Add m-esr52 at 52.6.0

16 files changed, 4336 insertions, 0 deletions

diff --git a/media/libvpx/vp9/decoder/vp9_decodeframe.c b/media/libvpx/vp9/decoder/vp9_decodeframe.c
new file mode 100644
index 000000000..30ca2d08a
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_decodeframe.c
@@ -0,0 +1,2077 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <stdlib.h>  // qsort()
+
+#include "./vp9_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/mem_ops.h"
+#include "vpx_scale/vpx_scale.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_idct.h"
+#include "vp9/common/vp9_thread_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_thread.h"
+#include "vp9/common/vp9_tile_common.h"
+
+#include "vp9/decoder/vp9_decodeframe.h"
+#include "vp9/decoder/vp9_detokenize.h"
+#include "vp9/decoder/vp9_decodemv.h"
+#include "vp9/decoder/vp9_decoder.h"
+#include "vp9/decoder/vp9_dsubexp.h"
+#include "vp9/decoder/vp9_read_bit_buffer.h"
+#include "vp9/decoder/vp9_reader.h"
+
+#define MAX_VP9_HEADER_SIZE 80
+
+static int is_compound_reference_allowed(const VP9_COMMON *cm) {
+  int i;
+  for (i = 1; i < REFS_PER_FRAME; ++i)
+    if (cm->ref_frame_sign_bias[i + 1] != cm->ref_frame_sign_bias[1])
+      return 1;
+
+  return 0;
+}
+
+static void setup_compound_reference_mode(VP9_COMMON *cm) {
+  if (cm->ref_frame_sign_bias[LAST_FRAME] ==
+          cm->ref_frame_sign_bias[GOLDEN_FRAME]) {
+    cm->comp_fixed_ref = ALTREF_FRAME;
+    cm->comp_var_ref[0] = LAST_FRAME;
+    cm->comp_var_ref[1] = GOLDEN_FRAME;
+  } else if (cm->ref_frame_sign_bias[LAST_FRAME] ==
+                 cm->ref_frame_sign_bias[ALTREF_FRAME]) {
+    cm->comp_fixed_ref = GOLDEN_FRAME;
+    cm->comp_var_ref[0] = LAST_FRAME;
+    cm->comp_var_ref[1] = ALTREF_FRAME;
+  } else {
+    cm->comp_fixed_ref = LAST_FRAME;
+    cm->comp_var_ref[0] = GOLDEN_FRAME;
+    cm->comp_var_ref[1] = ALTREF_FRAME;
+  }
+}
+
+static int read_is_valid(const uint8_t *start, size_t len, const uint8_t *end) {
+  return len != 0 && len <= (size_t)(end - start);
+}
+
+static int decode_unsigned_max(struct vp9_read_bit_buffer *rb, int max) {
+  const int data = vp9_rb_read_literal(rb, get_unsigned_bits(max));
+  return data > max ? max : data;
+}
+
+static TX_MODE read_tx_mode(vp9_reader *r) {
+  TX_MODE tx_mode = vp9_read_literal(r, 2);
+  if (tx_mode == ALLOW_32X32)
+    tx_mode += vp9_read_bit(r);
+  return tx_mode;
+}
+
+static void read_tx_mode_probs(struct tx_probs *tx_probs, vp9_reader *r) {
+  int i, j;
+
+  for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
+    for (j = 0; j < TX_SIZES - 3; ++j)
+      vp9_diff_update_prob(r, &tx_probs->p8x8[i][j]);
+
+  for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
+    for (j = 0; j < TX_SIZES - 2; ++j)
+      vp9_diff_update_prob(r, &tx_probs->p16x16[i][j]);
+
+  for (i = 0; i < TX_SIZE_CONTEXTS; ++i)
+    for (j = 0; j < TX_SIZES - 1; ++j)
+      vp9_diff_update_prob(r, &tx_probs->p32x32[i][j]);
+}
+
+static void read_switchable_interp_probs(FRAME_CONTEXT *fc, vp9_reader *r) {
+  int i, j;
+  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
+    for (i = 0; i < SWITCHABLE_FILTERS - 1; ++i)
+      vp9_diff_update_prob(r, &fc->switchable_interp_prob[j][i]);
+}
+
+static void read_inter_mode_probs(FRAME_CONTEXT *fc, vp9_reader *r) {
+  int i, j;
+  for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
+    for (j = 0; j < INTER_MODES - 1; ++j)
+      vp9_diff_update_prob(r, &fc->inter_mode_probs[i][j]);
+}
+
+static REFERENCE_MODE read_frame_reference_mode(const VP9_COMMON *cm,
+                                                vp9_reader *r) {
+  if (is_compound_reference_allowed(cm)) {
+    return vp9_read_bit(r) ? (vp9_read_bit(r) ? REFERENCE_MODE_SELECT
+                                              : COMPOUND_REFERENCE)
+                           : SINGLE_REFERENCE;
+  } else {
+    return SINGLE_REFERENCE;
+  }
+}
+
+static void read_frame_reference_mode_probs(VP9_COMMON *cm, vp9_reader *r) {
+  FRAME_CONTEXT *const fc = cm->fc;
+  int i;
+
+  if (cm->reference_mode == REFERENCE_MODE_SELECT)
+    for (i = 0; i < COMP_INTER_CONTEXTS; ++i)
+      vp9_diff_update_prob(r, &fc->comp_inter_prob[i]);
+
+  if (cm->reference_mode != COMPOUND_REFERENCE)
+    for (i = 0; i < REF_CONTEXTS; ++i) {
+      vp9_diff_update_prob(r, &fc->single_ref_prob[i][0]);
+      vp9_diff_update_prob(r, &fc->single_ref_prob[i][1]);
+    }
+
+  if (cm->reference_mode != SINGLE_REFERENCE)
+    for (i = 0; i < REF_CONTEXTS; ++i)
+      vp9_diff_update_prob(r, &fc->comp_ref_prob[i]);
+}
+
+static void update_mv_probs(vp9_prob *p, int n, vp9_reader *r) {
+  int i;
+  for (i = 0; i < n; ++i)
+    if (vp9_read(r, MV_UPDATE_PROB))
+      p[i] = (vp9_read_literal(r, 7) << 1) | 1;
+}
+
+static void read_mv_probs(nmv_context *ctx, int allow_hp, vp9_reader *r) {
+  int i, j;
+
+  update_mv_probs(ctx->joints, MV_JOINTS - 1, r);
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *const comp_ctx = &ctx->comps[i];
+    update_mv_probs(&comp_ctx->sign, 1, r);
+    update_mv_probs(comp_ctx->classes, MV_CLASSES - 1, r);
+    update_mv_probs(comp_ctx->class0, CLASS0_SIZE - 1, r);
+    update_mv_probs(comp_ctx->bits, MV_OFFSET_BITS, r);
+  }
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *const comp_ctx = &ctx->comps[i];
+    for (j = 0; j < CLASS0_SIZE; ++j)
+      update_mv_probs(comp_ctx->class0_fp[j], MV_FP_SIZE - 1, r);
+    update_mv_probs(comp_ctx->fp, 3, r);
+  }
+
+  if (allow_hp) {
+    for (i = 0; i < 2; ++i) {
+      nmv_component *const comp_ctx = &ctx->comps[i];
+      update_mv_probs(&comp_ctx->class0_hp, 1, r);
+      update_mv_probs(&comp_ctx->hp, 1, r);
+    }
+  }
+}
+
+static void inverse_transform_block(MACROBLOCKD* xd, int plane, int block,
+                                    TX_SIZE tx_size, uint8_t *dst, int stride,
+                                    int eob) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  if (eob > 0) {
+    TX_TYPE tx_type = DCT_DCT;
+    tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      if (xd->lossless) {
+        tx_type = DCT_DCT;
+        vp9_highbd_iwht4x4_add(dqcoeff, dst, stride, eob, xd->bd);
+      } else {
+        const PLANE_TYPE plane_type = pd->plane_type;
+        switch (tx_size) {
+          case TX_4X4:
+            tx_type = get_tx_type_4x4(plane_type, xd, block);
+            vp9_highbd_iht4x4_add(tx_type, dqcoeff, dst, stride, eob, xd->bd);
+            break;
+          case TX_8X8:
+            tx_type = get_tx_type(plane_type, xd);
+            vp9_highbd_iht8x8_add(tx_type, dqcoeff, dst, stride, eob, xd->bd);
+            break;
+          case TX_16X16:
+            tx_type = get_tx_type(plane_type, xd);
+            vp9_highbd_iht16x16_add(tx_type, dqcoeff, dst, stride, eob, xd->bd);
+            break;
+          case TX_32X32:
+            tx_type = DCT_DCT;
+            vp9_highbd_idct32x32_add(dqcoeff, dst, stride, eob, xd->bd);
+            break;
+          default:
+            assert(0 && "Invalid transform size");
+        }
+      }
+    } else {
+      if (xd->lossless) {
+        tx_type = DCT_DCT;
+        vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
+      } else {
+        const PLANE_TYPE plane_type = pd->plane_type;
+        switch (tx_size) {
+          case TX_4X4:
+            tx_type = get_tx_type_4x4(plane_type, xd, block);
+            vp9_iht4x4_add(tx_type, dqcoeff, dst, stride, eob);
+            break;
+          case TX_8X8:
+            tx_type = get_tx_type(plane_type, xd);
+            vp9_iht8x8_add(tx_type, dqcoeff, dst, stride, eob);
+            break;
+          case TX_16X16:
+            tx_type = get_tx_type(plane_type, xd);
+            vp9_iht16x16_add(tx_type, dqcoeff, dst, stride, eob);
+            break;
+          case TX_32X32:
+            tx_type = DCT_DCT;
+            vp9_idct32x32_add(dqcoeff, dst, stride, eob);
+            break;
+          default:
+            assert(0 && "Invalid transform size");
+            return;
+        }
+      }
+    }
+#else
+    if (xd->lossless) {
+      tx_type = DCT_DCT;
+      vp9_iwht4x4_add(dqcoeff, dst, stride, eob);
+    } else {
+      const PLANE_TYPE plane_type = pd->plane_type;
+      switch (tx_size) {
+        case TX_4X4:
+          tx_type = get_tx_type_4x4(plane_type, xd, block);
+          vp9_iht4x4_add(tx_type, dqcoeff, dst, stride, eob);
+          break;
+        case TX_8X8:
+          tx_type = get_tx_type(plane_type, xd);
+          vp9_iht8x8_add(tx_type, dqcoeff, dst, stride, eob);
+          break;
+        case TX_16X16:
+          tx_type = get_tx_type(plane_type, xd);
+          vp9_iht16x16_add(tx_type, dqcoeff, dst, stride, eob);
+          break;
+        case TX_32X32:
+          tx_type = DCT_DCT;
+          vp9_idct32x32_add(dqcoeff, dst, stride, eob);
+          break;
+        default:
+          assert(0 && "Invalid transform size");
+          return;
+      }
+    }
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+    if (eob == 1) {
+      memset(dqcoeff, 0, 2 * sizeof(dqcoeff[0]));
+    } else {
+      if (tx_type == DCT_DCT && tx_size <= TX_16X16 && eob <= 10)
+        memset(dqcoeff, 0, 4 * (4 << tx_size) * sizeof(dqcoeff[0]));
+      else if (tx_size == TX_32X32 && eob <= 34)
+        memset(dqcoeff, 0, 256 * sizeof(dqcoeff[0]));
+      else
+        memset(dqcoeff, 0, (16 << (tx_size << 1)) * sizeof(dqcoeff[0]));
+    }
+  }
+}
+
+struct intra_args {
+  MACROBLOCKD *xd;
+  vp9_reader *r;
+  int seg_id;
+};
+
+static void predict_and_reconstruct_intra_block(int plane, int block,
+                                                BLOCK_SIZE plane_bsize,
+                                                TX_SIZE tx_size, void *arg) {
+  struct intra_args *const args = (struct intra_args *)arg;
+  MACROBLOCKD *const xd = args->xd;
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  MODE_INFO *const mi = xd->mi[0];
+  const PREDICTION_MODE mode = (plane == 0) ? get_y_mode(mi, block)
+                                            : mi->mbmi.uv_mode;
+  int x, y;
+  uint8_t *dst;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
+  dst = &pd->dst.buf[4 * y * pd->dst.stride + 4 * x];
+
+  vp9_predict_intra_block(xd, block >> (tx_size << 1),
+                          b_width_log2_lookup[plane_bsize], tx_size, mode,
+                          dst, pd->dst.stride, dst, pd->dst.stride,
+                          x, y, plane);
+
+  if (!mi->mbmi.skip) {
+    const int eob = vp9_decode_block_tokens(xd, plane, block,
+                                            plane_bsize, x, y, tx_size,
+                                            args->r, args->seg_id);
+    inverse_transform_block(xd, plane, block, tx_size, dst, pd->dst.stride,
+                            eob);
+  }
+}
+
+struct inter_args {
+  MACROBLOCKD *xd;
+  vp9_reader *r;
+  int *eobtotal;
+  int seg_id;
+};
+
+static void reconstruct_inter_block(int plane, int block,
+                                    BLOCK_SIZE plane_bsize,
+                                    TX_SIZE tx_size, void *arg) {
+  struct inter_args *args = (struct inter_args *)arg;
+  MACROBLOCKD *const xd = args->xd;
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  int x, y, eob;
+  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
+  eob = vp9_decode_block_tokens(xd, plane, block, plane_bsize,
+                                x, y, tx_size, args->r, args->seg_id);
+  inverse_transform_block(xd, plane, block, tx_size,
+                          &pd->dst.buf[4 * y * pd->dst.stride + 4 * x],
+                          pd->dst.stride, eob);
+  *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];
+    }
+
+  set_skip_context(xd, mi_row, mi_col);
+
+  // 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 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 (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.");
+  }
+
+  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);
+  }
+
+  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);
+
+    // 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
+    }
+  }
+
+  xd->corrupted |= vp9_reader_has_error(r);
+}
+
+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 (counts)
+    ++counts->partition[ctx][p];
+
+  return p;
+}
+
+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 (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
+    return;
+
+  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");
+    }
+  }
+
+  // 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 read_coef_probs_common(vp9_coeff_probs_model *coef_probs,
+                                   vp9_reader *r) {
+  int i, j, k, l, m;
+
+  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]);
+}
+
+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);
+}
+
+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;
+
+  // 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 {
+      for (i = 0; i < PREDICTION_PROBS; i++)
+        seg->pred_probs[i] = MAX_PROB;
+    }
+  }
+
+  // Segmentation data update
+  seg->update_data = vp9_rb_read_bit(rb);
+  if (seg->update_data) {
+    seg->abs_delta = vp9_rb_read_bit(rb);
+
+    vp9_clearall_segfeatures(seg);
+
+    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 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);
+
+  // Read in loop filter deltas applied at the MB level based on mode or ref
+  // frame.
+  lf->mode_ref_delta_update = 0;
+
+  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;
+
+      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);
+
+      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);
+    }
+  }
+}
+
+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;
+}
+
+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;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  xd->bd = (int)cm->bit_depth;
+#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);
+    }
+  } 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);
+  }
+}
+
+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)];
+}
+
+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 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);
+}
+
+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 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;
+
+    // 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);
+  }
+}
+
+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);
+
+  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);
+
+  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;
+}
+
+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 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 (!found)
+    vp9_read_frame_size(rb, &width, &height);
+
+  if (width <= 0 || height <= 0)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid frame size");
+
+  // 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");
+  }
+
+  resize_context_buffers(cm, width, height);
+  setup_display_size(cm, rb);
+
+  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);
+
+  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;
+}
+
+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++;
+
+  if (cm->log2_tile_cols > 6)
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Invalid number of tile columns");
+
+  // rows
+  cm->log2_tile_rows = vp9_rb_read_bit(rb);
+  if (cm->log2_tile_rows)
+    cm->log2_tile_rows += vp9_rb_read_bit(rb);
+}
+
+typedef struct TileBuffer {
+  const uint8_t *data;
+  size_t size;
+  int col;  // only used with multi-threaded decoding
+} TileBuffer;
+
+// 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;
+
+  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");
+
+    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;
+
+    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;
+}
+
+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;
+
+  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 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);
+
+  memset(cm->above_seg_context, 0,
+         sizeof(*cm->above_seg_context) * aligned_cols);
+
+  get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers);
+
+  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;
+  }
+
+  // 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 (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;
+
+        // delay the loopfilter by 1 macroblock row.
+        if (lf_start < 0) continue;
+
+        // decoding has completed: finish up the loop filter in this thread.
+        if (mi_row + MI_BLOCK_SIZE >= cm->mi_rows) continue;
+
+        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);
+    }
+  }
+
+  // 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);
+  }
+
+  // Get last tile data.
+  tile_data = pbi->tile_data + tile_cols * tile_rows - 1;
+
+  if (pbi->frame_parallel_decode)
+    vp9_frameworker_broadcast(pbi->cur_buf, INT_MAX);
+  return vp9_reader_find_end(&tile_data->bit_reader);
+}
+
+static int tile_worker_hook(TileWorkerData *const tile_data,
+                            const TileInfo *const tile) {
+  int mi_row, mi_col;
+
+  if (setjmp(tile_data->error_info.jmp)) {
+    tile_data->error_info.setjmp = 0;
+    tile_data->xd.corrupted = 1;
+    return 0;
+  }
+
+  tile_data->error_info.setjmp = 1;
+  tile_data->xd.error_info = &tile_data->error_info;
+
+  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);
+    }
+  }
+  return !tile_data->xd.corrupted;
+}
+
+// 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 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;
+
+  assert(tile_cols <= (1 << 6));
+  assert(tile_rows == 1);
+  (void)tile_rows;
+
+  // TODO(jzern): See if we can remove the restriction of passing in max
+  // threads to the decoder.
+  if (pbi->num_tile_workers == 0) {
+    const int num_threads = pbi->max_threads & ~1;
+    int i;
+    // TODO(jzern): Allocate one less worker, as in the current code we only
+    // use num_threads - 1 workers.
+    CHECK_MEM_ERROR(cm, pbi->tile_workers,
+                    vpx_malloc(num_threads * sizeof(*pbi->tile_workers)));
+    // 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");
+      }
+    }
+  }
+
+  // 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];
+  }
+
+  // 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);
+
+  // 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;
+    }
+  }
+
+  // Initialize thread frame counts.
+  if (!cm->frame_parallel_decoding_mode) {
+    int i;
+
+    for (i = 0; i < num_workers; ++i) {
+      TileWorkerData *const tile_data =
+          (TileWorkerData*)pbi->tile_workers[i].data1;
+      vp9_zero(tile_data->counts);
+    }
+  }
+
+  n = 0;
+  while (n < tile_cols) {
+    int i;
+    for (i = 0; i < num_workers && n < tile_cols; ++i) {
+      VP9Worker *const worker = &pbi->tile_workers[i];
+      TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
+      TileInfo *const tile = (TileInfo*)worker->data2;
+      TileBuffer *const buf = &tile_buffers[0][n];
+
+      tile_data->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);
+
+      worker->had_error = 0;
+      if (i == num_workers - 1 || n == tile_cols - 1) {
+        winterface->execute(worker);
+      } else {
+        winterface->launch(worker);
+      }
+
+      if (buf->col == tile_cols - 1) {
+        final_worker = i;
+      }
+
+      ++n;
+    }
+
+    for (; i > 0; --i) {
+      VP9Worker *const worker = &pbi->tile_workers[i - 1];
+      // 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;
+    }
+
+    // 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 bit_reader_end;
+}
+
+static void error_handler(void *data) {
+  VP9_COMMON *const cm = (VP9_COMMON *)data;
+  vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME, "Truncated packet");
+}
+
+int vp9_read_sync_code(struct vp9_read_bit_buffer *const rb) {
+  return vp9_rb_read_literal(rb, 8) == VP9_SYNC_CODE_0 &&
+         vp9_rb_read_literal(rb, 8) == VP9_SYNC_CODE_1 &&
+         vp9_rb_read_literal(rb, 8) == VP9_SYNC_CODE_2;
+}
+
+BITSTREAM_PROFILE vp9_read_profile(struct vp9_read_bit_buffer *rb) {
+  int profile = vp9_rb_read_bit(rb);
+  profile |= vp9_rb_read_bit(rb) << 1;
+  if (profile > 2)
+    profile += vp9_rb_read_bit(rb);
+  return (BITSTREAM_PROFILE) profile;
+}
+
+static void read_bitdepth_colorspace_sampling(
+    VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
+  if (cm->profile >= PROFILE_2) {
+    cm->bit_depth = vp9_rb_read_bit(rb) ? VPX_BITS_12 : VPX_BITS_10;
+#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");
+    }
+  }
+}
+
+static size_t read_uncompressed_header(VP9Decoder *pbi,
+                                       struct vp9_read_bit_buffer *rb) {
+  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;
+
+  cm->last_frame_type = cm->frame_type;
+
+  if (vp9_rb_read_literal(rb, 2) != VP9_FRAME_MARKER)
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Invalid frame marker");
+
+  cm->profile = vp9_read_profile(rb);
+
+  if (cm->profile >= MAX_PROFILES)
+    vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                       "Unsupported bitstream profile");
+
+  cm->show_existing_frame = vp9_rb_read_bit(rb);
+  if (cm->show_existing_frame) {
+    // Show an existing frame directly.
+    const int frame_to_show = cm->ref_frame_map[vp9_rb_read_literal(rb, 3)];
+    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);
+    }
+
+    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->frame_parallel_decode) {
+      for (i = 0; i < REF_FRAMES; ++i)
+        cm->next_ref_frame_map[i] = cm->ref_frame_map[i];
+    }
+    return 0;
+  }
+
+  cm->frame_type = (FRAME_TYPE) vp9_rb_read_bit(rb);
+  cm->show_frame = vp9_rb_read_bit(rb);
+  cm->error_resilient_mode = vp9_rb_read_bit(rb);
+
+  if (cm->frame_type == KEY_FRAME) {
+    if (!vp9_read_sync_code(rb))
+      vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Invalid frame sync code");
+
+    read_bitdepth_colorspace_sampling(cm, rb);
+    pbi->refresh_frame_flags = (1 << REF_FRAMES) - 1;
+
+    for (i = 0; i < REFS_PER_FRAME; ++i) {
+      cm->frame_refs[i].idx = 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 {
+    cm->intra_only = cm->show_frame ? 0 : vp9_rb_read_bit(rb);
+
+    cm->reset_frame_context = cm->error_resilient_mode ?
+        0 : vp9_rb_read_literal(rb, 2);
+
+    if (cm->intra_only) {
+      if (!vp9_read_sync_code(rb))
+        vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
+                           "Invalid frame sync code");
+      if (cm->profile > PROFILE_0) {
+        read_bitdepth_colorspace_sampling(cm, rb);
+      } else {
+        // NOTE: The intra-only frame header does not include the specification
+        // of either the color format or color sub-sampling in profile 0. VP9
+        // specifies that the default color 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
+      }
+
+      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);
+      }
+
+      setup_frame_size_with_refs(cm, rb);
+
+      cm->allow_high_precision_mv = vp9_rb_read_bit(rb);
+      cm->interp_filter = read_interp_filter(rb);
+
+      for (i = 0; i < REFS_PER_FRAME; ++i) {
+        RefBuffer *const ref_buf = &cm->frame_refs[i];
+#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 (!cm->error_resilient_mode) {
+    cm->refresh_frame_context = vp9_rb_read_bit(rb);
+    cm->frame_parallel_decoding_mode = vp9_rb_read_bit(rb);
+  } else {
+    cm->refresh_frame_context = 0;
+    cm->frame_parallel_decoding_mode = 1;
+  }
+
+  // This flag will be overridden by the call to vp9_setup_past_independence
+  // below, forcing the use of context 0 for those frame types.
+  cm->frame_context_idx = vp9_rb_read_literal(rb, FRAME_CONTEXTS_LOG2);
+
+  // 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;
+  }
+
+  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 (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;
+}
+
+static int read_compressed_header(VP9Decoder *pbi, const uint8_t *data,
+                                  size_t partition_size) {
+  VP9_COMMON *const cm = &pbi->common;
+  MACROBLOCKD *const xd = &pbi->mb;
+  FRAME_CONTEXT *const fc = cm->fc;
+  vp9_reader r;
+  int k;
+
+  if (vp9_reader_init(&r, data, partition_size, pbi->decrypt_cb,
+                      pbi->decrypt_state))
+    vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
+                       "Failed to allocate bool decoder 0");
+
+  cm->tx_mode = xd->lossless ? ONLY_4X4 : read_tx_mode(&r);
+  if (cm->tx_mode == TX_MODE_SELECT)
+    read_tx_mode_probs(&fc->tx_probs, &r);
+  read_coef_probs(fc, cm->tx_mode, &r);
+
+  for (k = 0; k < SKIP_CONTEXTS; ++k)
+    vp9_diff_update_prob(&r, &fc->skip_probs[k]);
+
+  if (!frame_is_intra_only(cm)) {
+    nmv_context *const nmvc = &fc->nmvc;
+    int i, j;
+
+    read_inter_mode_probs(fc, &r);
+
+    if (cm->interp_filter == SWITCHABLE)
+      read_switchable_interp_probs(fc, &r);
+
+    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+      vp9_diff_update_prob(&r, &fc->intra_inter_prob[i]);
+
+    cm->reference_mode = read_frame_reference_mode(cm, &r);
+    if (cm->reference_mode != SINGLE_REFERENCE)
+      setup_compound_reference_mode(cm);
+    read_frame_reference_mode_probs(cm, &r);
+
+    for (j = 0; j < BLOCK_SIZE_GROUPS; j++)
+      for (i = 0; i < INTRA_MODES - 1; ++i)
+        vp9_diff_update_prob(&r, &fc->y_mode_prob[j][i]);
+
+    for (j = 0; j < PARTITION_CONTEXTS; ++j)
+      for (i = 0; i < PARTITION_TYPES - 1; ++i)
+        vp9_diff_update_prob(&r, &fc->partition_prob[j][i]);
+
+    read_mv_probs(nmvc, cm->allow_high_precision_mv, &r);
+  }
+
+  return vp9_reader_has_error(&r);
+}
+
+#ifdef NDEBUG
+#define debug_check_frame_counts(cm) (void)0
+#else  // !NDEBUG
+// Counts should only be incremented when frame_parallel_decoding_mode and
+// error_resilient_mode are disabled.
+static void debug_check_frame_counts(const VP9_COMMON *const cm) {
+  FRAME_COUNTS zero_counts;
+  vp9_zero(zero_counts);
+  assert(cm->frame_parallel_decoding_mode || cm->error_resilient_mode);
+  assert(!memcmp(cm->counts.y_mode, zero_counts.y_mode,
+                 sizeof(cm->counts.y_mode)));
+  assert(!memcmp(cm->counts.uv_mode, zero_counts.uv_mode,
+                 sizeof(cm->counts.uv_mode)));
+  assert(!memcmp(cm->counts.partition, zero_counts.partition,
+                 sizeof(cm->counts.partition)));
+  assert(!memcmp(cm->counts.coef, zero_counts.coef,
+                 sizeof(cm->counts.coef)));
+  assert(!memcmp(cm->counts.eob_branch, zero_counts.eob_branch,
+                 sizeof(cm->counts.eob_branch)));
+  assert(!memcmp(cm->counts.switchable_interp, zero_counts.switchable_interp,
+                 sizeof(cm->counts.switchable_interp)));
+  assert(!memcmp(cm->counts.inter_mode, zero_counts.inter_mode,
+                 sizeof(cm->counts.inter_mode)));
+  assert(!memcmp(cm->counts.intra_inter, zero_counts.intra_inter,
+                 sizeof(cm->counts.intra_inter)));
+  assert(!memcmp(cm->counts.comp_inter, zero_counts.comp_inter,
+                 sizeof(cm->counts.comp_inter)));
+  assert(!memcmp(cm->counts.single_ref, zero_counts.single_ref,
+                 sizeof(cm->counts.single_ref)));
+  assert(!memcmp(cm->counts.comp_ref, zero_counts.comp_ref,
+                 sizeof(cm->counts.comp_ref)));
+  assert(!memcmp(&cm->counts.tx, &zero_counts.tx, sizeof(cm->counts.tx)));
+  assert(!memcmp(cm->counts.skip, zero_counts.skip, sizeof(cm->counts.skip)));
+  assert(!memcmp(&cm->counts.mv, &zero_counts.mv, sizeof(cm->counts.mv)));
+}
+#endif  // NDEBUG
+
+static struct vp9_read_bit_buffer* init_read_bit_buffer(
+    VP9Decoder *pbi,
+    struct vp9_read_bit_buffer *rb,
+    const uint8_t *data,
+    const uint8_t *data_end,
+    uint8_t *clear_data /* buffer size MAX_VP9_HEADER_SIZE */) {
+  rb->bit_offset = 0;
+  rb->error_handler = error_handler;
+  rb->error_handler_data = &pbi->common;
+  if (pbi->decrypt_cb) {
+    const int n = (int)MIN(MAX_VP9_HEADER_SIZE, data_end - data);
+    pbi->decrypt_cb(pbi->decrypt_state, data, clear_data, n);
+    rb->bit_buffer = clear_data;
+    rb->bit_buffer_end = clear_data + n;
+  } else {
+    rb->bit_buffer = data;
+    rb->bit_buffer_end = data_end;
+  }
+  return rb;
+}
+
+void vp9_decode_frame(VP9Decoder *pbi,
+                      const uint8_t *data, const uint8_t *data_end,
+                      const uint8_t **p_data_end) {
+  VP9_COMMON *const cm = &pbi->common;
+  MACROBLOCKD *const xd = &pbi->mb;
+  struct vp9_read_bit_buffer rb = { NULL, NULL, 0, NULL, 0};
+  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");
+
+  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;
+
+  vp9_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);
+
+  *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.");
+
+  vp9_zero(cm->counts);
+
+  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.");
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
+  }
+
+  // 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;
+    }
+    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);
+  }
+
+  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.");
+
+    }
+  } else {
+    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
+  }
+
+  if (!xd->corrupted) {
+    if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
+      vp9_adapt_coef_probs(cm);
+
+      if (!frame_is_intra_only(cm)) {
+        vp9_adapt_mode_probs(cm);
+        vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
+      }
+    } else {
+      debug_check_frame_counts(cm);
+    }
+  } else {
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data is corrupted.");
+  }
+
+  // Non frame parallel update frame context here.
+  if (cm->refresh_frame_context && !context_updated)
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+}
+
+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;
+
+  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;
+
+    if (left > b_w)
+      left = b_w;
+
+    if (x + b_w > w)
+      right = x + b_w - w;
+
+    if (right > b_w)
+      right = b_w;
+
+    copy = b_w - left - right;
+
+    if (left)
+      memset(dst, ref_row[0], left);
+
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy);
+
+    if (right)
+      memset(dst + left + copy, ref_row[w - 1], right);
+
+    dst += dst_stride;
+    ++y;
+
+    if (y > 0 && y < h)
+      ref_row += src_stride;
+  } while (--b_h);
+}
+
+#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;
+
+    if (left > b_w)
+      left = b_w;
+
+    if (x + b_w > w)
+      right = x + b_w - w;
+
+    if (right > b_w)
+      right = b_w;
+
+    copy = b_w - left - right;
+
+    if (left)
+      vpx_memset16(dst, ref_row[0], left);
+
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy * sizeof(uint16_t));
+
+    if (right)
+      vpx_memset16(dst + left + copy, ref_row[w - 1], right);
+
+    dst += dst_stride;
+    ++y;
+
+    if (y > 0 && y < h)
+      ref_row += src_stride;
+  } while (--b_h);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+#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;
+
+  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;
+  }
+
+  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;
+
+  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);
+}
+#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;
+
+  // 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;
+  }
+
+  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);
+
+    // 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);
+
+    // 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;
+
+  // 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;
+
+  // 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,
+#endif
+                         w, h, ref, xs, ys);
+      return;
+    }
+  } 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);
+  }
+#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];
+
+    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);
+      }
+    }
+  }
+}
diff --git a/media/libvpx/vp9/decoder/vp9_decodeframe.h b/media/libvpx/vp9/decoder/vp9_decodeframe.h
new file mode 100644
index 000000000..8410c541e
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_decodeframe.h
@@ -0,0 +1,41 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_DECODER_VP9_DECODEFRAME_H_
+#define VP9_DECODER_VP9_DECODEFRAME_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct VP9Common;
+struct VP9Decoder;
+struct vp9_read_bit_buffer;
+
+void vp9_init_dequantizer(struct VP9Common *cm);
+
+void vp9_decode_frame(struct VP9Decoder *pbi,
+                      const uint8_t *data, const uint8_t *data_end,
+                      const uint8_t **p_data_end);
+
+int vp9_read_sync_code(struct vp9_read_bit_buffer *const rb);
+void vp9_read_frame_size(struct vp9_read_bit_buffer *rb,
+                         int *width, int *height);
+BITSTREAM_PROFILE vp9_read_profile(struct vp9_read_bit_buffer *rb);
+
+void vp9_dec_build_inter_predictors_sb(struct VP9Decoder *const pbi,
+                                       MACROBLOCKD *xd, int mi_row, int mi_col,
+                                       BLOCK_SIZE bsize);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DECODEFRAME_H_
diff --git a/media/libvpx/vp9/decoder/vp9_decodemv.c b/media/libvpx/vp9/decoder/vp9_decodemv.c
new file mode 100644
index 000000000..7ce3389e8
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_decodemv.c
@@ -0,0 +1,614 @@
+/*
+  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_entropymv.h"
+#include "vp9/common/vp9_mvref_common.h"
+#include "vp9/common/vp9_pred_common.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/common/vp9_seg_common.h"
+
+#include "vp9/decoder/vp9_decodemv.h"
+#include "vp9/decoder/vp9_decodeframe.h"
+#include "vp9/decoder/vp9_reader.h"
+
+static PREDICTION_MODE read_intra_mode(vp9_reader *r, const vp9_prob *p) {
+  return (PREDICTION_MODE)vp9_read_tree(r, vp9_intra_mode_tree, p);
+}
+
+static PREDICTION_MODE read_intra_mode_y(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                         vp9_reader *r, int size_group) {
+  const PREDICTION_MODE y_mode =
+      read_intra_mode(r, cm->fc->y_mode_prob[size_group]);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts)
+    ++counts->y_mode[size_group][y_mode];
+  return y_mode;
+}
+
+static PREDICTION_MODE read_intra_mode_uv(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                          vp9_reader *r,
+                                          PREDICTION_MODE y_mode) {
+  const PREDICTION_MODE uv_mode = read_intra_mode(r,
+                                         cm->fc->uv_mode_prob[y_mode]);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts)
+    ++counts->uv_mode[y_mode][uv_mode];
+  return uv_mode;
+}
+
+static PREDICTION_MODE read_inter_mode(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                       vp9_reader *r, int ctx) {
+  const int mode = vp9_read_tree(r, vp9_inter_mode_tree,
+                                 cm->fc->inter_mode_probs[ctx]);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts)
+    ++counts->inter_mode[ctx][mode];
+
+  return NEARESTMV + mode;
+}
+
+static int read_segment_id(vp9_reader *r, const struct segmentation *seg) {
+  return vp9_read_tree(r, vp9_segment_tree, seg->tree_probs);
+}
+
+static TX_SIZE read_selected_tx_size(VP9_COMMON *cm, MACROBLOCKD *xd,
+                                     TX_SIZE max_tx_size, vp9_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+  const int ctx = vp9_get_tx_size_context(xd);
+  const vp9_prob *tx_probs = get_tx_probs(max_tx_size, ctx, &cm->fc->tx_probs);
+  int tx_size = vp9_read(r, tx_probs[0]);
+  if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
+    tx_size += vp9_read(r, tx_probs[1]);
+    if (tx_size != TX_8X8 && max_tx_size >= TX_32X32)
+      tx_size += vp9_read(r, tx_probs[2]);
+  }
+
+  if (counts)
+    ++get_tx_counts(max_tx_size, ctx, &counts->tx)[tx_size];
+  return (TX_SIZE)tx_size;
+}
+
+static TX_SIZE read_tx_size(VP9_COMMON *cm, MACROBLOCKD *xd,
+                            int allow_select, vp9_reader *r) {
+  TX_MODE tx_mode = cm->tx_mode;
+  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
+  if (allow_select && tx_mode == TX_MODE_SELECT && bsize >= BLOCK_8X8)
+    return read_selected_tx_size(cm, xd, max_tx_size, r);
+  else
+    return MIN(max_tx_size, tx_mode_to_biggest_tx_size[tx_mode]);
+}
+
+static void set_segment_id(VP9_COMMON *cm, BLOCK_SIZE bsize,
+                           int mi_row, int mi_col, int segment_id) {
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = num_8x8_blocks_wide_lookup[bsize];
+  const int bh = num_8x8_blocks_high_lookup[bsize];
+  const int xmis = MIN(cm->mi_cols - mi_col, bw);
+  const int ymis = MIN(cm->mi_rows - mi_row, bh);
+  int x, y;
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+
+  for (y = 0; y < ymis; y++)
+    for (x = 0; x < xmis; x++)
+      cm->current_frame_seg_map[mi_offset + y * cm->mi_cols + x] = segment_id;
+}
+
+static void copy_segment_id(const VP9_COMMON *cm,
+                           const uint8_t *last_segment_ids,
+                           uint8_t *current_segment_ids,
+                           BLOCK_SIZE bsize, int mi_row, int mi_col) {
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = num_8x8_blocks_wide_lookup[bsize];
+  const int bh = num_8x8_blocks_high_lookup[bsize];
+  const int xmis = MIN(cm->mi_cols - mi_col, bw);
+  const int ymis = MIN(cm->mi_rows - mi_row, bh);
+  int x, y;
+
+  for (y = 0; y < ymis; y++)
+    for (x = 0; x < xmis; x++)
+      current_segment_ids[mi_offset + y * cm->mi_cols + x] =  last_segment_ids ?
+          last_segment_ids[mi_offset + y * cm->mi_cols + x] : 0;
+}
+
+static int read_intra_segment_id(VP9_COMMON *const cm, BLOCK_SIZE bsize,
+                                 int mi_row, int mi_col,
+                                 vp9_reader *r) {
+  struct segmentation *const seg = &cm->seg;
+  int segment_id;
+
+  if (!seg->enabled)
+    return 0;  // Default for disabled segmentation
+
+  if (!seg->update_map) {
+    copy_segment_id(cm, cm->last_frame_seg_map, cm->current_frame_seg_map,
+                    bsize, mi_row, mi_col);
+    return 0;
+  }
+
+  segment_id = read_segment_id(r, seg);
+  set_segment_id(cm, bsize, mi_row, mi_col, segment_id);
+  return segment_id;
+}
+
+static int read_inter_segment_id(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                                 int mi_row, int mi_col, vp9_reader *r) {
+  struct segmentation *const seg = &cm->seg;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  int predicted_segment_id, segment_id;
+
+  if (!seg->enabled)
+    return 0;  // Default for disabled segmentation
+
+  predicted_segment_id = cm->last_frame_seg_map ?
+      vp9_get_segment_id(cm, cm->last_frame_seg_map, bsize, mi_row, mi_col) : 0;
+
+  if (!seg->update_map) {
+    copy_segment_id(cm, cm->last_frame_seg_map, cm->current_frame_seg_map,
+                    bsize, mi_row, mi_col);
+    return predicted_segment_id;
+  }
+
+  if (seg->temporal_update) {
+    const vp9_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
+    mbmi->seg_id_predicted = vp9_read(r, pred_prob);
+    segment_id = mbmi->seg_id_predicted ? predicted_segment_id
+                                        : read_segment_id(r, seg);
+  } else {
+    segment_id = read_segment_id(r, seg);
+  }
+  set_segment_id(cm, bsize, mi_row, mi_col, segment_id);
+  return segment_id;
+}
+
+static int read_skip(VP9_COMMON *cm, const MACROBLOCKD *xd,
+                     int segment_id, vp9_reader *r) {
+  if (vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
+    return 1;
+  } else {
+    const int ctx = vp9_get_skip_context(xd);
+    const int skip = vp9_read(r, cm->fc->skip_probs[ctx]);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts)
+      ++counts->skip[ctx][skip];
+    return skip;
+  }
+}
+
+static void read_intra_frame_mode_info(VP9_COMMON *const cm,
+                                       MACROBLOCKD *const xd,
+                                       int mi_row, int mi_col, vp9_reader *r) {
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const MODE_INFO *above_mi = xd->above_mi;
+  const MODE_INFO *left_mi  = xd->left_mi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  int i;
+
+  mbmi->segment_id = read_intra_segment_id(cm, bsize, mi_row, mi_col, r);
+  mbmi->skip = read_skip(cm, xd, mbmi->segment_id, r);
+  mbmi->tx_size = read_tx_size(cm, xd, 1, r);
+  mbmi->ref_frame[0] = INTRA_FRAME;
+  mbmi->ref_frame[1] = NONE;
+
+  switch (bsize) {
+    case BLOCK_4X4:
+      for (i = 0; i < 4; ++i)
+        mi->bmi[i].as_mode =
+            read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, i));
+      mbmi->mode = mi->bmi[3].as_mode;
+      break;
+    case BLOCK_4X8:
+      mi->bmi[0].as_mode = mi->bmi[2].as_mode =
+          read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 0));
+      mi->bmi[1].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+          read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 1));
+      break;
+    case BLOCK_8X4:
+      mi->bmi[0].as_mode = mi->bmi[1].as_mode =
+          read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 0));
+      mi->bmi[2].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+          read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 2));
+      break;
+    default:
+      mbmi->mode = read_intra_mode(r,
+                                   get_y_mode_probs(mi, above_mi, left_mi, 0));
+  }
+
+  mbmi->uv_mode = read_intra_mode(r, vp9_kf_uv_mode_prob[mbmi->mode]);
+}
+
+static int read_mv_component(vp9_reader *r,
+                             const nmv_component *mvcomp, int usehp) {
+  int mag, d, fr, hp;
+  const int sign = vp9_read(r, mvcomp->sign);
+  const int mv_class = vp9_read_tree(r, vp9_mv_class_tree, mvcomp->classes);
+  const int class0 = mv_class == MV_CLASS_0;
+
+  // Integer part
+  if (class0) {
+    d = vp9_read_tree(r, vp9_mv_class0_tree, mvcomp->class0);
+  } else {
+    int i;
+    const int n = mv_class + CLASS0_BITS - 1;  // number of bits
+
+    d = 0;
+    for (i = 0; i < n; ++i)
+      d |= vp9_read(r, mvcomp->bits[i]) << i;
+  }
+
+  // Fractional part
+  fr = vp9_read_tree(r, vp9_mv_fp_tree, class0 ? mvcomp->class0_fp[d]
+                                               : mvcomp->fp);
+
+  // High precision part (if hp is not used, the default value of the hp is 1)
+  hp = usehp ? vp9_read(r, class0 ? mvcomp->class0_hp : mvcomp->hp)
+             : 1;
+
+  // Result
+  mag = vp9_get_mv_mag(mv_class, (d << 3) | (fr << 1) | hp) + 1;
+  return sign ? -mag : mag;
+}
+
+static INLINE void read_mv(vp9_reader *r, MV *mv, const MV *ref,
+                           const nmv_context *ctx,
+                           nmv_context_counts *counts, int allow_hp) {
+  const MV_JOINT_TYPE joint_type =
+      (MV_JOINT_TYPE)vp9_read_tree(r, vp9_mv_joint_tree, ctx->joints);
+  const int use_hp = allow_hp && vp9_use_mv_hp(ref);
+  MV diff = {0, 0};
+
+  if (mv_joint_vertical(joint_type))
+    diff.row = read_mv_component(r, &ctx->comps[0], use_hp);
+
+  if (mv_joint_horizontal(joint_type))
+    diff.col = read_mv_component(r, &ctx->comps[1], use_hp);
+
+  vp9_inc_mv(&diff, counts);
+
+  mv->row = ref->row + diff.row;
+  mv->col = ref->col + diff.col;
+}
+
+static REFERENCE_MODE read_block_reference_mode(VP9_COMMON *cm,
+                                                const MACROBLOCKD *xd,
+                                                vp9_reader *r) {
+  if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+    const int ctx = vp9_get_reference_mode_context(cm, xd);
+    const REFERENCE_MODE mode =
+        (REFERENCE_MODE)vp9_read(r, cm->fc->comp_inter_prob[ctx]);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts)
+      ++counts->comp_inter[ctx][mode];
+    return mode;  // SINGLE_REFERENCE or COMPOUND_REFERENCE
+  } else {
+    return cm->reference_mode;
+  }
+}
+
+// Read the referncence frame
+static void read_ref_frames(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                            vp9_reader *r,
+                            int segment_id, MV_REFERENCE_FRAME ref_frame[2]) {
+  FRAME_CONTEXT *const fc = cm->fc;
+  FRAME_COUNTS *counts = xd->counts;
+
+  if (vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    ref_frame[0] = (MV_REFERENCE_FRAME)vp9_get_segdata(&cm->seg, segment_id,
+                                                       SEG_LVL_REF_FRAME);
+    ref_frame[1] = NONE;
+  } else {
+    const REFERENCE_MODE mode = read_block_reference_mode(cm, xd, r);
+    // FIXME(rbultje) I'm pretty sure this breaks segmentation ref frame coding
+    if (mode == COMPOUND_REFERENCE) {
+      const int idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
+      const int ctx = vp9_get_pred_context_comp_ref_p(cm, xd);
+      const int bit = vp9_read(r, fc->comp_ref_prob[ctx]);
+      if (counts)
+        ++counts->comp_ref[ctx][bit];
+      ref_frame[idx] = cm->comp_fixed_ref;
+      ref_frame[!idx] = cm->comp_var_ref[bit];
+    } else if (mode == SINGLE_REFERENCE) {
+      const int ctx0 = vp9_get_pred_context_single_ref_p1(xd);
+      const int bit0 = vp9_read(r, fc->single_ref_prob[ctx0][0]);
+      if (counts)
+        ++counts->single_ref[ctx0][0][bit0];
+      if (bit0) {
+        const int ctx1 = vp9_get_pred_context_single_ref_p2(xd);
+        const int bit1 = vp9_read(r, fc->single_ref_prob[ctx1][1]);
+        if (counts)
+          ++counts->single_ref[ctx1][1][bit1];
+        ref_frame[0] = bit1 ? ALTREF_FRAME : GOLDEN_FRAME;
+      } else {
+        ref_frame[0] = LAST_FRAME;
+      }
+
+      ref_frame[1] = NONE;
+    } else {
+      assert(0 && "Invalid prediction mode.");
+    }
+  }
+}
+
+
+static INLINE INTERP_FILTER read_switchable_interp_filter(
+    VP9_COMMON *const cm, MACROBLOCKD *const xd,
+    vp9_reader *r) {
+  const int ctx = vp9_get_pred_context_switchable_interp(xd);
+  const INTERP_FILTER type =
+      (INTERP_FILTER)vp9_read_tree(r, vp9_switchable_interp_tree,
+                                   cm->fc->switchable_interp_prob[ctx]);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts)
+    ++counts->switchable_interp[ctx][type];
+  return type;
+}
+
+static void read_intra_block_mode_info(VP9_COMMON *const cm,
+                                       MACROBLOCKD *const xd, MODE_INFO *mi,
+                                       vp9_reader *r) {
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const BLOCK_SIZE bsize = mi->mbmi.sb_type;
+  int i;
+
+  mbmi->ref_frame[0] = INTRA_FRAME;
+  mbmi->ref_frame[1] = NONE;
+
+  switch (bsize) {
+    case BLOCK_4X4:
+      for (i = 0; i < 4; ++i)
+        mi->bmi[i].as_mode = read_intra_mode_y(cm, xd, r, 0);
+      mbmi->mode = mi->bmi[3].as_mode;
+      break;
+    case BLOCK_4X8:
+      mi->bmi[0].as_mode = mi->bmi[2].as_mode = read_intra_mode_y(cm, xd,
+                                                                  r, 0);
+      mi->bmi[1].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+          read_intra_mode_y(cm, xd, r, 0);
+      break;
+    case BLOCK_8X4:
+      mi->bmi[0].as_mode = mi->bmi[1].as_mode = read_intra_mode_y(cm, xd,
+                                                                  r, 0);
+      mi->bmi[2].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+          read_intra_mode_y(cm, xd, r, 0);
+      break;
+    default:
+      mbmi->mode = read_intra_mode_y(cm, xd, r, size_group_lookup[bsize]);
+  }
+
+  mbmi->uv_mode = read_intra_mode_uv(cm, xd, r, mbmi->mode);
+}
+
+static INLINE int is_mv_valid(const MV *mv) {
+  return mv->row > MV_LOW && mv->row < MV_UPP &&
+         mv->col > MV_LOW && mv->col < MV_UPP;
+}
+
+static INLINE int assign_mv(VP9_COMMON *cm, MACROBLOCKD *xd,
+                            PREDICTION_MODE mode,
+                            int_mv mv[2], int_mv ref_mv[2],
+                            int_mv nearest_mv[2], int_mv near_mv[2],
+                            int is_compound, int allow_hp, vp9_reader *r) {
+  int i;
+  int ret = 1;
+
+  switch (mode) {
+    case NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+      for (i = 0; i < 1 + is_compound; ++i) {
+        read_mv(r, &mv[i].as_mv, &ref_mv[i].as_mv, &cm->fc->nmvc, mv_counts,
+                allow_hp);
+        ret = ret && is_mv_valid(&mv[i].as_mv);
+      }
+      break;
+    }
+    case NEARESTMV: {
+      mv[0].as_int = nearest_mv[0].as_int;
+      if (is_compound)
+        mv[1].as_int = nearest_mv[1].as_int;
+      break;
+    }
+    case NEARMV: {
+      mv[0].as_int = near_mv[0].as_int;
+      if (is_compound)
+        mv[1].as_int = near_mv[1].as_int;
+      break;
+    }
+    case ZEROMV: {
+      mv[0].as_int = 0;
+      if (is_compound)
+        mv[1].as_int = 0;
+      break;
+    }
+    default: {
+      return 0;
+    }
+  }
+  return ret;
+}
+
+static int read_is_inter_block(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+                               int segment_id, vp9_reader *r) {
+  if (vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    return vp9_get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME) !=
+           INTRA_FRAME;
+  } else {
+    const int ctx = vp9_get_intra_inter_context(xd);
+    const int is_inter = vp9_read(r, cm->fc->intra_inter_prob[ctx]);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts)
+      ++counts->intra_inter[ctx][is_inter];
+    return is_inter;
+  }
+}
+
+static void fpm_sync(void *const data, int mi_row) {
+  VP9Decoder *const pbi = (VP9Decoder *)data;
+  vp9_frameworker_wait(pbi->frame_worker_owner, pbi->common.prev_frame,
+                       mi_row << MI_BLOCK_SIZE_LOG2);
+}
+
+static void read_inter_block_mode_info(VP9Decoder *const pbi,
+                                       MACROBLOCKD *const xd,
+                                       const TileInfo *const tile,
+                                       MODE_INFO *const mi,
+                                       int mi_row, int mi_col, vp9_reader *r) {
+  VP9_COMMON *const cm = &pbi->common;
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  const int allow_hp = cm->allow_high_precision_mv;
+  int_mv nearestmv[2], nearmv[2];
+  int inter_mode_ctx, ref, is_compound;
+
+  read_ref_frames(cm, xd, r, mbmi->segment_id, mbmi->ref_frame);
+  is_compound = has_second_ref(mbmi);
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
+    RefBuffer *ref_buf = &cm->frame_refs[frame - LAST_FRAME];
+    xd->block_refs[ref] = ref_buf;
+    if ((!vp9_is_valid_scale(&ref_buf->sf)))
+      vpx_internal_error(xd->error_info, VPX_CODEC_UNSUP_BITSTREAM,
+                         "Reference frame has invalid dimensions");
+    vp9_setup_pre_planes(xd, ref, ref_buf->buf, mi_row, mi_col,
+                         &ref_buf->sf);
+    vp9_find_mv_refs(cm, xd, tile, mi, frame, mbmi->ref_mvs[frame],
+                     mi_row, mi_col, fpm_sync, (void *)pbi);
+  }
+
+  inter_mode_ctx = mbmi->mode_context[mbmi->ref_frame[0]];
+
+  if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+    mbmi->mode = ZEROMV;
+    if (bsize < BLOCK_8X8) {
+        vpx_internal_error(xd->error_info, VPX_CODEC_UNSUP_BITSTREAM,
+                           "Invalid usage of segement feature on small blocks");
+        return;
+    }
+  } else {
+    if (bsize >= BLOCK_8X8)
+      mbmi->mode = read_inter_mode(cm, xd, r, inter_mode_ctx);
+  }
+
+  if (bsize < BLOCK_8X8 || mbmi->mode != ZEROMV) {
+    for (ref = 0; ref < 1 + is_compound; ++ref) {
+      vp9_find_best_ref_mvs(xd, allow_hp, mbmi->ref_mvs[mbmi->ref_frame[ref]],
+                            &nearestmv[ref], &nearmv[ref]);
+    }
+  }
+
+  mbmi->interp_filter = (cm->interp_filter == SWITCHABLE)
+                      ? read_switchable_interp_filter(cm, xd, r)
+                      : cm->interp_filter;
+
+  if (bsize < BLOCK_8X8) {
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];  // 1 or 2
+    const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];  // 1 or 2
+    int idx, idy;
+    PREDICTION_MODE b_mode;
+    int_mv nearest_sub8x8[2], near_sub8x8[2];
+    for (idy = 0; idy < 2; idy += num_4x4_h) {
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        int_mv block[2];
+        const int j = idy * 2 + idx;
+        b_mode = read_inter_mode(cm, xd, r, inter_mode_ctx);
+
+        if (b_mode == NEARESTMV || b_mode == NEARMV)
+          for (ref = 0; ref < 1 + is_compound; ++ref)
+            vp9_append_sub8x8_mvs_for_idx(cm, xd, tile, j, ref, mi_row, mi_col,
+                                          &nearest_sub8x8[ref],
+                                          &near_sub8x8[ref]);
+
+        if (!assign_mv(cm, xd, b_mode, block, nearestmv,
+                       nearest_sub8x8, near_sub8x8,
+                       is_compound, allow_hp, r)) {
+          xd->corrupted |= 1;
+          break;
+        };
+
+        mi->bmi[j].as_mv[0].as_int = block[0].as_int;
+        if (is_compound)
+          mi->bmi[j].as_mv[1].as_int = block[1].as_int;
+
+        if (num_4x4_h == 2)
+          mi->bmi[j + 2] = mi->bmi[j];
+        if (num_4x4_w == 2)
+          mi->bmi[j + 1] = mi->bmi[j];
+      }
+    }
+
+    mi->mbmi.mode = b_mode;
+
+    mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
+    mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
+  } else {
+    xd->corrupted |= !assign_mv(cm, xd, mbmi->mode, mbmi->mv, nearestmv,
+                                nearestmv, nearmv, is_compound, allow_hp, r);
+  }
+}
+
+static void read_inter_frame_mode_info(VP9Decoder *const pbi,
+                                       MACROBLOCKD *const xd,
+                                       const TileInfo *const tile,
+                                       int mi_row, int mi_col, vp9_reader *r) {
+  VP9_COMMON *const cm = &pbi->common;
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  int inter_block;
+
+  mbmi->mv[0].as_int = 0;
+  mbmi->mv[1].as_int = 0;
+  mbmi->segment_id = read_inter_segment_id(cm, xd, mi_row, mi_col, r);
+  mbmi->skip = read_skip(cm, xd, mbmi->segment_id, r);
+  inter_block = read_is_inter_block(cm, xd, mbmi->segment_id, r);
+  mbmi->tx_size = read_tx_size(cm, xd, !mbmi->skip || !inter_block, r);
+
+  if (inter_block)
+    read_inter_block_mode_info(pbi, xd, tile, mi, mi_row, mi_col, r);
+  else
+    read_intra_block_mode_info(cm, xd, mi, r);
+}
+
+void vp9_read_mode_info(VP9Decoder *const pbi, MACROBLOCKD *xd,
+                        const TileInfo *const tile,
+                        int mi_row, int mi_col, vp9_reader *r) {
+  VP9_COMMON *const cm = &pbi->common;
+  MODE_INFO *const mi = xd->mi[0];
+  const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
+  const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
+  const int x_mis = MIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = MIN(bh, cm->mi_rows - mi_row);
+  MV_REF* frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
+  int w, h;
+
+  if (frame_is_intra_only(cm))
+    read_intra_frame_mode_info(cm, xd, mi_row, mi_col, r);
+  else
+    read_inter_frame_mode_info(pbi, xd, tile, mi_row, mi_col, r);
+
+  for (h = 0; h < y_mis; ++h) {
+    MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
+    for (w = 0; w < x_mis; ++w) {
+      MV_REF *const mv = frame_mv + w;
+      mv->ref_frame[0] = mi->mbmi.ref_frame[0];
+      mv->ref_frame[1] = mi->mbmi.ref_frame[1];
+      mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
+      mv->mv[1].as_int = mi->mbmi.mv[1].as_int;
+    }
+  }
+}
diff --git a/media/libvpx/vp9/decoder/vp9_decodemv.h b/media/libvpx/vp9/decoder/vp9_decodemv.h
new file mode 100644
index 000000000..dd97d8da0
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_decodemv.h
@@ -0,0 +1,31 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_DECODER_VP9_DECODEMV_H_
+#define VP9_DECODER_VP9_DECODEMV_H_
+
+#include "vp9/decoder/vp9_decoder.h"
+#include "vp9/decoder/vp9_reader.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct TileInfo;
+
+void vp9_read_mode_info(VP9Decoder *const pbi, MACROBLOCKD *xd,
+                        const struct TileInfo *const tile,
+                        int mi_row, int mi_col, vp9_reader *r);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DECODEMV_H_
diff --git a/media/libvpx/vp9/decoder/vp9_decoder.c b/media/libvpx/vp9/decoder/vp9_decoder.c
new file mode 100644
index 000000000..cc60c596f
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_decoder.c
@@ -0,0 +1,509 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <stdio.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_scale_rtcd.h"
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/vpx_once.h"
+#include "vpx_ports/vpx_timer.h"
+#include "vpx_scale/vpx_scale.h"
+
+#include "vp9/common/vp9_alloccommon.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#if CONFIG_VP9_POSTPROC
+#include "vp9/common/vp9_postproc.h"
+#endif
+#include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_reconintra.h"
+#include "vp9/common/vp9_systemdependent.h"
+#include "vp9/common/vp9_thread.h"
+
+#include "vp9/decoder/vp9_decodeframe.h"
+#include "vp9/decoder/vp9_decoder.h"
+#include "vp9/decoder/vp9_detokenize.h"
+
+static void initialize_dec(void) {
+  static volatile int init_done = 0;
+
+  if (!init_done) {
+    vp9_rtcd();
+    vpx_dsp_rtcd();
+    vpx_scale_rtcd();
+    vp9_init_intra_predictors();
+    init_done = 1;
+  }
+}
+
+static void vp9_dec_setup_mi(VP9_COMMON *cm) {
+  cm->mi = cm->mip + cm->mi_stride + 1;
+  cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
+  memset(cm->mi_grid_base, 0,
+         cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
+}
+
+static int vp9_dec_alloc_mi(VP9_COMMON *cm, int mi_size) {
+  cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip));
+  if (!cm->mip)
+    return 1;
+  cm->mi_alloc_size = mi_size;
+  cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*));
+  if (!cm->mi_grid_base)
+    return 1;
+  return 0;
+}
+
+static void vp9_dec_free_mi(VP9_COMMON *cm) {
+  vpx_free(cm->mip);
+  cm->mip = NULL;
+  vpx_free(cm->mi_grid_base);
+  cm->mi_grid_base = NULL;
+}
+
+VP9Decoder *vp9_decoder_create(BufferPool *const pool) {
+  VP9Decoder *volatile const pbi = vpx_memalign(32, sizeof(*pbi));
+  VP9_COMMON *volatile const cm = pbi ? &pbi->common : NULL;
+
+  if (!cm)
+    return NULL;
+
+  vp9_zero(*pbi);
+
+  if (setjmp(cm->error.jmp)) {
+    cm->error.setjmp = 0;
+    vp9_decoder_remove(pbi);
+    return NULL;
+  }
+
+  cm->error.setjmp = 1;
+
+  CHECK_MEM_ERROR(cm, cm->fc,
+                  (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
+  CHECK_MEM_ERROR(cm, cm->frame_contexts,
+                  (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS,
+                  sizeof(*cm->frame_contexts)));
+
+  pbi->need_resync = 1;
+  once(initialize_dec);
+
+  // Initialize the references to not point to any frame buffers.
+  memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+  memset(&cm->next_ref_frame_map, -1, sizeof(cm->next_ref_frame_map));
+
+  cm->current_video_frame = 0;
+  pbi->ready_for_new_data = 1;
+  pbi->common.buffer_pool = pool;
+
+  cm->bit_depth = VPX_BITS_8;
+  cm->dequant_bit_depth = VPX_BITS_8;
+
+  cm->alloc_mi = vp9_dec_alloc_mi;
+  cm->free_mi = vp9_dec_free_mi;
+  cm->setup_mi = vp9_dec_setup_mi;
+
+  vp9_loop_filter_init(cm);
+
+  cm->error.setjmp = 0;
+
+  vp9_get_worker_interface()->init(&pbi->lf_worker);
+
+  return pbi;
+}
+
+void vp9_decoder_remove(VP9Decoder *pbi) {
+  int i;
+
+  vp9_get_worker_interface()->end(&pbi->lf_worker);
+  vpx_free(pbi->lf_worker.data1);
+  vpx_free(pbi->tile_data);
+  for (i = 0; i < pbi->num_tile_workers; ++i) {
+    VP9Worker *const worker = &pbi->tile_workers[i];
+    vp9_get_worker_interface()->end(worker);
+  }
+  vpx_free(pbi->tile_worker_data);
+  vpx_free(pbi->tile_worker_info);
+  vpx_free(pbi->tile_workers);
+
+  if (pbi->num_tile_workers > 0) {
+    vp9_loop_filter_dealloc(&pbi->lf_row_sync);
+  }
+
+  vpx_free(pbi);
+}
+
+static int equal_dimensions(const YV12_BUFFER_CONFIG *a,
+                            const YV12_BUFFER_CONFIG *b) {
+    return a->y_height == b->y_height && a->y_width == b->y_width &&
+           a->uv_height == b->uv_height && a->uv_width == b->uv_width;
+}
+
+vpx_codec_err_t vp9_copy_reference_dec(VP9Decoder *pbi,
+                                       VP9_REFFRAME ref_frame_flag,
+                                       YV12_BUFFER_CONFIG *sd) {
+  VP9_COMMON *cm = &pbi->common;
+
+  /* TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
+   * encoder is using the frame buffers for. This is just a stub to keep the
+   * vpxenc --test-decode functionality working, and will be replaced in a
+   * later commit that adds VP9-specific controls for this functionality.
+   */
+  if (ref_frame_flag == VP9_LAST_FLAG) {
+    const YV12_BUFFER_CONFIG *const cfg = get_ref_frame(cm, 0);
+    if (cfg == NULL) {
+      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                         "No 'last' reference frame");
+      return VPX_CODEC_ERROR;
+    }
+    if (!equal_dimensions(cfg, sd))
+      vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                         "Incorrect buffer dimensions");
+    else
+      vp8_yv12_copy_frame(cfg, sd);
+  } else {
+    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                       "Invalid reference frame");
+  }
+
+  return cm->error.error_code;
+}
+
+
+vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
+                                      VP9_REFFRAME ref_frame_flag,
+                                      YV12_BUFFER_CONFIG *sd) {
+  RefBuffer *ref_buf = NULL;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+
+  // TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
+  // encoder is using the frame buffers for. This is just a stub to keep the
+  // vpxenc --test-decode functionality working, and will be replaced in a
+  // later commit that adds VP9-specific controls for this functionality.
+  if (ref_frame_flag == VP9_LAST_FLAG) {
+    ref_buf = &cm->frame_refs[0];
+  } else if (ref_frame_flag == VP9_GOLD_FLAG) {
+    ref_buf = &cm->frame_refs[1];
+  } else if (ref_frame_flag == VP9_ALT_FLAG) {
+    ref_buf = &cm->frame_refs[2];
+  } else {
+    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                       "Invalid reference frame");
+    return cm->error.error_code;
+  }
+
+  if (!equal_dimensions(ref_buf->buf, sd)) {
+    vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
+                       "Incorrect buffer dimensions");
+  } else {
+    int *ref_fb_ptr = &ref_buf->idx;
+
+    // Find an empty frame buffer.
+    const int free_fb = get_free_fb(cm);
+    if (cm->new_fb_idx == INVALID_IDX)
+      return VPX_CODEC_MEM_ERROR;
+
+    // Decrease ref_count since it will be increased again in
+    // ref_cnt_fb() below.
+    --frame_bufs[free_fb].ref_count;
+
+    // Manage the reference counters and copy image.
+    ref_cnt_fb(frame_bufs, ref_fb_ptr, free_fb);
+    ref_buf->buf = &frame_bufs[*ref_fb_ptr].buf;
+    vp8_yv12_copy_frame(sd, ref_buf->buf);
+  }
+
+  return cm->error.error_code;
+}
+
+/* If any buffer updating is signaled it should be done here. */
+static void swap_frame_buffers(VP9Decoder *pbi) {
+  int ref_index = 0, mask;
+  VP9_COMMON *const cm = &pbi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+
+  lock_buffer_pool(pool);
+  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+    const int old_idx = cm->ref_frame_map[ref_index];
+    // Current thread releases the holding of reference frame.
+    decrease_ref_count(old_idx, frame_bufs, pool);
+
+    // Release the reference frame in reference map.
+    if ((mask & 1) && old_idx >= 0) {
+      decrease_ref_count(old_idx, frame_bufs, pool);
+    }
+    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
+    ++ref_index;
+  }
+
+  // Current thread releases the holding of reference frame.
+  for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
+    const int old_idx = cm->ref_frame_map[ref_index];
+    decrease_ref_count(old_idx, frame_bufs, pool);
+    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
+  }
+  unlock_buffer_pool(pool);
+  pbi->hold_ref_buf = 0;
+  cm->frame_to_show = get_frame_new_buffer(cm);
+
+  if (!pbi->frame_parallel_decode || !cm->show_frame) {
+    lock_buffer_pool(pool);
+    --frame_bufs[cm->new_fb_idx].ref_count;
+    unlock_buffer_pool(pool);
+  }
+
+  // Invalidate these references until the next frame starts.
+  for (ref_index = 0; ref_index < 3; ref_index++)
+    cm->frame_refs[ref_index].idx = -1;
+}
+
+int vp9_receive_compressed_data(VP9Decoder *pbi,
+                                size_t size, const uint8_t **psource) {
+  VP9_COMMON *volatile const cm = &pbi->common;
+  BufferPool *volatile const pool = cm->buffer_pool;
+  RefCntBuffer *volatile const frame_bufs = cm->buffer_pool->frame_bufs;
+  const uint8_t *source = *psource;
+  int retcode = 0;
+  cm->error.error_code = VPX_CODEC_OK;
+
+  if (size == 0) {
+    // This is used to signal that we are missing frames.
+    // We do not know if the missing frame(s) was supposed to update
+    // any of the reference buffers, but we act conservative and
+    // mark only the last buffer as corrupted.
+    //
+    // TODO(jkoleszar): Error concealment is undefined and non-normative
+    // at this point, but if it becomes so, [0] may not always be the correct
+    // thing to do here.
+    if (cm->frame_refs[0].idx > 0) {
+      assert(cm->frame_refs[0].buf != NULL);
+      cm->frame_refs[0].buf->corrupted = 1;
+    }
+  }
+
+  pbi->ready_for_new_data = 0;
+
+  // Check if the previous frame was a frame without any references to it.
+  // Release frame buffer if not decoding in frame parallel mode.
+  if (!pbi->frame_parallel_decode && cm->new_fb_idx >= 0
+      && frame_bufs[cm->new_fb_idx].ref_count == 0)
+    pool->release_fb_cb(pool->cb_priv,
+                        &frame_bufs[cm->new_fb_idx].raw_frame_buffer);
+  // Find a free frame buffer. Return error if can not find any.
+  cm->new_fb_idx = get_free_fb(cm);
+  if (cm->new_fb_idx == INVALID_IDX)
+    return VPX_CODEC_MEM_ERROR;
+
+  // Assign a MV array to the frame buffer.
+  cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
+
+  pbi->hold_ref_buf = 0;
+  if (pbi->frame_parallel_decode) {
+    VP9Worker *const worker = pbi->frame_worker_owner;
+    vp9_frameworker_lock_stats(worker);
+    frame_bufs[cm->new_fb_idx].frame_worker_owner = worker;
+    // Reset decoding progress.
+    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+    vp9_frameworker_unlock_stats(worker);
+  } else {
+    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
+  }
+
+
+  if (setjmp(cm->error.jmp)) {
+    const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+    int i;
+
+    cm->error.setjmp = 0;
+    pbi->ready_for_new_data = 1;
+
+    // Synchronize all threads immediately as a subsequent decode call may
+    // cause a resize invalidating some allocations.
+    winterface->sync(&pbi->lf_worker);
+    for (i = 0; i < pbi->num_tile_workers; ++i) {
+      winterface->sync(&pbi->tile_workers[i]);
+    }
+
+    lock_buffer_pool(pool);
+    // Release all the reference buffers if worker thread is holding them.
+    if (pbi->hold_ref_buf == 1) {
+      int ref_index = 0, mask;
+      for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+        const int old_idx = cm->ref_frame_map[ref_index];
+        // Current thread releases the holding of reference frame.
+        decrease_ref_count(old_idx, frame_bufs, pool);
+
+        // Release the reference frame in reference map.
+        if ((mask & 1) && old_idx >= 0) {
+          decrease_ref_count(old_idx, frame_bufs, pool);
+        }
+        ++ref_index;
+      }
+
+      // Current thread releases the holding of reference frame.
+      for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
+        const int old_idx = cm->ref_frame_map[ref_index];
+        decrease_ref_count(old_idx, frame_bufs, pool);
+      }
+      pbi->hold_ref_buf = 0;
+    }
+    // Release current frame.
+    decrease_ref_count(cm->new_fb_idx, frame_bufs, pool);
+    unlock_buffer_pool(pool);
+
+    vp9_clear_system_state();
+    return -1;
+  }
+
+  cm->error.setjmp = 1;
+  vp9_decode_frame(pbi, source, source + size, psource);
+
+  swap_frame_buffers(pbi);
+
+  vp9_clear_system_state();
+
+  if (!cm->show_existing_frame) {
+    cm->last_show_frame = cm->show_frame;
+    cm->prev_frame = cm->cur_frame;
+    if (cm->seg.enabled && !pbi->frame_parallel_decode)
+      vp9_swap_current_and_last_seg_map(cm);
+  }
+
+  // Update progress in frame parallel decode.
+  if (pbi->frame_parallel_decode) {
+    // Need to lock the mutex here as another thread may
+    // be accessing this buffer.
+    VP9Worker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    vp9_frameworker_lock_stats(worker);
+
+    if (cm->show_frame) {
+      cm->current_video_frame++;
+    }
+    frame_worker_data->frame_decoded = 1;
+    frame_worker_data->frame_context_ready = 1;
+    vp9_frameworker_signal_stats(worker);
+    vp9_frameworker_unlock_stats(worker);
+  } else {
+    cm->last_width = cm->width;
+    cm->last_height = cm->height;
+    if (cm->show_frame) {
+      cm->current_video_frame++;
+    }
+  }
+
+  cm->error.setjmp = 0;
+  return retcode;
+}
+
+int vp9_get_raw_frame(VP9Decoder *pbi, YV12_BUFFER_CONFIG *sd,
+                      vp9_ppflags_t *flags) {
+  VP9_COMMON *const cm = &pbi->common;
+  int ret = -1;
+#if !CONFIG_VP9_POSTPROC
+  (void)*flags;
+#endif
+
+  if (pbi->ready_for_new_data == 1)
+    return ret;
+
+  pbi->ready_for_new_data = 1;
+
+  /* no raw frame to show!!! */
+  if (!cm->show_frame)
+    return ret;
+
+  pbi->ready_for_new_data = 1;
+
+#if CONFIG_VP9_POSTPROC
+  if (!cm->show_existing_frame) {
+    ret = vp9_post_proc_frame(cm, sd, flags);
+  } else {
+    *sd = *cm->frame_to_show;
+    ret = 0;
+  }
+#else
+  *sd = *cm->frame_to_show;
+  ret = 0;
+#endif /*!CONFIG_POSTPROC*/
+  vp9_clear_system_state();
+  return ret;
+}
+
+vpx_codec_err_t vp9_parse_superframe_index(const uint8_t *data,
+                                           size_t data_sz,
+                                           uint32_t sizes[8], int *count,
+                                           vpx_decrypt_cb decrypt_cb,
+                                           void *decrypt_state) {
+  // A chunk ending with a byte matching 0xc0 is an invalid chunk unless
+  // it is a super frame index. If the last byte of real video compression
+  // data is 0xc0 the encoder must add a 0 byte. If we have the marker but
+  // not the associated matching marker byte at the front of the index we have
+  // an invalid bitstream and need to return an error.
+
+  uint8_t marker;
+
+  assert(data_sz);
+  marker = read_marker(decrypt_cb, decrypt_state, data + data_sz - 1);
+  *count = 0;
+
+  if ((marker & 0xe0) == 0xc0) {
+    const uint32_t frames = (marker & 0x7) + 1;
+    const uint32_t mag = ((marker >> 3) & 0x3) + 1;
+    const size_t index_sz = 2 + mag * frames;
+
+    // This chunk is marked as having a superframe index but doesn't have
+    // enough data for it, thus it's an invalid superframe index.
+    if (data_sz < index_sz)
+      return VPX_CODEC_CORRUPT_FRAME;
+
+    {
+      const uint8_t marker2 = read_marker(decrypt_cb, decrypt_state,
+                                          data + data_sz - index_sz);
+
+      // This chunk is marked as having a superframe index but doesn't have
+      // the matching marker byte at the front of the index therefore it's an
+      // invalid chunk.
+      if (marker != marker2)
+        return VPX_CODEC_CORRUPT_FRAME;
+    }
+
+    {
+      // Found a valid superframe index.
+      uint32_t i, j;
+      const uint8_t *x = &data[data_sz - index_sz + 1];
+
+      // Frames has a maximum of 8 and mag has a maximum of 4.
+      uint8_t clear_buffer[32];
+      assert(sizeof(clear_buffer) >= frames * mag);
+      if (decrypt_cb) {
+        decrypt_cb(decrypt_state, x, clear_buffer, frames * mag);
+        x = clear_buffer;
+      }
+
+      for (i = 0; i < frames; ++i) {
+        uint32_t this_sz = 0;
+
+        for (j = 0; j < mag; ++j)
+          this_sz |= (uint32_t)(*x++) << (j * 8);
+        sizes[i] = this_sz;
+      }
+      *count = frames;
+    }
+  }
+  return VPX_CODEC_OK;
+}
diff --git a/media/libvpx/vp9/decoder/vp9_decoder.h b/media/libvpx/vp9/decoder/vp9_decoder.h
new file mode 100644
index 000000000..c19f0ac3b
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_decoder.h
@@ -0,0 +1,136 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_DECODER_VP9_DECODER_H_
+#define VP9_DECODER_VP9_DECODER_H_
+
+#include "./vpx_config.h"
+
+#include "vpx/vpx_codec.h"
+#include "vpx_scale/yv12config.h"
+#include "vp9/common/vp9_thread_common.h"
+#include "vp9/common/vp9_onyxc_int.h"
+#include "vp9/common/vp9_ppflags.h"
+#include "vp9/common/vp9_thread.h"
+#include "vp9/decoder/vp9_dthread.h"
+#include "vp9/decoder/vp9_reader.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// TODO(hkuang): combine this with TileWorkerData.
+typedef struct TileData {
+  VP9_COMMON *cm;
+  vp9_reader bit_reader;
+  DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+} TileData;
+
+typedef struct TileWorkerData {
+  struct VP9Decoder *pbi;
+  vp9_reader bit_reader;
+  FRAME_COUNTS counts;
+  DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+  struct vpx_internal_error_info error_info;
+} TileWorkerData;
+
+typedef struct VP9Decoder {
+  DECLARE_ALIGNED(16, MACROBLOCKD, mb);
+
+  DECLARE_ALIGNED(16, VP9_COMMON, common);
+
+  int ready_for_new_data;
+
+  int refresh_frame_flags;
+
+  int frame_parallel_decode;  // frame-based threading.
+
+  // TODO(hkuang): Combine this with cur_buf in macroblockd as they are
+  // the same.
+  RefCntBuffer *cur_buf;   //  Current decoding frame buffer.
+
+  VP9Worker *frame_worker_owner;   // frame_worker that owns this pbi.
+  VP9Worker lf_worker;
+  VP9Worker *tile_workers;
+  TileWorkerData *tile_worker_data;
+  TileInfo *tile_worker_info;
+  int num_tile_workers;
+
+  TileData *tile_data;
+  int total_tiles;
+
+  VP9LfSync lf_row_sync;
+
+  vpx_decrypt_cb decrypt_cb;
+  void *decrypt_state;
+
+  int max_threads;
+  int inv_tile_order;
+  int need_resync;  // wait for key/intra-only frame.
+  int hold_ref_buf;  // hold the reference buffer.
+} VP9Decoder;
+
+int vp9_receive_compressed_data(struct VP9Decoder *pbi,
+                                size_t size, const uint8_t **dest);
+
+int vp9_get_raw_frame(struct VP9Decoder *pbi, YV12_BUFFER_CONFIG *sd,
+                      vp9_ppflags_t *flags);
+
+vpx_codec_err_t vp9_copy_reference_dec(struct VP9Decoder *pbi,
+                                       VP9_REFFRAME ref_frame_flag,
+                                       YV12_BUFFER_CONFIG *sd);
+
+vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
+                                      VP9_REFFRAME ref_frame_flag,
+                                      YV12_BUFFER_CONFIG *sd);
+
+static INLINE uint8_t read_marker(vpx_decrypt_cb decrypt_cb,
+                                  void *decrypt_state,
+                                  const uint8_t *data) {
+  if (decrypt_cb) {
+    uint8_t marker;
+    decrypt_cb(decrypt_state, data, &marker, 1);
+    return marker;
+  }
+  return *data;
+}
+
+// This function is exposed for use in tests, as well as the inlined function
+// "read_marker".
+vpx_codec_err_t vp9_parse_superframe_index(const uint8_t *data,
+                                           size_t data_sz,
+                                           uint32_t sizes[8], int *count,
+                                           vpx_decrypt_cb decrypt_cb,
+                                           void *decrypt_state);
+
+struct VP9Decoder *vp9_decoder_create(BufferPool *const pool);
+
+void vp9_decoder_remove(struct VP9Decoder *pbi);
+
+static INLINE void decrease_ref_count(int idx, RefCntBuffer *const frame_bufs,
+                                      BufferPool *const pool) {
+  if (idx >= 0) {
+    --frame_bufs[idx].ref_count;
+    // A worker may only get a free framebuffer index when calling get_free_fb.
+    // But the private buffer is not set up until finish decoding header.
+    // So any error happens during decoding header, the frame_bufs will not
+    // have valid priv buffer.
+    if (frame_bufs[idx].ref_count == 0 &&
+        frame_bufs[idx].raw_frame_buffer.priv) {
+      pool->release_fb_cb(pool->cb_priv, &frame_bufs[idx].raw_frame_buffer);
+    }
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DECODER_H_
diff --git a/media/libvpx/vp9/decoder/vp9_detokenize.c b/media/libvpx/vp9/decoder/vp9_detokenize.c
new file mode 100644
index 000000000..3304e64b2
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_detokenize.c
@@ -0,0 +1,226 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_mem/vpx_mem.h"
+#include "vpx_ports/mem.h"
+
+#include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/vp9_entropy.h"
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+#include "vp9/common/vp9_idct.h"
+#endif
+#include "vp9/common/vp9_scan.h"
+
+#include "vp9/decoder/vp9_detokenize.h"
+
+#define EOB_CONTEXT_NODE            0
+#define ZERO_CONTEXT_NODE           1
+#define ONE_CONTEXT_NODE            2
+#define LOW_VAL_CONTEXT_NODE        0
+#define TWO_CONTEXT_NODE            1
+#define THREE_CONTEXT_NODE          2
+#define HIGH_LOW_CONTEXT_NODE       3
+#define CAT_ONE_CONTEXT_NODE        4
+#define CAT_THREEFOUR_CONTEXT_NODE  5
+#define CAT_THREE_CONTEXT_NODE      6
+#define CAT_FIVE_CONTEXT_NODE       7
+
+#define INCREMENT_COUNT(token)                              \
+  do {                                                      \
+     if (counts)                                            \
+       ++coef_counts[band][ctx][token];                     \
+  } while (0)
+
+static INLINE int read_coeff(const vp9_prob *probs, int n, vp9_reader *r) {
+  int i, val = 0;
+  for (i = 0; i < n; ++i)
+    val = (val << 1) | vp9_read(r, probs[i]);
+  return val;
+}
+
+static int decode_coefs(const MACROBLOCKD *xd,
+                        PLANE_TYPE type,
+                        tran_low_t *dqcoeff, TX_SIZE tx_size, const int16_t *dq,
+                        int ctx, const int16_t *scan, const int16_t *nb,
+                        vp9_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+  const int max_eob = 16 << (tx_size << 1);
+  const FRAME_CONTEXT *const fc = xd->fc;
+  const int ref = is_inter_block(&xd->mi[0]->mbmi);
+  int band, c = 0;
+  const vp9_prob (*coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      fc->coef_probs[tx_size][type][ref];
+  const vp9_prob *prob;
+  unsigned int (*coef_counts)[COEFF_CONTEXTS][UNCONSTRAINED_NODES + 1];
+  unsigned int (*eob_branch_count)[COEFF_CONTEXTS];
+  uint8_t token_cache[32 * 32];
+  const uint8_t *band_translate = get_band_translate(tx_size);
+  const int dq_shift = (tx_size == TX_32X32);
+  int v, token;
+  int16_t dqv = dq[0];
+  const uint8_t *cat1_prob;
+  const uint8_t *cat2_prob;
+  const uint8_t *cat3_prob;
+  const uint8_t *cat4_prob;
+  const uint8_t *cat5_prob;
+  const uint8_t *cat6_prob;
+
+  if (counts) {
+    coef_counts = counts->coef[tx_size][type][ref];
+    eob_branch_count = counts->eob_branch[tx_size][type][ref];
+  }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (xd->bd > VPX_BITS_8) {
+    if (xd->bd == VPX_BITS_10) {
+      cat1_prob = vp9_cat1_prob_high10;
+      cat2_prob = vp9_cat2_prob_high10;
+      cat3_prob = vp9_cat3_prob_high10;
+      cat4_prob = vp9_cat4_prob_high10;
+      cat5_prob = vp9_cat5_prob_high10;
+      cat6_prob = vp9_cat6_prob_high10;
+    } else {
+      cat1_prob = vp9_cat1_prob_high12;
+      cat2_prob = vp9_cat2_prob_high12;
+      cat3_prob = vp9_cat3_prob_high12;
+      cat4_prob = vp9_cat4_prob_high12;
+      cat5_prob = vp9_cat5_prob_high12;
+      cat6_prob = vp9_cat6_prob_high12;
+    }
+  } else {
+    cat1_prob = vp9_cat1_prob;
+    cat2_prob = vp9_cat2_prob;
+    cat3_prob = vp9_cat3_prob;
+    cat4_prob = vp9_cat4_prob;
+    cat5_prob = vp9_cat5_prob;
+    cat6_prob = vp9_cat6_prob;
+  }
+#else
+  cat1_prob = vp9_cat1_prob;
+  cat2_prob = vp9_cat2_prob;
+  cat3_prob = vp9_cat3_prob;
+  cat4_prob = vp9_cat4_prob;
+  cat5_prob = vp9_cat5_prob;
+  cat6_prob = vp9_cat6_prob;
+#endif
+
+  while (c < max_eob) {
+    int val = -1;
+    band = *band_translate++;
+    prob = coef_probs[band][ctx];
+    if (counts)
+      ++eob_branch_count[band][ctx];
+    if (!vp9_read(r, prob[EOB_CONTEXT_NODE])) {
+      INCREMENT_COUNT(EOB_MODEL_TOKEN);
+      break;
+    }
+
+    while (!vp9_read(r, prob[ZERO_CONTEXT_NODE])) {
+      INCREMENT_COUNT(ZERO_TOKEN);
+      dqv = dq[1];
+      token_cache[scan[c]] = 0;
+      ++c;
+      if (c >= max_eob)
+        return c;  // zero tokens at the end (no eob token)
+      ctx = get_coef_context(nb, token_cache, c);
+      band = *band_translate++;
+      prob = coef_probs[band][ctx];
+    }
+
+    if (!vp9_read(r, prob[ONE_CONTEXT_NODE])) {
+      INCREMENT_COUNT(ONE_TOKEN);
+      token = ONE_TOKEN;
+      val = 1;
+    } else {
+      INCREMENT_COUNT(TWO_TOKEN);
+      token = vp9_read_tree(r, vp9_coef_con_tree,
+                            vp9_pareto8_full[prob[PIVOT_NODE] - 1]);
+      switch (token) {
+        case TWO_TOKEN:
+        case THREE_TOKEN:
+        case FOUR_TOKEN:
+          val = token;
+          break;
+        case CATEGORY1_TOKEN:
+          val = CAT1_MIN_VAL + read_coeff(cat1_prob, 1, r);
+          break;
+        case CATEGORY2_TOKEN:
+          val = CAT2_MIN_VAL + read_coeff(cat2_prob, 2, r);
+          break;
+        case CATEGORY3_TOKEN:
+          val = CAT3_MIN_VAL + read_coeff(cat3_prob, 3, r);
+          break;
+        case CATEGORY4_TOKEN:
+          val = CAT4_MIN_VAL + read_coeff(cat4_prob, 4, r);
+          break;
+        case CATEGORY5_TOKEN:
+          val = CAT5_MIN_VAL + read_coeff(cat5_prob, 5, r);
+          break;
+        case CATEGORY6_TOKEN:
+#if CONFIG_VP9_HIGHBITDEPTH
+          switch (xd->bd) {
+            case VPX_BITS_8:
+              val = CAT6_MIN_VAL + read_coeff(cat6_prob, 14, r);
+              break;
+            case VPX_BITS_10:
+              val = CAT6_MIN_VAL + read_coeff(cat6_prob, 16, r);
+              break;
+            case VPX_BITS_12:
+              val = CAT6_MIN_VAL + read_coeff(cat6_prob, 18, r);
+              break;
+            default:
+              assert(0);
+              return -1;
+          }
+#else
+          val = CAT6_MIN_VAL + read_coeff(cat6_prob, 14, r);
+#endif
+          break;
+      }
+    }
+    v = (val * dqv) >> dq_shift;
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+#if CONFIG_VP9_HIGHBITDEPTH
+    dqcoeff[scan[c]] = highbd_check_range((vp9_read_bit(r) ? -v : v),
+                                          xd->bd);
+#else
+    dqcoeff[scan[c]] = check_range(vp9_read_bit(r) ? -v : v);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+#else
+    dqcoeff[scan[c]] = vp9_read_bit(r) ? -v : v;
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+    token_cache[scan[c]] = vp9_pt_energy_class[token];
+    ++c;
+    ctx = get_coef_context(nb, token_cache, c);
+    dqv = dq[1];
+  }
+
+  return c;
+}
+
+int vp9_decode_block_tokens(MACROBLOCKD *xd,
+                            int plane, int block,
+                            BLOCK_SIZE plane_bsize, int x, int y,
+                            TX_SIZE tx_size, vp9_reader *r,
+                            int seg_id) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int16_t *const dequant = pd->seg_dequant[seg_id];
+  const int ctx = get_entropy_context(tx_size, pd->above_context + x,
+                                               pd->left_context + y);
+  const scan_order *so = get_scan(xd, tx_size, pd->plane_type, block);
+  const int eob = decode_coefs(xd, pd->plane_type,
+                               BLOCK_OFFSET(pd->dqcoeff, block), tx_size,
+                               dequant, ctx, so->scan, so->neighbors, r);
+  vp9_set_contexts(xd, pd, plane_bsize, tx_size, eob > 0, x, y);
+  return eob;
+}
+
+
diff --git a/media/libvpx/vp9/decoder/vp9_detokenize.h b/media/libvpx/vp9/decoder/vp9_detokenize.h
new file mode 100644
index 000000000..df1760668
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_detokenize.h
@@ -0,0 +1,32 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_DECODER_VP9_DETOKENIZE_H_
+#define VP9_DECODER_VP9_DETOKENIZE_H_
+
+#include "vp9/decoder/vp9_decoder.h"
+#include "vp9/decoder/vp9_reader.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int vp9_decode_block_tokens(MACROBLOCKD *xd,
+                            int plane, int block,
+                            BLOCK_SIZE plane_bsize, int x, int y,
+                            TX_SIZE tx_size, vp9_reader *r,
+                            int seg_id);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DETOKENIZE_H_
diff --git a/media/libvpx/vp9/decoder/vp9_dsubexp.c b/media/libvpx/vp9/decoder/vp9_dsubexp.c
new file mode 100644
index 000000000..c22617edb
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_dsubexp.c
@@ -0,0 +1,76 @@
+/*
+  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vp9/common/vp9_entropy.h"
+
+#include "vp9/decoder/vp9_dsubexp.h"
+
+static int inv_recenter_nonneg(int v, int m) {
+  if (v > 2 * m)
+    return v;
+
+  return v % 2 ? m - (v + 1) / 2 : m + v / 2;
+}
+
+static int decode_uniform(vp9_reader *r) {
+  const int l = 8;
+  const int m = (1 << l) - 191;
+  const int v = vp9_read_literal(r, l - 1);
+  return v < m ?  v : (v << 1) - m + vp9_read_bit(r);
+}
+
+static int inv_remap_prob(int v, int m) {
+  static int inv_map_table[MAX_PROB - 1] = {
+      6,  19,  32,  45,  58,  71,  84,  97, 110, 123, 136, 149, 162, 175, 188,
+    201, 214, 227, 240, 253,   0,   1,   2,   3,   4,   5,   7,   8,   9,  10,
+     11,  12,  13,  14,  15,  16,  17,  18,  20,  21,  22,  23,  24,  25,  26,
+     27,  28,  29,  30,  31,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,
+     43,  44,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  59,
+     60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  72,  73,  74,  75,
+     76,  77,  78,  79,  80,  81,  82,  83,  85,  86,  87,  88,  89,  90,  91,
+     92,  93,  94,  95,  96,  98,  99, 100, 101, 102, 103, 104, 105, 106, 107,
+    108, 109, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 124,
+    125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 137, 138, 139, 140,
+    141, 142, 143, 144, 145, 146, 147, 148, 150, 151, 152, 153, 154, 155, 156,
+    157, 158, 159, 160, 161, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172,
+    173, 174, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 189,
+    190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 202, 203, 204, 205,
+    206, 207, 208, 209, 210, 211, 212, 213, 215, 216, 217, 218, 219, 220, 221,
+    222, 223, 224, 225, 226, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237,
+    238, 239, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252
+  };
+  // The clamp is not necessary for conforming VP9 stream, it is added to
+  // prevent out of bound access for bad input data
+  v = clamp(v, 0, 253);
+  v = inv_map_table[v];
+  m--;
+  if ((m << 1) <= MAX_PROB) {
+    return 1 + inv_recenter_nonneg(v + 1, m);
+  } else {
+    return MAX_PROB - inv_recenter_nonneg(v + 1, MAX_PROB - 1 - m);
+  }
+}
+
+static int decode_term_subexp(vp9_reader *r) {
+  if (!vp9_read_bit(r))
+    return vp9_read_literal(r, 4);
+  if (!vp9_read_bit(r))
+    return vp9_read_literal(r, 4) + 16;
+  if (!vp9_read_bit(r))
+    return vp9_read_literal(r, 5) + 32;
+  return decode_uniform(r) + 64;
+}
+
+void vp9_diff_update_prob(vp9_reader *r, vp9_prob* p) {
+  if (vp9_read(r, DIFF_UPDATE_PROB)) {
+    const int delp = decode_term_subexp(r);
+    *p = (vp9_prob)inv_remap_prob(delp, *p);
+  }
+}
diff --git a/media/libvpx/vp9/decoder/vp9_dsubexp.h b/media/libvpx/vp9/decoder/vp9_dsubexp.h
new file mode 100644
index 000000000..436f434fb
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_dsubexp.h
@@ -0,0 +1,27 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#ifndef VP9_DECODER_VP9_DSUBEXP_H_
+#define VP9_DECODER_VP9_DSUBEXP_H_
+
+#include "vp9/decoder/vp9_reader.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void vp9_diff_update_prob(vp9_reader *r, vp9_prob* p);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_DSUBEXP_H_
diff --git a/media/libvpx/vp9/decoder/vp9_dthread.c b/media/libvpx/vp9/decoder/vp9_dthread.c
new file mode 100644
index 000000000..96a63bd9e
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_dthread.c
@@ -0,0 +1,189 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vpx_mem/vpx_mem.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/decoder/vp9_dthread.h"
+#include "vp9/decoder/vp9_decoder.h"
+
+// #define DEBUG_THREAD
+
+// TODO(hkuang): Clean up all the #ifdef in this file.
+void vp9_frameworker_lock_stats(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_lock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void vp9_frameworker_unlock_stats(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_unlock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void vp9_frameworker_signal_stats(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+
+// TODO(hkuang): Fix the pthread_cond_broadcast in windows wrapper.
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+  pthread_cond_signal(&worker_data->stats_cond);
+#else
+  pthread_cond_broadcast(&worker_data->stats_cond);
+#endif
+
+#else
+  (void)worker;
+#endif
+}
+
+// This macro prevents thread_sanitizer from reporting known concurrent writes.
+#if defined(__has_feature)
+#if __has_feature(thread_sanitizer)
+#define BUILDING_WITH_TSAN
+#endif
+#endif
+
+// TODO(hkuang): Remove worker parameter as it is only used in debug code.
+void vp9_frameworker_wait(VP9Worker *const worker, RefCntBuffer *const ref_buf,
+                          int row) {
+#if CONFIG_MULTITHREAD
+  if (!ref_buf)
+    return;
+
+#ifndef BUILDING_WITH_TSAN
+  // The following line of code will get harmless tsan error but it is the key
+  // to get best performance.
+  if (ref_buf->row >= row && ref_buf->buf.corrupted != 1) return;
+#endif
+
+  {
+    // Find the worker thread that owns the reference frame. If the reference
+    // frame has been fully decoded, it may not have owner.
+    VP9Worker *const ref_worker = ref_buf->frame_worker_owner;
+    FrameWorkerData *const ref_worker_data =
+        (FrameWorkerData *)ref_worker->data1;
+    const VP9Decoder *const pbi = ref_worker_data->pbi;
+
+#ifdef DEBUG_THREAD
+    {
+      FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+      printf("%d %p worker is waiting for %d %p worker (%d)  ref %d \r\n",
+             worker_data->worker_id, worker, ref_worker_data->worker_id,
+             ref_buf->frame_worker_owner, row, ref_buf->row);
+    }
+#endif
+
+    vp9_frameworker_lock_stats(ref_worker);
+    while (ref_buf->row < row && pbi->cur_buf == ref_buf &&
+           ref_buf->buf.corrupted != 1) {
+      pthread_cond_wait(&ref_worker_data->stats_cond,
+                        &ref_worker_data->stats_mutex);
+    }
+
+    if (ref_buf->buf.corrupted == 1) {
+      FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+      vp9_frameworker_unlock_stats(ref_worker);
+      vpx_internal_error(&worker_data->pbi->common.error,
+                         VPX_CODEC_CORRUPT_FRAME,
+                         "Worker %p failed to decode frame", worker);
+    }
+    vp9_frameworker_unlock_stats(ref_worker);
+  }
+#else
+  (void)worker;
+  (void)ref_buf;
+  (void)row;
+  (void)ref_buf;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void vp9_frameworker_broadcast(RefCntBuffer *const buf, int row) {
+#if CONFIG_MULTITHREAD
+  VP9Worker *worker = buf->frame_worker_owner;
+
+#ifdef DEBUG_THREAD
+  {
+    FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+    printf("%d %p worker decode to (%d) \r\n", worker_data->worker_id,
+           buf->frame_worker_owner, row);
+  }
+#endif
+
+  vp9_frameworker_lock_stats(worker);
+  buf->row = row;
+  vp9_frameworker_signal_stats(worker);
+  vp9_frameworker_unlock_stats(worker);
+#else
+  (void)buf;
+  (void)row;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void vp9_frameworker_copy_context(VP9Worker *const dst_worker,
+                                  VP9Worker *const src_worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const src_worker_data = (FrameWorkerData *)src_worker->data1;
+  FrameWorkerData *const dst_worker_data = (FrameWorkerData *)dst_worker->data1;
+  VP9_COMMON *const src_cm = &src_worker_data->pbi->common;
+  VP9_COMMON *const dst_cm = &dst_worker_data->pbi->common;
+  int i;
+
+  // Wait until source frame's context is ready.
+  vp9_frameworker_lock_stats(src_worker);
+  while (!src_worker_data->frame_context_ready) {
+    pthread_cond_wait(&src_worker_data->stats_cond,
+        &src_worker_data->stats_mutex);
+  }
+
+  dst_cm->last_frame_seg_map = src_cm->seg.enabled ?
+      src_cm->current_frame_seg_map : src_cm->last_frame_seg_map;
+  dst_worker_data->pbi->need_resync = src_worker_data->pbi->need_resync;
+  vp9_frameworker_unlock_stats(src_worker);
+
+  dst_cm->bit_depth = src_cm->bit_depth;
+#if CONFIG_VP9_HIGHBITDEPTH
+  dst_cm->use_highbitdepth = src_cm->use_highbitdepth;
+#endif
+  dst_cm->prev_frame = src_cm->show_existing_frame ?
+                       src_cm->prev_frame : src_cm->cur_frame;
+  dst_cm->last_width = !src_cm->show_existing_frame ?
+                       src_cm->width : src_cm->last_width;
+  dst_cm->last_height = !src_cm->show_existing_frame ?
+                        src_cm->height : src_cm->last_height;
+  dst_cm->subsampling_x = src_cm->subsampling_x;
+  dst_cm->subsampling_y = src_cm->subsampling_y;
+  dst_cm->frame_type = src_cm->frame_type;
+  dst_cm->last_show_frame = !src_cm->show_existing_frame ?
+                            src_cm->show_frame : src_cm->last_show_frame;
+  for (i = 0; i < REF_FRAMES; ++i)
+    dst_cm->ref_frame_map[i] = src_cm->next_ref_frame_map[i];
+
+  memcpy(dst_cm->lf_info.lfthr, src_cm->lf_info.lfthr,
+         (MAX_LOOP_FILTER + 1) * sizeof(loop_filter_thresh));
+  dst_cm->lf.last_sharpness_level = src_cm->lf.sharpness_level;
+  dst_cm->lf.filter_level = src_cm->lf.filter_level;
+  memcpy(dst_cm->lf.ref_deltas, src_cm->lf.ref_deltas, MAX_REF_LF_DELTAS);
+  memcpy(dst_cm->lf.mode_deltas, src_cm->lf.mode_deltas, MAX_MODE_LF_DELTAS);
+  dst_cm->seg = src_cm->seg;
+  memcpy(dst_cm->frame_contexts, src_cm->frame_contexts,
+         FRAME_CONTEXTS * sizeof(dst_cm->frame_contexts[0]));
+#else
+  (void) dst_worker;
+  (void) src_worker;
+#endif  // CONFIG_MULTITHREAD
+}
diff --git a/media/libvpx/vp9/decoder/vp9_dthread.h b/media/libvpx/vp9/decoder/vp9_dthread.h
new file mode 100644
index 000000000..979cb3d8b
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_dthread.h
@@ -0,0 +1,66 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_DECODER_VP9_DTHREAD_H_
+#define VP9_DECODER_VP9_DTHREAD_H_
+
+#include "./vpx_config.h"
+#include "vp9/common/vp9_thread.h"
+#include "vpx/internal/vpx_codec_internal.h"
+
+struct VP9Common;
+struct VP9Decoder;
+
+// WorkerData for the FrameWorker thread. It contains all the information of
+// the worker and decode structures for decoding a frame.
+typedef struct FrameWorkerData {
+  struct VP9Decoder *pbi;
+  const uint8_t *data;
+  const uint8_t *data_end;
+  size_t data_size;
+  void *user_priv;
+  int result;
+  int worker_id;
+  int received_frame;
+
+  // scratch_buffer is used in frame parallel mode only.
+  // It is used to make a copy of the compressed data.
+  uint8_t *scratch_buffer;
+  size_t scratch_buffer_size;
+
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t stats_mutex;
+  pthread_cond_t stats_cond;
+#endif
+
+  int frame_context_ready;  // Current frame's context is ready to read.
+  int frame_decoded;        // Finished decoding current frame.
+} FrameWorkerData;
+
+void vp9_frameworker_lock_stats(VP9Worker *const worker);
+void vp9_frameworker_unlock_stats(VP9Worker *const worker);
+void vp9_frameworker_signal_stats(VP9Worker *const worker);
+
+// Wait until ref_buf has been decoded to row in real pixel unit.
+// Note: worker may already finish decoding ref_buf and release it in order to
+// start decoding next frame. So need to check whether worker is still decoding
+// ref_buf.
+void vp9_frameworker_wait(VP9Worker *const worker, RefCntBuffer *const ref_buf,
+                          int row);
+
+// FrameWorker broadcasts its decoding progress so other workers that are
+// waiting on it can resume decoding.
+void vp9_frameworker_broadcast(RefCntBuffer *const buf, int row);
+
+// Copy necessary decoding context from src worker to dst worker.
+void vp9_frameworker_copy_context(VP9Worker *const dst_worker,
+                                  VP9Worker *const src_worker);
+
+#endif  // VP9_DECODER_VP9_DTHREAD_H_
diff --git a/media/libvpx/vp9/decoder/vp9_read_bit_buffer.c b/media/libvpx/vp9/decoder/vp9_read_bit_buffer.c
new file mode 100644
index 000000000..c3b38a9c7
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_read_bit_buffer.c
@@ -0,0 +1,41 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "vp9/decoder/vp9_read_bit_buffer.h"
+
+size_t vp9_rb_bytes_read(struct vp9_read_bit_buffer *rb) {
+  return (rb->bit_offset + 7) >> 3;
+}
+
+int vp9_rb_read_bit(struct vp9_read_bit_buffer *rb) {
+  const size_t off = rb->bit_offset;
+  const size_t p = off >> 3;
+  const int q = 7 - (int)(off & 0x7);
+  if (rb->bit_buffer + p < rb->bit_buffer_end) {
+    const int bit = (rb->bit_buffer[p] >> q) & 1;
+    rb->bit_offset = off + 1;
+    return bit;
+  } else {
+    rb->error_handler(rb->error_handler_data);
+    return 0;
+  }
+}
+
+int vp9_rb_read_literal(struct vp9_read_bit_buffer *rb, int bits) {
+  int value = 0, bit;
+  for (bit = bits - 1; bit >= 0; bit--)
+    value |= vp9_rb_read_bit(rb) << bit;
+  return value;
+}
+
+int vp9_rb_read_signed_literal(struct vp9_read_bit_buffer *rb,
+                               int bits) {
+  const int value = vp9_rb_read_literal(rb, bits);
+  return vp9_rb_read_bit(rb) ? -value : value;
+}
diff --git a/media/libvpx/vp9/decoder/vp9_read_bit_buffer.h b/media/libvpx/vp9/decoder/vp9_read_bit_buffer.h
new file mode 100644
index 000000000..fc88bd70a
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_read_bit_buffer.h
@@ -0,0 +1,45 @@
+/*
+ *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_DECODER_VP9_READ_BIT_BUFFER_H_
+#define VP9_DECODER_VP9_READ_BIT_BUFFER_H_
+
+#include <limits.h>
+
+#include "vpx/vpx_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*vp9_rb_error_handler)(void *data);
+
+struct vp9_read_bit_buffer {
+  const uint8_t *bit_buffer;
+  const uint8_t *bit_buffer_end;
+  size_t bit_offset;
+
+  void *error_handler_data;
+  vp9_rb_error_handler error_handler;
+};
+
+size_t vp9_rb_bytes_read(struct vp9_read_bit_buffer *rb);
+
+int vp9_rb_read_bit(struct vp9_read_bit_buffer *rb);
+
+int vp9_rb_read_literal(struct vp9_read_bit_buffer *rb, int bits);
+
+int vp9_rb_read_signed_literal(struct vp9_read_bit_buffer *rb, int bits);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_READ_BIT_BUFFER_H_
diff --git a/media/libvpx/vp9/decoder/vp9_reader.c b/media/libvpx/vp9/decoder/vp9_reader.c
new file mode 100644
index 000000000..6bb4f9f73
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_reader.c
@@ -0,0 +1,106 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_ports/mem.h"
+#include "vpx_mem/vpx_mem.h"
+
+#include "vp9/decoder/vp9_reader.h"
+
+// This is meant to be a large, positive constant that can still be efficiently
+// loaded as an immediate (on platforms like ARM, for example).
+// Even relatively modest values like 100 would work fine.
+#define LOTS_OF_BITS 0x40000000
+
+int vp9_reader_init(vp9_reader *r,
+                    const uint8_t *buffer,
+                    size_t size,
+                    vpx_decrypt_cb decrypt_cb,
+                    void *decrypt_state) {
+  if (size && !buffer) {
+    return 1;
+  } else {
+    r->buffer_end = buffer + size;
+    r->buffer = buffer;
+    r->value = 0;
+    r->count = -8;
+    r->range = 255;
+    r->decrypt_cb = decrypt_cb;
+    r->decrypt_state = decrypt_state;
+    vp9_reader_fill(r);
+    return vp9_read_bit(r) != 0;  // marker bit
+  }
+}
+
+void vp9_reader_fill(vp9_reader *r) {
+  const uint8_t *const buffer_end = r->buffer_end;
+  const uint8_t *buffer = r->buffer;
+  const uint8_t *buffer_start = buffer;
+  BD_VALUE value = r->value;
+  int count = r->count;
+  int shift = BD_VALUE_SIZE - CHAR_BIT - (count + CHAR_BIT);
+  int loop_end = 0;
+  const size_t bytes_left = buffer_end - buffer;
+  const size_t bits_left = bytes_left * CHAR_BIT;
+  const int x = (int)(shift + CHAR_BIT - bits_left);
+
+  if (r->decrypt_cb) {
+    size_t n = MIN(sizeof(r->clear_buffer), bytes_left);
+    r->decrypt_cb(r->decrypt_state, buffer, r->clear_buffer, (int)n);
+    buffer = r->clear_buffer;
+    buffer_start = r->clear_buffer;
+  }
+
+  if (x >= 0) {
+    count += LOTS_OF_BITS;
+    loop_end = x;
+  }
+
+  if (x < 0 || bits_left) {
+    while (shift >= loop_end) {
+      count += CHAR_BIT;
+      value |= (BD_VALUE)*buffer++ << shift;
+      shift -= CHAR_BIT;
+    }
+  }
+
+  // NOTE: Variable 'buffer' may not relate to 'r->buffer' after decryption,
+  // so we increase 'r->buffer' by the amount that 'buffer' moved, rather than
+  // assign 'buffer' to 'r->buffer'.
+  r->buffer += buffer - buffer_start;
+  r->value = value;
+  r->count = count;
+}
+
+const uint8_t *vp9_reader_find_end(vp9_reader *r) {
+  // Find the end of the coded buffer
+  while (r->count > CHAR_BIT && r->count < BD_VALUE_SIZE) {
+    r->count -= CHAR_BIT;
+    r->buffer--;
+  }
+  return r->buffer;
+}
+
+int vp9_reader_has_error(vp9_reader *r) {
+  // Check if we have reached the end of the buffer.
+  //
+  // Variable 'count' stores the number of bits in the 'value' buffer, minus
+  // 8. The top byte is part of the algorithm, and the remainder is buffered
+  // to be shifted into it. So if count == 8, the top 16 bits of 'value' are
+  // occupied, 8 for the algorithm and 8 in the buffer.
+  //
+  // When reading a byte from the user's buffer, count is filled with 8 and
+  // one byte is filled into the value buffer. When we reach the end of the
+  // data, count is additionally filled with LOTS_OF_BITS. So when
+  // count == LOTS_OF_BITS - 1, the user's data has been exhausted.
+  //
+  // 1 if we have tried to decode bits after the end of stream was encountered.
+  // 0 No error.
+  return r->count > BD_VALUE_SIZE && r->count < LOTS_OF_BITS;
+}
diff --git a/media/libvpx/vp9/decoder/vp9_reader.h b/media/libvpx/vp9/decoder/vp9_reader.h
new file mode 100644
index 000000000..a68a1d592
--- /dev/null
+++ b/media/libvpx/vp9/decoder/vp9_reader.h
@@ -0,0 +1,120 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_DECODER_VP9_READER_H_
+#define VP9_DECODER_VP9_READER_H_
+
+#include <stddef.h>
+#include <limits.h>
+
+#include "./vpx_config.h"
+#include "vpx_ports/mem.h"
+#include "vpx/vp8dx.h"
+#include "vpx/vpx_integer.h"
+
+#include "vp9/common/vp9_prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef size_t BD_VALUE;
+
+#define BD_VALUE_SIZE ((int)sizeof(BD_VALUE) * CHAR_BIT)
+
+typedef struct {
+  // Be careful when reordering this struct, it may impact the cache negatively.
+  BD_VALUE value;
+  unsigned int range;
+  int count;
+  const uint8_t *buffer_end;
+  const uint8_t *buffer;
+  vpx_decrypt_cb decrypt_cb;
+  void *decrypt_state;
+  uint8_t clear_buffer[sizeof(BD_VALUE) + 1];
+} vp9_reader;
+
+int vp9_reader_init(vp9_reader *r,
+                    const uint8_t *buffer,
+                    size_t size,
+                    vpx_decrypt_cb decrypt_cb,
+                    void *decrypt_state);
+
+void vp9_reader_fill(vp9_reader *r);
+
+int vp9_reader_has_error(vp9_reader *r);
+
+const uint8_t *vp9_reader_find_end(vp9_reader *r);
+
+static INLINE int vp9_read(vp9_reader *r, int prob) {
+  unsigned int bit = 0;
+  BD_VALUE value;
+  BD_VALUE bigsplit;
+  int count;
+  unsigned int range;
+  unsigned int split = (r->range * prob + (256 - prob)) >> CHAR_BIT;
+
+  if (r->count < 0)
+    vp9_reader_fill(r);
+
+  value = r->value;
+  count = r->count;
+
+  bigsplit = (BD_VALUE)split << (BD_VALUE_SIZE - CHAR_BIT);
+
+  range = split;
+
+  if (value >= bigsplit) {
+    range = r->range - split;
+    value = value - bigsplit;
+    bit = 1;
+  }
+
+  {
+    register unsigned int shift = vp9_norm[range];
+    range <<= shift;
+    value <<= shift;
+    count -= shift;
+  }
+  r->value = value;
+  r->count = count;
+  r->range = range;
+
+  return bit;
+}
+
+static INLINE int vp9_read_bit(vp9_reader *r) {
+  return vp9_read(r, 128);  // vp9_prob_half
+}
+
+static INLINE int vp9_read_literal(vp9_reader *r, int bits) {
+  int literal = 0, bit;
+
+  for (bit = bits - 1; bit >= 0; bit--)
+    literal |= vp9_read_bit(r) << bit;
+
+  return literal;
+}
+
+static INLINE int vp9_read_tree(vp9_reader *r, const vp9_tree_index *tree,
+                                const vp9_prob *probs) {
+  vp9_tree_index i = 0;
+
+  while ((i = tree[i + vp9_read(r, probs[i >> 1])]) > 0)
+    continue;
+
+  return -i;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_DECODER_VP9_READER_H_