Ejemplo n.º 1
0
/* Builds a literal prefix code into "depths" and "bits" based on the statistics
   of the "input" string and stores it into the bit stream.
   Note that the prefix code here is built from the pre-LZ77 input, therefore
   we can only approximate the statistics of the actual literal stream.
   Moreover, for long inputs we build a histogram from a sample of the input
   and thus have to assign a non-zero depth for each literal.
   Returns estimated compression ratio millibytes/char for encoding given input
   with generated code. */
static size_t BuildAndStoreLiteralPrefixCode(MemoryManager* m,
                                             const uint8_t* input,
                                             const size_t input_size,
                                             uint8_t depths[256],
                                             uint16_t bits[256],
                                             size_t* storage_ix,
                                             uint8_t* storage) {
  uint32_t histogram[256] = { 0 };
  size_t histogram_total;
  size_t i;
  if (input_size < (1 << 15)) {
    for (i = 0; i < input_size; ++i) {
      ++histogram[input[i]];
    }
    histogram_total = input_size;
    for (i = 0; i < 256; ++i) {
      /* We weigh the first 11 samples with weight 3 to account for the
         balancing effect of the LZ77 phase on the histogram. */
      const uint32_t adjust = 2 * BROTLI_MIN(uint32_t, histogram[i], 11u);
      histogram[i] += adjust;
      histogram_total += adjust;
    }
  } else {
    static const size_t kSampleRate = 29;
    for (i = 0; i < input_size; i += kSampleRate) {
      ++histogram[input[i]];
    }
    histogram_total = (input_size + kSampleRate - 1) / kSampleRate;
    for (i = 0; i < 256; ++i) {
      /* We add 1 to each population count to avoid 0 bit depths (since this is
         only a sample and we don't know if the symbol appears or not), and we
         weigh the first 11 samples with weight 3 to account for the balancing
         effect of the LZ77 phase on the histogram (more frequent symbols are
         more likely to be in backward references instead as literals). */
      const uint32_t adjust = 1 + 2 * BROTLI_MIN(uint32_t, histogram[i], 11u);
      histogram[i] += adjust;
      histogram_total += adjust;
    }
  }
  BrotliBuildAndStoreHuffmanTreeFast(m, histogram, histogram_total,
                                     /* max_bits = */ 8,
                                     depths, bits, storage_ix, storage);
  if (BROTLI_IS_OOM(m)) return 0;
  {
    size_t literal_ratio = 0;
    for (i = 0; i < 256; ++i) {
      if (histogram[i]) literal_ratio += histogram[i] * depths[i];
    }
    /* Estimated encoding ratio, millibytes per symbol. */
    return (literal_ratio * 125) / histogram_total;
  }
}
Ejemplo n.º 2
0
void BrotliSplitBlock(MemoryManager* m,
                      const Command* cmds,
                      const size_t num_commands,
                      const uint8_t* data,
                      const size_t pos,
                      const size_t mask,
                      const BrotliEncoderParams* params,
                      BlockSplit* literal_split,
                      BlockSplit* insert_and_copy_split,
                      BlockSplit* dist_split) {
  {
    size_t literals_count = CountLiterals(cmds, num_commands);
    uint8_t* literals = BROTLI_ALLOC(m, uint8_t, literals_count);
    if (BROTLI_IS_OOM(m)) return;
    /* Create a continuous array of literals. */
    CopyLiteralsToByteArray(cmds, num_commands, data, pos, mask, literals);
    /* Create the block split on the array of literals.
       Literal histograms have alphabet size 256. */
    SplitByteVectorLiteral(
        m, literals, literals_count,
        kSymbolsPerLiteralHistogram, kMaxLiteralHistograms,
        kLiteralStrideLength, kLiteralBlockSwitchCost, params,
        literal_split);
    if (BROTLI_IS_OOM(m)) return;
    BROTLI_FREE(m, literals);
  }

  {
    /* Compute prefix codes for commands. */
    uint16_t* insert_and_copy_codes = BROTLI_ALLOC(m, uint16_t, num_commands);
    size_t i;
    if (BROTLI_IS_OOM(m)) return;
    for (i = 0; i < num_commands; ++i) {
      insert_and_copy_codes[i] = cmds[i].cmd_prefix_;
    }
    /* Create the block split on the array of command prefixes. */
    SplitByteVectorCommand(
        m, insert_and_copy_codes, num_commands,
        kSymbolsPerCommandHistogram, kMaxCommandHistograms,
        kCommandStrideLength, kCommandBlockSwitchCost, params,
        insert_and_copy_split);
    if (BROTLI_IS_OOM(m)) return;
    /* TODO: reuse for distances? */
    BROTLI_FREE(m, insert_and_copy_codes);
  }

  {
    /* Create a continuous array of distance prefixes. */
    uint16_t* distance_prefixes = BROTLI_ALLOC(m, uint16_t, num_commands);
    size_t j = 0;
    size_t i;
    if (BROTLI_IS_OOM(m)) return;
    for (i = 0; i < num_commands; ++i) {
      const Command* cmd = &cmds[i];
      if (CommandCopyLen(cmd) && cmd->cmd_prefix_ >= 128) {
        distance_prefixes[j++] = cmd->dist_prefix_;
      }
    }
    /* Create the block split on the array of distance prefixes. */
    SplitByteVectorDistance(
        m, distance_prefixes, j,
        kSymbolsPerDistanceHistogram, kMaxCommandHistograms,
        kCommandStrideLength, kDistanceBlockSwitchCost, params,
        dist_split);
    if (BROTLI_IS_OOM(m)) return;
    BROTLI_FREE(m, distance_prefixes);
  }
}