libvpx: Pull from upstream

source/libvpx/vp9/decoder/vp9_decodeframe.c

Index: source/libvpx/vp9/decoder/vp9_decodeframe.c
diff --git a/source/libvpx/vp9/decoder/vp9_decodeframe.c b/source/libvpx/vp9/decoder/vp9_decodeframe.c
index 30ca2d08a7ee4286332527c8d9b75d7e02f66b43..9311d8dad7d13e6738eaa46f2cb6bace2f4d45a7 100644
--- a/source/libvpx/vp9/decoder/vp9_decodeframe.c
+++ b/source/libvpx/vp9/decoder/vp9_decodeframe.c
@@ -346,1732 +346,1734 @@ static void reconstruct_inter_block(int plane, int block,
   *args->eobtotal += eob;
 }
 
-static MB_MODE_INFO *set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd,
-                                 const TileInfo *const tile,
-                                 BLOCK_SIZE bsize, int mi_row, int mi_col) {
-  const int bw = num_8x8_blocks_wide_lookup[bsize];
-  const int bh = num_8x8_blocks_high_lookup[bsize];
-  const int x_mis = MIN(bw, cm->mi_cols - mi_col);
-  const int y_mis = MIN(bh, cm->mi_rows - mi_row);
-  const int offset = mi_row * cm->mi_stride + mi_col;
-  int x, y;
-
-  xd->mi = cm->mi_grid_visible + offset;
-  xd->mi[0] = &cm->mi[offset];
-  xd->mi[0]->mbmi.sb_type = bsize;
-  for (y = 0; y < y_mis; ++y)
-    for (x = !y; x < x_mis; ++x) {
-      xd->mi[y * cm->mi_stride + x] = xd->mi[0];
-    }
+static void build_mc_border(const uint8_t *src, int src_stride,
+                            uint8_t *dst, int dst_stride,
+                            int x, int y, int b_w, int b_h, int w, int h) {
+  // Get a pointer to the start of the real data for this row.
+  const uint8_t *ref_row = src - x - y * src_stride;
 
-  set_skip_context(xd, mi_row, mi_col);
+  if (y >= h)
+    ref_row += (h - 1) * src_stride;
+  else if (y > 0)
+    ref_row += y * src_stride;
 
-  // Distance of Mb to the various image edges. These are specified to 8th pel
-  // as they are always compared to values that are in 1/8th pel units
-  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
+  do {
+    int right = 0, copy;
+    int left = x < 0 ? -x : 0;
 
-  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
-  return &xd->mi[0]->mbmi;
-}
+    if (left > b_w)
+      left = b_w;
 
-static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd,
-                         const TileInfo *const tile,
-                         int mi_row, int mi_col,
-                         vp9_reader *r, BLOCK_SIZE bsize) {
-  VP9_COMMON *const cm = &pbi->common;
-  const int less8x8 = bsize < BLOCK_8X8;
-  MB_MODE_INFO *mbmi = set_offsets(cm, xd, tile, bsize, mi_row, mi_col);
+    if (x + b_w > w)
+      right = x + b_w - w;
 
-  if (bsize >= BLOCK_8X8 && (cm->subsampling_x || cm->subsampling_y)) {
-    const BLOCK_SIZE uv_subsize =
-        ss_size_lookup[bsize][cm->subsampling_x][cm->subsampling_y];
-    if (uv_subsize == BLOCK_INVALID)
-      vpx_internal_error(xd->error_info,
-                         VPX_CODEC_CORRUPT_FRAME, "Invalid block size.");
-  }
+    if (right > b_w)
+      right = b_w;
 
-  vp9_read_mode_info(pbi, xd, tile, mi_row, mi_col, r);
+    copy = b_w - left - right;
 
-  if (less8x8)
-    bsize = BLOCK_8X8;
+    if (left)
+      memset(dst, ref_row[0], left);
 
-  if (mbmi->skip) {
-    reset_skip_context(xd, bsize);
-  }
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy);
 
-  if (!is_inter_block(mbmi)) {
-    struct intra_args arg = {xd, r, mbmi->segment_id};
-    vp9_foreach_transformed_block(xd, bsize,
-                                  predict_and_reconstruct_intra_block, &arg);
-  } else {
-    // Prediction
-    vp9_dec_build_inter_predictors_sb(pbi, xd, mi_row, mi_col, bsize);
+    if (right)
+      memset(dst + left + copy, ref_row[w - 1], right);
 
-    // Reconstruction
-    if (!mbmi->skip) {
-      int eobtotal = 0;
-      struct inter_args arg = {xd, r, &eobtotal, mbmi->segment_id};
-      vp9_foreach_transformed_block(xd, bsize, reconstruct_inter_block, &arg);
-      if (!less8x8 && eobtotal == 0)
-        mbmi->skip = 1;  // skip loopfilter
-    }
-  }
+    dst += dst_stride;
+    ++y;
 
-  xd->corrupted |= vp9_reader_has_error(r);
+    if (y > 0 && y < h)
+      ref_row += src_stride;
+  } while (--b_h);
 }
 
-static PARTITION_TYPE read_partition(VP9_COMMON *cm, MACROBLOCKD *xd,
-                                     int hbs,
-                                     int mi_row, int mi_col, BLOCK_SIZE bsize,
-                                     vp9_reader *r) {
-  const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
-  const vp9_prob *const probs = get_partition_probs(cm, ctx);
-  const int has_rows = (mi_row + hbs) < cm->mi_rows;
-  const int has_cols = (mi_col + hbs) < cm->mi_cols;
-  FRAME_COUNTS *counts = xd->counts;
-  PARTITION_TYPE p;
-
-  if (has_rows && has_cols)
-    p = (PARTITION_TYPE)vp9_read_tree(r, vp9_partition_tree, probs);
-  else if (!has_rows && has_cols)
-    p = vp9_read(r, probs[1]) ? PARTITION_SPLIT : PARTITION_HORZ;
-  else if (has_rows && !has_cols)
-    p = vp9_read(r, probs[2]) ? PARTITION_SPLIT : PARTITION_VERT;
-  else
-    p = PARTITION_SPLIT;
+#if CONFIG_VP9_HIGHBITDEPTH
+static void high_build_mc_border(const uint8_t *src8, int src_stride,
+                                 uint16_t *dst, int dst_stride,
+                                 int x, int y, int b_w, int b_h,
+                                 int w, int h) {
+  // Get a pointer to the start of the real data for this row.
+  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  const uint16_t *ref_row = src - x - y * src_stride;
 
-  if (counts)
-    ++counts->partition[ctx][p];
+  if (y >= h)
+    ref_row += (h - 1) * src_stride;
+  else if (y > 0)
+    ref_row += y * src_stride;
 
-  return p;
-}
+  do {
+    int right = 0, copy;
+    int left = x < 0 ? -x : 0;
 
-static void decode_partition(VP9Decoder *const pbi, MACROBLOCKD *const xd,
-                             const TileInfo *const tile,
-                             int mi_row, int mi_col,
-                             vp9_reader* r, BLOCK_SIZE bsize) {
-  VP9_COMMON *const cm = &pbi->common;
-  const int hbs = num_8x8_blocks_wide_lookup[bsize] / 2;
-  PARTITION_TYPE partition;
-  BLOCK_SIZE subsize;
+    if (left > b_w)
+      left = b_w;
 
-  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
-    return;
+    if (x + b_w > w)
+      right = x + b_w - w;
 
-  partition = read_partition(cm, xd, hbs, mi_row, mi_col, bsize, r);
-  subsize = get_subsize(bsize, partition);
-  if (bsize == BLOCK_8X8) {
-    decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
-  } else {
-    switch (partition) {
-      case PARTITION_NONE:
-        decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
-        break;
-      case PARTITION_HORZ:
-        decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
-        if (mi_row + hbs < cm->mi_rows)
-          decode_block(pbi, xd, tile, mi_row + hbs, mi_col, r, subsize);
-        break;
-      case PARTITION_VERT:
-        decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
-        if (mi_col + hbs < cm->mi_cols)
-          decode_block(pbi, xd, tile, mi_row, mi_col + hbs, r, subsize);
-        break;
-      case PARTITION_SPLIT:
-        decode_partition(pbi, xd, tile, mi_row, mi_col, r, subsize);
-        decode_partition(pbi, xd, tile, mi_row, mi_col + hbs, r, subsize);
-        decode_partition(pbi, xd, tile, mi_row + hbs, mi_col, r, subsize);
-        decode_partition(pbi, xd, tile, mi_row + hbs, mi_col + hbs, r, subsize);
-        break;
-      default:
-        assert(0 && "Invalid partition type");
-    }
-  }
+    if (right > b_w)
+      right = b_w;
 
-  // update partition context
-  if (bsize >= BLOCK_8X8 &&
-      (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
-    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
-}
+    copy = b_w - left - right;
 
-static void setup_token_decoder(const uint8_t *data,
-                                const uint8_t *data_end,
-                                size_t read_size,
-                                struct vpx_internal_error_info *error_info,
-                                vp9_reader *r,
-                                vpx_decrypt_cb decrypt_cb,
-                                void *decrypt_state) {
-  // Validate the calculated partition length. If the buffer
-  // described by the partition can't be fully read, then restrict
-  // it to the portion that can be (for EC mode) or throw an error.
-  if (!read_is_valid(data, read_size, data_end))
-    vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
-                       "Truncated packet or corrupt tile length");
+    if (left)
+      vpx_memset16(dst, ref_row[0], left);
 
-  if (vp9_reader_init(r, data, read_size, decrypt_cb, decrypt_state))
-    vpx_internal_error(error_info, VPX_CODEC_MEM_ERROR,
-                       "Failed to allocate bool decoder %d", 1);
-}
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy * sizeof(uint16_t));
 
-static void read_coef_probs_common(vp9_coeff_probs_model *coef_probs,
-                                   vp9_reader *r) {
-  int i, j, k, l, m;
+    if (right)
+      vpx_memset16(dst + left + copy, ref_row[w - 1], right);
 
-  if (vp9_read_bit(r))
-    for (i = 0; i < PLANE_TYPES; ++i)
-      for (j = 0; j < REF_TYPES; ++j)
-        for (k = 0; k < COEF_BANDS; ++k)
-          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
-            for (m = 0; m < UNCONSTRAINED_NODES; ++m)
-              vp9_diff_update_prob(r, &coef_probs[i][j][k][l][m]);
-}
+    dst += dst_stride;
+    ++y;
 
-static void read_coef_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode,
-                            vp9_reader *r) {
-    const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
-    TX_SIZE tx_size;
-    for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
-      read_coef_probs_common(fc->coef_probs[tx_size], r);
+    if (y > 0 && y < h)
+      ref_row += src_stride;
+  } while (--b_h);
 }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
 
-static void setup_segmentation(struct segmentation *seg,
-                               struct vp9_read_bit_buffer *rb) {
-  int i, j;
-
-  seg->update_map = 0;
-  seg->update_data = 0;
-
-  seg->enabled = vp9_rb_read_bit(rb);
-  if (!seg->enabled)
-    return;
+#if CONFIG_VP9_HIGHBITDEPTH
+static void extend_and_predict(const uint8_t *buf_ptr1, int pre_buf_stride,
+                               int x0, int y0, int b_w, int b_h,
+                               int frame_width, int frame_height,
+                               int border_offset,
+                               uint8_t *const dst, int dst_buf_stride,
+                               int subpel_x, int subpel_y,
+                               const InterpKernel *kernel,
+                               const struct scale_factors *sf,
+                               MACROBLOCKD *xd,
+                               int w, int h, int ref, int xs, int ys) {
+  DECLARE_ALIGNED(16, uint16_t, mc_buf_high[80 * 2 * 80 * 2]);
+  const uint8_t *buf_ptr;
 
-  // Segmentation map update
-  seg->update_map = vp9_rb_read_bit(rb);
-  if (seg->update_map) {
-    for (i = 0; i < SEG_TREE_PROBS; i++)
-      seg->tree_probs[i] = vp9_rb_read_bit(rb) ? vp9_rb_read_literal(rb, 8)
-                                               : MAX_PROB;
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    high_build_mc_border(buf_ptr1, pre_buf_stride, mc_buf_high, b_w,
+                         x0, y0, b_w, b_h, frame_width, frame_height);
+    buf_ptr = CONVERT_TO_BYTEPTR(mc_buf_high) + border_offset;
+  } else {
+    build_mc_border(buf_ptr1, pre_buf_stride, (uint8_t *)mc_buf_high, b_w,
+                    x0, y0, b_w, b_h, frame_width, frame_height);
+    buf_ptr = ((uint8_t *)mc_buf_high) + border_offset;
+  }
 
-    seg->temporal_update = vp9_rb_read_bit(rb);
-    if (seg->temporal_update) {
-      for (i = 0; i < PREDICTION_PROBS; i++)
-        seg->pred_probs[i] = vp9_rb_read_bit(rb) ? vp9_rb_read_literal(rb, 8)
-                                                 : MAX_PROB;
-    } else {
-      for (i = 0; i < PREDICTION_PROBS; i++)
-        seg->pred_probs[i] = MAX_PROB;
-    }
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    high_inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
+                         subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
+  } else {
+    inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
+                    subpel_y, sf, w, h, ref, kernel, xs, ys);
   }
+}
+#else
+static void extend_and_predict(const uint8_t *buf_ptr1, int pre_buf_stride,
+                               int x0, int y0, int b_w, int b_h,
+                               int frame_width, int frame_height,
+                               int border_offset,
+                               uint8_t *const dst, int dst_buf_stride,
+                               int subpel_x, int subpel_y,
+                               const InterpKernel *kernel,
+                               const struct scale_factors *sf,
+                               int w, int h, int ref, int xs, int ys) {
+  DECLARE_ALIGNED(16, uint8_t, mc_buf[80 * 2 * 80 * 2]);
+  const uint8_t *buf_ptr;
 
-  // Segmentation data update
-  seg->update_data = vp9_rb_read_bit(rb);
-  if (seg->update_data) {
-    seg->abs_delta = vp9_rb_read_bit(rb);
+  build_mc_border(buf_ptr1, pre_buf_stride, mc_buf, b_w,
+                  x0, y0, b_w, b_h, frame_width, frame_height);
+  buf_ptr = mc_buf + border_offset;
 
-    vp9_clearall_segfeatures(seg);
+  inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
+                  subpel_y, sf, w, h, ref, kernel, xs, ys);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
 
-    for (i = 0; i < MAX_SEGMENTS; i++) {
-      for (j = 0; j < SEG_LVL_MAX; j++) {
-        int data = 0;
-        const int feature_enabled = vp9_rb_read_bit(rb);
-        if (feature_enabled) {
-          vp9_enable_segfeature(seg, i, j);
-          data = decode_unsigned_max(rb, vp9_seg_feature_data_max(j));
-          if (vp9_is_segfeature_signed(j))
-            data = vp9_rb_read_bit(rb) ? -data : data;
-        }
-        vp9_set_segdata(seg, i, j, data);
-      }
-    }
+static void dec_build_inter_predictors(VP9Decoder *const pbi, MACROBLOCKD *xd,
+                                       int plane, int bw, int bh, int x,
+                                       int y, int w, int h, int mi_x, int mi_y,
+                                       const InterpKernel *kernel,
+                                       const struct scale_factors *sf,
+                                       struct buf_2d *pre_buf,
+                                       struct buf_2d *dst_buf, const MV* mv,
+                                       RefCntBuffer *ref_frame_buf,
+                                       int is_scaled, int ref) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
+  MV32 scaled_mv;
+  int xs, ys, x0, y0, x0_16, y0_16, frame_width, frame_height,
+      buf_stride, subpel_x, subpel_y;
+  uint8_t *ref_frame, *buf_ptr;
+
+  // Get reference frame pointer, width and height.
+  if (plane == 0) {
+    frame_width = ref_frame_buf->buf.y_crop_width;
+    frame_height = ref_frame_buf->buf.y_crop_height;
+    ref_frame = ref_frame_buf->buf.y_buffer;
+  } else {
+    frame_width = ref_frame_buf->buf.uv_crop_width;
+    frame_height = ref_frame_buf->buf.uv_crop_height;
+    ref_frame = plane == 1 ? ref_frame_buf->buf.u_buffer
+                         : ref_frame_buf->buf.v_buffer;
   }
-}
 
-static void setup_loopfilter(struct loopfilter *lf,
-                             struct vp9_read_bit_buffer *rb) {
-  lf->filter_level = vp9_rb_read_literal(rb, 6);
-  lf->sharpness_level = vp9_rb_read_literal(rb, 3);
+  if (is_scaled) {
+    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, mv, bw, bh,
+                                               pd->subsampling_x,
+                                               pd->subsampling_y);
+    // Co-ordinate of containing block to pixel precision.
+    int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
+    int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
 
-  // Read in loop filter deltas applied at the MB level based on mode or ref
-  // frame.
-  lf->mode_ref_delta_update = 0;
+    // Co-ordinate of the block to 1/16th pixel precision.
+    x0_16 = (x_start + x) << SUBPEL_BITS;
+    y0_16 = (y_start + y) << SUBPEL_BITS;
 
-  lf->mode_ref_delta_enabled = vp9_rb_read_bit(rb);
-  if (lf->mode_ref_delta_enabled) {
-    lf->mode_ref_delta_update = vp9_rb_read_bit(rb);
-    if (lf->mode_ref_delta_update) {
-      int i;
+    // Co-ordinate of current block in reference frame
+    // to 1/16th pixel precision.
+    x0_16 = sf->scale_value_x(x0_16, sf);
+    y0_16 = sf->scale_value_y(y0_16, sf);
 
-      for (i = 0; i < MAX_REF_LF_DELTAS; i++)
-        if (vp9_rb_read_bit(rb))
-          lf->ref_deltas[i] = vp9_rb_read_signed_literal(rb, 6);
+    // Map the top left corner of the block into the reference frame.
+    x0 = sf->scale_value_x(x_start + x, sf);
+    y0 = sf->scale_value_y(y_start + y, sf);
 
-      for (i = 0; i < MAX_MODE_LF_DELTAS; i++)
-        if (vp9_rb_read_bit(rb))
-          lf->mode_deltas[i] = vp9_rb_read_signed_literal(rb, 6);
-    }
+    // Scale the MV and incorporate the sub-pixel offset of the block
+    // in the reference frame.
+    scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
+    xs = sf->x_step_q4;
+    ys = sf->y_step_q4;
+  } else {
+    // Co-ordinate of containing block to pixel precision.
+    x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
+    y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+
+    // Co-ordinate of the block to 1/16th pixel precision.
+    x0_16 = x0 << SUBPEL_BITS;
+    y0_16 = y0 << SUBPEL_BITS;
+
+    scaled_mv.row = mv->row * (1 << (1 - pd->subsampling_y));
+    scaled_mv.col = mv->col * (1 << (1 - pd->subsampling_x));
+    xs = ys = 16;
   }
-}
+  subpel_x = scaled_mv.col & SUBPEL_MASK;
+  subpel_y = scaled_mv.row & SUBPEL_MASK;
 
-static INLINE int read_delta_q(struct vp9_read_bit_buffer *rb) {
-  return vp9_rb_read_bit(rb) ? vp9_rb_read_signed_literal(rb, 4) : 0;
-}
+  // Calculate the top left corner of the best matching block in the
+  // reference frame.
+  x0 += scaled_mv.col >> SUBPEL_BITS;
+  y0 += scaled_mv.row >> SUBPEL_BITS;
+  x0_16 += scaled_mv.col;
+  y0_16 += scaled_mv.row;
 
-static void setup_quantization(VP9_COMMON *const cm, MACROBLOCKD *const xd,
-                               struct vp9_read_bit_buffer *rb) {
-  cm->base_qindex = vp9_rb_read_literal(rb, QINDEX_BITS);
-  cm->y_dc_delta_q = read_delta_q(rb);
-  cm->uv_dc_delta_q = read_delta_q(rb);
-  cm->uv_ac_delta_q = read_delta_q(rb);
-  cm->dequant_bit_depth = cm->bit_depth;
-  xd->lossless = cm->base_qindex == 0 &&
-                 cm->y_dc_delta_q == 0 &&
-                 cm->uv_dc_delta_q == 0 &&
-                 cm->uv_ac_delta_q == 0;
+  // Get reference block pointer.
+  buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
+  buf_stride = pre_buf->stride;
+
+  // Do border extension if there is motion or the
+  // width/height is not a multiple of 8 pixels.
+  if (is_scaled || scaled_mv.col || scaled_mv.row ||
+      (frame_width & 0x7) || (frame_height & 0x7)) {
+    int y1 = (y0_16 + (h - 1) * ys) >> SUBPEL_BITS;
+
+    // Get reference block bottom right horizontal coordinate.
+    int x1 = (x0_16 + (w - 1) * xs) >> SUBPEL_BITS;
+    int x_pad = 0, y_pad = 0;
+
+    if (subpel_x || (sf->x_step_q4 != SUBPEL_SHIFTS)) {
+      x0 -= VP9_INTERP_EXTEND - 1;
+      x1 += VP9_INTERP_EXTEND;
+      x_pad = 1;
+    }
 
+    if (subpel_y || (sf->y_step_q4 != SUBPEL_SHIFTS)) {
+      y0 -= VP9_INTERP_EXTEND - 1;
+      y1 += VP9_INTERP_EXTEND;
+      y_pad = 1;
+    }
+
+    // Wait until reference block is ready. Pad 7 more pixels as last 7
+    // pixels of each superblock row can be changed by next superblock row.
+    if (pbi->frame_parallel_decode)
+      vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
+                           MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
+
+    // Skip border extension if block is inside the frame.
+    if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width - 1 ||
+        y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
+      // Extend the border.
+      const uint8_t *const buf_ptr1 = ref_frame + y0 * buf_stride + x0;
+      const int b_w = x1 - x0 + 1;
+      const int b_h = y1 - y0 + 1;
+      const int border_offset = y_pad * 3 * b_w + x_pad * 3;
+
+      extend_and_predict(buf_ptr1, buf_stride, x0, y0, b_w, b_h,
+                         frame_width, frame_height, border_offset,
+                         dst, dst_buf->stride,
+                         subpel_x, subpel_y,
+                         kernel, sf,
 #if CONFIG_VP9_HIGHBITDEPTH
-  xd->bd = (int)cm->bit_depth;
+                         xd,
 #endif
-}
-
-static void setup_segmentation_dequant(VP9_COMMON *const cm) {
-  // Build y/uv dequant values based on segmentation.
-  if (cm->seg.enabled) {
-    int i;
-    for (i = 0; i < MAX_SEGMENTS; ++i) {
-      const int qindex = vp9_get_qindex(&cm->seg, i, cm->base_qindex);
-      cm->y_dequant[i][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q,
-                                         cm->bit_depth);
-      cm->y_dequant[i][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
-      cm->uv_dequant[i][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
-                                          cm->bit_depth);
-      cm->uv_dequant[i][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
-                                          cm->bit_depth);
+                         w, h, ref, xs, ys);
+      return;
     }
   } else {
-    const int qindex = cm->base_qindex;
-    // When segmentation is disabled, only the first value is used.  The
-    // remaining are don't cares.
-    cm->y_dequant[0][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
-    cm->y_dequant[0][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
-    cm->uv_dequant[0][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
-                                        cm->bit_depth);
-    cm->uv_dequant[0][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
-                                        cm->bit_depth);
+    // Wait until reference block is ready. Pad 7 more pixels as last 7
+    // pixels of each superblock row can be changed by next superblock row.
+     if (pbi->frame_parallel_decode) {
+       const int y1 = (y0_16 + (h - 1) * ys) >> SUBPEL_BITS;
+       vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
+                            MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
+     }
+  }
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    high_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                         subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
+  } else {
+    inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                    subpel_y, sf, w, h, ref, kernel, xs, ys);
   }
+#else
+  inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                  subpel_y, sf, w, h, ref, kernel, xs, ys);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
 }
 
-static INTERP_FILTER read_interp_filter(struct vp9_read_bit_buffer *rb) {
-  const INTERP_FILTER literal_to_filter[] = { EIGHTTAP_SMOOTH,
-                                              EIGHTTAP,
-                                              EIGHTTAP_SHARP,
-                                              BILINEAR };
-  return vp9_rb_read_bit(rb) ? SWITCHABLE
-                             : literal_to_filter[vp9_rb_read_literal(rb, 2)];
+static void dec_build_inter_predictors_sb(VP9Decoder *const pbi,
+                                          MACROBLOCKD *xd,
+                                          int mi_row, int mi_col,
+                                          BLOCK_SIZE bsize) {
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+  const MODE_INFO *mi = xd->mi[0];
+  const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
+  const BLOCK_SIZE sb_type = mi->mbmi.sb_type;
+  const int is_compound = has_second_ref(&mi->mbmi);
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
+                                                        &xd->plane[plane]);
+    struct macroblockd_plane *const pd = &xd->plane[plane];
+    struct buf_2d *const dst_buf = &pd->dst;
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+
+    const int bw = 4 * num_4x4_w;
+    const int bh = 4 * num_4x4_h;
+    int ref;
+
+    for (ref = 0; ref < 1 + is_compound; ++ref) {
+      const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
+      struct buf_2d *const pre_buf = &pd->pre[ref];
+      const int idx = xd->block_refs[ref]->idx;
+      BufferPool *const pool = pbi->common.buffer_pool;
+      RefCntBuffer *const ref_frame_buf = &pool->frame_bufs[idx];
+      const int is_scaled = vp9_is_scaled(sf);
+
+      if (sb_type < BLOCK_8X8) {
+        int i = 0, x, y;
+        assert(bsize == BLOCK_8X8);
+        for (y = 0; y < num_4x4_h; ++y) {
+          for (x = 0; x < num_4x4_w; ++x) {
+            const MV mv = average_split_mvs(pd, mi, ref, i++);
+            dec_build_inter_predictors(pbi, xd, plane, bw, bh,
+                                       4 * x, 4 * y, 4, 4, mi_x, mi_y, kernel,
+                                       sf, pre_buf, dst_buf, &mv,
+                                       ref_frame_buf, is_scaled, ref);
+          }
+        }
+      } else {
+        const MV mv = mi->mbmi.mv[ref].as_mv;
+        dec_build_inter_predictors(pbi, xd, plane, bw, bh,
+                                   0, 0, bw, bh, mi_x, mi_y, kernel,
+                                   sf, pre_buf, dst_buf, &mv, ref_frame_buf,
+                                   is_scaled, ref);
+      }
+    }
+  }
 }
 
-void vp9_read_frame_size(struct vp9_read_bit_buffer *rb,
-                         int *width, int *height) {
-  *width = vp9_rb_read_literal(rb, 16) + 1;
-  *height = vp9_rb_read_literal(rb, 16) + 1;
-}
+static MB_MODE_INFO *set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                                 const TileInfo *const tile,
+                                 BLOCK_SIZE bsize, int mi_row, int mi_col) {
+  const int bw = num_8x8_blocks_wide_lookup[bsize];
+  const int bh = num_8x8_blocks_high_lookup[bsize];
+  const int x_mis = MIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = MIN(bh, cm->mi_rows - mi_row);
+  const int offset = mi_row * cm->mi_stride + mi_col;
+  int x, y;
+
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = &cm->mi[offset];
+  xd->mi[0]->mbmi.sb_type = bsize;
+  for (y = 0; y < y_mis; ++y)
+    for (x = !y; x < x_mis; ++x) {
+      xd->mi[y * cm->mi_stride + x] = xd->mi[0];
+    }
 
-static void setup_display_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
-  cm->display_width = cm->width;
-  cm->display_height = cm->height;
-  if (vp9_rb_read_bit(rb))
-    vp9_read_frame_size(rb, &cm->display_width, &cm->display_height);
-}
+  set_skip_context(xd, mi_row, mi_col);
 
-static void resize_mv_buffer(VP9_COMMON *cm) {
-  vpx_free(cm->cur_frame->mvs);
-  cm->cur_frame->mi_rows = cm->mi_rows;
-  cm->cur_frame->mi_cols = cm->mi_cols;
-  cm->cur_frame->mvs = (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
-                                            sizeof(*cm->cur_frame->mvs));
+  // Distance of Mb to the various image edges. These are specified to 8th pel
+  // as they are always compared to values that are in 1/8th pel units
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
+
+  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
+  return &xd->mi[0]->mbmi;
 }
 
-static void resize_context_buffers(VP9_COMMON *cm, int width, int height) {
-#if CONFIG_SIZE_LIMIT
-  if (width > DECODE_WIDTH_LIMIT || height > DECODE_HEIGHT_LIMIT)
-    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Dimensions of %dx%d beyond allowed size of %dx%d.",
-                       width, height, DECODE_WIDTH_LIMIT, DECODE_HEIGHT_LIMIT);
-#endif
-  if (cm->width != width || cm->height != height) {
-    const int new_mi_rows =
-        ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2) >> MI_SIZE_LOG2;
-    const int new_mi_cols =
-        ALIGN_POWER_OF_TWO(width,  MI_SIZE_LOG2) >> MI_SIZE_LOG2;
+static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd,
+                         const TileInfo *const tile,
+                         int mi_row, int mi_col,
+                         vp9_reader *r, BLOCK_SIZE bsize) {
+  VP9_COMMON *const cm = &pbi->common;
+  const int less8x8 = bsize < BLOCK_8X8;
+  MB_MODE_INFO *mbmi = set_offsets(cm, xd, tile, bsize, mi_row, mi_col);
 
-    // Allocations in vp9_alloc_context_buffers() depend on individual
-    // dimensions as well as the overall size.
-    if (new_mi_cols > cm->mi_cols || new_mi_rows > cm->mi_rows) {
-      if (vp9_alloc_context_buffers(cm, width, height))
-        vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
-                           "Failed to allocate context buffers");
-    } else {
-      vp9_set_mb_mi(cm, width, height);
-    }
-    vp9_init_context_buffers(cm);
-    cm->width = width;
-    cm->height = height;
+  if (bsize >= BLOCK_8X8 && (cm->subsampling_x || cm->subsampling_y)) {
+    const BLOCK_SIZE uv_subsize =
+        ss_size_lookup[bsize][cm->subsampling_x][cm->subsampling_y];
+    if (uv_subsize == BLOCK_INVALID)
+      vpx_internal_error(xd->error_info,
+                         VPX_CODEC_CORRUPT_FRAME, "Invalid block size.");
   }
-  if (cm->cur_frame->mvs == NULL || cm->mi_rows > cm->cur_frame->mi_rows ||
-      cm->mi_cols > cm->cur_frame->mi_cols) {
-    resize_mv_buffer(cm);
+
+  vp9_read_mode_info(pbi, xd, tile, mi_row, mi_col, r);
+
+  if (less8x8)
+    bsize = BLOCK_8X8;
+
+  if (mbmi->skip) {
+    reset_skip_context(xd, bsize);
   }
-}
 
-static void setup_frame_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
-  int width, height;
-  BufferPool *const pool = cm->buffer_pool;
-  vp9_read_frame_size(rb, &width, &height);
-  resize_context_buffers(cm, width, height);
-  setup_display_size(cm, rb);
+  if (!is_inter_block(mbmi)) {
+    struct intra_args arg = {xd, r, mbmi->segment_id};
+    vp9_foreach_transformed_block(xd, bsize,
+                                  predict_and_reconstruct_intra_block, &arg);
+  } else {
+    // Prediction
+    dec_build_inter_predictors_sb(pbi, xd, mi_row, mi_col, bsize);
 
-  lock_buffer_pool(pool);
-  if (vp9_realloc_frame_buffer(
-          get_frame_new_buffer(cm), cm->width, cm->height,
-          cm->subsampling_x, cm->subsampling_y,
-#if CONFIG_VP9_HIGHBITDEPTH
-          cm->use_highbitdepth,
-#endif
-          VP9_DEC_BORDER_IN_PIXELS,
-          cm->byte_alignment,
-          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
-          pool->cb_priv)) {
-    unlock_buffer_pool(pool);
-    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
-                       "Failed to allocate frame buffer");
+    // Reconstruction
+    if (!mbmi->skip) {
+      int eobtotal = 0;
+      struct inter_args arg = {xd, r, &eobtotal, mbmi->segment_id};
+      vp9_foreach_transformed_block(xd, bsize, reconstruct_inter_block, &arg);
+      if (!less8x8 && eobtotal == 0)
+        mbmi->skip = 1;  // skip loopfilter
+    }
   }
-  unlock_buffer_pool(pool);
 
-  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
-  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
-  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
-  pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
+  xd->corrupted |= vp9_reader_has_error(r);
 }
 
-static INLINE int valid_ref_frame_img_fmt(vpx_bit_depth_t ref_bit_depth,
-                                          int ref_xss, int ref_yss,
-                                          vpx_bit_depth_t this_bit_depth,
-                                          int this_xss, int this_yss) {
-  return ref_bit_depth == this_bit_depth && ref_xss == this_xss &&
-         ref_yss == this_yss;
-}
+static PARTITION_TYPE read_partition(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                     BLOCK_SIZE bsize, vp9_reader *r,
+                                     int has_rows, int has_cols) {
+  const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
+  const vp9_prob *const probs = get_partition_probs(xd, ctx);
+  FRAME_COUNTS *counts = xd->counts;
+  PARTITION_TYPE p;
 
-static void setup_frame_size_with_refs(VP9_COMMON *cm,
-                                       struct vp9_read_bit_buffer *rb) {
-  int width, height;
-  int found = 0, i;
-  int has_valid_ref_frame = 0;
-  BufferPool *const pool = cm->buffer_pool;
-  for (i = 0; i < REFS_PER_FRAME; ++i) {
-    if (vp9_rb_read_bit(rb)) {
-      YV12_BUFFER_CONFIG *const buf = cm->frame_refs[i].buf;
-      width = buf->y_crop_width;
-      height = buf->y_crop_height;
-      found = 1;
-      break;
-    }
-  }
+  if (has_rows && has_cols)
+    p = (PARTITION_TYPE)vp9_read_tree(r, vp9_partition_tree, probs);
+  else if (!has_rows && has_cols)
+    p = vp9_read(r, probs[1]) ? PARTITION_SPLIT : PARTITION_HORZ;
+  else if (has_rows && !has_cols)
+    p = vp9_read(r, probs[2]) ? PARTITION_SPLIT : PARTITION_VERT;
+  else
+    p = PARTITION_SPLIT;
 
-  if (!found)
-    vp9_read_frame_size(rb, &width, &height);
+  if (counts)
+    ++counts->partition[ctx][p];
 
-  if (width <= 0 || height <= 0)
-    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Invalid frame size");
+  return p;
+}
 
-  // Check to make sure at least one of frames that this frame references
-  // has valid dimensions.
-  for (i = 0; i < REFS_PER_FRAME; ++i) {
-    RefBuffer *const ref_frame = &cm->frame_refs[i];
-    has_valid_ref_frame |= valid_ref_frame_size(ref_frame->buf->y_crop_width,
-                                                ref_frame->buf->y_crop_height,
-                                                width, height);
-  }
-  if (!has_valid_ref_frame)
-    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Referenced frame has invalid size");
-  for (i = 0; i < REFS_PER_FRAME; ++i) {
-    RefBuffer *const ref_frame = &cm->frame_refs[i];
-    if (!valid_ref_frame_img_fmt(
-            ref_frame->buf->bit_depth,
-            ref_frame->buf->subsampling_x,
-            ref_frame->buf->subsampling_y,
-            cm->bit_depth,
-            cm->subsampling_x,
-            cm->subsampling_y))
-      vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                         "Referenced frame has incompatible color format");
-  }
+static void decode_partition(VP9Decoder *const pbi, MACROBLOCKD *const xd,
+                             const TileInfo *const tile,
+                             int mi_row, int mi_col,
+                             vp9_reader* r, BLOCK_SIZE bsize) {
+  VP9_COMMON *const cm = &pbi->common;
+  const int hbs = num_8x8_blocks_wide_lookup[bsize] / 2;
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+  const int has_rows = (mi_row + hbs) < cm->mi_rows;
+  const int has_cols = (mi_col + hbs) < cm->mi_cols;
 
-  resize_context_buffers(cm, width, height);
-  setup_display_size(cm, rb);
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
 
-  lock_buffer_pool(pool);
-  if (vp9_realloc_frame_buffer(
-          get_frame_new_buffer(cm), cm->width, cm->height,
-          cm->subsampling_x, cm->subsampling_y,
-#if CONFIG_VP9_HIGHBITDEPTH
-          cm->use_highbitdepth,
-#endif
-          VP9_DEC_BORDER_IN_PIXELS,
-          cm->byte_alignment,
-          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
-          pool->cb_priv)) {
-    unlock_buffer_pool(pool);
-    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
-                       "Failed to allocate frame buffer");
+  partition = read_partition(xd, mi_row, mi_col, bsize, r, has_rows, has_cols);
+  subsize = get_subsize(bsize, partition);
+  if (bsize == BLOCK_8X8) {
+    decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
+        break;
+      case PARTITION_HORZ:
+        decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
+        if (has_rows)
+          decode_block(pbi, xd, tile, mi_row + hbs, mi_col, r, subsize);
+        break;
+      case PARTITION_VERT:
+        decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
+        if (has_cols)
+          decode_block(pbi, xd, tile, mi_row, mi_col + hbs, r, subsize);
+        break;
+      case PARTITION_SPLIT:
+        decode_partition(pbi, xd, tile, mi_row, mi_col, r, subsize);
+        decode_partition(pbi, xd, tile, mi_row, mi_col + hbs, r, subsize);
+        decode_partition(pbi, xd, tile, mi_row + hbs, mi_col, r, subsize);
+        decode_partition(pbi, xd, tile, mi_row + hbs, mi_col + hbs, r, subsize);
+        break;
+      default:
+        assert(0 && "Invalid partition type");
+    }
   }
-  unlock_buffer_pool(pool);
 
-  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
-  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
-  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
-  pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
+  // update partition context
+  if (bsize >= BLOCK_8X8 &&
+      (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+}
+
+static void setup_token_decoder(const uint8_t *data,
+                                const uint8_t *data_end,
+                                size_t read_size,
+                                struct vpx_internal_error_info *error_info,
+                                vp9_reader *r,
+                                vpx_decrypt_cb decrypt_cb,
+                                void *decrypt_state) {
+  // Validate the calculated partition length. If the buffer
+  // described by the partition can't be fully read, then restrict
+  // it to the portion that can be (for EC mode) or throw an error.
+  if (!read_is_valid(data, read_size, data_end))
+    vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt tile length");
+
+  if (vp9_reader_init(r, data, read_size, decrypt_cb, decrypt_state))
+    vpx_internal_error(error_info, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate bool decoder %d", 1);
 }
 
-static void setup_tile_info(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
-  int min_log2_tile_cols, max_log2_tile_cols, max_ones;
-  vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
-
-  // columns
-  max_ones = max_log2_tile_cols - min_log2_tile_cols;
-  cm->log2_tile_cols = min_log2_tile_cols;
-  while (max_ones-- && vp9_rb_read_bit(rb))
-    cm->log2_tile_cols++;
+static void read_coef_probs_common(vp9_coeff_probs_model *coef_probs,
+                                   vp9_reader *r) {
+  int i, j, k, l, m;
 
-  if (cm->log2_tile_cols > 6)
-    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Invalid number of tile columns");
+  if (vp9_read_bit(r))
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
+            for (m = 0; m < UNCONSTRAINED_NODES; ++m)
+              vp9_diff_update_prob(r, &coef_probs[i][j][k][l][m]);
+}
 
-  // rows
-  cm->log2_tile_rows = vp9_rb_read_bit(rb);
-  if (cm->log2_tile_rows)
-    cm->log2_tile_rows += vp9_rb_read_bit(rb);
+static void read_coef_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode,
+                            vp9_reader *r) {
+    const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+    TX_SIZE tx_size;
+    for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
+      read_coef_probs_common(fc->coef_probs[tx_size], r);
 }
 
-typedef struct TileBuffer {
-  const uint8_t *data;
-  size_t size;
-  int col;  // only used with multi-threaded decoding
-} TileBuffer;
+static void setup_segmentation(struct segmentation *seg,
+                               struct vp9_read_bit_buffer *rb) {
+  int i, j;
 
-// Reads the next tile returning its size and adjusting '*data' accordingly
-// based on 'is_last'.
-static void get_tile_buffer(const uint8_t *const data_end,
-                            int is_last,
-                            struct vpx_internal_error_info *error_info,
-                            const uint8_t **data,
-                            vpx_decrypt_cb decrypt_cb, void *decrypt_state,
-                            TileBuffer *buf) {
-  size_t size;
+  seg->update_map = 0;
+  seg->update_data = 0;
 
-  if (!is_last) {
-    if (!read_is_valid(*data, 4, data_end))
-      vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
-                         "Truncated packet or corrupt tile length");
+  seg->enabled = vp9_rb_read_bit(rb);
+  if (!seg->enabled)
+    return;
 
-    if (decrypt_cb) {
-      uint8_t be_data[4];
-      decrypt_cb(decrypt_state, *data, be_data, 4);
-      size = mem_get_be32(be_data);
+  // Segmentation map update
+  seg->update_map = vp9_rb_read_bit(rb);
+  if (seg->update_map) {
+    for (i = 0; i < SEG_TREE_PROBS; i++)
+      seg->tree_probs[i] = vp9_rb_read_bit(rb) ? vp9_rb_read_literal(rb, 8)
+                                               : MAX_PROB;
+
+    seg->temporal_update = vp9_rb_read_bit(rb);
+    if (seg->temporal_update) {
+      for (i = 0; i < PREDICTION_PROBS; i++)
+        seg->pred_probs[i] = vp9_rb_read_bit(rb) ? vp9_rb_read_literal(rb, 8)
+                                                 : MAX_PROB;
     } else {
-      size = mem_get_be32(*data);
+      for (i = 0; i < PREDICTION_PROBS; i++)
+        seg->pred_probs[i] = MAX_PROB;
     }
-    *data += 4;
-
-    if (size > (size_t)(data_end - *data))
-      vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
-                         "Truncated packet or corrupt tile size");
-  } else {
-    size = data_end - *data;
   }
 
-  buf->data = *data;
-  buf->size = size;
-
-  *data += size;
-}
+  // Segmentation data update
+  seg->update_data = vp9_rb_read_bit(rb);
+  if (seg->update_data) {
+    seg->abs_delta = vp9_rb_read_bit(rb);
 
-static void get_tile_buffers(VP9Decoder *pbi,
-                             const uint8_t *data, const uint8_t *data_end,
-                             int tile_cols, int tile_rows,
-                             TileBuffer (*tile_buffers)[1 << 6]) {
-  int r, c;
+    vp9_clearall_segfeatures(seg);
 
-  for (r = 0; r < tile_rows; ++r) {
-    for (c = 0; c < tile_cols; ++c) {
-      const int is_last = (r == tile_rows - 1) && (c == tile_cols - 1);
-      TileBuffer *const buf = &tile_buffers[r][c];
-      buf->col = c;
-      get_tile_buffer(data_end, is_last, &pbi->common.error, &data,
-                      pbi->decrypt_cb, pbi->decrypt_state, buf);
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+      for (j = 0; j < SEG_LVL_MAX; j++) {
+        int data = 0;
+        const int feature_enabled = vp9_rb_read_bit(rb);
+        if (feature_enabled) {
+          vp9_enable_segfeature(seg, i, j);
+          data = decode_unsigned_max(rb, vp9_seg_feature_data_max(j));
+          if (vp9_is_segfeature_signed(j))
+            data = vp9_rb_read_bit(rb) ? -data : data;
+        }
+        vp9_set_segdata(seg, i, j, data);
+      }
     }
   }
 }
 
-static const uint8_t *decode_tiles(VP9Decoder *pbi,
-                                   const uint8_t *data,
-                                   const uint8_t *data_end) {
-  VP9_COMMON *const cm = &pbi->common;
-  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
-  const int aligned_cols = mi_cols_aligned_to_sb(cm->mi_cols);
-  const int tile_cols = 1 << cm->log2_tile_cols;
-  const int tile_rows = 1 << cm->log2_tile_rows;
-  TileBuffer tile_buffers[4][1 << 6];
-  int tile_row, tile_col;
-  int mi_row, mi_col;
-  TileData *tile_data = NULL;
-
-  if (cm->lf.filter_level && !cm->skip_loop_filter &&
-      pbi->lf_worker.data1 == NULL) {
-    CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
-                    vpx_memalign(32, sizeof(LFWorkerData)));
-    pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
-    if (pbi->max_threads > 1 && !winterface->reset(&pbi->lf_worker)) {
-      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
-                         "Loop filter thread creation failed");
-    }
-  }
-
-  if (cm->lf.filter_level && !cm->skip_loop_filter) {
-    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
-    // Be sure to sync as we might be resuming after a failed frame decode.
-    winterface->sync(&pbi->lf_worker);
-    vp9_loop_filter_data_reset(lf_data, get_frame_new_buffer(cm), cm,
-                               pbi->mb.plane);
-  }
-
-  assert(tile_rows <= 4);
-  assert(tile_cols <= (1 << 6));
-
-  // Note: this memset assumes above_context[0], [1] and [2]
-  // are allocated as part of the same buffer.
-  memset(cm->above_context, 0,
-         sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_cols);
+static void setup_loopfilter(struct loopfilter *lf,
+                             struct vp9_read_bit_buffer *rb) {
+  lf->filter_level = vp9_rb_read_literal(rb, 6);
+  lf->sharpness_level = vp9_rb_read_literal(rb, 3);
 
-  memset(cm->above_seg_context, 0,
-         sizeof(*cm->above_seg_context) * aligned_cols);
+  // Read in loop filter deltas applied at the MB level based on mode or ref
+  // frame.
+  lf->mode_ref_delta_update = 0;
 
-  get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
+  lf->mode_ref_delta_enabled = vp9_rb_read_bit(rb);
+  if (lf->mode_ref_delta_enabled) {
+    lf->mode_ref_delta_update = vp9_rb_read_bit(rb);
+    if (lf->mode_ref_delta_update) {
+      int i;
 
-  if (pbi->tile_data == NULL ||
-      (tile_cols * tile_rows) != pbi->total_tiles) {
-    vpx_free(pbi->tile_data);
-    CHECK_MEM_ERROR(
-        cm,
-        pbi->tile_data,
-        vpx_memalign(32, tile_cols * tile_rows * (sizeof(*pbi->tile_data))));
-    pbi->total_tiles = tile_rows * tile_cols;
-  }
+      for (i = 0; i < MAX_REF_LF_DELTAS; i++)
+        if (vp9_rb_read_bit(rb))
+          lf->ref_deltas[i] = vp9_rb_read_signed_literal(rb, 6);
 
-  // Load all tile information into tile_data.
-  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
-    for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
-      TileInfo tile;
-      const TileBuffer *const buf = &tile_buffers[tile_row][tile_col];
-      tile_data = pbi->tile_data + tile_cols * tile_row + tile_col;
-      tile_data->cm = cm;
-      tile_data->xd = pbi->mb;
-      tile_data->xd.corrupted = 0;
-      tile_data->xd.counts = cm->frame_parallel_decoding_mode ?
-                             NULL : &cm->counts;
-      vp9_tile_init(&tile, tile_data->cm, tile_row, tile_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);
+      for (i = 0; i < MAX_MODE_LF_DELTAS; i++)
+        if (vp9_rb_read_bit(rb))
+          lf->mode_deltas[i] = vp9_rb_read_signed_literal(rb, 6);
     }
   }
+}
 
-  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
-    TileInfo tile;
-    vp9_tile_set_row(&tile, cm, tile_row);
-    for (mi_row = tile.mi_row_start; mi_row < tile.mi_row_end;
-         mi_row += MI_BLOCK_SIZE) {
-      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
-        const int col = pbi->inv_tile_order ?
-                        tile_cols - tile_col - 1 : tile_col;
-        tile_data = pbi->tile_data + tile_cols * tile_row + col;
-        vp9_tile_set_col(&tile, tile_data->cm, col);
-        vp9_zero(tile_data->xd.left_context);
-        vp9_zero(tile_data->xd.left_seg_context);
-        for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
-             mi_col += MI_BLOCK_SIZE) {
-          decode_partition(pbi, &tile_data->xd, &tile, mi_row,
-                           mi_col, &tile_data->bit_reader, BLOCK_64X64);
-        }
-        pbi->mb.corrupted |= tile_data->xd.corrupted;
-        if (pbi->mb.corrupted)
-            vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                               "Failed to decode tile data");
-      }
-      // Loopfilter one row.
-      if (cm->lf.filter_level && !cm->skip_loop_filter) {
-        const int lf_start = mi_row - MI_BLOCK_SIZE;
-        LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+static INLINE int read_delta_q(struct vp9_read_bit_buffer *rb) {
+  return vp9_rb_read_bit(rb) ? vp9_rb_read_signed_literal(rb, 4) : 0;
+}
 
-        // delay the loopfilter by 1 macroblock row.
-        if (lf_start < 0) continue;
+static void setup_quantization(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                               struct vp9_read_bit_buffer *rb) {
+  cm->base_qindex = vp9_rb_read_literal(rb, QINDEX_BITS);
+  cm->y_dc_delta_q = read_delta_q(rb);
+  cm->uv_dc_delta_q = read_delta_q(rb);
+  cm->uv_ac_delta_q = read_delta_q(rb);
+  cm->dequant_bit_depth = cm->bit_depth;
+  xd->lossless = cm->base_qindex == 0 &&
+                 cm->y_dc_delta_q == 0 &&
+                 cm->uv_dc_delta_q == 0 &&
+                 cm->uv_ac_delta_q == 0;
 
-        // decoding has completed: finish up the loop filter in this thread.
-        if (mi_row + MI_BLOCK_SIZE >= cm->mi_rows) continue;
+#if CONFIG_VP9_HIGHBITDEPTH
+  xd->bd = (int)cm->bit_depth;
+#endif
+}
 
-        winterface->sync(&pbi->lf_worker);
-        lf_data->start = lf_start;
-        lf_data->stop = mi_row;
-        if (pbi->max_threads > 1) {
-          winterface->launch(&pbi->lf_worker);
-        } else {
-          winterface->execute(&pbi->lf_worker);
-        }
-      }
-      // After loopfiltering, the last 7 row pixels in each superblock row may
-      // still be changed by the longest loopfilter of the next superblock
-      // row.
-      if (pbi->frame_parallel_decode)
-        vp9_frameworker_broadcast(pbi->cur_buf,
-                                  mi_row << MI_BLOCK_SIZE_LOG2);
+static void setup_segmentation_dequant(VP9_COMMON *const cm) {
+  // Build y/uv dequant values based on segmentation.
+  if (cm->seg.enabled) {
+    int i;
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      const int qindex = vp9_get_qindex(&cm->seg, i, cm->base_qindex);
+      cm->y_dequant[i][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q,
+                                         cm->bit_depth);
+      cm->y_dequant[i][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
+      cm->uv_dequant[i][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
+                                          cm->bit_depth);
+      cm->uv_dequant[i][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
+                                          cm->bit_depth);
     }
+  } else {
+    const int qindex = cm->base_qindex;
+    // When segmentation is disabled, only the first value is used.  The
+    // remaining are don't cares.
+    cm->y_dequant[0][0] = vp9_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
+    cm->y_dequant[0][1] = vp9_ac_quant(qindex, 0, cm->bit_depth);
+    cm->uv_dequant[0][0] = vp9_dc_quant(qindex, cm->uv_dc_delta_q,
+                                        cm->bit_depth);
+    cm->uv_dequant[0][1] = vp9_ac_quant(qindex, cm->uv_ac_delta_q,
+                                        cm->bit_depth);
   }
+}
 
-  // Loopfilter remaining rows in the frame.
-  if (cm->lf.filter_level && !cm->skip_loop_filter) {
-    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
-    winterface->sync(&pbi->lf_worker);
-    lf_data->start = lf_data->stop;
-    lf_data->stop = cm->mi_rows;
-    winterface->execute(&pbi->lf_worker);
-  }
+static INTERP_FILTER read_interp_filter(struct vp9_read_bit_buffer *rb) {
+  const INTERP_FILTER literal_to_filter[] = { EIGHTTAP_SMOOTH,
+                                              EIGHTTAP,
+                                              EIGHTTAP_SHARP,
+                                              BILINEAR };
+  return vp9_rb_read_bit(rb) ? SWITCHABLE
+                             : literal_to_filter[vp9_rb_read_literal(rb, 2)];
+}
 
-  // Get last tile data.
-  tile_data = pbi->tile_data + tile_cols * tile_rows - 1;
+static void setup_display_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
+  cm->display_width = cm->width;
+  cm->display_height = cm->height;
+  if (vp9_rb_read_bit(rb))
+    vp9_read_frame_size(rb, &cm->display_width, &cm->display_height);
+}
 
-  if (pbi->frame_parallel_decode)
-    vp9_frameworker_broadcast(pbi->cur_buf, INT_MAX);
-  return vp9_reader_find_end(&tile_data->bit_reader);
+static void resize_mv_buffer(VP9_COMMON *cm) {
+  vpx_free(cm->cur_frame->mvs);
+  cm->cur_frame->mi_rows = cm->mi_rows;
+  cm->cur_frame->mi_cols = cm->mi_cols;
+  cm->cur_frame->mvs = (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
+                                            sizeof(*cm->cur_frame->mvs));
 }
 
-static int tile_worker_hook(TileWorkerData *const tile_data,
-                            const TileInfo *const tile) {
-  int mi_row, mi_col;
+static void resize_context_buffers(VP9_COMMON *cm, int width, int height) {
+#if CONFIG_SIZE_LIMIT
+  if (width > DECODE_WIDTH_LIMIT || height > DECODE_HEIGHT_LIMIT)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Dimensions of %dx%d beyond allowed size of %dx%d.",
+                       width, height, DECODE_WIDTH_LIMIT, DECODE_HEIGHT_LIMIT);
+#endif
+  if (cm->width != width || cm->height != height) {
+    const int new_mi_rows =
+        ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2) >> MI_SIZE_LOG2;
+    const int new_mi_cols =
+        ALIGN_POWER_OF_TWO(width,  MI_SIZE_LOG2) >> MI_SIZE_LOG2;
 
-  if (setjmp(tile_data->error_info.jmp)) {
-    tile_data->error_info.setjmp = 0;
-    tile_data->xd.corrupted = 1;
-    return 0;
+    // Allocations in vp9_alloc_context_buffers() depend on individual
+    // dimensions as well as the overall size.
+    if (new_mi_cols > cm->mi_cols || new_mi_rows > cm->mi_rows) {
+      if (vp9_alloc_context_buffers(cm, width, height))
+        vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate context buffers");
+    } else {
+      vp9_set_mb_mi(cm, width, height);
+    }
+    vp9_init_context_buffers(cm);
+    cm->width = width;
+    cm->height = height;
+  }
+  if (cm->cur_frame->mvs == NULL || cm->mi_rows > cm->cur_frame->mi_rows ||
+      cm->mi_cols > cm->cur_frame->mi_cols) {
+    resize_mv_buffer(cm);
   }
+}
 
-  tile_data->error_info.setjmp = 1;
-  tile_data->xd.error_info = &tile_data->error_info;
+static void setup_frame_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
+  int width, height;
+  BufferPool *const pool = cm->buffer_pool;
+  vp9_read_frame_size(rb, &width, &height);
+  resize_context_buffers(cm, width, height);
+  setup_display_size(cm, rb);
 
-  for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
-       mi_row += MI_BLOCK_SIZE) {
-    vp9_zero(tile_data->xd.left_context);
-    vp9_zero(tile_data->xd.left_seg_context);
-    for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
-         mi_col += MI_BLOCK_SIZE) {
-      decode_partition(tile_data->pbi, &tile_data->xd,
-                       tile, mi_row, mi_col, &tile_data->bit_reader,
-                       BLOCK_64X64);
-    }
+  lock_buffer_pool(pool);
+  if (vp9_realloc_frame_buffer(
+          get_frame_new_buffer(cm), cm->width, cm->height,
+          cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+          cm->use_highbitdepth,
+#endif
+          VP9_DEC_BORDER_IN_PIXELS,
+          cm->byte_alignment,
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
+          pool->cb_priv)) {
+    unlock_buffer_pool(pool);
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate frame buffer");
   }
-  return !tile_data->xd.corrupted;
+  unlock_buffer_pool(pool);
+
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
+  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
 }
 
-// sorts in descending order
-static int compare_tile_buffers(const void *a, const void *b) {
-  const TileBuffer *const buf1 = (const TileBuffer*)a;
-  const TileBuffer *const buf2 = (const TileBuffer*)b;
-  return (int)(buf2->size - buf1->size);
+static INLINE int valid_ref_frame_img_fmt(vpx_bit_depth_t ref_bit_depth,
+                                          int ref_xss, int ref_yss,
+                                          vpx_bit_depth_t this_bit_depth,
+                                          int this_xss, int this_yss) {
+  return ref_bit_depth == this_bit_depth && ref_xss == this_xss &&
+         ref_yss == this_yss;
 }
 
-static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
-                                      const uint8_t *data,
-                                      const uint8_t *data_end) {
-  VP9_COMMON *const cm = &pbi->common;
-  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
-  const uint8_t *bit_reader_end = NULL;
-  const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
-  const int tile_cols = 1 << cm->log2_tile_cols;
-  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;
+static void setup_frame_size_with_refs(VP9_COMMON *cm,
+                                       struct vp9_read_bit_buffer *rb) {
+  int width, height;
+  int found = 0, i;
+  int has_valid_ref_frame = 0;
+  BufferPool *const pool = cm->buffer_pool;
+  for (i = 0; i < REFS_PER_FRAME; ++i) {
+    if (vp9_rb_read_bit(rb)) {
+      YV12_BUFFER_CONFIG *const buf = cm->frame_refs[i].buf;
+      width = buf->y_crop_width;
+      height = buf->y_crop_height;
+      found = 1;
+      break;
+    }
+  }
 
-  assert(tile_cols <= (1 << 6));
-  assert(tile_rows == 1);
-  (void)tile_rows;
+  if (!found)
+    vp9_read_frame_size(rb, &width, &height);
 
-  // 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)));
-    // Ensure tile data offsets will be properly aligned. This may fail on
-    // platforms without DECLARE_ALIGNED().
-    assert((sizeof(*pbi->tile_worker_data) % 16) == 0);
-    CHECK_MEM_ERROR(cm, pbi->tile_worker_data,
-                    vpx_memalign(32, num_threads *
-                                 sizeof(*pbi->tile_worker_data)));
-    CHECK_MEM_ERROR(cm, pbi->tile_worker_info,
-                    vpx_malloc(num_threads * sizeof(*pbi->tile_worker_info)));
-    for (i = 0; i < num_threads; ++i) {
-      VP9Worker *const worker = &pbi->tile_workers[i];
-      ++pbi->num_tile_workers;
+  if (width <= 0 || height <= 0)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid frame size");
 
-      winterface->init(worker);
-      if (i < num_threads - 1 && !winterface->reset(worker)) {
-        vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
-                           "Tile decoder thread creation failed");
-      }
-    }
+  // Check to make sure at least one of frames that this frame references
+  // has valid dimensions.
+  for (i = 0; i < REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    has_valid_ref_frame |= valid_ref_frame_size(ref_frame->buf->y_crop_width,
+                                                ref_frame->buf->y_crop_height,
+                                                width, height);
   }
-
-  // Reset tile decoding hook
-  for (n = 0; n < num_workers; ++n) {
-    VP9Worker *const worker = &pbi->tile_workers[n];
-    winterface->sync(worker);
-    worker->hook = (VP9WorkerHook)tile_worker_hook;
-    worker->data1 = &pbi->tile_worker_data[n];
-    worker->data2 = &pbi->tile_worker_info[n];
+  if (!has_valid_ref_frame)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Referenced frame has invalid size");
+  for (i = 0; i < REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    if (!valid_ref_frame_img_fmt(
+            ref_frame->buf->bit_depth,
+            ref_frame->buf->subsampling_x,
+            ref_frame->buf->subsampling_y,
+            cm->bit_depth,
+            cm->subsampling_x,
+            cm->subsampling_y))
+      vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                         "Referenced frame has incompatible color format");
   }
 
-  // Note: this memset assumes above_context[0], [1] and [2]
-  // are allocated as part of the same buffer.
-  memset(cm->above_context, 0,
-         sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_mi_cols);
-  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);
+  resize_context_buffers(cm, width, height);
+  setup_display_size(cm, rb);
 
-  // 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;
-    }
+  lock_buffer_pool(pool);
+  if (vp9_realloc_frame_buffer(
+          get_frame_new_buffer(cm), cm->width, cm->height,
+          cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+          cm->use_highbitdepth,
+#endif
+          VP9_DEC_BORDER_IN_PIXELS,
+          cm->byte_alignment,
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
+          pool->cb_priv)) {
+    unlock_buffer_pool(pool);
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate frame buffer");
   }
+  unlock_buffer_pool(pool);
 
-  // Initialize thread frame counts.
-  if (!cm->frame_parallel_decoding_mode) {
-    int i;
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
+  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
+}
 
-    for (i = 0; i < num_workers; ++i) {
-      TileWorkerData *const tile_data =
-          (TileWorkerData*)pbi->tile_workers[i].data1;
-      vp9_zero(tile_data->counts);
-    }
-  }
+static void setup_tile_info(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
+  int min_log2_tile_cols, max_log2_tile_cols, max_ones;
+  vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
 
-  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];
+  // columns
+  max_ones = max_log2_tile_cols - min_log2_tile_cols;
+  cm->log2_tile_cols = min_log2_tile_cols;
+  while (max_ones-- && vp9_rb_read_bit(rb))
+    cm->log2_tile_cols++;
 
-      tile_data->pbi = pbi;
-      tile_data->xd = pbi->mb;
-      tile_data->xd.corrupted = 0;
-      tile_data->xd.counts = cm->frame_parallel_decoding_mode ?
-                             0 : &tile_data->counts;
-      vp9_tile_init(tile, 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);
+  if (cm->log2_tile_cols > 6)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid number of tile columns");
 
-      worker->had_error = 0;
-      if (i == num_workers - 1 || n == tile_cols - 1) {
-        winterface->execute(worker);
-      } else {
-        winterface->launch(worker);
-      }
+  // rows
+  cm->log2_tile_rows = vp9_rb_read_bit(rb);
+  if (cm->log2_tile_rows)
+    cm->log2_tile_rows += vp9_rb_read_bit(rb);
+}
 
-      if (buf->col == tile_cols - 1) {
-        final_worker = i;
-      }
+typedef struct TileBuffer {
+  const uint8_t *data;
+  size_t size;
+  int col;  // only used with multi-threaded decoding
+} TileBuffer;
 
-      ++n;
-    }
+// Reads the next tile returning its size and adjusting '*data' accordingly
+// based on 'is_last'.
+static void get_tile_buffer(const uint8_t *const data_end,
+                            int is_last,
+                            struct vpx_internal_error_info *error_info,
+                            const uint8_t **data,
+                            vpx_decrypt_cb decrypt_cb, void *decrypt_state,
+                            TileBuffer *buf) {
+  size_t size;
 
-    for (; i > 0; --i) {
-      VP9Worker *const worker = &pbi->tile_workers[i - 1];
-      // TODO(jzern): The tile may have specific error data associated with
-      // its vpx_internal_error_info which could be propagated to the main info
-      // in cm. Additionally once the threads have been synced and an error is
-      // detected, there's no point in continuing to decode tiles.
-      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;
-    }
+  if (!is_last) {
+    if (!read_is_valid(*data, 4, data_end))
+      vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
+                         "Truncated packet or corrupt tile length");
 
-    // Accumulate thread frame counts.
-    if (n >= tile_cols && !cm->frame_parallel_decoding_mode) {
-      for (i = 0; i < num_workers; ++i) {
-        TileWorkerData *const tile_data =
-            (TileWorkerData*)pbi->tile_workers[i].data1;
-        vp9_accumulate_frame_counts(cm, &tile_data->counts, 1);
-      }
+    if (decrypt_cb) {
+      uint8_t be_data[4];
+      decrypt_cb(decrypt_state, *data, be_data, 4);
+      size = mem_get_be32(be_data);
+    } else {
+      size = mem_get_be32(*data);
     }
-  }
+    *data += 4;
 
-  return bit_reader_end;
-}
+    if (size > (size_t)(data_end - *data))
+      vpx_internal_error(error_info, VPX_CODEC_CORRUPT_FRAME,
+                         "Truncated packet or corrupt tile size");
+  } else {
+    size = data_end - *data;
+  }
 
-static void error_handler(void *data) {
-  VP9_COMMON *const cm = (VP9_COMMON *)data;
-  vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME, "Truncated packet");
-}
+  buf->data = *data;
+  buf->size = size;
 
-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;
+  *data += size;
 }
 
-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 get_tile_buffers(VP9Decoder *pbi,
+                             const uint8_t *data, const uint8_t *data_end,
+                             int tile_cols, int tile_rows,
+                             TileBuffer (*tile_buffers)[1 << 6]) {
+  int r, c;
 
-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;
-#if CONFIG_VP9_HIGHBITDEPTH
-    cm->use_highbitdepth = 1;
-#endif
-  } else {
-    cm->bit_depth = VPX_BITS_8;
-#if CONFIG_VP9_HIGHBITDEPTH
-    cm->use_highbitdepth = 0;
-#endif
-  }
-  cm->color_space = vp9_rb_read_literal(rb, 3);
-  if (cm->color_space != VPX_CS_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");
+  for (r = 0; r < tile_rows; ++r) {
+    for (c = 0; c < tile_cols; ++c) {
+      const int is_last = (r == tile_rows - 1) && (c == tile_cols - 1);
+      TileBuffer *const buf = &tile_buffers[r][c];
+      buf->col = c;
+      get_tile_buffer(data_end, is_last, &pbi->common.error, &data,
+                      pbi->decrypt_cb, pbi->decrypt_state, buf);
     }
   }
 }
 
-static size_t read_uncompressed_header(VP9Decoder *pbi,
-                                       struct vp9_read_bit_buffer *rb) {
+static const uint8_t *decode_tiles(VP9Decoder *pbi,
+                                   const uint8_t *data,
+                                   const uint8_t *data_end) {
   VP9_COMMON *const cm = &pbi->common;
-  BufferPool *const pool = cm->buffer_pool;
-  RefCntBuffer *const frame_bufs = pool->frame_bufs;
-  int i, mask, ref_index = 0;
-  size_t sz;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+  const int aligned_cols = mi_cols_aligned_to_sb(cm->mi_cols);
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  TileBuffer tile_buffers[4][1 << 6];
+  int tile_row, tile_col;
+  int mi_row, mi_col;
+  TileData *tile_data = NULL;
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter &&
+      pbi->lf_worker.data1 == NULL) {
+    CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
+                    vpx_memalign(32, sizeof(LFWorkerData)));
+    pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
+    if (pbi->max_threads > 1 && !winterface->reset(&pbi->lf_worker)) {
+      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                         "Loop filter thread creation failed");
+    }
+  }
 
-  cm->last_frame_type = cm->frame_type;
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+    // Be sure to sync as we might be resuming after a failed frame decode.
+    winterface->sync(&pbi->lf_worker);
+    vp9_loop_filter_data_reset(lf_data, get_frame_new_buffer(cm), cm,
+                               pbi->mb.plane);
+  }
 
-  if (vp9_rb_read_literal(rb, 2) != VP9_FRAME_MARKER)
-      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
-                         "Invalid frame marker");
+  assert(tile_rows <= 4);
+  assert(tile_cols <= (1 << 6));
 
-  cm->profile = vp9_read_profile(rb);
+  // Note: this memset assumes above_context[0], [1] and [2]
+  // are allocated as part of the same buffer.
+  memset(cm->above_context, 0,
+         sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_cols);
 
-  if (cm->profile >= MAX_PROFILES)
-    vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
-                       "Unsupported bitstream profile");
+  memset(cm->above_seg_context, 0,
+         sizeof(*cm->above_seg_context) * aligned_cols);
 
-  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)];
-    lock_buffer_pool(pool);
-    if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1) {
-      unlock_buffer_pool(pool);
-      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
-                         "Buffer %d does not contain a decoded frame",
-                         frame_to_show);
-    }
+  get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
 
-    ref_cnt_fb(frame_bufs, &cm->new_fb_idx, frame_to_show);
-    unlock_buffer_pool(pool);
-    pbi->refresh_frame_flags = 0;
-    cm->lf.filter_level = 0;
-    cm->show_frame = 1;
+  if (pbi->tile_data == NULL ||
+      (tile_cols * tile_rows) != pbi->total_tiles) {
+    vpx_free(pbi->tile_data);
+    CHECK_MEM_ERROR(
+        cm,
+        pbi->tile_data,
+        vpx_memalign(32, tile_cols * tile_rows * (sizeof(*pbi->tile_data))));
+    pbi->total_tiles = tile_rows * tile_cols;
+  }
 
-    if (pbi->frame_parallel_decode) {
-      for (i = 0; i < REF_FRAMES; ++i)
-        cm->next_ref_frame_map[i] = cm->ref_frame_map[i];
+  // Load all tile information into tile_data.
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
+    for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+      TileInfo tile;
+      const TileBuffer *const buf = &tile_buffers[tile_row][tile_col];
+      tile_data = pbi->tile_data + tile_cols * tile_row + tile_col;
+      tile_data->cm = cm;
+      tile_data->xd = pbi->mb;
+      tile_data->xd.corrupted = 0;
+      tile_data->xd.counts = cm->frame_parallel_decoding_mode ?
+                             NULL : &cm->counts;
+      vp9_tile_init(&tile, tile_data->cm, tile_row, tile_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);
     }
-    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);
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
+    TileInfo tile;
+    vp9_tile_set_row(&tile, cm, tile_row);
+    for (mi_row = tile.mi_row_start; mi_row < tile.mi_row_end;
+         mi_row += MI_BLOCK_SIZE) {
+      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+        const int col = pbi->inv_tile_order ?
+                        tile_cols - tile_col - 1 : tile_col;
+        tile_data = pbi->tile_data + tile_cols * tile_row + col;
+        vp9_tile_set_col(&tile, tile_data->cm, col);
+        vp9_zero(tile_data->xd.left_context);
+        vp9_zero(tile_data->xd.left_seg_context);
+        for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
+             mi_col += MI_BLOCK_SIZE) {
+          decode_partition(pbi, &tile_data->xd, &tile, mi_row,
+                           mi_col, &tile_data->bit_reader, BLOCK_64X64);
+        }
+        pbi->mb.corrupted |= tile_data->xd.corrupted;
+        if (pbi->mb.corrupted)
+            vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                               "Failed to decode tile data");
+      }
+      // Loopfilter one row.
+      if (cm->lf.filter_level && !cm->skip_loop_filter) {
+        const int lf_start = mi_row - MI_BLOCK_SIZE;
+        LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
 
-  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");
+        // delay the loopfilter by 1 macroblock row.
+        if (lf_start < 0) continue;
 
-    read_bitdepth_colorspace_sampling(cm, rb);
-    pbi->refresh_frame_flags = (1 << REF_FRAMES) - 1;
+        // decoding has completed: finish up the loop filter in this thread.
+        if (mi_row + MI_BLOCK_SIZE >= cm->mi_rows) continue;
 
-    for (i = 0; i < REFS_PER_FRAME; ++i) {
-      cm->frame_refs[i].idx = INVALID_IDX;
-      cm->frame_refs[i].buf = NULL;
+        winterface->sync(&pbi->lf_worker);
+        lf_data->start = lf_start;
+        lf_data->stop = mi_row;
+        if (pbi->max_threads > 1) {
+          winterface->launch(&pbi->lf_worker);
+        } else {
+          winterface->execute(&pbi->lf_worker);
+        }
+      }
+      // After loopfiltering, the last 7 row pixels in each superblock row may
+      // still be changed by the longest loopfilter of the next superblock
+      // row.
+      if (pbi->frame_parallel_decode)
+        vp9_frameworker_broadcast(pbi->cur_buf,
+                                  mi_row << MI_BLOCK_SIZE_LOG2);
     }
+  }
 
-    setup_frame_size(cm, rb);
-    if (pbi->need_resync) {
-      memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
-      pbi->need_resync = 0;
-    }
-  } else {
-    cm->intra_only = cm->show_frame ? 0 : vp9_rb_read_bit(rb);
+  // Loopfilter remaining rows in the frame.
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+    winterface->sync(&pbi->lf_worker);
+    lf_data->start = lf_data->stop;
+    lf_data->stop = cm->mi_rows;
+    winterface->execute(&pbi->lf_worker);
+  }
 
-    cm->reset_frame_context = cm->error_resilient_mode ?
-        0 : vp9_rb_read_literal(rb, 2);
+  // Get last tile data.
+  tile_data = pbi->tile_data + tile_cols * tile_rows - 1;
 
-    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 format should be YUV 4:2:0 in this
-        // case (normative).
-        cm->color_space = VPX_CS_BT_601;
-        cm->subsampling_y = cm->subsampling_x = 1;
-        cm->bit_depth = VPX_BITS_8;
-#if CONFIG_VP9_HIGHBITDEPTH
-        cm->use_highbitdepth = 0;
-#endif
-      }
+  if (pbi->frame_parallel_decode)
+    vp9_frameworker_broadcast(pbi->cur_buf, INT_MAX);
+  return vp9_reader_find_end(&tile_data->bit_reader);
+}
 
-      pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
-      setup_frame_size(cm, rb);
-      if (pbi->need_resync) {
-        memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
-        pbi->need_resync = 0;
-      }
-    } else if (pbi->need_resync != 1) {  /* Skip if need resync */
-      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 = &frame_bufs[idx].buf;
-        cm->ref_frame_sign_bias[LAST_FRAME + i] = vp9_rb_read_bit(rb);
-      }
+static int tile_worker_hook(TileWorkerData *const tile_data,
+                            const TileInfo *const tile) {
+  int mi_row, mi_col;
 
-      setup_frame_size_with_refs(cm, rb);
+  if (setjmp(tile_data->error_info.jmp)) {
+    tile_data->error_info.setjmp = 0;
+    tile_data->xd.corrupted = 1;
+    return 0;
+  }
 
-      cm->allow_high_precision_mv = vp9_rb_read_bit(rb);
-      cm->interp_filter = read_interp_filter(rb);
+  tile_data->error_info.setjmp = 1;
+  tile_data->xd.error_info = &tile_data->error_info;
 
-      for (i = 0; i < REFS_PER_FRAME; ++i) {
-        RefBuffer *const ref_buf = &cm->frame_refs[i];
-#if CONFIG_VP9_HIGHBITDEPTH
-        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,
-                                          cm->use_highbitdepth);
-#else
-        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);
-#endif
-      }
+  for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
+       mi_row += MI_BLOCK_SIZE) {
+    vp9_zero(tile_data->xd.left_context);
+    vp9_zero(tile_data->xd.left_seg_context);
+    for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
+         mi_col += MI_BLOCK_SIZE) {
+      decode_partition(tile_data->pbi, &tile_data->xd,
+                       tile, mi_row, mi_col, &tile_data->bit_reader,
+                       BLOCK_64X64);
     }
   }
-#if CONFIG_VP9_HIGHBITDEPTH
-  get_frame_new_buffer(cm)->bit_depth = cm->bit_depth;
-#endif
-  get_frame_new_buffer(cm)->color_space = cm->color_space;
+  return !tile_data->xd.corrupted;
+}
 
-  if (pbi->need_resync) {
-    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Keyframe / intra-only frame required to reset decoder"
-                       " state");
-  }
+// sorts in descending order
+static int compare_tile_buffers(const void *a, const void *b) {
+  const TileBuffer *const buf1 = (const TileBuffer*)a;
+  const TileBuffer *const buf2 = (const TileBuffer*)b;
+  return (int)(buf2->size - buf1->size);
+}
 
-  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;
-  }
+static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
+                                      const uint8_t *data,
+                                      const uint8_t *data_end) {
+  VP9_COMMON *const cm = &pbi->common;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+  const uint8_t *bit_reader_end = NULL;
+  const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  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;
 
-  // 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);
+  assert(tile_cols <= (1 << 6));
+  assert(tile_rows == 1);
+  (void)tile_rows;
 
-  // Generate next_ref_frame_map.
-  lock_buffer_pool(pool);
-  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
-    if (mask & 1) {
-      cm->next_ref_frame_map[ref_index] = cm->new_fb_idx;
-      ++frame_bufs[cm->new_fb_idx].ref_count;
-    } else {
-      cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+  // 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;
+    CHECK_MEM_ERROR(cm, pbi->tile_workers,
+                    vpx_malloc(num_threads * sizeof(*pbi->tile_workers)));
+    // Ensure tile data offsets will be properly aligned. This may fail on
+    // platforms without DECLARE_ALIGNED().
+    assert((sizeof(*pbi->tile_worker_data) % 16) == 0);
+    CHECK_MEM_ERROR(cm, pbi->tile_worker_data,
+                    vpx_memalign(32, num_threads *
+                                 sizeof(*pbi->tile_worker_data)));
+    CHECK_MEM_ERROR(cm, pbi->tile_worker_info,
+                    vpx_malloc(num_threads * sizeof(*pbi->tile_worker_info)));
+    for (i = 0; i < num_threads; ++i) {
+      VP9Worker *const worker = &pbi->tile_workers[i];
+      ++pbi->num_tile_workers;
+
+      winterface->init(worker);
+      if (i < num_threads - 1 && !winterface->reset(worker)) {
+        vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                           "Tile decoder thread creation failed");
+      }
     }
-    // Current thread holds the reference frame.
-    if (cm->ref_frame_map[ref_index] >= 0)
-      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
-    ++ref_index;
   }
 
-  for (; ref_index < REF_FRAMES; ++ref_index) {
-    cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
-    // Current thread holds the reference frame.
-    if (cm->ref_frame_map[ref_index] >= 0)
-      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+  // Reset tile decoding hook
+  for (n = 0; n < num_workers; ++n) {
+    VP9Worker *const worker = &pbi->tile_workers[n];
+    winterface->sync(worker);
+    worker->hook = (VP9WorkerHook)tile_worker_hook;
+    worker->data1 = &pbi->tile_worker_data[n];
+    worker->data2 = &pbi->tile_worker_info[n];
   }
-  unlock_buffer_pool(pool);
-  pbi->hold_ref_buf = 1;
-
-  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_segmentation_dequant(cm);
-
-  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;
-}
+  // Note: this memset assumes above_context[0], [1] and [2]
+  // are allocated as part of the same buffer.
+  memset(cm->above_context, 0,
+         sizeof(*cm->above_context) * MAX_MB_PLANE * 2 * aligned_mi_cols);
+  memset(cm->above_seg_context, 0,
+         sizeof(*cm->above_seg_context) * aligned_mi_cols);
 
-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;
+  // Load tile data into tile_buffers
+  get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
 
-  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");
+  // Sort the buffers based on size in descending order.
+  qsort(tile_buffers[0], tile_cols, sizeof(tile_buffers[0][0]),
+        compare_tile_buffers);
 
-  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);
+  // 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;
+    }
+  }
 
-  for (k = 0; k < SKIP_CONTEXTS; ++k)
-    vp9_diff_update_prob(&r, &fc->skip_probs[k]);
+  // Initialize thread frame counts.
+  if (!cm->frame_parallel_decoding_mode) {
+    int i;
 
-  if (!frame_is_intra_only(cm)) {
-    nmv_context *const nmvc = &fc->nmvc;
-    int i, j;
+    for (i = 0; i < num_workers; ++i) {
+      TileWorkerData *const tile_data =
+          (TileWorkerData*)pbi->tile_workers[i].data1;
+      vp9_zero(tile_data->counts);
+    }
+  }
 
-    read_inter_mode_probs(fc, &r);
+  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];
 
-    if (cm->interp_filter == SWITCHABLE)
-      read_switchable_interp_probs(fc, &r);
+      tile_data->pbi = pbi;
+      tile_data->xd = pbi->mb;
+      tile_data->xd.corrupted = 0;
+      tile_data->xd.counts = cm->frame_parallel_decoding_mode ?
+                             0 : &tile_data->counts;
+      vp9_tile_init(tile, 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);
 
-    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
-      vp9_diff_update_prob(&r, &fc->intra_inter_prob[i]);
+      worker->had_error = 0;
+      if (i == num_workers - 1 || n == tile_cols - 1) {
+        winterface->execute(worker);
+      } else {
+        winterface->launch(worker);
+      }
 
-    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);
+      if (buf->col == tile_cols - 1) {
+        final_worker = i;
+      }
 
-    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]);
+      ++n;
+    }
 
-    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]);
+    for (; i > 0; --i) {
+      VP9Worker *const worker = &pbi->tile_workers[i - 1];
+      // TODO(jzern): The tile may have specific error data associated with
+      // its vpx_internal_error_info which could be propagated to the main info
+      // in cm. Additionally once the threads have been synced and an error is
+      // detected, there's no point in continuing to decode tiles.
+      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;
+    }
 
-    read_mv_probs(nmvc, cm->allow_high_precision_mv, &r);
+    // Accumulate thread frame counts.
+    if (n >= tile_cols && !cm->frame_parallel_decoding_mode) {
+      for (i = 0; i < num_workers; ++i) {
+        TileWorkerData *const tile_data =
+            (TileWorkerData*)pbi->tile_workers[i].data1;
+        vp9_accumulate_frame_counts(cm, &tile_data->counts, 1);
+      }
+    }
   }
 
-  return vp9_reader_has_error(&r);
+  return bit_reader_end;
 }
 
-#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)));
+static void error_handler(void *data) {
+  VP9_COMMON *const cm = (VP9_COMMON *)data;
+  vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME, "Truncated packet");
 }
-#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;
+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;
+#if CONFIG_VP9_HIGHBITDEPTH
+    cm->use_highbitdepth = 1;
+#endif
   } else {
-    rb->bit_buffer = data;
-    rb->bit_buffer_end = data_end;
+    cm->bit_depth = VPX_BITS_8;
+#if CONFIG_VP9_HIGHBITDEPTH
+    cm->use_highbitdepth = 0;
+#endif
+  }
+  cm->color_space = vp9_rb_read_literal(rb, 3);
+  if (cm->color_space != VPX_CS_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");
+    }
   }
-  return rb;
 }
 
-void vp9_decode_frame(VP9Decoder *pbi,
-                      const uint8_t *data, const uint8_t *data_end,
-                      const uint8_t **p_data_end) {
+static size_t read_uncompressed_header(VP9Decoder *pbi,
+                                       struct vp9_read_bit_buffer *rb) {
   VP9_COMMON *const cm = &pbi->common;
-  MACROBLOCKD *const xd = &pbi->mb;
-  struct vp9_read_bit_buffer rb = { NULL, NULL, 0, NULL, 0};
-  int context_updated = 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");
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = pool->frame_bufs;
+  int i, mask, ref_index = 0;
+  size_t sz;
 
-  cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
-                           cm->width == cm->last_width &&
-                           cm->height == cm->last_height &&
-                           !cm->intra_only &&
-                           cm->last_show_frame;
+  cm->last_frame_type = cm->frame_type;
+  cm->last_intra_only = cm->intra_only;
 
-  vp9_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);
+  if (vp9_rb_read_literal(rb, 2) != VP9_FRAME_MARKER)
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Invalid frame marker");
 
-  *cm->fc = cm->frame_contexts[cm->frame_context_idx];
-  if (!cm->fc->initialized)
-    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Uninitialized entropy context.");
+  cm->profile = vp9_read_profile(rb);
 
-  vp9_zero(cm->counts);
+  if (cm->profile >= MAX_PROFILES)
+    vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                       "Unsupported bitstream profile");
 
-  xd->corrupted = 0;
-  new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
-  if (new_fb->corrupted)
-    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Decode failed. Frame data header is corrupted.");
+  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)];
+    lock_buffer_pool(pool);
+    if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1) {
+      unlock_buffer_pool(pool);
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Buffer %d does not contain a decoded frame",
+                         frame_to_show);
+    }
 
-  if (cm->lf.filter_level && !cm->skip_loop_filter) {
-    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
-  }
+    ref_cnt_fb(frame_bufs, &cm->new_fb_idx, frame_to_show);
+    unlock_buffer_pool(pool);
+    pbi->refresh_frame_flags = 0;
+    cm->lf.filter_level = 0;
+    cm->show_frame = 1;
 
-  // If encoded in frame parallel mode, frame context is ready after decoding
-  // the frame header.
-  if (pbi->frame_parallel_decode && cm->frame_parallel_decoding_mode) {
-    VP9Worker *const worker = pbi->frame_worker_owner;
-    FrameWorkerData *const frame_worker_data = worker->data1;
-    if (cm->refresh_frame_context) {
-      context_updated = 1;
-      cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+    if (pbi->frame_parallel_decode) {
+      for (i = 0; i < REF_FRAMES; ++i)
+        cm->next_ref_frame_map[i] = cm->ref_frame_map[i];
     }
-    vp9_frameworker_lock_stats(worker);
-    pbi->cur_buf->row = -1;
-    pbi->cur_buf->col = -1;
-    frame_worker_data->frame_context_ready = 1;
-    // Signal the main thread that context is ready.
-    vp9_frameworker_signal_stats(worker);
-    vp9_frameworker_unlock_stats(worker);
+    return 0;
   }
 
-  if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1) {
-    // Multi-threaded tile decoder
-    *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
-    if (!xd->corrupted) {
-      if (!cm->skip_loop_filter) {
-        // If multiple threads are used to decode tiles, then we use those
-        // threads to do parallel loopfiltering.
-        vp9_loop_filter_frame_mt(new_fb, cm, pbi->mb.plane,
-                                 cm->lf.filter_level, 0, 0, pbi->tile_workers,
-                                 pbi->num_tile_workers, &pbi->lf_row_sync);
-      }
-    } else {
-      vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                         "Decode failed. Frame data is corrupted.");
+  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 = INVALID_IDX;
+      cm->frame_refs[i].buf = NULL;
+    }
 
+    setup_frame_size(cm, rb);
+    if (pbi->need_resync) {
+      memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+      pbi->need_resync = 0;
     }
   } else {
-    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
-  }
+    cm->intra_only = cm->show_frame ? 0 : vp9_rb_read_bit(rb);
 
-  if (!xd->corrupted) {
-    if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
-      vp9_adapt_coef_probs(cm);
+    cm->reset_frame_context = cm->error_resilient_mode ?
+        0 : vp9_rb_read_literal(rb, 2);
 
-      if (!frame_is_intra_only(cm)) {
-        vp9_adapt_mode_probs(cm);
-        vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
+    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 format should be YUV 4:2:0 in this
+        // case (normative).
+        cm->color_space = VPX_CS_BT_601;
+        cm->subsampling_y = cm->subsampling_x = 1;
+        cm->bit_depth = VPX_BITS_8;
+#if CONFIG_VP9_HIGHBITDEPTH
+        cm->use_highbitdepth = 0;
+#endif
       }
-    } else {
-      debug_check_frame_counts(cm);
-    }
-  } else {
-    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
-                       "Decode failed. Frame data is corrupted.");
-  }
 
-  // Non frame parallel update frame context here.
-  if (cm->refresh_frame_context && !context_updated)
-    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
-}
+      pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
+      setup_frame_size(cm, rb);
+      if (pbi->need_resync) {
+        memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+        pbi->need_resync = 0;
+      }
+    } else if (pbi->need_resync != 1) {  /* Skip if need resync */
+      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 = &frame_bufs[idx].buf;
+        cm->ref_frame_sign_bias[LAST_FRAME + i] = vp9_rb_read_bit(rb);
+      }
 
-static void build_mc_border(const uint8_t *src, int src_stride,
-                            uint8_t *dst, int dst_stride,
-                            int x, int y, int b_w, int b_h, int w, int h) {
-  // Get a pointer to the start of the real data for this row.
-  const uint8_t *ref_row = src - x - y * src_stride;
+      setup_frame_size_with_refs(cm, rb);
 
-  if (y >= h)
-    ref_row += (h - 1) * src_stride;
-  else if (y > 0)
-    ref_row += y * src_stride;
+      cm->allow_high_precision_mv = vp9_rb_read_bit(rb);
+      cm->interp_filter = read_interp_filter(rb);
 
-  do {
-    int right = 0, copy;
-    int left = x < 0 ? -x : 0;
+      for (i = 0; i < REFS_PER_FRAME; ++i) {
+        RefBuffer *const ref_buf = &cm->frame_refs[i];
+#if CONFIG_VP9_HIGHBITDEPTH
+        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,
+                                          cm->use_highbitdepth);
+#else
+        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);
+#endif
+      }
+    }
+  }
+#if CONFIG_VP9_HIGHBITDEPTH
+  get_frame_new_buffer(cm)->bit_depth = cm->bit_depth;
+#endif
+  get_frame_new_buffer(cm)->color_space = cm->color_space;
+
+  if (pbi->need_resync) {
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Keyframe / intra-only frame required to reset decoder"
+                       " state");
+  }
 
-    if (left > b_w)
-      left = b_w;
+  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;
+  }
 
-    if (x + b_w > w)
-      right = x + b_w - w;
+  // 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 (right > b_w)
-      right = b_w;
+  // Generate next_ref_frame_map.
+  lock_buffer_pool(pool);
+  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+    if (mask & 1) {
+      cm->next_ref_frame_map[ref_index] = cm->new_fb_idx;
+      ++frame_bufs[cm->new_fb_idx].ref_count;
+    } else {
+      cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+    }
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+    ++ref_index;
+  }
 
-    copy = b_w - left - right;
+  for (; ref_index < REF_FRAMES; ++ref_index) {
+    cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+  }
+  unlock_buffer_pool(pool);
+  pbi->hold_ref_buf = 1;
 
-    if (left)
-      memset(dst, ref_row[0], left);
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode)
+    vp9_setup_past_independence(cm);
 
-    if (copy)
-      memcpy(dst + left, ref_row + x + left, copy);
+  setup_loopfilter(&cm->lf, rb);
+  setup_quantization(cm, &pbi->mb, rb);
+  setup_segmentation(&cm->seg, rb);
+  setup_segmentation_dequant(cm);
 
-    if (right)
-      memset(dst + left + copy, ref_row[w - 1], right);
+  setup_tile_info(cm, rb);
+  sz = vp9_rb_read_literal(rb, 16);
 
-    dst += dst_stride;
-    ++y;
+  if (sz == 0)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid header size");
 
-    if (y > 0 && y < h)
-      ref_row += src_stride;
-  } while (--b_h);
+  return sz;
 }
 
-#if CONFIG_VP9_HIGHBITDEPTH
-static void high_build_mc_border(const uint8_t *src8, int src_stride,
-                                 uint16_t *dst, int dst_stride,
-                                 int x, int y, int b_w, int b_h,
-                                 int w, int h) {
-  // Get a pointer to the start of the real data for this row.
-  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
-  const uint16_t *ref_row = src - x - y * src_stride;
-
-  if (y >= h)
-    ref_row += (h - 1) * src_stride;
-  else if (y > 0)
-    ref_row += y * src_stride;
-
-  do {
-    int right = 0, copy;
-    int left = x < 0 ? -x : 0;
+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 (left > b_w)
-      left = b_w;
+  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");
 
-    if (x + b_w > w)
-      right = x + b_w - w;
+  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);
 
-    if (right > b_w)
-      right = b_w;
+  for (k = 0; k < SKIP_CONTEXTS; ++k)
+    vp9_diff_update_prob(&r, &fc->skip_probs[k]);
 
-    copy = b_w - left - right;
+  if (!frame_is_intra_only(cm)) {
+    nmv_context *const nmvc = &fc->nmvc;
+    int i, j;
 
-    if (left)
-      vpx_memset16(dst, ref_row[0], left);
+    read_inter_mode_probs(fc, &r);
 
-    if (copy)
-      memcpy(dst + left, ref_row + x + left, copy * sizeof(uint16_t));
+    if (cm->interp_filter == SWITCHABLE)
+      read_switchable_interp_probs(fc, &r);
 
-    if (right)
-      vpx_memset16(dst + left + copy, ref_row[w - 1], right);
+    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+      vp9_diff_update_prob(&r, &fc->intra_inter_prob[i]);
 
-    dst += dst_stride;
-    ++y;
+    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);
 
-    if (y > 0 && y < h)
-      ref_row += src_stride;
-  } while (--b_h);
-}
-#endif  // CONFIG_VP9_HIGHBITDEPTH
+    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]);
 
-#if CONFIG_VP9_HIGHBITDEPTH
-static void extend_and_predict(const uint8_t *buf_ptr1, int pre_buf_stride,
-                               int x0, int y0, int b_w, int b_h,
-                               int frame_width, int frame_height,
-                               int border_offset,
-                               uint8_t *const dst, int dst_buf_stride,
-                               int subpel_x, int subpel_y,
-                               const InterpKernel *kernel,
-                               const struct scale_factors *sf,
-                               MACROBLOCKD *xd,
-                               int w, int h, int ref, int xs, int ys) {
-  DECLARE_ALIGNED(16, uint16_t, mc_buf_high[80 * 2 * 80 * 2]);
-  const uint8_t *buf_ptr;
+    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]);
 
-  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
-    high_build_mc_border(buf_ptr1, pre_buf_stride, mc_buf_high, b_w,
-                         x0, y0, b_w, b_h, frame_width, frame_height);
-    buf_ptr = CONVERT_TO_BYTEPTR(mc_buf_high) + border_offset;
-  } else {
-    build_mc_border(buf_ptr1, pre_buf_stride, (uint8_t *)mc_buf_high, b_w,
-                    x0, y0, b_w, b_h, frame_width, frame_height);
-    buf_ptr = ((uint8_t *)mc_buf_high) + border_offset;
+    read_mv_probs(nmvc, cm->allow_high_precision_mv, &r);
   }
 
-  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
-    high_inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
-                         subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
-  } else {
-    inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
-                    subpel_y, sf, w, h, ref, kernel, xs, ys);
-  }
+  return vp9_reader_has_error(&r);
 }
-#else
-static void extend_and_predict(const uint8_t *buf_ptr1, int pre_buf_stride,
-                               int x0, int y0, int b_w, int b_h,
-                               int frame_width, int frame_height,
-                               int border_offset,
-                               uint8_t *const dst, int dst_buf_stride,
-                               int subpel_x, int subpel_y,
-                               const InterpKernel *kernel,
-                               const struct scale_factors *sf,
-                               int w, int h, int ref, int xs, int ys) {
-  DECLARE_ALIGNED(16, uint8_t, mc_buf[80 * 2 * 80 * 2]);
-  const uint8_t *buf_ptr;
 
-  build_mc_border(buf_ptr1, pre_buf_stride, mc_buf, b_w,
-                  x0, y0, b_w, b_h, frame_width, frame_height);
-  buf_ptr = mc_buf + border_offset;
-
-  inter_predictor(buf_ptr, b_w, dst, dst_buf_stride, subpel_x,
-                  subpel_y, sf, w, h, ref, kernel, xs, ys);
+#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  // CONFIG_VP9_HIGHBITDEPTH
-
-static void dec_build_inter_predictors(VP9Decoder *const pbi, MACROBLOCKD *xd,
-                                       int plane, int bw, int bh, int x,
-                                       int y, int w, int h, int mi_x, int mi_y,
-                                       const InterpKernel *kernel,
-                                       const struct scale_factors *sf,
-                                       struct buf_2d *pre_buf,
-                                       struct buf_2d *dst_buf, const MV* mv,
-                                       RefCntBuffer *ref_frame_buf,
-                                       int is_scaled, int ref) {
-  struct macroblockd_plane *const pd = &xd->plane[plane];
-  uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
-  MV32 scaled_mv;
-  int xs, ys, x0, y0, x0_16, y0_16, frame_width, frame_height,
-      buf_stride, subpel_x, subpel_y;
-  uint8_t *ref_frame, *buf_ptr;
+#endif  // NDEBUG
 
-  // Get reference frame pointer, width and height.
-  if (plane == 0) {
-    frame_width = ref_frame_buf->buf.y_crop_width;
-    frame_height = ref_frame_buf->buf.y_crop_height;
-    ref_frame = ref_frame_buf->buf.y_buffer;
+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[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 {
-    frame_width = ref_frame_buf->buf.uv_crop_width;
-    frame_height = ref_frame_buf->buf.uv_crop_height;
-    ref_frame = plane == 1 ? ref_frame_buf->buf.u_buffer
-                         : ref_frame_buf->buf.v_buffer;
+    rb->bit_buffer = data;
+    rb->bit_buffer_end = data_end;
   }
+  return rb;
+}
 
-  if (is_scaled) {
-    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, mv, bw, bh,
-                                               pd->subsampling_x,
-                                               pd->subsampling_y);
-    // Co-ordinate of containing block to pixel precision.
-    int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
-    int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
-
-    // Co-ordinate of the block to 1/16th pixel precision.
-    x0_16 = (x_start + x) << SUBPEL_BITS;
-    y0_16 = (y_start + y) << SUBPEL_BITS;
+//------------------------------------------------------------------------------
 
-    // Co-ordinate of current block in reference frame
-    // to 1/16th pixel precision.
-    x0_16 = sf->scale_value_x(x0_16, sf);
-    y0_16 = sf->scale_value_y(y0_16, sf);
+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;
+}
 
-    // Map the top left corner of the block into the reference frame.
-    x0 = sf->scale_value_x(x_start + x, sf);
-    y0 = sf->scale_value_y(y_start + y, sf);
+void vp9_read_frame_size(struct vp9_read_bit_buffer *rb,
+                         int *width, int *height) {
+  *width = vp9_rb_read_literal(rb, 16) + 1;
+  *height = vp9_rb_read_literal(rb, 16) + 1;
+}
 
-    // Scale the MV and incorporate the sub-pixel offset of the block
-    // in the reference frame.
-    scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
-    xs = sf->x_step_q4;
-    ys = sf->y_step_q4;
-  } else {
-    // Co-ordinate of containing block to pixel precision.
-    x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
-    y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+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;
+}
 
-    // Co-ordinate of the block to 1/16th pixel precision.
-    x0_16 = x0 << SUBPEL_BITS;
-    y0_16 = y0 << SUBPEL_BITS;
+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;
+  int context_updated = 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;
 
-    scaled_mv.row = mv->row * (1 << (1 - pd->subsampling_y));
-    scaled_mv.col = mv->col * (1 << (1 - pd->subsampling_x));
-    xs = ys = 16;
+  if (!first_partition_size) {
+    // showing a frame directly
+    *p_data_end = data + (cm->profile <= PROFILE_2 ? 1 : 2);
+    return;
   }
-  subpel_x = scaled_mv.col & SUBPEL_MASK;
-  subpel_y = scaled_mv.row & SUBPEL_MASK;
 
-  // Calculate the top left corner of the best matching block in the
-  // reference frame.
-  x0 += scaled_mv.col >> SUBPEL_BITS;
-  y0 += scaled_mv.row >> SUBPEL_BITS;
-  x0_16 += scaled_mv.col;
-  y0_16 += scaled_mv.row;
-
-  // Get reference block pointer.
-  buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
-  buf_stride = pre_buf->stride;
+  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");
 
-  // Do border extension if there is motion or the
-  // width/height is not a multiple of 8 pixels.
-  if (is_scaled || scaled_mv.col || scaled_mv.row ||
-      (frame_width & 0x7) || (frame_height & 0x7)) {
-    int y1 = (y0_16 + (h - 1) * ys) >> SUBPEL_BITS;
+  cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
+                           cm->width == cm->last_width &&
+                           cm->height == cm->last_height &&
+                           !cm->last_intra_only &&
+                           cm->last_show_frame &&
+                           (cm->last_frame_type != KEY_FRAME);
 
-    // Get reference block bottom right horizontal coordinate.
-    int x1 = (x0_16 + (w - 1) * xs) >> SUBPEL_BITS;
-    int x_pad = 0, y_pad = 0;
+  vp9_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);
 
-    if (subpel_x || (sf->x_step_q4 != SUBPEL_SHIFTS)) {
-      x0 -= VP9_INTERP_EXTEND - 1;
-      x1 += VP9_INTERP_EXTEND;
-      x_pad = 1;
-    }
+  *cm->fc = cm->frame_contexts[cm->frame_context_idx];
+  if (!cm->fc->initialized)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Uninitialized entropy context.");
 
-    if (subpel_y || (sf->y_step_q4 != SUBPEL_SHIFTS)) {
-      y0 -= VP9_INTERP_EXTEND - 1;
-      y1 += VP9_INTERP_EXTEND;
-      y_pad = 1;
-    }
+  vp9_zero(cm->counts);
 
-    // Wait until reference block is ready. Pad 7 more pixels as last 7
-    // pixels of each superblock row can be changed by next superblock row.
-    if (pbi->frame_parallel_decode)
-      vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
-                           MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
+  xd->corrupted = 0;
+  new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
+  if (new_fb->corrupted)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data header is corrupted.");
 
-    // Skip border extension if block is inside the frame.
-    if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width - 1 ||
-        y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
-      // Extend the border.
-      const uint8_t *const buf_ptr1 = ref_frame + y0 * buf_stride + x0;
-      const int b_w = x1 - x0 + 1;
-      const int b_h = y1 - y0 + 1;
-      const int border_offset = y_pad * 3 * b_w + x_pad * 3;
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
+  }
 
-      extend_and_predict(buf_ptr1, buf_stride, x0, y0, b_w, b_h,
-                         frame_width, frame_height, border_offset,
-                         dst, dst_buf->stride,
-                         subpel_x, subpel_y,
-                         kernel, sf,
-#if CONFIG_VP9_HIGHBITDEPTH
-                         xd,
-#endif
-                         w, h, ref, xs, ys);
-      return;
+  // If encoded in frame parallel mode, frame context is ready after decoding
+  // the frame header.
+  if (pbi->frame_parallel_decode && cm->frame_parallel_decoding_mode) {
+    VP9Worker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    if (cm->refresh_frame_context) {
+      context_updated = 1;
+      cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
     }
-  } else {
-    // Wait until reference block is ready. Pad 7 more pixels as last 7
-    // pixels of each superblock row can be changed by next superblock row.
-     if (pbi->frame_parallel_decode) {
-       const int y1 = (y0_16 + (h - 1) * ys) >> SUBPEL_BITS;
-       vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
-                            MAX(0, (y1 + 7)) << (plane == 0 ? 0 : 1));
-     }
-  }
-#if CONFIG_VP9_HIGHBITDEPTH
-  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
-    high_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
-                         subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
-  } else {
-    inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
-                    subpel_y, sf, w, h, ref, kernel, xs, ys);
+    vp9_frameworker_lock_stats(worker);
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+    frame_worker_data->frame_context_ready = 1;
+    // Signal the main thread that context is ready.
+    vp9_frameworker_signal_stats(worker);
+    vp9_frameworker_unlock_stats(worker);
   }
-#else
-  inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
-                  subpel_y, sf, w, h, ref, kernel, xs, ys);
-#endif  // CONFIG_VP9_HIGHBITDEPTH
-}
-
-void vp9_dec_build_inter_predictors_sb(VP9Decoder *const pbi, MACROBLOCKD *xd,
-                                       int mi_row, int mi_col,
-                                       BLOCK_SIZE bsize) {
-  int plane;
-  const int mi_x = mi_col * MI_SIZE;
-  const int mi_y = mi_row * MI_SIZE;
-  const MODE_INFO *mi = xd->mi[0];
-  const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
-  const BLOCK_SIZE sb_type = mi->mbmi.sb_type;
-  const int is_compound = has_second_ref(&mi->mbmi);
 
-  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
-    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
-                                                        &xd->plane[plane]);
-    struct macroblockd_plane *const pd = &xd->plane[plane];
-    struct buf_2d *const dst_buf = &pd->dst;
-    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
-    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+  if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1) {
+    // Multi-threaded tile decoder
+    *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
+    if (!xd->corrupted) {
+      if (!cm->skip_loop_filter) {
+        // If multiple threads are used to decode tiles, then we use those
+        // threads to do parallel loopfiltering.
+        vp9_loop_filter_frame_mt(new_fb, cm, pbi->mb.plane,
+                                 cm->lf.filter_level, 0, 0, pbi->tile_workers,
+                                 pbi->num_tile_workers, &pbi->lf_row_sync);
+      }
+    } else {
+      vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                         "Decode failed. Frame data is corrupted.");
 
-    const int bw = 4 * num_4x4_w;
-    const int bh = 4 * num_4x4_h;
-    int ref;
+    }
+  } else {
+    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
+  }
 
-    for (ref = 0; ref < 1 + is_compound; ++ref) {
-      const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
-      struct buf_2d *const pre_buf = &pd->pre[ref];
-      const int idx = xd->block_refs[ref]->idx;
-      BufferPool *const pool = pbi->common.buffer_pool;
-      RefCntBuffer *const ref_frame_buf = &pool->frame_bufs[idx];
-      const int is_scaled = vp9_is_scaled(sf);
+  if (!xd->corrupted) {
+    if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
+      vp9_adapt_coef_probs(cm);
 
-      if (sb_type < BLOCK_8X8) {
-        int i = 0, x, y;
-        assert(bsize == BLOCK_8X8);
-        for (y = 0; y < num_4x4_h; ++y) {
-          for (x = 0; x < num_4x4_w; ++x) {
-            const MV mv = average_split_mvs(pd, mi, ref, i++);
-            dec_build_inter_predictors(pbi, xd, plane, bw, bh,
-                                       4 * x, 4 * y, 4, 4, mi_x, mi_y, kernel,
-                                       sf, pre_buf, dst_buf, &mv,
-                                       ref_frame_buf, is_scaled, ref);
-          }
-        }
-      } else {
-        const MV mv = mi->mbmi.mv[ref].as_mv;
-        dec_build_inter_predictors(pbi, xd, plane, bw, bh,
-                                   0, 0, bw, bh, mi_x, mi_y, kernel,
-                                   sf, pre_buf, dst_buf, &mv, ref_frame_buf,
-                                   is_scaled, ref);
+      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.");
   }
+
+  // Non frame parallel update frame context here.
+  if (cm->refresh_frame_context && !context_updated)
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
 }