Import aom library

This is the reference implementation for the Alliance for Open Media's av1 video code.

The commit used was 4d668d7feb1f8abd809d1bca0418570a7f142a36.

23 files changed, 10749 insertions, 0 deletions

diff --git a/third_party/aom/av1/decoder/accounting.c b/third_party/aom/av1/decoder/accounting.c
new file mode 100644
index 000000000..ba243c9e1
--- /dev/null
+++ b/third_party/aom/av1/decoder/accounting.c
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_integer.h"
+#include "./accounting.h"
+
+static int aom_accounting_hash(const char *str) {
+  uint32_t val;
+  const unsigned char *ustr;
+  val = 0;
+  ustr = (const unsigned char *)str;
+  /* This is about the worst hash one can design, but it should be good enough
+     here. */
+  while (*ustr) val += *ustr++;
+  return val % AOM_ACCOUNTING_HASH_SIZE;
+}
+
+/* Dictionary lookup based on an open-addressing hash table. */
+int aom_accounting_dictionary_lookup(Accounting *accounting, const char *str) {
+  int hash;
+  int len;
+  AccountingDictionary *dictionary;
+  dictionary = &accounting->syms.dictionary;
+  hash = aom_accounting_hash(str);
+  while (accounting->hash_dictionary[hash] != -1) {
+    if (strcmp(dictionary->strs[accounting->hash_dictionary[hash]], str) == 0) {
+      return accounting->hash_dictionary[hash];
+    }
+    hash++;
+    if (hash == AOM_ACCOUNTING_HASH_SIZE) hash = 0;
+  }
+  /* No match found. */
+  assert(dictionary->num_strs + 1 < MAX_SYMBOL_TYPES);
+  accounting->hash_dictionary[hash] = dictionary->num_strs;
+  len = strlen(str);
+  dictionary->strs[dictionary->num_strs] = malloc(len + 1);
+  snprintf(dictionary->strs[dictionary->num_strs], len + 1, "%s", str);
+  dictionary->num_strs++;
+  return dictionary->num_strs - 1;
+}
+
+void aom_accounting_init(Accounting *accounting) {
+  int i;
+  accounting->num_syms_allocated = 1000;
+  accounting->syms.syms =
+      malloc(sizeof(AccountingSymbol) * accounting->num_syms_allocated);
+  accounting->syms.dictionary.num_strs = 0;
+  assert(AOM_ACCOUNTING_HASH_SIZE > 2 * MAX_SYMBOL_TYPES);
+  for (i = 0; i < AOM_ACCOUNTING_HASH_SIZE; i++)
+    accounting->hash_dictionary[i] = -1;
+  aom_accounting_reset(accounting);
+}
+
+void aom_accounting_reset(Accounting *accounting) {
+  accounting->syms.num_syms = 0;
+  accounting->syms.num_binary_syms = 0;
+  accounting->syms.num_multi_syms = 0;
+  accounting->context.x = -1;
+  accounting->context.y = -1;
+  accounting->last_tell_frac = 0;
+}
+
+void aom_accounting_clear(Accounting *accounting) {
+  int i;
+  AccountingDictionary *dictionary;
+  free(accounting->syms.syms);
+  dictionary = &accounting->syms.dictionary;
+  for (i = 0; i < dictionary->num_strs; i++) {
+    free(dictionary->strs[i]);
+  }
+}
+
+void aom_accounting_set_context(Accounting *accounting, int16_t x, int16_t y) {
+  accounting->context.x = x;
+  accounting->context.y = y;
+}
+
+void aom_accounting_record(Accounting *accounting, const char *str,
+                           uint32_t bits) {
+  AccountingSymbol sym;
+  // Reuse previous symbol if it has the same context and symbol id.
+  if (accounting->syms.num_syms) {
+    AccountingSymbol *last_sym;
+    last_sym = &accounting->syms.syms[accounting->syms.num_syms - 1];
+    if (memcmp(&last_sym->context, &accounting->context,
+               sizeof(AccountingSymbolContext)) == 0) {
+      uint32_t id;
+      id = aom_accounting_dictionary_lookup(accounting, str);
+      if (id == last_sym->id) {
+        last_sym->bits += bits;
+        last_sym->samples++;
+        return;
+      }
+    }
+  }
+  sym.context = accounting->context;
+  sym.samples = 1;
+  sym.bits = bits;
+  sym.id = aom_accounting_dictionary_lookup(accounting, str);
+  assert(sym.id <= 255);
+  if (accounting->syms.num_syms == accounting->num_syms_allocated) {
+    accounting->num_syms_allocated *= 2;
+    accounting->syms.syms =
+        realloc(accounting->syms.syms,
+                sizeof(AccountingSymbol) * accounting->num_syms_allocated);
+    assert(accounting->syms.syms != NULL);
+  }
+  accounting->syms.syms[accounting->syms.num_syms++] = sym;
+}
+
+void aom_accounting_dump(Accounting *accounting) {
+  int i;
+  AccountingSymbol *sym;
+  printf("\n----- Number of recorded syntax elements = %d -----\n",
+         accounting->syms.num_syms);
+  printf("----- Total number of symbol calls = %d (%d binary) -----\n",
+         accounting->syms.num_multi_syms + accounting->syms.num_binary_syms,
+         accounting->syms.num_binary_syms);
+  for (i = 0; i < accounting->syms.num_syms; i++) {
+    sym = &accounting->syms.syms[i];
+    printf("%s x: %d, y: %d bits: %f samples: %d\n",
+           accounting->syms.dictionary.strs[sym->id], sym->context.x,
+           sym->context.y, (float)sym->bits / 8.0, sym->samples);
+  }
+}
diff --git a/third_party/aom/av1/decoder/accounting.h b/third_party/aom/av1/decoder/accounting.h
new file mode 100644
index 000000000..889865b2e
--- /dev/null
+++ b/third_party/aom/av1/decoder/accounting.h
@@ -0,0 +1,83 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_ACCOUNTING_H_
+#define AOM_ACCOUNTING_H_
+#include <stdlib.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+#define AOM_ACCOUNTING_HASH_SIZE (1021)
+
+/* Max number of entries for symbol types in the dictionary (increase as
+   necessary). */
+#define MAX_SYMBOL_TYPES (256)
+
+/*The resolution of fractional-precision bit usage measurements, i.e.,
+   3 => 1/8th bits.*/
+#define AOM_ACCT_BITRES (3)
+
+typedef struct {
+  int16_t x;
+  int16_t y;
+} AccountingSymbolContext;
+
+typedef struct {
+  AccountingSymbolContext context;
+  uint32_t id;
+  /** Number of bits in units of 1/8 bit. */
+  uint32_t bits;
+  uint32_t samples;
+} AccountingSymbol;
+
+/** Dictionary for translating strings into id. */
+typedef struct {
+  char *(strs[MAX_SYMBOL_TYPES]);
+  int num_strs;
+} AccountingDictionary;
+
+typedef struct {
+  /** All recorded symbols decoded. */
+  AccountingSymbol *syms;
+  /** Number of syntax actually recorded. */
+  int num_syms;
+  /** Raw symbol decoding calls for non-binary values. */
+  int num_multi_syms;
+  /** Raw binary symbol decoding calls. */
+  int num_binary_syms;
+  /** Dictionary for translating strings into id. */
+  AccountingDictionary dictionary;
+} AccountingSymbols;
+
+typedef struct Accounting Accounting;
+
+struct Accounting {
+  AccountingSymbols syms;
+  /** Size allocated for symbols (not all may be used). */
+  int num_syms_allocated;
+  int16_t hash_dictionary[AOM_ACCOUNTING_HASH_SIZE];
+  AccountingSymbolContext context;
+  uint32_t last_tell_frac;
+};
+
+void aom_accounting_init(Accounting *accounting);
+void aom_accounting_reset(Accounting *accounting);
+void aom_accounting_clear(Accounting *accounting);
+void aom_accounting_set_context(Accounting *accounting, int16_t x, int16_t y);
+int aom_accounting_dictionary_lookup(Accounting *accounting, const char *str);
+void aom_accounting_record(Accounting *accounting, const char *str,
+                           uint32_t bits);
+void aom_accounting_dump(Accounting *accounting);
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_ACCOUNTING_H_
diff --git a/third_party/aom/av1/decoder/decint.h b/third_party/aom/av1/decoder/decint.h
new file mode 100644
index 000000000..e887ad5e0
--- /dev/null
+++ b/third_party/aom/av1/decoder/decint.h
@@ -0,0 +1,35 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_decint_H)
+# define _decint_H (1)
+# include "av1/common/pvq_state.h"
+# include "aom_dsp/bitreader.h"
+# include "aom_dsp/entdec.h"
+
+typedef struct daala_dec_ctx daala_dec_ctx;
+
+typedef struct daala_dec_ctx od_dec_ctx;
+
+
+struct daala_dec_ctx {
+  /* Stores context-adaptive CDFs for PVQ. */
+  od_state state;
+  /* AOM entropy decoder. */
+  aom_reader *r;
+  int use_activity_masking;
+  /* Mode of quantization matrice : FLAT (0) or HVS (1) */
+  int qm;
+};
+
+#endif
diff --git a/third_party/aom/av1/decoder/decodeframe.c b/third_party/aom/av1/decoder/decodeframe.c
new file mode 100644
index 000000000..289d38670
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodeframe.c
@@ -0,0 +1,5159 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdlib.h>  // qsort()
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "./aom_scale_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom/aom_codec.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/bitreader.h"
+#include "aom_dsp/bitreader_buffer.h"
+#include "aom_dsp/binary_codes_reader.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/mem_ops.h"
+#include "aom_scale/aom_scale.h"
+#include "aom_util/aom_thread.h"
+
+#if CONFIG_BITSTREAM_DEBUG
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#include "av1/common/alloccommon.h"
+#if CONFIG_CDEF
+#include "av1/common/cdef.h"
+#include "av1/common/clpf.h"
+#endif
+#if CONFIG_INSPECTION
+#include "av1/decoder/inspection.h"
+#endif
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/entropymv.h"
+#include "av1/common/idct.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+#include "av1/common/seg_common.h"
+#include "av1/common/thread_common.h"
+#include "av1/common/tile_common.h"
+
+#include "av1/decoder/decodeframe.h"
+#include "av1/decoder/decodemv.h"
+#include "av1/decoder/decoder.h"
+#if CONFIG_LV_MAP
+#include "av1/decoder/decodetxb.h"
+#endif
+#include "av1/decoder/detokenize.h"
+#include "av1/decoder/dsubexp.h"
+
+#if CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION
+#include "av1/common/warped_motion.h"
+#endif  // CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION
+
+#define MAX_AV1_HEADER_SIZE 80
+#define ACCT_STR __func__
+
+#if CONFIG_PVQ
+#include "av1/common/partition.h"
+#include "av1/common/pvq.h"
+#include "av1/common/scan.h"
+#include "av1/decoder/decint.h"
+#include "av1/decoder/pvq_decoder.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/hybrid_fwd_txfm.h"
+#endif
+
+#if CONFIG_CFL
+#include "av1/common/cfl.h"
+#endif
+
+static struct aom_read_bit_buffer *init_read_bit_buffer(
+    AV1Decoder *pbi, struct aom_read_bit_buffer *rb, const uint8_t *data,
+    const uint8_t *data_end, uint8_t clear_data[MAX_AV1_HEADER_SIZE]);
+static int read_compressed_header(AV1Decoder *pbi, const uint8_t *data,
+                                  size_t partition_size);
+static size_t read_uncompressed_header(AV1Decoder *pbi,
+                                       struct aom_read_bit_buffer *rb);
+
+static int is_compound_reference_allowed(const AV1_COMMON *cm) {
+#if CONFIG_LOWDELAY_COMPOUND  // Normative in decoder
+  return !frame_is_intra_only(cm);
+#else
+  int i;
+  if (frame_is_intra_only(cm)) return 0;
+  for (i = 1; i < INTER_REFS_PER_FRAME; ++i)
+    if (cm->ref_frame_sign_bias[i + 1] != cm->ref_frame_sign_bias[1]) return 1;
+
+  return 0;
+#endif
+}
+
+static void setup_compound_reference_mode(AV1_COMMON *cm) {
+#if CONFIG_EXT_REFS
+  cm->comp_fwd_ref[0] = LAST_FRAME;
+  cm->comp_fwd_ref[1] = LAST2_FRAME;
+  cm->comp_fwd_ref[2] = LAST3_FRAME;
+  cm->comp_fwd_ref[3] = GOLDEN_FRAME;
+
+  cm->comp_bwd_ref[0] = BWDREF_FRAME;
+  cm->comp_bwd_ref[1] = ALTREF_FRAME;
+#else
+  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;
+  }
+#endif  // CONFIG_EXT_REFS
+}
+
+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 aom_read_bit_buffer *rb, int max) {
+  const int data = aom_rb_read_literal(rb, get_unsigned_bits(max));
+  return data > max ? max : data;
+}
+
+static TX_MODE read_tx_mode(AV1_COMMON *cm, MACROBLOCKD *xd,
+                            struct aom_read_bit_buffer *rb) {
+  int i, all_lossless = 1;
+#if CONFIG_TX64X64
+  TX_MODE tx_mode;
+#endif
+
+  if (cm->seg.enabled) {
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      if (!xd->lossless[i]) {
+        all_lossless = 0;
+        break;
+      }
+    }
+  } else {
+    all_lossless = xd->lossless[0];
+  }
+
+  if (all_lossless) return ONLY_4X4;
+#if CONFIG_TX64X64
+  tx_mode = aom_rb_read_bit(rb) ? TX_MODE_SELECT : aom_rb_read_literal(rb, 2);
+  if (tx_mode == ALLOW_32X32) tx_mode += aom_rb_read_bit(rb);
+  return tx_mode;
+#else
+  return aom_rb_read_bit(rb) ? TX_MODE_SELECT : aom_rb_read_literal(rb, 2);
+#endif  // CONFIG_TX64X64
+}
+
+#if !CONFIG_EC_ADAPT
+static void read_tx_size_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j, k;
+  for (i = 0; i < MAX_TX_DEPTH; ++i)
+    for (j = 0; j < TX_SIZE_CONTEXTS; ++j)
+      for (k = 0; k < i + 1; ++k)
+        av1_diff_update_prob(r, &fc->tx_size_probs[i][j][k], ACCT_STR);
+}
+#endif
+
+#if !CONFIG_EC_ADAPT
+static void read_switchable_interp_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j;
+  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j) {
+    for (i = 0; i < SWITCHABLE_FILTERS - 1; ++i)
+      av1_diff_update_prob(r, &fc->switchable_interp_prob[j][i], ACCT_STR);
+  }
+}
+#endif
+
+static void read_inter_mode_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+#if CONFIG_REF_MV
+  int i;
+  for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i)
+    av1_diff_update_prob(r, &fc->newmv_prob[i], ACCT_STR);
+  for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i)
+    av1_diff_update_prob(r, &fc->zeromv_prob[i], ACCT_STR);
+  for (i = 0; i < REFMV_MODE_CONTEXTS; ++i)
+    av1_diff_update_prob(r, &fc->refmv_prob[i], ACCT_STR);
+  for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
+    av1_diff_update_prob(r, &fc->drl_prob[i], ACCT_STR);
+#else
+#if !CONFIG_EC_ADAPT
+  int i, j;
+  for (i = 0; i < INTER_MODE_CONTEXTS; ++i) {
+    for (j = 0; j < INTER_MODES - 1; ++j)
+      av1_diff_update_prob(r, &fc->inter_mode_probs[i][j], ACCT_STR);
+  }
+#else
+  (void)fc;
+  (void)r;
+#endif
+#endif
+}
+
+#if CONFIG_EXT_INTER
+static void read_inter_compound_mode_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j;
+  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+    for (j = 0; j < INTER_MODE_CONTEXTS; ++j) {
+      for (i = 0; i < INTER_COMPOUND_MODES - 1; ++i) {
+        av1_diff_update_prob(r, &fc->inter_compound_mode_probs[j][i], ACCT_STR);
+      }
+    }
+  }
+}
+#endif  // CONFIG_EXT_INTER
+#if !CONFIG_EC_ADAPT
+#if !CONFIG_EXT_TX
+static void read_ext_tx_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j, k;
+  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+      for (j = 0; j < TX_TYPES; ++j) {
+        for (k = 0; k < TX_TYPES - 1; ++k)
+          av1_diff_update_prob(r, &fc->intra_ext_tx_prob[i][j][k], ACCT_STR);
+      }
+    }
+  }
+  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+      for (k = 0; k < TX_TYPES - 1; ++k)
+        av1_diff_update_prob(r, &fc->inter_ext_tx_prob[i][k], ACCT_STR);
+    }
+  }
+}
+#endif
+#endif
+
+static REFERENCE_MODE read_frame_reference_mode(
+    const AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  if (is_compound_reference_allowed(cm)) {
+#if CONFIG_REF_ADAPT
+    return aom_rb_read_bit(rb) ? REFERENCE_MODE_SELECT : SINGLE_REFERENCE;
+#else
+    return aom_rb_read_bit(rb)
+               ? REFERENCE_MODE_SELECT
+               : (aom_rb_read_bit(rb) ? COMPOUND_REFERENCE : SINGLE_REFERENCE);
+#endif  // CONFIG_REF_ADAPT
+  } else {
+    return SINGLE_REFERENCE;
+  }
+}
+
+static void read_frame_reference_mode_probs(AV1_COMMON *cm, aom_reader *r) {
+  FRAME_CONTEXT *const fc = cm->fc;
+  int i, j;
+
+  if (cm->reference_mode == REFERENCE_MODE_SELECT)
+    for (i = 0; i < COMP_INTER_CONTEXTS; ++i)
+      av1_diff_update_prob(r, &fc->comp_inter_prob[i], ACCT_STR);
+
+  if (cm->reference_mode != COMPOUND_REFERENCE) {
+    for (i = 0; i < REF_CONTEXTS; ++i) {
+      for (j = 0; j < (SINGLE_REFS - 1); ++j) {
+        av1_diff_update_prob(r, &fc->single_ref_prob[i][j], ACCT_STR);
+      }
+    }
+  }
+
+  if (cm->reference_mode != SINGLE_REFERENCE) {
+    for (i = 0; i < REF_CONTEXTS; ++i) {
+#if CONFIG_EXT_REFS
+      for (j = 0; j < (FWD_REFS - 1); ++j)
+        av1_diff_update_prob(r, &fc->comp_ref_prob[i][j], ACCT_STR);
+      for (j = 0; j < (BWD_REFS - 1); ++j)
+        av1_diff_update_prob(r, &fc->comp_bwdref_prob[i][j], ACCT_STR);
+#else
+      for (j = 0; j < (COMP_REFS - 1); ++j)
+        av1_diff_update_prob(r, &fc->comp_ref_prob[i][j], ACCT_STR);
+#endif  // CONFIG_EXT_REFS
+    }
+  }
+}
+
+static void update_mv_probs(aom_prob *p, int n, aom_reader *r) {
+  int i;
+  for (i = 0; i < n; ++i) av1_diff_update_prob(r, &p[i], ACCT_STR);
+}
+
+static void read_mv_probs(nmv_context *ctx, int allow_hp, aom_reader *r) {
+  int i;
+
+#if !CONFIG_EC_ADAPT
+  int 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, MV_FP_SIZE - 1, r);
+  }
+#endif  // !CONFIG_EC_ADAPT
+
+  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,
+                                    const TX_TYPE tx_type,
+                                    const TX_SIZE tx_size, uint8_t *dst,
+                                    int stride, int16_t scan_line, int eob) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *const dqcoeff = pd->dqcoeff;
+  av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, dst, stride, eob);
+  memset(dqcoeff, 0, (scan_line + 1) * sizeof(dqcoeff[0]));
+}
+
+#if CONFIG_PVQ
+static int av1_pvq_decode_helper(MACROBLOCKD *xd, tran_low_t *ref_coeff,
+                                 tran_low_t *dqcoeff, int16_t *quant, int pli,
+                                 int bs, TX_TYPE tx_type, int xdec,
+                                 PVQ_SKIP_TYPE ac_dc_coded) {
+  unsigned int flags;  // used for daala's stream analyzer.
+  int off;
+  const int is_keyframe = 0;
+  const int has_dc_skip = 1;
+  int coeff_shift = 3 - av1_get_tx_scale(bs);
+  int hbd_downshift = 0;
+  int rounding_mask;
+  // DC quantizer for PVQ
+  int pvq_dc_quant;
+  int lossless = (quant[0] == 0);
+  const int blk_size = tx_size_wide[bs];
+  int eob = 0;
+  int i;
+  od_dec_ctx *dec = &xd->daala_dec;
+  int use_activity_masking = dec->use_activity_masking;
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_coeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+
+  od_coeff ref_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX];
+  od_coeff out_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX];
+
+#if CONFIG_HIGHBITDEPTH
+  hbd_downshift = xd->bd - 8;
+#endif  // CONFIG_HIGHBITDEPTH
+
+  od_raster_to_coding_order(ref_coeff_pvq, blk_size, tx_type, ref_coeff,
+                            blk_size);
+
+  assert(OD_COEFF_SHIFT >= 4);
+  if (lossless)
+    pvq_dc_quant = 1;
+  else {
+    if (use_activity_masking)
+      pvq_dc_quant = OD_MAXI(
+          1, (quant[0] << (OD_COEFF_SHIFT - 3) >> hbd_downshift) *
+                     dec->state.pvq_qm_q4[pli][od_qm_get_index(bs, 0)] >>
+                 4);
+    else
+      pvq_dc_quant =
+          OD_MAXI(1, quant[0] << (OD_COEFF_SHIFT - 3) >> hbd_downshift);
+  }
+
+  off = od_qm_offset(bs, xdec);
+
+  // copy int16 inputs to int32
+  for (i = 0; i < blk_size * blk_size; i++) {
+    ref_int32[i] =
+        AOM_SIGNED_SHL(ref_coeff_pvq[i], OD_COEFF_SHIFT - coeff_shift) >>
+        hbd_downshift;
+  }
+
+  od_pvq_decode(dec, ref_int32, out_int32,
+                OD_MAXI(1, quant[1] << (OD_COEFF_SHIFT - 3) >> hbd_downshift),
+                pli, bs, OD_PVQ_BETA[use_activity_masking][pli][bs],
+                is_keyframe, &flags, ac_dc_coded, dec->state.qm + off,
+                dec->state.qm_inv + off);
+
+  if (!has_dc_skip || out_int32[0]) {
+    out_int32[0] =
+        has_dc_skip + generic_decode(dec->r, &dec->state.adapt->model_dc[pli],
+                                     &dec->state.adapt->ex_dc[pli][bs][0], 2,
+                                     "dc:mag");
+    if (out_int32[0]) out_int32[0] *= aom_read_bit(dec->r, "dc:sign") ? -1 : 1;
+  }
+  out_int32[0] = out_int32[0] * pvq_dc_quant + ref_int32[0];
+
+  // copy int32 result back to int16
+  assert(OD_COEFF_SHIFT > coeff_shift);
+  rounding_mask = (1 << (OD_COEFF_SHIFT - coeff_shift - 1)) - 1;
+  for (i = 0; i < blk_size * blk_size; i++) {
+    out_int32[i] = AOM_SIGNED_SHL(out_int32[i], hbd_downshift);
+    dqcoeff_pvq[i] = (out_int32[i] + (out_int32[i] < 0) + rounding_mask) >>
+                     (OD_COEFF_SHIFT - coeff_shift);
+  }
+
+  od_coding_order_to_raster(dqcoeff, blk_size, tx_type, dqcoeff_pvq, blk_size);
+
+  eob = blk_size * blk_size;
+
+  return eob;
+}
+
+static PVQ_SKIP_TYPE read_pvq_skip(AV1_COMMON *cm, MACROBLOCKD *const xd,
+                                   int plane, TX_SIZE tx_size) {
+  // decode ac/dc coded flag. bit0: DC coded, bit1 : AC coded
+  // NOTE : we don't use 5 symbols for luma here in aom codebase,
+  // since block partition is taken care of by aom.
+  // So, only AC/DC skip info is coded
+  const int ac_dc_coded = aom_read_symbol(
+      xd->daala_dec.r,
+      xd->daala_dec.state.adapt->skip_cdf[2 * tx_size + (plane != 0)], 4,
+      "skip");
+  if (ac_dc_coded < 0 || ac_dc_coded > 3) {
+    aom_internal_error(&cm->error, AOM_CODEC_INVALID_PARAM,
+                       "Invalid PVQ Skip Type");
+  }
+  return ac_dc_coded;
+}
+
+static int av1_pvq_decode_helper2(AV1_COMMON *cm, MACROBLOCKD *const xd,
+                                  MB_MODE_INFO *const mbmi, int plane, int row,
+                                  int col, TX_SIZE tx_size, TX_TYPE tx_type) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  // transform block size in pixels
+  int tx_blk_size = tx_size_wide[tx_size];
+  int i, j;
+  tran_low_t *pvq_ref_coeff = pd->pvq_ref_coeff;
+  const int diff_stride = tx_blk_size;
+  int16_t *pred = pd->pred;
+  tran_low_t *const dqcoeff = pd->dqcoeff;
+  uint8_t *dst;
+  int eob;
+  const PVQ_SKIP_TYPE ac_dc_coded = read_pvq_skip(cm, xd, plane, tx_size);
+
+  eob = 0;
+  dst = &pd->dst.buf[4 * row * pd->dst.stride + 4 * col];
+
+  if (ac_dc_coded) {
+    int xdec = pd->subsampling_x;
+    int seg_id = mbmi->segment_id;
+    int16_t *quant;
+    FWD_TXFM_PARAM fwd_txfm_param;
+    // ToDo(yaowu): correct this with optimal number from decoding process.
+    const int max_scan_line = tx_size_2d[tx_size];
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      for (j = 0; j < tx_blk_size; j++)
+        for (i = 0; i < tx_blk_size; i++)
+          pred[diff_stride * j + i] =
+              CONVERT_TO_SHORTPTR(dst)[pd->dst.stride * j + i];
+    } else {
+#endif
+      for (j = 0; j < tx_blk_size; j++)
+        for (i = 0; i < tx_blk_size; i++)
+          pred[diff_stride * j + i] = dst[pd->dst.stride * j + i];
+#if CONFIG_HIGHBITDEPTH
+    }
+#endif
+
+    fwd_txfm_param.tx_type = tx_type;
+    fwd_txfm_param.tx_size = tx_size;
+    fwd_txfm_param.lossless = xd->lossless[seg_id];
+
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      fwd_txfm_param.bd = xd->bd;
+      av1_highbd_fwd_txfm(pred, pvq_ref_coeff, diff_stride, &fwd_txfm_param);
+    } else {
+#endif  // CONFIG_HIGHBITDEPTH
+      av1_fwd_txfm(pred, pvq_ref_coeff, diff_stride, &fwd_txfm_param);
+#if CONFIG_HIGHBITDEPTH
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+
+    quant = &pd->seg_dequant[seg_id][0];  // aom's quantizer
+
+    eob = av1_pvq_decode_helper(xd, pvq_ref_coeff, dqcoeff, quant, plane,
+                                tx_size, tx_type, xdec, ac_dc_coded);
+
+    inverse_transform_block(xd, plane, tx_type, tx_size, dst, pd->dst.stride,
+                            max_scan_line, eob);
+  }
+
+  return eob;
+}
+#endif
+
+static int get_block_idx(const MACROBLOCKD *xd, int plane, int row, int col) {
+  const int bsize = xd->mi[0]->mbmi.sb_type;
+  const struct macroblockd_plane *pd = &xd->plane[plane];
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+  const BLOCK_SIZE plane_bsize =
+      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif  // CONFIG_CHROMA_2X2
+#else
+  const BLOCK_SIZE plane_bsize =
+      get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
+#endif
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  const TX_SIZE tx_size = get_tx_size(plane, xd);
+  const uint8_t txh_unit = tx_size_high_unit[tx_size];
+  return row * max_blocks_wide + col * txh_unit;
+}
+
+static void predict_and_reconstruct_intra_block(
+    AV1_COMMON *cm, MACROBLOCKD *const xd, aom_reader *const r,
+    MB_MODE_INFO *const mbmi, int plane, int row, int col, TX_SIZE tx_size) {
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  const int block_idx = get_block_idx(xd, plane, row, col);
+#if CONFIG_PVQ
+  (void)r;
+#endif
+  av1_predict_intra_block_facade(xd, plane, block_idx, col, row, tx_size);
+
+  if (!mbmi->skip) {
+#if !CONFIG_PVQ
+    struct macroblockd_plane *const pd = &xd->plane[plane];
+#if CONFIG_LV_MAP
+    int16_t max_scan_line = 0;
+    int eob;
+    av1_read_coeffs_txb_facade(cm, xd, r, row, col, block_idx, plane,
+                               pd->dqcoeff, &max_scan_line, &eob);
+    // tx_type will be read out in av1_read_coeffs_txb_facade
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+#else   // CONFIG_LV_MAP
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+    const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 0);
+    int16_t max_scan_line = 0;
+    const int eob =
+        av1_decode_block_tokens(cm, xd, plane, scan_order, col, row, tx_size,
+                                tx_type, &max_scan_line, r, mbmi->segment_id);
+#endif  // CONFIG_LV_MAP
+    if (eob) {
+      uint8_t *dst =
+          &pd->dst.buf[(row * pd->dst.stride + col) << tx_size_wide_log2[0]];
+      inverse_transform_block(xd, plane, tx_type, tx_size, dst, pd->dst.stride,
+                              max_scan_line, eob);
+    }
+#else
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+    av1_pvq_decode_helper2(cm, xd, mbmi, plane, row, col, tx_size, tx_type);
+#endif
+  }
+#if CONFIG_CFL
+  if (plane == AOM_PLANE_Y) {
+    struct macroblockd_plane *const pd = &xd->plane[plane];
+    uint8_t *dst =
+        &pd->dst.buf[(row * pd->dst.stride + col) << tx_size_wide_log2[0]];
+    cfl_store(xd->cfl, dst, pd->dst.stride, row, col, tx_size);
+  }
+#endif
+}
+
+#if CONFIG_VAR_TX && !CONFIG_COEF_INTERLEAVE
+static void decode_reconstruct_tx(AV1_COMMON *cm, MACROBLOCKD *const xd,
+                                  aom_reader *r, MB_MODE_INFO *const mbmi,
+                                  int plane, BLOCK_SIZE plane_bsize,
+                                  int blk_row, int blk_col, TX_SIZE tx_size,
+                                  int *eob_total) {
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
+  const int tx_row = blk_row >> (1 - pd->subsampling_y);
+  const int tx_col = blk_col >> (1 - pd->subsampling_x);
+  const TX_SIZE plane_tx_size =
+      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
+            : mbmi->inter_tx_size[tx_row][tx_col];
+  // Scale to match transform block unit.
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  if (tx_size == plane_tx_size) {
+    PLANE_TYPE plane_type = get_plane_type(plane);
+    int block_idx = get_block_idx(xd, plane, blk_row, blk_col);
+#if CONFIG_LV_MAP
+    (void)segment_id;
+    int16_t max_scan_line = 0;
+    int eob;
+    av1_read_coeffs_txb_facade(cm, xd, r, row, col, block_idx, plane,
+                               pd->dqcoeff, &max_scan_line, &eob);
+    // tx_type will be read out in av1_read_coeffs_txb_facade
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, plane_tx_size);
+#else   // CONFIG_LV_MAP
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, plane_tx_size);
+    const SCAN_ORDER *sc = get_scan(cm, plane_tx_size, tx_type, 1);
+    int16_t max_scan_line = 0;
+    const int eob = av1_decode_block_tokens(
+        cm, xd, plane, sc, blk_col, blk_row, plane_tx_size, tx_type,
+        &max_scan_line, r, mbmi->segment_id);
+#endif  // CONFIG_LV_MAP
+    inverse_transform_block(xd, plane, tx_type, plane_tx_size,
+                            &pd->dst.buf[(blk_row * pd->dst.stride + blk_col)
+                                         << tx_size_wide_log2[0]],
+                            pd->dst.stride, max_scan_line, eob);
+    *eob_total += eob;
+  } else {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+
+    assert(bsl > 0);
+
+    for (i = 0; i < 4; ++i) {
+      const int offsetr = blk_row + (i >> 1) * bsl;
+      const int offsetc = blk_col + (i & 0x01) * bsl;
+
+      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
+
+      decode_reconstruct_tx(cm, xd, r, mbmi, plane, plane_bsize, offsetr,
+                            offsetc, sub_txs, eob_total);
+    }
+  }
+}
+#endif  // CONFIG_VAR_TX
+
+#if !CONFIG_VAR_TX || CONFIG_SUPERTX || CONFIG_COEF_INTERLEAVE || \
+    (!CONFIG_VAR_TX && CONFIG_EXT_TX && CONFIG_RECT_TX)
+static int reconstruct_inter_block(AV1_COMMON *cm, MACROBLOCKD *const xd,
+                                   aom_reader *const r, int segment_id,
+                                   int plane, int row, int col,
+                                   TX_SIZE tx_size) {
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  int block_idx = get_block_idx(xd, plane, row, col);
+#if CONFIG_PVQ
+  int eob;
+  (void)r;
+  (void)segment_id;
+#else
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+#endif
+
+#if !CONFIG_PVQ
+#if CONFIG_LV_MAP
+  (void)segment_id;
+  int16_t max_scan_line = 0;
+  int eob;
+  av1_read_coeffs_txb_facade(cm, xd, r, row, col, block_idx, plane, pd->dqcoeff,
+                             &max_scan_line, &eob);
+  // tx_type will be read out in av1_read_coeffs_txb_facade
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+#else   // CONFIG_LV_MAP
+  int16_t max_scan_line = 0;
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+  const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 1);
+  const int eob =
+      av1_decode_block_tokens(cm, xd, plane, scan_order, col, row, tx_size,
+                              tx_type, &max_scan_line, r, segment_id);
+#endif  // CONFIG_LV_MAP
+  uint8_t *dst =
+      &pd->dst.buf[(row * pd->dst.stride + col) << tx_size_wide_log2[0]];
+  if (eob)
+    inverse_transform_block(xd, plane, tx_type, tx_size, dst, pd->dst.stride,
+                            max_scan_line, eob);
+#else
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+  eob = av1_pvq_decode_helper2(cm, xd, &xd->mi[0]->mbmi, plane, row, col,
+                               tx_size, tx_type);
+#endif
+  return eob;
+}
+#endif  // !CONFIG_VAR_TX || CONFIG_SUPER_TX
+
+static void set_offsets(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                        BLOCK_SIZE bsize, int mi_row, int mi_col, int bw,
+                        int bh, int x_mis, int y_mis) {
+  const int offset = mi_row * cm->mi_stride + mi_col;
+  int x, y;
+  const TileInfo *const tile = &xd->tile;
+
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = &cm->mi[offset];
+  // TODO(slavarnway): Generate sb_type based on bwl and bhl, instead of
+  // passing bsize from decode_partition().
+  xd->mi[0]->mbmi.sb_type = bsize;
+#if CONFIG_RD_DEBUG
+  xd->mi[0]->mbmi.mi_row = mi_row;
+  xd->mi[0]->mbmi.mi_col = mi_col;
+#endif
+  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_plane_n4(xd, bw, bh);
+  set_skip_context(xd, mi_row, mi_col);
+
+#if CONFIG_VAR_TX
+  xd->max_tx_size = max_txsize_lookup[bsize];
+#endif
+
+  // 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,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                       mi_col);
+}
+
+#if CONFIG_SUPERTX
+static MB_MODE_INFO *set_offsets_extend(AV1_COMMON *const cm,
+                                        MACROBLOCKD *const xd,
+                                        const TileInfo *const tile,
+                                        BLOCK_SIZE bsize_pred, int mi_row_pred,
+                                        int mi_col_pred, int mi_row_ori,
+                                        int mi_col_ori) {
+  // Used in supertx
+  // (mi_row_ori, mi_col_ori): location for mv
+  // (mi_row_pred, mi_col_pred, bsize_pred): region to predict
+  const int bw = mi_size_wide[bsize_pred];
+  const int bh = mi_size_high[bsize_pred];
+  const int offset = mi_row_ori * cm->mi_stride + mi_col_ori;
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = cm->mi + offset;
+  set_mi_row_col(xd, tile, mi_row_pred, bh, mi_col_pred, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  xd->up_available = (mi_row_ori > tile->mi_row_start);
+  xd->left_available = (mi_col_ori > tile->mi_col_start);
+
+  set_plane_n4(xd, bw, bh);
+
+  return &xd->mi[0]->mbmi;
+}
+
+#if CONFIG_SUPERTX
+static MB_MODE_INFO *set_mb_offsets(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                    BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                    int bw, int bh, int x_mis, int y_mis) {
+  const int offset = mi_row * cm->mi_stride + mi_col;
+  const TileInfo *const tile = &xd->tile;
+  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_mi_row_col(xd, tile, mi_row, bh, mi_col, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+  return &xd->mi[0]->mbmi;
+}
+#endif
+
+static void set_offsets_topblock(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                 const TileInfo *const tile, BLOCK_SIZE bsize,
+                                 int mi_row, int mi_col) {
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int offset = mi_row * cm->mi_stride + mi_col;
+
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = cm->mi + offset;
+
+  set_plane_n4(xd, bw, bh);
+
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                       mi_col);
+}
+
+static void set_param_topblock(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                               BLOCK_SIZE bsize, int mi_row, int mi_col,
+                               int txfm, int skip) {
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = AOMMIN(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;
+
+  for (y = 0; y < y_mis; ++y)
+    for (x = 0; x < x_mis; ++x) {
+      xd->mi[y * cm->mi_stride + x]->mbmi.skip = skip;
+      xd->mi[y * cm->mi_stride + x]->mbmi.tx_type = txfm;
+    }
+#if CONFIG_VAR_TX
+  xd->above_txfm_context = cm->above_txfm_context + mi_col;
+  xd->left_txfm_context =
+      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+  set_txfm_ctxs(xd->mi[0]->mbmi.tx_size, bw, bh, skip, xd);
+#endif
+}
+
+static void set_ref(AV1_COMMON *const cm, MACROBLOCKD *const xd, int idx,
+                    int mi_row, int mi_col) {
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  RefBuffer *ref_buffer = &cm->frame_refs[mbmi->ref_frame[idx] - LAST_FRAME];
+  xd->block_refs[idx] = ref_buffer;
+  if (!av1_is_valid_scale(&ref_buffer->sf))
+    aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                       "Invalid scale factors");
+  av1_setup_pre_planes(xd, idx, ref_buffer->buf, mi_row, mi_col,
+                       &ref_buffer->sf);
+  aom_merge_corrupted_flag(&xd->corrupted, ref_buffer->buf->corrupted);
+}
+
+static void dec_predict_b_extend(
+    AV1Decoder *const pbi, MACROBLOCKD *const xd, const TileInfo *const tile,
+    int block, int mi_row_ori, int mi_col_ori, int mi_row_pred, int mi_col_pred,
+    int mi_row_top, int mi_col_top, uint8_t *dst_buf[3], int dst_stride[3],
+    BLOCK_SIZE bsize_top, BLOCK_SIZE bsize_pred, int b_sub8x8, int bextend) {
+  // Used in supertx
+  // (mi_row_ori, mi_col_ori): location for mv
+  // (mi_row_pred, mi_col_pred, bsize_pred): region to predict
+  // (mi_row_top, mi_col_top, bsize_top): region of the top partition size
+  // block: sub location of sub8x8 blocks
+  // b_sub8x8: 1: ori is sub8x8; 0: ori is not sub8x8
+  // bextend: 1: region to predict is an extension of ori; 0: not
+  int r = (mi_row_pred - mi_row_top) * MI_SIZE;
+  int c = (mi_col_pred - mi_col_top) * MI_SIZE;
+  const int mi_width_top = mi_size_wide[bsize_top];
+  const int mi_height_top = mi_size_high[bsize_top];
+  MB_MODE_INFO *mbmi;
+  AV1_COMMON *const cm = &pbi->common;
+
+  if (mi_row_pred < mi_row_top || mi_col_pred < mi_col_top ||
+      mi_row_pred >= mi_row_top + mi_height_top ||
+      mi_col_pred >= mi_col_top + mi_width_top || mi_row_pred >= cm->mi_rows ||
+      mi_col_pred >= cm->mi_cols)
+    return;
+
+  mbmi = set_offsets_extend(cm, xd, tile, bsize_pred, mi_row_pred, mi_col_pred,
+                            mi_row_ori, mi_col_ori);
+  set_ref(cm, xd, 0, mi_row_pred, mi_col_pred);
+  if (has_second_ref(&xd->mi[0]->mbmi))
+    set_ref(cm, xd, 1, mi_row_pred, mi_col_pred);
+
+  if (!bextend) mbmi->tx_size = max_txsize_lookup[bsize_top];
+
+  xd->plane[0].dst.stride = dst_stride[0];
+  xd->plane[1].dst.stride = dst_stride[1];
+  xd->plane[2].dst.stride = dst_stride[2];
+  xd->plane[0].dst.buf = dst_buf[0] +
+                         (r >> xd->plane[0].subsampling_y) * dst_stride[0] +
+                         (c >> xd->plane[0].subsampling_x);
+  xd->plane[1].dst.buf = dst_buf[1] +
+                         (r >> xd->plane[1].subsampling_y) * dst_stride[1] +
+                         (c >> xd->plane[1].subsampling_x);
+  xd->plane[2].dst.buf = dst_buf[2] +
+                         (r >> xd->plane[2].subsampling_y) * dst_stride[2] +
+                         (c >> xd->plane[2].subsampling_x);
+
+  if (!b_sub8x8)
+    av1_build_inter_predictors_sb_extend(xd,
+#if CONFIG_EXT_INTER
+                                         mi_row_ori, mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                         mi_row_pred, mi_col_pred, bsize_pred);
+  else
+    av1_build_inter_predictors_sb_sub8x8_extend(xd,
+#if CONFIG_EXT_INTER
+                                                mi_row_ori, mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                                mi_row_pred, mi_col_pred,
+                                                bsize_pred, block);
+}
+
+static void dec_extend_dir(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+                           const TileInfo *const tile, int block,
+                           BLOCK_SIZE bsize, BLOCK_SIZE top_bsize, int mi_row,
+                           int mi_col, int mi_row_top, int mi_col_top,
+                           uint8_t *dst_buf[3], int dst_stride[3], int dir) {
+  // dir: 0-lower, 1-upper, 2-left, 3-right
+  //      4-lowerleft, 5-upperleft, 6-lowerright, 7-upperright
+  const int mi_width = mi_size_wide[bsize];
+  const int mi_height = mi_size_high[bsize];
+  int xss = xd->plane[1].subsampling_x;
+  int yss = xd->plane[1].subsampling_y;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+  int b_sub8x8 = (bsize < BLOCK_8X8) && !unify_bsize ? 1 : 0;
+  BLOCK_SIZE extend_bsize;
+  int mi_row_pred, mi_col_pred;
+
+  int wide_unit, high_unit;
+  int i, j;
+  int ext_offset = 0;
+
+  if (dir == 0 || dir == 1) {
+    extend_bsize =
+        (mi_width == mi_size_wide[BLOCK_8X8] || bsize < BLOCK_8X8 || xss < yss)
+            ? BLOCK_8X8
+            : BLOCK_16X8;
+#if CONFIG_CB4X4
+    if (bsize < BLOCK_8X8) {
+      extend_bsize = BLOCK_4X4;
+      ext_offset = mi_size_wide[BLOCK_8X8];
+    }
+#endif
+
+    wide_unit = mi_size_wide[extend_bsize];
+    high_unit = mi_size_high[extend_bsize];
+
+    mi_row_pred = mi_row + ((dir == 0) ? mi_height : -(mi_height + ext_offset));
+    mi_col_pred = mi_col;
+
+    for (j = 0; j < mi_height + ext_offset; j += high_unit)
+      for (i = 0; i < mi_width + ext_offset; i += wide_unit)
+        dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col,
+                             mi_row_pred + j, mi_col_pred + i, mi_row_top,
+                             mi_col_top, dst_buf, dst_stride, top_bsize,
+                             extend_bsize, b_sub8x8, 1);
+  } else if (dir == 2 || dir == 3) {
+    extend_bsize =
+        (mi_height == mi_size_high[BLOCK_8X8] || bsize < BLOCK_8X8 || yss < xss)
+            ? BLOCK_8X8
+            : BLOCK_8X16;
+#if CONFIG_CB4X4
+    if (bsize < BLOCK_8X8) {
+      extend_bsize = BLOCK_4X4;
+      ext_offset = mi_size_wide[BLOCK_8X8];
+    }
+#endif
+
+    wide_unit = mi_size_wide[extend_bsize];
+    high_unit = mi_size_high[extend_bsize];
+
+    mi_row_pred = mi_row;
+    mi_col_pred = mi_col + ((dir == 3) ? mi_width : -(mi_width + ext_offset));
+
+    for (j = 0; j < mi_height + ext_offset; j += high_unit)
+      for (i = 0; i < mi_width + ext_offset; i += wide_unit)
+        dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col,
+                             mi_row_pred + j, mi_col_pred + i, mi_row_top,
+                             mi_col_top, dst_buf, dst_stride, top_bsize,
+                             extend_bsize, b_sub8x8, 1);
+  } else {
+    extend_bsize = BLOCK_8X8;
+#if CONFIG_CB4X4
+    if (bsize < BLOCK_8X8) {
+      extend_bsize = BLOCK_4X4;
+      ext_offset = mi_size_wide[BLOCK_8X8];
+    }
+#endif
+    wide_unit = mi_size_wide[extend_bsize];
+    high_unit = mi_size_high[extend_bsize];
+
+    mi_row_pred = mi_row + ((dir == 4 || dir == 6) ? mi_height
+                                                   : -(mi_height + ext_offset));
+    mi_col_pred =
+        mi_col + ((dir == 6 || dir == 7) ? mi_width : -(mi_width + ext_offset));
+
+    for (j = 0; j < mi_height + ext_offset; j += high_unit)
+      for (i = 0; i < mi_width + ext_offset; i += wide_unit)
+        dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col,
+                             mi_row_pred + j, mi_col_pred + i, mi_row_top,
+                             mi_col_top, dst_buf, dst_stride, top_bsize,
+                             extend_bsize, b_sub8x8, 1);
+  }
+}
+
+static void dec_extend_all(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+                           const TileInfo *const tile, int block,
+                           BLOCK_SIZE bsize, BLOCK_SIZE top_bsize, int mi_row,
+                           int mi_col, int mi_row_top, int mi_col_top,
+                           uint8_t *dst_buf[3], int dst_stride[3]) {
+  for (int i = 0; i < 8; ++i) {
+    dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
+                   mi_row_top, mi_col_top, dst_buf, dst_stride, i);
+  }
+}
+
+static void dec_predict_sb_complex(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+                                   const TileInfo *const tile, int mi_row,
+                                   int mi_col, int mi_row_top, int mi_col_top,
+                                   BLOCK_SIZE bsize, BLOCK_SIZE top_bsize,
+                                   uint8_t *dst_buf[3], int dst_stride[3]) {
+  const AV1_COMMON *const cm = &pbi->common;
+  const int hbs = mi_size_wide[bsize] / 2;
+  const PARTITION_TYPE partition = get_partition(cm, mi_row, mi_col, bsize);
+  const BLOCK_SIZE subsize = get_subsize(bsize, partition);
+#if CONFIG_EXT_PARTITION_TYPES
+  const BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+  int i;
+  const int mi_offset = mi_row * cm->mi_stride + mi_col;
+  uint8_t *dst_buf1[3], *dst_buf2[3], *dst_buf3[3];
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf3[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
+  int dst_stride1[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
+  int dst_stride2[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
+  int dst_stride3[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    int len = sizeof(uint16_t);
+    dst_buf1[0] = CONVERT_TO_BYTEPTR(tmp_buf1);
+    dst_buf1[1] = CONVERT_TO_BYTEPTR(tmp_buf1 + MAX_TX_SQUARE * len);
+    dst_buf1[2] = CONVERT_TO_BYTEPTR(tmp_buf1 + 2 * MAX_TX_SQUARE * len);
+    dst_buf2[0] = CONVERT_TO_BYTEPTR(tmp_buf2);
+    dst_buf2[1] = CONVERT_TO_BYTEPTR(tmp_buf2 + MAX_TX_SQUARE * len);
+    dst_buf2[2] = CONVERT_TO_BYTEPTR(tmp_buf2 + 2 * MAX_TX_SQUARE * len);
+    dst_buf3[0] = CONVERT_TO_BYTEPTR(tmp_buf3);
+    dst_buf3[1] = CONVERT_TO_BYTEPTR(tmp_buf3 + MAX_TX_SQUARE * len);
+    dst_buf3[2] = CONVERT_TO_BYTEPTR(tmp_buf3 + 2 * MAX_TX_SQUARE * len);
+  } else {
+#endif
+    dst_buf1[0] = tmp_buf1;
+    dst_buf1[1] = tmp_buf1 + MAX_TX_SQUARE;
+    dst_buf1[2] = tmp_buf1 + 2 * MAX_TX_SQUARE;
+    dst_buf2[0] = tmp_buf2;
+    dst_buf2[1] = tmp_buf2 + MAX_TX_SQUARE;
+    dst_buf2[2] = tmp_buf2 + 2 * MAX_TX_SQUARE;
+    dst_buf3[0] = tmp_buf3;
+    dst_buf3[1] = tmp_buf3 + MAX_TX_SQUARE;
+    dst_buf3[2] = tmp_buf3 + 2 * MAX_TX_SQUARE;
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  xd->mi = cm->mi_grid_visible + mi_offset;
+  xd->mi[0] = cm->mi + mi_offset;
+
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].dst.buf = dst_buf[i];
+    xd->plane[i].dst.stride = dst_stride[i];
+  }
+
+  switch (partition) {
+    case PARTITION_NONE:
+      assert(bsize < top_bsize);
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, bsize, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dst_buf, dst_stride);
+      break;
+    case PARTITION_HORZ:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        // For sub8x8, predict in 8x8 unit
+        // First half
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, BLOCK_8X8, 1, 0);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+
+        // Second half
+        dec_predict_b_extend(pbi, xd, tile, 2, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 2, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+        // weighted average to smooth the boundary
+        xd->plane[0].dst.buf = dst_buf[0];
+        xd->plane[0].dst.stride = dst_stride[0];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[0], dst_stride[0], dst_buf1[0], dst_stride1[0], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            0);
+      } else {
+        // First half
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, subsize, 0, 0);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+        else
+          dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride, 0);
+
+        if (mi_row + hbs < cm->mi_rows) {
+          // Second half
+          dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col,
+                               mi_row + hbs, mi_col, mi_row_top, mi_col_top,
+                               dst_buf1, dst_stride1, top_bsize, subsize, 0, 0);
+          if (bsize < top_bsize)
+            dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1);
+          else
+            dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, 1);
+
+          // weighted average to smooth the boundary
+          for (i = 0; i < MAX_MB_PLANE; i++) {
+            xd->plane[i].dst.buf = dst_buf[i];
+            xd->plane[i].dst.stride = dst_stride[i];
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_HORZ, i);
+          }
+        }
+      }
+      break;
+    case PARTITION_VERT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        // First half
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, BLOCK_8X8, 1, 0);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+
+        // Second half
+        dec_predict_b_extend(pbi, xd, tile, 1, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 1, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+        // Smooth
+        xd->plane[0].dst.buf = dst_buf[0];
+        xd->plane[0].dst.stride = dst_stride[0];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[0], dst_stride[0], dst_buf1[0], dst_stride1[0], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            0);
+      } else {
+        // First half
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, subsize, 0, 0);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+        else
+          dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride, 3);
+
+        // Second half
+        if (mi_col + hbs < cm->mi_cols) {
+          dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col + hbs, mi_row,
+                               mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                               dst_stride1, top_bsize, subsize, 0, 0);
+          if (bsize < top_bsize)
+            dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1);
+          else
+            dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, 2);
+
+          // Smooth
+          for (i = 0; i < MAX_MB_PLANE; i++) {
+            xd->plane[i].dst.buf = dst_buf[i];
+            xd->plane[i].dst.stride = dst_stride[i];
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_VERT, i);
+          }
+        }
+      }
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, BLOCK_8X8, 1, 0);
+        dec_predict_b_extend(pbi, xd, tile, 1, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        dec_predict_b_extend(pbi, xd, tile, 2, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf2, dst_stride2,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        dec_predict_b_extend(pbi, xd, tile, 3, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf3, dst_stride3,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        if (bsize < top_bsize) {
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+          dec_extend_all(pbi, xd, tile, 1, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+          dec_extend_all(pbi, xd, tile, 2, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf2, dst_stride2);
+          dec_extend_all(pbi, xd, tile, 3, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf3, dst_stride3);
+        }
+      } else {
+        dec_predict_sb_complex(pbi, xd, tile, mi_row, mi_col, mi_row_top,
+                               mi_col_top, subsize, top_bsize, dst_buf,
+                               dst_stride);
+        if (mi_row < cm->mi_rows && mi_col + hbs < cm->mi_cols)
+          dec_predict_sb_complex(pbi, xd, tile, mi_row, mi_col + hbs,
+                                 mi_row_top, mi_col_top, subsize, top_bsize,
+                                 dst_buf1, dst_stride1);
+        if (mi_row + hbs < cm->mi_rows && mi_col < cm->mi_cols)
+          dec_predict_sb_complex(pbi, xd, tile, mi_row + hbs, mi_col,
+                                 mi_row_top, mi_col_top, subsize, top_bsize,
+                                 dst_buf2, dst_stride2);
+        if (mi_row + hbs < cm->mi_rows && mi_col + hbs < cm->mi_cols)
+          dec_predict_sb_complex(pbi, xd, tile, mi_row + hbs, mi_col + hbs,
+                                 mi_row_top, mi_col_top, subsize, top_bsize,
+                                 dst_buf3, dst_stride3);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+#if !CONFIG_CB4X4
+        if (bsize == BLOCK_8X8 && i != 0)
+          continue;  // Skip <4x4 chroma smoothing
+#endif
+        if (mi_row < cm->mi_rows && mi_col + hbs < cm->mi_cols) {
+          av1_build_masked_inter_predictor_complex(
+              xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
+              mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+              PARTITION_VERT, i);
+          if (mi_row + hbs < cm->mi_rows) {
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf2[i], dst_stride2[i], dst_buf3[i], dst_stride3[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_VERT, i);
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_HORZ, i);
+          }
+        } else if (mi_row + hbs < cm->mi_rows && mi_col < cm->mi_cols) {
+          av1_build_masked_inter_predictor_complex(
+              xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i],
+              mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+              PARTITION_HORZ, i);
+        }
+      }
+      break;
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_HORZ_A:
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dst_buf, dst_stride);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col + hbs, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row, mi_col + hbs,
+                     mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf2,
+                           dst_stride2, top_bsize, subsize, 0, 0);
+      if (bsize < top_bsize)
+        dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
+                       mi_col, mi_row_top, mi_col_top, dst_buf2, dst_stride2);
+      else
+        dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
+                       mi_col, mi_row_top, mi_col_top, dst_buf2, dst_stride2,
+                       1);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            i);
+      }
+      break;
+    case PARTITION_VERT_A:
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dst_buf, dst_stride);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row + hbs, mi_col,
+                     mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col + hbs, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                           dst_stride2, top_bsize, subsize, 0, 0);
+      if (bsize < top_bsize)
+        dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row,
+                       mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                       dst_stride2);
+      else
+        dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row,
+                       mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                       dst_stride2, 2);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            i);
+      }
+      break;
+    case PARTITION_HORZ_B:
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, subsize, 0, 0);
+      if (bsize < top_bsize)
+        dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride);
+      else
+        dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride, 0);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row + hbs, mi_col,
+                     mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col + hbs,
+                           mi_row + hbs, mi_col + hbs, mi_row_top, mi_col_top,
+                           dst_buf2, dst_stride2, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row + hbs,
+                     mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                     dst_stride2);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf1[i];
+        xd->plane[i].dst.stride = dst_stride1[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf1[i], dst_stride1[i], dst_buf2[i], dst_stride2[i],
+            mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+            PARTITION_VERT, i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            i);
+      }
+      break;
+    case PARTITION_VERT_B:
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, subsize, 0, 0);
+      if (bsize < top_bsize)
+        dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride);
+      else
+        dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride, 3);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col + hbs, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row, mi_col + hbs,
+                     mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col + hbs,
+                           mi_row + hbs, mi_col + hbs, mi_row_top, mi_col_top,
+                           dst_buf2, dst_stride2, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row + hbs,
+                     mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                     dst_stride2);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf1[i];
+        xd->plane[i].dst.stride = dst_stride1[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf1[i], dst_stride1[i], dst_buf2[i], dst_stride2[i],
+            mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+            PARTITION_HORZ, i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            i);
+      }
+      break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+    default: assert(0);
+  }
+}
+
+static void set_segment_id_supertx(const AV1_COMMON *const cm, int mi_row,
+                                   int mi_col, BLOCK_SIZE bsize) {
+  const struct segmentation *seg = &cm->seg;
+  const int miw = AOMMIN(mi_size_wide[bsize], cm->mi_cols - mi_col);
+  const int mih = AOMMIN(mi_size_high[bsize], cm->mi_rows - mi_row);
+  const int mi_offset = mi_row * cm->mi_stride + mi_col;
+  MODE_INFO **const mip = cm->mi_grid_visible + mi_offset;
+  int r, c;
+  int seg_id_supertx = MAX_SEGMENTS;
+
+  if (!seg->enabled) {
+    seg_id_supertx = 0;
+  } else {
+    // Find the minimum segment_id
+    for (r = 0; r < mih; r++)
+      for (c = 0; c < miw; c++)
+        seg_id_supertx =
+            AOMMIN(mip[r * cm->mi_stride + c]->mbmi.segment_id, seg_id_supertx);
+    assert(0 <= seg_id_supertx && seg_id_supertx < MAX_SEGMENTS);
+  }
+
+  // Assign the the segment_id back to segment_id_supertx
+  for (r = 0; r < mih; r++)
+    for (c = 0; c < miw; c++)
+      mip[r * cm->mi_stride + c]->mbmi.segment_id_supertx = seg_id_supertx;
+}
+#endif  // CONFIG_SUPERTX
+
+static void decode_mbmi_block(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+#if CONFIG_SUPERTX
+                              int supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                              int mi_row, int mi_col, aom_reader *r,
+#if CONFIG_EXT_PARTITION_TYPES
+                              PARTITION_TYPE partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                              BLOCK_SIZE bsize) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = AOMMIN(bh, cm->mi_rows - mi_row);
+
+#if CONFIG_ACCOUNTING
+  aom_accounting_set_context(&pbi->accounting, mi_col, mi_row);
+#endif
+#if CONFIG_SUPERTX
+  if (supertx_enabled) {
+    set_mb_offsets(cm, xd, bsize, mi_row, mi_col, bw, bh, x_mis, y_mis);
+  } else {
+    set_offsets(cm, xd, bsize, mi_row, mi_col, bw, bh, x_mis, y_mis);
+  }
+#if CONFIG_EXT_PARTITION_TYPES
+  xd->mi[0]->mbmi.partition = partition;
+#endif
+  av1_read_mode_info(pbi, xd, supertx_enabled, mi_row, mi_col, r, x_mis, y_mis);
+#else
+  set_offsets(cm, xd, bsize, mi_row, mi_col, bw, bh, x_mis, y_mis);
+#if CONFIG_EXT_PARTITION_TYPES
+  xd->mi[0]->mbmi.partition = partition;
+#endif
+  av1_read_mode_info(pbi, xd, mi_row, mi_col, r, x_mis, y_mis);
+#endif  // CONFIG_SUPERTX
+
+  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)
+      aom_internal_error(xd->error_info, AOM_CODEC_CORRUPT_FRAME,
+                         "Invalid block size.");
+  }
+
+#if CONFIG_SUPERTX
+  xd->mi[0]->mbmi.segment_id_supertx = MAX_SEGMENTS;
+#endif  // CONFIG_SUPERTX
+
+  int reader_corrupted_flag = aom_reader_has_error(r);
+  aom_merge_corrupted_flag(&xd->corrupted, reader_corrupted_flag);
+}
+
+static void decode_token_and_recon_block(AV1Decoder *const pbi,
+                                         MACROBLOCKD *const xd, int mi_row,
+                                         int mi_col, aom_reader *r,
+                                         BLOCK_SIZE bsize) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = AOMMIN(bh, cm->mi_rows - mi_row);
+
+  set_offsets(cm, xd, bsize, mi_row, mi_col, bw, bh, x_mis, y_mis);
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+
+#if CONFIG_DELTA_Q
+  if (cm->delta_q_present_flag) {
+    int i;
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+#if CONFIG_EXT_DELTA_Q
+      xd->plane[0].seg_dequant[i][0] =
+          av1_dc_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->y_dc_delta_q, cm->bit_depth);
+      xd->plane[0].seg_dequant[i][1] = av1_ac_quant(
+          av1_get_qindex(&cm->seg, i, xd->current_qindex), 0, cm->bit_depth);
+      xd->plane[1].seg_dequant[i][0] =
+          av1_dc_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->uv_dc_delta_q, cm->bit_depth);
+      xd->plane[1].seg_dequant[i][1] =
+          av1_ac_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->uv_ac_delta_q, cm->bit_depth);
+      xd->plane[2].seg_dequant[i][0] =
+          av1_dc_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->uv_dc_delta_q, cm->bit_depth);
+      xd->plane[2].seg_dequant[i][1] =
+          av1_ac_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->uv_ac_delta_q, cm->bit_depth);
+#else
+      xd->plane[0].seg_dequant[i][0] =
+          av1_dc_quant(xd->current_qindex, cm->y_dc_delta_q, cm->bit_depth);
+      xd->plane[0].seg_dequant[i][1] =
+          av1_ac_quant(xd->current_qindex, 0, cm->bit_depth);
+      xd->plane[1].seg_dequant[i][0] =
+          av1_dc_quant(xd->current_qindex, cm->uv_dc_delta_q, cm->bit_depth);
+      xd->plane[1].seg_dequant[i][1] =
+          av1_ac_quant(xd->current_qindex, cm->uv_ac_delta_q, cm->bit_depth);
+      xd->plane[2].seg_dequant[i][0] =
+          av1_dc_quant(xd->current_qindex, cm->uv_dc_delta_q, cm->bit_depth);
+      xd->plane[2].seg_dequant[i][1] =
+          av1_ac_quant(xd->current_qindex, cm->uv_ac_delta_q, cm->bit_depth);
+#endif
+    }
+  }
+#endif
+
+#if CONFIG_CB4X4
+  if (mbmi->skip) reset_skip_context(xd, bsize);
+#else
+  if (mbmi->skip) reset_skip_context(xd, AOMMAX(BLOCK_8X8, bsize));
+#endif
+
+#if CONFIG_COEF_INTERLEAVE
+  {
+    const struct macroblockd_plane *const pd_y = &xd->plane[0];
+    const struct macroblockd_plane *const pd_c = &xd->plane[1];
+    const TX_SIZE tx_log2_y = mbmi->tx_size;
+    const TX_SIZE tx_log2_c = get_uv_tx_size(mbmi, pd_c);
+    const int tx_sz_y = (1 << tx_log2_y);
+    const int tx_sz_c = (1 << tx_log2_c);
+    const int num_4x4_w_y = pd_y->n4_w;
+    const int num_4x4_h_y = pd_y->n4_h;
+    const int num_4x4_w_c = pd_c->n4_w;
+    const int num_4x4_h_c = pd_c->n4_h;
+    const int max_4x4_w_y = get_max_4x4_size(num_4x4_w_y, xd->mb_to_right_edge,
+                                             pd_y->subsampling_x);
+    const int max_4x4_h_y = get_max_4x4_size(num_4x4_h_y, xd->mb_to_bottom_edge,
+                                             pd_y->subsampling_y);
+    const int max_4x4_w_c = get_max_4x4_size(num_4x4_w_c, xd->mb_to_right_edge,
+                                             pd_c->subsampling_x);
+    const int max_4x4_h_c = get_max_4x4_size(num_4x4_h_c, xd->mb_to_bottom_edge,
+                                             pd_c->subsampling_y);
+
+    // The max_4x4_w/h may be smaller than tx_sz under some corner cases,
+    // i.e. when the SB is splitted by tile boundaries.
+    const int tu_num_w_y = (max_4x4_w_y + tx_sz_y - 1) / tx_sz_y;
+    const int tu_num_h_y = (max_4x4_h_y + tx_sz_y - 1) / tx_sz_y;
+    const int tu_num_w_c = (max_4x4_w_c + tx_sz_c - 1) / tx_sz_c;
+    const int tu_num_h_c = (max_4x4_h_c + tx_sz_c - 1) / tx_sz_c;
+    const int tu_num_c = tu_num_w_c * tu_num_h_c;
+
+    if (!is_inter_block(mbmi)) {
+      int tu_idx_c = 0;
+      int row_y, col_y, row_c, col_c;
+      int plane;
+
+#if CONFIG_PALETTE
+      for (plane = 0; plane <= 1; ++plane) {
+        if (mbmi->palette_mode_info.palette_size[plane])
+          av1_decode_palette_tokens(xd, plane, r);
+      }
+#endif
+
+      for (row_y = 0; row_y < tu_num_h_y; row_y++) {
+        for (col_y = 0; col_y < tu_num_w_y; col_y++) {
+          // luma
+          predict_and_reconstruct_intra_block(
+              cm, xd, r, mbmi, 0, row_y * tx_sz_y, col_y * tx_sz_y, tx_log2_y);
+          // chroma
+          if (tu_idx_c < tu_num_c) {
+            row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+            col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+            predict_and_reconstruct_intra_block(cm, xd, r, mbmi, 1, row_c,
+                                                col_c, tx_log2_c);
+            predict_and_reconstruct_intra_block(cm, xd, r, mbmi, 2, row_c,
+                                                col_c, tx_log2_c);
+            tu_idx_c++;
+          }
+        }
+      }
+
+      // In 422 case, it's possilbe that Chroma has more TUs than Luma
+      while (tu_idx_c < tu_num_c) {
+        row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+        col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+        predict_and_reconstruct_intra_block(cm, xd, r, mbmi, 1, row_c, col_c,
+                                            tx_log2_c);
+        predict_and_reconstruct_intra_block(cm, xd, r, mbmi, 2, row_c, col_c,
+                                            tx_log2_c);
+        tu_idx_c++;
+      }
+    } else {
+      // Prediction
+      av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL,
+                                    AOMMAX(bsize, BLOCK_8X8));
+
+      // Reconstruction
+      if (!mbmi->skip) {
+        int eobtotal = 0;
+        int tu_idx_c = 0;
+        int row_y, col_y, row_c, col_c;
+
+        for (row_y = 0; row_y < tu_num_h_y; row_y++) {
+          for (col_y = 0; col_y < tu_num_w_y; col_y++) {
+            // luma
+            eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id, 0,
+                                                row_y * tx_sz_y,
+                                                col_y * tx_sz_y, tx_log2_y);
+            // chroma
+            if (tu_idx_c < tu_num_c) {
+              row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+              col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+              eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id,
+                                                  1, row_c, col_c, tx_log2_c);
+              eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id,
+                                                  2, row_c, col_c, tx_log2_c);
+              tu_idx_c++;
+            }
+          }
+        }
+
+        // In 422 case, it's possilbe that Chroma has more TUs than Luma
+        while (tu_idx_c < tu_num_c) {
+          row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+          col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+          eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id, 1,
+                                              row_c, col_c, tx_log2_c);
+          eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id, 2,
+                                              row_c, col_c, tx_log2_c);
+          tu_idx_c++;
+        }
+
+        // TODO(CONFIG_COEF_INTERLEAVE owners): bring eob == 0 corner case
+        // into line with the defaut configuration
+        if (bsize >= BLOCK_8X8 && eobtotal == 0) mbmi->skip = 1;
+      }
+    }
+  }
+#else  // CONFIG_COEF_INTERLEAVE
+  if (!is_inter_block(mbmi)) {
+    int plane;
+#if CONFIG_PALETTE
+    for (plane = 0; plane <= 1; ++plane) {
+      if (mbmi->palette_mode_info.palette_size[plane])
+        av1_decode_palette_tokens(xd, plane, r);
+    }
+#endif  // CONFIG_PALETTE
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+      const struct macroblockd_plane *const pd = &xd->plane[plane];
+      const TX_SIZE tx_size = get_tx_size(plane, xd);
+      const int stepr = tx_size_high_unit[tx_size];
+      const int stepc = tx_size_wide_unit[tx_size];
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+      const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+      const BLOCK_SIZE plane_bsize =
+          AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif  // CONFIG_CHROMA_2X2
+#else
+      const BLOCK_SIZE plane_bsize =
+          get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
+#endif
+      int row, col;
+      const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+      const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+#if CONFIG_CB4X4
+      if (!is_chroma_reference(mi_row, mi_col, bsize, pd->subsampling_x,
+                               pd->subsampling_y))
+        continue;
+#endif
+
+      for (row = 0; row < max_blocks_high; row += stepr)
+        for (col = 0; col < max_blocks_wide; col += stepc)
+          predict_and_reconstruct_intra_block(cm, xd, r, mbmi, plane, row, col,
+                                              tx_size);
+    }
+  } else {
+    int ref;
+
+    for (ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
+      const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
+      if (frame < LAST_FRAME) {
+#if CONFIG_INTRABC
+        assert(is_intrabc_block(mbmi));
+        assert(frame == INTRA_FRAME);
+        assert(ref == 0);
+#else
+        assert(0);
+#endif  // CONFIG_INTRABC
+      } else {
+        RefBuffer *ref_buf = &cm->frame_refs[frame - LAST_FRAME];
+
+        xd->block_refs[ref] = ref_buf;
+        if ((!av1_is_valid_scale(&ref_buf->sf)))
+          aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
+                             "Reference frame has invalid dimensions");
+        av1_setup_pre_planes(xd, ref, ref_buf->buf, mi_row, mi_col,
+                             &ref_buf->sf);
+      }
+    }
+
+#if CONFIG_CB4X4
+    av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL, bsize);
+#else
+    av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL,
+                                  AOMMAX(bsize, BLOCK_8X8));
+#endif
+
+#if CONFIG_MOTION_VAR
+    if (mbmi->motion_mode == OBMC_CAUSAL) {
+#if CONFIG_NCOBMC
+      av1_build_ncobmc_inter_predictors_sb(cm, xd, mi_row, mi_col);
+#else
+      av1_build_obmc_inter_predictors_sb(cm, xd, mi_row, mi_col);
+#endif
+    }
+#endif  // CONFIG_MOTION_VAR
+
+    // Reconstruction
+    if (!mbmi->skip) {
+      int eobtotal = 0;
+      int plane;
+
+      for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+        const struct macroblockd_plane *const pd = &xd->plane[plane];
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+        const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+        const BLOCK_SIZE plane_bsize =
+            AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif  // CONFIG_CHROMA_2X2
+#else
+        const BLOCK_SIZE plane_bsize =
+            get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
+#endif
+        const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+        const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+        int row, col;
+
+#if CONFIG_CB4X4
+        if (!is_chroma_reference(mi_row, mi_col, bsize, pd->subsampling_x,
+                                 pd->subsampling_y))
+          continue;
+#endif
+
+#if CONFIG_VAR_TX
+        const TX_SIZE max_tx_size = get_vartx_max_txsize(mbmi, plane_bsize);
+        const int bh_var_tx = tx_size_high_unit[max_tx_size];
+        const int bw_var_tx = tx_size_wide_unit[max_tx_size];
+        for (row = 0; row < max_blocks_high; row += bh_var_tx)
+          for (col = 0; col < max_blocks_wide; col += bw_var_tx)
+            decode_reconstruct_tx(cm, xd, r, mbmi, plane, plane_bsize, row, col,
+                                  max_tx_size, &eobtotal);
+#else
+        const TX_SIZE tx_size = get_tx_size(plane, xd);
+        const int stepr = tx_size_high_unit[tx_size];
+        const int stepc = tx_size_wide_unit[tx_size];
+        for (row = 0; row < max_blocks_high; row += stepr)
+          for (col = 0; col < max_blocks_wide; col += stepc)
+            eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id,
+                                                plane, row, col, tx_size);
+#endif
+      }
+    }
+  }
+#endif  // CONFIG_COEF_INTERLEAVE
+
+  int reader_corrupted_flag = aom_reader_has_error(r);
+  aom_merge_corrupted_flag(&xd->corrupted, reader_corrupted_flag);
+}
+
+#if CONFIG_NCOBMC && CONFIG_MOTION_VAR
+static void detoken_and_recon_sb(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+                                 int mi_row, int mi_col, aom_reader *r,
+                                 BLOCK_SIZE bsize) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int hbs = mi_size_wide[bsize] >> 1;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+  BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+  const int has_rows = (mi_row + hbs) < cm->mi_rows;
+  const int has_cols = (mi_col + hbs) < cm->mi_cols;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  partition = get_partition(cm, mi_row, mi_col, bsize);
+  subsize = subsize_lookup[partition][bsize];
+
+  if (!hbs && !unify_bsize) {
+    xd->bmode_blocks_wl = 1 >> !!(partition & PARTITION_VERT);
+    xd->bmode_blocks_hl = 1 >> !!(partition & PARTITION_HORZ);
+    decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, bsize);
+        break;
+      case PARTITION_HORZ:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+        if (has_rows)
+          decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col, r,
+                                       subsize);
+        break;
+      case PARTITION_VERT:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+        if (has_cols)
+          decode_token_and_recon_block(pbi, xd, mi_row, mi_col + hbs, r,
+                                       subsize);
+        break;
+      case PARTITION_SPLIT:
+        detoken_and_recon_sb(pbi, xd, mi_row, mi_col, r, subsize);
+        detoken_and_recon_sb(pbi, xd, mi_row, mi_col + hbs, r, subsize);
+        detoken_and_recon_sb(pbi, xd, mi_row + hbs, mi_col, r, subsize);
+        detoken_and_recon_sb(pbi, xd, mi_row + hbs, mi_col + hbs, r, subsize);
+        break;
+#if CONFIG_EXT_PARTITION_TYPES
+      case PARTITION_HORZ_A:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col + hbs, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col, r, subsize);
+        break;
+      case PARTITION_HORZ_B:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col + hbs, r,
+                                     bsize2);
+        break;
+      case PARTITION_VERT_A:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col + hbs, r, subsize);
+        break;
+      case PARTITION_VERT_B:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col + hbs, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col + hbs, r,
+                                     bsize2);
+        break;
+#endif
+      default: assert(0 && "Invalid partition type");
+    }
+  }
+}
+#endif
+
+static void decode_block(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+#if CONFIG_SUPERTX
+                         int supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         int mi_row, int mi_col, aom_reader *r,
+#if CONFIG_EXT_PARTITION_TYPES
+                         PARTITION_TYPE partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                         BLOCK_SIZE bsize) {
+  decode_mbmi_block(pbi, xd,
+#if CONFIG_SUPERTX
+                    supertx_enabled,
+#endif
+                    mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                    partition,
+#endif
+                    bsize);
+#if !(CONFIG_MOTION_VAR && CONFIG_NCOBMC)
+#if CONFIG_SUPERTX
+  if (!supertx_enabled)
+#endif  // CONFIG_SUPERTX
+    decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, bsize);
+#endif
+}
+
+static PARTITION_TYPE read_partition(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                     int mi_row, int mi_col, aom_reader *r,
+                                     int has_rows, int has_cols,
+                                     BLOCK_SIZE bsize) {
+#if CONFIG_UNPOISON_PARTITION_CTX
+  const int ctx =
+      partition_plane_context(xd, mi_row, mi_col, has_rows, has_cols, bsize);
+  const aom_prob *const probs =
+      ctx < PARTITION_CONTEXTS ? cm->fc->partition_prob[ctx] : NULL;
+  FRAME_COUNTS *const counts = ctx < PARTITION_CONTEXTS ? xd->counts : NULL;
+#else
+  const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
+  const aom_prob *const probs = cm->fc->partition_prob[ctx];
+  FRAME_COUNTS *const counts = xd->counts;
+#endif
+  PARTITION_TYPE p;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob *partition_cdf = (ctx >= 0) ? ec_ctx->partition_cdf[ctx] : NULL;
+#endif
+
+  if (has_rows && has_cols)
+#if CONFIG_EXT_PARTITION_TYPES
+    if (bsize <= BLOCK_8X8)
+#if CONFIG_EC_MULTISYMBOL
+      p = (PARTITION_TYPE)aom_read_symbol(r, partition_cdf, PARTITION_TYPES,
+                                          ACCT_STR);
+#else
+      p = (PARTITION_TYPE)aom_read_tree(r, av1_partition_tree, probs, ACCT_STR);
+#endif
+    else
+#if CONFIG_EC_MULTISYMBOL
+      p = (PARTITION_TYPE)aom_read_symbol(r, partition_cdf, EXT_PARTITION_TYPES,
+                                          ACCT_STR);
+#else
+      p = (PARTITION_TYPE)aom_read_tree(r, av1_ext_partition_tree, probs,
+                                        ACCT_STR);
+#endif
+#else
+#if CONFIG_EC_MULTISYMBOL
+    p = (PARTITION_TYPE)aom_read_symbol(r, partition_cdf, PARTITION_TYPES,
+                                        ACCT_STR);
+#else
+    p = (PARTITION_TYPE)aom_read_tree(r, av1_partition_tree, probs, ACCT_STR);
+#endif
+#endif  // CONFIG_EXT_PARTITION_TYPES
+  else if (!has_rows && has_cols)
+    p = aom_read(r, probs[1], ACCT_STR) ? PARTITION_SPLIT : PARTITION_HORZ;
+  else if (has_rows && !has_cols)
+    p = aom_read(r, probs[2], ACCT_STR) ? PARTITION_SPLIT : PARTITION_VERT;
+  else
+    p = PARTITION_SPLIT;
+
+  if (counts) ++counts->partition[ctx][p];
+
+  return p;
+}
+
+#if CONFIG_SUPERTX
+static int read_skip(AV1_COMMON *cm, const MACROBLOCKD *xd, int segment_id,
+                     aom_reader *r) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
+    return 1;
+  } else {
+    const int ctx = av1_get_skip_context(xd);
+    const int skip = aom_read(r, cm->fc->skip_probs[ctx], ACCT_STR);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts) ++counts->skip[ctx][skip];
+    return skip;
+  }
+}
+#endif  // CONFIG_SUPERTX
+
+// TODO(slavarnway): eliminate bsize and subsize in future commits
+static void decode_partition(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+#if CONFIG_SUPERTX
+                             int supertx_enabled,
+#endif
+                             int mi_row, int mi_col, aom_reader *r,
+                             BLOCK_SIZE bsize, int n4x4_l2) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int n8x8_l2 = n4x4_l2 - 1;
+  const int num_8x8_wh = mi_size_wide[bsize];
+  const int hbs = num_8x8_wh >> 1;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+#if CONFIG_EXT_PARTITION_TYPES
+  BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+  const int has_rows = (mi_row + hbs) < cm->mi_rows;
+  const int has_cols = (mi_col + hbs) < cm->mi_cols;
+#if CONFIG_SUPERTX
+  const int read_token = !supertx_enabled;
+  int skip = 0;
+  TX_SIZE supertx_size = max_txsize_lookup[bsize];
+  const TileInfo *const tile = &xd->tile;
+  int txfm = DCT_DCT;
+#endif  // CONFIG_SUPERTX
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  partition = (bsize < BLOCK_8X8) ? PARTITION_NONE
+                                  : read_partition(cm, xd, mi_row, mi_col, r,
+                                                   has_rows, has_cols, bsize);
+  subsize = subsize_lookup[partition][bsize];  // get_subsize(bsize, partition);
+
+#if CONFIG_PVQ
+  assert(partition < PARTITION_TYPES);
+  assert(subsize < BLOCK_SIZES);
+#endif
+#if CONFIG_SUPERTX
+  if (!frame_is_intra_only(cm) && partition != PARTITION_NONE &&
+      bsize <= MAX_SUPERTX_BLOCK_SIZE && !supertx_enabled && !xd->lossless[0]) {
+    const int supertx_context = partition_supertx_context_lookup[partition];
+    supertx_enabled = aom_read(
+        r, cm->fc->supertx_prob[supertx_context][supertx_size], ACCT_STR);
+    if (xd->counts)
+      xd->counts->supertx[supertx_context][supertx_size][supertx_enabled]++;
+#if CONFIG_VAR_TX
+    if (supertx_enabled) xd->supertx_size = supertx_size;
+#endif
+  }
+#endif  // CONFIG_SUPERTX
+  if (!hbs && !unify_bsize) {
+    // calculate bmode block dimensions (log 2)
+    xd->bmode_blocks_wl = 1 >> !!(partition & PARTITION_VERT);
+    xd->bmode_blocks_hl = 1 >> !!(partition & PARTITION_HORZ);
+    decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                 supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                 mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                 partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                 subsize);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                     mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                     partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                     subsize);
+        break;
+      case PARTITION_HORZ:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                     mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                     partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                     subsize);
+        if (has_rows)
+          decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                       supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                       mi_row + hbs, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                       partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                       subsize);
+        break;
+      case PARTITION_VERT:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                     mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                     partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                     subsize);
+        if (has_cols)
+          decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                       supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                       mi_row, mi_col + hbs, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                       partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                       subsize);
+        break;
+      case PARTITION_SPLIT:
+        decode_partition(pbi, xd,
+#if CONFIG_SUPERTX
+                         supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         mi_row, mi_col, r, subsize, n8x8_l2);
+        decode_partition(pbi, xd,
+#if CONFIG_SUPERTX
+                         supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         mi_row, mi_col + hbs, r, subsize, n8x8_l2);
+        decode_partition(pbi, xd,
+#if CONFIG_SUPERTX
+                         supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         mi_row + hbs, mi_col, r, subsize, n8x8_l2);
+        decode_partition(pbi, xd,
+#if CONFIG_SUPERTX
+                         supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         mi_row + hbs, mi_col + hbs, r, subsize, n8x8_l2);
+        break;
+#if CONFIG_EXT_PARTITION_TYPES
+      case PARTITION_HORZ_A:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col + hbs, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col, r, partition, subsize);
+        break;
+      case PARTITION_HORZ_B:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col, r, partition, subsize);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col + hbs, r, partition, bsize2);
+        break;
+      case PARTITION_VERT_A:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col + hbs, r, partition, subsize);
+        break;
+      case PARTITION_VERT_B:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col, r, partition, subsize);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col + hbs, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col + hbs, r, partition, bsize2);
+        break;
+#endif
+      default: assert(0 && "Invalid partition type");
+    }
+  }
+
+#if CONFIG_SUPERTX
+  if (supertx_enabled && read_token) {
+    uint8_t *dst_buf[3];
+    int dst_stride[3], i;
+    int offset = mi_row * cm->mi_stride + mi_col;
+
+    set_segment_id_supertx(cm, mi_row, mi_col, bsize);
+
+#if CONFIG_DELTA_Q
+    if (cm->delta_q_present_flag) {
+      for (i = 0; i < MAX_SEGMENTS; i++) {
+        xd->plane[0].seg_dequant[i][0] =
+            av1_dc_quant(xd->current_qindex, cm->y_dc_delta_q, cm->bit_depth);
+        xd->plane[0].seg_dequant[i][1] =
+            av1_ac_quant(xd->current_qindex, 0, cm->bit_depth);
+        xd->plane[1].seg_dequant[i][0] =
+            av1_dc_quant(xd->current_qindex, cm->uv_dc_delta_q, cm->bit_depth);
+        xd->plane[1].seg_dequant[i][1] =
+            av1_ac_quant(xd->current_qindex, cm->uv_ac_delta_q, cm->bit_depth);
+        xd->plane[2].seg_dequant[i][0] =
+            av1_dc_quant(xd->current_qindex, cm->uv_dc_delta_q, cm->bit_depth);
+        xd->plane[2].seg_dequant[i][1] =
+            av1_ac_quant(xd->current_qindex, cm->uv_ac_delta_q, cm->bit_depth);
+      }
+    }
+#endif
+
+    xd->mi = cm->mi_grid_visible + offset;
+    xd->mi[0] = cm->mi + offset;
+    set_mi_row_col(xd, tile, mi_row, mi_size_high[bsize], mi_col,
+                   mi_size_wide[bsize],
+#if CONFIG_DEPENDENT_HORZTILES
+                   cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                   cm->mi_rows, cm->mi_cols);
+    set_skip_context(xd, mi_row, mi_col);
+    skip = read_skip(cm, xd, xd->mi[0]->mbmi.segment_id_supertx, r);
+    if (skip) {
+      reset_skip_context(xd, bsize);
+    } else {
+#if CONFIG_EXT_TX
+      if (get_ext_tx_types(supertx_size, bsize, 1, cm->reduced_tx_set_used) >
+          1) {
+        const int eset =
+            get_ext_tx_set(supertx_size, bsize, 1, cm->reduced_tx_set_used);
+        if (eset > 0) {
+          txfm = aom_read_tree(r, av1_ext_tx_inter_tree[eset],
+                               cm->fc->inter_ext_tx_prob[eset][supertx_size],
+                               ACCT_STR);
+          if (xd->counts) ++xd->counts->inter_ext_tx[eset][supertx_size][txfm];
+        }
+      }
+#else
+      if (supertx_size < TX_32X32) {
+        txfm = aom_read_tree(r, av1_ext_tx_tree,
+                             cm->fc->inter_ext_tx_prob[supertx_size], ACCT_STR);
+        if (xd->counts) ++xd->counts->inter_ext_tx[supertx_size][txfm];
+      }
+#endif  // CONFIG_EXT_TX
+    }
+
+    av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                         mi_col);
+    for (i = 0; i < MAX_MB_PLANE; i++) {
+      dst_buf[i] = xd->plane[i].dst.buf;
+      dst_stride[i] = xd->plane[i].dst.stride;
+    }
+    dec_predict_sb_complex(pbi, xd, tile, mi_row, mi_col, mi_row, mi_col, bsize,
+                           bsize, dst_buf, dst_stride);
+
+    if (!skip) {
+      int eobtotal = 0;
+      MB_MODE_INFO *mbmi;
+      set_offsets_topblock(cm, xd, tile, bsize, mi_row, mi_col);
+      mbmi = &xd->mi[0]->mbmi;
+      mbmi->tx_type = txfm;
+      assert(mbmi->segment_id_supertx != MAX_SEGMENTS);
+      for (i = 0; i < MAX_MB_PLANE; ++i) {
+        const struct macroblockd_plane *const pd = &xd->plane[i];
+        int row, col;
+        const TX_SIZE tx_size = get_tx_size(i, xd);
+        const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+        const int stepr = tx_size_high_unit[tx_size];
+        const int stepc = tx_size_wide_unit[tx_size];
+        const int max_blocks_wide = max_block_wide(xd, plane_bsize, i);
+        const int max_blocks_high = max_block_high(xd, plane_bsize, i);
+
+        for (row = 0; row < max_blocks_high; row += stepr)
+          for (col = 0; col < max_blocks_wide; col += stepc)
+            eobtotal += reconstruct_inter_block(
+                cm, xd, r, mbmi->segment_id_supertx, i, row, col, tx_size);
+      }
+      if ((unify_bsize || !(subsize < BLOCK_8X8)) && eobtotal == 0) skip = 1;
+    }
+    set_param_topblock(cm, xd, bsize, mi_row, mi_col, txfm, skip);
+  }
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_EXT_PARTITION_TYPES
+  update_ext_partition_context(xd, mi_row, mi_col, subsize, bsize, partition);
+#else
+  // update partition context
+  if (bsize >= BLOCK_8X8 &&
+      (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+#if CONFIG_CDEF
+#if CONFIG_EXT_PARTITION
+  if (cm->sb_size == BLOCK_128X128 && bsize == BLOCK_128X128) {
+    if (!sb_all_skip(cm, mi_row, mi_col)) {
+      cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]->mbmi.cdef_strength =
+          aom_read_literal(r, cm->cdef_bits, ACCT_STR);
+    } else {
+      cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]->mbmi.cdef_strength =
+          0;
+    }
+  } else if (cm->sb_size == BLOCK_64X64 && bsize == BLOCK_64X64) {
+#else
+  if (bsize == BLOCK_64X64) {
+#endif
+    if (!sb_all_skip(cm, mi_row, mi_col)) {
+      cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]->mbmi.cdef_strength =
+          aom_read_literal(r, cm->cdef_bits, ACCT_STR);
+    } else {
+      cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]->mbmi.cdef_strength =
+          -1;
+    }
+  }
+#endif  // CONFIG_CDEF
+}
+
+static void setup_bool_decoder(const uint8_t *data, const uint8_t *data_end,
+                               const size_t read_size,
+                               struct aom_internal_error_info *error_info,
+                               aom_reader *r,
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+                               int window_size,
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+                               aom_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))
+    aom_internal_error(error_info, AOM_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt tile length");
+
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  r->window_size = window_size;
+#endif
+  if (aom_reader_init(r, data, read_size, decrypt_cb, decrypt_state))
+    aom_internal_error(error_info, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate bool decoder %d", 1);
+}
+
+#if !CONFIG_PVQ && !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET) && !CONFIG_LV_MAP
+static void read_coef_probs_common(av1_coeff_probs_model *coef_probs,
+                                   aom_reader *r) {
+  int i, j, k, l, m;
+#if CONFIG_EC_ADAPT
+  const int node_limit = UNCONSTRAINED_NODES - 1;
+#else
+  const int node_limit = UNCONSTRAINED_NODES;
+#endif
+
+  if (aom_read_bit(r, ACCT_STR))
+    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 < node_limit; ++m)
+              av1_diff_update_prob(r, &coef_probs[i][j][k][l][m], ACCT_STR);
+}
+
+static void read_coef_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode, aom_reader *r) {
+  const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+  TX_SIZE tx_size;
+  for (tx_size = 0; tx_size <= max_tx_size; ++tx_size)
+    read_coef_probs_common(fc->coef_probs[tx_size], r);
+}
+#endif
+
+static void setup_segmentation(AV1_COMMON *const cm,
+                               struct aom_read_bit_buffer *rb) {
+  struct segmentation *const seg = &cm->seg;
+  int i, j;
+
+  seg->update_map = 0;
+  seg->update_data = 0;
+
+  seg->enabled = aom_rb_read_bit(rb);
+  if (!seg->enabled) return;
+
+  // Segmentation map update
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
+    seg->update_map = 1;
+  } else {
+    seg->update_map = aom_rb_read_bit(rb);
+  }
+  if (seg->update_map) {
+    if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
+      seg->temporal_update = 0;
+    } else {
+      seg->temporal_update = aom_rb_read_bit(rb);
+    }
+  }
+
+  // Segmentation data update
+  seg->update_data = aom_rb_read_bit(rb);
+  if (seg->update_data) {
+    seg->abs_delta = aom_rb_read_bit(rb);
+
+    av1_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 = aom_rb_read_bit(rb);
+        if (feature_enabled) {
+          av1_enable_segfeature(seg, i, j);
+          data = decode_unsigned_max(rb, av1_seg_feature_data_max(j));
+          if (av1_is_segfeature_signed(j))
+            data = aom_rb_read_bit(rb) ? -data : data;
+        }
+        av1_set_segdata(seg, i, j, data);
+      }
+    }
+  }
+}
+
+#if CONFIG_LOOP_RESTORATION
+static void decode_restoration_mode(AV1_COMMON *cm,
+                                    struct aom_read_bit_buffer *rb) {
+  int p;
+  RestorationInfo *rsi = &cm->rst_info[0];
+  if (aom_rb_read_bit(rb)) {
+    rsi->frame_restoration_type =
+        aom_rb_read_bit(rb) ? RESTORE_SGRPROJ : RESTORE_WIENER;
+  } else {
+    rsi->frame_restoration_type =
+        aom_rb_read_bit(rb) ? RESTORE_SWITCHABLE : RESTORE_NONE;
+  }
+  for (p = 1; p < MAX_MB_PLANE; ++p) {
+    cm->rst_info[p].frame_restoration_type =
+        aom_rb_read_bit(rb) ? RESTORE_WIENER : RESTORE_NONE;
+  }
+
+  cm->rst_info[0].restoration_tilesize = RESTORATION_TILESIZE_MAX;
+  cm->rst_info[1].restoration_tilesize = RESTORATION_TILESIZE_MAX;
+  cm->rst_info[2].restoration_tilesize = RESTORATION_TILESIZE_MAX;
+  if (cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[2].frame_restoration_type != RESTORE_NONE) {
+    rsi = &cm->rst_info[0];
+    rsi->restoration_tilesize >>= aom_rb_read_bit(rb);
+    if (rsi->restoration_tilesize != RESTORATION_TILESIZE_MAX) {
+      rsi->restoration_tilesize >>= aom_rb_read_bit(rb);
+    }
+    cm->rst_info[1].restoration_tilesize = cm->rst_info[0].restoration_tilesize;
+    cm->rst_info[2].restoration_tilesize = cm->rst_info[0].restoration_tilesize;
+  }
+}
+
+static void read_wiener_filter(WienerInfo *wiener_info,
+                               WienerInfo *ref_wiener_info, aom_reader *rb) {
+  wiener_info->vfilter[0] = wiener_info->vfilter[WIENER_WIN - 1] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
+          WIENER_FILT_TAP0_SUBEXP_K,
+          ref_wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV) +
+      WIENER_FILT_TAP0_MINV;
+  wiener_info->vfilter[1] = wiener_info->vfilter[WIENER_WIN - 2] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
+          WIENER_FILT_TAP1_SUBEXP_K,
+          ref_wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV) +
+      WIENER_FILT_TAP1_MINV;
+  wiener_info->vfilter[2] = wiener_info->vfilter[WIENER_WIN - 3] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
+          WIENER_FILT_TAP2_SUBEXP_K,
+          ref_wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV) +
+      WIENER_FILT_TAP2_MINV;
+  // The central element has an implicit +WIENER_FILT_STEP
+  wiener_info->vfilter[WIENER_HALFWIN] =
+      -2 * (wiener_info->vfilter[0] + wiener_info->vfilter[1] +
+            wiener_info->vfilter[2]);
+
+  wiener_info->hfilter[0] = wiener_info->hfilter[WIENER_WIN - 1] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
+          WIENER_FILT_TAP0_SUBEXP_K,
+          ref_wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV) +
+      WIENER_FILT_TAP0_MINV;
+  wiener_info->hfilter[1] = wiener_info->hfilter[WIENER_WIN - 2] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
+          WIENER_FILT_TAP1_SUBEXP_K,
+          ref_wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV) +
+      WIENER_FILT_TAP1_MINV;
+  wiener_info->hfilter[2] = wiener_info->hfilter[WIENER_WIN - 3] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
+          WIENER_FILT_TAP2_SUBEXP_K,
+          ref_wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV) +
+      WIENER_FILT_TAP2_MINV;
+  // The central element has an implicit +WIENER_FILT_STEP
+  wiener_info->hfilter[WIENER_HALFWIN] =
+      -2 * (wiener_info->hfilter[0] + wiener_info->hfilter[1] +
+            wiener_info->hfilter[2]);
+  memcpy(ref_wiener_info, wiener_info, sizeof(*wiener_info));
+}
+
+static void read_sgrproj_filter(SgrprojInfo *sgrproj_info,
+                                SgrprojInfo *ref_sgrproj_info, aom_reader *rb) {
+  sgrproj_info->ep = aom_read_literal(rb, SGRPROJ_PARAMS_BITS, ACCT_STR);
+  sgrproj_info->xqd[0] =
+      aom_read_primitive_refsubexpfin(
+          rb, SGRPROJ_PRJ_MAX0 - SGRPROJ_PRJ_MIN0 + 1, SGRPROJ_PRJ_SUBEXP_K,
+          ref_sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0) +
+      SGRPROJ_PRJ_MIN0;
+  sgrproj_info->xqd[1] =
+      aom_read_primitive_refsubexpfin(
+          rb, SGRPROJ_PRJ_MAX1 - SGRPROJ_PRJ_MIN1 + 1, SGRPROJ_PRJ_SUBEXP_K,
+          ref_sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1) +
+      SGRPROJ_PRJ_MIN1;
+  memcpy(ref_sgrproj_info, sgrproj_info, sizeof(*sgrproj_info));
+}
+
+static void decode_restoration(AV1_COMMON *cm, aom_reader *rb) {
+  int i, p;
+  SgrprojInfo ref_sgrproj_info;
+  WienerInfo ref_wiener_info;
+  set_default_wiener(&ref_wiener_info);
+  set_default_sgrproj(&ref_sgrproj_info);
+  const int ntiles = av1_get_rest_ntiles(cm->width, cm->height,
+                                         cm->rst_info[0].restoration_tilesize,
+                                         NULL, NULL, NULL, NULL);
+  const int ntiles_uv = av1_get_rest_ntiles(
+      ROUND_POWER_OF_TWO(cm->width, cm->subsampling_x),
+      ROUND_POWER_OF_TWO(cm->height, cm->subsampling_y),
+      cm->rst_info[1].restoration_tilesize, NULL, NULL, NULL, NULL);
+  RestorationInfo *rsi = &cm->rst_info[0];
+  if (rsi->frame_restoration_type != RESTORE_NONE) {
+    if (rsi->frame_restoration_type == RESTORE_SWITCHABLE) {
+      for (i = 0; i < ntiles; ++i) {
+        rsi->restoration_type[i] =
+            aom_read_tree(rb, av1_switchable_restore_tree,
+                          cm->fc->switchable_restore_prob, ACCT_STR);
+        if (rsi->restoration_type[i] == RESTORE_WIENER) {
+          read_wiener_filter(&rsi->wiener_info[i], &ref_wiener_info, rb);
+        } else if (rsi->restoration_type[i] == RESTORE_SGRPROJ) {
+          read_sgrproj_filter(&rsi->sgrproj_info[i], &ref_sgrproj_info, rb);
+        }
+      }
+    } else if (rsi->frame_restoration_type == RESTORE_WIENER) {
+      for (i = 0; i < ntiles; ++i) {
+        if (aom_read(rb, RESTORE_NONE_WIENER_PROB, ACCT_STR)) {
+          rsi->restoration_type[i] = RESTORE_WIENER;
+          read_wiener_filter(&rsi->wiener_info[i], &ref_wiener_info, rb);
+        } else {
+          rsi->restoration_type[i] = RESTORE_NONE;
+        }
+      }
+    } else if (rsi->frame_restoration_type == RESTORE_SGRPROJ) {
+      for (i = 0; i < ntiles; ++i) {
+        if (aom_read(rb, RESTORE_NONE_SGRPROJ_PROB, ACCT_STR)) {
+          rsi->restoration_type[i] = RESTORE_SGRPROJ;
+          read_sgrproj_filter(&rsi->sgrproj_info[i], &ref_sgrproj_info, rb);
+        } else {
+          rsi->restoration_type[i] = RESTORE_NONE;
+        }
+      }
+    }
+  }
+  for (p = 1; p < MAX_MB_PLANE; ++p) {
+    set_default_wiener(&ref_wiener_info);
+    rsi = &cm->rst_info[p];
+    if (rsi->frame_restoration_type == RESTORE_WIENER) {
+      for (i = 0; i < ntiles_uv; ++i) {
+        if (ntiles_uv > 1)
+          rsi->restoration_type[i] =
+              aom_read(rb, RESTORE_NONE_WIENER_PROB, ACCT_STR) ? RESTORE_WIENER
+                                                               : RESTORE_NONE;
+        else
+          rsi->restoration_type[i] = RESTORE_WIENER;
+        if (rsi->restoration_type[i] == RESTORE_WIENER) {
+          read_wiener_filter(&rsi->wiener_info[i], &ref_wiener_info, rb);
+        }
+      }
+    }
+  }
+}
+#endif  // CONFIG_LOOP_RESTORATION
+
+static void setup_loopfilter(AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  struct loopfilter *lf = &cm->lf;
+  lf->filter_level = aom_rb_read_literal(rb, 6);
+  lf->sharpness_level = aom_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 = aom_rb_read_bit(rb);
+  if (lf->mode_ref_delta_enabled) {
+    lf->mode_ref_delta_update = aom_rb_read_bit(rb);
+    if (lf->mode_ref_delta_update) {
+      int i;
+
+      for (i = 0; i < TOTAL_REFS_PER_FRAME; i++)
+        if (aom_rb_read_bit(rb))
+          lf->ref_deltas[i] = aom_rb_read_inv_signed_literal(rb, 6);
+
+      for (i = 0; i < MAX_MODE_LF_DELTAS; i++)
+        if (aom_rb_read_bit(rb))
+          lf->mode_deltas[i] = aom_rb_read_inv_signed_literal(rb, 6);
+    }
+  }
+}
+
+#if CONFIG_CDEF
+static void setup_cdef(AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  int i;
+  cm->cdef_dering_damping = aom_rb_read_literal(rb, 1) + 5;
+  cm->cdef_clpf_damping = aom_rb_read_literal(rb, 2) + 3;
+  cm->cdef_bits = aom_rb_read_literal(rb, 2);
+  cm->nb_cdef_strengths = 1 << cm->cdef_bits;
+  for (i = 0; i < cm->nb_cdef_strengths; i++) {
+    cm->cdef_strengths[i] = aom_rb_read_literal(rb, CDEF_STRENGTH_BITS);
+    cm->cdef_uv_strengths[i] = aom_rb_read_literal(rb, CDEF_STRENGTH_BITS);
+  }
+}
+#endif  // CONFIG_CDEF
+
+static INLINE int read_delta_q(struct aom_read_bit_buffer *rb) {
+  return aom_rb_read_bit(rb) ? aom_rb_read_inv_signed_literal(rb, 6) : 0;
+}
+
+static void setup_quantization(AV1_COMMON *const cm,
+                               struct aom_read_bit_buffer *rb) {
+  cm->base_qindex = aom_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;
+#if CONFIG_AOM_QM
+  cm->using_qmatrix = aom_rb_read_bit(rb);
+  if (cm->using_qmatrix) {
+    cm->min_qmlevel = aom_rb_read_literal(rb, QM_LEVEL_BITS);
+    cm->max_qmlevel = aom_rb_read_literal(rb, QM_LEVEL_BITS);
+  } else {
+    cm->min_qmlevel = 0;
+    cm->max_qmlevel = 0;
+  }
+#endif
+}
+
+static void setup_segmentation_dequant(AV1_COMMON *const cm) {
+  // Build y/uv dequant values based on segmentation.
+  int i = 0;
+#if CONFIG_AOM_QM
+  int lossless;
+  int j = 0;
+  int qmlevel;
+  int using_qm = cm->using_qmatrix;
+  int minqm = cm->min_qmlevel;
+  int maxqm = cm->max_qmlevel;
+#endif
+#if CONFIG_NEW_QUANT
+  int b;
+  int dq;
+#endif  //  CONFIG_NEW_QUANT
+  if (cm->seg.enabled) {
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      const int qindex = av1_get_qindex(&cm->seg, i, cm->base_qindex);
+      cm->y_dequant[i][0] =
+          av1_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
+      cm->y_dequant[i][1] = av1_ac_quant(qindex, 0, cm->bit_depth);
+      cm->uv_dequant[i][0] =
+          av1_dc_quant(qindex, cm->uv_dc_delta_q, cm->bit_depth);
+      cm->uv_dequant[i][1] =
+          av1_ac_quant(qindex, cm->uv_ac_delta_q, cm->bit_depth);
+#if CONFIG_AOM_QM
+      lossless = qindex == 0 && cm->y_dc_delta_q == 0 &&
+                 cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
+      // NB: depends on base index so there is only 1 set per frame
+      // No quant weighting when lossless or signalled not using QM
+      qmlevel = (lossless || using_qm == 0)
+                    ? NUM_QM_LEVELS - 1
+                    : aom_get_qmlevel(cm->base_qindex, minqm, maxqm);
+      for (j = 0; j < TX_SIZES; ++j) {
+        cm->y_iqmatrix[i][1][j] = aom_iqmatrix(cm, qmlevel, 0, j, 1);
+        cm->y_iqmatrix[i][0][j] = aom_iqmatrix(cm, qmlevel, 0, j, 0);
+        cm->uv_iqmatrix[i][1][j] = aom_iqmatrix(cm, qmlevel, 1, j, 1);
+        cm->uv_iqmatrix[i][0][j] = aom_iqmatrix(cm, qmlevel, 1, j, 0);
+      }
+#endif  // CONFIG_AOM_QM
+#if CONFIG_NEW_QUANT
+      for (dq = 0; dq < QUANT_PROFILES; dq++) {
+        for (b = 0; b < COEF_BANDS; ++b) {
+          av1_get_dequant_val_nuq(cm->y_dequant[i][b != 0], b,
+                                  cm->y_dequant_nuq[i][dq][b], NULL, dq);
+          av1_get_dequant_val_nuq(cm->uv_dequant[i][b != 0], b,
+                                  cm->uv_dequant_nuq[i][dq][b], NULL, dq);
+        }
+      }
+#endif  //  CONFIG_NEW_QUANT
+    }
+  } 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] = av1_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
+    cm->y_dequant[0][1] = av1_ac_quant(qindex, 0, cm->bit_depth);
+    cm->uv_dequant[0][0] =
+        av1_dc_quant(qindex, cm->uv_dc_delta_q, cm->bit_depth);
+    cm->uv_dequant[0][1] =
+        av1_ac_quant(qindex, cm->uv_ac_delta_q, cm->bit_depth);
+#if CONFIG_AOM_QM
+    lossless = qindex == 0 && cm->y_dc_delta_q == 0 && cm->uv_dc_delta_q == 0 &&
+               cm->uv_ac_delta_q == 0;
+    // No quant weighting when lossless or signalled not using QM
+    qmlevel = (lossless || using_qm == 0)
+                  ? NUM_QM_LEVELS - 1
+                  : aom_get_qmlevel(cm->base_qindex, minqm, maxqm);
+    for (j = 0; j < TX_SIZES; ++j) {
+      cm->y_iqmatrix[i][1][j] = aom_iqmatrix(cm, qmlevel, 0, j, 1);
+      cm->y_iqmatrix[i][0][j] = aom_iqmatrix(cm, qmlevel, 0, j, 0);
+      cm->uv_iqmatrix[i][1][j] = aom_iqmatrix(cm, qmlevel, 1, j, 1);
+      cm->uv_iqmatrix[i][0][j] = aom_iqmatrix(cm, qmlevel, 1, j, 0);
+    }
+#endif
+#if CONFIG_NEW_QUANT
+    for (dq = 0; dq < QUANT_PROFILES; dq++) {
+      for (b = 0; b < COEF_BANDS; ++b) {
+        av1_get_dequant_val_nuq(cm->y_dequant[0][b != 0], b,
+                                cm->y_dequant_nuq[0][dq][b], NULL, dq);
+        av1_get_dequant_val_nuq(cm->uv_dequant[0][b != 0], b,
+                                cm->uv_dequant_nuq[0][dq][b], NULL, dq);
+      }
+    }
+#endif  //  CONFIG_NEW_QUANT
+  }
+}
+
+static InterpFilter read_frame_interp_filter(struct aom_read_bit_buffer *rb) {
+  return aom_rb_read_bit(rb) ? SWITCHABLE
+                             : aom_rb_read_literal(rb, LOG_SWITCHABLE_FILTERS);
+}
+
+static void setup_render_size(AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  cm->render_width = cm->width;
+  cm->render_height = cm->height;
+  if (aom_rb_read_bit(rb))
+    av1_read_frame_size(rb, &cm->render_width, &cm->render_height);
+}
+
+#if CONFIG_FRAME_SUPERRES
+// TODO(afergs): make "struct aom_read_bit_buffer *const rb"?
+static void setup_superres_size(AV1_COMMON *const cm,
+                                struct aom_read_bit_buffer *rb, int *width,
+                                int *height) {
+  // TODO(afergs): Test this behaviour
+  // Frame superres is probably in compatible with this render resolution
+  assert(cm->width == cm->render_width && cm->height == cm->render_height);
+
+  cm->superres_width = cm->width;
+  cm->superres_height = cm->height;
+  if (aom_rb_read_bit(rb)) {
+    cm->superres_scale_numerator =
+        (uint8_t)aom_rb_read_literal(rb, SUPERRES_SCALE_BITS);
+    cm->superres_scale_numerator += SUPERRES_SCALE_NUMERATOR_MIN;
+    // Don't edit cm->width or cm->height directly, or the buffers won't get
+    // resized correctly
+    // TODO(afergs): Should the render resolution not be modified? It's the same
+    // by default (ie. when it isn't sent)...
+    // resize_context_buffers() will change cm->width to equal cm->render_width,
+    // then they'll be the same again
+    *width = cm->render_width =
+        cm->width * cm->superres_scale_numerator / SUPERRES_SCALE_DENOMINATOR;
+    *height = cm->render_height =
+        cm->height * cm->superres_scale_numerator / SUPERRES_SCALE_DENOMINATOR;
+  } else {
+    // 1:1 scaling - ie. no scaling, scale not provided
+    cm->superres_scale_numerator = SUPERRES_SCALE_DENOMINATOR;
+  }
+}
+#endif  // CONFIG_FRAME_SUPERRES
+
+static void resize_mv_buffer(AV1_COMMON *cm) {
+  aom_free(cm->cur_frame->mvs);
+  cm->cur_frame->mi_rows = cm->mi_rows;
+  cm->cur_frame->mi_cols = cm->mi_cols;
+  CHECK_MEM_ERROR(cm, cm->cur_frame->mvs,
+                  (MV_REF *)aom_calloc(cm->mi_rows * cm->mi_cols,
+                                       sizeof(*cm->cur_frame->mvs)));
+}
+
+static void resize_context_buffers(AV1_COMMON *cm, int width, int height) {
+#if CONFIG_SIZE_LIMIT
+  if (width > DECODE_WIDTH_LIMIT || height > DECODE_HEIGHT_LIMIT)
+    aom_internal_error(&cm->error, AOM_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 av1_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 (av1_alloc_context_buffers(cm, width, height))
+        aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                           "Failed to allocate context buffers");
+    } else {
+      av1_set_mb_mi(cm, width, height);
+    }
+    av1_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(AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  int width, height;
+  BufferPool *const pool = cm->buffer_pool;
+  av1_read_frame_size(rb, &width, &height);
+  setup_render_size(cm, rb);
+#if CONFIG_FRAME_SUPERRES
+  setup_superres_size(cm, rb, &width, &height);
+#endif  // CONFIG_FRAME_SUPERRES
+  resize_context_buffers(cm, width, height);
+
+  lock_buffer_pool(pool);
+  if (aom_realloc_frame_buffer(
+          get_frame_new_buffer(cm), cm->width, cm->height, cm->subsampling_x,
+          cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+          cm->use_highbitdepth,
+#endif
+          AOM_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);
+    aom_internal_error(&cm->error, AOM_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;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_range = cm->color_range;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_width = cm->render_width;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_height = cm->render_height;
+}
+
+static INLINE int valid_ref_frame_img_fmt(aom_bit_depth_t ref_bit_depth,
+                                          int ref_xss, int ref_yss,
+                                          aom_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(AV1_COMMON *cm,
+                                       struct aom_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 < INTER_REFS_PER_FRAME; ++i) {
+    if (aom_rb_read_bit(rb)) {
+      YV12_BUFFER_CONFIG *const buf = cm->frame_refs[i].buf;
+      width = buf->y_crop_width;
+      height = buf->y_crop_height;
+      cm->render_width = buf->render_width;
+      cm->render_height = buf->render_height;
+      found = 1;
+      break;
+    }
+  }
+
+  if (!found) {
+    av1_read_frame_size(rb, &width, &height);
+    setup_render_size(cm, rb);
+  }
+
+  if (width <= 0 || height <= 0)
+    aom_internal_error(&cm->error, AOM_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 < INTER_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)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Referenced frame has invalid size");
+  for (i = 0; i < INTER_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))
+      aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                         "Referenced frame has incompatible color format");
+  }
+
+  resize_context_buffers(cm, width, height);
+
+  lock_buffer_pool(pool);
+  if (aom_realloc_frame_buffer(
+          get_frame_new_buffer(cm), cm->width, cm->height, cm->subsampling_x,
+          cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+          cm->use_highbitdepth,
+#endif
+          AOM_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);
+    aom_internal_error(&cm->error, AOM_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;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_range = cm->color_range;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_width = cm->render_width;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_height = cm->render_height;
+}
+
+static void read_tile_info(AV1Decoder *const pbi,
+                           struct aom_read_bit_buffer *const rb) {
+  AV1_COMMON *const cm = &pbi->common;
+#if CONFIG_EXT_TILE
+  cm->tile_encoding_mode = aom_rb_read_literal(rb, 1);
+// Read the tile width/height
+#if CONFIG_EXT_PARTITION
+  if (cm->sb_size == BLOCK_128X128) {
+    cm->tile_width = aom_rb_read_literal(rb, 5) + 1;
+    cm->tile_height = aom_rb_read_literal(rb, 5) + 1;
+  } else
+#endif  // CONFIG_EXT_PARTITION
+  {
+    cm->tile_width = aom_rb_read_literal(rb, 6) + 1;
+    cm->tile_height = aom_rb_read_literal(rb, 6) + 1;
+  }
+
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  cm->loop_filter_across_tiles_enabled = aom_rb_read_bit(rb);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+  cm->tile_width <<= cm->mib_size_log2;
+  cm->tile_height <<= cm->mib_size_log2;
+
+  cm->tile_width = AOMMIN(cm->tile_width, cm->mi_cols);
+  cm->tile_height = AOMMIN(cm->tile_height, cm->mi_rows);
+
+  // Get the number of tiles
+  cm->tile_cols = 1;
+  while (cm->tile_cols * cm->tile_width < cm->mi_cols) ++cm->tile_cols;
+
+  cm->tile_rows = 1;
+  while (cm->tile_rows * cm->tile_height < cm->mi_rows) ++cm->tile_rows;
+
+  if (cm->tile_cols * cm->tile_rows > 1) {
+    // Read the number of bytes used to store tile size
+    pbi->tile_col_size_bytes = aom_rb_read_literal(rb, 2) + 1;
+    pbi->tile_size_bytes = aom_rb_read_literal(rb, 2) + 1;
+  }
+
+#if CONFIG_DEPENDENT_HORZTILES
+  if (cm->tile_rows <= 1)
+    cm->dependent_horz_tiles = aom_rb_read_bit(rb);
+  else
+    cm->dependent_horz_tiles = 0;
+#endif
+#else
+  int min_log2_tile_cols, max_log2_tile_cols, max_ones;
+  av1_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-- && aom_rb_read_bit(rb)) cm->log2_tile_cols++;
+
+  if (cm->log2_tile_cols > 6)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Invalid number of tile columns");
+
+  // rows
+  cm->log2_tile_rows = aom_rb_read_bit(rb);
+  if (cm->log2_tile_rows) cm->log2_tile_rows += aom_rb_read_bit(rb);
+#if CONFIG_DEPENDENT_HORZTILES
+  if (cm->log2_tile_rows != 0)
+    cm->dependent_horz_tiles = aom_rb_read_bit(rb);
+  else
+    cm->dependent_horz_tiles = 0;
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  cm->loop_filter_across_tiles_enabled = aom_rb_read_bit(rb);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+  cm->tile_cols = 1 << cm->log2_tile_cols;
+  cm->tile_rows = 1 << cm->log2_tile_rows;
+
+  cm->tile_width = ALIGN_POWER_OF_TWO(cm->mi_cols, MAX_MIB_SIZE_LOG2);
+  cm->tile_width >>= cm->log2_tile_cols;
+  cm->tile_height = ALIGN_POWER_OF_TWO(cm->mi_rows, MAX_MIB_SIZE_LOG2);
+  cm->tile_height >>= cm->log2_tile_rows;
+
+  // round to integer multiples of superblock size
+  cm->tile_width = ALIGN_POWER_OF_TWO(cm->tile_width, MAX_MIB_SIZE_LOG2);
+  cm->tile_height = ALIGN_POWER_OF_TWO(cm->tile_height, MAX_MIB_SIZE_LOG2);
+
+// tile size magnitude
+#if !CONFIG_TILE_GROUPS
+  if (cm->tile_rows > 1 || cm->tile_cols > 1)
+#endif
+    pbi->tile_size_bytes = aom_rb_read_literal(rb, 2) + 1;
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_TILE_GROUPS
+  // Store an index to the location of the tile group information
+  pbi->tg_size_bit_offset = rb->bit_offset;
+  pbi->tg_size = 1 << (cm->log2_tile_rows + cm->log2_tile_cols);
+  if (cm->log2_tile_rows + cm->log2_tile_cols > 0) {
+    pbi->tg_start =
+        aom_rb_read_literal(rb, cm->log2_tile_rows + cm->log2_tile_cols);
+    pbi->tg_size =
+        1 + aom_rb_read_literal(rb, cm->log2_tile_rows + cm->log2_tile_cols);
+  }
+#endif
+}
+
+static int mem_get_varsize(const uint8_t *src, int sz) {
+  switch (sz) {
+    case 1: return src[0];
+    case 2: return mem_get_le16(src);
+    case 3: return mem_get_le24(src);
+    case 4: return mem_get_le32(src);
+    default: assert("Invalid size" && 0); return -1;
+  }
+}
+
+#if CONFIG_EXT_TILE
+// 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,
+                            struct aom_internal_error_info *error_info,
+                            const uint8_t **data, aom_decrypt_cb decrypt_cb,
+                            void *decrypt_state,
+                            TileBufferDec (*const tile_buffers)[MAX_TILE_COLS],
+                            int tile_size_bytes, int col, int row,
+                            unsigned int tile_encoding_mode) {
+  size_t size;
+
+  size_t copy_size = 0;
+  const uint8_t *copy_data = NULL;
+
+  if (!read_is_valid(*data, tile_size_bytes, data_end))
+    aom_internal_error(error_info, AOM_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt tile length");
+  if (decrypt_cb) {
+    uint8_t be_data[4];
+    decrypt_cb(decrypt_state, *data, be_data, tile_size_bytes);
+
+    // Only read number of bytes in cm->tile_size_bytes.
+    size = mem_get_varsize(be_data, tile_size_bytes);
+  } else {
+    size = mem_get_varsize(*data, tile_size_bytes);
+  }
+
+  // If cm->tile_encoding_mode = 1 (i.e. TILE_VR), then the top bit of the tile
+  // header indicates copy mode.
+  if (tile_encoding_mode && (size >> (tile_size_bytes * 8 - 1)) == 1) {
+    // The remaining bits in the top byte signal the row offset
+    int offset = (size >> (tile_size_bytes - 1) * 8) & 0x7f;
+
+    // Currently, only use tiles in same column as reference tiles.
+    copy_data = tile_buffers[row - offset][col].data;
+    copy_size = tile_buffers[row - offset][col].size;
+    size = 0;
+  }
+
+  *data += tile_size_bytes;
+
+  if (size > (size_t)(data_end - *data))
+    aom_internal_error(error_info, AOM_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt tile size");
+
+  if (size > 0) {
+    tile_buffers[row][col].data = *data;
+    tile_buffers[row][col].size = size;
+  } else {
+    tile_buffers[row][col].data = copy_data;
+    tile_buffers[row][col].size = copy_size;
+  }
+
+  *data += size;
+
+  tile_buffers[row][col].raw_data_end = *data;
+}
+
+static void get_tile_buffers(
+    AV1Decoder *pbi, const uint8_t *data, const uint8_t *data_end,
+    TileBufferDec (*const tile_buffers)[MAX_TILE_COLS]) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  const int have_tiles = tile_cols * tile_rows > 1;
+
+  if (!have_tiles) {
+    const size_t tile_size = data_end - data;
+    tile_buffers[0][0].data = data;
+    tile_buffers[0][0].size = tile_size;
+    tile_buffers[0][0].raw_data_end = NULL;
+  } else {
+    // We locate only the tile buffers that are required, which are the ones
+    // specified by pbi->dec_tile_col and pbi->dec_tile_row. Also, we always
+    // need the last (bottom right) tile buffer, as we need to know where the
+    // end of the compressed frame buffer is for proper superframe decoding.
+
+    const uint8_t *tile_col_data_end[MAX_TILE_COLS];
+    const uint8_t *const data_start = data;
+
+    const int dec_tile_row = AOMMIN(pbi->dec_tile_row, tile_rows);
+    const int single_row = pbi->dec_tile_row >= 0;
+    const int tile_rows_start = single_row ? dec_tile_row : 0;
+    const int tile_rows_end = single_row ? tile_rows_start + 1 : tile_rows;
+    const int dec_tile_col = AOMMIN(pbi->dec_tile_col, tile_cols);
+    const int single_col = pbi->dec_tile_col >= 0;
+    const int tile_cols_start = single_col ? dec_tile_col : 0;
+    const int tile_cols_end = single_col ? tile_cols_start + 1 : tile_cols;
+
+    const int tile_col_size_bytes = pbi->tile_col_size_bytes;
+    const int tile_size_bytes = pbi->tile_size_bytes;
+
+    size_t tile_col_size;
+    int r, c;
+
+    // Read tile column sizes for all columns (we need the last tile buffer)
+    for (c = 0; c < tile_cols; ++c) {
+      const int is_last = c == tile_cols - 1;
+      if (!is_last) {
+        tile_col_size = mem_get_varsize(data, tile_col_size_bytes);
+        data += tile_col_size_bytes;
+        tile_col_data_end[c] = data + tile_col_size;
+      } else {
+        tile_col_size = data_end - data;
+        tile_col_data_end[c] = data_end;
+      }
+      data += tile_col_size;
+    }
+
+    data = data_start;
+
+    // Read the required tile sizes.
+    for (c = tile_cols_start; c < tile_cols_end; ++c) {
+      const int is_last = c == tile_cols - 1;
+
+      if (c > 0) data = tile_col_data_end[c - 1];
+
+      if (!is_last) data += tile_col_size_bytes;
+
+      // Get the whole of the last column, otherwise stop at the required tile.
+      for (r = 0; r < (is_last ? tile_rows : tile_rows_end); ++r) {
+        tile_buffers[r][c].col = c;
+
+        get_tile_buffer(tile_col_data_end[c], &pbi->common.error, &data,
+                        pbi->decrypt_cb, pbi->decrypt_state, tile_buffers,
+                        tile_size_bytes, c, r, cm->tile_encoding_mode);
+      }
+    }
+
+    // If we have not read the last column, then read it to get the last tile.
+    if (tile_cols_end != tile_cols) {
+      c = tile_cols - 1;
+
+      data = tile_col_data_end[c - 1];
+
+      for (r = 0; r < tile_rows; ++r) {
+        tile_buffers[r][c].col = c;
+
+        get_tile_buffer(tile_col_data_end[c], &pbi->common.error, &data,
+                        pbi->decrypt_cb, pbi->decrypt_state, tile_buffers,
+                        tile_size_bytes, c, r, cm->tile_encoding_mode);
+      }
+    }
+  }
+}
+#else
+// 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,
+                            const int tile_size_bytes, int is_last,
+                            struct aom_internal_error_info *error_info,
+                            const uint8_t **data, aom_decrypt_cb decrypt_cb,
+                            void *decrypt_state, TileBufferDec *const buf) {
+  size_t size;
+
+  if (!is_last) {
+    if (!read_is_valid(*data, tile_size_bytes, data_end))
+      aom_internal_error(error_info, AOM_CODEC_CORRUPT_FRAME,
+                         "Truncated packet or corrupt tile length");
+
+    if (decrypt_cb) {
+      uint8_t be_data[4];
+      decrypt_cb(decrypt_state, *data, be_data, tile_size_bytes);
+      size = mem_get_varsize(be_data, tile_size_bytes);
+    } else {
+      size = mem_get_varsize(*data, tile_size_bytes);
+    }
+    *data += tile_size_bytes;
+
+    if (size > (size_t)(data_end - *data))
+      aom_internal_error(error_info, AOM_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(
+    AV1Decoder *pbi, const uint8_t *data, const uint8_t *data_end,
+    TileBufferDec (*const tile_buffers)[MAX_TILE_COLS]) {
+  AV1_COMMON *const cm = &pbi->common;
+#if CONFIG_TILE_GROUPS
+  int r, c;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  int tc = 0;
+  int first_tile_in_tg = 0;
+  struct aom_read_bit_buffer rb_tg_hdr;
+  uint8_t clear_data[MAX_AV1_HEADER_SIZE];
+  const int num_tiles = tile_rows * tile_cols;
+  const int num_bits = OD_ILOG(num_tiles) - 1;
+  const size_t hdr_size = pbi->uncomp_hdr_size + pbi->first_partition_size;
+  const int tg_size_bit_offset = pbi->tg_size_bit_offset;
+#if CONFIG_DEPENDENT_HORZTILES
+  int tile_group_start_col = 0;
+  int tile_group_start_row = 0;
+#endif
+
+  for (r = 0; r < tile_rows; ++r) {
+    for (c = 0; c < tile_cols; ++c, ++tc) {
+      TileBufferDec *const buf = &tile_buffers[r][c];
+      const int is_last = (r == tile_rows - 1) && (c == tile_cols - 1);
+      const size_t hdr_offset = (tc && tc == first_tile_in_tg) ? hdr_size : 0;
+
+      buf->col = c;
+      if (hdr_offset) {
+        init_read_bit_buffer(pbi, &rb_tg_hdr, data, data_end, clear_data);
+        rb_tg_hdr.bit_offset = tg_size_bit_offset;
+        if (num_tiles) {
+          pbi->tg_start = aom_rb_read_literal(&rb_tg_hdr, num_bits);
+          pbi->tg_size = 1 + aom_rb_read_literal(&rb_tg_hdr, num_bits);
+#if CONFIG_DEPENDENT_HORZTILES
+          tile_group_start_row = r;
+          tile_group_start_col = c;
+#endif
+        }
+      }
+      first_tile_in_tg += tc == first_tile_in_tg ? pbi->tg_size : 0;
+      data += hdr_offset;
+      get_tile_buffer(data_end, pbi->tile_size_bytes, is_last,
+                      &pbi->common.error, &data, pbi->decrypt_cb,
+                      pbi->decrypt_state, buf);
+#if CONFIG_DEPENDENT_HORZTILES
+      cm->tile_group_start_row[r][c] = tile_group_start_row;
+      cm->tile_group_start_col[r][c] = tile_group_start_col;
+#endif
+    }
+  }
+#else
+  int r, c;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+
+  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);
+      TileBufferDec *const buf = &tile_buffers[r][c];
+      buf->col = c;
+      get_tile_buffer(data_end, pbi->tile_size_bytes, is_last, &cm->error,
+                      &data, pbi->decrypt_cb, pbi->decrypt_state, buf);
+    }
+  }
+#endif
+}
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_PVQ
+static void daala_dec_init(AV1_COMMON *const cm, daala_dec_ctx *daala_dec,
+                           aom_reader *r) {
+  daala_dec->r = r;
+
+  // TODO(yushin) : activity masking info needs be signaled by a bitstream
+  daala_dec->use_activity_masking = AV1_PVQ_ENABLE_ACTIVITY_MASKING;
+
+#if !CONFIG_DAALA_DIST
+  daala_dec->use_activity_masking = 0;
+#endif
+
+  if (daala_dec->use_activity_masking)
+    daala_dec->qm = OD_HVS_QM;
+  else
+    daala_dec->qm = OD_FLAT_QM;
+
+  od_init_qm(daala_dec->state.qm, daala_dec->state.qm_inv,
+             daala_dec->qm == OD_HVS_QM ? OD_QM8_Q4_HVS : OD_QM8_Q4_FLAT);
+
+  if (daala_dec->use_activity_masking) {
+    int pli;
+    int use_masking = daala_dec->use_activity_masking;
+    int segment_id = 0;
+    int qindex = av1_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+
+    for (pli = 0; pli < MAX_MB_PLANE; pli++) {
+      int i;
+      int q;
+
+      q = qindex;
+      if (q <= OD_DEFAULT_QMS[use_masking][0][pli].interp_q << OD_COEFF_SHIFT) {
+        od_interp_qm(&daala_dec->state.pvq_qm_q4[pli][0], q,
+                     &OD_DEFAULT_QMS[use_masking][0][pli], NULL);
+      } else {
+        i = 0;
+        while (OD_DEFAULT_QMS[use_masking][i + 1][pli].qm_q4 != NULL &&
+               q > OD_DEFAULT_QMS[use_masking][i + 1][pli].interp_q
+                       << OD_COEFF_SHIFT) {
+          i++;
+        }
+        od_interp_qm(&daala_dec->state.pvq_qm_q4[pli][0], q,
+                     &OD_DEFAULT_QMS[use_masking][i][pli],
+                     &OD_DEFAULT_QMS[use_masking][i + 1][pli]);
+      }
+    }
+  }
+}
+#endif  // #if CONFIG_PVQ
+
+static const uint8_t *decode_tiles(AV1Decoder *pbi, const uint8_t *data,
+                                   const uint8_t *data_end) {
+  AV1_COMMON *const cm = &pbi->common;
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  const int n_tiles = tile_cols * tile_rows;
+  TileBufferDec(*const tile_buffers)[MAX_TILE_COLS] = pbi->tile_buffers;
+#if CONFIG_EXT_TILE
+  const int dec_tile_row = AOMMIN(pbi->dec_tile_row, tile_rows);
+  const int single_row = pbi->dec_tile_row >= 0;
+  const int tile_rows_start = single_row ? dec_tile_row : 0;
+  const int tile_rows_end = single_row ? dec_tile_row + 1 : tile_rows;
+  const int dec_tile_col = AOMMIN(pbi->dec_tile_col, tile_cols);
+  const int single_col = pbi->dec_tile_col >= 0;
+  const int tile_cols_start = single_col ? dec_tile_col : 0;
+  const int tile_cols_end = single_col ? tile_cols_start + 1 : tile_cols;
+  const int inv_col_order = pbi->inv_tile_order && !single_col;
+  const int inv_row_order = pbi->inv_tile_order && !single_row;
+#else
+  const int tile_rows_start = 0;
+  const int tile_rows_end = tile_rows;
+  const int tile_cols_start = 0;
+  const int tile_cols_end = tile_cols;
+  const int inv_col_order = pbi->inv_tile_order;
+  const int inv_row_order = pbi->inv_tile_order;
+#endif  // CONFIG_EXT_TILE
+  int tile_row, tile_col;
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  cm->do_subframe_update = n_tiles == 1;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter &&
+      pbi->lf_worker.data1 == NULL) {
+    CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
+                    aom_memalign(32, sizeof(LFWorkerData)));
+    pbi->lf_worker.hook = (AVxWorkerHook)av1_loop_filter_worker;
+    if (pbi->max_threads > 1 && !winterface->reset(&pbi->lf_worker)) {
+      aom_internal_error(&cm->error, AOM_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);
+    av1_loop_filter_data_reset(lf_data, get_frame_new_buffer(cm), cm,
+                               pbi->mb.plane);
+  }
+
+  assert(tile_rows <= MAX_TILE_ROWS);
+  assert(tile_cols <= MAX_TILE_COLS);
+
+  get_tile_buffers(pbi, data, data_end, tile_buffers);
+
+  if (pbi->tile_data == NULL || n_tiles != pbi->allocated_tiles) {
+    aom_free(pbi->tile_data);
+    CHECK_MEM_ERROR(cm, pbi->tile_data,
+                    aom_memalign(32, n_tiles * (sizeof(*pbi->tile_data))));
+    pbi->allocated_tiles = n_tiles;
+  }
+#if CONFIG_ACCOUNTING
+  if (pbi->acct_enabled) {
+    aom_accounting_reset(&pbi->accounting);
+  }
+#endif
+  // Load all tile information into tile_data.
+  for (tile_row = tile_rows_start; tile_row < tile_rows_end; ++tile_row) {
+    for (tile_col = tile_cols_start; tile_col < tile_cols_end; ++tile_col) {
+      const TileBufferDec *const buf = &tile_buffers[tile_row][tile_col];
+      TileData *const td = pbi->tile_data + tile_cols * tile_row + tile_col;
+
+      td->cm = cm;
+      td->xd = pbi->mb;
+      td->xd.corrupted = 0;
+      td->xd.counts =
+          cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD
+              ? &cm->counts
+              : NULL;
+      av1_zero(td->dqcoeff);
+#if CONFIG_PVQ
+      av1_zero(td->pvq_ref_coeff);
+#endif
+      av1_tile_init(&td->xd.tile, td->cm, tile_row, tile_col);
+      setup_bool_decoder(buf->data, data_end, buf->size, &cm->error,
+                         &td->bit_reader,
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+                         1 << cm->ans_window_size_log2,
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+                         pbi->decrypt_cb, pbi->decrypt_state);
+#if CONFIG_ACCOUNTING
+      if (pbi->acct_enabled) {
+        td->bit_reader.accounting = &pbi->accounting;
+      } else {
+        td->bit_reader.accounting = NULL;
+      }
+#endif
+      av1_init_macroblockd(cm, &td->xd,
+#if CONFIG_PVQ
+                           td->pvq_ref_coeff,
+#endif
+#if CONFIG_CFL
+                           &td->cfl,
+#endif
+                           td->dqcoeff);
+
+#if CONFIG_EC_ADAPT
+      // Initialise the tile context from the frame context
+      td->tctx = *cm->fc;
+      td->xd.tile_ctx = &td->tctx;
+#endif
+
+#if CONFIG_PVQ
+      daala_dec_init(cm, &td->xd.daala_dec, &td->bit_reader);
+      td->xd.daala_dec.state.adapt = &td->tctx.pvq_context;
+#endif
+
+#if CONFIG_PALETTE
+      td->xd.plane[0].color_index_map = td->color_index_map[0];
+      td->xd.plane[1].color_index_map = td->color_index_map[1];
+#endif  // CONFIG_PALETTE
+    }
+  }
+
+  for (tile_row = tile_rows_start; tile_row < tile_rows_end; ++tile_row) {
+    const int row = inv_row_order ? tile_rows - 1 - tile_row : tile_row;
+    int mi_row = 0;
+    TileInfo tile_info;
+
+    av1_tile_set_row(&tile_info, cm, row);
+
+    for (tile_col = tile_cols_start; tile_col < tile_cols_end; ++tile_col) {
+      const int col = inv_col_order ? tile_cols - 1 - tile_col : tile_col;
+      TileData *const td = pbi->tile_data + tile_cols * row + col;
+#if CONFIG_ACCOUNTING
+      if (pbi->acct_enabled) {
+        td->bit_reader.accounting->last_tell_frac =
+            aom_reader_tell_frac(&td->bit_reader);
+      }
+#endif
+
+      av1_tile_set_col(&tile_info, cm, col);
+
+#if CONFIG_DEPENDENT_HORZTILES
+#if CONFIG_TILE_GROUPS
+      av1_tile_set_tg_boundary(&tile_info, cm, tile_row, tile_col);
+      if (!cm->dependent_horz_tiles || tile_row == 0 ||
+          tile_info.tg_horz_boundary) {
+#else
+      if (!cm->dependent_horz_tiles || tile_row == 0) {
+#endif
+        av1_zero_above_context(cm, tile_info.mi_col_start,
+                               tile_info.mi_col_end);
+      }
+#else
+      av1_zero_above_context(cm, tile_info.mi_col_start, tile_info.mi_col_end);
+#endif
+
+      for (mi_row = tile_info.mi_row_start; mi_row < tile_info.mi_row_end;
+           mi_row += cm->mib_size) {
+        int mi_col;
+
+        av1_zero_left_context(&td->xd);
+
+        for (mi_col = tile_info.mi_col_start; mi_col < tile_info.mi_col_end;
+             mi_col += cm->mib_size) {
+          av1_update_boundary_info(cm, &tile_info, mi_row, mi_col);
+          decode_partition(pbi, &td->xd,
+#if CONFIG_SUPERTX
+                           0,
+#endif  // CONFIG_SUPERTX
+                           mi_row, mi_col, &td->bit_reader, cm->sb_size,
+                           b_width_log2_lookup[cm->sb_size]);
+#if CONFIG_NCOBMC && CONFIG_MOTION_VAR
+          detoken_and_recon_sb(pbi, &td->xd, mi_row, mi_col, &td->bit_reader,
+                               cm->sb_size);
+#endif
+        }
+        aom_merge_corrupted_flag(&pbi->mb.corrupted, td->xd.corrupted);
+        if (pbi->mb.corrupted)
+          aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                             "Failed to decode tile data");
+#if CONFIG_SUBFRAME_PROB_UPDATE
+        if (cm->do_subframe_update &&
+            cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+          const int mi_rows_per_update =
+              MI_SIZE * AOMMAX(cm->mi_rows / MI_SIZE / COEF_PROBS_BUFS, 1);
+          if ((mi_row + MI_SIZE) % mi_rows_per_update == 0 &&
+              mi_row + MI_SIZE < cm->mi_rows &&
+              cm->coef_probs_update_idx < COEF_PROBS_BUFS - 1) {
+            av1_partial_adapt_probs(cm, mi_row, mi_col);
+            ++cm->coef_probs_update_idx;
+          }
+        }
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+      }
+    }
+
+    assert(mi_row > 0);
+
+// when Parallel deblocking is enabled, deblocking should not
+// be interleaved with decoding. Instead, deblocking should be done
+// after the entire frame is decoded.
+#if !CONFIG_VAR_TX && !CONFIG_PARALLEL_DEBLOCKING && !CONFIG_CB4X4
+    // Loopfilter one tile row.
+    // Note: If out-of-order tile decoding is used(for example, inv_row_order
+    // = 1), the loopfiltering has be done after all tile rows are decoded.
+    if (!inv_row_order && cm->lf.filter_level && !cm->skip_loop_filter) {
+      LFWorkerData *const lf_data = (LFWorkerData *)pbi->lf_worker.data1;
+      const int lf_start = AOMMAX(0, tile_info.mi_row_start - cm->mib_size);
+      const int lf_end = tile_info.mi_row_end - cm->mib_size;
+
+      // Delay the loopfilter if the first tile row is only
+      // a single superblock high.
+      if (lf_end <= 0) continue;
+
+      // Decoding has completed. Finish up the loop filter in this thread.
+      if (tile_info.mi_row_end >= cm->mi_rows) continue;
+
+      winterface->sync(&pbi->lf_worker);
+      lf_data->start = lf_start;
+      lf_data->stop = lf_end;
+      if (pbi->max_threads > 1) {
+        winterface->launch(&pbi->lf_worker);
+      } else {
+        winterface->execute(&pbi->lf_worker);
+      }
+    }
+#endif  // !CONFIG_VAR_TX && !CONFIG_PARALLEL_DEBLOCKING
+
+    // 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 (cm->frame_parallel_decode)
+      av1_frameworker_broadcast(pbi->cur_buf, mi_row << cm->mib_size_log2);
+  }
+
+#if CONFIG_VAR_TX || CONFIG_CB4X4
+  // Loopfilter the whole frame.
+  av1_loop_filter_frame(get_frame_new_buffer(cm), cm, &pbi->mb,
+                        cm->lf.filter_level, 0, 0);
+#else
+#if CONFIG_PARALLEL_DEBLOCKING
+  // Loopfilter all rows in the frame 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 = 0;
+    lf_data->stop = cm->mi_rows;
+    winterface->execute(&pbi->lf_worker);
+  }
+#else
+  // 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);
+  }
+#endif  // CONFIG_PARALLEL_DEBLOCKING
+#endif  // CONFIG_VAR_TX
+  if (cm->frame_parallel_decode)
+    av1_frameworker_broadcast(pbi->cur_buf, INT_MAX);
+
+#if CONFIG_EXT_TILE
+  if (n_tiles == 1) {
+#if CONFIG_ANS
+    return data_end;
+#else
+    // Find the end of the single tile buffer
+    return aom_reader_find_end(&pbi->tile_data->bit_reader);
+#endif  // CONFIG_ANS
+  } else {
+    // Return the end of the last tile buffer
+    return tile_buffers[tile_rows - 1][tile_cols - 1].raw_data_end;
+  }
+#else
+#if CONFIG_ANS
+  return data_end;
+#else
+  {
+    // Get last tile data.
+    TileData *const td = pbi->tile_data + tile_cols * tile_rows - 1;
+    return aom_reader_find_end(&td->bit_reader);
+  }
+#endif  // CONFIG_ANS
+#endif  // CONFIG_EXT_TILE
+}
+
+static int tile_worker_hook(TileWorkerData *const tile_data,
+                            const TileInfo *const tile) {
+  AV1Decoder *const pbi = tile_data->pbi;
+  const AV1_COMMON *const cm = &pbi->common;
+  int mi_row, mi_col;
+
+  if (setjmp(tile_data->error_info.jmp)) {
+    tile_data->error_info.setjmp = 0;
+    aom_merge_corrupted_flag(&tile_data->xd.corrupted, 1);
+    return 0;
+  }
+
+  tile_data->error_info.setjmp = 1;
+  tile_data->xd.error_info = &tile_data->error_info;
+#if CONFIG_DEPENDENT_HORZTILES
+#if CONFIG_TILE_GROUPS
+  if (!cm->dependent_horz_tiles || tile->tg_horz_boundary) {
+#else
+  if (!cm->dependent_horz_tiles) {
+#endif
+    av1_zero_above_context(&pbi->common, tile->mi_col_start, tile->mi_col_end);
+  }
+#else
+  av1_zero_above_context(&pbi->common, tile->mi_col_start, tile->mi_col_end);
+#endif
+
+  for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
+       mi_row += cm->mib_size) {
+    av1_zero_left_context(&tile_data->xd);
+
+    for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
+         mi_col += cm->mib_size) {
+      decode_partition(pbi, &tile_data->xd,
+#if CONFIG_SUPERTX
+                       0,
+#endif
+                       mi_row, mi_col, &tile_data->bit_reader, cm->sb_size,
+                       b_width_log2_lookup[cm->sb_size]);
+#if CONFIG_NCOBMC && CONFIG_MOTION_VAR
+      detoken_and_recon_sb(pbi, &tile_data->xd, mi_row, mi_col,
+                           &tile_data->bit_reader, cm->sb_size);
+#endif
+    }
+  }
+  return !tile_data->xd.corrupted;
+}
+
+// sorts in descending order
+static int compare_tile_buffers(const void *a, const void *b) {
+  const TileBufferDec *const buf1 = (const TileBufferDec *)a;
+  const TileBufferDec *const buf2 = (const TileBufferDec *)b;
+  return (int)(buf2->size - buf1->size);
+}
+
+static const uint8_t *decode_tiles_mt(AV1Decoder *pbi, const uint8_t *data,
+                                      const uint8_t *data_end) {
+  AV1_COMMON *const cm = &pbi->common;
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  const int num_workers = AOMMIN(pbi->max_threads & ~1, tile_cols);
+  TileBufferDec(*const tile_buffers)[MAX_TILE_COLS] = pbi->tile_buffers;
+#if CONFIG_EXT_TILE
+  const int dec_tile_row = AOMMIN(pbi->dec_tile_row, tile_rows);
+  const int single_row = pbi->dec_tile_row >= 0;
+  const int tile_rows_start = single_row ? dec_tile_row : 0;
+  const int tile_rows_end = single_row ? dec_tile_row + 1 : tile_rows;
+  const int dec_tile_col = AOMMIN(pbi->dec_tile_col, tile_cols);
+  const int single_col = pbi->dec_tile_col >= 0;
+  const int tile_cols_start = single_col ? dec_tile_col : 0;
+  const int tile_cols_end = single_col ? tile_cols_start + 1 : tile_cols;
+#else
+  const int tile_rows_start = 0;
+  const int tile_rows_end = tile_rows;
+  const int tile_cols_start = 0;
+  const int tile_cols_end = tile_cols;
+#endif  // CONFIG_EXT_TILE
+  int tile_row, tile_col;
+  int i;
+
+#if !(CONFIG_ANS || CONFIG_EXT_TILE)
+  int final_worker = -1;
+#endif  // !(CONFIG_ANS || CONFIG_EXT_TILE)
+
+  assert(tile_rows <= MAX_TILE_ROWS);
+  assert(tile_cols <= MAX_TILE_COLS);
+
+  assert(tile_cols * tile_rows > 1);
+
+  // 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;
+    CHECK_MEM_ERROR(cm, pbi->tile_workers,
+                    aom_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,
+        aom_memalign(32, num_threads * sizeof(*pbi->tile_worker_data)));
+    CHECK_MEM_ERROR(cm, pbi->tile_worker_info,
+                    aom_malloc(num_threads * sizeof(*pbi->tile_worker_info)));
+    for (i = 0; i < num_threads; ++i) {
+      AVxWorker *const worker = &pbi->tile_workers[i];
+      ++pbi->num_tile_workers;
+
+      winterface->init(worker);
+      if (i < num_threads - 1 && !winterface->reset(worker)) {
+        aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                           "Tile decoder thread creation failed");
+      }
+    }
+  }
+
+  // Reset tile decoding hook
+  for (i = 0; i < num_workers; ++i) {
+    AVxWorker *const worker = &pbi->tile_workers[i];
+    winterface->sync(worker);
+    worker->hook = (AVxWorkerHook)tile_worker_hook;
+    worker->data1 = &pbi->tile_worker_data[i];
+    worker->data2 = &pbi->tile_worker_info[i];
+  }
+
+  // Initialize thread frame counts.
+  if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+    for (i = 0; i < num_workers; ++i) {
+      TileWorkerData *const twd = (TileWorkerData *)pbi->tile_workers[i].data1;
+      av1_zero(twd->counts);
+    }
+  }
+
+  // Load tile data into tile_buffers
+  get_tile_buffers(pbi, data, data_end, tile_buffers);
+
+  for (tile_row = tile_rows_start; tile_row < tile_rows_end; ++tile_row) {
+    // Sort the buffers in this tile row based on size in descending order.
+    qsort(&tile_buffers[tile_row][tile_cols_start],
+          tile_cols_end - tile_cols_start, sizeof(tile_buffers[0][0]),
+          compare_tile_buffers);
+
+    // Rearrange the tile buffers in this tile row 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;
+      for (group_start = tile_cols_start; group_start < tile_cols_end;
+           group_start += num_workers) {
+        const int group_end = AOMMIN(group_start + num_workers, tile_cols);
+        const TileBufferDec largest = tile_buffers[tile_row][group_start];
+        memmove(&tile_buffers[tile_row][group_start],
+                &tile_buffers[tile_row][group_start + 1],
+                (group_end - group_start - 1) * sizeof(tile_buffers[0][0]));
+        tile_buffers[tile_row][group_end - 1] = largest;
+      }
+    }
+
+    for (tile_col = tile_cols_start; tile_col < tile_cols_end;) {
+      // Launch workers for individual columns
+      for (i = 0; i < num_workers && tile_col < tile_cols_end;
+           ++i, ++tile_col) {
+        TileBufferDec *const buf = &tile_buffers[tile_row][tile_col];
+        AVxWorker *const worker = &pbi->tile_workers[i];
+        TileWorkerData *const twd = (TileWorkerData *)worker->data1;
+        TileInfo *const tile_info = (TileInfo *)worker->data2;
+
+        twd->pbi = pbi;
+        twd->xd = pbi->mb;
+        twd->xd.corrupted = 0;
+        twd->xd.counts =
+            cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD
+                ? &twd->counts
+                : NULL;
+        av1_zero(twd->dqcoeff);
+        av1_tile_init(tile_info, cm, tile_row, buf->col);
+        av1_tile_init(&twd->xd.tile, cm, tile_row, buf->col);
+        setup_bool_decoder(buf->data, data_end, buf->size, &cm->error,
+                           &twd->bit_reader,
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+                           1 << cm->ans_window_size_log2,
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+                           pbi->decrypt_cb, pbi->decrypt_state);
+        av1_init_macroblockd(cm, &twd->xd,
+#if CONFIG_PVQ
+                             twd->pvq_ref_coeff,
+#endif
+#if CONFIG_CFL
+                             &twd->cfl,
+#endif
+                             twd->dqcoeff);
+#if CONFIG_PVQ
+        daala_dec_init(cm, &twd->xd.daala_dec, &twd->bit_reader);
+        twd->xd.daala_dec.state.adapt = &twd->tctx.pvq_context;
+#endif
+#if CONFIG_EC_ADAPT
+        // Initialise the tile context from the frame context
+        twd->tctx = *cm->fc;
+        twd->xd.tile_ctx = &twd->tctx;
+#endif
+#if CONFIG_PALETTE
+        twd->xd.plane[0].color_index_map = twd->color_index_map[0];
+        twd->xd.plane[1].color_index_map = twd->color_index_map[1];
+#endif  // CONFIG_PALETTE
+
+        worker->had_error = 0;
+        if (i == num_workers - 1 || tile_col == tile_cols_end - 1) {
+          winterface->execute(worker);
+        } else {
+          winterface->launch(worker);
+        }
+
+#if !(CONFIG_ANS || CONFIG_EXT_TILE)
+        if (tile_row == tile_rows - 1 && buf->col == tile_cols - 1) {
+          final_worker = i;
+        }
+#endif  // !(CONFIG_ANS || CONFIG_EXT_TILE)
+      }
+
+      // Sync all workers
+      for (; i > 0; --i) {
+        AVxWorker *const worker = &pbi->tile_workers[i - 1];
+        // TODO(jzern): The tile may have specific error data associated with
+        // its aom_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);
+      }
+    }
+  }
+
+  // Accumulate thread frame counts.
+  if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+    for (i = 0; i < num_workers; ++i) {
+      TileWorkerData *const twd = (TileWorkerData *)pbi->tile_workers[i].data1;
+      av1_accumulate_frame_counts(&cm->counts, &twd->counts);
+    }
+  }
+
+#if CONFIG_EXT_TILE
+  // Return the end of the last tile buffer
+  return tile_buffers[tile_rows - 1][tile_cols - 1].raw_data_end;
+#else
+#if CONFIG_ANS
+  return data_end;
+#else
+  assert(final_worker != -1);
+  {
+    TileWorkerData *const twd =
+        (TileWorkerData *)pbi->tile_workers[final_worker].data1;
+    return aom_reader_find_end(&twd->bit_reader);
+  }
+#endif  // CONFIG_ANS
+#endif  // CONFIG_EXT_TILE
+}
+
+static void error_handler(void *data) {
+  AV1_COMMON *const cm = (AV1_COMMON *)data;
+  aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME, "Truncated packet");
+}
+
+static void read_bitdepth_colorspace_sampling(AV1_COMMON *cm,
+                                              struct aom_read_bit_buffer *rb) {
+  if (cm->profile >= PROFILE_2) {
+    cm->bit_depth = aom_rb_read_bit(rb) ? AOM_BITS_12 : AOM_BITS_10;
+  } else {
+    cm->bit_depth = AOM_BITS_8;
+  }
+
+#if CONFIG_HIGHBITDEPTH
+  if (cm->bit_depth > AOM_BITS_8) {
+    cm->use_highbitdepth = 1;
+  } else {
+#if CONFIG_LOWBITDEPTH
+    cm->use_highbitdepth = 0;
+#else
+    cm->use_highbitdepth = 1;
+#endif
+  }
+#endif
+
+  cm->color_space = aom_rb_read_literal(rb, 3);
+  if (cm->color_space != AOM_CS_SRGB) {
+    // [16,235] (including xvycc) vs [0,255] range
+    cm->color_range = aom_rb_read_bit(rb);
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      cm->subsampling_x = aom_rb_read_bit(rb);
+      cm->subsampling_y = aom_rb_read_bit(rb);
+      if (cm->subsampling_x == 1 && cm->subsampling_y == 1)
+        aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                           "4:2:0 color not supported in profile 1 or 3");
+      if (aom_rb_read_bit(rb))
+        aom_internal_error(&cm->error, AOM_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 (aom_rb_read_bit(rb))
+        aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                           "Reserved bit set");
+    } else {
+      aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                         "4:4:4 color not supported in profile 0 or 2");
+    }
+  }
+}
+
+#if CONFIG_REFERENCE_BUFFER
+void read_sequence_header(SequenceHeader *seq_params) {
+  /* Placeholder for actually reading from the bitstream */
+  seq_params->frame_id_numbers_present_flag = FRAME_ID_NUMBERS_PRESENT_FLAG;
+  seq_params->frame_id_length_minus7 = FRAME_ID_LENGTH_MINUS7;
+  seq_params->delta_frame_id_length_minus2 = DELTA_FRAME_ID_LENGTH_MINUS2;
+}
+#endif
+
+static size_t read_uncompressed_header(AV1Decoder *pbi,
+                                       struct aom_read_bit_buffer *rb) {
+  AV1_COMMON *const cm = &pbi->common;
+  MACROBLOCKD *const xd = &pbi->mb;
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = pool->frame_bufs;
+  int i, mask, ref_index = 0;
+  size_t sz;
+
+#if CONFIG_REFERENCE_BUFFER
+  /* TODO: Move outside frame loop or inside key-frame branch */
+  read_sequence_header(&pbi->seq_params);
+#endif
+
+  cm->last_frame_type = cm->frame_type;
+  cm->last_intra_only = cm->intra_only;
+
+#if CONFIG_EXT_REFS
+  // NOTE: By default all coded frames to be used as a reference
+  cm->is_reference_frame = 1;
+#endif  // CONFIG_EXT_REFS
+
+  if (aom_rb_read_literal(rb, 2) != AOM_FRAME_MARKER)
+    aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                       "Invalid frame marker");
+
+  cm->profile = av1_read_profile(rb);
+
+  const BITSTREAM_PROFILE MAX_SUPPORTED_PROFILE =
+      CONFIG_HIGHBITDEPTH ? MAX_PROFILES : PROFILE_2;
+
+  if (cm->profile >= MAX_SUPPORTED_PROFILE)
+    aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                       "Unsupported bitstream profile");
+
+  cm->show_existing_frame = aom_rb_read_bit(rb);
+
+  if (cm->show_existing_frame) {
+    // Show an existing frame directly.
+    const int existing_frame_idx = aom_rb_read_literal(rb, 3);
+    const int frame_to_show = cm->ref_frame_map[existing_frame_idx];
+#if CONFIG_REFERENCE_BUFFER
+    if (pbi->seq_params.frame_id_numbers_present_flag) {
+      int frame_id_length = pbi->seq_params.frame_id_length_minus7 + 7;
+      int display_frame_id = aom_rb_read_literal(rb, frame_id_length);
+      /* Compare display_frame_id with ref_frame_id and check valid for
+      * referencing */
+      if (display_frame_id != cm->ref_frame_id[existing_frame_idx] ||
+          cm->valid_for_referencing[existing_frame_idx] == 0)
+        aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                           "Reference buffer frame ID mismatch");
+    }
+#endif
+    lock_buffer_pool(pool);
+    if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1) {
+      unlock_buffer_pool(pool);
+      aom_internal_error(&cm->error, AOM_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);
+
+    cm->lf.filter_level = 0;
+    cm->show_frame = 1;
+    pbi->refresh_frame_flags = 0;
+
+    if (cm->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)aom_rb_read_bit(rb);
+  cm->show_frame = aom_rb_read_bit(rb);
+  cm->error_resilient_mode = aom_rb_read_bit(rb);
+#if CONFIG_REFERENCE_BUFFER
+  if (pbi->seq_params.frame_id_numbers_present_flag) {
+    int frame_id_length = pbi->seq_params.frame_id_length_minus7 + 7;
+    int diff_len = pbi->seq_params.delta_frame_id_length_minus2 + 2;
+    int prev_frame_id = 0;
+    if (cm->frame_type != KEY_FRAME) {
+      prev_frame_id = cm->current_frame_id;
+    }
+    cm->current_frame_id = aom_rb_read_literal(rb, frame_id_length);
+
+    if (cm->frame_type != KEY_FRAME) {
+      int diff_frame_id;
+      if (cm->current_frame_id > prev_frame_id) {
+        diff_frame_id = cm->current_frame_id - prev_frame_id;
+      } else {
+        diff_frame_id =
+            (1 << frame_id_length) + cm->current_frame_id - prev_frame_id;
+      }
+      /* Check current_frame_id for conformance */
+      if (prev_frame_id == cm->current_frame_id ||
+          diff_frame_id >= (1 << (frame_id_length - 1))) {
+        aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                           "Invalid value of current_frame_id");
+      }
+    }
+    /* Check if some frames need to be marked as not valid for referencing */
+    for (i = 0; i < REF_FRAMES; i++) {
+      if (cm->frame_type == KEY_FRAME) {
+        cm->valid_for_referencing[i] = 0;
+      } else if (cm->current_frame_id - (1 << diff_len) > 0) {
+        if (cm->ref_frame_id[i] > cm->current_frame_id ||
+            cm->ref_frame_id[i] < cm->current_frame_id - (1 << diff_len))
+          cm->valid_for_referencing[i] = 0;
+      } else {
+        if (cm->ref_frame_id[i] > cm->current_frame_id &&
+            cm->ref_frame_id[i] <
+                (1 << frame_id_length) + cm->current_frame_id - (1 << diff_len))
+          cm->valid_for_referencing[i] = 0;
+      }
+    }
+  }
+#endif
+  if (cm->frame_type == KEY_FRAME) {
+    if (!av1_read_sync_code(rb))
+      aom_internal_error(&cm->error, AOM_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 < INTER_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;
+    }
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+    cm->ans_window_size_log2 = aom_rb_read_literal(rb, 4) + 8;
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+#if CONFIG_PALETTE
+    cm->allow_screen_content_tools = aom_rb_read_bit(rb);
+#endif  // CONFIG_PALETTE
+  } else {
+    cm->intra_only = cm->show_frame ? 0 : aom_rb_read_bit(rb);
+#if CONFIG_PALETTE
+    if (cm->intra_only) cm->allow_screen_content_tools = aom_rb_read_bit(rb);
+#endif  // CONFIG_PALETTE
+    if (cm->error_resilient_mode) {
+      cm->reset_frame_context = RESET_FRAME_CONTEXT_ALL;
+    } else {
+      if (cm->intra_only) {
+        cm->reset_frame_context = aom_rb_read_bit(rb)
+                                      ? RESET_FRAME_CONTEXT_ALL
+                                      : RESET_FRAME_CONTEXT_CURRENT;
+      } else {
+        cm->reset_frame_context = aom_rb_read_bit(rb)
+                                      ? RESET_FRAME_CONTEXT_CURRENT
+                                      : RESET_FRAME_CONTEXT_NONE;
+        if (cm->reset_frame_context == RESET_FRAME_CONTEXT_CURRENT)
+          cm->reset_frame_context = aom_rb_read_bit(rb)
+                                        ? RESET_FRAME_CONTEXT_ALL
+                                        : RESET_FRAME_CONTEXT_CURRENT;
+      }
+    }
+
+    if (cm->intra_only) {
+      if (!av1_read_sync_code(rb))
+        aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                           "Invalid frame sync code");
+
+      read_bitdepth_colorspace_sampling(cm, rb);
+
+      pbi->refresh_frame_flags = aom_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;
+      }
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+      cm->ans_window_size_log2 = aom_rb_read_literal(rb, 4) + 8;
+#endif
+    } else if (pbi->need_resync != 1) { /* Skip if need resync */
+      pbi->refresh_frame_flags = aom_rb_read_literal(rb, REF_FRAMES);
+
+#if CONFIG_EXT_REFS
+      if (!pbi->refresh_frame_flags) {
+        // NOTE: "pbi->refresh_frame_flags == 0" indicates that the coded frame
+        //       will not be used as a reference
+        cm->is_reference_frame = 0;
+      }
+#endif  // CONFIG_EXT_REFS
+
+      for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+        const int ref = aom_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] = aom_rb_read_bit(rb);
+#if CONFIG_REFERENCE_BUFFER
+        if (pbi->seq_params.frame_id_numbers_present_flag) {
+          int frame_id_length = pbi->seq_params.frame_id_length_minus7 + 7;
+          int diff_len = pbi->seq_params.delta_frame_id_length_minus2 + 2;
+          int delta_frame_id_minus1 = aom_rb_read_literal(rb, diff_len);
+          int ref_frame_id =
+              ((cm->current_frame_id - (delta_frame_id_minus1 + 1) +
+                (1 << frame_id_length)) %
+               (1 << frame_id_length));
+          /* Compare values derived from delta_frame_id_minus1 and
+          * refresh_frame_flags. Also, check valid for referencing */
+          if (ref_frame_id != cm->ref_frame_id[ref] ||
+              cm->valid_for_referencing[ref] == 0)
+            aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                               "Reference buffer frame ID mismatch");
+        }
+#endif
+      }
+
+#if CONFIG_FRAME_SIZE
+      if (cm->error_resilient_mode == 0) {
+        setup_frame_size_with_refs(cm, rb);
+      } else {
+        setup_frame_size(cm, rb);
+      }
+#else
+      setup_frame_size_with_refs(cm, rb);
+#endif
+
+      cm->allow_high_precision_mv = aom_rb_read_bit(rb);
+      cm->interp_filter = read_frame_interp_filter(rb);
+#if CONFIG_TEMPMV_SIGNALING
+      if (!cm->error_resilient_mode) {
+        cm->use_prev_frame_mvs = aom_rb_read_bit(rb);
+      }
+#endif
+      for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+        RefBuffer *const ref_buf = &cm->frame_refs[i];
+#if CONFIG_HIGHBITDEPTH
+        av1_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
+        av1_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_TEMPMV_SIGNALING
+  cm->cur_frame->intra_only = cm->frame_type == KEY_FRAME || cm->intra_only;
+#endif
+
+#if CONFIG_REFERENCE_BUFFER
+  if (pbi->seq_params.frame_id_numbers_present_flag) {
+    /* If bitmask is set, update reference frame id values and
+    mark frames as valid for reference */
+    int refresh_frame_flags =
+        cm->frame_type == KEY_FRAME ? 0xFF : pbi->refresh_frame_flags;
+    for (i = 0; i < REF_FRAMES; i++) {
+      if ((refresh_frame_flags >> i) & 1) {
+        cm->ref_frame_id[i] = cm->current_frame_id;
+        cm->valid_for_referencing[i] = 1;
+      }
+    }
+  }
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+  get_frame_new_buffer(cm)->bit_depth = cm->bit_depth;
+#endif
+  get_frame_new_buffer(cm)->color_space = cm->color_space;
+  get_frame_new_buffer(cm)->color_range = cm->color_range;
+  get_frame_new_buffer(cm)->render_width = cm->render_width;
+  get_frame_new_buffer(cm)->render_height = cm->render_height;
+
+  if (pbi->need_resync) {
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Keyframe / intra-only frame required to reset decoder"
+                       " state");
+  }
+
+  if (!cm->error_resilient_mode) {
+    cm->refresh_frame_context = aom_rb_read_bit(rb)
+                                    ? REFRESH_FRAME_CONTEXT_FORWARD
+                                    : REFRESH_FRAME_CONTEXT_BACKWARD;
+  } else {
+    cm->refresh_frame_context = REFRESH_FRAME_CONTEXT_FORWARD;
+  }
+
+  // This flag will be overridden by the call to av1_setup_past_independence
+  // below, forcing the use of context 0 for those frame types.
+  cm->frame_context_idx = aom_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)
+    av1_setup_past_independence(cm);
+
+#if CONFIG_EXT_PARTITION
+  set_sb_size(cm, aom_rb_read_bit(rb) ? BLOCK_128X128 : BLOCK_64X64);
+#else
+  set_sb_size(cm, BLOCK_64X64);
+#endif  // CONFIG_EXT_PARTITION
+
+  setup_loopfilter(cm, rb);
+#if CONFIG_CDEF
+  setup_cdef(cm, rb);
+#endif
+#if CONFIG_LOOP_RESTORATION
+  decode_restoration_mode(cm, rb);
+#endif  // CONFIG_LOOP_RESTORATION
+  setup_quantization(cm, rb);
+#if CONFIG_HIGHBITDEPTH
+  xd->bd = (int)cm->bit_depth;
+#endif
+
+#if CONFIG_Q_ADAPT_PROBS
+  av1_default_coef_probs(cm);
+  if (cm->frame_type == KEY_FRAME || cm->error_resilient_mode ||
+      cm->reset_frame_context == RESET_FRAME_CONTEXT_ALL) {
+    for (i = 0; i < FRAME_CONTEXTS; ++i) cm->frame_contexts[i] = *cm->fc;
+  } else if (cm->reset_frame_context == RESET_FRAME_CONTEXT_CURRENT) {
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+  }
+#endif  // CONFIG_Q_ADAPT_PROBS
+
+  setup_segmentation(cm, rb);
+
+#if CONFIG_DELTA_Q
+  {
+    struct segmentation *const seg = &cm->seg;
+    int segment_quantizer_active = 0;
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+      if (segfeature_active(seg, i, SEG_LVL_ALT_Q)) {
+        segment_quantizer_active = 1;
+      }
+    }
+
+    cm->delta_q_res = 1;
+#if CONFIG_EXT_DELTA_Q
+    cm->delta_lf_res = 1;
+#endif
+    if (segment_quantizer_active == 0 && cm->base_qindex > 0) {
+      cm->delta_q_present_flag = aom_rb_read_bit(rb);
+    } else {
+      cm->delta_q_present_flag = 0;
+    }
+    if (cm->delta_q_present_flag) {
+      xd->prev_qindex = cm->base_qindex;
+      cm->delta_q_res = 1 << aom_rb_read_literal(rb, 2);
+#if CONFIG_EXT_DELTA_Q
+      if (segment_quantizer_active) {
+        assert(seg->abs_delta == SEGMENT_DELTADATA);
+      }
+      cm->delta_lf_present_flag = aom_rb_read_bit(rb);
+      if (cm->delta_lf_present_flag) {
+        xd->prev_delta_lf_from_base = 0;
+        cm->delta_lf_res = 1 << aom_rb_read_literal(rb, 2);
+      } else {
+        cm->delta_lf_present_flag = 0;
+      }
+#endif  // CONFIG_EXT_DELTA_Q
+    }
+  }
+#endif
+
+  for (i = 0; i < MAX_SEGMENTS; ++i) {
+    const int qindex = cm->seg.enabled
+                           ? av1_get_qindex(&cm->seg, i, cm->base_qindex)
+                           : cm->base_qindex;
+    xd->lossless[i] = qindex == 0 && cm->y_dc_delta_q == 0 &&
+                      cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
+    xd->qindex[i] = qindex;
+  }
+
+  setup_segmentation_dequant(cm);
+  cm->tx_mode = read_tx_mode(cm, xd, rb);
+  cm->reference_mode = read_frame_reference_mode(cm, rb);
+
+#if CONFIG_EXT_TX
+  cm->reduced_tx_set_used = aom_rb_read_bit(rb);
+#endif  // CONFIG_EXT_TX
+
+  read_tile_info(pbi, rb);
+  sz = aom_rb_read_literal(rb, 16);
+
+  if (sz == 0)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Invalid header size");
+  return sz;
+}
+
+#if CONFIG_EXT_TX
+#if !CONFIG_EC_ADAPT
+static void read_ext_tx_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j, k;
+  int s;
+  for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
+    if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+      for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+        if (!use_inter_ext_tx_for_txsize[s][i]) continue;
+        for (j = 0; j < num_ext_tx_set[ext_tx_set_type_inter[s]] - 1; ++j)
+          av1_diff_update_prob(r, &fc->inter_ext_tx_prob[s][i][j], ACCT_STR);
+      }
+    }
+  }
+
+  for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
+    if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+      for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+        if (!use_intra_ext_tx_for_txsize[s][i]) continue;
+        for (j = 0; j < INTRA_MODES; ++j)
+          for (k = 0; k < num_ext_tx_set[ext_tx_set_type_intra[s]] - 1; ++k)
+            av1_diff_update_prob(r, &fc->intra_ext_tx_prob[s][i][j][k],
+                                 ACCT_STR);
+      }
+    }
+  }
+}
+#endif  // !CONFIG_EC_ADAPT
+#else
+
+#endif  // CONFIG_EXT_TX
+#if CONFIG_SUPERTX
+static void read_supertx_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j;
+  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+    for (i = 0; i < PARTITION_SUPERTX_CONTEXTS; ++i) {
+      for (j = TX_8X8; j < TX_SIZES; ++j) {
+        av1_diff_update_prob(r, &fc->supertx_prob[i][j], ACCT_STR);
+      }
+    }
+  }
+}
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_GLOBAL_MOTION
+static void read_global_motion_params(WarpedMotionParams *params,
+                                      WarpedMotionParams *ref_params,
+                                      aom_prob *probs, aom_reader *r,
+                                      int allow_hp) {
+  TransformationType type =
+      aom_read_tree(r, av1_global_motion_types_tree, probs, ACCT_STR);
+  int trans_bits;
+  int trans_dec_factor;
+  int trans_prec_diff;
+  set_default_warp_params(params);
+  params->wmtype = type;
+  switch (type) {
+    case HOMOGRAPHY:
+    case HORTRAPEZOID:
+    case VERTRAPEZOID:
+      if (type != HORTRAPEZOID)
+        params->wmmat[6] =
+            aom_read_signed_primitive_refsubexpfin(
+                r, GM_ROW3HOMO_MAX + 1, SUBEXPFIN_K,
+                (ref_params->wmmat[6] >> GM_ROW3HOMO_PREC_DIFF)) *
+            GM_ROW3HOMO_DECODE_FACTOR;
+      if (type != VERTRAPEZOID)
+        params->wmmat[7] =
+            aom_read_signed_primitive_refsubexpfin(
+                r, GM_ROW3HOMO_MAX + 1, SUBEXPFIN_K,
+                (ref_params->wmmat[7] >> GM_ROW3HOMO_PREC_DIFF)) *
+            GM_ROW3HOMO_DECODE_FACTOR;
+    case AFFINE:
+    case ROTZOOM:
+      params->wmmat[2] = aom_read_signed_primitive_refsubexpfin(
+                             r, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+                             (ref_params->wmmat[2] >> GM_ALPHA_PREC_DIFF) -
+                                 (1 << GM_ALPHA_PREC_BITS)) *
+                             GM_ALPHA_DECODE_FACTOR +
+                         (1 << WARPEDMODEL_PREC_BITS);
+      if (type != VERTRAPEZOID)
+        params->wmmat[3] = aom_read_signed_primitive_refsubexpfin(
+                               r, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+                               (ref_params->wmmat[3] >> GM_ALPHA_PREC_DIFF)) *
+                           GM_ALPHA_DECODE_FACTOR;
+      if (type >= AFFINE) {
+        if (type != HORTRAPEZOID)
+          params->wmmat[4] = aom_read_signed_primitive_refsubexpfin(
+                                 r, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+                                 (ref_params->wmmat[4] >> GM_ALPHA_PREC_DIFF)) *
+                             GM_ALPHA_DECODE_FACTOR;
+        params->wmmat[5] = aom_read_signed_primitive_refsubexpfin(
+                               r, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+                               (ref_params->wmmat[5] >> GM_ALPHA_PREC_DIFF) -
+                                   (1 << GM_ALPHA_PREC_BITS)) *
+                               GM_ALPHA_DECODE_FACTOR +
+                           (1 << WARPEDMODEL_PREC_BITS);
+      } else {
+        params->wmmat[4] = -params->wmmat[3];
+        params->wmmat[5] = params->wmmat[2];
+      }
+    // fallthrough intended
+    case TRANSLATION:
+      trans_bits = (type == TRANSLATION) ? GM_ABS_TRANS_ONLY_BITS - !allow_hp
+                                         : GM_ABS_TRANS_BITS;
+      trans_dec_factor = (type == TRANSLATION)
+                             ? GM_TRANS_ONLY_DECODE_FACTOR * (1 << !allow_hp)
+                             : GM_TRANS_DECODE_FACTOR;
+      trans_prec_diff = (type == TRANSLATION)
+                            ? GM_TRANS_ONLY_PREC_DIFF + !allow_hp
+                            : GM_TRANS_PREC_DIFF;
+      params->wmmat[0] = aom_read_signed_primitive_refsubexpfin(
+                             r, (1 << trans_bits) + 1, SUBEXPFIN_K,
+                             (ref_params->wmmat[0] >> trans_prec_diff)) *
+                         trans_dec_factor;
+      params->wmmat[1] = aom_read_signed_primitive_refsubexpfin(
+                             r, (1 << trans_bits) + 1, SUBEXPFIN_K,
+                             (ref_params->wmmat[1] >> trans_prec_diff)) *
+                         trans_dec_factor;
+    case IDENTITY: break;
+    default: assert(0);
+  }
+  if (params->wmtype <= AFFINE)
+    if (!get_shear_params(params)) assert(0);
+}
+
+static void read_global_motion(AV1_COMMON *cm, aom_reader *r) {
+  int frame;
+  for (frame = LAST_FRAME; frame <= ALTREF_FRAME; ++frame) {
+    read_global_motion_params(
+        &cm->global_motion[frame], &cm->prev_frame->global_motion[frame],
+        cm->fc->global_motion_types_prob, r, cm->allow_high_precision_mv);
+    /*
+    printf("Dec Ref %d [%d/%d]: %d %d %d %d\n",
+           frame, cm->current_video_frame, cm->show_frame,
+           cm->global_motion[frame].wmmat[0],
+           cm->global_motion[frame].wmmat[1],
+           cm->global_motion[frame].wmmat[2],
+           cm->global_motion[frame].wmmat[3]);
+           */
+  }
+  memcpy(cm->cur_frame->global_motion, cm->global_motion,
+         TOTAL_REFS_PER_FRAME * sizeof(WarpedMotionParams));
+}
+#endif  // CONFIG_GLOBAL_MOTION
+
+static int read_compressed_header(AV1Decoder *pbi, const uint8_t *data,
+                                  size_t partition_size) {
+  AV1_COMMON *const cm = &pbi->common;
+#if CONFIG_SUPERTX
+  MACROBLOCKD *const xd = &pbi->mb;
+#endif
+  FRAME_CONTEXT *const fc = cm->fc;
+  aom_reader r;
+  int k, i;
+#if !CONFIG_EC_ADAPT || \
+    (CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION || CONFIG_EXT_INTER)
+  int j;
+#endif
+
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  r.window_size = 1 << cm->ans_window_size_log2;
+#endif
+  if (aom_reader_init(&r, data, partition_size, pbi->decrypt_cb,
+                      pbi->decrypt_state))
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate bool decoder 0");
+
+#if CONFIG_LOOP_RESTORATION
+  if (cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[2].frame_restoration_type != RESTORE_NONE) {
+    av1_alloc_restoration_buffers(cm);
+    decode_restoration(cm, &r);
+  }
+#endif
+
+#if !CONFIG_EC_ADAPT
+  if (cm->tx_mode == TX_MODE_SELECT) read_tx_size_probs(fc, &r);
+#endif
+
+#if CONFIG_LV_MAP
+  av1_read_txb_probs(fc, cm->tx_mode, &r);
+#else  // CONFIG_LV_MAP
+#if !CONFIG_PVQ
+#if !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+  read_coef_probs(fc, cm->tx_mode, &r);
+#endif  // !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+#endif  // !CONFIG_PVQ
+#endif  // CONFIG_LV_MAP
+
+#if CONFIG_VAR_TX
+  for (k = 0; k < TXFM_PARTITION_CONTEXTS; ++k)
+    av1_diff_update_prob(&r, &fc->txfm_partition_prob[k], ACCT_STR);
+#endif  // CONFIG_VAR_TX
+  for (k = 0; k < SKIP_CONTEXTS; ++k)
+    av1_diff_update_prob(&r, &fc->skip_probs[k], ACCT_STR);
+
+#if CONFIG_DELTA_Q && !CONFIG_EC_ADAPT
+#if CONFIG_EXT_DELTA_Q
+  if (cm->delta_q_present_flag) {
+    for (k = 0; k < DELTA_Q_PROBS; ++k)
+      av1_diff_update_prob(&r, &fc->delta_q_prob[k], ACCT_STR);
+  }
+  if (cm->delta_lf_present_flag) {
+    for (k = 0; k < DELTA_LF_PROBS; ++k)
+      av1_diff_update_prob(&r, &fc->delta_lf_prob[k], ACCT_STR);
+  }
+#else
+  for (k = 0; k < DELTA_Q_PROBS; ++k)
+    av1_diff_update_prob(&r, &fc->delta_q_prob[k], ACCT_STR);
+#endif
+#endif
+
+#if !CONFIG_EC_ADAPT
+  if (cm->seg.enabled && cm->seg.update_map) {
+    if (cm->seg.temporal_update) {
+      for (k = 0; k < PREDICTION_PROBS; k++)
+        av1_diff_update_prob(&r, &cm->fc->seg.pred_probs[k], ACCT_STR);
+    }
+    for (k = 0; k < MAX_SEGMENTS - 1; k++)
+      av1_diff_update_prob(&r, &cm->fc->seg.tree_probs[k], ACCT_STR);
+  }
+
+  for (j = 0; j < INTRA_MODES; j++) {
+    for (i = 0; i < INTRA_MODES - 1; ++i)
+      av1_diff_update_prob(&r, &fc->uv_mode_prob[j][i], ACCT_STR);
+  }
+
+#if CONFIG_EXT_PARTITION_TYPES
+  for (j = 0; j < PARTITION_PLOFFSET; ++j)
+    for (i = 0; i < PARTITION_TYPES - 1; ++i)
+      av1_diff_update_prob(&r, &fc->partition_prob[j][i], ACCT_STR);
+  for (; j < PARTITION_CONTEXTS_PRIMARY; ++j)
+    for (i = 0; i < EXT_PARTITION_TYPES - 1; ++i)
+      av1_diff_update_prob(&r, &fc->partition_prob[j][i], ACCT_STR);
+#else
+  for (j = 0; j < PARTITION_CONTEXTS_PRIMARY; ++j)
+    for (i = 0; i < PARTITION_TYPES - 1; ++i)
+      av1_diff_update_prob(&r, &fc->partition_prob[j][i], ACCT_STR);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+#if CONFIG_UNPOISON_PARTITION_CTX
+  for (; j < PARTITION_CONTEXTS_PRIMARY + PARTITION_BLOCK_SIZES; ++j)
+    av1_diff_update_prob(&r, &fc->partition_prob[j][PARTITION_VERT], ACCT_STR);
+  for (; j < PARTITION_CONTEXTS_PRIMARY + 2 * PARTITION_BLOCK_SIZES; ++j)
+    av1_diff_update_prob(&r, &fc->partition_prob[j][PARTITION_HORZ], ACCT_STR);
+#endif  // CONFIG_UNPOISON_PARTITION_CTX
+
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+  for (i = 0; i < INTRA_FILTERS + 1; ++i)
+    for (j = 0; j < INTRA_FILTERS - 1; ++j)
+      av1_diff_update_prob(&r, &fc->intra_filter_probs[i][j], ACCT_STR);
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+#endif  // !CONFIG_EC_ADAPT
+
+  if (frame_is_intra_only(cm)) {
+    av1_copy(cm->kf_y_prob, av1_kf_y_mode_prob);
+#if CONFIG_EC_MULTISYMBOL
+    av1_copy(cm->fc->kf_y_cdf, av1_kf_y_mode_cdf);
+#endif
+#if !CONFIG_EC_ADAPT
+    for (k = 0; k < INTRA_MODES; k++)
+      for (j = 0; j < INTRA_MODES; j++)
+        for (i = 0; i < INTRA_MODES - 1; ++i)
+          av1_diff_update_prob(&r, &cm->kf_y_prob[k][j][i], ACCT_STR);
+#endif
+  } else {
+#if !CONFIG_REF_MV
+    nmv_context *const nmvc = &fc->nmvc;
+#endif
+    read_inter_mode_probs(fc, &r);
+
+#if CONFIG_EXT_INTER
+    read_inter_compound_mode_probs(fc, &r);
+    if (cm->reference_mode != COMPOUND_REFERENCE) {
+      for (i = 0; i < BLOCK_SIZE_GROUPS; i++) {
+        if (is_interintra_allowed_bsize_group(i)) {
+          av1_diff_update_prob(&r, &fc->interintra_prob[i], ACCT_STR);
+        }
+      }
+      for (i = 0; i < BLOCK_SIZE_GROUPS; i++) {
+        for (j = 0; j < INTERINTRA_MODES - 1; j++)
+          av1_diff_update_prob(&r, &fc->interintra_mode_prob[i][j], ACCT_STR);
+      }
+      for (i = 0; i < BLOCK_SIZES; i++) {
+        if (is_interintra_allowed_bsize(i) && is_interintra_wedge_used(i)) {
+          av1_diff_update_prob(&r, &fc->wedge_interintra_prob[i], ACCT_STR);
+        }
+      }
+    }
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+    if (cm->reference_mode != SINGLE_REFERENCE) {
+      for (i = 0; i < BLOCK_SIZES; i++) {
+        for (j = 0; j < COMPOUND_TYPES - 1; j++) {
+          av1_diff_update_prob(&r, &fc->compound_type_prob[i][j], ACCT_STR);
+        }
+      }
+    }
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    for (i = BLOCK_8X8; i < BLOCK_SIZES; ++i) {
+      for (j = 0; j < MOTION_MODES - 1; ++j)
+        av1_diff_update_prob(&r, &fc->motion_mode_prob[i][j], ACCT_STR);
+    }
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if !CONFIG_EC_ADAPT
+    if (cm->interp_filter == SWITCHABLE) read_switchable_interp_probs(fc, &r);
+#endif
+
+    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+      av1_diff_update_prob(&r, &fc->intra_inter_prob[i], ACCT_STR);
+
+    if (cm->reference_mode != SINGLE_REFERENCE)
+      setup_compound_reference_mode(cm);
+    read_frame_reference_mode_probs(cm, &r);
+
+#if !CONFIG_EC_ADAPT
+    for (j = 0; j < BLOCK_SIZE_GROUPS; j++) {
+      for (i = 0; i < INTRA_MODES - 1; ++i)
+        av1_diff_update_prob(&r, &fc->y_mode_prob[j][i], ACCT_STR);
+    }
+#endif
+
+#if CONFIG_REF_MV
+    for (i = 0; i < NMV_CONTEXTS; ++i)
+      read_mv_probs(&fc->nmvc[i], cm->allow_high_precision_mv, &r);
+#else
+    read_mv_probs(nmvc, cm->allow_high_precision_mv, &r);
+#endif
+#if !CONFIG_EC_ADAPT
+    read_ext_tx_probs(fc, &r);
+#endif  // EC_ADAPT
+#if CONFIG_SUPERTX
+    if (!xd->lossless[0]) read_supertx_probs(fc, &r);
+#endif
+#if CONFIG_GLOBAL_MOTION
+    read_global_motion(cm, &r);
+#endif  // EC_ADAPT, DAALA_EC
+  }
+#if CONFIG_EC_MULTISYMBOL && !CONFIG_EC_ADAPT
+#if CONFIG_NEW_TOKENSET
+  av1_coef_head_cdfs(fc);
+#endif
+  /* Make tail distribution from head */
+  av1_coef_pareto_cdfs(fc);
+#if CONFIG_REF_MV
+  for (i = 0; i < NMV_CONTEXTS; ++i) av1_set_mv_cdfs(&fc->nmvc[i]);
+#else
+  av1_set_mv_cdfs(&fc->nmvc);
+#endif
+  av1_set_mode_cdfs(cm);
+#endif  // CONFIG_EC_MULTISYMBOL && !CONFIG_EC_ADAPT
+
+  return aom_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 AV1_COMMON *const cm) {
+  FRAME_COUNTS zero_counts;
+  av1_zero(zero_counts);
+  assert(cm->refresh_frame_context != REFRESH_FRAME_CONTEXT_BACKWARD ||
+         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)));
+#if CONFIG_EC_MULTISYMBOL
+  assert(!memcmp(cm->counts.blockz_count, zero_counts.blockz_count,
+                 sizeof(cm->counts.blockz_count)));
+#endif
+  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)));
+#if CONFIG_EXT_INTER
+  assert(!memcmp(cm->counts.inter_compound_mode,
+                 zero_counts.inter_compound_mode,
+                 sizeof(cm->counts.inter_compound_mode)));
+  assert(!memcmp(cm->counts.interintra, zero_counts.interintra,
+                 sizeof(cm->counts.interintra)));
+  assert(!memcmp(cm->counts.wedge_interintra, zero_counts.wedge_interintra,
+                 sizeof(cm->counts.wedge_interintra)));
+  assert(!memcmp(cm->counts.compound_interinter,
+                 zero_counts.compound_interinter,
+                 sizeof(cm->counts.compound_interinter)));
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  assert(!memcmp(cm->counts.motion_mode, zero_counts.motion_mode,
+                 sizeof(cm->counts.motion_mode)));
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  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)));
+#if CONFIG_EXT_REFS
+  assert(!memcmp(cm->counts.comp_bwdref, zero_counts.comp_bwdref,
+                 sizeof(cm->counts.comp_bwdref)));
+#endif  // CONFIG_EXT_REFS
+  assert(!memcmp(&cm->counts.tx_size, &zero_counts.tx_size,
+                 sizeof(cm->counts.tx_size)));
+  assert(!memcmp(cm->counts.skip, zero_counts.skip, sizeof(cm->counts.skip)));
+#if CONFIG_REF_MV
+  assert(
+      !memcmp(&cm->counts.mv[0], &zero_counts.mv[0], sizeof(cm->counts.mv[0])));
+  assert(
+      !memcmp(&cm->counts.mv[1], &zero_counts.mv[1], sizeof(cm->counts.mv[0])));
+#else
+  assert(!memcmp(&cm->counts.mv, &zero_counts.mv, sizeof(cm->counts.mv)));
+#endif
+  assert(!memcmp(cm->counts.inter_ext_tx, zero_counts.inter_ext_tx,
+                 sizeof(cm->counts.inter_ext_tx)));
+  assert(!memcmp(cm->counts.intra_ext_tx, zero_counts.intra_ext_tx,
+                 sizeof(cm->counts.intra_ext_tx)));
+}
+#endif  // NDEBUG
+
+static struct aom_read_bit_buffer *init_read_bit_buffer(
+    AV1Decoder *pbi, struct aom_read_bit_buffer *rb, const uint8_t *data,
+    const uint8_t *data_end, uint8_t clear_data[MAX_AV1_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)AOMMIN(MAX_AV1_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;
+}
+
+//------------------------------------------------------------------------------
+
+int av1_read_sync_code(struct aom_read_bit_buffer *const rb) {
+  return aom_rb_read_literal(rb, 8) == AV1_SYNC_CODE_0 &&
+         aom_rb_read_literal(rb, 8) == AV1_SYNC_CODE_1 &&
+         aom_rb_read_literal(rb, 8) == AV1_SYNC_CODE_2;
+}
+
+void av1_read_frame_size(struct aom_read_bit_buffer *rb, int *width,
+                         int *height) {
+  *width = aom_rb_read_literal(rb, 16) + 1;
+  *height = aom_rb_read_literal(rb, 16) + 1;
+}
+
+BITSTREAM_PROFILE av1_read_profile(struct aom_read_bit_buffer *rb) {
+  int profile = aom_rb_read_bit(rb);
+  profile |= aom_rb_read_bit(rb) << 1;
+  if (profile > 2) profile += aom_rb_read_bit(rb);
+  return (BITSTREAM_PROFILE)profile;
+}
+
+#if CONFIG_EC_ADAPT
+static void make_update_tile_list_dec(AV1Decoder *pbi, int tile_rows,
+                                      int tile_cols, FRAME_CONTEXT *ec_ctxs[]) {
+  int i;
+  for (i = 0; i < tile_rows * tile_cols; ++i)
+    ec_ctxs[i] = &pbi->tile_data[i].tctx;
+}
+#endif
+
+void av1_decode_frame(AV1Decoder *pbi, const uint8_t *data,
+                      const uint8_t *data_end, const uint8_t **p_data_end) {
+  AV1_COMMON *const cm = &pbi->common;
+  MACROBLOCKD *const xd = &pbi->mb;
+  struct aom_read_bit_buffer rb;
+  int context_updated = 0;
+  uint8_t clear_data[MAX_AV1_HEADER_SIZE];
+  size_t first_partition_size;
+  YV12_BUFFER_CONFIG *new_fb;
+
+#if CONFIG_ADAPT_SCAN
+  av1_deliver_eob_threshold(cm, xd);
+#endif
+#if CONFIG_BITSTREAM_DEBUG
+  bitstream_queue_set_frame_read(cm->current_video_frame * 2 + cm->show_frame);
+#endif
+
+  first_partition_size = read_uncompressed_header(
+      pbi, init_read_bit_buffer(pbi, &rb, data, data_end, clear_data));
+
+#if CONFIG_EXT_TILE
+  // If cm->tile_encoding_mode == TILE_NORMAL, the independent decoding of a
+  // single tile or a section of a frame is not allowed.
+  if (!cm->tile_encoding_mode &&
+      (pbi->dec_tile_row >= 0 || pbi->dec_tile_col >= 0)) {
+    pbi->dec_tile_row = -1;
+    pbi->dec_tile_col = -1;
+  }
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_TILE_GROUPS
+  pbi->first_partition_size = first_partition_size;
+  pbi->uncomp_hdr_size = aom_rb_bytes_read(&rb);
+#endif
+  new_fb = get_frame_new_buffer(cm);
+  xd->cur_buf = new_fb;
+#if CONFIG_GLOBAL_MOTION
+  int i;
+  for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
+    set_default_warp_params(&cm->global_motion[i]);
+    set_default_warp_params(&cm->cur_frame->global_motion[i]);
+  }
+  xd->global_motion = cm->global_motion;
+#endif  // CONFIG_GLOBAL_MOTION
+
+  if (!first_partition_size) {
+    // showing a frame directly
+    *p_data_end = data + aom_rb_bytes_read(&rb);
+    return;
+  }
+
+  data += aom_rb_bytes_read(&rb);
+  if (!read_is_valid(data, first_partition_size, data_end))
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt header length");
+
+#if CONFIG_REF_MV
+  cm->setup_mi(cm);
+#endif
+
+#if CONFIG_TEMPMV_SIGNALING
+  if (cm->use_prev_frame_mvs) {
+    RefBuffer *last_fb_ref_buf = &cm->frame_refs[LAST_FRAME - LAST_FRAME];
+    cm->prev_frame = &cm->buffer_pool->frame_bufs[last_fb_ref_buf->idx];
+    assert(!cm->error_resilient_mode &&
+           cm->width == last_fb_ref_buf->buf->y_width &&
+           cm->height == last_fb_ref_buf->buf->y_height &&
+           !cm->prev_frame->intra_only);
+  }
+#else
+  cm->use_prev_frame_mvs =
+      !cm->error_resilient_mode && cm->width == cm->last_width &&
+      cm->height == cm->last_height && !cm->last_intra_only &&
+      cm->last_show_frame && (cm->last_frame_type != KEY_FRAME);
+#endif
+#if CONFIG_EXT_REFS
+  // NOTE(zoeliu): As cm->prev_frame can take neither a frame of
+  //               show_exisiting_frame=1, nor can it take a frame not used as
+  //               a reference, it is probable that by the time it is being
+  //               referred to, the frame buffer it originally points to may
+  //               already get expired and have been reassigned to the current
+  //               newly coded frame. Hence, we need to check whether this is
+  //               the case, and if yes, we have 2 choices:
+  //               (1) Simply disable the use of previous frame mvs; or
+  //               (2) Have cm->prev_frame point to one reference frame buffer,
+  //                   e.g. LAST_FRAME.
+  if (cm->use_prev_frame_mvs && !dec_is_ref_frame_buf(pbi, cm->prev_frame)) {
+    // Reassign the LAST_FRAME buffer to cm->prev_frame.
+    RefBuffer *last_fb_ref_buf = &cm->frame_refs[LAST_FRAME - LAST_FRAME];
+    cm->prev_frame = &cm->buffer_pool->frame_bufs[last_fb_ref_buf->idx];
+  }
+#endif  // CONFIG_EXT_REFS
+
+  av1_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);
+
+  *cm->fc = cm->frame_contexts[cm->frame_context_idx];
+  if (!cm->fc->initialized)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Uninitialized entropy context.");
+
+  av1_zero(cm->counts);
+
+  xd->corrupted = 0;
+  new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
+  if (new_fb->corrupted)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data header is corrupted.");
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    av1_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 (cm->frame_parallel_decode &&
+      cm->refresh_frame_context != REFRESH_FRAME_CONTEXT_BACKWARD) {
+    AVxWorker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_FORWARD) {
+      context_updated = 1;
+      cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+    }
+    av1_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.
+    av1_frameworker_signal_stats(worker);
+    av1_frameworker_unlock_stats(worker);
+  }
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  av1_copy(cm->starting_coef_probs, cm->fc->coef_probs);
+  cm->coef_probs_update_idx = 0;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+  if (pbi->max_threads > 1 && !CONFIG_CB4X4 &&
+#if CONFIG_EXT_TILE
+      pbi->dec_tile_col < 0 &&  // Decoding all columns
+#endif                          // CONFIG_EXT_TILE
+      cm->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.
+        av1_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 {
+      aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                         "Decode failed. Frame data is corrupted.");
+    }
+  } else {
+    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
+  }
+
+#if CONFIG_CDEF
+  if (!cm->skip_loop_filter) {
+    av1_cdef_frame(&pbi->cur_buf->buf, cm, &pbi->mb);
+  }
+#endif  // CONFIG_CDEF
+
+#if CONFIG_LOOP_RESTORATION
+  if (cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[2].frame_restoration_type != RESTORE_NONE) {
+    av1_loop_restoration_frame(new_fb, cm, cm->rst_info, 7, 0, NULL);
+  }
+#endif  // CONFIG_LOOP_RESTORATION
+
+  if (!xd->corrupted) {
+    if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+#if CONFIG_EC_ADAPT
+      FRAME_CONTEXT **tile_ctxs = aom_malloc(cm->tile_rows * cm->tile_cols *
+                                             sizeof(&pbi->tile_data[0].tctx));
+      aom_cdf_prob **cdf_ptrs =
+          aom_malloc(cm->tile_rows * cm->tile_cols *
+                     sizeof(&pbi->tile_data[0].tctx.partition_cdf[0][0]));
+      make_update_tile_list_dec(pbi, cm->tile_rows, cm->tile_cols, tile_ctxs);
+#endif
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+      cm->partial_prob_update = 0;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+      av1_adapt_coef_probs(cm);
+      av1_adapt_intra_frame_probs(cm);
+#if CONFIG_EC_ADAPT
+      av1_average_tile_coef_cdfs(pbi->common.fc, tile_ctxs, cdf_ptrs,
+                                 cm->tile_rows * cm->tile_cols);
+      av1_average_tile_intra_cdfs(pbi->common.fc, tile_ctxs, cdf_ptrs,
+                                  cm->tile_rows * cm->tile_cols);
+#if CONFIG_PVQ
+      av1_average_tile_pvq_cdfs(pbi->common.fc, tile_ctxs,
+                                cm->tile_rows * cm->tile_cols);
+#endif  // CONFIG_PVQ
+#endif  // CONFIG_EC_ADAPT
+#if CONFIG_ADAPT_SCAN
+      av1_adapt_scan_order(cm);
+#endif  // CONFIG_ADAPT_SCAN
+
+      if (!frame_is_intra_only(cm)) {
+        av1_adapt_inter_frame_probs(cm);
+        av1_adapt_mv_probs(cm, cm->allow_high_precision_mv);
+#if CONFIG_EC_ADAPT
+        av1_average_tile_inter_cdfs(&pbi->common, pbi->common.fc, tile_ctxs,
+                                    cdf_ptrs, cm->tile_rows * cm->tile_cols);
+        av1_average_tile_mv_cdfs(pbi->common.fc, tile_ctxs, cdf_ptrs,
+                                 cm->tile_rows * cm->tile_cols);
+#endif
+      }
+#if CONFIG_EC_ADAPT
+      aom_free(tile_ctxs);
+      aom_free(cdf_ptrs);
+#endif
+    } else {
+      debug_check_frame_counts(cm);
+    }
+  } else {
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data is corrupted.");
+  }
+
+#if CONFIG_INSPECTION
+  if (pbi->inspect_cb != NULL) {
+    (*pbi->inspect_cb)(pbi, pbi->inspect_ctx);
+  }
+#endif
+
+  // Non frame parallel update frame context here.
+  if (!cm->error_resilient_mode && !context_updated)
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+}
diff --git a/third_party/aom/av1/decoder/decodeframe.h b/third_party/aom/av1/decoder/decodeframe.h
new file mode 100644
index 000000000..a904658b0
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodeframe.h
@@ -0,0 +1,39 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DECODEFRAME_H_
+#define AV1_DECODER_DECODEFRAME_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1Decoder;
+struct aom_read_bit_buffer;
+
+#if CONFIG_REFERENCE_BUFFER
+/* Placeholder for now */
+void read_sequence_header(SequenceHeader *seq_params);
+#endif
+
+int av1_read_sync_code(struct aom_read_bit_buffer *const rb);
+void av1_read_frame_size(struct aom_read_bit_buffer *rb, int *width,
+                         int *height);
+BITSTREAM_PROFILE av1_read_profile(struct aom_read_bit_buffer *rb);
+
+void av1_decode_frame(struct AV1Decoder *pbi, const uint8_t *data,
+                      const uint8_t *data_end, const uint8_t **p_data_end);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_DECODER_DECODEFRAME_H_
diff --git a/third_party/aom/av1/decoder/decodemv.c b/third_party/aom/av1/decoder/decodemv.c
new file mode 100644
index 000000000..ec0f87751
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodemv.c
@@ -0,0 +1,2405 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/entropymv.h"
+#include "av1/common/mvref_common.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/reconinter.h"
+#if CONFIG_EXT_INTRA
+#include "av1/common/reconintra.h"
+#endif  // CONFIG_EXT_INTRA
+#include "av1/common/seg_common.h"
+#if CONFIG_WARPED_MOTION
+#include "av1/common/warped_motion.h"
+#endif  // CONFIG_WARPED_MOTION
+
+#include "av1/decoder/decodeframe.h"
+#include "av1/decoder/decodemv.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+
+#define ACCT_STR __func__
+#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
+static INLINE int read_uniform(aom_reader *r, int n) {
+  const int l = get_unsigned_bits(n);
+  const int m = (1 << l) - n;
+  const int v = aom_read_literal(r, l - 1, ACCT_STR);
+  assert(l != 0);
+  if (v < m)
+    return v;
+  else
+    return (v << 1) - m + aom_read_literal(r, 1, ACCT_STR);
+}
+#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
+
+#if CONFIG_EC_MULTISYMBOL
+static PREDICTION_MODE read_intra_mode(aom_reader *r, aom_cdf_prob *cdf) {
+  return (PREDICTION_MODE)
+      av1_intra_mode_inv[aom_read_symbol(r, cdf, INTRA_MODES, ACCT_STR)];
+}
+#else
+static PREDICTION_MODE read_intra_mode(aom_reader *r, const aom_prob *p) {
+  return (PREDICTION_MODE)aom_read_tree(r, av1_intra_mode_tree, p, ACCT_STR);
+}
+#endif
+
+#if CONFIG_DELTA_Q
+static int read_delta_qindex(AV1_COMMON *cm, MACROBLOCKD *xd, aom_reader *r,
+                             MB_MODE_INFO *const mbmi, int mi_col, int mi_row) {
+  FRAME_COUNTS *counts = xd->counts;
+  int sign, abs, reduced_delta_qindex = 0;
+  BLOCK_SIZE bsize = mbmi->sb_type;
+  const int b_col = mi_col & MAX_MIB_MASK;
+  const int b_row = mi_row & MAX_MIB_MASK;
+  const int read_delta_q_flag = (b_col == 0 && b_row == 0);
+  int rem_bits, thr;
+  int i, smallval;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  if ((bsize != BLOCK_LARGEST || mbmi->skip == 0) && read_delta_q_flag) {
+#if !CONFIG_EC_MULTISYMBOL
+    int bit = 1;
+    abs = 0;
+    while (abs < DELTA_Q_SMALL && bit) {
+      bit = aom_read(r, ec_ctx->delta_q_prob[abs], ACCT_STR);
+      abs += bit;
+    }
+#else
+    abs = aom_read_symbol(r, ec_ctx->delta_q_cdf, DELTA_Q_PROBS + 1, ACCT_STR);
+#endif
+    smallval = (abs < DELTA_Q_SMALL);
+    if (counts) {
+      for (i = 0; i < abs; ++i) counts->delta_q[i][1]++;
+      if (smallval) counts->delta_q[abs][0]++;
+    }
+
+    if (!smallval) {
+      rem_bits = aom_read_literal(r, 3, ACCT_STR);
+      thr = (1 << rem_bits) + 1;
+      abs = aom_read_literal(r, rem_bits, ACCT_STR) + thr;
+    }
+
+    if (abs) {
+      sign = aom_read_bit(r, ACCT_STR);
+    } else {
+      sign = 1;
+    }
+
+    reduced_delta_qindex = sign ? -abs : abs;
+  }
+  return reduced_delta_qindex;
+}
+#if CONFIG_EXT_DELTA_Q
+static int read_delta_lflevel(AV1_COMMON *cm, MACROBLOCKD *xd, aom_reader *r,
+                              MB_MODE_INFO *const mbmi, int mi_col,
+                              int mi_row) {
+  FRAME_COUNTS *counts = xd->counts;
+  int sign, abs, reduced_delta_lflevel = 0;
+  BLOCK_SIZE bsize = mbmi->sb_type;
+  const int b_col = mi_col & MAX_MIB_MASK;
+  const int b_row = mi_row & MAX_MIB_MASK;
+  const int read_delta_lf_flag = (b_col == 0 && b_row == 0);
+  int rem_bits, thr;
+  int i, smallval;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  if ((bsize != BLOCK_64X64 || mbmi->skip == 0) && read_delta_lf_flag) {
+#if !CONFIG_EC_MULTISYMBOL
+    int bit = 1;
+    abs = 0;
+    while (abs < DELTA_LF_SMALL && bit) {
+      bit = aom_read(r, ec_ctx->delta_lf_prob[abs], ACCT_STR);
+      abs += bit;
+    }
+#else
+    abs =
+        aom_read_symbol(r, ec_ctx->delta_lf_cdf, DELTA_LF_PROBS + 1, ACCT_STR);
+#endif
+    smallval = (abs < DELTA_LF_SMALL);
+    if (counts) {
+      for (i = 0; i < abs; ++i) counts->delta_lf[i][1]++;
+      if (smallval) counts->delta_lf[abs][0]++;
+    }
+    if (!smallval) {
+      rem_bits = aom_read_literal(r, 3, ACCT_STR);
+      thr = (1 << rem_bits) + 1;
+      abs = aom_read_literal(r, rem_bits, ACCT_STR) + thr;
+    }
+
+    if (abs) {
+      sign = aom_read_bit(r, ACCT_STR);
+    } else {
+      sign = 1;
+    }
+
+    reduced_delta_lflevel = sign ? -abs : abs;
+  }
+  return reduced_delta_lflevel;
+}
+#endif
+#endif
+
+static PREDICTION_MODE read_intra_mode_y(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                         aom_reader *r, int size_group) {
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  const PREDICTION_MODE y_mode =
+#if CONFIG_EC_MULTISYMBOL
+      read_intra_mode(r, ec_ctx->y_mode_cdf[size_group]);
+#else
+      read_intra_mode(r, cm->fc->y_mode_prob[size_group]);
+#endif
+  FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_ADAPT
+  (void)cm;
+#endif
+  if (counts) ++counts->y_mode[size_group][y_mode];
+  return y_mode;
+}
+
+static PREDICTION_MODE read_intra_mode_uv(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                          aom_reader *r,
+                                          PREDICTION_MODE y_mode) {
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  const PREDICTION_MODE uv_mode =
+#if CONFIG_EC_MULTISYMBOL
+      read_intra_mode(r, ec_ctx->uv_mode_cdf[y_mode]);
+#else
+      read_intra_mode(r, cm->fc->uv_mode_prob[y_mode]);
+#endif
+  FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_ADAPT
+  (void)cm;
+#endif
+  if (counts) ++counts->uv_mode[y_mode][uv_mode];
+  return uv_mode;
+}
+
+#if CONFIG_EXT_INTER
+static INTERINTRA_MODE read_interintra_mode(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                            aom_reader *r, int size_group) {
+  const INTERINTRA_MODE ii_mode = (INTERINTRA_MODE)aom_read_tree(
+      r, av1_interintra_mode_tree, cm->fc->interintra_mode_prob[size_group],
+      ACCT_STR);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts) ++counts->interintra_mode[size_group][ii_mode];
+  return ii_mode;
+}
+#endif  // CONFIG_EXT_INTER
+
+static PREDICTION_MODE read_inter_mode(FRAME_CONTEXT *ec_ctx, MACROBLOCKD *xd,
+                                       aom_reader *r, int16_t ctx) {
+#if CONFIG_REF_MV
+  FRAME_COUNTS *counts = xd->counts;
+  int16_t mode_ctx = ctx & NEWMV_CTX_MASK;
+  aom_prob mode_prob = ec_ctx->newmv_prob[mode_ctx];
+
+  if (aom_read(r, mode_prob, ACCT_STR) == 0) {
+    if (counts) ++counts->newmv_mode[mode_ctx][0];
+    return NEWMV;
+  }
+  if (counts) ++counts->newmv_mode[mode_ctx][1];
+
+  if (ctx & (1 << ALL_ZERO_FLAG_OFFSET)) return ZEROMV;
+
+  mode_ctx = (ctx >> ZEROMV_OFFSET) & ZEROMV_CTX_MASK;
+
+  mode_prob = ec_ctx->zeromv_prob[mode_ctx];
+  if (aom_read(r, mode_prob, ACCT_STR) == 0) {
+    if (counts) ++counts->zeromv_mode[mode_ctx][0];
+    return ZEROMV;
+  }
+  if (counts) ++counts->zeromv_mode[mode_ctx][1];
+
+  mode_ctx = (ctx >> REFMV_OFFSET) & REFMV_CTX_MASK;
+
+  if (ctx & (1 << SKIP_NEARESTMV_OFFSET)) mode_ctx = 6;
+  if (ctx & (1 << SKIP_NEARMV_OFFSET)) mode_ctx = 7;
+  if (ctx & (1 << SKIP_NEARESTMV_SUB8X8_OFFSET)) mode_ctx = 8;
+
+  mode_prob = ec_ctx->refmv_prob[mode_ctx];
+
+  if (aom_read(r, mode_prob, ACCT_STR) == 0) {
+    if (counts) ++counts->refmv_mode[mode_ctx][0];
+
+    return NEARESTMV;
+  } else {
+    if (counts) ++counts->refmv_mode[mode_ctx][1];
+    return NEARMV;
+  }
+
+  // Invalid prediction mode.
+  assert(0);
+#else
+#if CONFIG_EC_MULTISYMBOL
+  const int mode = av1_inter_mode_inv[aom_read_symbol(
+      r, ec_ctx->inter_mode_cdf[ctx], INTER_MODES, ACCT_STR)];
+#else
+  const int mode = aom_read_tree(r, av1_inter_mode_tree,
+                                 ec_ctx->inter_mode_probs[ctx], ACCT_STR);
+#endif
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts) ++counts->inter_mode[ctx][mode];
+
+  return NEARESTMV + mode;
+#endif
+}
+
+#if CONFIG_REF_MV
+static void read_drl_idx(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                         MB_MODE_INFO *mbmi, aom_reader *r) {
+  uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+  mbmi->ref_mv_idx = 0;
+
+#if CONFIG_EXT_INTER
+  if (mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV) {
+#else
+  if (mbmi->mode == NEWMV) {
+#endif
+    int idx;
+    for (idx = 0; idx < 2; ++idx) {
+      if (xd->ref_mv_count[ref_frame_type] > idx + 1) {
+        uint8_t drl_ctx = av1_drl_ctx(xd->ref_mv_stack[ref_frame_type], idx);
+        aom_prob drl_prob = cm->fc->drl_prob[drl_ctx];
+        if (!aom_read(r, drl_prob, ACCT_STR)) {
+          mbmi->ref_mv_idx = idx;
+          if (xd->counts) ++xd->counts->drl_mode[drl_ctx][0];
+          return;
+        }
+        mbmi->ref_mv_idx = idx + 1;
+        if (xd->counts) ++xd->counts->drl_mode[drl_ctx][1];
+      }
+    }
+  }
+
+  if (have_nearmv_in_inter_mode(mbmi->mode)) {
+    int idx;
+    // Offset the NEARESTMV mode.
+    // TODO(jingning): Unify the two syntax decoding loops after the NEARESTMV
+    // mode is factored in.
+    for (idx = 1; idx < 3; ++idx) {
+      if (xd->ref_mv_count[ref_frame_type] > idx + 1) {
+        uint8_t drl_ctx = av1_drl_ctx(xd->ref_mv_stack[ref_frame_type], idx);
+        aom_prob drl_prob = cm->fc->drl_prob[drl_ctx];
+        if (!aom_read(r, drl_prob, ACCT_STR)) {
+          mbmi->ref_mv_idx = idx - 1;
+          if (xd->counts) ++xd->counts->drl_mode[drl_ctx][0];
+          return;
+        }
+        mbmi->ref_mv_idx = idx;
+        if (xd->counts) ++xd->counts->drl_mode[drl_ctx][1];
+      }
+    }
+  }
+}
+#endif
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+static MOTION_MODE read_motion_mode(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                    MODE_INFO *mi, aom_reader *r) {
+  MB_MODE_INFO *mbmi = &mi->mbmi;
+  const MOTION_MODE last_motion_mode_allowed = motion_mode_allowed(
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      0, xd->global_motion,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      mi);
+  int motion_mode;
+  FRAME_COUNTS *counts = xd->counts;
+
+  if (last_motion_mode_allowed == SIMPLE_TRANSLATION) return SIMPLE_TRANSLATION;
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  if (last_motion_mode_allowed == OBMC_CAUSAL) {
+    motion_mode = aom_read(r, cm->fc->obmc_prob[mbmi->sb_type], ACCT_STR);
+    if (counts) ++counts->obmc[mbmi->sb_type][motion_mode];
+    return (MOTION_MODE)(SIMPLE_TRANSLATION + motion_mode);
+  } else {
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+    motion_mode =
+        aom_read_tree(r, av1_motion_mode_tree,
+                      cm->fc->motion_mode_prob[mbmi->sb_type], ACCT_STR);
+    if (counts) ++counts->motion_mode[mbmi->sb_type][motion_mode];
+    return (MOTION_MODE)(SIMPLE_TRANSLATION + motion_mode);
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  }
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+}
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_EXT_INTER
+static PREDICTION_MODE read_inter_compound_mode(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                                aom_reader *r, int16_t ctx) {
+  const int mode =
+      aom_read_tree(r, av1_inter_compound_mode_tree,
+                    cm->fc->inter_compound_mode_probs[ctx], ACCT_STR);
+  FRAME_COUNTS *counts = xd->counts;
+
+  if (counts) ++counts->inter_compound_mode[ctx][mode];
+
+  assert(is_inter_compound_mode(NEAREST_NEARESTMV + mode));
+  return NEAREST_NEARESTMV + mode;
+}
+#endif  // CONFIG_EXT_INTER
+
+static int read_segment_id(aom_reader *r, struct segmentation_probs *segp) {
+#if CONFIG_EC_MULTISYMBOL
+  return aom_read_symbol(r, segp->tree_cdf, MAX_SEGMENTS, ACCT_STR);
+#else
+  return aom_read_tree(r, av1_segment_tree, segp->tree_probs, ACCT_STR);
+#endif
+}
+
+#if CONFIG_VAR_TX
+static void read_tx_size_vartx(AV1_COMMON *cm, MACROBLOCKD *xd,
+                               MB_MODE_INFO *mbmi, FRAME_COUNTS *counts,
+                               TX_SIZE tx_size, int depth, int blk_row,
+                               int blk_col, aom_reader *r) {
+  int is_split = 0;
+  const int tx_row = blk_row >> 1;
+  const int tx_col = blk_col >> 1;
+  const int max_blocks_high = max_block_high(xd, mbmi->sb_type, 0);
+  const int max_blocks_wide = max_block_wide(xd, mbmi->sb_type, 0);
+  int ctx = txfm_partition_context(xd->above_txfm_context + tx_col,
+                                   xd->left_txfm_context + tx_row,
+                                   mbmi->sb_type, tx_size);
+  TX_SIZE(*const inter_tx_size)
+  [MAX_MIB_SIZE] =
+      (TX_SIZE(*)[MAX_MIB_SIZE]) & mbmi->inter_tx_size[tx_row][tx_col];
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  if (depth == MAX_VARTX_DEPTH) {
+    int idx, idy;
+    inter_tx_size[0][0] = tx_size;
+    for (idy = 0; idy < tx_size_high_unit[tx_size] / 2; ++idy)
+      for (idx = 0; idx < tx_size_wide_unit[tx_size] / 2; ++idx)
+        inter_tx_size[idy][idx] = tx_size;
+    mbmi->tx_size = tx_size;
+    mbmi->min_tx_size = AOMMIN(mbmi->min_tx_size, get_min_tx_size(tx_size));
+    if (counts) ++counts->txfm_partition[ctx][0];
+    txfm_partition_update(xd->above_txfm_context + tx_col,
+                          xd->left_txfm_context + tx_row, tx_size, tx_size);
+    return;
+  }
+
+  is_split = aom_read(r, cm->fc->txfm_partition_prob[ctx], ACCT_STR);
+
+  if (is_split) {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+
+    if (counts) ++counts->txfm_partition[ctx][1];
+
+    if (tx_size == TX_8X8) {
+      int idx, idy;
+      inter_tx_size[0][0] = sub_txs;
+      for (idy = 0; idy < tx_size_high_unit[tx_size] / 2; ++idy)
+        for (idx = 0; idx < tx_size_wide_unit[tx_size] / 2; ++idx)
+          inter_tx_size[idy][idx] = inter_tx_size[0][0];
+      mbmi->tx_size = sub_txs;
+      mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+      txfm_partition_update(xd->above_txfm_context + tx_col,
+                            xd->left_txfm_context + tx_row, sub_txs, tx_size);
+      return;
+    }
+
+    assert(bsl > 0);
+    for (i = 0; i < 4; ++i) {
+      int offsetr = blk_row + (i >> 1) * bsl;
+      int offsetc = blk_col + (i & 0x01) * bsl;
+      read_tx_size_vartx(cm, xd, mbmi, counts, sub_txs, depth + 1, offsetr,
+                         offsetc, r);
+    }
+  } else {
+    int idx, idy;
+    inter_tx_size[0][0] = tx_size;
+    for (idy = 0; idy < tx_size_high_unit[tx_size] / 2; ++idy)
+      for (idx = 0; idx < tx_size_wide_unit[tx_size] / 2; ++idx)
+        inter_tx_size[idy][idx] = tx_size;
+    mbmi->tx_size = tx_size;
+    mbmi->min_tx_size = AOMMIN(mbmi->min_tx_size, get_min_tx_size(tx_size));
+    if (counts) ++counts->txfm_partition[ctx][0];
+    txfm_partition_update(xd->above_txfm_context + tx_col,
+                          xd->left_txfm_context + tx_row, tx_size, tx_size);
+  }
+}
+#endif
+
+static TX_SIZE read_selected_tx_size(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                     int tx_size_cat, aom_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+  const int ctx = get_tx_size_context(xd);
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  const int depth =
+#if CONFIG_EC_MULTISYMBOL
+      aom_read_symbol(r, ec_ctx->tx_size_cdf[tx_size_cat][ctx], tx_size_cat + 2,
+                      ACCT_STR);
+#else
+      aom_read_tree(r, av1_tx_size_tree[tx_size_cat],
+                    ec_ctx->tx_size_probs[tx_size_cat][ctx], ACCT_STR);
+#endif
+  const TX_SIZE tx_size = depth_to_tx_size(depth);
+#if CONFIG_RECT_TX
+  assert(!is_rect_tx(tx_size));
+#endif  // CONFIG_RECT_TX
+  if (counts) ++counts->tx_size[tx_size_cat][ctx][depth];
+  return tx_size;
+}
+
+static TX_SIZE read_tx_size(AV1_COMMON *cm, MACROBLOCKD *xd, int is_inter,
+                            int allow_select_inter, aom_reader *r) {
+  const TX_MODE tx_mode = cm->tx_mode;
+  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  if (xd->lossless[xd->mi[0]->mbmi.segment_id]) return TX_4X4;
+#if CONFIG_CB4X4 && (CONFIG_VAR_TX || CONFIG_EXT_TX) && CONFIG_RECT_TX
+  if (bsize > BLOCK_4X4) {
+#else
+  if (bsize >= BLOCK_8X8) {
+#endif  // CONFIG_CB4X4 && CONFIG_VAR_TX
+    if ((!is_inter || allow_select_inter) && tx_mode == TX_MODE_SELECT) {
+      const int32_t tx_size_cat = is_inter ? inter_tx_size_cat_lookup[bsize]
+                                           : intra_tx_size_cat_lookup[bsize];
+      const TX_SIZE coded_tx_size =
+          read_selected_tx_size(cm, xd, tx_size_cat, r);
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+      if (coded_tx_size > max_txsize_lookup[bsize]) {
+        assert(coded_tx_size == max_txsize_lookup[bsize] + 1);
+        return max_txsize_rect_lookup[bsize];
+      }
+#else
+      assert(coded_tx_size <= max_txsize_lookup[bsize]);
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+      return coded_tx_size;
+    } else {
+      return tx_size_from_tx_mode(bsize, tx_mode, is_inter);
+    }
+  } else {
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+    assert(IMPLIES(tx_mode == ONLY_4X4, bsize == BLOCK_4X4));
+    return max_txsize_rect_lookup[bsize];
+#else
+    return TX_4X4;
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+  }
+}
+
+static int dec_get_segment_id(const AV1_COMMON *cm, const uint8_t *segment_ids,
+                              int mi_offset, int x_mis, int y_mis) {
+  int x, y, segment_id = INT_MAX;
+
+  for (y = 0; y < y_mis; y++)
+    for (x = 0; x < x_mis; x++)
+      segment_id =
+          AOMMIN(segment_id, segment_ids[mi_offset + y * cm->mi_cols + x]);
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+  return segment_id;
+}
+
+static void set_segment_id(AV1_COMMON *cm, int mi_offset, int x_mis, int y_mis,
+                           int segment_id) {
+  int x, y;
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+
+  for (y = 0; y < y_mis; y++)
+    for (x = 0; x < x_mis; x++)
+      cm->current_frame_seg_map[mi_offset + y * cm->mi_cols + x] = segment_id;
+}
+
+static int read_intra_segment_id(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                 int mi_offset, int x_mis, int y_mis,
+                                 aom_reader *r) {
+  struct segmentation *const seg = &cm->seg;
+  FRAME_COUNTS *counts = xd->counts;
+  struct segmentation_probs *const segp = &cm->fc->seg;
+  int segment_id;
+
+  if (!seg->enabled) return 0;  // Default for disabled segmentation
+
+  assert(seg->update_map && !seg->temporal_update);
+
+  segment_id = read_segment_id(r, segp);
+  if (counts) ++counts->seg.tree_total[segment_id];
+  set_segment_id(cm, mi_offset, x_mis, y_mis, segment_id);
+  return segment_id;
+}
+
+static void copy_segment_id(const AV1_COMMON *cm,
+                            const uint8_t *last_segment_ids,
+                            uint8_t *current_segment_ids, int mi_offset,
+                            int x_mis, int y_mis) {
+  int x, y;
+
+  for (y = 0; y < y_mis; y++)
+    for (x = 0; x < x_mis; 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_inter_segment_id(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                 int mi_row, int mi_col, aom_reader *r) {
+  struct segmentation *const seg = &cm->seg;
+  FRAME_COUNTS *counts = xd->counts;
+  struct segmentation_probs *const segp = &cm->fc->seg;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int predicted_segment_id, segment_id;
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = mi_size_wide[mbmi->sb_type];
+  const int bh = mi_size_high[mbmi->sb_type];
+
+  // TODO(slavarnway): move x_mis, y_mis into xd ?????
+  const int x_mis = AOMMIN(cm->mi_cols - mi_col, bw);
+  const int y_mis = AOMMIN(cm->mi_rows - mi_row, bh);
+
+  if (!seg->enabled) return 0;  // Default for disabled segmentation
+
+  predicted_segment_id = cm->last_frame_seg_map
+                             ? dec_get_segment_id(cm, cm->last_frame_seg_map,
+                                                  mi_offset, x_mis, y_mis)
+                             : 0;
+
+  if (!seg->update_map) {
+    copy_segment_id(cm, cm->last_frame_seg_map, cm->current_frame_seg_map,
+                    mi_offset, x_mis, y_mis);
+    return predicted_segment_id;
+  }
+
+  if (seg->temporal_update) {
+    const int ctx = av1_get_pred_context_seg_id(xd);
+    const aom_prob pred_prob = segp->pred_probs[ctx];
+    mbmi->seg_id_predicted = aom_read(r, pred_prob, ACCT_STR);
+    if (counts) ++counts->seg.pred[ctx][mbmi->seg_id_predicted];
+    if (mbmi->seg_id_predicted) {
+      segment_id = predicted_segment_id;
+    } else {
+      segment_id = read_segment_id(r, segp);
+      if (counts) ++counts->seg.tree_mispred[segment_id];
+    }
+  } else {
+    segment_id = read_segment_id(r, segp);
+    if (counts) ++counts->seg.tree_total[segment_id];
+  }
+  set_segment_id(cm, mi_offset, x_mis, y_mis, segment_id);
+  return segment_id;
+}
+
+static int read_skip(AV1_COMMON *cm, const MACROBLOCKD *xd, int segment_id,
+                     aom_reader *r) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
+    return 1;
+  } else {
+    const int ctx = av1_get_skip_context(xd);
+    const int skip = aom_read(r, cm->fc->skip_probs[ctx], ACCT_STR);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts) ++counts->skip[ctx][skip];
+    return skip;
+  }
+}
+
+#if CONFIG_PALETTE
+static void read_palette_mode_info(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                   aom_reader *r) {
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  int i, n;
+  PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+
+  if (mbmi->mode == DC_PRED) {
+    int palette_y_mode_ctx = 0;
+    if (above_mi)
+      palette_y_mode_ctx +=
+          (above_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+    if (left_mi)
+      palette_y_mode_ctx +=
+          (left_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+    if (aom_read(r, av1_default_palette_y_mode_prob[bsize - BLOCK_8X8]
+                                                   [palette_y_mode_ctx],
+                 ACCT_STR)) {
+      pmi->palette_size[0] =
+          aom_read_tree(r, av1_palette_size_tree,
+                        av1_default_palette_y_size_prob[bsize - BLOCK_8X8],
+                        ACCT_STR) +
+          2;
+      n = pmi->palette_size[0];
+#if CONFIG_PALETTE_DELTA_ENCODING
+      const int min_bits = cm->bit_depth - 3;
+      int bits = min_bits + aom_read_literal(r, 2, ACCT_STR);
+      pmi->palette_colors[0] = aom_read_literal(r, cm->bit_depth, ACCT_STR);
+      for (i = 1; i < n; ++i) {
+        pmi->palette_colors[i] = pmi->palette_colors[i - 1] +
+                                 aom_read_literal(r, bits, ACCT_STR) + 1;
+        bits = AOMMIN(
+            bits, av1_ceil_log2((1 << cm->bit_depth) - pmi->palette_colors[i]));
+      }
+#else
+      for (i = 0; i < n; ++i)
+        pmi->palette_colors[i] = aom_read_literal(r, cm->bit_depth, ACCT_STR);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+      xd->plane[0].color_index_map[0] = read_uniform(r, n);
+      assert(xd->plane[0].color_index_map[0] < n);
+    }
+  }
+
+  if (mbmi->uv_mode == DC_PRED) {
+    const int palette_uv_mode_ctx = (pmi->palette_size[0] > 0);
+    if (aom_read(r, av1_default_palette_uv_mode_prob[palette_uv_mode_ctx],
+                 ACCT_STR)) {
+      pmi->palette_size[1] =
+          aom_read_tree(r, av1_palette_size_tree,
+                        av1_default_palette_uv_size_prob[bsize - BLOCK_8X8],
+                        ACCT_STR) +
+          2;
+      n = pmi->palette_size[1];
+#if CONFIG_PALETTE_DELTA_ENCODING
+      // U channel colors.
+      const int min_bits_u = cm->bit_depth - 3;
+      int bits = min_bits_u + aom_read_literal(r, 2, ACCT_STR);
+      pmi->palette_colors[PALETTE_MAX_SIZE] =
+          aom_read_literal(r, cm->bit_depth, ACCT_STR);
+      for (i = 1; i < n; ++i) {
+        pmi->palette_colors[PALETTE_MAX_SIZE + i] =
+            pmi->palette_colors[PALETTE_MAX_SIZE + i - 1] +
+            aom_read_literal(r, bits, ACCT_STR);
+        bits = AOMMIN(bits,
+                      av1_ceil_log2(1 + (1 << cm->bit_depth) -
+                                    pmi->palette_colors[PALETTE_MAX_SIZE + i]));
+      }
+      // V channel colors.
+      if (aom_read_bit(r, ACCT_STR)) {  // Delta encoding.
+        const int min_bits_v = cm->bit_depth - 4;
+        const int max_val = 1 << cm->bit_depth;
+        bits = min_bits_v + aom_read_literal(r, 2, ACCT_STR);
+        pmi->palette_colors[2 * PALETTE_MAX_SIZE] =
+            aom_read_literal(r, cm->bit_depth, ACCT_STR);
+        for (i = 1; i < n; ++i) {
+          int delta = aom_read_literal(r, bits, ACCT_STR);
+          if (delta && aom_read_bit(r, ACCT_STR)) delta = -delta;
+          int val =
+              (int)pmi->palette_colors[2 * PALETTE_MAX_SIZE + i - 1] + delta;
+          if (val < 0) val += max_val;
+          if (val >= max_val) val -= max_val;
+          pmi->palette_colors[2 * PALETTE_MAX_SIZE + i] = val;
+        }
+      } else {
+        for (i = 0; i < n; ++i) {
+          pmi->palette_colors[2 * PALETTE_MAX_SIZE + i] =
+              aom_read_literal(r, cm->bit_depth, ACCT_STR);
+        }
+      }
+#else
+      for (i = 0; i < n; ++i) {
+        pmi->palette_colors[PALETTE_MAX_SIZE + i] =
+            aom_read_literal(r, cm->bit_depth, ACCT_STR);
+        pmi->palette_colors[2 * PALETTE_MAX_SIZE + i] =
+            aom_read_literal(r, cm->bit_depth, ACCT_STR);
+      }
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+      xd->plane[1].color_index_map[0] = read_uniform(r, n);
+      assert(xd->plane[1].color_index_map[0] < n);
+    }
+  }
+}
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+static void read_filter_intra_mode_info(AV1_COMMON *const cm,
+                                        MACROBLOCKD *const xd, aom_reader *r) {
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  FRAME_COUNTS *counts = xd->counts;
+  FILTER_INTRA_MODE_INFO *filter_intra_mode_info =
+      &mbmi->filter_intra_mode_info;
+
+  if (mbmi->mode == DC_PRED
+#if CONFIG_PALETTE
+      && mbmi->palette_mode_info.palette_size[0] == 0
+#endif  // CONFIG_PALETTE
+      ) {
+    filter_intra_mode_info->use_filter_intra_mode[0] =
+        aom_read(r, cm->fc->filter_intra_probs[0], ACCT_STR);
+    if (filter_intra_mode_info->use_filter_intra_mode[0]) {
+      filter_intra_mode_info->filter_intra_mode[0] =
+          read_uniform(r, FILTER_INTRA_MODES);
+    }
+    if (counts) {
+      ++counts
+            ->filter_intra[0][filter_intra_mode_info->use_filter_intra_mode[0]];
+    }
+  }
+  if (mbmi->uv_mode == DC_PRED
+#if CONFIG_PALETTE
+      && mbmi->palette_mode_info.palette_size[1] == 0
+#endif  // CONFIG_PALETTE
+      ) {
+    filter_intra_mode_info->use_filter_intra_mode[1] =
+        aom_read(r, cm->fc->filter_intra_probs[1], ACCT_STR);
+    if (filter_intra_mode_info->use_filter_intra_mode[1]) {
+      filter_intra_mode_info->filter_intra_mode[1] =
+          read_uniform(r, FILTER_INTRA_MODES);
+    }
+    if (counts) {
+      ++counts
+            ->filter_intra[1][filter_intra_mode_info->use_filter_intra_mode[1]];
+    }
+  }
+}
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_EXT_INTRA
+static void read_intra_angle_info(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                  aom_reader *r) {
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+#if CONFIG_INTRA_INTERP
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *const ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *const ec_ctx = cm->fc;
+#endif  // CONFIG_EC_ADAPT
+  const int ctx = av1_get_pred_context_intra_interp(xd);
+  int p_angle;
+#endif  // CONFIG_INTRA_INTERP
+
+  (void)cm;
+  if (bsize < BLOCK_8X8) return;
+
+  if (av1_is_directional_mode(mbmi->mode, bsize)) {
+    mbmi->angle_delta[0] =
+        read_uniform(r, 2 * MAX_ANGLE_DELTA + 1) - MAX_ANGLE_DELTA;
+#if CONFIG_INTRA_INTERP
+    p_angle = mode_to_angle_map[mbmi->mode] + mbmi->angle_delta[0] * ANGLE_STEP;
+    if (av1_is_intra_filter_switchable(p_angle)) {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_MULTISYMBOL
+      mbmi->intra_filter = aom_read_symbol(r, ec_ctx->intra_filter_cdf[ctx],
+                                           INTRA_FILTERS, ACCT_STR);
+#else
+      mbmi->intra_filter = aom_read_tree(
+          r, av1_intra_filter_tree, ec_ctx->intra_filter_probs[ctx], ACCT_STR);
+#endif  // CONFIG_EC_MULTISYMBOL
+      if (counts) ++counts->intra_filter[ctx][mbmi->intra_filter];
+    } else {
+      mbmi->intra_filter = INTRA_FILTER_LINEAR;
+    }
+#endif  // CONFIG_INTRA_INTERP
+  }
+
+  if (av1_is_directional_mode(mbmi->uv_mode, bsize)) {
+    mbmi->angle_delta[1] =
+        read_uniform(r, 2 * MAX_ANGLE_DELTA + 1) - MAX_ANGLE_DELTA;
+  }
+}
+#endif  // CONFIG_EXT_INTRA
+
+void av1_read_tx_type(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                      int supertx_enabled,
+#endif
+#if CONFIG_TXK_SEL
+                      int block, int plane,
+#endif
+                      aom_reader *r) {
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const int inter_block = is_inter_block(mbmi);
+#if CONFIG_VAR_TX
+  const TX_SIZE tx_size = inter_block ? mbmi->min_tx_size : mbmi->tx_size;
+#else
+  const TX_SIZE tx_size = mbmi->tx_size;
+#endif
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+#if !CONFIG_TXK_SEL
+  TX_TYPE *tx_type = &mbmi->tx_type;
+#else
+  // only y plane's tx_type is transmitted
+  if (plane > 0) return;
+  TX_TYPE *tx_type = &mbmi->txk_type[block];
+#endif
+
+  if (!FIXED_TX_TYPE) {
+#if CONFIG_EXT_TX
+    const TX_SIZE square_tx_size = txsize_sqr_map[tx_size];
+    if (get_ext_tx_types(tx_size, mbmi->sb_type, inter_block,
+                         cm->reduced_tx_set_used) > 1 &&
+        ((!cm->seg.enabled && cm->base_qindex > 0) ||
+         (cm->seg.enabled && xd->qindex[mbmi->segment_id] > 0)) &&
+        !mbmi->skip &&
+#if CONFIG_SUPERTX
+        !supertx_enabled &&
+#endif  // CONFIG_SUPERTX
+        !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+      const int eset = get_ext_tx_set(tx_size, mbmi->sb_type, inter_block,
+                                      cm->reduced_tx_set_used);
+      FRAME_COUNTS *counts = xd->counts;
+
+      if (inter_block) {
+        if (eset > 0) {
+#if CONFIG_EC_MULTISYMBOL
+          *tx_type = av1_ext_tx_inter_inv[eset][aom_read_symbol(
+              r, ec_ctx->inter_ext_tx_cdf[eset][square_tx_size],
+              ext_tx_cnt_inter[eset], ACCT_STR)];
+#else
+          *tx_type = aom_read_tree(
+              r, av1_ext_tx_inter_tree[eset],
+              ec_ctx->inter_ext_tx_prob[eset][square_tx_size], ACCT_STR);
+#endif
+          if (counts) ++counts->inter_ext_tx[eset][square_tx_size][*tx_type];
+        }
+      } else if (ALLOW_INTRA_EXT_TX) {
+        if (eset > 0) {
+#if CONFIG_EC_MULTISYMBOL
+          *tx_type = av1_ext_tx_intra_inv[eset][aom_read_symbol(
+              r, ec_ctx->intra_ext_tx_cdf[eset][square_tx_size][mbmi->mode],
+              ext_tx_cnt_intra[eset], ACCT_STR)];
+#else
+          *tx_type = aom_read_tree(
+              r, av1_ext_tx_intra_tree[eset],
+              ec_ctx->intra_ext_tx_prob[eset][square_tx_size][mbmi->mode],
+              ACCT_STR);
+#endif
+          if (counts)
+            ++counts->intra_ext_tx[eset][square_tx_size][mbmi->mode][*tx_type];
+        }
+      }
+    } else {
+      *tx_type = DCT_DCT;
+    }
+#else
+
+    if (tx_size < TX_32X32 &&
+        ((!cm->seg.enabled && cm->base_qindex > 0) ||
+         (cm->seg.enabled && xd->qindex[mbmi->segment_id] > 0)) &&
+        !mbmi->skip &&
+#if CONFIG_SUPERTX
+        !supertx_enabled &&
+#endif  // CONFIG_SUPERTX
+        !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+      FRAME_COUNTS *counts = xd->counts;
+
+      if (inter_block) {
+#if CONFIG_EC_MULTISYMBOL
+        *tx_type = av1_ext_tx_inv[aom_read_symbol(
+            r, ec_ctx->inter_ext_tx_cdf[tx_size], TX_TYPES, ACCT_STR)];
+#else
+        *tx_type = aom_read_tree(r, av1_ext_tx_tree,
+                                 ec_ctx->inter_ext_tx_prob[tx_size], ACCT_STR);
+#endif
+        if (counts) ++counts->inter_ext_tx[tx_size][*tx_type];
+      } else {
+        const TX_TYPE tx_type_nom = intra_mode_to_tx_type_context[mbmi->mode];
+#if CONFIG_EC_MULTISYMBOL
+        *tx_type = av1_ext_tx_inv[aom_read_symbol(
+            r, ec_ctx->intra_ext_tx_cdf[tx_size][tx_type_nom], TX_TYPES,
+            ACCT_STR)];
+#else
+        *tx_type = aom_read_tree(
+            r, av1_ext_tx_tree, ec_ctx->intra_ext_tx_prob[tx_size][tx_type_nom],
+            ACCT_STR);
+#endif
+        if (counts) ++counts->intra_ext_tx[tx_size][tx_type_nom][*tx_type];
+      }
+    } else {
+      *tx_type = DCT_DCT;
+    }
+#endif  // CONFIG_EXT_TX
+  }
+}
+
+#if CONFIG_INTRABC
+static INLINE void read_mv(aom_reader *r, MV *mv, const MV *ref,
+                           nmv_context *ctx, nmv_context_counts *counts,
+                           int allow_hp);
+
+static INLINE int is_mv_valid(const MV *mv);
+
+static INLINE int assign_dv(AV1_COMMON *cm, MACROBLOCKD *xd, int_mv *mv,
+                            const int_mv *ref_mv, int mi_row, int mi_col,
+                            BLOCK_SIZE bsize, aom_reader *r) {
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+  FRAME_COUNTS *counts = xd->counts;
+  nmv_context_counts *const dv_counts = counts ? &counts->dv : NULL;
+  read_mv(r, &mv->as_mv, &ref_mv->as_mv, &ec_ctx->ndvc, dv_counts, 0);
+  int valid = is_mv_valid(&mv->as_mv) &&
+              is_dv_valid(mv->as_mv, &xd->tile, mi_row, mi_col, bsize);
+  // TODO(aconverse@google.com): additional validation
+  return valid;
+}
+#endif  // CONFIG_INTRABC
+
+static void read_intra_frame_mode_info(AV1_COMMON *const cm,
+                                       MACROBLOCKD *const xd, int mi_row,
+                                       int mi_col, aom_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;
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+
+  // TODO(slavarnway): move x_mis, y_mis into xd ?????
+  const int x_mis = AOMMIN(cm->mi_cols - mi_col, bw);
+  const int y_mis = AOMMIN(cm->mi_rows - mi_row, bh);
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  mbmi->segment_id = read_intra_segment_id(cm, xd, mi_offset, x_mis, y_mis, r);
+  mbmi->skip = read_skip(cm, xd, mbmi->segment_id, r);
+
+#if CONFIG_DELTA_Q
+  if (cm->delta_q_present_flag) {
+    xd->current_qindex =
+        xd->prev_qindex +
+        read_delta_qindex(cm, xd, r, mbmi, mi_col, mi_row) * cm->delta_q_res;
+    /* Normative: Clamp to [1,MAXQ] to not interfere with lossless mode */
+    xd->current_qindex = clamp(xd->current_qindex, 1, MAXQ);
+    xd->prev_qindex = xd->current_qindex;
+#if CONFIG_EXT_DELTA_Q
+    if (cm->delta_lf_present_flag) {
+      mbmi->current_delta_lf_from_base = xd->current_delta_lf_from_base =
+          xd->prev_delta_lf_from_base +
+          read_delta_lflevel(cm, xd, r, mbmi, mi_col, mi_row) *
+              cm->delta_lf_res;
+      xd->prev_delta_lf_from_base = xd->current_delta_lf_from_base;
+    }
+#endif
+  }
+#endif
+
+  mbmi->tx_size = read_tx_size(cm, xd, 0, 1, r);
+  mbmi->ref_frame[0] = INTRA_FRAME;
+  mbmi->ref_frame[1] = NONE_FRAME;
+
+#if CONFIG_INTRABC
+  if (bsize >= BLOCK_8X8 && cm->allow_screen_content_tools) {
+    mbmi->use_intrabc = aom_read(r, INTRABC_PROB, ACCT_STR);
+    if (mbmi->use_intrabc) {
+      int_mv dv_ref;
+      mbmi->mode = mbmi->uv_mode = DC_PRED;
+#if CONFIG_DUAL_FILTER
+      for (int idx = 0; idx < 4; ++idx) mbmi->interp_filter[idx] = BILINEAR;
+#else
+      mbmi->interp_filter = BILINEAR;
+#endif
+      av1_find_ref_dv(&dv_ref, mi_row, mi_col);
+      xd->corrupted |=
+          !assign_dv(cm, xd, &mbmi->mv[0], &dv_ref, mi_row, mi_col, bsize, r);
+      return;
+    }
+  }
+#endif  // CONFIG_INTRABC
+
+#if CONFIG_CB4X4
+  (void)i;
+  mbmi->mode =
+#if CONFIG_EC_MULTISYMBOL
+      read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 0));
+#else
+      read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 0));
+#endif
+#else
+  switch (bsize) {
+    case BLOCK_4X4:
+      for (i = 0; i < 4; ++i)
+        mi->bmi[i].as_mode =
+#if CONFIG_EC_MULTISYMBOL
+            read_intra_mode(r,
+                            get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, i));
+#else
+            read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, i));
+#endif
+      mbmi->mode = mi->bmi[3].as_mode;
+      break;
+    case BLOCK_4X8:
+      mi->bmi[0].as_mode = mi->bmi[2].as_mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 0));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 0));
+#endif
+      mi->bmi[1].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 1));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 1));
+#endif
+      break;
+    case BLOCK_8X4:
+      mi->bmi[0].as_mode = mi->bmi[1].as_mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 0));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 0));
+#endif
+      mi->bmi[2].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 2));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 2));
+#endif
+      break;
+    default:
+      mbmi->mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 0));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 0));
+#endif
+  }
+#endif
+
+#if CONFIG_CB4X4
+  if (is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x,
+                          xd->plane[1].subsampling_y))
+    mbmi->uv_mode = read_intra_mode_uv(cm, xd, r, mbmi->mode);
+#else
+  mbmi->uv_mode = read_intra_mode_uv(cm, xd, r, mbmi->mode);
+#endif
+
+#if CONFIG_EXT_INTRA
+  read_intra_angle_info(cm, xd, r);
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[0] = 0;
+  mbmi->palette_mode_info.palette_size[1] = 0;
+  if (bsize >= BLOCK_8X8 && cm->allow_screen_content_tools)
+    read_palette_mode_info(cm, xd, r);
+#endif  // CONFIG_PALETTE
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+  if (bsize >= BLOCK_8X8 || CONFIG_CB4X4)
+    read_filter_intra_mode_info(cm, xd, r);
+#endif  // CONFIG_FILTER_INTRA
+
+#if !CONFIG_TXK_SEL
+  av1_read_tx_type(cm, xd,
+#if CONFIG_SUPERTX
+                   0,
+#endif
+                   r);
+#endif  // !CONFIG_TXK_SEL
+}
+
+static int read_mv_component(aom_reader *r, nmv_component *mvcomp, int usehp) {
+  int mag, d, fr, hp;
+  const int sign = aom_read(r, mvcomp->sign, ACCT_STR);
+  const int mv_class =
+#if CONFIG_EC_MULTISYMBOL
+      aom_read_symbol(r, mvcomp->class_cdf, MV_CLASSES, ACCT_STR);
+#else
+      aom_read_tree(r, av1_mv_class_tree, mvcomp->classes, ACCT_STR);
+#endif
+  const int class0 = mv_class == MV_CLASS_0;
+
+  // Integer part
+  if (class0) {
+    d = aom_read(r, mvcomp->class0[0], ACCT_STR);
+    mag = 0;
+  } else {
+    int i;
+    const int n = mv_class + CLASS0_BITS - 1;  // number of bits
+
+    d = 0;
+    for (i = 0; i < n; ++i) d |= aom_read(r, mvcomp->bits[i], ACCT_STR) << i;
+    mag = CLASS0_SIZE << (mv_class + 2);
+  }
+
+// Fractional part
+#if CONFIG_EC_MULTISYMBOL
+  fr = aom_read_symbol(r, class0 ? mvcomp->class0_fp_cdf[d] : mvcomp->fp_cdf,
+                       MV_FP_SIZE, ACCT_STR);
+#else
+  fr = aom_read_tree(r, av1_mv_fp_tree,
+                     class0 ? mvcomp->class0_fp[d] : mvcomp->fp, ACCT_STR);
+#endif
+
+  // High precision part (if hp is not used, the default value of the hp is 1)
+  hp = usehp ? aom_read(r, class0 ? mvcomp->class0_hp : mvcomp->hp, ACCT_STR)
+             : 1;
+
+  // Result
+  mag += ((d << 3) | (fr << 1) | hp) + 1;
+  return sign ? -mag : mag;
+}
+
+static INLINE void read_mv(aom_reader *r, MV *mv, const MV *ref,
+                           nmv_context *ctx, nmv_context_counts *counts,
+                           int allow_hp) {
+  MV_JOINT_TYPE joint_type;
+  MV diff = { 0, 0 };
+  joint_type =
+#if CONFIG_EC_MULTISYMBOL
+      (MV_JOINT_TYPE)aom_read_symbol(r, ctx->joint_cdf, MV_JOINTS, ACCT_STR);
+#else
+      (MV_JOINT_TYPE)aom_read_tree(r, av1_mv_joint_tree, ctx->joints, ACCT_STR);
+#endif
+
+  if (mv_joint_vertical(joint_type))
+    diff.row = read_mv_component(r, &ctx->comps[0], allow_hp);
+
+  if (mv_joint_horizontal(joint_type))
+    diff.col = read_mv_component(r, &ctx->comps[1], allow_hp);
+
+  av1_inc_mv(&diff, counts, allow_hp);
+
+  mv->row = ref->row + diff.row;
+  mv->col = ref->col + diff.col;
+}
+
+static REFERENCE_MODE read_block_reference_mode(AV1_COMMON *cm,
+                                                const MACROBLOCKD *xd,
+                                                aom_reader *r) {
+#if !SUB8X8_COMP_REF
+  if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8) return SINGLE_REFERENCE;
+#endif
+  if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+    const int ctx = av1_get_reference_mode_context(cm, xd);
+    const REFERENCE_MODE mode =
+        (REFERENCE_MODE)aom_read(r, cm->fc->comp_inter_prob[ctx], ACCT_STR);
+    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(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                            aom_reader *r, int segment_id,
+                            MV_REFERENCE_FRAME ref_frame[2]) {
+  FRAME_CONTEXT *const fc = cm->fc;
+  FRAME_COUNTS *counts = xd->counts;
+
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    ref_frame[0] = (MV_REFERENCE_FRAME)get_segdata(&cm->seg, segment_id,
+                                                   SEG_LVL_REF_FRAME);
+    ref_frame[1] = NONE_FRAME;
+  } 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) {
+#if CONFIG_LOWDELAY_COMPOUND  // Normative in decoder (for low delay)
+      const int idx = 1;
+#else
+#if CONFIG_EXT_REFS
+      const int idx = cm->ref_frame_sign_bias[cm->comp_bwd_ref[0]];
+#else
+      const int idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
+#endif  // CONFIG_EXT_REFS
+#endif
+      const int ctx = av1_get_pred_context_comp_ref_p(cm, xd);
+
+      const int bit = aom_read(r, fc->comp_ref_prob[ctx][0], ACCT_STR);
+      if (counts) ++counts->comp_ref[ctx][0][bit];
+
+#if CONFIG_EXT_REFS
+      // Decode forward references.
+      if (!bit) {
+        const int ctx1 = av1_get_pred_context_comp_ref_p1(cm, xd);
+        const int bit1 = aom_read(r, fc->comp_ref_prob[ctx1][1], ACCT_STR);
+        if (counts) ++counts->comp_ref[ctx1][1][bit1];
+        ref_frame[!idx] = cm->comp_fwd_ref[bit1 ? 0 : 1];
+      } else {
+        const int ctx2 = av1_get_pred_context_comp_ref_p2(cm, xd);
+        const int bit2 = aom_read(r, fc->comp_ref_prob[ctx2][2], ACCT_STR);
+        if (counts) ++counts->comp_ref[ctx2][2][bit2];
+        ref_frame[!idx] = cm->comp_fwd_ref[bit2 ? 3 : 2];
+      }
+
+      // Decode backward references.
+      {
+        const int ctx_bwd = av1_get_pred_context_comp_bwdref_p(cm, xd);
+        const int bit_bwd =
+            aom_read(r, fc->comp_bwdref_prob[ctx_bwd][0], ACCT_STR);
+        if (counts) ++counts->comp_bwdref[ctx_bwd][0][bit_bwd];
+        ref_frame[idx] = cm->comp_bwd_ref[bit_bwd];
+      }
+#else
+      ref_frame[!idx] = cm->comp_var_ref[bit];
+      ref_frame[idx] = cm->comp_fixed_ref;
+#endif  // CONFIG_EXT_REFS
+    } else if (mode == SINGLE_REFERENCE) {
+#if CONFIG_EXT_REFS
+      const int ctx0 = av1_get_pred_context_single_ref_p1(xd);
+      const int bit0 = aom_read(r, fc->single_ref_prob[ctx0][0], ACCT_STR);
+      if (counts) ++counts->single_ref[ctx0][0][bit0];
+
+      if (bit0) {
+        const int ctx1 = av1_get_pred_context_single_ref_p2(xd);
+        const int bit1 = aom_read(r, fc->single_ref_prob[ctx1][1], ACCT_STR);
+        if (counts) ++counts->single_ref[ctx1][1][bit1];
+        ref_frame[0] = bit1 ? ALTREF_FRAME : BWDREF_FRAME;
+      } else {
+        const int ctx2 = av1_get_pred_context_single_ref_p3(xd);
+        const int bit2 = aom_read(r, fc->single_ref_prob[ctx2][2], ACCT_STR);
+        if (counts) ++counts->single_ref[ctx2][2][bit2];
+        if (bit2) {
+          const int ctx4 = av1_get_pred_context_single_ref_p5(xd);
+          const int bit4 = aom_read(r, fc->single_ref_prob[ctx4][4], ACCT_STR);
+          if (counts) ++counts->single_ref[ctx4][4][bit4];
+          ref_frame[0] = bit4 ? GOLDEN_FRAME : LAST3_FRAME;
+        } else {
+          const int ctx3 = av1_get_pred_context_single_ref_p4(xd);
+          const int bit3 = aom_read(r, fc->single_ref_prob[ctx3][3], ACCT_STR);
+          if (counts) ++counts->single_ref[ctx3][3][bit3];
+          ref_frame[0] = bit3 ? LAST2_FRAME : LAST_FRAME;
+        }
+      }
+#else
+      const int ctx0 = av1_get_pred_context_single_ref_p1(xd);
+      const int bit0 = aom_read(r, fc->single_ref_prob[ctx0][0], ACCT_STR);
+      if (counts) ++counts->single_ref[ctx0][0][bit0];
+
+      if (bit0) {
+        const int ctx1 = av1_get_pred_context_single_ref_p2(xd);
+        const int bit1 = aom_read(r, fc->single_ref_prob[ctx1][1], ACCT_STR);
+        if (counts) ++counts->single_ref[ctx1][1][bit1];
+        ref_frame[0] = bit1 ? ALTREF_FRAME : GOLDEN_FRAME;
+      } else {
+        ref_frame[0] = LAST_FRAME;
+      }
+#endif  // CONFIG_EXT_REFS
+
+      ref_frame[1] = NONE_FRAME;
+    } else {
+      assert(0 && "Invalid prediction mode.");
+    }
+  }
+}
+
+static INLINE void read_mb_interp_filter(AV1_COMMON *const cm,
+                                         MACROBLOCKD *const xd,
+                                         MB_MODE_INFO *const mbmi,
+                                         aom_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  if (!av1_is_interp_needed(xd)) {
+    set_default_interp_filters(mbmi, cm->interp_filter);
+    return;
+  }
+
+#if CONFIG_DUAL_FILTER
+  if (cm->interp_filter != SWITCHABLE) {
+    int dir;
+
+    for (dir = 0; dir < 4; ++dir) mbmi->interp_filter[dir] = cm->interp_filter;
+  } else {
+    int dir;
+
+    for (dir = 0; dir < 2; ++dir) {
+      const int ctx = av1_get_pred_context_switchable_interp(xd, dir);
+      mbmi->interp_filter[dir] = EIGHTTAP_REGULAR;
+
+      if (has_subpel_mv_component(xd->mi[0], xd, dir) ||
+          (mbmi->ref_frame[1] > INTRA_FRAME &&
+           has_subpel_mv_component(xd->mi[0], xd, dir + 2))) {
+#if CONFIG_EC_MULTISYMBOL
+        mbmi->interp_filter[dir] =
+            (InterpFilter)av1_switchable_interp_inv[aom_read_symbol(
+                r, ec_ctx->switchable_interp_cdf[ctx], SWITCHABLE_FILTERS,
+                ACCT_STR)];
+#else
+        mbmi->interp_filter[dir] = (InterpFilter)aom_read_tree(
+            r, av1_switchable_interp_tree, ec_ctx->switchable_interp_prob[ctx],
+            ACCT_STR);
+#endif
+        if (counts) ++counts->switchable_interp[ctx][mbmi->interp_filter[dir]];
+      }
+    }
+    // The index system works as:
+    // (0, 1) -> (vertical, horizontal) filter types for the first ref frame.
+    // (2, 3) -> (vertical, horizontal) filter types for the second ref frame.
+    mbmi->interp_filter[2] = mbmi->interp_filter[0];
+    mbmi->interp_filter[3] = mbmi->interp_filter[1];
+  }
+#else  // CONFIG_DUAL_FILTER
+  if (cm->interp_filter != SWITCHABLE) {
+    mbmi->interp_filter = cm->interp_filter;
+  } else {
+    const int ctx = av1_get_pred_context_switchable_interp(xd);
+#if CONFIG_EC_MULTISYMBOL
+    mbmi->interp_filter =
+        (InterpFilter)av1_switchable_interp_inv[aom_read_symbol(
+            r, ec_ctx->switchable_interp_cdf[ctx], SWITCHABLE_FILTERS,
+            ACCT_STR)];
+#else
+    mbmi->interp_filter = (InterpFilter)aom_read_tree(
+        r, av1_switchable_interp_tree, ec_ctx->switchable_interp_prob[ctx],
+        ACCT_STR);
+#endif
+    if (counts) ++counts->switchable_interp[ctx][mbmi->interp_filter];
+  }
+#endif  // CONFIG_DUAL_FILTER
+}
+
+static void read_intra_block_mode_info(AV1_COMMON *const cm, const int mi_row,
+                                       const int mi_col, MACROBLOCKD *const xd,
+                                       MODE_INFO *mi, aom_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_FRAME;
+
+#if CONFIG_CB4X4
+  (void)i;
+  mbmi->mode = read_intra_mode_y(cm, xd, r, size_group_lookup[bsize]);
+#else
+  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]);
+  }
+#endif
+
+#if CONFIG_CB4X4
+  if (is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x,
+                          xd->plane[1].subsampling_y))
+    mbmi->uv_mode = read_intra_mode_uv(cm, xd, r, mbmi->mode);
+#else
+  mbmi->uv_mode = read_intra_mode_uv(cm, xd, r, mbmi->mode);
+  (void)mi_row;
+  (void)mi_col;
+#endif
+
+#if CONFIG_EXT_INTRA
+  read_intra_angle_info(cm, xd, r);
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[0] = 0;
+  mbmi->palette_mode_info.palette_size[1] = 0;
+  if (bsize >= BLOCK_8X8 && cm->allow_screen_content_tools)
+    read_palette_mode_info(cm, xd, r);
+#endif  // CONFIG_PALETTE
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+  if (bsize >= BLOCK_8X8 || CONFIG_CB4X4)
+    read_filter_intra_mode_info(cm, xd, r);
+#endif  // CONFIG_FILTER_INTRA
+}
+
+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(AV1_COMMON *cm, MACROBLOCKD *xd,
+                            PREDICTION_MODE mode,
+                            MV_REFERENCE_FRAME ref_frame[2], int block,
+                            int_mv mv[2], int_mv ref_mv[2],
+                            int_mv nearest_mv[2], int_mv near_mv[2], int mi_row,
+                            int mi_col, int is_compound, int allow_hp,
+                            aom_reader *r) {
+  int i;
+  int ret = 1;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+#if CONFIG_REF_MV
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+#if CONFIG_CB4X4
+  int_mv *pred_mv = mbmi->pred_mv;
+  (void)block;
+#else
+  int_mv *pred_mv =
+      (bsize >= BLOCK_8X8) ? mbmi->pred_mv : xd->mi[0]->bmi[block].pred_mv;
+#endif  // CONFIG_CB4X4
+#else
+  (void)block;
+#endif  // CONFIG_REF_MV
+  (void)ref_frame;
+  (void)cm;
+  (void)mi_row;
+  (void)mi_col;
+  (void)bsize;
+
+  switch (mode) {
+    case NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if !CONFIG_REF_MV
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      for (i = 0; i < 1 + is_compound; ++i) {
+#if CONFIG_REF_MV
+        int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+        int nmv_ctx =
+            av1_nmv_ctx(xd->ref_mv_count[rf_type], xd->ref_mv_stack[rf_type], i,
+                        mbmi->ref_mv_idx);
+        nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+        nmv_context_counts *const mv_counts =
+            counts ? &counts->mv[nmv_ctx] : NULL;
+#endif
+        read_mv(r, &mv[i].as_mv, &ref_mv[i].as_mv, nmvc, mv_counts, allow_hp);
+        ret = ret && is_mv_valid(&mv[i].as_mv);
+
+#if CONFIG_REF_MV
+        pred_mv[i].as_int = ref_mv[i].as_int;
+#endif
+      }
+      break;
+    }
+    case NEARESTMV: {
+      mv[0].as_int = nearest_mv[0].as_int;
+      if (is_compound) mv[1].as_int = nearest_mv[1].as_int;
+
+#if CONFIG_REF_MV
+      pred_mv[0].as_int = nearest_mv[0].as_int;
+      if (is_compound) pred_mv[1].as_int = nearest_mv[1].as_int;
+#endif
+      break;
+    }
+    case NEARMV: {
+      mv[0].as_int = near_mv[0].as_int;
+      if (is_compound) mv[1].as_int = near_mv[1].as_int;
+
+#if CONFIG_REF_MV
+      pred_mv[0].as_int = near_mv[0].as_int;
+      if (is_compound) pred_mv[1].as_int = near_mv[1].as_int;
+#endif
+      break;
+    }
+    case ZEROMV: {
+#if CONFIG_GLOBAL_MOTION
+      mv[0].as_int = gm_get_motion_vector(&cm->global_motion[ref_frame[0]],
+                                          cm->allow_high_precision_mv, bsize,
+                                          mi_col, mi_row, block)
+                         .as_int;
+      if (is_compound)
+        mv[1].as_int = gm_get_motion_vector(&cm->global_motion[ref_frame[1]],
+                                            cm->allow_high_precision_mv, bsize,
+                                            mi_col, mi_row, block)
+                           .as_int;
+#else
+      mv[0].as_int = 0;
+      if (is_compound) mv[1].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+
+#if CONFIG_REF_MV
+      pred_mv[0].as_int = mv[0].as_int;
+      if (is_compound) pred_mv[1].as_int = mv[1].as_int;
+#endif
+      break;
+    }
+#if CONFIG_EXT_INTER
+    case NEW_NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if !CONFIG_REF_MV
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      assert(is_compound);
+      for (i = 0; i < 2; ++i) {
+#if CONFIG_REF_MV
+        int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+        int nmv_ctx =
+            av1_nmv_ctx(xd->ref_mv_count[rf_type], xd->ref_mv_stack[rf_type], i,
+                        mbmi->ref_mv_idx);
+        nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+        nmv_context_counts *const mv_counts =
+            counts ? &counts->mv[nmv_ctx] : NULL;
+#endif
+        read_mv(r, &mv[i].as_mv, &ref_mv[i].as_mv, nmvc, mv_counts, allow_hp);
+        ret = ret && is_mv_valid(&mv[i].as_mv);
+      }
+      break;
+    }
+    case NEAREST_NEARESTMV: {
+      assert(is_compound);
+      mv[0].as_int = nearest_mv[0].as_int;
+      mv[1].as_int = nearest_mv[1].as_int;
+      break;
+    }
+    case NEAREST_NEARMV: {
+      assert(is_compound);
+      mv[0].as_int = nearest_mv[0].as_int;
+      mv[1].as_int = near_mv[1].as_int;
+      break;
+    }
+    case NEAR_NEARESTMV: {
+      assert(is_compound);
+      mv[0].as_int = near_mv[0].as_int;
+      mv[1].as_int = nearest_mv[1].as_int;
+      break;
+    }
+    case NEAR_NEARMV: {
+      assert(is_compound);
+      mv[0].as_int = near_mv[0].as_int;
+      mv[1].as_int = near_mv[1].as_int;
+      break;
+    }
+    case NEW_NEARESTMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx = av1_nmv_ctx(xd->ref_mv_count[rf_type],
+                                xd->ref_mv_stack[rf_type], 0, mbmi->ref_mv_idx);
+      nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+      nmv_context_counts *const mv_counts =
+          counts ? &counts->mv[nmv_ctx] : NULL;
+#else
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      read_mv(r, &mv[0].as_mv, &ref_mv[0].as_mv, nmvc, mv_counts, allow_hp);
+      assert(is_compound);
+      ret = ret && is_mv_valid(&mv[0].as_mv);
+      mv[1].as_int = nearest_mv[1].as_int;
+      break;
+    }
+    case NEAREST_NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx = av1_nmv_ctx(xd->ref_mv_count[rf_type],
+                                xd->ref_mv_stack[rf_type], 1, mbmi->ref_mv_idx);
+      nmv_context_counts *const mv_counts =
+          counts ? &counts->mv[nmv_ctx] : NULL;
+      nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+#else
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      mv[0].as_int = nearest_mv[0].as_int;
+      read_mv(r, &mv[1].as_mv, &ref_mv[1].as_mv, nmvc, mv_counts, allow_hp);
+      assert(is_compound);
+      ret = ret && is_mv_valid(&mv[1].as_mv);
+      break;
+    }
+    case NEAR_NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx = av1_nmv_ctx(xd->ref_mv_count[rf_type],
+                                xd->ref_mv_stack[rf_type], 1, mbmi->ref_mv_idx);
+      nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+      nmv_context_counts *const mv_counts =
+          counts ? &counts->mv[nmv_ctx] : NULL;
+#else
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      mv[0].as_int = near_mv[0].as_int;
+      read_mv(r, &mv[1].as_mv, &ref_mv[1].as_mv, nmvc, mv_counts, allow_hp);
+      assert(is_compound);
+
+      ret = ret && is_mv_valid(&mv[1].as_mv);
+      break;
+    }
+    case NEW_NEARMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx = av1_nmv_ctx(xd->ref_mv_count[rf_type],
+                                xd->ref_mv_stack[rf_type], 0, mbmi->ref_mv_idx);
+      nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+      nmv_context_counts *const mv_counts =
+          counts ? &counts->mv[nmv_ctx] : NULL;
+#else
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      read_mv(r, &mv[0].as_mv, &ref_mv[0].as_mv, nmvc, mv_counts, allow_hp);
+      assert(is_compound);
+      ret = ret && is_mv_valid(&mv[0].as_mv);
+      mv[1].as_int = near_mv[1].as_int;
+      break;
+    }
+    case ZERO_ZEROMV: {
+      assert(is_compound);
+#if CONFIG_GLOBAL_MOTION
+      mv[0].as_int = gm_get_motion_vector(&cm->global_motion[ref_frame[0]],
+                                          cm->allow_high_precision_mv, bsize,
+                                          mi_col, mi_row, block)
+                         .as_int;
+      mv[1].as_int = gm_get_motion_vector(&cm->global_motion[ref_frame[1]],
+                                          cm->allow_high_precision_mv, bsize,
+                                          mi_col, mi_row, block)
+                         .as_int;
+#else
+      mv[0].as_int = 0;
+      mv[1].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+      break;
+    }
+#endif  // CONFIG_EXT_INTER
+    default: { return 0; }
+  }
+  return ret;
+}
+
+static int read_is_inter_block(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                               int segment_id, aom_reader *r) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    return get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME) != INTRA_FRAME;
+  } else {
+    const int ctx = av1_get_intra_inter_context(xd);
+    const int is_inter = aom_read(r, cm->fc->intra_inter_prob[ctx], ACCT_STR);
+    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) {
+  AV1Decoder *const pbi = (AV1Decoder *)data;
+  av1_frameworker_wait(pbi->frame_worker_owner, pbi->common.prev_frame,
+                       mi_row << pbi->common.mib_size_log2);
+}
+
+static void read_inter_block_mode_info(AV1Decoder *const pbi,
+                                       MACROBLOCKD *const xd,
+                                       MODE_INFO *const mi,
+#if (CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION || CONFIG_EXT_INTER) && \
+    CONFIG_SUPERTX
+                                       int mi_row, int mi_col, aom_reader *r,
+                                       int supertx_enabled) {
+#else
+                                       int mi_row, int mi_col, aom_reader *r) {
+#endif  // CONFIG_MOTION_VAR && CONFIG_SUPERTX
+  AV1_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;
+  const int unify_bsize = CONFIG_CB4X4;
+  int_mv nearestmv[2], nearmv[2];
+  int_mv ref_mvs[MODE_CTX_REF_FRAMES][MAX_MV_REF_CANDIDATES];
+  int ref, is_compound;
+  int16_t inter_mode_ctx[MODE_CTX_REF_FRAMES];
+#if CONFIG_REF_MV && CONFIG_EXT_INTER
+  int16_t compound_inter_mode_ctx[MODE_CTX_REF_FRAMES];
+#endif  // CONFIG_REF_MV && CONFIG_EXT_INTER
+  int16_t mode_ctx = 0;
+#if CONFIG_WARPED_MOTION
+  int pts[SAMPLES_ARRAY_SIZE], pts_inref[SAMPLES_ARRAY_SIZE];
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[0] = 0;
+  mbmi->palette_mode_info.palette_size[1] = 0;
+#endif  // CONFIG_PALETTE
+
+  memset(ref_mvs, 0, sizeof(ref_mvs));
+
+  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) {
+    MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
+
+    av1_find_mv_refs(cm, xd, mi, frame,
+#if CONFIG_REF_MV
+                     &xd->ref_mv_count[frame], xd->ref_mv_stack[frame],
+#if CONFIG_EXT_INTER
+                     compound_inter_mode_ctx,
+#endif  // CONFIG_EXT_INTER
+#endif
+                     ref_mvs[frame], mi_row, mi_col, fpm_sync, (void *)pbi,
+                     inter_mode_ctx);
+  }
+
+#if CONFIG_REF_MV
+  if (is_compound) {
+    MV_REFERENCE_FRAME ref_frame = av1_ref_frame_type(mbmi->ref_frame);
+    av1_find_mv_refs(cm, xd, mi, ref_frame, &xd->ref_mv_count[ref_frame],
+                     xd->ref_mv_stack[ref_frame],
+#if CONFIG_EXT_INTER
+                     compound_inter_mode_ctx,
+#endif  // CONFIG_EXT_INTER
+                     ref_mvs[ref_frame], mi_row, mi_col, fpm_sync, (void *)pbi,
+                     inter_mode_ctx);
+
+    if (xd->ref_mv_count[ref_frame] < 2) {
+      MV_REFERENCE_FRAME rf[2];
+      int_mv zeromv[2];
+      av1_set_ref_frame(rf, ref_frame);
+#if CONFIG_GLOBAL_MOTION
+      zeromv[0].as_int = gm_get_motion_vector(&cm->global_motion[rf[0]],
+                                              cm->allow_high_precision_mv,
+                                              bsize, mi_col, mi_row, 0)
+                             .as_int;
+      zeromv[1].as_int = (rf[1] != NONE_FRAME)
+                             ? gm_get_motion_vector(&cm->global_motion[rf[1]],
+                                                    cm->allow_high_precision_mv,
+                                                    bsize, mi_col, mi_row, 0)
+                                   .as_int
+                             : 0;
+#else
+      zeromv[0].as_int = zeromv[1].as_int = 0;
+#endif
+      for (ref = 0; ref < 2; ++ref) {
+        if (rf[ref] == NONE_FRAME) continue;
+        lower_mv_precision(&ref_mvs[rf[ref]][0].as_mv, allow_hp);
+        lower_mv_precision(&ref_mvs[rf[ref]][1].as_mv, allow_hp);
+        if (ref_mvs[rf[ref]][0].as_int != zeromv[ref].as_int ||
+            ref_mvs[rf[ref]][1].as_int != zeromv[ref].as_int)
+          inter_mode_ctx[ref_frame] &= ~(1 << ALL_ZERO_FLAG_OFFSET);
+      }
+    }
+  }
+
+#if CONFIG_EXT_INTER
+  if (is_compound)
+    mode_ctx = compound_inter_mode_ctx[mbmi->ref_frame[0]];
+  else
+#endif  // CONFIG_EXT_INTER
+    mode_ctx =
+        av1_mode_context_analyzer(inter_mode_ctx, mbmi->ref_frame, bsize, -1);
+  mbmi->ref_mv_idx = 0;
+#else
+  mode_ctx = inter_mode_ctx[mbmi->ref_frame[0]];
+#endif
+
+  if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+    mbmi->mode = ZEROMV;
+    if (bsize < BLOCK_8X8 && !unify_bsize) {
+      aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
+                         "Invalid usage of segement feature on small blocks");
+      return;
+    }
+  } else {
+    if (bsize >= BLOCK_8X8 || unify_bsize) {
+#if CONFIG_EXT_INTER
+      if (is_compound)
+        mbmi->mode = read_inter_compound_mode(cm, xd, r, mode_ctx);
+      else
+#endif  // CONFIG_EXT_INTER
+        mbmi->mode = read_inter_mode(ec_ctx, xd, r, mode_ctx);
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+      if (mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV ||
+          have_nearmv_in_inter_mode(mbmi->mode))
+#else
+      if (mbmi->mode == NEARMV || mbmi->mode == NEWMV)
+#endif
+        read_drl_idx(cm, xd, mbmi, r);
+#endif
+    }
+  }
+
+#if CONFIG_EXT_INTER
+  if ((bsize < BLOCK_8X8 && unify_bsize) ||
+      (mbmi->mode != ZEROMV && mbmi->mode != ZERO_ZEROMV)) {
+#else
+  if ((bsize < BLOCK_8X8 && !unify_bsize) || mbmi->mode != ZEROMV) {
+#endif  // CONFIG_EXT_INTER
+    for (ref = 0; ref < 1 + is_compound; ++ref) {
+      av1_find_best_ref_mvs(allow_hp, ref_mvs[mbmi->ref_frame[ref]],
+                            &nearestmv[ref], &nearmv[ref]);
+    }
+  }
+
+#if CONFIG_REF_MV
+  if (mbmi->ref_mv_idx > 0) {
+    int_mv cur_mv =
+        xd->ref_mv_stack[mbmi->ref_frame[0]][1 + mbmi->ref_mv_idx].this_mv;
+    nearmv[0] = cur_mv;
+  }
+
+#if CONFIG_EXT_INTER
+  if (is_compound && (bsize >= BLOCK_8X8 || unify_bsize) &&
+      mbmi->mode != ZERO_ZEROMV) {
+#else
+  if (is_compound && (bsize >= BLOCK_8X8 || unify_bsize) &&
+      mbmi->mode != NEWMV && mbmi->mode != ZEROMV) {
+#endif  // CONFIG_EXT_INTER
+    uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+
+#if CONFIG_EXT_INTER
+    if (xd->ref_mv_count[ref_frame_type] > 0) {
+#else
+    if (xd->ref_mv_count[ref_frame_type] == 1 && mbmi->mode == NEARESTMV) {
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_EXT_INTER
+      if (mbmi->mode == NEAREST_NEARESTMV) {
+#endif  // CONFIG_EXT_INTER
+        nearestmv[0] = xd->ref_mv_stack[ref_frame_type][0].this_mv;
+        nearestmv[1] = xd->ref_mv_stack[ref_frame_type][0].comp_mv;
+        lower_mv_precision(&nearestmv[0].as_mv, allow_hp);
+        lower_mv_precision(&nearestmv[1].as_mv, allow_hp);
+#if CONFIG_EXT_INTER
+      } else if (mbmi->mode == NEAREST_NEWMV || mbmi->mode == NEAREST_NEARMV) {
+        nearestmv[0] = xd->ref_mv_stack[ref_frame_type][0].this_mv;
+        lower_mv_precision(&nearestmv[0].as_mv, allow_hp);
+      } else if (mbmi->mode == NEW_NEARESTMV || mbmi->mode == NEAR_NEARESTMV) {
+        nearestmv[1] = xd->ref_mv_stack[ref_frame_type][0].comp_mv;
+        lower_mv_precision(&nearestmv[1].as_mv, allow_hp);
+      }
+#endif  // CONFIG_EXT_INTER
+    }
+
+#if CONFIG_EXT_INTER
+    if (xd->ref_mv_count[ref_frame_type] > 1) {
+      int ref_mv_idx = 1 + mbmi->ref_mv_idx;
+      if (compound_ref0_mode(mbmi->mode) == NEARMV) {
+        nearmv[0] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+        lower_mv_precision(&nearmv[0].as_mv, allow_hp);
+      }
+
+      if (compound_ref1_mode(mbmi->mode) == NEARMV) {
+        nearmv[1] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+        lower_mv_precision(&nearmv[1].as_mv, allow_hp);
+      }
+    }
+#else
+    if (xd->ref_mv_count[ref_frame_type] > 1) {
+      int ref_mv_idx = 1 + mbmi->ref_mv_idx;
+      nearestmv[0] = xd->ref_mv_stack[ref_frame_type][0].this_mv;
+      nearestmv[1] = xd->ref_mv_stack[ref_frame_type][0].comp_mv;
+      nearmv[0] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+      nearmv[1] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+    }
+#endif  // CONFIG_EXT_INTER
+  }
+#endif
+
+#if !CONFIG_DUAL_FILTER && !CONFIG_WARPED_MOTION && !CONFIG_GLOBAL_MOTION
+  read_mb_interp_filter(cm, xd, mbmi, r);
+#endif  // !CONFIG_DUAL_FILTER && !CONFIG_WARPED_MOTION
+
+  if (bsize < BLOCK_8X8 && !unify_bsize) {
+    const int num_4x4_w = 1 << xd->bmode_blocks_wl;
+    const int num_4x4_h = 1 << xd->bmode_blocks_hl;
+    int idx, idy;
+    PREDICTION_MODE b_mode;
+    int_mv nearest_sub8x8[2], near_sub8x8[2];
+#if CONFIG_EXT_INTER
+    int_mv ref_mv[2][2];
+#endif  // CONFIG_EXT_INTER
+    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;
+        int_mv ref_mv_s8[2];
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+        if (!is_compound)
+#endif  // CONFIG_EXT_INTER
+          mode_ctx = av1_mode_context_analyzer(inter_mode_ctx, mbmi->ref_frame,
+                                               bsize, j);
+#endif
+#if CONFIG_EXT_INTER
+        if (is_compound)
+          b_mode = read_inter_compound_mode(cm, xd, r, mode_ctx);
+        else
+#endif  // CONFIG_EXT_INTER
+          b_mode = read_inter_mode(ec_ctx, xd, r, mode_ctx);
+
+#if CONFIG_EXT_INTER
+        if (b_mode != ZEROMV && b_mode != ZERO_ZEROMV) {
+#else
+        if (b_mode != ZEROMV) {
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_REF_MV
+          CANDIDATE_MV ref_mv_stack[2][MAX_REF_MV_STACK_SIZE];
+          uint8_t ref_mv_count[2];
+#endif
+          for (ref = 0; ref < 1 + is_compound; ++ref)
+#if CONFIG_EXT_INTER
+          {
+            int_mv mv_ref_list[MAX_MV_REF_CANDIDATES];
+            av1_update_mv_context(cm, xd, mi, mbmi->ref_frame[ref], mv_ref_list,
+                                  j, mi_row, mi_col, NULL);
+#endif  // CONFIG_EXT_INTER
+            av1_append_sub8x8_mvs_for_idx(cm, xd, j, ref, mi_row, mi_col,
+#if CONFIG_REF_MV
+                                          ref_mv_stack[ref], &ref_mv_count[ref],
+#endif
+#if CONFIG_EXT_INTER
+                                          mv_ref_list,
+#endif  // CONFIG_EXT_INTER
+                                          &nearest_sub8x8[ref],
+                                          &near_sub8x8[ref]);
+#if CONFIG_EXT_INTER
+            if (have_newmv_in_inter_mode(b_mode)) {
+              mv_ref_list[0].as_int = nearest_sub8x8[ref].as_int;
+              mv_ref_list[1].as_int = near_sub8x8[ref].as_int;
+              av1_find_best_ref_mvs(allow_hp, mv_ref_list, &ref_mv[0][ref],
+                                    &ref_mv[1][ref]);
+            }
+          }
+#endif  // CONFIG_EXT_INTER
+        }
+
+        for (ref = 0; ref < 1 + is_compound && b_mode != ZEROMV; ++ref) {
+#if CONFIG_REF_MV
+          ref_mv_s8[ref] = nearest_sub8x8[ref];
+          lower_mv_precision(&ref_mv_s8[ref].as_mv, allow_hp);
+#else
+          ref_mv_s8[ref] = nearestmv[ref];
+#endif
+        }
+#if CONFIG_EXT_INTER
+        (void)ref_mv_s8;
+#endif
+
+        if (!assign_mv(cm, xd, b_mode, mbmi->ref_frame, j, block,
+#if CONFIG_EXT_INTER
+                       ref_mv[0],
+#else   // !CONFIG_EXT_INTER
+                       ref_mv_s8,
+#endif  // CONFIG_EXT_INTER
+                       nearest_sub8x8, near_sub8x8, mi_row, mi_col, is_compound,
+                       allow_hp, r)) {
+          aom_merge_corrupted_flag(&xd->corrupted, 1);
+          break;
+        };
+
+        mi->bmi[j].as_mv[0].as_int = block[0].as_int;
+        mi->bmi[j].as_mode = b_mode;
+        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];
+      }
+    }
+
+#if CONFIG_REF_MV
+    mbmi->pred_mv[0].as_int = mi->bmi[3].pred_mv[0].as_int;
+    mbmi->pred_mv[1].as_int = mi->bmi[3].pred_mv[1].as_int;
+#endif
+    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 {
+    int_mv ref_mv[2];
+    ref_mv[0] = nearestmv[0];
+    ref_mv[1] = nearestmv[1];
+
+#if CONFIG_EXT_INTER
+    if (is_compound) {
+#if CONFIG_REF_MV
+      int ref_mv_idx = mbmi->ref_mv_idx;
+      // Special case: NEAR_NEWMV and NEW_NEARMV modes use
+      // 1 + mbmi->ref_mv_idx (like NEARMV) instead of
+      // mbmi->ref_mv_idx (like NEWMV)
+      if (mbmi->mode == NEAR_NEWMV || mbmi->mode == NEW_NEARMV)
+        ref_mv_idx = 1 + mbmi->ref_mv_idx;
+#endif
+
+      if (compound_ref0_mode(mbmi->mode) == NEWMV) {
+#if CONFIG_REF_MV
+        uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+        if (xd->ref_mv_count[ref_frame_type] > 1) {
+          ref_mv[0] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+          clamp_mv_ref(&ref_mv[0].as_mv, xd->n8_w << MI_SIZE_LOG2,
+                       xd->n8_h << MI_SIZE_LOG2, xd);
+        }
+#endif
+        nearestmv[0] = ref_mv[0];
+      }
+      if (compound_ref1_mode(mbmi->mode) == NEWMV) {
+#if CONFIG_REF_MV
+        uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+        if (xd->ref_mv_count[ref_frame_type] > 1) {
+          ref_mv[1] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+          clamp_mv_ref(&ref_mv[1].as_mv, xd->n8_w << MI_SIZE_LOG2,
+                       xd->n8_h << MI_SIZE_LOG2, xd);
+        }
+#endif
+        nearestmv[1] = ref_mv[1];
+      }
+    } else {
+#endif  // CONFIG_EXT_INTER
+      if (mbmi->mode == NEWMV) {
+        for (ref = 0; ref < 1 + is_compound; ++ref) {
+#if CONFIG_REF_MV
+          uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+          if (xd->ref_mv_count[ref_frame_type] > 1) {
+            ref_mv[ref] =
+                (ref == 0)
+                    ? xd->ref_mv_stack[ref_frame_type][mbmi->ref_mv_idx].this_mv
+                    : xd->ref_mv_stack[ref_frame_type][mbmi->ref_mv_idx]
+                          .comp_mv;
+            clamp_mv_ref(&ref_mv[ref].as_mv, xd->n8_w << MI_SIZE_LOG2,
+                         xd->n8_h << MI_SIZE_LOG2, xd);
+          }
+#endif
+          nearestmv[ref] = ref_mv[ref];
+        }
+      }
+#if CONFIG_EXT_INTER
+    }
+#endif  // CONFIG_EXT_INTER
+
+    int mv_corrupted_flag =
+        !assign_mv(cm, xd, mbmi->mode, mbmi->ref_frame, 0, mbmi->mv, ref_mv,
+                   nearestmv, nearmv, mi_row, mi_col, is_compound, allow_hp, r);
+    aom_merge_corrupted_flag(&xd->corrupted, mv_corrupted_flag);
+  }
+
+#if CONFIG_EXT_INTER
+  mbmi->use_wedge_interintra = 0;
+  if (cm->reference_mode != COMPOUND_REFERENCE &&
+#if CONFIG_SUPERTX
+      !supertx_enabled &&
+#endif
+      is_interintra_allowed(mbmi)) {
+    const int bsize_group = size_group_lookup[bsize];
+    const int interintra =
+        aom_read(r, cm->fc->interintra_prob[bsize_group], ACCT_STR);
+    if (xd->counts) xd->counts->interintra[bsize_group][interintra]++;
+    assert(mbmi->ref_frame[1] == NONE_FRAME);
+    if (interintra) {
+      const INTERINTRA_MODE interintra_mode =
+          read_interintra_mode(cm, xd, r, bsize_group);
+      mbmi->ref_frame[1] = INTRA_FRAME;
+      mbmi->interintra_mode = interintra_mode;
+#if CONFIG_EXT_INTRA
+      mbmi->angle_delta[0] = 0;
+      mbmi->angle_delta[1] = 0;
+#if CONFIG_INTRA_INTERP
+      mbmi->intra_filter = INTRA_FILTER_LINEAR;
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+      mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+      mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+#endif  // CONFIG_FILTER_INTRA
+      if (is_interintra_wedge_used(bsize)) {
+        mbmi->use_wedge_interintra =
+            aom_read(r, cm->fc->wedge_interintra_prob[bsize], ACCT_STR);
+        if (xd->counts)
+          xd->counts->wedge_interintra[bsize][mbmi->use_wedge_interintra]++;
+        if (mbmi->use_wedge_interintra) {
+          mbmi->interintra_wedge_index =
+              aom_read_literal(r, get_wedge_bits_lookup(bsize), ACCT_STR);
+          mbmi->interintra_wedge_sign = 0;
+        }
+      }
+    }
+  }
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  mbmi->motion_mode = SIMPLE_TRANSLATION;
+#if CONFIG_WARPED_MOTION
+  if (mbmi->sb_type >= BLOCK_8X8 && !has_second_ref(mbmi))
+    mbmi->num_proj_ref[0] = findSamples(cm, xd, mi_row, mi_col, pts, pts_inref);
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+  av1_count_overlappable_neighbors(cm, xd, mi_row, mi_col);
+#endif
+
+#if CONFIG_SUPERTX
+  if (!supertx_enabled) {
+#endif  // CONFIG_SUPERTX
+#if CONFIG_EXT_INTER
+    if (mbmi->ref_frame[1] != INTRA_FRAME)
+#endif  // CONFIG_EXT_INTER
+      mbmi->motion_mode = read_motion_mode(cm, xd, mi, r);
+#if CONFIG_WARPED_MOTION
+    if (mbmi->motion_mode == WARPED_CAUSAL) {
+      mbmi->wm_params[0].wmtype = DEFAULT_WMTYPE;
+      if (find_projection(mbmi->num_proj_ref[0], pts, pts_inref, bsize,
+                          mbmi->mv[0].as_mv.row, mbmi->mv[0].as_mv.col,
+                          &mbmi->wm_params[0], mi_row, mi_col)) {
+        assert(0 && "Invalid Warped Model.");
+      }
+    }
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_SUPERTX
+  }
+#endif  // CONFIG_SUPERTX
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_EXT_INTER
+  mbmi->interinter_compound_type = COMPOUND_AVERAGE;
+  if (cm->reference_mode != SINGLE_REFERENCE &&
+      is_inter_compound_mode(mbmi->mode)
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      && mbmi->motion_mode == SIMPLE_TRANSLATION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      ) {
+    if (is_any_masked_compound_used(bsize)) {
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+      mbmi->interinter_compound_type =
+          aom_read_tree(r, av1_compound_type_tree,
+                        cm->fc->compound_type_prob[bsize], ACCT_STR);
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#if CONFIG_WEDGE
+      if (mbmi->interinter_compound_type == COMPOUND_WEDGE) {
+        mbmi->wedge_index =
+            aom_read_literal(r, get_wedge_bits_lookup(bsize), ACCT_STR);
+        mbmi->wedge_sign = aom_read_bit(r, ACCT_STR);
+      }
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+      if (mbmi->interinter_compound_type == COMPOUND_SEG) {
+        mbmi->mask_type = aom_read_literal(r, MAX_SEG_MASK_BITS, ACCT_STR);
+      }
+#endif  // CONFIG_COMPOUND_SEGMENT
+    } else {
+      mbmi->interinter_compound_type = COMPOUND_AVERAGE;
+    }
+    if (xd->counts)
+      xd->counts->compound_interinter[bsize][mbmi->interinter_compound_type]++;
+  }
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_DUAL_FILTER || CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION
+  read_mb_interp_filter(cm, xd, mbmi, r);
+#endif  // CONFIG_DUAL_FILTER || CONFIG_WARPED_MOTION
+}
+
+static void read_inter_frame_mode_info(AV1Decoder *const pbi,
+                                       MACROBLOCKD *const xd,
+#if CONFIG_SUPERTX
+                                       int supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                                       int mi_row, int mi_col, aom_reader *r) {
+  AV1_COMMON *const cm = &pbi->common;
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  int inter_block = 1;
+#if CONFIG_VAR_TX
+  BLOCK_SIZE bsize = mbmi->sb_type;
+#endif  // CONFIG_VAR_TX
+
+  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);
+#if CONFIG_SUPERTX
+  if (!supertx_enabled)
+#endif  // CONFIG_SUPERTX
+    mbmi->skip = read_skip(cm, xd, mbmi->segment_id, r);
+
+#if CONFIG_DELTA_Q
+  if (cm->delta_q_present_flag) {
+    xd->current_qindex =
+        xd->prev_qindex +
+        read_delta_qindex(cm, xd, r, mbmi, mi_col, mi_row) * cm->delta_q_res;
+    /* Normative: Clamp to [1,MAXQ] to not interfere with lossless mode */
+    xd->current_qindex = clamp(xd->current_qindex, 1, MAXQ);
+    xd->prev_qindex = xd->current_qindex;
+#if CONFIG_EXT_DELTA_Q
+    if (cm->delta_lf_present_flag) {
+      mbmi->current_delta_lf_from_base = xd->current_delta_lf_from_base =
+          xd->prev_delta_lf_from_base +
+          read_delta_lflevel(cm, xd, r, mbmi, mi_col, mi_row) *
+              cm->delta_lf_res;
+      xd->prev_delta_lf_from_base = xd->current_delta_lf_from_base;
+    }
+#endif
+  }
+#endif
+
+#if CONFIG_SUPERTX
+  if (!supertx_enabled) {
+#endif  // CONFIG_SUPERTX
+    inter_block = read_is_inter_block(cm, xd, mbmi->segment_id, r);
+
+#if CONFIG_VAR_TX
+    xd->above_txfm_context = cm->above_txfm_context + mi_col;
+    xd->left_txfm_context =
+        xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+
+    if (cm->tx_mode == TX_MODE_SELECT &&
+#if CONFIG_CB4X4
+        bsize > BLOCK_4X4 &&
+#else
+        bsize >= BLOCK_8X8 &&
+#endif
+        !mbmi->skip && inter_block) {
+      const TX_SIZE max_tx_size = max_txsize_rect_lookup[bsize];
+      const int bh = tx_size_high_unit[max_tx_size];
+      const int bw = tx_size_wide_unit[max_tx_size];
+      const int width = block_size_wide[bsize] >> tx_size_wide_log2[0];
+      const int height = block_size_high[bsize] >> tx_size_wide_log2[0];
+      int idx, idy;
+
+      mbmi->min_tx_size = TX_SIZES_ALL;
+      for (idy = 0; idy < height; idy += bh)
+        for (idx = 0; idx < width; idx += bw)
+          read_tx_size_vartx(cm, xd, mbmi, xd->counts, max_tx_size,
+                             height != width, idy, idx, r);
+    } else {
+      mbmi->tx_size = read_tx_size(cm, xd, inter_block, !mbmi->skip, r);
+
+      if (inter_block) {
+        const int width = block_size_wide[bsize] >> tx_size_wide_log2[0];
+        const int height = block_size_high[bsize] >> tx_size_high_log2[0];
+        int idx, idy;
+        for (idy = 0; idy < height; ++idy)
+          for (idx = 0; idx < width; ++idx)
+            mbmi->inter_tx_size[idy >> 1][idx >> 1] = mbmi->tx_size;
+      }
+      mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+      set_txfm_ctxs(mbmi->tx_size, xd->n8_w, xd->n8_h, mbmi->skip, xd);
+    }
+#else
+  mbmi->tx_size = read_tx_size(cm, xd, inter_block, !mbmi->skip, r);
+#endif  // CONFIG_VAR_TX
+#if CONFIG_SUPERTX
+  }
+#if CONFIG_VAR_TX
+  else if (inter_block) {
+    const int width = num_4x4_blocks_wide_lookup[bsize];
+    const int height = num_4x4_blocks_high_lookup[bsize];
+    int idx, idy;
+    xd->mi[0]->mbmi.tx_size = xd->supertx_size;
+    for (idy = 0; idy < height; ++idy)
+      for (idx = 0; idx < width; ++idx)
+        xd->mi[0]->mbmi.inter_tx_size[idy >> 1][idx >> 1] = xd->supertx_size;
+  }
+#endif  // CONFIG_VAR_TX
+#endif  // CONFIG_SUPERTX
+
+  if (inter_block)
+    read_inter_block_mode_info(pbi, xd,
+#if (CONFIG_MOTION_VAR || CONFIG_EXT_INTER || CONFIG_WARPED_MOTION) && \
+    CONFIG_SUPERTX
+
+                               mi, mi_row, mi_col, r, supertx_enabled);
+#else
+                               mi, mi_row, mi_col, r);
+#endif  // CONFIG_MOTION_VAR && CONFIG_SUPERTX
+  else
+    read_intra_block_mode_info(cm, mi_row, mi_col, xd, mi, r);
+
+#if !CONFIG_TXK_SEL
+  av1_read_tx_type(cm, xd,
+#if CONFIG_SUPERTX
+                   supertx_enabled,
+#endif
+                   r);
+#endif  // !CONFIG_TXK_SEL
+}
+
+void av1_read_mode_info(AV1Decoder *const pbi, MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                        int supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                        int mi_row, int mi_col, aom_reader *r, int x_mis,
+                        int y_mis) {
+  AV1_COMMON *const cm = &pbi->common;
+  MODE_INFO *const mi = xd->mi[0];
+  MV_REF *frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
+  int w, h;
+
+#if CONFIG_INTRABC
+  mi->mbmi.use_intrabc = 0;
+#endif  // CONFIG_INTRABC
+
+  if (frame_is_intra_only(cm)) {
+    read_intra_frame_mode_info(cm, xd, mi_row, mi_col, r);
+#if CONFIG_REF_MV
+    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] = NONE_FRAME;
+        mv->ref_frame[1] = NONE_FRAME;
+      }
+    }
+#endif
+  } else {
+    read_inter_frame_mode_info(pbi, xd,
+#if CONFIG_SUPERTX
+                               supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                               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;
+#if CONFIG_REF_MV
+        mv->pred_mv[0].as_int = mi->mbmi.pred_mv[0].as_int;
+        mv->pred_mv[1].as_int = mi->mbmi.pred_mv[1].as_int;
+#endif
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/decoder/decodemv.h b/third_party/aom/av1/decoder/decodemv.h
new file mode 100644
index 000000000..ceaee1d6b
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodemv.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DECODEMV_H_
+#define AV1_DECODER_DECODEMV_H_
+
+#include "aom_dsp/bitreader.h"
+
+#include "av1/decoder/decoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_read_mode_info(AV1Decoder *const pbi, MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                        int supertx_enabled,
+#endif
+
+                        int mi_row, int mi_col, aom_reader *r, int x_mis,
+                        int y_mis);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+void av1_read_tx_type(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                      int supertx_enabled,
+#endif
+#if CONFIG_TXK_SEL
+                      int block, int plane,
+#endif
+                      aom_reader *r);
+
+#endif  // AV1_DECODER_DECODEMV_H_
diff --git a/third_party/aom/av1/decoder/decoder.c b/third_party/aom/av1/decoder/decoder.c
new file mode 100644
index 000000000..1bd91086e
--- /dev/null
+++ b/third_party/aom/av1/decoder/decoder.c
@@ -0,0 +1,583 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <stdio.h>
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "./aom_scale_rtcd.h"
+
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/system_state.h"
+#include "aom_ports/aom_once.h"
+#include "aom_ports/aom_timer.h"
+#include "aom_scale/aom_scale.h"
+#include "aom_util/aom_thread.h"
+
+#include "av1/common/alloccommon.h"
+#include "av1/common/av1_loopfilter.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+
+#include "av1/decoder/decodeframe.h"
+#include "av1/decoder/decoder.h"
+
+#if !CONFIG_PVQ
+#include "av1/decoder/detokenize.h"
+#endif
+
+static void initialize_dec(void) {
+  static volatile int init_done = 0;
+
+  if (!init_done) {
+    av1_rtcd();
+    aom_dsp_rtcd();
+    aom_scale_rtcd();
+    av1_init_intra_predictors();
+#if CONFIG_EXT_INTER
+    av1_init_wedge_masks();
+#endif  // CONFIG_EXT_INTER
+    init_done = 1;
+#if CONFIG_EC_MULTISYMBOL
+    av1_indices_from_tree(av1_intra_mode_ind, av1_intra_mode_inv,
+                          av1_intra_mode_tree);
+    av1_indices_from_tree(av1_switchable_interp_ind, av1_switchable_interp_inv,
+                          av1_switchable_interp_tree);
+#if CONFIG_EXT_TX
+    int s;
+    for (s = 1; s < EXT_TX_SETS_INTRA; ++s)
+      av1_indices_from_tree(av1_ext_tx_intra_ind[s], av1_ext_tx_intra_inv[s],
+                            av1_ext_tx_intra_tree[s]);
+    for (s = 1; s < EXT_TX_SETS_INTER; ++s)
+      av1_indices_from_tree(av1_ext_tx_inter_ind[s], av1_ext_tx_inter_inv[s],
+                            av1_ext_tx_inter_tree[s]);
+#else
+    av1_indices_from_tree(av1_ext_tx_ind, av1_ext_tx_inv, av1_ext_tx_tree);
+#endif
+    av1_indices_from_tree(av1_inter_mode_ind, av1_inter_mode_inv,
+                          av1_inter_mode_tree);
+#endif
+  }
+}
+
+static void av1_dec_setup_mi(AV1_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 av1_dec_alloc_mi(AV1_COMMON *cm, int mi_size) {
+  cm->mip = aom_calloc(mi_size, sizeof(*cm->mip));
+  if (!cm->mip) return 1;
+  cm->mi_alloc_size = mi_size;
+  cm->mi_grid_base = (MODE_INFO **)aom_calloc(mi_size, sizeof(MODE_INFO *));
+  if (!cm->mi_grid_base) return 1;
+  return 0;
+}
+
+static void av1_dec_free_mi(AV1_COMMON *cm) {
+  aom_free(cm->mip);
+  cm->mip = NULL;
+  aom_free(cm->mi_grid_base);
+  cm->mi_grid_base = NULL;
+}
+
+AV1Decoder *av1_decoder_create(BufferPool *const pool) {
+  AV1Decoder *volatile const pbi = aom_memalign(32, sizeof(*pbi));
+  AV1_COMMON *volatile const cm = pbi ? &pbi->common : NULL;
+
+  if (!cm) return NULL;
+
+  av1_zero(*pbi);
+
+  if (setjmp(cm->error.jmp)) {
+    cm->error.setjmp = 0;
+    av1_decoder_remove(pbi);
+    return NULL;
+  }
+
+  cm->error.setjmp = 1;
+
+  CHECK_MEM_ERROR(cm, cm->fc,
+                  (FRAME_CONTEXT *)aom_memalign(32, sizeof(*cm->fc)));
+  CHECK_MEM_ERROR(cm, cm->frame_contexts,
+                  (FRAME_CONTEXT *)aom_memalign(
+                      32, FRAME_CONTEXTS * sizeof(*cm->frame_contexts)));
+  memset(cm->fc, 0, sizeof(*cm->fc));
+  memset(cm->frame_contexts, 0, 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 = AOM_BITS_8;
+  cm->dequant_bit_depth = AOM_BITS_8;
+
+  cm->alloc_mi = av1_dec_alloc_mi;
+  cm->free_mi = av1_dec_free_mi;
+  cm->setup_mi = av1_dec_setup_mi;
+
+  av1_loop_filter_init(cm);
+
+#if CONFIG_AOM_QM
+  aom_qm_init(cm);
+#endif
+#if CONFIG_LOOP_RESTORATION
+  av1_loop_restoration_precal();
+#endif  // CONFIG_LOOP_RESTORATION
+#if CONFIG_ACCOUNTING
+  pbi->acct_enabled = 1;
+  aom_accounting_init(&pbi->accounting);
+#endif
+
+  cm->error.setjmp = 0;
+
+  aom_get_worker_interface()->init(&pbi->lf_worker);
+
+  return pbi;
+}
+
+void av1_decoder_remove(AV1Decoder *pbi) {
+  int i;
+
+  if (!pbi) return;
+
+  aom_get_worker_interface()->end(&pbi->lf_worker);
+  aom_free(pbi->lf_worker.data1);
+  aom_free(pbi->tile_data);
+  for (i = 0; i < pbi->num_tile_workers; ++i) {
+    AVxWorker *const worker = &pbi->tile_workers[i];
+    aom_get_worker_interface()->end(worker);
+  }
+  aom_free(pbi->tile_worker_data);
+  aom_free(pbi->tile_worker_info);
+  aom_free(pbi->tile_workers);
+
+  if (pbi->num_tile_workers > 0) {
+    av1_loop_filter_dealloc(&pbi->lf_row_sync);
+  }
+
+#if CONFIG_ACCOUNTING
+  aom_accounting_clear(&pbi->accounting);
+#endif
+
+  aom_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;
+}
+
+aom_codec_err_t av1_copy_reference_dec(AV1Decoder *pbi,
+                                       AOM_REFFRAME ref_frame_flag,
+                                       YV12_BUFFER_CONFIG *sd) {
+  AV1_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
+   * aomenc --test-decode functionality working, and will be replaced in a
+   * later commit that adds AV1-specific controls for this functionality.
+   */
+  if (ref_frame_flag == AOM_LAST_FLAG) {
+    const YV12_BUFFER_CONFIG *const cfg = get_ref_frame(cm, 0);
+    if (cfg == NULL) {
+      aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                         "No 'last' reference frame");
+      return AOM_CODEC_ERROR;
+    }
+    if (!equal_dimensions(cfg, sd))
+      aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                         "Incorrect buffer dimensions");
+    else
+      aom_yv12_copy_frame(cfg, sd);
+  } else {
+    aom_internal_error(&cm->error, AOM_CODEC_ERROR, "Invalid reference frame");
+  }
+
+  return cm->error.error_code;
+}
+
+aom_codec_err_t av1_set_reference_dec(AV1_COMMON *cm,
+                                      AOM_REFFRAME ref_frame_flag,
+                                      YV12_BUFFER_CONFIG *sd) {
+  int idx;
+  YV12_BUFFER_CONFIG *ref_buf = NULL;
+
+  // 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
+  // aomenc --test-decode functionality working, and will be replaced in a
+  // later commit that adds AV1-specific controls for this functionality.
+
+  // (Yunqing) The set_reference control depends on the following setting in
+  // encoder.
+  //   cpi->lst_fb_idx = 0;
+  // #if CONFIG_EXT_REFS
+  //   cpi->lst2_fb_idx = 1;
+  //   cpi->lst3_fb_idx = 2;
+  //   cpi->gld_fb_idx = 3;
+  //   cpi->bwd_fb_idx = 4;
+  //   cpi->alt_fb_idx = 5;
+  // #else  // CONFIG_EXT_REFS
+  //   cpi->gld_fb_idx = 1;
+  //   cpi->alt_fb_idx = 2;
+  // #endif  // CONFIG_EXT_REFS
+
+  // TODO(zoeliu): To revisit following code and reconsider what assumption we
+  // may take on the reference frame buffer virtual indexes
+  if (ref_frame_flag == AOM_LAST_FLAG) {
+    idx = cm->ref_frame_map[0];
+#if CONFIG_EXT_REFS
+  } else if (ref_frame_flag == AOM_LAST2_FLAG) {
+    idx = cm->ref_frame_map[1];
+  } else if (ref_frame_flag == AOM_LAST3_FLAG) {
+    idx = cm->ref_frame_map[2];
+  } else if (ref_frame_flag == AOM_GOLD_FLAG) {
+    idx = cm->ref_frame_map[3];
+  } else if (ref_frame_flag == AOM_BWD_FLAG) {
+    idx = cm->ref_frame_map[4];
+  } else if (ref_frame_flag == AOM_ALT_FLAG) {
+    idx = cm->ref_frame_map[5];
+#else
+  } else if (ref_frame_flag == AOM_GOLD_FLAG) {
+    idx = cm->ref_frame_map[1];
+  } else if (ref_frame_flag == AOM_ALT_FLAG) {
+    idx = cm->ref_frame_map[2];
+#endif  // CONFIG_EXT_REFS
+  } else {
+    aom_internal_error(&cm->error, AOM_CODEC_ERROR, "Invalid reference frame");
+    return cm->error.error_code;
+  }
+
+  if (idx < 0 || idx >= FRAME_BUFFERS) {
+    aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                       "Invalid reference frame map");
+    return cm->error.error_code;
+  }
+
+  // Get the destination reference buffer.
+  ref_buf = &cm->buffer_pool->frame_bufs[idx].buf;
+
+  if (!equal_dimensions(ref_buf, sd)) {
+    aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                       "Incorrect buffer dimensions");
+  } else {
+    // Overwrite the reference frame buffer.
+    aom_yv12_copy_frame(sd, ref_buf);
+  }
+
+  return cm->error.error_code;
+}
+
+/* If any buffer updating is signaled it should be done here. */
+static void swap_frame_buffers(AV1Decoder *pbi) {
+  int ref_index = 0, mask;
+  AV1_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 holding in the reference map for the decoding
+    // of the next frame.
+    if (mask & 1) 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);
+
+  // TODO(zoeliu): To fix the ref frame buffer update for the scenario of
+  //               cm->frame_parellel_decode == 1
+  if (!cm->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 < INTER_REFS_PER_FRAME; ref_index++) {
+    cm->frame_refs[ref_index].idx = INVALID_IDX;
+    cm->frame_refs[ref_index].buf = NULL;
+  }
+}
+
+int av1_receive_compressed_data(AV1Decoder *pbi, size_t size,
+                                const uint8_t **psource) {
+  AV1_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 = AOM_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;
+
+  // Find a free buffer for the new frame, releasing the reference previously
+  // held.
+
+  // Check if the previous frame was a frame without any references to it.
+  // Release frame buffer if not decoding in frame parallel mode.
+  if (!cm->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 AOM_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 (cm->frame_parallel_decode) {
+    AVxWorker *const worker = pbi->frame_worker_owner;
+    av1_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;
+    av1_frameworker_unlock_stats(worker);
+  } else {
+    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
+  }
+
+  if (setjmp(cm->error.jmp)) {
+    const AVxWorkerInterface *const winterface = aom_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 holding in the reference map for the
+        // decoding of the next frame.
+        if (mask & 1) 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);
+
+    aom_clear_system_state();
+    return -1;
+  }
+
+  cm->error.setjmp = 1;
+  av1_decode_frame(pbi, source, source + size, psource);
+
+  swap_frame_buffers(pbi);
+
+#if CONFIG_EXT_TILE
+  // For now, we only extend the frame borders when the whole frame is decoded.
+  // Later, if needed, extend the border for the decoded tile on the frame
+  // border.
+  if (pbi->dec_tile_row == -1 && pbi->dec_tile_col == -1)
+#endif  // CONFIG_EXT_TILE
+    aom_extend_frame_inner_borders(cm->frame_to_show);
+
+  aom_clear_system_state();
+
+  if (!cm->show_existing_frame) {
+    cm->last_show_frame = cm->show_frame;
+
+#if CONFIG_EXT_REFS
+    // NOTE: It is not supposed to ref to any frame not used as reference
+    if (cm->is_reference_frame)
+#endif  // CONFIG_EXT_REFS
+      cm->prev_frame = cm->cur_frame;
+
+    if (cm->seg.enabled && !cm->frame_parallel_decode)
+      av1_swap_current_and_last_seg_map(cm);
+  }
+
+  // Update progress in frame parallel decode.
+  if (cm->frame_parallel_decode) {
+    // Need to lock the mutex here as another thread may
+    // be accessing this buffer.
+    AVxWorker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    av1_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;
+    av1_frameworker_signal_stats(worker);
+    av1_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 av1_get_raw_frame(AV1Decoder *pbi, YV12_BUFFER_CONFIG *sd) {
+  AV1_COMMON *const cm = &pbi->common;
+  int ret = -1;
+  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;
+  *sd = *cm->frame_to_show;
+  ret = 0;
+  aom_clear_system_state();
+  return ret;
+}
+
+int av1_get_frame_to_show(AV1Decoder *pbi, YV12_BUFFER_CONFIG *frame) {
+  AV1_COMMON *const cm = &pbi->common;
+
+  if (!cm->show_frame || !cm->frame_to_show) return -1;
+
+  *frame = *cm->frame_to_show;
+  return 0;
+}
+
+aom_codec_err_t av1_parse_superframe_index(const uint8_t *data, size_t data_sz,
+                                           uint32_t sizes[8], int *count,
+                                           aom_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;
+  size_t frame_sz_sum = 0;
+
+  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 - 1);
+
+    // 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 AOM_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 AOM_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[28];
+      assert(sizeof(clear_buffer) >= (frames - 1) * mag);
+      if (decrypt_cb) {
+        decrypt_cb(decrypt_state, x, clear_buffer, (frames - 1) * mag);
+        x = clear_buffer;
+      }
+
+      for (i = 0; i < frames - 1; ++i) {
+        uint32_t this_sz = 0;
+
+        for (j = 0; j < mag; ++j) this_sz |= (*x++) << (j * 8);
+        this_sz += 1;
+        sizes[i] = this_sz;
+        frame_sz_sum += this_sz;
+      }
+      sizes[i] = (uint32_t)(data_sz - index_sz - frame_sz_sum);
+      *count = frames;
+    }
+  }
+  return AOM_CODEC_OK;
+}
diff --git a/third_party/aom/av1/decoder/decoder.h b/third_party/aom/av1/decoder/decoder.h
new file mode 100644
index 000000000..4a90b4ad5
--- /dev/null
+++ b/third_party/aom/av1/decoder/decoder.h
@@ -0,0 +1,224 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DECODER_H_
+#define AV1_DECODER_DECODER_H_
+
+#include "./aom_config.h"
+
+#include "aom/aom_codec.h"
+#include "aom_dsp/bitreader.h"
+#include "aom_scale/yv12config.h"
+#include "aom_util/aom_thread.h"
+
+#include "av1/common/thread_common.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/decoder/dthread.h"
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+#if CONFIG_INSPECTION
+#include "av1/decoder/inspection.h"
+#endif
+
+#if CONFIG_PVQ
+#include "aom_dsp/entdec.h"
+#include "av1/decoder/decint.h"
+#include "av1/encoder/encodemb.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// TODO(hkuang): combine this with TileWorkerData.
+typedef struct TileData {
+  AV1_COMMON *cm;
+  aom_reader bit_reader;
+  DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+  /* dqcoeff are shared by all the planes. So planes must be decoded serially */
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[MAX_TX_SQUARE]);
+#if CONFIG_PVQ
+  /* forward transformed predicted image, a reference for PVQ */
+  DECLARE_ALIGNED(16, tran_low_t, pvq_ref_coeff[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+#endif
+#if CONFIG_CFL
+  CFL_CTX cfl;
+#endif
+#if CONFIG_EC_ADAPT
+  DECLARE_ALIGNED(16, FRAME_CONTEXT, tctx);
+#endif
+#if CONFIG_PALETTE
+  DECLARE_ALIGNED(16, uint8_t, color_index_map[2][MAX_SB_SQUARE]);
+#endif  // CONFIG_PALETTE
+} TileData;
+
+typedef struct TileWorkerData {
+  struct AV1Decoder *pbi;
+  aom_reader bit_reader;
+  FRAME_COUNTS counts;
+  DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+  /* dqcoeff are shared by all the planes. So planes must be decoded serially */
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[MAX_TX_SQUARE]);
+#if CONFIG_PVQ
+  /* forward transformed predicted image, a reference for PVQ */
+  DECLARE_ALIGNED(16, tran_low_t, pvq_ref_coeff[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+#endif
+#if CONFIG_CFL
+  CFL_CTX cfl;
+#endif
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT tctx;
+#endif
+#if CONFIG_PALETTE
+  DECLARE_ALIGNED(16, uint8_t, color_index_map[2][MAX_SB_SQUARE]);
+#endif  // CONFIG_PALETTE
+  struct aom_internal_error_info error_info;
+} TileWorkerData;
+
+typedef struct TileBufferDec {
+  const uint8_t *data;
+  size_t size;
+  const uint8_t *raw_data_end;  // The end of the raw tile buffer in the
+                                // bit stream.
+  int col;                      // only used with multi-threaded decoding
+} TileBufferDec;
+
+typedef struct AV1Decoder {
+  DECLARE_ALIGNED(16, MACROBLOCKD, mb);
+
+  DECLARE_ALIGNED(16, AV1_COMMON, common);
+
+  int ready_for_new_data;
+
+  int refresh_frame_flags;
+
+  // TODO(hkuang): Combine this with cur_buf in macroblockd as they are
+  // the same.
+  RefCntBuffer *cur_buf;  //  Current decoding frame buffer.
+
+  AVxWorker *frame_worker_owner;  // frame_worker that owns this pbi.
+  AVxWorker lf_worker;
+  AVxWorker *tile_workers;
+  TileWorkerData *tile_worker_data;
+  TileInfo *tile_worker_info;
+  int num_tile_workers;
+
+  TileData *tile_data;
+  int allocated_tiles;
+
+  TileBufferDec tile_buffers[MAX_TILE_ROWS][MAX_TILE_COLS];
+
+  AV1LfSync lf_row_sync;
+
+  aom_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.
+
+  int tile_size_bytes;
+#if CONFIG_EXT_TILE
+  int tile_col_size_bytes;
+  int dec_tile_row, dec_tile_col;
+#endif  // CONFIG_EXT_TILE
+#if CONFIG_ACCOUNTING
+  int acct_enabled;
+  Accounting accounting;
+#endif
+  size_t uncomp_hdr_size;       // Size of the uncompressed header
+  size_t first_partition_size;  // Size of the compressed header
+#if CONFIG_TILE_GROUPS
+  int tg_size;   // Number of tiles in the current tilegroup
+  int tg_start;  // First tile in the current tilegroup
+  int tg_size_bit_offset;
+#endif
+#if CONFIG_REFERENCE_BUFFER
+  SequenceHeader seq_params;
+#endif
+#if CONFIG_INSPECTION
+  aom_inspect_cb inspect_cb;
+  void *inspect_ctx;
+#endif
+} AV1Decoder;
+
+int av1_receive_compressed_data(struct AV1Decoder *pbi, size_t size,
+                                const uint8_t **dest);
+
+int av1_get_raw_frame(struct AV1Decoder *pbi, YV12_BUFFER_CONFIG *sd);
+
+int av1_get_frame_to_show(struct AV1Decoder *pbi, YV12_BUFFER_CONFIG *frame);
+
+aom_codec_err_t av1_copy_reference_dec(struct AV1Decoder *pbi,
+                                       AOM_REFFRAME ref_frame_flag,
+                                       YV12_BUFFER_CONFIG *sd);
+
+aom_codec_err_t av1_set_reference_dec(AV1_COMMON *cm,
+                                      AOM_REFFRAME ref_frame_flag,
+                                      YV12_BUFFER_CONFIG *sd);
+
+static INLINE uint8_t read_marker(aom_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".
+aom_codec_err_t av1_parse_superframe_index(const uint8_t *data, size_t data_sz,
+                                           uint32_t sizes[8], int *count,
+                                           aom_decrypt_cb decrypt_cb,
+                                           void *decrypt_state);
+
+struct AV1Decoder *av1_decoder_create(BufferPool *const pool);
+
+void av1_decoder_remove(struct AV1Decoder *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);
+    }
+  }
+}
+
+#if CONFIG_EXT_REFS
+static INLINE int dec_is_ref_frame_buf(AV1Decoder *const pbi,
+                                       RefCntBuffer *frame_buf) {
+  AV1_COMMON *const cm = &pbi->common;
+  int i;
+  for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    if (ref_frame->idx == INVALID_IDX) continue;
+    if (frame_buf == &cm->buffer_pool->frame_bufs[ref_frame->idx]) break;
+  }
+  return (i < INTER_REFS_PER_FRAME);
+}
+#endif  // CONFIG_EXT_REFS
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_DECODER_DECODER_H_
diff --git a/third_party/aom/av1/decoder/decodetxb.c b/third_party/aom/av1/decoder/decodetxb.c
new file mode 100644
index 000000000..e1db09775
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodetxb.c
@@ -0,0 +1,286 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/scan.h"
+#include "av1/common/idct.h"
+#include "av1/common/txb_common.h"
+#include "av1/decoder/decodemv.h"
+#include "av1/decoder/decodetxb.h"
+#include "av1/decoder/dsubexp.h"
+
+#define ACCT_STR __func__
+
+static int read_golomb(MACROBLOCKD *xd, aom_reader *r) {
+  int x = 1;
+  int length = 0;
+  int i = 0;
+
+  while (!i) {
+    i = aom_read_bit(r, ACCT_STR);
+    ++length;
+    if (length >= 32) {
+      aom_internal_error(xd->error_info, AOM_CODEC_CORRUPT_FRAME,
+                         "Invalid length in read_golomb");
+      break;
+    }
+  }
+
+  for (i = 0; i < length - 1; ++i) {
+    x <<= 1;
+    x += aom_read_bit(r, ACCT_STR);
+  }
+
+  return x - 1;
+}
+
+uint8_t av1_read_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+                            aom_reader *r, int block, int plane,
+                            tran_low_t *tcoeffs, TXB_CTX *txb_ctx,
+                            int16_t *max_scan_line, int *eob) {
+  FRAME_COUNTS *counts = xd->counts;
+  TX_SIZE tx_size = get_tx_size(plane, xd);
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  aom_prob *nz_map = cm->fc->nz_map[tx_size][plane_type];
+  aom_prob *eob_flag = cm->fc->eob_flag[tx_size][plane_type];
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const int seg_eob = tx_size_2d[tx_size];
+  int c = 0;
+  int update_eob = -1;
+  const int16_t *const dequant = xd->plane[plane].seg_dequant[mbmi->segment_id];
+  const int shift = av1_get_tx_scale(tx_size);
+  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
+  int cul_level = 0;
+  unsigned int(*nz_map_count)[SIG_COEF_CONTEXTS][2];
+  uint8_t txb_mask[32 * 32] = { 0 };
+
+  nz_map_count = (counts) ? &counts->nz_map[tx_size][plane_type] : NULL;
+
+  memset(tcoeffs, 0, sizeof(*tcoeffs) * seg_eob);
+
+  int all_zero =
+      aom_read(r, cm->fc->txb_skip[tx_size][txb_ctx->txb_skip_ctx], ACCT_STR);
+  if (xd->counts)
+    ++xd->counts->txb_skip[tx_size][txb_ctx->txb_skip_ctx][all_zero];
+
+  *eob = 0;
+  if (all_zero) {
+    *max_scan_line = 0;
+    return 0;
+  }
+
+#if CONFIG_TXK_SEL
+  av1_read_tx_type(cm, xd, block, plane, r);
+#endif
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int16_t *scan = scan_order->scan;
+
+  for (c = 0; c < seg_eob; ++c) {
+    int is_nz;
+    int coeff_ctx = get_nz_map_ctx(tcoeffs, txb_mask, scan[c], bwl);
+    int eob_ctx = get_eob_ctx(tcoeffs, scan[c], bwl);
+
+    if (c < seg_eob - 1)
+      is_nz = aom_read(r, nz_map[coeff_ctx], tx_size);
+    else
+      is_nz = 1;
+
+    // set non-zero coefficient map.
+    tcoeffs[scan[c]] = is_nz;
+
+    if (c == seg_eob - 1) {
+      ++c;
+      break;
+    }
+
+    if (counts) ++(*nz_map_count)[coeff_ctx][is_nz];
+
+    if (is_nz) {
+      int is_eob = aom_read(r, eob_flag[eob_ctx], tx_size);
+      if (counts) ++counts->eob_flag[tx_size][plane_type][eob_ctx][is_eob];
+      if (is_eob) break;
+    }
+    txb_mask[scan[c]] = 1;
+  }
+
+  *eob = AOMMIN(seg_eob, c + 1);
+  *max_scan_line = *eob;
+
+  int i;
+  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+    aom_prob *coeff_base = cm->fc->coeff_base[tx_size][plane_type][i];
+
+    update_eob = 0;
+    for (c = *eob - 1; c >= 0; --c) {
+      tran_low_t *v = &tcoeffs[scan[c]];
+      int sign;
+      int ctx;
+
+      if (*v <= i) continue;
+
+      ctx = get_base_ctx(tcoeffs, scan[c], bwl, i + 1);
+
+      if (aom_read(r, coeff_base[ctx], tx_size)) {
+        *v = i + 1;
+        cul_level += i + 1;
+
+        if (counts) ++counts->coeff_base[tx_size][plane_type][i][ctx][1];
+
+        if (c == 0) {
+          int dc_sign_ctx = txb_ctx->dc_sign_ctx;
+          sign = aom_read(r, cm->fc->dc_sign[plane_type][dc_sign_ctx], tx_size);
+          if (counts) ++counts->dc_sign[plane_type][dc_sign_ctx][sign];
+        } else {
+          sign = aom_read_bit(r, ACCT_STR);
+        }
+        if (sign) *v = -(*v);
+        continue;
+      }
+      *v = i + 2;
+      if (counts) ++counts->coeff_base[tx_size][plane_type][i][ctx][0];
+
+      // update the eob flag for coefficients with magnitude above 1.
+      update_eob = AOMMAX(update_eob, c);
+    }
+  }
+
+  for (c = update_eob; c >= 0; --c) {
+    tran_low_t *v = &tcoeffs[scan[c]];
+    int sign;
+    int idx;
+    int ctx;
+
+    if (*v <= NUM_BASE_LEVELS) continue;
+
+    if (c == 0) {
+      int dc_sign_ctx = txb_ctx->dc_sign_ctx;
+      sign = aom_read(r, cm->fc->dc_sign[plane_type][dc_sign_ctx], tx_size);
+      if (counts) ++counts->dc_sign[plane_type][dc_sign_ctx][sign];
+    } else {
+      sign = aom_read_bit(r, ACCT_STR);
+    }
+
+    ctx = get_level_ctx(tcoeffs, scan[c], bwl);
+
+    if (cm->fc->coeff_lps[tx_size][plane_type][ctx] == 0) exit(0);
+
+    for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
+      if (aom_read(r, cm->fc->coeff_lps[tx_size][plane_type][ctx], tx_size)) {
+        *v = (idx + 1 + NUM_BASE_LEVELS);
+        if (sign) *v = -(*v);
+        cul_level += abs(*v);
+
+        if (counts) ++counts->coeff_lps[tx_size][plane_type][ctx][1];
+        break;
+      }
+      if (counts) ++counts->coeff_lps[tx_size][plane_type][ctx][0];
+    }
+    if (idx < COEFF_BASE_RANGE) continue;
+
+    // decode 0-th order Golomb code
+    *v = read_golomb(xd, r) + COEFF_BASE_RANGE + 1 + NUM_BASE_LEVELS;
+    if (sign) *v = -(*v);
+    cul_level += abs(*v);
+  }
+
+  for (c = 0; c < *eob; ++c) {
+    int16_t dqv = (c == 0) ? dequant[0] : dequant[1];
+    tran_low_t *v = &tcoeffs[scan[c]];
+    int sign = (*v) < 0;
+    *v = (abs(*v) * dqv) >> shift;
+    if (sign) *v = -(*v);
+  }
+
+  cul_level = AOMMIN(63, cul_level);
+
+  // DC value
+  set_dc_sign(&cul_level, tcoeffs[0]);
+
+  return cul_level;
+}
+
+uint8_t av1_read_coeffs_txb_facade(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                   aom_reader *r, int row, int col, int block,
+                                   int plane, tran_low_t *tcoeffs,
+                                   int16_t *max_scan_line, int *eob) {
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct macroblockd_plane *pd = &xd->plane[plane];
+
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+  const BLOCK_SIZE plane_bsize =
+      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif  // CONFIG_CHROMA_2X2
+#else   // CONFIG_CB4X4
+  const BLOCK_SIZE plane_bsize =
+      get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
+#endif  // CONFIG_CB4X4
+
+  TX_SIZE tx_size = get_tx_size(plane, xd);
+  TXB_CTX txb_ctx;
+  get_txb_ctx(plane_bsize, tx_size, plane, pd->above_context + col,
+              pd->left_context + row, &txb_ctx);
+  uint8_t cul_level = av1_read_coeffs_txb(cm, xd, r, block, plane, tcoeffs,
+                                          &txb_ctx, max_scan_line, eob);
+#if CONFIG_ADAPT_SCAN
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  if (xd->counts && *eob > 0)
+    av1_update_scan_count_facade(cm, xd->counts, tx_size, tx_type, pd->dqcoeff,
+                                 *eob);
+#endif
+  av1_set_contexts(xd, pd, plane, tx_size, cul_level, col, row);
+  return cul_level;
+}
+
+static void read_txb_probs(FRAME_CONTEXT *fc, const TX_SIZE tx_size,
+                           aom_reader *r) {
+  int plane, ctx, level;
+
+  if (aom_read_bit(r, ACCT_STR) == 0) return;
+
+  for (ctx = 0; ctx < TXB_SKIP_CONTEXTS; ++ctx)
+    av1_diff_update_prob(r, &fc->txb_skip[tx_size][ctx], ACCT_STR);
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < SIG_COEF_CONTEXTS; ++ctx)
+      av1_diff_update_prob(r, &fc->nz_map[tx_size][plane][ctx], ACCT_STR);
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < EOB_COEF_CONTEXTS; ++ctx)
+      av1_diff_update_prob(r, &fc->eob_flag[tx_size][plane][ctx], ACCT_STR);
+
+  for (level = 0; level < NUM_BASE_LEVELS; ++level)
+    for (plane = 0; plane < PLANE_TYPES; ++plane)
+      for (ctx = 0; ctx < COEFF_BASE_CONTEXTS; ++ctx)
+        av1_diff_update_prob(r, &fc->coeff_base[tx_size][plane][level][ctx],
+                             ACCT_STR);
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx)
+      av1_diff_update_prob(r, &fc->coeff_lps[tx_size][plane][ctx], ACCT_STR);
+}
+
+void av1_read_txb_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode, aom_reader *r) {
+  const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+  TX_SIZE tx_size;
+  int ctx, plane;
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < DC_SIGN_CONTEXTS; ++ctx)
+      av1_diff_update_prob(r, &fc->dc_sign[plane][ctx], ACCT_STR);
+
+  for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
+    read_txb_probs(fc, tx_size, r);
+}
diff --git a/third_party/aom/av1/decoder/decodetxb.h b/third_party/aom/av1/decoder/decodetxb.h
new file mode 100644
index 000000000..ee1bf6a3d
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodetxb.h
@@ -0,0 +1,31 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef DECODETXB_H_
+#define DECODETXB_H_
+
+#include "./aom_config.h"
+#include "av1/common/blockd.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/txb_common.h"
+#include "aom_dsp/bitreader.h"
+
+uint8_t av1_read_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+                            aom_reader *r, int block, int plane,
+                            tran_low_t *tcoeffs, TXB_CTX *txb_ctx,
+                            int16_t *max_scan_line, int *eob);
+
+uint8_t av1_read_coeffs_txb_facade(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                   aom_reader *r, int row, int col, int block,
+                                   int plane, tran_low_t *tcoeffs,
+                                   int16_t *max_scan_line, int *eob);
+void av1_read_txb_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode, aom_reader *r);
+#endif  //  DECODETXB_H_
diff --git a/third_party/aom/av1/decoder/detokenize.c b/third_party/aom/av1/decoder/detokenize.c
new file mode 100644
index 000000000..494f1681f
--- /dev/null
+++ b/third_party/aom/av1/decoder/detokenize.c
@@ -0,0 +1,467 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#if !CONFIG_PVQ
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#endif  // !CONFIG_PVQ
+
+#include "av1/common/blockd.h"
+
+#define ACCT_STR __func__
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/common/idct.h"
+#include "av1/decoder/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)
+
+#if CONFIG_NEW_MULTISYMBOL
+#define READ_COEFF(prob_name, cdf_name, num, r) read_coeff(cdf_name, num, r);
+static INLINE int read_coeff(const aom_cdf_prob *const *cdf, int n,
+                             aom_reader *r) {
+  int val = 0;
+  int i = 0;
+  int count = 0;
+  while (count < n) {
+    const int size = AOMMIN(n - count, 4);
+    val |= aom_read_cdf(r, cdf[i++], 1 << size, ACCT_STR) << count;
+    count += size;
+  }
+  return val;
+}
+#else
+#define READ_COEFF(prob_name, cdf_name, num, r) read_coeff(prob_name, num, r);
+static INLINE int read_coeff(const aom_prob *probs, int n, aom_reader *r) {
+  int i, val = 0;
+  for (i = 0; i < n; ++i) val = (val << 1) | aom_read(r, probs[i], ACCT_STR);
+  return val;
+}
+
+#endif
+
+static int token_to_value(aom_reader *const r, int token, TX_SIZE tx_size,
+                          int bit_depth) {
+#if !CONFIG_HIGHBITDEPTH
+  assert(bit_depth == 8);
+#endif  // !CONFIG_HIGHBITDEPTH
+
+  switch (token) {
+    case ZERO_TOKEN:
+    case ONE_TOKEN:
+    case TWO_TOKEN:
+    case THREE_TOKEN:
+    case FOUR_TOKEN: return token;
+    case CATEGORY1_TOKEN:
+      return CAT1_MIN_VAL + READ_COEFF(av1_cat1_prob, av1_cat1_cdf, 1, r);
+    case CATEGORY2_TOKEN:
+      return CAT2_MIN_VAL + READ_COEFF(av1_cat2_prob, av1_cat2_cdf, 2, r);
+    case CATEGORY3_TOKEN:
+      return CAT3_MIN_VAL + READ_COEFF(av1_cat3_prob, av1_cat3_cdf, 3, r);
+    case CATEGORY4_TOKEN:
+      return CAT4_MIN_VAL + READ_COEFF(av1_cat4_prob, av1_cat4_cdf, 4, r);
+    case CATEGORY5_TOKEN:
+      return CAT5_MIN_VAL + READ_COEFF(av1_cat5_prob, av1_cat5_cdf, 5, r);
+    case CATEGORY6_TOKEN: {
+      const int skip_bits = (int)sizeof(av1_cat6_prob) -
+                            av1_get_cat6_extrabits_size(tx_size, bit_depth);
+      return CAT6_MIN_VAL + READ_COEFF(av1_cat6_prob + skip_bits, av1_cat6_cdf,
+                                       18 - skip_bits, r);
+    }
+    default:
+      assert(0);  // Invalid token.
+      return -1;
+  }
+}
+
+static int decode_coefs(MACROBLOCKD *xd, PLANE_TYPE type, tran_low_t *dqcoeff,
+                        TX_SIZE tx_size, TX_TYPE tx_type, const int16_t *dq,
+#if CONFIG_NEW_QUANT
+                        dequant_val_type_nuq *dq_val,
+#endif  // CONFIG_NEW_QUANT
+#if CONFIG_AOM_QM
+                        const qm_val_t *iqm[2][TX_SIZES],
+#endif  // CONFIG_AOM_QM
+                        int ctx, const int16_t *scan, const int16_t *nb,
+                        int16_t *max_scan_line, aom_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *const ec_ctx = xd->fc;
+#endif
+  const int max_eob = tx_size_2d[tx_size];
+  const int ref = is_inter_block(&xd->mi[0]->mbmi);
+#if CONFIG_AOM_QM
+  const qm_val_t *iqmatrix = iqm[!ref][tx_size];
+#endif  // CONFIG_AOM_QM
+  int band, c = 0;
+  const int tx_size_ctx = txsize_sqr_map[tx_size];
+#if CONFIG_NEW_TOKENSET
+  aom_cdf_prob(*coef_head_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_head_cdfs[tx_size_ctx][type][ref];
+  aom_cdf_prob(*coef_tail_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_tail_cdfs[tx_size_ctx][type][ref];
+  int val = 0;
+
+#if !CONFIG_EC_ADAPT
+  unsigned int *blockz_count;
+  unsigned int(*coef_counts)[COEFF_CONTEXTS][UNCONSTRAINED_NODES + 1] = NULL;
+  unsigned int(*eob_branch_count)[COEFF_CONTEXTS] = NULL;
+#endif
+#else
+  aom_prob(*coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      ec_ctx->coef_probs[tx_size_ctx][type][ref];
+  const aom_prob *prob;
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob(*coef_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_cdfs[tx_size_ctx][type][ref];
+  aom_cdf_prob(*cdf)[CDF_SIZE(ENTROPY_TOKENS)];
+#endif  // CONFIG_EC_MULTISYMBOL
+  unsigned int(*coef_counts)[COEFF_CONTEXTS][UNCONSTRAINED_NODES + 1] = NULL;
+  unsigned int(*eob_branch_count)[COEFF_CONTEXTS] = NULL;
+#endif  // CONFIG_NEW_TOKENSET
+  uint8_t token_cache[MAX_TX_SQUARE];
+  const uint8_t *band_translate = get_band_translate(tx_size);
+  int dq_shift;
+  int v, token;
+  int16_t dqv = dq[0];
+#if CONFIG_NEW_QUANT
+  const tran_low_t *dqv_val = &dq_val[0][0];
+#endif  // CONFIG_NEW_QUANT
+  (void)tx_type;
+#if CONFIG_AOM_QM
+  (void)iqmatrix;
+#endif  // CONFIG_AOM_QM
+
+  if (counts) {
+#if !CONFIG_NEW_TOKENSET || !CONFIG_EC_ADAPT
+    coef_counts = counts->coef[tx_size_ctx][type][ref];
+    eob_branch_count = counts->eob_branch[tx_size_ctx][type][ref];
+#endif
+#if CONFIG_NEW_TOKENSET && !CONFIG_EC_ADAPT
+    blockz_count = counts->blockz_count[tx_size_ctx][type][ref][ctx];
+#endif
+  }
+
+  dq_shift = av1_get_tx_scale(tx_size);
+
+#if CONFIG_NEW_TOKENSET
+  band = *band_translate++;
+
+  int more_data = 1;
+  while (more_data) {
+    int comb_token;
+    int last_pos = (c + 1 == max_eob);
+    int first_pos = (c == 0);
+
+#if CONFIG_NEW_QUANT
+    dqv_val = &dq_val[band][0];
+#endif  // CONFIG_NEW_QUANT
+
+    comb_token = last_pos ? 2 * aom_read_bit(r, ACCT_STR) + 2
+                          : aom_read_symbol(r, coef_head_cdfs[band][ctx],
+                                            HEAD_TOKENS + first_pos, ACCT_STR) +
+                                !first_pos;
+    if (first_pos) {
+#if !CONFIG_EC_ADAPT
+      if (counts) ++blockz_count[comb_token != 0];
+#endif
+      if (comb_token == 0) return 0;
+    }
+    token = comb_token >> 1;
+
+    while (!token) {
+      *max_scan_line = AOMMAX(*max_scan_line, scan[c]);
+      token_cache[scan[c]] = 0;
+#if !CONFIG_EC_ADAPT
+      if (counts && !last_pos) {
+        ++coef_counts[band][ctx][ZERO_TOKEN];
+      }
+#endif
+      ++c;
+      dqv = dq[1];
+      ctx = get_coef_context(nb, token_cache, c);
+      band = *band_translate++;
+
+      last_pos = (c + 1 == max_eob);
+
+      comb_token = last_pos ? 2 * aom_read_bit(r, ACCT_STR) + 2
+                            : aom_read_symbol(r, coef_head_cdfs[band][ctx],
+                                              HEAD_TOKENS, ACCT_STR) +
+                                  1;
+      token = comb_token >> 1;
+    }
+
+    more_data = comb_token & 1;
+#if !CONFIG_EC_ADAPT
+    if (counts && !last_pos) {
+      ++coef_counts[band][ctx][token];
+      ++eob_branch_count[band][ctx];
+      if (!more_data) ++coef_counts[band][ctx][EOB_MODEL_TOKEN];
+    }
+#endif
+
+    if (token > ONE_TOKEN)
+      token +=
+          aom_read_symbol(r, coef_tail_cdfs[band][ctx], TAIL_TOKENS, ACCT_STR);
+#if CONFIG_NEW_QUANT
+    dqv_val = &dq_val[band][0];
+#endif  // CONFIG_NEW_QUANT
+
+    *max_scan_line = AOMMAX(*max_scan_line, scan[c]);
+    token_cache[scan[c]] = av1_pt_energy_class[token];
+
+    val = token_to_value(r, token, tx_size,
+#if CONFIG_HIGHBITDEPTH
+                         xd->bd);
+#else
+                         8);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_NEW_QUANT
+    v = av1_dequant_abscoeff_nuq(val, dqv, dqv_val);
+    v = dq_shift ? ROUND_POWER_OF_TWO(v, dq_shift) : v;
+#else
+#if CONFIG_AOM_QM
+    dqv = ((iqmatrix[scan[c]] * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >>
+          AOM_QM_BITS;
+#endif
+    v = (val * dqv) >> dq_shift;
+#endif
+
+    v = aom_read_bit(r, ACCT_STR) ? -v : v;
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+#if CONFIG_HIGHBITDEPTH
+    check_range(v, xd->bd);
+#else
+    check_range(v, 8);
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+
+    dqcoeff[scan[c]] = v;
+
+    ++c;
+    more_data &= (c < max_eob);
+    if (!more_data) break;
+    dqv = dq[1];
+    ctx = get_coef_context(nb, token_cache, c);
+    band = *band_translate++;
+
+#else  // CONFIG_NEW_TOKENSET
+  while (c < max_eob) {
+    int val = -1;
+    band = *band_translate++;
+    prob = coef_probs[band][ctx];
+    if (counts) ++eob_branch_count[band][ctx];
+    if (!aom_read(r, prob[EOB_CONTEXT_NODE], ACCT_STR)) {
+      INCREMENT_COUNT(EOB_MODEL_TOKEN);
+      break;
+    }
+
+#if CONFIG_NEW_QUANT
+    dqv_val = &dq_val[band][0];
+#endif  // CONFIG_NEW_QUANT
+
+    while (!aom_read(r, prob[ZERO_CONTEXT_NODE], ACCT_STR)) {
+      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 CONFIG_NEW_QUANT
+      dqv_val = &dq_val[band][0];
+#endif  // CONFIG_NEW_QUANT
+    }
+
+    *max_scan_line = AOMMAX(*max_scan_line, scan[c]);
+
+#if CONFIG_EC_MULTISYMBOL
+    cdf = &coef_cdfs[band][ctx];
+    token = ONE_TOKEN +
+            aom_read_symbol(r, *cdf, CATEGORY6_TOKEN - ONE_TOKEN + 1, ACCT_STR);
+    INCREMENT_COUNT(ONE_TOKEN + (token > ONE_TOKEN));
+    assert(token != ZERO_TOKEN);
+    val = token_to_value(r, token, tx_size,
+#if CONFIG_HIGHBITDEPTH
+                         xd->bd);
+#else
+                         8);
+#endif  // CONFIG_HIGHBITDEPTH
+#else   // CONFIG_EC_MULTISYMBOL
+    if (!aom_read(r, prob[ONE_CONTEXT_NODE], ACCT_STR)) {
+      INCREMENT_COUNT(ONE_TOKEN);
+      token = ONE_TOKEN;
+      val = 1;
+    } else {
+      INCREMENT_COUNT(TWO_TOKEN);
+      token = aom_read_tree(r, av1_coef_con_tree,
+                            av1_pareto8_full[prob[PIVOT_NODE] - 1], ACCT_STR);
+      assert(token != ZERO_TOKEN && token != ONE_TOKEN);
+      val = token_to_value(r, token, tx_size,
+#if CONFIG_HIGHBITDEPTH
+                           xd->bd);
+#else
+                           8);
+#endif  // CONFIG_HIGHBITDEPTH
+    }
+#endif  // CONFIG_EC_MULTISYMBOL
+#if CONFIG_NEW_QUANT
+    v = av1_dequant_abscoeff_nuq(val, dqv, dqv_val);
+    v = dq_shift ? ROUND_POWER_OF_TWO(v, dq_shift) : v;
+#else
+#if CONFIG_AOM_QM
+    dqv = ((iqmatrix[scan[c]] * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >>
+          AOM_QM_BITS;
+#endif
+    v = (val * dqv) >> dq_shift;
+#endif  // CONFIG_NEW_QUANT
+
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+#if CONFIG_HIGHBITDEPTH
+    dqcoeff[scan[c]] =
+        highbd_check_range((aom_read_bit(r, ACCT_STR) ? -v : v), xd->bd);
+#else
+    dqcoeff[scan[c]] = check_range(aom_read_bit(r, ACCT_STR) ? -v : v, 8);
+#endif  // CONFIG_HIGHBITDEPTH
+#else
+    dqcoeff[scan[c]] = aom_read_bit(r, ACCT_STR) ? -v : v;
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+    token_cache[scan[c]] = av1_pt_energy_class[token];
+    ++c;
+    ctx = get_coef_context(nb, token_cache, c);
+    dqv = dq[1];
+#endif  // CONFIG_NEW_TOKENSET
+  }
+
+  return c;
+}
+#endif  // !CONFIG_PVQ
+
+#if CONFIG_PALETTE
+void av1_decode_palette_tokens(MACROBLOCKD *const xd, int plane,
+                               aom_reader *r) {
+  const MODE_INFO *const mi = xd->mi[0];
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  uint8_t color_order[PALETTE_MAX_SIZE];
+  const int n = mbmi->palette_mode_info.palette_size[plane];
+  int i, j;
+  uint8_t *const color_map = xd->plane[plane].color_index_map;
+  const aom_prob(
+      *const prob)[PALETTE_COLOR_INDEX_CONTEXTS][PALETTE_COLORS - 1] =
+      plane ? av1_default_palette_uv_color_index_prob
+            : av1_default_palette_y_color_index_prob;
+  int plane_block_width, plane_block_height, rows, cols;
+  av1_get_block_dimensions(mbmi->sb_type, plane, xd, &plane_block_width,
+                           &plane_block_height, &rows, &cols);
+  assert(plane == 0 || plane == 1);
+
+#if CONFIG_PALETTE_THROUGHPUT
+  // Run wavefront on the palette map index decoding.
+  for (i = 1; i < rows + cols - 1; ++i) {
+    for (j = AOMMIN(i, cols - 1); j >= AOMMAX(0, i - rows + 1); --j) {
+      const int color_ctx = av1_get_palette_color_index_context(
+          color_map, plane_block_width, (i - j), j, n, color_order, NULL);
+      const int color_idx =
+          aom_read_tree(r, av1_palette_color_index_tree[n - 2],
+                        prob[n - 2][color_ctx], ACCT_STR);
+      assert(color_idx >= 0 && color_idx < n);
+      color_map[(i - j) * plane_block_width + j] = color_order[color_idx];
+    }
+  }
+  // Copy last column to extra columns.
+  if (cols < plane_block_width) {
+    for (i = 0; i < plane_block_height; ++i) {
+      memset(color_map + i * plane_block_width + cols,
+             color_map[i * plane_block_width + cols - 1],
+             (plane_block_width - cols));
+    }
+  }
+#else
+  for (i = 0; i < rows; ++i) {
+    for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
+      const int color_ctx = av1_get_palette_color_index_context(
+          color_map, plane_block_width, i, j, n, color_order, NULL);
+      const int color_idx =
+          aom_read_tree(r, av1_palette_color_index_tree[n - PALETTE_MIN_SIZE],
+                        prob[n - PALETTE_MIN_SIZE][color_ctx], ACCT_STR);
+      assert(color_idx >= 0 && color_idx < n);
+      color_map[i * plane_block_width + j] = color_order[color_idx];
+    }
+    memset(color_map + i * plane_block_width + cols,
+           color_map[i * plane_block_width + cols - 1],
+           (plane_block_width - cols));  // Copy last column to extra columns.
+  }
+#endif  // CONFIG_PALETTE_THROUGHPUT
+  // Copy last row to extra rows.
+  for (i = rows; i < plane_block_height; ++i) {
+    memcpy(color_map + i * plane_block_width,
+           color_map + (rows - 1) * plane_block_width, plane_block_width);
+  }
+}
+#endif  // CONFIG_PALETTE
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+int av1_decode_block_tokens(AV1_COMMON *cm, MACROBLOCKD *const xd, int plane,
+                            const SCAN_ORDER *sc, int x, int y, TX_SIZE tx_size,
+                            TX_TYPE tx_type, int16_t *max_scan_line,
+                            aom_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);
+#if CONFIG_NEW_QUANT
+  const int ref = is_inter_block(&xd->mi[0]->mbmi);
+  int dq =
+      get_dq_profile_from_ctx(xd->qindex[seg_id], ctx, ref, pd->plane_type);
+#endif  //  CONFIG_NEW_QUANT
+
+  const int eob =
+      decode_coefs(xd, pd->plane_type, pd->dqcoeff, tx_size, tx_type, dequant,
+#if CONFIG_NEW_QUANT
+                   pd->seg_dequant_nuq[seg_id][dq],
+#endif  // CONFIG_NEW_QUANT
+#if CONFIG_AOM_QM
+                   pd->seg_iqmatrix[seg_id],
+#endif  // CONFIG_AOM_QM
+                   ctx, sc->scan, sc->neighbors, max_scan_line, r);
+  av1_set_contexts(xd, pd, plane, tx_size, eob > 0, x, y);
+#if CONFIG_ADAPT_SCAN
+  if (xd->counts)
+    av1_update_scan_count_facade(cm, xd->counts, tx_size, tx_type, pd->dqcoeff,
+                                 eob);
+#else
+  (void)cm;
+#endif
+  return eob;
+}
+#endif  // !CONFIG_PVQ
diff --git a/third_party/aom/av1/decoder/detokenize.h b/third_party/aom/av1/decoder/detokenize.h
new file mode 100644
index 000000000..ba4066603
--- /dev/null
+++ b/third_party/aom/av1/decoder/detokenize.h
@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DETOKENIZE_H_
+#define AV1_DECODER_DETOKENIZE_H_
+
+#include "./aom_config.h"
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+#include "av1/decoder/decoder.h"
+#include "av1/common/scan.h"
+#endif  // !CONFIG_PVQ
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_PALETTE
+void av1_decode_palette_tokens(MACROBLOCKD *const xd, int plane, aom_reader *r);
+#endif  // CONFIG_PALETTE
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+int av1_decode_block_tokens(AV1_COMMON *cm, MACROBLOCKD *const xd, int plane,
+                            const SCAN_ORDER *sc, int x, int y, TX_SIZE tx_size,
+                            TX_TYPE tx_type, int16_t *max_scan_line,
+                            aom_reader *r, int seg_id);
+#endif  // !CONFIG_PVQ
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+#endif  // AV1_DECODER_DETOKENIZE_H_
diff --git a/third_party/aom/av1/decoder/dsubexp.c b/third_party/aom/av1/decoder/dsubexp.c
new file mode 100644
index 000000000..5171f1144
--- /dev/null
+++ b/third_party/aom/av1/decoder/dsubexp.c
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/entropy.h"
+
+#include "av1/decoder/dsubexp.h"
+
+static int inv_recenter_nonneg(int v, int m) {
+  if (v > 2 * m) return v;
+
+  return (v & 1) ? m - ((v + 1) >> 1) : m + (v >> 1);
+}
+
+#define decode_uniform(r, ACCT_STR_NAME) \
+  decode_uniform_(r ACCT_STR_ARG(ACCT_STR_NAME))
+#define decode_term_subexp(r, ACCT_STR_NAME) \
+  decode_term_subexp_(r ACCT_STR_ARG(ACCT_STR_NAME))
+
+static int decode_uniform_(aom_reader *r ACCT_STR_PARAM) {
+  const int l = 8;
+  const int m = (1 << l) - 190;
+  const int v = aom_read_literal(r, l - 1, ACCT_STR_NAME);
+  return v < m ? v : (v << 1) - m + aom_read_bit(r, ACCT_STR_NAME);
+}
+
+static int inv_remap_prob(int v, int m) {
+  /* clang-format off */
+  static uint8_t inv_map_table[MAX_PROB - 1] = {
+      7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176, 189,
+    202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,  10,  11,
+     12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,  25,  26,  27,
+     28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,
+     44,  45,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  60,
+     61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  73,  74,  75,  76,
+     77,  78,  79,  80,  81,  82,  83,  84,  86,  87,  88,  89,  90,  91,  92,
+     93,  94,  95,  96,  97,  99, 100, 101, 102, 103, 104, 105, 106, 107, 108,
+    109, 110, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 125,
+    126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 138, 139, 140, 141,
+    142, 143, 144, 145, 146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157,
+    158, 159, 160, 161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173,
+    174, 175, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190,
+    191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,
+    207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221, 222,
+    223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238,
+    239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253
+  }; /* clang-format on */
+  assert(v < (int)(sizeof(inv_map_table) / sizeof(inv_map_table[0])));
+  v = inv_map_table[v];
+  m--;
+  if ((m << 1) <= MAX_PROB) {
+    return 1 + inv_recenter_nonneg(v, m);
+  } else {
+    return MAX_PROB - inv_recenter_nonneg(v, MAX_PROB - 1 - m);
+  }
+}
+
+static int decode_term_subexp_(aom_reader *r ACCT_STR_PARAM) {
+  if (!aom_read_bit(r, ACCT_STR_NAME))
+    return aom_read_literal(r, 4, ACCT_STR_NAME);
+  if (!aom_read_bit(r, ACCT_STR_NAME))
+    return aom_read_literal(r, 4, ACCT_STR_NAME) + 16;
+  if (!aom_read_bit(r, ACCT_STR_NAME))
+    return aom_read_literal(r, 5, ACCT_STR_NAME) + 32;
+  return decode_uniform(r, ACCT_STR_NAME) + 64;
+}
+
+void av1_diff_update_prob_(aom_reader *r, aom_prob *p ACCT_STR_PARAM) {
+  if (aom_read(r, DIFF_UPDATE_PROB, ACCT_STR_NAME)) {
+    const int delp = decode_term_subexp(r, ACCT_STR_NAME);
+    *p = (aom_prob)inv_remap_prob(delp, *p);
+  }
+}
diff --git a/third_party/aom/av1/decoder/dsubexp.h b/third_party/aom/av1/decoder/dsubexp.h
new file mode 100644
index 000000000..4bc38578c
--- /dev/null
+++ b/third_party/aom/av1/decoder/dsubexp.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DSUBEXP_H_
+#define AV1_DECODER_DSUBEXP_H_
+
+#include "aom_dsp/bitreader.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_ACCOUNTING
+#define av1_diff_update_prob(r, p, str) av1_diff_update_prob_(r, p, str)
+#else
+#define av1_diff_update_prob(r, p, str) av1_diff_update_prob_(r, p)
+#endif
+
+void av1_diff_update_prob_(aom_reader *r, aom_prob *p ACCT_STR_PARAM);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+#endif  // AV1_DECODER_DSUBEXP_H_
diff --git a/third_party/aom/av1/decoder/dthread.c b/third_party/aom/av1/decoder/dthread.c
new file mode 100644
index 000000000..50f8ed192
--- /dev/null
+++ b/third_party/aom/av1/decoder/dthread.c
@@ -0,0 +1,194 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "aom_mem/aom_mem.h"
+#include "av1/common/reconinter.h"
+#include "av1/decoder/dthread.h"
+#include "av1/decoder/decoder.h"
+
+// #define DEBUG_THREAD
+
+// TODO(hkuang): Clean up all the #ifdef in this file.
+void av1_frameworker_lock_stats(AVxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_lock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void av1_frameworker_unlock_stats(AVxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_unlock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void av1_frameworker_signal_stats(AVxWorker *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 av1_frameworker_wait(AVxWorker *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.
+    AVxWorker *const ref_worker = ref_buf->frame_worker_owner;
+    FrameWorkerData *const ref_worker_data =
+        (FrameWorkerData *)ref_worker->data1;
+    const AV1Decoder *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
+
+    av1_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;
+      av1_frameworker_unlock_stats(ref_worker);
+      aom_internal_error(&worker_data->pbi->common.error,
+                         AOM_CODEC_CORRUPT_FRAME,
+                         "Worker %p failed to decode frame", worker);
+    }
+    av1_frameworker_unlock_stats(ref_worker);
+  }
+#else
+  (void)worker;
+  (void)ref_buf;
+  (void)row;
+  (void)ref_buf;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void av1_frameworker_broadcast(RefCntBuffer *const buf, int row) {
+#if CONFIG_MULTITHREAD
+  AVxWorker *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
+
+  av1_frameworker_lock_stats(worker);
+  buf->row = row;
+  av1_frameworker_signal_stats(worker);
+  av1_frameworker_unlock_stats(worker);
+#else
+  (void)buf;
+  (void)row;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void av1_frameworker_copy_context(AVxWorker *const dst_worker,
+                                  AVxWorker *const src_worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const src_worker_data = (FrameWorkerData *)src_worker->data1;
+  FrameWorkerData *const dst_worker_data = (FrameWorkerData *)dst_worker->data1;
+  AV1_COMMON *const src_cm = &src_worker_data->pbi->common;
+  AV1_COMMON *const dst_cm = &dst_worker_data->pbi->common;
+  int i;
+
+  // Wait until source frame's context is ready.
+  av1_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;
+  av1_frameworker_unlock_stats(src_worker);
+
+  dst_cm->bit_depth = src_cm->bit_depth;
+#if CONFIG_HIGHBITDEPTH
+  dst_cm->use_highbitdepth = src_cm->use_highbitdepth;
+#endif
+#if CONFIG_EXT_REFS
+// TODO(zoeliu): To handle parallel decoding
+#endif  // CONFIG_EXT_REFS
+  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, TOTAL_REFS_PER_FRAME);
+  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/third_party/aom/av1/decoder/dthread.h b/third_party/aom/av1/decoder/dthread.h
new file mode 100644
index 000000000..c17053d9c
--- /dev/null
+++ b/third_party/aom/av1/decoder/dthread.h
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DTHREAD_H_
+#define AV1_DECODER_DTHREAD_H_
+
+#include "./aom_config.h"
+#include "aom_util/aom_thread.h"
+#include "aom/internal/aom_codec_internal.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1Common;
+struct AV1Decoder;
+
+// WorkerData for the FrameWorker thread. It contains all the information of
+// the worker and decode structures for decoding a frame.
+typedef struct FrameWorkerData {
+  struct AV1Decoder *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 av1_frameworker_lock_stats(AVxWorker *const worker);
+void av1_frameworker_unlock_stats(AVxWorker *const worker);
+void av1_frameworker_signal_stats(AVxWorker *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 av1_frameworker_wait(AVxWorker *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 av1_frameworker_broadcast(RefCntBuffer *const buf, int row);
+
+// Copy necessary decoding context from src worker to dst worker.
+void av1_frameworker_copy_context(AVxWorker *const dst_worker,
+                                  AVxWorker *const src_worker);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_DECODER_DTHREAD_H_
diff --git a/third_party/aom/av1/decoder/generic_decoder.c b/third_party/aom/av1/decoder/generic_decoder.c
new file mode 100644
index 000000000..0c7d71b9f
--- /dev/null
+++ b/third_party/aom/av1/decoder/generic_decoder.c
@@ -0,0 +1,110 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+
+#include "aom_dsp/bitreader.h"
+#include "av1/common/generic_code.h"
+#include "av1/common/odintrin.h"
+#include "pvq_decoder.h"
+
+/** Decodes a value from 0 to N-1 (with N up to 16) based on a cdf and adapts
+ * the cdf accordingly.
+ *
+ * @param [in,out] r     multi-symbol entropy decoder
+ * @param [in,out] cdf   CDF of the variable (Q15)
+ * @param [in]     n     number of values possible
+ * @param [in,out] count number of symbols encoded with that cdf so far
+ * @param [in]     rate  adaptation rate shift (smaller is faster)
+ * @return decoded variable
+ */
+int aom_decode_cdf_adapt_q15_(aom_reader *r, uint16_t *cdf, int n,
+ int *count, int rate ACCT_STR_PARAM) {
+  int val;
+  int i;
+  if (*count == 0) {
+    int ft;
+    ft = cdf[n - 1];
+    for (i = 0; i < n; i++) {
+      cdf[i] = AOM_ICDF(cdf[i]*32768/ft);
+    }
+  }
+  val = aom_read_cdf(r, cdf, n, ACCT_STR_NAME);
+  aom_cdf_adapt_q15(val, cdf, n, count, rate);
+  return val;
+}
+
+/** Encodes a random variable using a "generic" model, assuming that the
+ * distribution is one-sided (zero and up), has a single mode, and decays
+ * exponentially past the model.
+ *
+ * @param [in,out] r     multi-symbol entropy decoder
+ * @param [in,out] model generic probability model
+ * @param [in]     x     variable being encoded
+ * @param [in,out] ExQ16 expectation of x (adapted)
+ * @param [in]     integration integration period of ExQ16 (leaky average over
+ * 1<<integration samples)
+ *
+ * @retval decoded variable x
+ */
+int generic_decode_(aom_reader *r, generic_encoder *model,
+ int *ex_q16, int integration ACCT_STR_PARAM) {
+  int lg_q1;
+  int shift;
+  int id;
+  uint16_t *cdf;
+  int xs;
+  int lsb;
+  int x;
+  lsb = 0;
+  lg_q1 = log_ex(*ex_q16);
+  /* If expectation is too large, shift x to ensure that
+     all we have past xs=15 is the exponentially decaying tail
+     of the distribution. */
+  shift = OD_MAXI(0, (lg_q1 - 5) >> 1);
+  /* Choose the cdf to use: we have two per "octave" of ExQ16. */
+  id = OD_MINI(GENERIC_TABLES - 1, lg_q1);
+  cdf = model->cdf[id];
+  xs = aom_read_symbol_pvq(r, cdf, 16, ACCT_STR_NAME);
+  if (xs == 15) {
+    int e;
+    unsigned decay;
+    /* Estimate decay based on the assumption that the distribution is close
+       to Laplacian for large values. We should probably have an adaptive
+       estimate instead. Note: The 2* is a kludge that's not fully understood
+       yet. */
+    OD_ASSERT(*ex_q16 < INT_MAX >> 1);
+    e = ((2**ex_q16 >> 8) + (1 << shift >> 1)) >> shift;
+    decay = OD_MAXI(2, OD_MINI(254, 256*e/(e + 256)));
+    xs += aom_laplace_decode_special(r, decay, ACCT_STR_NAME);
+  }
+  if (shift != 0) {
+    int special;
+    /* Because of the rounding, there's only half the number of possibilities
+       for xs=0 */
+    special = xs == 0;
+    if (shift - special > 0) {
+      lsb = aom_read_literal(r, shift - special, ACCT_STR_NAME);
+    }
+    lsb -= !special << (shift - 1);
+  }
+  x = (xs << shift) + lsb;
+  generic_model_update(ex_q16, x, integration);
+  OD_LOG((OD_LOG_ENTROPY_CODER, OD_LOG_DEBUG,
+   "dec: %d %d %d %d %d %x", *ex_q16, x, shift, id, xs, dec->rng));
+  return x;
+}
diff --git a/third_party/aom/av1/decoder/inspection.c b/third_party/aom/av1/decoder/inspection.c
new file mode 100644
index 000000000..2e8a61087
--- /dev/null
+++ b/third_party/aom/av1/decoder/inspection.c
@@ -0,0 +1,103 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "av1/decoder/decoder.h"
+#include "av1/decoder/inspection.h"
+#include "av1/common/enums.h"
+#if CONFIG_CDEF
+#include "av1/common/cdef.h"
+#endif
+
+void ifd_init(insp_frame_data *fd, int frame_width, int frame_height) {
+  fd->mi_cols = ALIGN_POWER_OF_TWO(frame_width, 3) >> MI_SIZE_LOG2;
+  fd->mi_rows = ALIGN_POWER_OF_TWO(frame_height, 3) >> MI_SIZE_LOG2;
+  fd->mi_grid = (insp_mi_data *)aom_malloc(sizeof(insp_mi_data) * fd->mi_rows *
+                                           fd->mi_cols);
+}
+
+void ifd_clear(insp_frame_data *fd) {
+  aom_free(fd->mi_grid);
+  fd->mi_grid = NULL;
+}
+
+/* TODO(negge) This function may be called by more than one thread when using
+               a multi-threaded decoder and this may cause a data race. */
+int ifd_inspect(insp_frame_data *fd, void *decoder) {
+  struct AV1Decoder *pbi = (struct AV1Decoder *)decoder;
+  AV1_COMMON *const cm = &pbi->common;
+  // TODO(negge): Should this function just call ifd_clear() and ifd_init()?
+  if (fd->mi_rows != cm->mi_rows || fd->mi_cols != cm->mi_cols) {
+    return 0;
+  }
+  fd->show_frame = cm->show_frame;
+  fd->frame_type = cm->frame_type;
+  fd->base_qindex = cm->base_qindex;
+  fd->tile_mi_cols = cm->tile_width;
+  fd->tile_mi_rows = cm->tile_height;
+#if CONFIG_ACCOUNTING
+  fd->accounting = &pbi->accounting;
+#endif
+#if CONFIG_CDEF
+// TODO(negge): copy per frame CDEF data
+#endif
+  int i, j;
+  for (i = 0; i < MAX_SEGMENTS; i++) {
+    for (j = 0; j < 2; j++) {
+      fd->y_dequant[i][j] = cm->y_dequant[i][j];
+      fd->uv_dequant[i][j] = cm->uv_dequant[i][j];
+    }
+  }
+  for (j = 0; j < cm->mi_rows; j++) {
+    for (i = 0; i < cm->mi_cols; i++) {
+      const MB_MODE_INFO *mbmi =
+          &cm->mi_grid_visible[j * cm->mi_stride + i]->mbmi;
+      insp_mi_data *mi = &fd->mi_grid[j * cm->mi_cols + i];
+      // Segment
+      mi->segment_id = mbmi->segment_id;
+      // Motion Vectors
+      mi->mv[0].row = mbmi->mv[0].as_mv.row;
+      mi->mv[0].col = mbmi->mv[0].as_mv.col;
+      mi->mv[1].row = mbmi->mv[1].as_mv.row;
+      mi->mv[1].col = mbmi->mv[1].as_mv.col;
+      // Reference Frames
+      mi->ref_frame[0] = mbmi->ref_frame[0];
+      mi->ref_frame[1] = mbmi->ref_frame[1];
+      // Prediction Mode
+      mi->mode = mbmi->mode;
+      // Prediction Mode for Chromatic planes
+      if (mi->mode < INTRA_MODES) {
+        mi->uv_mode = mbmi->uv_mode;
+      } else {
+        mi->uv_mode = INTRA_INVALID;
+      }
+      // Block Size
+      mi->sb_type = mbmi->sb_type;
+      // Skip Flag
+      mi->skip = mbmi->skip;
+#if CONFIG_DUAL_FILTER
+      mi->filter[0] = mbmi->interp_filter[0];
+      mi->filter[1] = mbmi->interp_filter[1];
+#else
+      mi->filter = mbmi->interp_filter;
+#endif
+      // Transform
+      mi->tx_type = mbmi->tx_type;
+      mi->tx_size = mbmi->tx_size;
+
+#if CONFIG_CDEF
+      mi->cdef_level = cm->cdef_strengths[mbmi->cdef_strength] / CLPF_STRENGTHS;
+      mi->cdef_strength =
+          cm->cdef_strengths[mbmi->cdef_strength] % CLPF_STRENGTHS;
+      mi->cdef_strength += mi->cdef_strength == 3;
+#endif
+    }
+  }
+  return 1;
+}
diff --git a/third_party/aom/av1/decoder/inspection.h b/third_party/aom/av1/decoder/inspection.h
new file mode 100644
index 000000000..d6cf4319a
--- /dev/null
+++ b/third_party/aom/av1/decoder/inspection.h
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_INSPECTION_H_
+#define AOM_INSPECTION_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+
+typedef void (*aom_inspect_cb)(void *decoder, void *data);
+
+typedef struct insp_mv insp_mv;
+
+struct insp_mv {
+  int16_t row;
+  int16_t col;
+};
+
+typedef struct insp_mi_data insp_mi_data;
+
+struct insp_mi_data {
+  insp_mv mv[2];
+  int8_t ref_frame[2];
+  int8_t mode;
+  int8_t uv_mode;
+  int8_t sb_type;
+  int8_t skip;
+  int8_t segment_id;
+#if CONFIG_DUAL_FILTER
+  int8_t filter[2];
+#else
+  int8_t filter;
+#endif
+  int8_t tx_type;
+  int8_t tx_size;
+#if CONFIG_CDEF
+  int8_t cdef_level;
+  int8_t cdef_strength;
+#endif
+};
+
+typedef struct insp_frame_data insp_frame_data;
+
+struct insp_frame_data {
+#if CONFIG_ACCOUNTING
+  Accounting *accounting;
+#endif
+  insp_mi_data *mi_grid;
+  int show_frame;
+  int frame_type;
+  int base_qindex;
+  int mi_rows;
+  int mi_cols;
+  int tile_mi_rows;
+  int tile_mi_cols;
+  int16_t y_dequant[MAX_SEGMENTS][2];
+  int16_t uv_dequant[MAX_SEGMENTS][2];
+#if CONFIG_CDEF
+// TODO(negge): add per frame CDEF data
+#endif
+};
+
+void ifd_init(insp_frame_data *fd, int frame_width, int frame_height);
+void ifd_clear(insp_frame_data *fd);
+int ifd_inspect(insp_frame_data *fd, void *decoder);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_INSPECTION_H_
diff --git a/third_party/aom/av1/decoder/laplace_decoder.c b/third_party/aom/av1/decoder/laplace_decoder.c
new file mode 100644
index 000000000..b6cf50bc7
--- /dev/null
+++ b/third_party/aom/av1/decoder/laplace_decoder.c
@@ -0,0 +1,121 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+
+#include "aom_dsp/bitreader.h"
+#include "av1/common/pvq.h"
+#include "pvq_decoder.h"
+
+#define aom_decode_pvq_split(r, adapt, sum, ctx, ACCT_STR_NAME) \
+  aom_decode_pvq_split_(r, adapt, sum, ctx ACCT_STR_ARG(ACCT_STR_NAME))
+
+static int aom_decode_pvq_split_(aom_reader *r, od_pvq_codeword_ctx *adapt,
+ int sum, int ctx ACCT_STR_PARAM) {
+  int shift;
+  int count;
+  int msbs;
+  int fctx;
+  count = 0;
+  if (sum == 0) return 0;
+  shift = OD_MAXI(0, OD_ILOG(sum) - 3);
+  fctx = 7*ctx + (sum >> shift) - 1;
+  msbs = aom_read_symbol_pvq(r, adapt->pvq_split_cdf[fctx], (sum >> shift) + 1,
+      ACCT_STR_NAME);
+  if (shift) count = aom_read_literal(r, shift, ACCT_STR_NAME);
+  count += msbs << shift;
+  if (count > sum) {
+    count = sum;
+#if CONFIG_DAALA_EC
+    r->ec.error = 1;
+#else
+# error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  }
+  return count;
+}
+
+void aom_decode_band_pvq_splits(aom_reader *r, od_pvq_codeword_ctx *adapt,
+ od_coeff *y, int n, int k, int level) {
+  int mid;
+  int count_right;
+  if (n == 1) {
+    y[0] = k;
+  }
+  else if (k == 0) {
+    OD_CLEAR(y, n);
+  }
+  else if (k == 1 && n <= 16) {
+    int cdf_id;
+    int pos;
+    cdf_id = od_pvq_k1_ctx(n, level == 0);
+    OD_CLEAR(y, n);
+    pos = aom_read_symbol_pvq(r, adapt->pvq_k1_cdf[cdf_id], n, "pvq:k1");
+    y[pos] = 1;
+  }
+  else {
+    mid = n >> 1;
+    count_right = aom_decode_pvq_split(r, adapt, k, od_pvq_size_ctx(n),
+     "pvq:split");
+    aom_decode_band_pvq_splits(r, adapt, y, mid, k - count_right, level + 1);
+    aom_decode_band_pvq_splits(r, adapt, y + mid, n - mid, count_right,
+     level + 1);
+  }
+}
+
+/** Decodes the tail of a Laplace-distributed variable, i.e. it doesn't
+ * do anything special for the zero case.
+ *
+ * @param [dec] range decoder
+ * @param [decay] decay factor of the distribution, i.e. pdf ~= decay^x
+ *
+ * @retval decoded variable x
+ */
+int aom_laplace_decode_special_(aom_reader *r, unsigned decay ACCT_STR_PARAM) {
+  int pos;
+  int shift;
+  int xs;
+  int sym;
+  const uint16_t *cdf;
+  shift = 0;
+  /* We don't want a large decay value because that would require too many
+     symbols. */
+  while (decay > 235) {
+    decay = (decay*decay + 128) >> 8;
+    shift++;
+  }
+  decay = OD_MINI(decay, 254);
+  decay = OD_MAXI(decay, 2);
+  cdf = EXP_CDF_TABLE[(decay + 1) >> 1];
+  OD_LOG((OD_LOG_PVQ, OD_LOG_DEBUG, "decay = %d\n", decay));
+  xs = 0;
+  do {
+    sym = OD_MINI(xs, 15);
+    {
+      int i;
+      OD_LOG((OD_LOG_PVQ, OD_LOG_DEBUG, "%d %d %d", xs, shift, sym));
+      for (i = 0; i < 16; i++) {
+        OD_LOG_PARTIAL((OD_LOG_PVQ, OD_LOG_DEBUG, "%d ", cdf[i]));
+      }
+      OD_LOG_PARTIAL((OD_LOG_PVQ, OD_LOG_DEBUG, "\n"));
+    }
+    sym = aom_read_cdf(r, cdf, 16, ACCT_STR_NAME);
+    xs += sym;
+  } while (sym >= 15);
+  if (shift) pos = (xs << shift) + aom_read_literal(r, shift, ACCT_STR_NAME);
+  else pos = xs;
+  return pos;
+}
diff --git a/third_party/aom/av1/decoder/pvq_decoder.c b/third_party/aom/av1/decoder/pvq_decoder.c
new file mode 100644
index 000000000..d9a8e8056
--- /dev/null
+++ b/third_party/aom/av1/decoder/pvq_decoder.c
@@ -0,0 +1,378 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+#include <stdlib.h>
+#include "./aom_config.h"
+#include "aom_dsp/bitreader.h"
+#include "aom_dsp/entcode.h"
+#include "aom_dsp/entdec.h"
+#include "av1/common/odintrin.h"
+#include "av1/common/partition.h"
+#include "av1/common/pvq_state.h"
+#include "av1/decoder/decint.h"
+#include "av1/decoder/pvq_decoder.h"
+#include "aom_ports/system_state.h"
+
+int aom_read_symbol_pvq_(aom_reader *r, aom_cdf_prob *cdf, int nsymbs
+ ACCT_STR_PARAM) {
+  if (cdf[0] == 0)
+    aom_cdf_init_q15_1D(cdf, nsymbs, CDF_SIZE(nsymbs));
+  return aom_read_symbol(r, cdf, nsymbs, ACCT_STR_NAME);
+}
+
+static void aom_decode_pvq_codeword(aom_reader *r, od_pvq_codeword_ctx *ctx,
+ od_coeff *y, int n, int k) {
+  int i;
+  aom_decode_band_pvq_splits(r, ctx, y, n, k, 0);
+  for (i = 0; i < n; i++) {
+    if (y[i] && aom_read_bit(r, "pvq:sign")) y[i] = -y[i];
+  }
+}
+
+/** Inverse of neg_interleave; decodes the interleaved gain.
+ *
+ * @param [in]      x      quantized/interleaved gain to decode
+ * @param [in]      ref    quantized gain of the reference
+ * @return                 original quantized gain value
+ */
+static int neg_deinterleave(int x, int ref) {
+  if (x < 2*ref-1) {
+    if (x & 1) return ref - 1 - (x >> 1);
+    else return ref + (x >> 1);
+  }
+  else return x+1;
+}
+
+/** Synthesizes one parition of coefficient values from a PVQ-encoded
+ * vector.
+ *
+ * @param [out]     xcoeff  output coefficient partition (x in math doc)
+ * @param [in]      ypulse  PVQ-encoded values (y in math doc); in the noref
+ *                          case, this vector has n entries, in the
+ *                          reference case it contains n-1 entries
+ *                          (the m-th entry is not included)
+ * @param [in]      ref     reference vector (prediction)
+ * @param [in]      n       number of elements in this partition
+ * @param [in]      gr      gain of the reference vector (prediction)
+ * @param [in]      noref   indicates presence or lack of prediction
+ * @param [in]      g       decoded quantized vector gain
+ * @param [in]      theta   decoded theta (prediction error)
+ * @param [in]      qm      QM with magnitude compensation
+ * @param [in]      qm_inv  Inverse of QM with magnitude compensation
+ */
+static void pvq_synthesis(od_coeff *xcoeff, od_coeff *ypulse, od_val16 *r16,
+ int n, od_val32 gr, int noref, od_val32 g, od_val32 theta, const int16_t *qm_inv,
+ int shift) {
+  int s;
+  int m;
+  /* Sign of the Householder reflection vector */
+  s = 0;
+  /* Direction of the Householder reflection vector */
+  m = noref ? 0 : od_compute_householder(r16, n, gr, &s, shift);
+  od_pvq_synthesis_partial(xcoeff, ypulse, r16, n, noref, g, theta, m, s,
+   qm_inv);
+}
+
+typedef struct {
+  od_coeff *ref;
+  int nb_coeffs;
+  int allow_flip;
+} cfl_ctx;
+
+/** Decodes a single vector of integers (eg, a partition within a
+ *  coefficient block) encoded using PVQ
+ *
+ * @param [in,out] ec          range encoder
+ * @param [in]     q0          scale/quantizer
+ * @param [in]     n           number of coefficients in partition
+ * @param [in,out] model       entropy decoder state
+ * @param [in,out] adapt       adaptation context
+ * @param [in,out] exg         ExQ16 expectation of decoded gain value
+ * @param [in,out] ext         ExQ16 expectation of decoded theta value
+ * @param [in]     ref         'reference' (prediction) vector
+ * @param [out]    out         decoded partition
+ * @param [out]    noref       boolean indicating absence of reference
+ * @param [in]     beta        per-band activity masking beta param
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [in]     pli         plane index
+ * @param [in]     cdf_ctx     selects which cdf context to use
+ * @param [in,out] skip_rest   whether to skip further bands in each direction
+ * @param [in]     band        index of the band being decoded
+ * @param [in]     band        index of the band being decoded
+ * @param [out]    skip        skip flag with range [0,1]
+ * @param [in]     qm          QM with magnitude compensation
+ * @param [in]     qm_inv      Inverse of QM with magnitude compensation
+ */
+static void pvq_decode_partition(aom_reader *r,
+                                 int q0,
+                                 int n,
+                                 generic_encoder model[3],
+                                 od_adapt_ctx *adapt,
+                                 int *exg,
+                                 int *ext,
+                                 od_coeff *ref,
+                                 od_coeff *out,
+                                 int *noref,
+                                 od_val16 beta,
+                                 int is_keyframe,
+                                 int pli,
+                                 int cdf_ctx,
+                                 cfl_ctx *cfl,
+                                 int has_skip,
+                                 int *skip_rest,
+                                 int band,
+                                 int *skip,
+                                 const int16_t *qm,
+                                 const int16_t *qm_inv) {
+  int k;
+  od_val32 qcg;
+  int itheta;
+  od_val32 theta;
+  od_val32 gr;
+  od_val32 gain_offset;
+  od_coeff y[MAXN];
+  int qg;
+  int id;
+  int i;
+  od_val16 ref16[MAXN];
+  int rshift;
+  theta = 0;
+  gr = 0;
+  gain_offset = 0;
+  /* Skip is per-direction. For band=0, we can use any of the flags. */
+  if (skip_rest[(band + 2) % 3]) {
+    qg = 0;
+    if (is_keyframe) {
+      itheta = -1;
+      *noref = 1;
+    }
+    else {
+      itheta = 0;
+      *noref = 0;
+    }
+  }
+  else {
+    /* Jointly decode gain, itheta and noref for small values. Then we handle
+       larger gain. */
+    id = aom_read_symbol_pvq(r, &adapt->pvq.pvq_gaintheta_cdf[cdf_ctx][0],
+     8 + 7*has_skip, "pvq:gaintheta");
+    if (!is_keyframe && id >= 10) id++;
+    if (is_keyframe && id >= 8) id++;
+    if (id >= 8) {
+      id -= 8;
+      skip_rest[0] = skip_rest[1] = skip_rest[2] = 1;
+    }
+    qg = id & 1;
+    itheta = (id >> 1) - 1;
+    *noref = (itheta == -1);
+  }
+  /* The CfL flip bit is only decoded on the first band that has noref=0. */
+  if (cfl->allow_flip && !*noref) {
+    int flip;
+    flip = aom_read_bit(r, "cfl:flip");
+    if (flip) {
+      for (i = 0; i < cfl->nb_coeffs; i++) cfl->ref[i] = -cfl->ref[i];
+    }
+    cfl->allow_flip = 0;
+  }
+  if (qg > 0) {
+    int tmp;
+    tmp = *exg;
+    qg = 1 + generic_decode(r, &model[!*noref], &tmp, 2, "pvq:gain");
+    OD_IIR_DIADIC(*exg, qg << 16, 2);
+  }
+  *skip = 0;
+#if defined(OD_FLOAT_PVQ)
+  rshift = 0;
+#else
+  /* Shift needed to make the reference fit in 15 bits, so that the Householder
+     vector can fit in 16 bits. */
+  rshift = OD_MAXI(0, od_vector_log_mag(ref, n) - 14);
+#endif
+  for (i = 0; i < n; i++) {
+#if defined(OD_FLOAT_PVQ)
+    ref16[i] = ref[i]*(double)qm[i]*OD_QM_SCALE_1;
+#else
+    ref16[i] = OD_SHR_ROUND(ref[i]*qm[i], OD_QM_SHIFT + rshift);
+#endif
+  }
+  if(!*noref){
+    /* we have a reference; compute its gain */
+    od_val32 cgr;
+    int icgr;
+    int cfl_enabled;
+    cfl_enabled = pli != 0 && is_keyframe && !OD_DISABLE_CFL;
+    cgr = od_pvq_compute_gain(ref16, n, q0, &gr, beta, rshift);
+    if (cfl_enabled) cgr = OD_CGAIN_SCALE;
+#if defined(OD_FLOAT_PVQ)
+    icgr = (int)floor(.5 + cgr);
+#else
+    icgr = OD_SHR_ROUND(cgr, OD_CGAIN_SHIFT);
+#endif
+    /* quantized gain is interleave encoded when there's a reference;
+       deinterleave it now */
+    if (is_keyframe) qg = neg_deinterleave(qg, icgr);
+    else {
+      qg = neg_deinterleave(qg, icgr + 1) - 1;
+      if (qg == 0) *skip = (icgr ? OD_PVQ_SKIP_ZERO : OD_PVQ_SKIP_COPY);
+    }
+    if (qg == icgr && itheta == 0 && !cfl_enabled) *skip = OD_PVQ_SKIP_COPY;
+    gain_offset = cgr - OD_SHL(icgr, OD_CGAIN_SHIFT);
+    qcg = OD_SHL(qg, OD_CGAIN_SHIFT) + gain_offset;
+    /* read and decode first-stage PVQ error theta */
+    if (itheta > 1) {
+      int tmp;
+      tmp = *ext;
+      itheta = 2 + generic_decode(r, &model[2], &tmp, 2, "pvq:theta");
+      OD_IIR_DIADIC(*ext, itheta << 16, 2);
+    }
+    theta = od_pvq_compute_theta(itheta, od_pvq_compute_max_theta(qcg, beta));
+  }
+  else{
+    itheta = 0;
+    if (!is_keyframe) qg++;
+    qcg = OD_SHL(qg, OD_CGAIN_SHIFT);
+    if (qg == 0) *skip = OD_PVQ_SKIP_ZERO;
+  }
+
+  k = od_pvq_compute_k(qcg, itheta, *noref, n, beta);
+  if (k != 0) {
+    /* when noref==0, y is actually size n-1 */
+    aom_decode_pvq_codeword(r, &adapt->pvq.pvq_codeword_ctx, y,
+     n - !*noref, k);
+  }
+  else {
+    OD_CLEAR(y, n);
+  }
+  if (*skip) {
+    if (*skip == OD_PVQ_SKIP_COPY) OD_COPY(out, ref, n);
+    else OD_CLEAR(out, n);
+  }
+  else {
+    od_val32 g;
+    g = od_gain_expand(qcg, q0, beta);
+    pvq_synthesis(out, y, ref16, n, gr, *noref, g, theta, qm_inv, rshift);
+  }
+  /* If OD_PVQ_SKIP_ZERO or OD_PVQ_SKIP_COPY, set skip to 1 for visualization */
+  if (*skip) *skip = 1;
+}
+
+/** Decodes a coefficient block (except for DC) encoded using PVQ
+ *
+ * @param [in,out] dec         daala decoder context
+ * @param [in]     ref         'reference' (prediction) vector
+ * @param [out]    out         decoded partition
+ * @param [in]     q0          quantizer
+ * @param [in]     pli         plane index
+ * @param [in]     bs          log of the block size minus two
+ * @param [in]     beta        per-band activity masking beta param
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [out]    flags       bitmask of the per band skip and noref flags
+ * @param [in]     ac_dc_coded skip flag for the block (range 0-3)
+ * @param [in]     qm          QM with magnitude compensation
+ * @param [in]     qm_inv      Inverse of QM with magnitude compensation
+ */
+void od_pvq_decode(daala_dec_ctx *dec,
+                   od_coeff *ref,
+                   od_coeff *out,
+                   int q0,
+                   int pli,
+                   int bs,
+                   const od_val16 *beta,
+                   int is_keyframe,
+                   unsigned int *flags,
+                   PVQ_SKIP_TYPE ac_dc_coded,
+                   const int16_t *qm,
+                   const int16_t *qm_inv){
+
+  int noref[PVQ_MAX_PARTITIONS];
+  int skip[PVQ_MAX_PARTITIONS];
+  int *exg;
+  int *ext;
+  int nb_bands;
+  int i;
+  const int *off;
+  int size[PVQ_MAX_PARTITIONS];
+  generic_encoder *model;
+  int skip_rest[3] = {0};
+  cfl_ctx cfl;
+  const unsigned char *pvq_qm;
+  int use_masking;
+
+  aom_clear_system_state();
+
+  /*Default to skip=1 and noref=0 for all bands.*/
+  for (i = 0; i < PVQ_MAX_PARTITIONS; i++) {
+    noref[i] = 0;
+    skip[i] = 1;
+  }
+
+  use_masking = dec->use_activity_masking;
+
+  if (use_masking)
+    pvq_qm = &dec->state.pvq_qm_q4[pli][0];
+  else
+    pvq_qm = 0;
+
+  exg = &dec->state.adapt->pvq.pvq_exg[pli][bs][0];
+  ext = dec->state.adapt->pvq.pvq_ext + bs*PVQ_MAX_PARTITIONS;
+  model = dec->state.adapt->pvq.pvq_param_model;
+  nb_bands = OD_BAND_OFFSETS[bs][0];
+  off = &OD_BAND_OFFSETS[bs][1];
+  out[0] = ac_dc_coded & DC_CODED;
+  if (ac_dc_coded < AC_CODED) {
+    if (is_keyframe) for (i = 1; i < 1 << (2*bs + 4); i++) out[i] = 0;
+    else for (i = 1; i < 1 << (2*bs + 4); i++) out[i] = ref[i];
+  }
+  else {
+    for (i = 0; i < nb_bands; i++) size[i] = off[i+1] - off[i];
+    cfl.ref = ref;
+    cfl.nb_coeffs = off[nb_bands];
+    cfl.allow_flip = pli != 0 && is_keyframe;
+    for (i = 0; i < nb_bands; i++) {
+      int q;
+
+      if (use_masking)
+        q = OD_MAXI(1, q0 * pvq_qm[od_qm_get_index(bs, i + 1)] >> 4);
+      else
+        q = OD_MAXI(1, q0);
+
+      pvq_decode_partition(dec->r, q, size[i],
+       model, dec->state.adapt, exg + i, ext + i, ref + off[i], out + off[i],
+       &noref[i], beta[i], is_keyframe, pli,
+       (pli != 0)*OD_TXSIZES*PVQ_MAX_PARTITIONS + bs*PVQ_MAX_PARTITIONS + i,
+       &cfl, i == 0 && (i < nb_bands - 1), skip_rest, i, &skip[i],
+       qm + off[i], qm_inv + off[i]);
+      if (i == 0 && !skip_rest[0] && bs > 0) {
+        int skip_dir;
+        int j;
+        skip_dir = aom_read_symbol(dec->r,
+         &dec->state.adapt->pvq.pvq_skip_dir_cdf[(pli != 0) + 2*(bs - 1)][0], 7,
+         "pvq:skiprest");
+        for (j = 0; j < 3; j++) skip_rest[j] = !!(skip_dir & (1 << j));
+      }
+    }
+  }
+  *flags = 0;
+  for (i = nb_bands - 1; i >= 0; i--) {
+    *flags <<= 1;
+    *flags |= noref[i]&1;
+    *flags <<= 1;
+    *flags |= skip[i]&1;
+  }
+}
diff --git a/third_party/aom/av1/decoder/pvq_decoder.h b/third_party/aom/av1/decoder/pvq_decoder.h
new file mode 100644
index 000000000..98970663b
--- /dev/null
+++ b/third_party/aom/av1/decoder/pvq_decoder.h
@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_pvq_decoder_H)
+# define _pvq_decoder_H (1)
+# include "aom_dsp/bitreader.h"
+# include "aom_dsp/entdec.h"
+# include "av1/common/pvq.h"
+# include "av1/decoder/decint.h"
+
+#define aom_read_symbol_pvq(r, cdf, nsymbs, ACCT_STR_NAME) \
+  aom_read_symbol_pvq_(r, cdf, nsymbs ACCT_STR_ARG(ACCT_STR_NAME))
+
+int aom_read_symbol_pvq_(aom_reader *r, aom_cdf_prob *cdf, int nsymbs
+  ACCT_STR_PARAM);
+
+void aom_decode_band_pvq_splits(aom_reader *r, od_pvq_codeword_ctx *adapt,
+ od_coeff *y, int n, int k, int level);
+
+#define aom_laplace_decode_special(r, decay, ACCT_STR_NAME) \
+  aom_laplace_decode_special_(r, decay ACCT_STR_ARG(ACCT_STR_NAME))
+
+int aom_laplace_decode_special_(aom_reader *r, unsigned decay ACCT_STR_PARAM);
+
+void od_pvq_decode(daala_dec_ctx *dec, od_coeff *ref, od_coeff *out, int q0,
+    int pli, int bs, const od_val16 *beta, int is_keyframe,
+    unsigned int *flags, PVQ_SKIP_TYPE ac_dc_coded, const int16_t *qm,
+    const int16_t *qm_inv);
+
+#endif