Updates to Brotli compression format, decoder and encoder

This commit contains a batch of changes that were made to the Brotli
compression algorithm in the last three weeks. Most important changes:

  * Added UTF8 context model for good text compression.
  * Simplified context modeling by having only 4 context modes.
  * Per-block context mode selection.
  * Faster backward copying and bit reading functions.
  * More efficient histogram coding.
  * Streaming support for the decoder and encoder.
This commit is contained in:
Zoltan Szabadka
2013-11-15 19:02:17 +01:00
parent 79e99afe46
commit 1571db36a9
23 changed files with 1647 additions and 870 deletions
+12 -10
View File
@@ -22,37 +22,39 @@
namespace brotli {
void EstimateBitCostsForLiterals(size_t len, const uint8_t *data, float *cost) {
void EstimateBitCostsForLiterals(size_t pos, size_t len, size_t mask,
const uint8_t *data, float *cost) {
int histogram[256] = { 0 };
int window_half = 2000;
int in_window = std::min(static_cast<size_t>(window_half), len);
// Bootstrap histogram.
for (int i = 0; i < in_window; ++i) {
++histogram[data[i]];
++histogram[data[(pos + i) & mask]];
}
// Compute bit costs with sliding window.
for (int i = 0; i < len; ++i) {
if (i - window_half >= 0) {
// Remove a byte in the past.
--histogram[data[i - window_half]];
--histogram[data[(pos + i - window_half) & mask]];
--in_window;
}
if (i + window_half < len) {
// Add a byte in the future.
++histogram[data[i + window_half]];
++histogram[data[(pos + i + window_half) & mask]];
++in_window;
}
int histo = histogram[data[i]];
int masked_pos = (pos + i) & mask;
int histo = histogram[data[masked_pos]];
if (histo == 0) {
histo = 1;
}
cost[i] = log2(static_cast<double>(in_window) / histo);
cost[i] += 0.03;
if (cost[i] < 1.0) {
cost[i] *= 0.5;
cost[i] += 0.5;
cost[masked_pos] = log2(static_cast<double>(in_window) / histo);
cost[masked_pos] += 0.03;
if (cost[masked_pos] < 1.0) {
cost[masked_pos] *= 0.5;
cost[masked_pos] += 0.5;
}
}
}