diff --git a/src/dec/idec.c b/src/dec/idec.c
index 765a59b3..392f5d70 100644
--- a/src/dec/idec.c
+++ b/src/dec/idec.c
@@ -72,15 +72,14 @@ struct WebPIDecoder {
   MemBuffer mem_;          // input memory buffer.
   WebPDecBuffer output_;   // output buffer (when no external one is supplied)
   size_t chunk_size_;      // Compressed VP8/VP8L size extracted from Header.
+
+  int last_mb_y_;          // last row reached for intra-mode decoding
 };
 
 // MB context to restore in case VP8DecodeMB() fails
 typedef struct {
   VP8MB left_;
   VP8MB info_;
-  uint8_t intra_t_[4];
-  uint8_t intra_l_[4];
-  VP8BitReader br_;
   VP8BitReader token_br_;
 } MBContext;
 
@@ -264,30 +263,16 @@ static VP8StatusCode FinishDecoding(WebPIDecoder* const idec) {
 
 static void SaveContext(const VP8Decoder* dec, const VP8BitReader* token_br,
                         MBContext* const context) {
-  const VP8BitReader* const br = &dec->br_;
-  const VP8MB* const left = dec->mb_info_ - 1;
-  const VP8MB* const info = dec->mb_info_ + dec->mb_x_;
-
-  context->left_ = *left;
-  context->info_ = *info;
-  context->br_ = *br;
+  context->left_ = dec->mb_info_[-1];
+  context->info_ = dec->mb_info_[dec->mb_x_];
   context->token_br_ = *token_br;
-  memcpy(context->intra_t_, dec->intra_t_ + 4 * dec->mb_x_, 4);
-  memcpy(context->intra_l_, dec->intra_l_, 4);
 }
 
 static void RestoreContext(const MBContext* context, VP8Decoder* const dec,
                            VP8BitReader* const token_br) {
-  VP8BitReader* const br = &dec->br_;
-  VP8MB* const left = dec->mb_info_ - 1;
-  VP8MB* const info = dec->mb_info_ + dec->mb_x_;
-
-  *left = context->left_;
-  *info = context->info_;
-  *br = context->br_;
+  dec->mb_info_[-1] = context->left_;
+  dec->mb_info_[dec->mb_x_] = context->info_;
   *token_br = context->token_br_;
-  memcpy(dec->intra_t_ + 4 * dec->mb_x_, context->intra_t_, 4);
-  memcpy(dec->intra_l_, context->intra_l_, 4);
 }
 
 //------------------------------------------------------------------------------
@@ -459,16 +444,26 @@ static VP8StatusCode DecodeRemaining(WebPIDecoder* const idec) {
 
   assert(dec->ready_);
   for (; dec->mb_y_ < dec->mb_h_; ++dec->mb_y_) {
-    VP8BitReader* token_br = &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
+    if (idec->last_mb_y_ != dec->mb_y_) {
+      if (!VP8ParseIntraModeRow(&dec->br_, dec)) {
+        // note: normally, error shouldn't occur since we already have the whole
+        // partition0 available here in DecodeRemaining(). Reaching EOF while
+        // reading intra modes really means a BITSTREAM_ERROR.
+        return IDecError(idec, VP8_STATUS_BITSTREAM_ERROR);
+      }
+      idec->last_mb_y_ = dec->mb_y_;
+    }
     for (; dec->mb_x_ < dec->mb_w_; ++dec->mb_x_) {
+      VP8BitReader* const token_br =
+          &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
       MBContext context;
       SaveContext(dec, token_br, &context);
       if (!VP8DecodeMB(dec, token_br)) {
-        RestoreContext(&context, dec, token_br);
         // We shouldn't fail when MAX_MB data was available
         if (dec->num_parts_ == 1 && MemDataSize(&idec->mem_) > MAX_MB_SIZE) {
           return IDecError(idec, VP8_STATUS_BITSTREAM_ERROR);
         }
+        RestoreContext(&context, dec, token_br);
         return VP8_STATUS_SUSPENDED;
       }
       // Release buffer only if there is only one partition
@@ -585,6 +580,8 @@ WebPIDecoder* WebPINewDecoder(WebPDecBuffer* output_buffer) {
   idec->state_ = STATE_WEBP_HEADER;
   idec->chunk_size_ = 0;
 
+  idec->last_mb_y_ = -1;
+
   InitMemBuffer(&idec->mem_);
   WebPInitDecBuffer(&idec->output_);
   VP8InitIo(&idec->io_);
diff --git a/src/dec/tree.c b/src/dec/tree.c
index bf9b7c55..22d83ee5 100644
--- a/src/dec/tree.c
+++ b/src/dec/tree.c
@@ -278,10 +278,23 @@ void VP8ResetProba(VP8Proba* const proba) {
   // proba->bands_[][] is initialized later
 }
 
-void VP8ParseIntraMode(VP8BitReader* const br, VP8Decoder* const dec) {
-  uint8_t* const top = dec->intra_t_ + 4 * dec->mb_x_;
+static void ParseIntraMode(VP8BitReader* const br,
+                           VP8Decoder* const dec, int mb_x) {
+  uint8_t* const top = dec->intra_t_ + 4 * mb_x;
   uint8_t* const left = dec->intra_l_;
-  VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
+  VP8MBData* const block = dec->mb_data_ + mb_x;
+
+  // Note: we don't save segment map (yet), as we don't expect
+  // to decode more than 1 keyframe.
+  if (dec->segment_hdr_.update_map_) {
+    // Hardcoded tree parsing
+    block->segment_ = !VP8GetBit(br, dec->proba_.segments_[0])
+                    ? VP8GetBit(br, dec->proba_.segments_[1])
+                    : 2 + VP8GetBit(br, dec->proba_.segments_[2]);
+  } else {
+    block->segment_ = 0;  // default for intra
+  }
+  if (dec->use_skip_proba_) block->skip_ = VP8GetBit(br, dec->skip_p_);
 
   block->is_i4x4_ = !VP8GetBit(br, 145);   // decide for B_PRED first
   if (!block->is_i4x4_) {
@@ -332,6 +345,14 @@ void VP8ParseIntraMode(VP8BitReader* const br, VP8Decoder* const dec) {
                  : VP8GetBit(br, 183) ? TM_PRED : H_PRED;
 }
 
+int VP8ParseIntraModeRow(VP8BitReader* const br, VP8Decoder* const dec) {
+  int mb_x;
+  for (mb_x = 0; mb_x < dec->mb_w_; ++mb_x) {
+    ParseIntraMode(br, dec, mb_x);
+  }
+  return !dec->br_.eof_;
+}
+
 //------------------------------------------------------------------------------
 // Paragraph 13
 
diff --git a/src/dec/vp8.c b/src/dec/vp8.c
index bfd0e8f9..e079f2f9 100644
--- a/src/dec/vp8.c
+++ b/src/dec/vp8.c
@@ -317,7 +317,6 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
 
     VP8ResetProba(&dec->proba_);
     ResetSegmentHeader(&dec->segment_hdr_);
-    dec->segment_ = 0;    // default for intra
   }
 
   // Check if we have all the partition #0 available, and initialize dec->br_
@@ -479,8 +478,8 @@ static int ParseResiduals(VP8Decoder* const dec,
                           VP8MB* const mb, VP8BitReader* const token_br) {
   VP8BandProbas (* const bands)[NUM_BANDS] = dec->proba_.bands_;
   const VP8BandProbas* ac_proba;
-  const VP8QuantMatrix* const q = &dec->dqm_[dec->segment_];
   VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
+  const VP8QuantMatrix* const q = &dec->dqm_[block->segment_];
   int16_t* dst = block->coeffs_;
   VP8MB* const left_mb = dec->mb_info_ - 1;
   uint8_t tnz, lnz;
@@ -570,26 +569,10 @@ static int ParseResiduals(VP8Decoder* const dec,
 // Main loop
 
 int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br) {
-  VP8BitReader* const br = &dec->br_;
   VP8MB* const left = dec->mb_info_ - 1;
   VP8MB* const mb = dec->mb_info_ + dec->mb_x_;
   VP8MBData* const block = dec->mb_data_ + dec->mb_x_;
-  int skip;
-
-  // Note: we don't save segment map (yet), as we don't expect
-  // to decode more than 1 keyframe.
-  if (dec->segment_hdr_.update_map_) {
-    // Hardcoded tree parsing
-    dec->segment_ = !VP8GetBit(br, dec->proba_.segments_[0]) ?
-        VP8GetBit(br, dec->proba_.segments_[1]) :
-        2 + VP8GetBit(br, dec->proba_.segments_[2]);
-  }
-  skip = dec->use_skip_proba_ ? VP8GetBit(br, dec->skip_p_) : 0;
-
-  VP8ParseIntraMode(br, dec);
-  if (br->eof_) {
-    return 0;
-  }
+  int skip = dec->use_skip_proba_ ? block->skip_ : 0;
 
   if (!skip) {
     skip = ParseResiduals(dec, mb, token_br);
@@ -604,7 +587,7 @@ int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br) {
 
   if (dec->filter_type_ > 0) {  // store filter info
     VP8FInfo* const finfo = dec->f_info_ + dec->mb_x_;
-    *finfo = dec->fstrengths_[dec->segment_][block->is_i4x4_];
+    *finfo = dec->fstrengths_[block->segment_][block->is_i4x4_];
     finfo->f_inner_ |= !skip;
   }
 
@@ -624,6 +607,10 @@ static int ParseFrame(VP8Decoder* const dec, VP8Io* io) {
     // Parse bitstream for this row.
     VP8BitReader* const token_br =
         &dec->parts_[dec->mb_y_ & (dec->num_parts_ - 1)];
+    if (!VP8ParseIntraModeRow(&dec->br_, dec)) {
+      return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
+                         "Premature end-of-partition0 encountered.");
+    }
     for (; dec->mb_x_ < dec->mb_w_; ++dec->mb_x_) {
       if (!VP8DecodeMB(dec, token_br)) {
         return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
diff --git a/src/dec/vp8i.h b/src/dec/vp8i.h
index 3f4cf297..1efe3f5c 100644
--- a/src/dec/vp8i.h
+++ b/src/dec/vp8i.h
@@ -195,6 +195,8 @@ typedef struct {
   uint32_t non_zero_y_;
   uint32_t non_zero_uv_;
   uint8_t dither_;      // local dithering strength (deduced from non_zero_*)
+  uint8_t skip_;
+  uint8_t segment_;
 } VP8MBData;
 
 // Persistent information needed by the parallel processing
@@ -265,7 +267,6 @@ struct VP8Decoder {
   uint8_t* intra_t_;      // top intra modes values: 4 * mb_w_
   uint8_t  intra_l_[4];   // left intra modes values
 
-  uint8_t segment_;       // segment of the currently parsed block
   VP8TopSamples* yuv_t_;  // top y/u/v samples
 
   VP8MB* mb_info_;        // contextual macroblock info (mb_w_ + 1)
@@ -313,7 +314,8 @@ int VP8SetError(VP8Decoder* const dec,
 // in tree.c
 void VP8ResetProba(VP8Proba* const proba);
 void VP8ParseProba(VP8BitReader* const br, VP8Decoder* const dec);
-void VP8ParseIntraMode(VP8BitReader* const br,  VP8Decoder* const dec);
+// parses one row of intra mode data in partition 0, returns !eof
+int VP8ParseIntraModeRow(VP8BitReader* const br, VP8Decoder* const dec);
 
 // in quant.c
 void VP8ParseQuant(VP8Decoder* const dec);