUM_REF_FRAMES_LOG2);
1.1158 + cm->active_ref_idx[i] = cm->ref_frame_map[ref];
1.1159 + cm->ref_frame_sign_bias[LAST_FRAME + i] = vp9_rb_read_bit(rb);
1.1160 + }
1.1161 +
1.1162 + setup_frame_size_with_refs(pbi, rb);
1.1163 +
1.1164 + cm->allow_high_precision_mv = vp9_rb_read_bit(rb);
1.1165 + cm->mcomp_filter_type = read_interp_filter_type(rb);
1.1166 +
1.1167 + for (i = 0; i < ALLOWED_REFS_PER_FRAME; ++i)
1.1168 + vp9_setup_scale_factors(cm, i);
1.1169 + }
1.1170 + }
1.1171 +
1.1172 + if (!cm->error_resilient_mode) {
1.1173 + cm->refresh_frame_context = vp9_rb_read_bit(rb);
1.1174 + cm->frame_parallel_decoding_mode = vp9_rb_read_bit(rb);
1.1175 + } else {
1.1176 + cm->refresh_frame_context = 0;
1.1177 + cm->frame_parallel_decoding_mode = 1;
1.1178 + }
1.1179 +
1.1180 + // This flag will be overridden by the call to vp9_setup_past_independence
1.1181 + // below, forcing the use of context 0 for those frame types.
1.1182 + cm->frame_context_idx = vp9_rb_read_literal(rb, NUM_FRAME_CONTEXTS_LOG2);
1.1183 +
1.1184 + if (frame_is_intra_only(cm) || cm->error_resilient_mode)
1.1185 + vp9_setup_past_independence(cm);
1.1186 +
1.1187 + setup_loopfilter(&cm->lf, rb);
1.1188 + setup_quantization(cm, &pbi->mb, rb);
1.1189 + setup_segmentation(&cm->seg, rb);
1.1190 +
1.1191 + setup_tile_info(cm, rb);
1.1192 + sz = vp9_rb_read_literal(rb, 16);
1.1193 +
1.1194 + if (sz == 0)
1.1195 + vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1.1196 + "Invalid header size");
1.1197 +
1.1198 + return sz;
1.1199 +}
1.1200 +
1.1201 +static int read_compressed_header(VP9D_COMP *pbi, const uint8_t *data,
1.1202 + size_t partition_size) {
1.1203 + VP9_COMMON *const cm = &pbi->common;
1.1204 + MACROBLOCKD *const xd = &pbi->mb;
1.1205 + FRAME_CONTEXT *const fc = &cm->fc;
1.1206 + vp9_reader r;
1.1207 + int k;
1.1208 +
1.1209 + if (vp9_reader_init(&r, data, partition_size))
1.1210 + vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
1.1211 + "Failed to allocate bool decoder 0");
1.1212 +
1.1213 + cm->tx_mode = xd->lossless ? ONLY_4X4 : read_tx_mode(&r);
1.1214 + if (cm->tx_mode == TX_MODE_SELECT)
1.1215 + read_tx_probs(&fc->tx_probs, &r);
1.1216 + read_coef_probs(fc, cm->tx_mode, &r);
1.1217 +
1.1218 + for (k = 0; k < MBSKIP_CONTEXTS; ++k)
1.1219 + vp9_diff_update_prob(&r, &fc->mbskip_probs[k]);
1.1220 +
1.1221 + if (!frame_is_intra_only(cm)) {
1.1222 + nmv_context *const nmvc = &fc->nmvc;
1.1223 + int i, j;
1.1224 +
1.1225 + read_inter_mode_probs(fc, &r);
1.1226 +
1.1227 + if (cm->mcomp_filter_type == SWITCHABLE)
1.1228 + read_switchable_interp_probs(fc, &r);
1.1229 +
1.1230 + for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
1.1231 + vp9_diff_update_prob(&r, &fc->intra_inter_prob[i]);
1.1232 +
1.1233 + read_comp_pred(cm, &r);
1.1234 +
1.1235 + for (j = 0; j < BLOCK_SIZE_GROUPS; j++)
1.1236 + for (i = 0; i < INTRA_MODES - 1; ++i)
1.1237 + vp9_diff_update_prob(&r, &fc->y_mode_prob[j][i]);
1.1238 +
1.1239 + for (j = 0; j < PARTITION_CONTEXTS; ++j)
1.1240 + for (i = 0; i < PARTITION_TYPES - 1; ++i)
1.1241 + vp9_diff_update_prob(&r, &fc->partition_prob[j][i]);
1.1242 +
1.1243 + read_mv_probs(nmvc, cm->allow_high_precision_mv, &r);
1.1244 + }
1.1245 +
1.1246 + return vp9_reader_has_error(&r);
1.1247 +}
1.1248 +
1.1249 +void vp9_init_dequantizer(VP9_COMMON *cm) {
1.1250 + int q;
1.1251 +
1.1252 + for (q = 0; q < QINDEX_RANGE; q++) {
1.1253 + cm->y_dequant[q][0] = vp9_dc_quant(q, cm->y_dc_delta_q);
1.1254 + cm->y_dequant[q][1] = vp9_ac_quant(q, 0);
1.1255 +
1.1256 + cm->uv_dequant[q][0] = vp9_dc_quant(q, cm->uv_dc_delta_q);
1.1257 + cm->uv_dequant[q][1] = vp9_ac_quant(q, cm->uv_ac_delta_q);
1.1258 + }
1.1259 +}
1.1260 +
1.1261 +#ifdef NDEBUG
1.1262 +#define debug_check_frame_counts(cm) (void)0
1.1263 +#else // !NDEBUG
1.1264 +// Counts should only be incremented when frame_parallel_decoding_mode and
1.1265 +// error_resilient_mode are disabled.
1.1266 +static void debug_check_frame_counts(const VP9_COMMON *const cm) {
1.1267 + FRAME_COUNTS zero_counts;
1.1268 + vp9_zero(zero_counts);
1.1269 + assert(cm->frame_parallel_decoding_mode || cm->error_resilient_mode);
1.1270 + assert(!memcmp(cm->counts.y_mode, zero_counts.y_mode,
1.1271 + sizeof(cm->counts.y_mode)));
1.1272 + assert(!memcmp(cm->counts.uv_mode, zero_counts.uv_mode,
1.1273 + sizeof(cm->counts.uv_mode)));
1.1274 + assert(!memcmp(cm->counts.partition, zero_counts.partition,
1.1275 + sizeof(cm->counts.partition)));
1.1276 + assert(!memcmp(cm->counts.coef, zero_counts.coef,
1.1277 + sizeof(cm->counts.coef)));
1.1278 + assert(!memcmp(cm->counts.eob_branch, zero_counts.eob_branch,
1.1279 + sizeof(cm->counts.eob_branch)));
1.1280 + assert(!memcmp(cm->counts.switchable_interp, zero_counts.switchable_interp,
1.1281 + sizeof(cm->counts.switchable_interp)));
1.1282 + assert(!memcmp(cm->counts.inter_mode, zero_counts.inter_mode,
1.1283 + sizeof(cm->counts.inter_mode)));
1.1284 + assert(!memcmp(cm->counts.intra_inter, zero_counts.intra_inter,
1.1285 + sizeof(cm->counts.intra_inter)));
1.1286 + assert(!memcmp(cm->counts.comp_inter, zero_counts.comp_inter,
1.1287 + sizeof(cm->counts.comp_inter)));
1.1288 + assert(!memcmp(cm->counts.single_ref, zero_counts.single_ref,
1.1289 + sizeof(cm->counts.single_ref)));
1.1290 + assert(!memcmp(cm->counts.comp_ref, zero_counts.comp_ref,
1.1291 + sizeof(cm->counts.comp_ref)));
1.1292 + assert(!memcmp(&cm->counts.tx, &zero_counts.tx, sizeof(cm->counts.tx)));
1.1293 + assert(!memcmp(cm->counts.mbskip, zero_counts.mbskip,
1.1294 + sizeof(cm->counts.mbskip)));
1.1295 + assert(!memcmp(&cm->counts.mv, &zero_counts.mv, sizeof(cm->counts.mv)));
1.1296 +}
1.1297 +#endif // NDEBUG
1.1298 +
1.1299 +int vp9_decode_frame(VP9D_COMP *pbi, const uint8_t **p_data_end) {
1.1300 + int i;
1.1301 + VP9_COMMON *const cm = &pbi->common;
1.1302 + MACROBLOCKD *const xd = &pbi->mb;
1.1303 +
1.1304 + const uint8_t *data = pbi->source;
1.1305 + const uint8_t *const data_end = pbi->source + pbi->source_sz;
1.1306 +
1.1307 + struct vp9_read_bit_buffer rb = { data, data_end, 0, cm, error_handler };
1.1308 + const size_t first_partition_size = read_uncompressed_header(pbi, &rb);
1.1309 + const int keyframe = cm->frame_type == KEY_FRAME;
1.1310 + const int tile_rows = 1 << cm->log2_tile_rows;
1.1311 + const int tile_cols = 1 << cm->log2_tile_cols;
1.1312 + YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
1.1313 +
1.1314 + if (!first_partition_size) {
1.1315 + // showing a frame directly
1.1316 + *p_data_end = data + 1;
1.1317 + return 0;
1.1318 + }
1.1319 +
1.1320 + if (!pbi->decoded_key_frame && !keyframe)
1.1321 + return -1;
1.1322 +
1.1323 + data += vp9_rb_bytes_read(&rb);
1.1324 + if (!read_is_valid(data, first_partition_size, data_end))
1.1325 + vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1.1326 + "Truncated packet or corrupt header length");
1.1327 +
1.1328 + pbi->do_loopfilter_inline =
1.1329 + (cm->log2_tile_rows | cm->log2_tile_cols) == 0 && cm->lf.filter_level;
1.1330 + if (pbi->do_loopfilter_inline && pbi->lf_worker.data1 == NULL) {
1.1331 + CHECK_MEM_ERROR(cm, pbi->lf_worker.data1, vpx_malloc(sizeof(LFWorkerData)));
1.1332 + pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
1.1333 + if (pbi->oxcf.max_threads > 1 && !vp9_worker_reset(&pbi->lf_worker)) {
1.1334 + vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1.1335 + "Loop filter thread creation failed");
1.1336 + }
1.1337 + }
1.1338 +
1.1339 + alloc_tile_storage(pbi, tile_rows, tile_cols);
1.1340 +
1.1341 + xd->mode_info_stride = cm->mode_info_stride;
1.1342 + set_prev_mi(cm);
1.1343 +
1.1344 + setup_plane_dequants(cm, xd, cm->base_qindex);
1.1345 + setup_block_dptrs(xd, cm->subsampling_x, cm->subsampling_y);
1.1346 +
1.1347 + cm->fc = cm->frame_contexts[cm->frame_context_idx];
1.1348 + vp9_zero(cm->counts);
1.1349 + for (i = 0; i < MAX_MB_PLANE; ++i)
1.1350 + vpx_memset(xd->plane[i].dqcoeff, 0, 64 * 64 * sizeof(int16_t));
1.1351 +
1.1352 + xd->corrupted = 0;
1.1353 + new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
1.1354 +
1.1355 + // TODO(jzern): remove frame_parallel_decoding_mode restriction for
1.1356 + // single-frame tile decoding.
1.1357 + if (pbi->oxcf.max_threads > 1 && tile_rows == 1 && tile_cols > 1 &&
1.1358 + cm->frame_parallel_decoding_mode) {
1.1359 + *p_data_end = decode_tiles_mt(pbi, data + first_partition_size);
1.1360 + } else {
1.1361 + *p_data_end = decode_tiles(pbi, data + first_partition_size);
1.1362 + }
1.1363 +
1.1364 + cm->last_width = cm->width;
1.1365 + cm->last_height = cm->height;
1.1366 +
1.1367 + new_fb->corrupted |= xd->corrupted;
1.1368 +
1.1369 + if (!pbi->decoded_key_frame) {
1.1370 + if (keyframe && !new_fb->corrupted)
1.1371 + pbi->decoded_key_frame = 1;
1.1372 + else
1.1373 + vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
1.1374 + "A stream must start with a complete key frame");
1.1375 + }
1.1376 +
1.1377 + if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
1.1378 + vp9_adapt_coef_probs(cm);
1.1379 +
1.1380 + if (!frame_is_intra_only(cm)) {
1.1381 + vp9_adapt_mode_probs(cm);
1.1382 + vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
1.1383 + }
1.1384 + } else {
1.1385 + debug_check_frame_counts(cm);
1.1386 + }
1.1387 +
1.1388 + if (cm->refresh_frame_context)
1.1389 + cm->frame_contexts[cm->frame_context_idx] = cm->fc;
1.1390 +
1.1391 + return 0;
1.1392 +}
