Sync-patch with libwebp (ChangeId: Ia5475247)android-4.3_r3.1android-4.3_r3android-4.3_r2.3android-4.3_r2.2android-4.3_r2.1android-4.3_r2android-4.3_r1.1android-4.3_r1android-4.3_r0.9.1android-4.3_r0.9android-4.3.1_r1tools_r22.2jb-mr2.0.0-releasejb-mr2.0-releasejb-mr2-releasejb-mr2-dev

Added 16bit swapping of RGB565 / RGB4444 colorspace.
Added ARM/NEON code for decoder/encoder modules.
Speedup in WebP compression (method 3 and above).

Change-Id: I95a697338bef7c3ea08054eb5f850a97d1889eb9

11 files changed, 313 insertions, 141 deletions

diff --git a/src/utils/bit_reader.c b/src/utils/bit_reader.c
index 1afb1db8..d6cfd864 100644
--- a/src/utils/bit_reader.c
+++ b/src/utils/bit_reader.c
@@ -15,7 +15,11 @@
 extern "C" {
 #endif
 
-#define MK(X) (((bit_t)(X) << (BITS)) | (MASK))
+#ifndef USE_RIGHT_JUSTIFY
+#define MK(X) (((range_t)(X) << (BITS)) | (MASK))
+#else
+#define MK(X) ((range_t)(X))
+#endif
 
 //------------------------------------------------------------------------------
 // VP8BitReader
@@ -29,7 +33,7 @@ void VP8InitBitReader(VP8BitReader* const br,
   br->buf_     = start;
   br->buf_end_ = end;
   br->value_   = 0;
-  br->missing_ = 8;   // to load the very first 8bits
+  br->bits_    = -8;   // to load the very first 8bits
   br->eof_     = 0;
 }
 
@@ -46,7 +50,7 @@ const uint8_t kVP8Log2Range[128] = {
 };
 
 // range = (range << kVP8Log2Range[range]) + trailing 1's
-const bit_t kVP8NewRange[128] = {
+const range_t kVP8NewRange[128] = {
   MK(127), MK(127), MK(191), MK(127), MK(159), MK(191), MK(223), MK(127),
   MK(143), MK(159), MK(175), MK(191), MK(207), MK(223), MK(239), MK(127),
   MK(135), MK(143), MK(151), MK(159), MK(167), MK(175), MK(183), MK(191),
@@ -71,9 +75,19 @@ void VP8LoadFinalBytes(VP8BitReader* const br) {
   assert(br != NULL && br->buf_ != NULL);
   // Only read 8bits at a time
   if (br->buf_ < br->buf_end_) {
-    br->value_ |= (bit_t)(*br->buf_++) << ((BITS) - 8 + br->missing_);
-    br->missing_ -= 8;
-  } else {
+#ifndef USE_RIGHT_JUSTIFY
+    br->value_ |= (bit_t)(*br->buf_++) << ((BITS) - 8 - br->bits_);
+#else
+    br->value_ = (bit_t)(*br->buf_++) | (br->value_ << 8);
+#endif
+    br->bits_ += 8;
+  } else if (!br->eof_) {
+#ifdef USE_RIGHT_JUSTIFY
+    // These are not strictly needed, but it makes the behaviour
+    // consistent for both USE_RIGHT_JUSTIFY and !USE_RIGHT_JUSTIFY.
+    br->value_ <<= 8;
+    br->bits_ += 8;
+#endif
     br->eof_ = 1;
   }
 }
@@ -99,6 +113,10 @@ int32_t VP8GetSignedValue(VP8BitReader* const br, int bits) {
 
 #define MAX_NUM_BIT_READ 25
 
+#define LBITS 64      // Number of bits prefetched.
+#define WBITS 32      // Minimum number of bytes needed after VP8LFillBitWindow.
+#define LOG8_WBITS 4  // Number of bytes needed to store WBITS bits.
+
 static const uint32_t kBitMask[MAX_NUM_BIT_READ] = {
   0, 1, 3, 7, 15, 31, 63, 127, 255, 511, 1023, 2047, 4095, 8191, 16383, 32767,
   65535, 131071, 262143, 524287, 1048575, 2097151, 4194303, 8388607, 16777215
@@ -120,7 +138,7 @@ void VP8LInitBitReader(VP8LBitReader* const br,
   br->eos_ = 0;
   br->error_ = 0;
   for (i = 0; i < sizeof(br->val_) && i < br->len_; ++i) {
-    br->val_ |= ((uint64_t)br->buf_[br->pos_]) << (8 * i);
+    br->val_ |= ((vp8l_val_t)br->buf_[br->pos_]) << (8 * i);
     ++br->pos_;
   }
 }
@@ -135,91 +153,56 @@ void VP8LBitReaderSetBuffer(VP8LBitReader* const br,
   br->len_ = len;
 }
 
+// If not at EOS, reload up to LBITS byte-by-byte
 static void ShiftBytes(VP8LBitReader* const br) {
   while (br->bit_pos_ >= 8 && br->pos_ < br->len_) {
     br->val_ >>= 8;
-    br->val_ |= ((uint64_t)br->buf_[br->pos_]) << 56;
+    br->val_ |= ((vp8l_val_t)br->buf_[br->pos_]) << (LBITS - 8);
     ++br->pos_;
     br->bit_pos_ -= 8;
   }
 }
 
 void VP8LFillBitWindow(VP8LBitReader* const br) {
-  if (br->bit_pos_ >= 32) {
-#if defined(__x86_64__) || defined(_M_X64)
-    if (br->pos_ + 8 < br->len_) {
-      br->val_ >>= 32;
+  if (br->bit_pos_ >= WBITS) {
+#if (defined(__x86_64__) || defined(_M_X64))
+    if (br->pos_ + sizeof(br->val_) < br->len_) {
+      br->val_ >>= WBITS;
+      br->bit_pos_ -= WBITS;
       // The expression below needs a little-endian arch to work correctly.
       // This gives a large speedup for decoding speed.
-      br->val_ |= *(const uint64_t *)(br->buf_ + br->pos_) << 32;
-      br->pos_ += 4;
-      br->bit_pos_ -= 32;
-    } else {
-      // Slow path.
-      ShiftBytes(br);
+      br->val_ |= *(const vp8l_val_t*)(br->buf_ + br->pos_) << (LBITS - WBITS);
+      br->pos_ += LOG8_WBITS;
+      return;
     }
-#else
-    // Always the slow path.
-    ShiftBytes(br);
 #endif
-  }
-  if (br->pos_ == br->len_ && br->bit_pos_ == 64) {
-    br->eos_ = 1;
-  }
-}
-
-uint32_t VP8LReadOneBit(VP8LBitReader* const br) {
-  const uint32_t val = (br->val_ >> br->bit_pos_) & 1;
-  // Flag an error at end_of_stream.
-  if (!br->eos_) {
-    ++br->bit_pos_;
-    if (br->bit_pos_ >= 32) {
-      ShiftBytes(br);
-    }
-    // After this last bit is read, check if eos needs to be flagged.
-    if (br->pos_ == br->len_ && br->bit_pos_ == 64) {
+    ShiftBytes(br);       // Slow path.
+    if (br->pos_ == br->len_ && br->bit_pos_ == LBITS) {
       br->eos_ = 1;
     }
-  } else {
-    br->error_ = 1;
   }
-  return val;
 }
 
 uint32_t VP8LReadBits(VP8LBitReader* const br, int n_bits) {
-  uint32_t val = 0;
   assert(n_bits >= 0);
   // Flag an error if end_of_stream or n_bits is more than allowed limit.
   if (!br->eos_ && n_bits < MAX_NUM_BIT_READ) {
+    const uint32_t val =
+        (uint32_t)(br->val_ >> br->bit_pos_) & kBitMask[n_bits];
+    const int new_bits = br->bit_pos_ + n_bits;
+    br->bit_pos_ = new_bits;
     // If this read is going to cross the read buffer, set the eos flag.
     if (br->pos_ == br->len_) {
-      if ((br->bit_pos_ + n_bits) >= 64) {
+      if (new_bits >= LBITS) {
         br->eos_ = 1;
-        if ((br->bit_pos_ + n_bits) > 64) return val;
-      }
-    }
-    val = (br->val_ >> br->bit_pos_) & kBitMask[n_bits];
-    br->bit_pos_ += n_bits;
-    if (br->bit_pos_ >= 40) {
-      if (br->pos_ + 5 < br->len_) {
-        br->val_ >>= 40;
-        br->val_ |=
-            (((uint64_t)br->buf_[br->pos_ + 0]) << 24) |
-            (((uint64_t)br->buf_[br->pos_ + 1]) << 32) |
-            (((uint64_t)br->buf_[br->pos_ + 2]) << 40) |
-            (((uint64_t)br->buf_[br->pos_ + 3]) << 48) |
-            (((uint64_t)br->buf_[br->pos_ + 4]) << 56);
-        br->pos_ += 5;
-        br->bit_pos_ -= 40;
-      }
-      if (br->bit_pos_ >= 8) {
-        ShiftBytes(br);
       }
     }
+    ShiftBytes(br);
+    return val;
   } else {
     br->error_ = 1;
+    return 0;
   }
-  return val;
 }
 
 //------------------------------------------------------------------------------
diff --git a/src/utils/bit_reader.h b/src/utils/bit_reader.h
index 11a40a55..d98332e1 100644
--- a/src/utils/bit_reader.h
+++ b/src/utils/bit_reader.h
@@ -24,11 +24,72 @@
 extern "C" {
 #endif
 
-#define BITS 32     // can be 32, 16 or 8
-#define MASK ((((bit_t)1) << (BITS)) - 1)
+// The Boolean decoder needs to maintain infinite precision on the value_ field.
+// However, since range_ is only 8bit, we only need an active window of 8 bits
+// for value_. Left bits (MSB) gets zeroed and shifted away when value_ falls
+// below 128, range_ is updated, and fresh bits read from the bitstream are
+// brought in as LSB.
+// To avoid reading the fresh bits one by one (slow), we cache a few of them
+// ahead (actually, we cache BITS of them ahead. See below). There's two
+// strategies regarding how to shift these looked-ahead fresh bits into the
+// 8bit window of value_: either we shift them in, while keeping the position of
+// the window fixed. Or we slide the window to the right while keeping the cache
+// bits at a fixed, right-justified, position.
+//
+//  Example, for BITS=16: here is the content of value_ for both strategies:
+//
+//          !USE_RIGHT_JUSTIFY            ||        USE_RIGHT_JUSTIFY
+//                                        ||
+//   <- 8b -><- 8b -><- BITS bits  ->     ||  <- 8b+3b -><- 8b -><- 13 bits ->
+//   [unused][value_][cached bits][0]     ||  [unused...][value_][cached bits]
+//  [........00vvvvvvBBBBBBBBBBBBB000]LSB || [...........00vvvvvvBBBBBBBBBBBBB]
+//                                        ||
+// After calling VP8Shift(), where we need to shift away two zeros:
+//  [........vvvvvvvvBBBBBBBBBBB00000]LSB || [.............vvvvvvvvBBBBBBBBBBB]
+//                                        ||
+// Just before we need to call VP8LoadNewBytes(), the situation is:
+//  [........vvvvvv000000000000000000]LSB || [..........................vvvvvv]
+//                                        ||
+// And just after calling VP8LoadNewBytes():
+//  [........vvvvvvvvBBBBBBBBBBBBBBBB]LSB || [........vvvvvvvvBBBBBBBBBBBBBBBB]
+//
+// -> we're back to height active 'value_' bits (marked 'v') and BITS cached
+// bits (marked 'B')
+//
+// The right-justify strategy tends to use less shifts and is often faster.
+
+//------------------------------------------------------------------------------
+// BITS can be either 32, 24, 16 or 8.
+// Pick values that fit natural register size.
+
+#if !defined(WEBP_REFERENCE_IMPLEMENTATION)
+
+#define USE_RIGHT_JUSTIFY
+
+#if defined(__i386__) || defined(_M_IX86)      // x86 32bit
+#define BITS 16
+#elif defined(__arm__) || defined(_M_ARM)     // ARM
+#define BITS 24
+#else                      // reasonable default
+#define BITS 24
+#endif
+
+#else     // reference choices
+
+#define USE_RIGHT_JUSTIFY
+#define BITS 8
+
+#endif
+
+//------------------------------------------------------------------------------
+// Derived types and constants
+
 #if (BITS == 32)
 typedef uint64_t bit_t;   // natural register type
 typedef uint32_t lbit_t;  // natural type for memory I/O
+#elif (BITS == 24)
+typedef uint32_t bit_t;
+typedef uint32_t lbit_t;
 #elif (BITS == 16)
 typedef uint32_t bit_t;
 typedef uint16_t lbit_t;
@@ -37,8 +98,15 @@ typedef uint32_t bit_t;
 typedef uint8_t lbit_t;
 #endif
 
+#ifndef USE_RIGHT_JUSTIFY
+typedef bit_t range_t;     // type for storing range_
+#define MASK ((((bit_t)1) << (BITS)) - 1)
+#else
+typedef uint32_t range_t;  // range_ only uses 8bits here. No need for bit_t.
+#endif
+
 //------------------------------------------------------------------------------
-// Bitreader and code-tree reader
+// Bitreader
 
 typedef struct VP8BitReader VP8BitReader;
 struct VP8BitReader {
@@ -47,9 +115,9 @@ struct VP8BitReader {
   int eof_;                   // true if input is exhausted
 
   // boolean decoder
-  bit_t range_;            // current range minus 1. In [127, 254] interval.
-  bit_t value_;            // current value
-  int missing_;            // number of missing bits in value_ (8bit)
+  range_t range_;            // current range minus 1. In [127, 254] interval.
+  bit_t value_;              // current value
+  int bits_;                 // number of valid bits left
 };
 
 // Initialize the bit reader and the boolean decoder.
@@ -67,12 +135,12 @@ int32_t VP8GetSignedValue(VP8BitReader* const br, int num_bits);
 
 // Read a bit with proba 'prob'. Speed-critical function!
 extern const uint8_t kVP8Log2Range[128];
-extern const bit_t kVP8NewRange[128];
+extern const range_t kVP8NewRange[128];
 
 void VP8LoadFinalBytes(VP8BitReader* const br);    // special case for the tail
 
 static WEBP_INLINE void VP8LoadNewBytes(VP8BitReader* const br) {
-  assert(br && br->buf_);
+  assert(br != NULL && br->buf_ != NULL);
   // Read 'BITS' bits at a time if possible.
   if (br->buf_ + sizeof(lbit_t) <= br->buf_end_) {
     // convert memory type to register type (with some zero'ing!)
@@ -80,64 +148,103 @@ static WEBP_INLINE void VP8LoadNewBytes(VP8BitReader* const br) {
     lbit_t in_bits = *(lbit_t*)br->buf_;
     br->buf_ += (BITS) >> 3;
 #if !defined(__BIG_ENDIAN__)
-#if (BITS == 32)
+#if (BITS == 32) || (BITS == 24)
 #if defined(__i386__) || defined(__x86_64__)
     __asm__ volatile("bswap %k0" : "=r"(in_bits) : "0"(in_bits));
-    bits = (bit_t)in_bits;   // 32b -> 64b zero-extension
+    bits = (bit_t)in_bits;   // 24b/32b -> 32b/64b zero-extension
 #elif defined(_MSC_VER)
     bits = _byteswap_ulong(in_bits);
 #else
     bits = (bit_t)(in_bits >> 24) | ((in_bits >> 8) & 0xff00)
          | ((in_bits << 8) & 0xff0000)  | (in_bits << 24);
 #endif  // x86
+#if (BITS == 24)
+    bits >>= 8;
+#endif
 #elif (BITS == 16)
     // gcc will recognize a 'rorw $8, ...' here:
     bits = (bit_t)(in_bits >> 8) | ((in_bits & 0xff) << 8);
+#else   // BITS == 8
+    bits = (bit_t)in_bits;
 #endif
-#else    // LITTLE_ENDIAN
+#else    // BIG_ENDIAN
     bits = (bit_t)in_bits;
 #endif
-    br->value_ |= bits << br->missing_;
-    br->missing_ -= (BITS);
+#ifndef USE_RIGHT_JUSTIFY
+    br->value_ |= bits << (-br->bits_);
+#else
+    br->value_ = bits | (br->value_ << (BITS));
+#endif
+    br->bits_ += (BITS);
   } else {
     VP8LoadFinalBytes(br);    // no need to be inlined
   }
 }
 
-static WEBP_INLINE int VP8BitUpdate(VP8BitReader* const br, bit_t split) {
-  const bit_t value_split = split | (MASK);
-  if (br->missing_ > 0) {  // Make sure we have a least BITS bits in 'value_'
+static WEBP_INLINE int VP8BitUpdate(VP8BitReader* const br, range_t split) {
+  if (br->bits_ < 0) {  // Make sure we have a least BITS bits in 'value_'
     VP8LoadNewBytes(br);
   }
-  if (br->value_ > value_split) {
-    br->range_ -= value_split + 1;
-    br->value_ -= value_split + 1;
+#ifndef USE_RIGHT_JUSTIFY
+  split |= (MASK);
+  if (br->value_ > split) {
+    br->range_ -= split + 1;
+    br->value_ -= split + 1;
     return 1;
   } else {
-    br->range_ = value_split;
+    br->range_ = split;
     return 0;
   }
+#else
+  {
+    const int pos = br->bits_;
+    const range_t value = (range_t)(br->value_ >> pos);
+    if (value > split) {
+      br->range_ -= split + 1;
+      br->value_ -= (bit_t)(split + 1) << pos;
+      return 1;
+    } else {
+      br->range_ = split;
+      return 0;
+    }
+  }
+#endif
 }
 
 static WEBP_INLINE void VP8Shift(VP8BitReader* const br) {
+#ifndef USE_RIGHT_JUSTIFY
   // range_ is in [0..127] interval here.
-  const int idx = br->range_ >> (BITS);
+  const bit_t idx = br->range_ >> (BITS);
   const int shift = kVP8Log2Range[idx];
   br->range_ = kVP8NewRange[idx];
   br->value_ <<= shift;
-  br->missing_ += shift;
+  br->bits_ -= shift;
+#else
+  const int shift = kVP8Log2Range[br->range_];
+  assert(br->range_ < (range_t)128);
+  br->range_ = kVP8NewRange[br->range_];
+  br->bits_ -= shift;
+#endif
 }
-
 static WEBP_INLINE int VP8GetBit(VP8BitReader* const br, int prob) {
+#ifndef USE_RIGHT_JUSTIFY
   // It's important to avoid generating a 64bit x 64bit multiply here.
   // We just need an 8b x 8b after all.
-  const bit_t split =
-      (bit_t)((uint32_t)(br->range_ >> (BITS)) * prob) << ((BITS) - 8);
+  const range_t split =
+      (range_t)((uint32_t)(br->range_ >> (BITS)) * prob) << ((BITS) - 8);
+  const int bit = VP8BitUpdate(br, split);
+  if (br->range_ <= (((range_t)0x7e << (BITS)) | (MASK))) {
+    VP8Shift(br);
+  }
+  return bit;
+#else
+  const range_t split = (br->range_ * prob) >> 8;
   const int bit = VP8BitUpdate(br, split);
-  if (br->range_ <= (((bit_t)0x7e << (BITS)) | (MASK))) {
+  if (br->range_ <= (range_t)0x7e) {
     VP8Shift(br);
   }
   return bit;
+#endif
 }
 
 static WEBP_INLINE int VP8GetSigned(VP8BitReader* const br, int v) {
@@ -149,16 +256,18 @@ static WEBP_INLINE int VP8GetSigned(VP8BitReader* const br, int v) {
 
 
 // -----------------------------------------------------------------------------
-// Bitreader
+// Bitreader for lossless format
+
+typedef uint64_t vp8l_val_t;  // right now, this bit-reader can only use 64bit.
 
 typedef struct {
-  uint64_t       val_;
-  const uint8_t* buf_;
-  size_t         len_;
-  size_t         pos_;
-  int            bit_pos_;
-  int            eos_;
-  int            error_;
+  vp8l_val_t     val_;        // pre-fetched bits
+  const uint8_t* buf_;        // input byte buffer
+  size_t         len_;        // buffer length
+  size_t         pos_;        // byte position in buf_
+  int            bit_pos_;    // current bit-reading position in val_
+  int            eos_;        // bitstream is finished
+  int            error_;      // an error occurred (buffer overflow attempt...)
 } VP8LBitReader;
 
 void VP8LInitBitReader(VP8LBitReader* const br,
@@ -174,17 +283,14 @@ void VP8LBitReaderSetBuffer(VP8LBitReader* const br,
 // Flags eos if this read attempt is going to cross the read buffer.
 uint32_t VP8LReadBits(VP8LBitReader* const br, int n_bits);
 
-// Reads one bit from Read Buffer. Flags an error in case end_of_stream.
-// Flags eos after reading last bit from the buffer.
-uint32_t VP8LReadOneBit(VP8LBitReader* const br);
-
-// VP8LReadOneBitUnsafe is faster than VP8LReadOneBit, but it can be called only
-// 32 times after the last VP8LFillBitWindow. Any subsequent calls
-// (without VP8LFillBitWindow) will return invalid data.
-static WEBP_INLINE uint32_t VP8LReadOneBitUnsafe(VP8LBitReader* const br) {
-  const uint32_t val = (br->val_ >> br->bit_pos_) & 1;
-  ++br->bit_pos_;
-  return val;
+// Return the prefetched bits, so they can be looked up.
+static WEBP_INLINE uint32_t VP8LPrefetchBits(VP8LBitReader* const br) {
+  return (uint32_t)(br->val_ >> br->bit_pos_);
+}
+
+// Discard 'num_bits' bits from the cache.
+static WEBP_INLINE void VP8LDiscardBits(VP8LBitReader* const br, int num_bits) {
+  br->bit_pos_ += num_bits;
 }
 
 // Advances the Read buffer by 4 bytes to make room for reading next 32 bits.
diff --git a/src/utils/huffman_encode.c b/src/utils/huffman_encode.c
index 2686c665..49187592 100644
--- a/src/utils/huffman_encode.c
+++ b/src/utils/huffman_encode.c
@@ -138,13 +138,8 @@ static int CompareHuffmanTrees(const void* ptr1, const void* ptr2) {
   } else if (t1->total_count_ < t2->total_count_) {
     return 1;
   } else {
-    if (t1->value_ < t2->value_) {
-      return -1;
-    }
-    if (t1->value_ > t2->value_) {
-      return 1;
-    }
-    return 0;
+    assert(t1->value_ != t2->value_);
+    return (t1->value_ < t2->value_) ? -1 : 1;
   }
 }
 
@@ -193,6 +188,10 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
     }
   }
 
+  if (tree_size_orig == 0) {   // pretty optimal already!
+    return 1;
+  }
+
   // 3 * tree_size is enough to cover all the nodes representing a
   // population and all the inserted nodes combining two existing nodes.
   // The tree pool needs 2 * (tree_size_orig - 1) entities, and the
@@ -234,7 +233,7 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
         tree_pool[tree_pool_size++] = tree[tree_size - 1];
         tree_pool[tree_pool_size++] = tree[tree_size - 2];
         count = tree_pool[tree_pool_size - 1].total_count_ +
-            tree_pool[tree_pool_size - 2].total_count_;
+                tree_pool[tree_pool_size - 2].total_count_;
         tree_size -= 2;
         {
           // Search for the insertion point.
diff --git a/src/utils/quant_levels.c b/src/utils/quant_levels.c
index f6884392..649aae65 100644
--- a/src/utils/quant_levels.c
+++ b/src/utils/quant_levels.c
@@ -140,15 +140,6 @@ int QuantizeLevels(uint8_t* const data, int width, int height,
   return 1;
 }
 
-int DequantizeLevels(uint8_t* const data, int width, int height) {
-  if (data == NULL || width <= 0 || height <= 0) return 0;
-  // TODO(skal): implement gradient smoothing.
-  (void)data;
-  (void)width;
-  (void)height;
-  return 1;
-}
-
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
diff --git a/src/utils/quant_levels.h b/src/utils/quant_levels.h
index 8dd3afee..f86cbe59 100644
--- a/src/utils/quant_levels.h
+++ b/src/utils/quant_levels.h
@@ -27,11 +27,6 @@ extern "C" {
 int QuantizeLevels(uint8_t* const data, int width, int height, int num_levels,
                    uint64_t* const sse);
 
-// Apply post-processing to input 'data' of size 'width'x'height' assuming
-// that the source was quantized to a reduced number of levels.
-// Returns false in case of error (data is NULL, invalid parameters, ...).
-int DequantizeLevels(uint8_t* const data, int width, int height);
-
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
diff --git a/src/utils/quant_levels_dec.c b/src/utils/quant_levels_dec.c
new file mode 100644
index 00000000..95142b1b
--- /dev/null
+++ b/src/utils/quant_levels_dec.c
@@ -0,0 +1,28 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+// TODO(skal): implement gradient smoothing.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./quant_levels_dec.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+int DequantizeLevels(uint8_t* const data, int width, int height) {
+  if (data == NULL || width <= 0 || height <= 0) return 0;
+  (void)data;
+  (void)width;
+  (void)height;
+  return 1;
+}
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}    // extern "C"
+#endif
diff --git a/src/utils/quant_levels_dec.h b/src/utils/quant_levels_dec.h
new file mode 100644
index 00000000..de60ef95
--- /dev/null
+++ b/src/utils/quant_levels_dec.h
@@ -0,0 +1,30 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+// Alpha plane de-quantization utility
+//
+// Author:  Vikas Arora (vikasa@google.com)
+
+#ifndef WEBP_UTILS_QUANT_LEVELS_DEC_H_
+#define WEBP_UTILS_QUANT_LEVELS_DEC_H_
+
+#include "webp/types.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+// Apply post-processing to input 'data' of size 'width'x'height' assuming
+// that the source was quantized to a reduced number of levels.
+// Returns false in case of error (data is NULL, invalid parameters, ...).
+int DequantizeLevels(uint8_t* const data, int width, int height);
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_QUANT_LEVELS_DEC_H_ */
diff --git a/src/utils/thread.c b/src/utils/thread.c
index ce89cf9d..a14af559 100644
--- a/src/utils/thread.c
+++ b/src/utils/thread.c
@@ -9,10 +9,6 @@
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
-#ifdef HAVE_CONFIG_H
-#include "config.h"
-#endif
-
 #include <assert.h>
 #include <string.h>   // for memset()
 #include "./thread.h"
diff --git a/src/utils/thread.h b/src/utils/thread.h
index 3191890b..9afe0967 100644
--- a/src/utils/thread.h
+++ b/src/utils/thread.h
@@ -12,6 +12,10 @@
 #ifndef WEBP_UTILS_THREAD_H_
 #define WEBP_UTILS_THREAD_H_
 
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
 #endif
@@ -63,13 +67,13 @@ typedef struct {
 
 // Must be called first, before any other method.
 void WebPWorkerInit(WebPWorker* const worker);
-// Must be called initialize the object and spawn the thread. Re-entrant.
+// Must be called to initialize the object and spawn the thread. Re-entrant.
 // Will potentially launch the thread. Returns false in case of error.
 int WebPWorkerReset(WebPWorker* const worker);
-// Make sure the previous work is finished. Returns true if worker->had_error
-// was not set and not error condition was triggered by the working thread.
+// Makes sure the previous work is finished. Returns true if worker->had_error
+// was not set and no error condition was triggered by the working thread.
 int WebPWorkerSync(WebPWorker* const worker);
-// Trigger the thread to call hook() with data1 and data2 argument. These
+// Triggers the thread to call hook() with data1 and data2 argument. These
 // hook/data1/data2 can be changed at any time before calling this function,
 // but not be changed afterward until the next call to WebPWorkerSync().
 void WebPWorkerLaunch(WebPWorker* const worker);
diff --git a/src/utils/utils.c b/src/utils/utils.c
index 673b7e28..b1db2f9d 100644
--- a/src/utils/utils.c
+++ b/src/utils/utils.c
@@ -19,7 +19,8 @@ extern "C" {
 //------------------------------------------------------------------------------
 // Checked memory allocation
 
-static int CheckSizeArguments(uint64_t nmemb, size_t size) {
+// Returns 0 in case of overflow of nmemb * size.
+static int CheckSizeArgumentsOverflow(uint64_t nmemb, size_t size) {
   const uint64_t total_size = nmemb * size;
   if (nmemb == 0) return 1;
   if ((uint64_t)size > WEBP_MAX_ALLOCABLE_MEMORY / nmemb) return 0;
@@ -28,12 +29,14 @@ static int CheckSizeArguments(uint64_t nmemb, size_t size) {
 }
 
 void* WebPSafeMalloc(uint64_t nmemb, size_t size) {
-  if (!CheckSizeArguments(nmemb, size)) return NULL;
+  if (!CheckSizeArgumentsOverflow(nmemb, size)) return NULL;
+  assert(nmemb * size > 0);
   return malloc((size_t)(nmemb * size));
 }
 
 void* WebPSafeCalloc(uint64_t nmemb, size_t size) {
-  if (!CheckSizeArguments(nmemb, size)) return NULL;
+  if (!CheckSizeArgumentsOverflow(nmemb, size)) return NULL;
+  assert(nmemb * size > 0);
   return calloc((size_t)nmemb, size);
 }
 
diff --git a/src/utils/utils.h b/src/utils/utils.h
index aa445695..32dfb8a9 100644
--- a/src/utils/utils.h
+++ b/src/utils/utils.h
@@ -7,11 +7,14 @@
 //
 // Misc. common utility functions
 //
-// Author: Skal (pascal.massimino@gmail.com)
+// Authors: Skal (pascal.massimino@gmail.com)
+//          Urvang (urvang@google.com)
 
 #ifndef WEBP_UTILS_UTILS_H_
 #define WEBP_UTILS_UTILS_H_
 
+#include <assert.h>
+
 #include "webp/types.h"
 
 #if defined(__cplusplus) || defined(c_plusplus)
@@ -36,6 +39,40 @@ void* WebPSafeMalloc(uint64_t nmemb, size_t size);
 void* WebPSafeCalloc(uint64_t nmemb, size_t size);
 
 //------------------------------------------------------------------------------
+// Reading/writing data.
+
+// Read 16, 24 or 32 bits stored in little-endian order.
+static WEBP_INLINE int GetLE16(const uint8_t* const data) {
+  return (int)(data[0] << 0) | (data[1] << 8);
+}
+
+static WEBP_INLINE int GetLE24(const uint8_t* const data) {
+  return GetLE16(data) | (data[2] << 16);
+}
+
+static WEBP_INLINE uint32_t GetLE32(const uint8_t* const data) {
+  return (uint32_t)GetLE16(data) | (GetLE16(data + 2) << 16);
+}
+
+// Store 16, 24 or 32 bits in little-endian order.
+static WEBP_INLINE void PutLE16(uint8_t* const data, int val) {
+  assert(val < (1 << 16));
+  data[0] = (val >> 0);
+  data[1] = (val >> 8);
+}
+
+static WEBP_INLINE void PutLE24(uint8_t* const data, int val) {
+  assert(val < (1 << 24));
+  PutLE16(data, val & 0xffff);
+  data[2] = (val >> 16);
+}
+
+static WEBP_INLINE void PutLE32(uint8_t* const data, uint32_t val) {
+  PutLE16(data, (int)(val & 0xffff));
+  PutLE16(data + 2, (int)(val >> 16));
+}
+
+//------------------------------------------------------------------------------
 
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"