vp9_decodeframe.c

  const int tile_rows = 1 << cm->log2_tile_rows;
  const int num_workers = MIN(pbi->max_threads & ~1, tile_cols);
  TileBuffer tile_buffers[1][1 << 6];
  int n;
  int final_worker = -1;

  assert(tile_cols <= (1 << 6));
  assert(tile_rows == 1);
  (void)tile_rows;

  // TODO(jzern): See if we can remove the restriction of passing in max
  // threads to the decoder.
  if (pbi->num_tile_workers == 0) {
    const int num_threads = pbi->max_threads & ~1;
    int i;
    // TODO(jzern): Allocate one less worker, as in the current code we only
    // use num_threads - 1 workers.
    CHECK_MEM_ERROR(cm, pbi->tile_workers,
                    vpx_malloc(num_threads * sizeof(*pbi->tile_workers)));
    for (i = 0; i < num_threads; ++i) {
      VP9Worker *const worker = &pbi->tile_workers[i];
      ++pbi->num_tile_workers;

      winterface->init(worker);
      CHECK_MEM_ERROR(cm, worker->data1,
                      vpx_memalign(32, sizeof(TileWorkerData)));
      CHECK_MEM_ERROR(cm, worker->data2, vpx_malloc(sizeof(TileInfo)));
      if (i < num_threads - 1 && !winterface->reset(worker)) {
        vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                           "Tile decoder thread creation failed");
      }
    }
  }

  // Reset tile decoding hook
  for (n = 0; n < num_workers; ++n) {
    winterface->sync(&pbi->tile_workers[n]);
    pbi->tile_workers[n].hook = (VP9WorkerHook)tile_worker_hook;
  }

  // Note: this memset assumes above_context[0], [1] and [2]
  // are allocated as part of the same buffer.
  vpx_memset(cm->above_context, 0,
             sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_mi_cols);
  vpx_memset(cm->above_seg_context, 0,
             sizeof(*cm->above_seg_context) * aligned_mi_cols);

  // Load tile data into tile_buffers
  get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);

  // Sort the buffers based on size in descending order.
  qsort(tile_buffers[0], tile_cols, sizeof(tile_buffers[0][0]),
        compare_tile_buffers);

  // Rearrange the tile buffers such that per-tile group the largest, and
  // presumably the most difficult, tile will be decoded in the main thread.
  // This should help minimize the number of instances where the main thread is
  // waiting for a worker to complete.
  {
    int group_start = 0;
    while (group_start < tile_cols) {
      const TileBuffer largest = tile_buffers[0][group_start];
      const int group_end = MIN(group_start + num_workers, tile_cols) - 1;
      memmove(tile_buffers[0] + group_start, tile_buffers[0] + group_start + 1,
              (group_end - group_start) * sizeof(tile_buffers[0][0]));
      tile_buffers[0][group_end] = largest;
      group_start = group_end + 1;
    }
  }

  n = 0;
  while (n < tile_cols) {
    int i;
    for (i = 0; i < num_workers && n < tile_cols; ++i) {
      VP9Worker *const worker = &pbi->tile_workers[i];
      TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
      TileInfo *const tile = (TileInfo*)worker->data2;
      TileBuffer *const buf = &tile_buffers[0][n];

      tile_data->cm = cm;
      tile_data->xd = pbi->mb;
      tile_data->xd.corrupted = 0;
      vp9_tile_init(tile, tile_data->cm, 0, buf->col);
      setup_token_decoder(buf->data, data_end, buf->size, &cm->error,
                          &tile_data->bit_reader, pbi->decrypt_cb,
                          pbi->decrypt_state);
      init_macroblockd(cm, &tile_data->xd);
      vp9_zero(tile_data->xd.dqcoeff);

      worker->had_error = 0;
      if (i == num_workers - 1 || n == tile_cols - 1) {
        winterface->execute(worker);
      } else {
        winterface->launch(worker);
      }

      if (buf->col == tile_cols - 1) {
        final_worker = i;
      }

      ++n;
    }

    for (; i > 0; --i) {
      VP9Worker *const worker = &pbi->tile_workers[i - 1];
      pbi->mb.corrupted |= !winterface->sync(worker);
    }
    if (final_worker > -1) {
      TileWorkerData *const tile_data =
          (TileWorkerData*)pbi->tile_workers[final_worker].data1;
      bit_reader_end = vp9_reader_find_end(&tile_data->bit_reader);
      final_worker = -1;
    }
  }

  return bit_reader_end;
}

static void error_handler(void *data) {
  VP9_COMMON *const cm = (VP9_COMMON *)data;
  vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME, "Truncated packet");
}

int vp9_read_sync_code(struct vp9_read_bit_buffer *const rb) {
  return vp9_rb_read_literal(rb, 8) == VP9_SYNC_CODE_0 &&
         vp9_rb_read_literal(rb, 8) == VP9_SYNC_CODE_1 &&
         vp9_rb_read_literal(rb, 8) == VP9_SYNC_CODE_2;
}

BITSTREAM_PROFILE vp9_read_profile(struct vp9_read_bit_buffer *rb) {
  int profile = vp9_rb_read_bit(rb);
  profile |= vp9_rb_read_bit(rb) << 1;
  if (profile > 2)
    profile += vp9_rb_read_bit(rb);
  return (BITSTREAM_PROFILE) profile;
}

static void read_bitdepth_colorspace_sampling(
    VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
  if (cm->profile >= PROFILE_2)
    cm->bit_depth = vp9_rb_read_bit(rb) ? VPX_BITS_12 : VPX_BITS_10;
  cm->color_space = (COLOR_SPACE)vp9_rb_read_literal(rb, 3);
  if (cm->color_space != SRGB) {
    vp9_rb_read_bit(rb);  // [16,235] (including xvycc) vs [0,255] range
    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
      cm->subsampling_x = vp9_rb_read_bit(rb);
      cm->subsampling_y = vp9_rb_read_bit(rb);
      if (cm->subsampling_x == 1 && cm->subsampling_y == 1)
        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                           "4:2:0 color not supported in profile 1 or 3");
      if (vp9_rb_read_bit(rb))
        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                           "Reserved bit set");
    } else {
      cm->subsampling_y = cm->subsampling_x = 1;
    }
  } else {
    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
      // Note if colorspace is SRGB then 4:4:4 chroma sampling is assumed.
      // 4:2:2 or 4:4:0 chroma sampling is not allowed.
      cm->subsampling_y = cm->subsampling_x = 0;
      if (vp9_rb_read_bit(rb))
        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                           "Reserved bit set");
    } else {
      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                         "4:4:4 color not supported in profile 0 or 2");
    }
  }
}

static size_t read_uncompressed_header(VP9Decoder *pbi,
                                       struct vp9_read_bit_buffer *rb) {
  VP9_COMMON *const cm = &pbi->common;
  size_t sz;
  int i;

  cm->last_frame_type = cm->frame_type;

  if (vp9_rb_read_literal(rb, 2) != VP9_FRAME_MARKER)
      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                         "Invalid frame marker");

  cm->profile = vp9_read_profile(rb);

  if (cm->profile >= MAX_PROFILES)
    vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                       "Unsupported bitstream profile");

  cm->show_existing_frame = vp9_rb_read_bit(rb);
  if (cm->show_existing_frame) {
    // Show an existing frame directly.
    const int frame_to_show = cm->ref_frame_map[vp9_rb_read_literal(rb, 3)];

    if (frame_to_show < 0 || cm->frame_bufs[frame_to_show].ref_count < 1)
      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                         "Buffer %d does not contain a decoded frame",
                         frame_to_show);

    ref_cnt_fb(cm->frame_bufs, &cm->new_fb_idx, frame_to_show);
    pbi->refresh_frame_flags = 0;
    cm->lf.filter_level = 0;
    cm->show_frame = 1;
    return 0;
  }

  cm->frame_type = (FRAME_TYPE) vp9_rb_read_bit(rb);
  cm->show_frame = vp9_rb_read_bit(rb);
  cm->error_resilient_mode = vp9_rb_read_bit(rb);

  if (cm->frame_type == KEY_FRAME) {
    if (!vp9_read_sync_code(rb))
      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                         "Invalid frame sync code");

    read_bitdepth_colorspace_sampling(cm, rb);
    pbi->refresh_frame_flags = (1 << REF_FRAMES) - 1;

    for (i = 0; i < REFS_PER_FRAME; ++i) {
      cm->frame_refs[i].idx = -1;
      cm->frame_refs[i].buf = NULL;
    }

    setup_frame_size(cm, rb);
  } else {
    cm->intra_only = cm->show_frame ? 0 : vp9_rb_read_bit(rb);

    cm->reset_frame_context = cm->error_resilient_mode ?
        0 : vp9_rb_read_literal(rb, 2);

    if (cm->intra_only) {
      if (!vp9_read_sync_code(rb))
        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                           "Invalid frame sync code");
      if (cm->profile > PROFILE_0) {
        read_bitdepth_colorspace_sampling(cm, rb);
      } else {
        // NOTE: The intra-only frame header does not include the specification
        // of either the color format or color sub-sampling in profile 0. VP9
        // specifies that the default color space should be YUV 4:2:0 in this
        // case (normative).
        cm->color_space = BT_601;
        cm->subsampling_y = cm->subsampling_x = 1;
      }

      pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
      setup_frame_size(cm, rb);
    } else {
      pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
      for (i = 0; i < REFS_PER_FRAME; ++i) {
        const int ref = vp9_rb_read_literal(rb, REF_FRAMES_LOG2);
        const int idx = cm->ref_frame_map[ref];
        RefBuffer *const ref_frame = &cm->frame_refs[i];
        ref_frame->idx = idx;
        ref_frame->buf = &cm->frame_bufs[idx].buf;
        cm->ref_frame_sign_bias[LAST_FRAME + i] = vp9_rb_read_bit(rb);
      }

      setup_frame_size_with_refs(cm, rb);

      cm->allow_high_precision_mv = vp9_rb_read_bit(rb);
      cm->interp_filter = read_interp_filter(rb);

      for (i = 0; i < REFS_PER_FRAME; ++i) {
        RefBuffer *const ref_buf = &cm->frame_refs[i];
        vp9_setup_scale_factors_for_frame(&ref_buf->sf,
                                          ref_buf->buf->y_crop_width,
                                          ref_buf->buf->y_crop_height,
                                          cm->width, cm->height);
        if (vp9_is_scaled(&ref_buf->sf))
          vp9_extend_frame_borders(ref_buf->buf);
      }
    }
  }

  if (!cm->error_resilient_mode) {
    cm->refresh_frame_context = vp9_rb_read_bit(rb);
    cm->frame_parallel_decoding_mode = vp9_rb_read_bit(rb);
  } else {
    cm->refresh_frame_context = 0;
    cm->frame_parallel_decoding_mode = 1;
  }

  // This flag will be overridden by the call to vp9_setup_past_independence
  // below, forcing the use of context 0 for those frame types.
  cm->frame_context_idx = vp9_rb_read_literal(rb, FRAME_CONTEXTS_LOG2);

  if (frame_is_intra_only(cm) || cm->error_resilient_mode)
    vp9_setup_past_independence(cm);

  setup_loopfilter(&cm->lf, rb);
  setup_quantization(cm, &pbi->mb, rb);
  setup_segmentation(&cm->seg, rb);

  setup_tile_info(cm, rb);
  sz = vp9_rb_read_literal(rb, 16);

  if (sz == 0)
    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                       "Invalid header size");

  return sz;
}

static int read_compressed_header(VP9Decoder *pbi, const uint8_t *data,
                                  size_t partition_size) {
  VP9_COMMON *const cm = &pbi->common;
  MACROBLOCKD *const xd = &pbi->mb;
  FRAME_CONTEXT *const fc = &cm->fc;
  vp9_reader r;
  int k;

  if (vp9_reader_init(&r, data, partition_size, pbi->decrypt_cb,
                      pbi->decrypt_state))
    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                       "Failed to allocate bool decoder 0");

  cm->tx_mode = xd->lossless ? ONLY_4X4 : read_tx_mode(&r);
  if (cm->tx_mode == TX_MODE_SELECT)
    read_tx_mode_probs(&fc->tx_probs, &r);
  read_coef_probs(fc, cm->tx_mode, &r);

  for (k = 0; k < SKIP_CONTEXTS; ++k)
    vp9_diff_update_prob(&r, &fc->skip_probs[k]);

  if (!frame_is_intra_only(cm)) {
    nmv_context *const nmvc = &fc->nmvc;
    int i, j;

    read_inter_mode_probs(fc, &r);

    if (cm->interp_filter == SWITCHABLE)
      read_switchable_interp_probs(fc, &r);

    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
      vp9_diff_update_prob(&r, &fc->intra_inter_prob[i]);

    cm->reference_mode = read_frame_reference_mode(cm, &r);
    if (cm->reference_mode != SINGLE_REFERENCE)
      setup_compound_reference_mode(cm);
    read_frame_reference_mode_probs(cm, &r);

    for (j = 0; j < BLOCK_SIZE_GROUPS; j++)
      for (i = 0; i < INTRA_MODES - 1; ++i)
        vp9_diff_update_prob(&r, &fc->y_mode_prob[j][i]);

    for (j = 0; j < PARTITION_CONTEXTS; ++j)
      for (i = 0; i < PARTITION_TYPES - 1; ++i)
        vp9_diff_update_prob(&r, &fc->partition_prob[j][i]);

    read_mv_probs(nmvc, cm->allow_high_precision_mv, &r);
  }

  return vp9_reader_has_error(&r);
}

void vp9_init_dequantizer(VP9_COMMON *cm) {
  int q;

  for (q = 0; q < QINDEX_RANGE; q++) {
    cm->y_dequant[q][0] = vp9_dc_quant(q, cm->y_dc_delta_q);
    cm->y_dequant[q][1] = vp9_ac_quant(q, 0);

    cm->uv_dequant[q][0] = vp9_dc_quant(q, cm->uv_dc_delta_q);
    cm->uv_dequant[q][1] = vp9_ac_quant(q, cm->uv_ac_delta_q);
  }
}

#ifdef NDEBUG
#define debug_check_frame_counts(cm) (void)0
#else  // !NDEBUG
// Counts should only be incremented when frame_parallel_decoding_mode and
// error_resilient_mode are disabled.
static void debug_check_frame_counts(const VP9_COMMON *const cm) {
  FRAME_COUNTS zero_counts;
  vp9_zero(zero_counts);
  assert(cm->frame_parallel_decoding_mode || cm->error_resilient_mode);
  assert(!memcmp(cm->counts.y_mode, zero_counts.y_mode,
                 sizeof(cm->counts.y_mode)));
  assert(!memcmp(cm->counts.uv_mode, zero_counts.uv_mode,
                 sizeof(cm->counts.uv_mode)));
  assert(!memcmp(cm->counts.partition, zero_counts.partition,
                 sizeof(cm->counts.partition)));
  assert(!memcmp(cm->counts.coef, zero_counts.coef,
                 sizeof(cm->counts.coef)));
  assert(!memcmp(cm->counts.eob_branch, zero_counts.eob_branch,
                 sizeof(cm->counts.eob_branch)));
  assert(!memcmp(cm->counts.switchable_interp, zero_counts.switchable_interp,
                 sizeof(cm->counts.switchable_interp)));
  assert(!memcmp(cm->counts.inter_mode, zero_counts.inter_mode,
                 sizeof(cm->counts.inter_mode)));
  assert(!memcmp(cm->counts.intra_inter, zero_counts.intra_inter,
                 sizeof(cm->counts.intra_inter)));
  assert(!memcmp(cm->counts.comp_inter, zero_counts.comp_inter,
                 sizeof(cm->counts.comp_inter)));
  assert(!memcmp(cm->counts.single_ref, zero_counts.single_ref,
                 sizeof(cm->counts.single_ref)));
  assert(!memcmp(cm->counts.comp_ref, zero_counts.comp_ref,
                 sizeof(cm->counts.comp_ref)));
  assert(!memcmp(&cm->counts.tx, &zero_counts.tx, sizeof(cm->counts.tx)));
  assert(!memcmp(cm->counts.skip, zero_counts.skip, sizeof(cm->counts.skip)));
  assert(!memcmp(&cm->counts.mv, &zero_counts.mv, sizeof(cm->counts.mv)));
}
#endif  // NDEBUG

static struct vp9_read_bit_buffer* init_read_bit_buffer(
    VP9Decoder *pbi,
    struct vp9_read_bit_buffer *rb,
    const uint8_t *data,
    const uint8_t *data_end,
    uint8_t *clear_data /* buffer size MAX_VP9_HEADER_SIZE */) {
  rb->bit_offset = 0;
  rb->error_handler = error_handler;
  rb->error_handler_data = &pbi->common;
  if (pbi->decrypt_cb) {
    const int n = (int)MIN(MAX_VP9_HEADER_SIZE, data_end - data);
    pbi->decrypt_cb(pbi->decrypt_state, data, clear_data, n);
    rb->bit_buffer = clear_data;
    rb->bit_buffer_end = clear_data + n;
  } else {
    rb->bit_buffer = data;
    rb->bit_buffer_end = data_end;
  }
  return rb;
}

void vp9_decode_frame(VP9Decoder *pbi,
                      const uint8_t *data, const uint8_t *data_end,
                      const uint8_t **p_data_end) {
  VP9_COMMON *const cm = &pbi->common;
  MACROBLOCKD *const xd = &pbi->mb;
  struct vp9_read_bit_buffer rb = { NULL, NULL, 0, NULL, 0};

  uint8_t clear_data[MAX_VP9_HEADER_SIZE];
  const size_t first_partition_size = read_uncompressed_header(pbi,
      init_read_bit_buffer(pbi, &rb, data, data_end, clear_data));
  const int tile_rows = 1 << cm->log2_tile_rows;
  const int tile_cols = 1 << cm->log2_tile_cols;
  YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
  xd->cur_buf = new_fb;

  if (!first_partition_size) {
    // showing a frame directly
    *p_data_end = data + (cm->profile <= PROFILE_2 ? 1 : 2);
    return;
  }

  data += vp9_rb_bytes_read(&rb);
  if (!read_is_valid(data, first_partition_size, data_end))
    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                       "Truncated packet or corrupt header length");

  init_macroblockd(cm, &pbi->mb);

  if (!cm->error_resilient_mode)
    set_prev_mi(cm);
  else
    cm->prev_mi = NULL;

  setup_plane_dequants(cm, xd, cm->base_qindex);
  vp9_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);

  cm->fc = cm->frame_contexts[cm->frame_context_idx];
  vp9_zero(cm->counts);
  vp9_zero(xd->dqcoeff);

  xd->corrupted = 0;
  new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);

  // TODO(jzern): remove frame_parallel_decoding_mode restriction for
  // single-frame tile decoding.
  if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1 &&
      cm->frame_parallel_decoding_mode) {
    *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
    if (!xd->corrupted) {
      // If multiple threads are used to decode tiles, then we use those threads
      // to do parallel loopfiltering.
      vp9_loop_filter_frame_mt(new_fb, pbi, cm, cm->lf.filter_level, 0);
    }
  } else {
    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
  }

  new_fb->corrupted |= xd->corrupted;

  if (!new_fb->corrupted) {
    if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
      vp9_adapt_coef_probs(cm);

      if (!frame_is_intra_only(cm)) {
        vp9_adapt_mode_probs(cm);
        vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
      }
    } else {
      debug_check_frame_counts(cm);
    }
  } else {
    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                       "Decode failed. Frame data is corrupted.");
  }

  if (cm->refresh_frame_context)
    cm->frame_contexts[cm->frame_context_idx] = cm->fc;
}