AuroraMiddleware/brotli
1
0
Fork 0
mirror of https://github.com/google/brotli.git synced 2024-11-25 04:50:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #83 from szabadka/master

Browse Source 
Encoder implementation using input/output classes.
This commit is contained in:
szabadka 2015-04-23 15:30:51 +02:00
commit 5ea92475af
4 changed files with 340 additions and 152 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
enc/context.h
View File

@ -139,7 +139,7 @@ static const uint8_t kUTF8ContextLookup[512] = {
};
// Context lookup table for small signed integers.
static const int kSigned3BitContextLookup[] = {
static const uint8_t kSigned3BitContextLookup[] = {
0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,

371
enc/encode.cc
View File

@ -37,6 +37,8 @@
namespace brotli {
static const double kMinUTF8Ratio = 0.75;
int ParseAsUTF8(int* symbol, const uint8_t* input, int size) {
// ASCII
if ((input[0] & 0x80) == 0) {
@ -130,8 +132,14 @@ uint8_t* BrotliCompressor::GetBrotliStorage(size_t size) {
BrotliCompressor::BrotliCompressor(BrotliParams params)
: params_(params),
hashers_(new Hashers()),
input_pos_(0) {
input_pos_(0),
num_commands_(0),
last_insert_len_(0),
last_flush_pos_(0),
last_processed_pos_(0),
storage_size_(0) {
// Sanitize params.
params_.quality = std::max(0, params_.quality);
if (params_.lgwin < kMinWindowBits) {
params_.lgwin = kMinWindowBits;
} else if (params_.lgwin > kMaxWindowBits) {
@ -168,9 +176,11 @@ BrotliCompressor::BrotliCompressor(BrotliParams params)
literal_cost_.reset(new float[literal_cost_mask_ + 1]);
}
// Initialize storage.
storage_size_ = 1 << 16;
storage_.reset(new uint8_t[storage_size_]);
// Allocate command buffer.
cmd_buffer_size_ = std::max(1 << 18, 1 << params_.lgblock);
commands_.reset(new brotli::Command[cmd_buffer_size_]);
// Initialize last byte with stream header.
if (params_.lgwin == 16) {
last_byte_ = 0;
last_byte_bits_ = 1;
@ -186,10 +196,19 @@ BrotliCompressor::BrotliCompressor(BrotliParams params)
dist_cache_[3] = 16;
// Initialize hashers.
switch (params_.mode) {
case BrotliParams::MODE_TEXT: hash_type_ = 8; break;
case BrotliParams::MODE_FONT: hash_type_ = 9; break;
default: break;
switch (params_.quality) {
case 0:
case 1: hash_type_ = 1; break;
case 2:
case 3: hash_type_ = 2; break;
case 4: hash_type_ = 3; break;
case 5:
case 6: hash_type_ = 4; break;
case 7: hash_type_ = 5; break;
case 8: hash_type_ = 6; break;
case 9: hash_type_ = 7; break;
default: // quality > 9
hash_type_ = (params_.mode == BrotliParams::MODE_TEXT) ? 8 : 9;
}
hashers_->Init(hash_type_);
if (params_.mode == BrotliParams::MODE_TEXT &&
@ -201,7 +220,7 @@ BrotliCompressor::BrotliCompressor(BrotliParams params)
BrotliCompressor::~BrotliCompressor() {
}
StaticDictionary *BrotliCompressor::static_dictionary_ = NULL;
StaticDictionary* BrotliCompressor::static_dictionary_ = NULL;
void BrotliCompressor::StoreDictionaryWordHashes(bool enable_transforms) {
if (static_dictionary_ == NULL) {
@ -211,123 +230,240 @@ void BrotliCompressor::StoreDictionaryWordHashes(bool enable_transforms) {
hashers_->SetStaticDictionary(static_dictionary_);
}
bool BrotliCompressor::WriteMetaBlock(const size_t input_size,
const uint8_t* input_buffer,
const bool is_last,
size_t* encoded_size,
uint8_t* encoded_buffer) {
if (input_size > input_block_size()) {
void BrotliCompressor::CopyInputToRingBuffer(const size_t input_size,
const uint8_t* input_buffer) {
ringbuffer_->Write(input_buffer, input_size);
input_pos_ += input_size;
// Erase a few more bytes in the ring buffer to make hashing not
// depend on uninitialized data. This makes compression deterministic
// and it prevents uninitialized memory warnings in Valgrind. Even
// without erasing, the output would be valid (but nondeterministic).
//
// Background information: The compressor stores short (at most 8 bytes)
// substrings of the input already read in a hash table, and detects
// repetitions by looking up such substrings in the hash table. If it
// can find a substring, it checks whether the substring is really there
// in the ring buffer (or it's just a hash collision). Should the hash
// table become corrupt, this check makes sure that the output is
// still valid, albeit the compression ratio would be bad.
//
// The compressor populates the hash table from the ring buffer as it's
// reading new bytes from the input. However, at the last few indexes of
// the ring buffer, there are not enough bytes to build full-length
// substrings from. Since the hash table always contains full-length
// substrings, we erase with dummy 0s here to make sure that those
// substrings will contain 0s at the end instead of uninitialized
// data.
//
// Please note that erasing is not necessary (because the
// memory region is already initialized since he ring buffer
// has a `tail' that holds a copy of the beginning,) so we
// skip erasing if we have already gone around at least once in
// the ring buffer.
size_t pos = ringbuffer_->position();
// Only clear during the first round of ringbuffer writes. On
// subsequent rounds data in the ringbuffer would be affected.
if (pos <= ringbuffer_->mask()) {
// This is the first time when the ring buffer is being written.
// We clear 3 bytes just after the bytes that have been copied from
// the input buffer.
//
// The ringbuffer has a "tail" that holds a copy of the beginning,
// but only once the ring buffer has been fully written once, i.e.,
// pos <= mask. For the first time, we need to write values
// in this tail (where index may be larger than mask), so that
// we have exactly defined behavior and don't read un-initialized
// memory. Due to performance reasons, hashing reads data using a
// LOAD32, which can go 3 bytes beyond the bytes written in the
// ringbuffer.
memset(ringbuffer_->start() + pos, 0, 3);
}
}
bool BrotliCompressor::WriteBrotliData(const bool is_last,
const bool force_flush,
size_t* out_size,
uint8_t** output) {
const size_t bytes = input_pos_ - last_processed_pos_;
const uint8_t* data = ringbuffer_->start();
const size_t mask = ringbuffer_->mask();
if (bytes > input_block_size()) {
return false;
}
static const double kMinUTF8Ratio = 0.75;
bool utf8_mode = false;
std::vector<Command> commands((input_size + 1) >> 1);
// Save the state of the distance cache in case we need to restore it for
// emitting an uncompressed block.
int saved_dist_cache[4];
memcpy(saved_dist_cache, dist_cache_, sizeof(dist_cache_));
if (input_size > 0) {
ringbuffer_->Write(input_buffer, input_size);
utf8_mode =
bool utf8_mode =
params_.enable_context_modeling &&
IsMostlyUTF8(&ringbuffer_->start()[input_pos_ & ringbuffer_->mask()],
input_size, kMinUTF8Ratio);
IsMostlyUTF8(&data[last_processed_pos_ & mask], bytes, kMinUTF8Ratio);
if (literal_cost_.get()) {
if (utf8_mode) {
EstimateBitCostsForLiteralsUTF8(input_pos_, input_size,
ringbuffer_->mask(),
literal_cost_mask_,
ringbuffer_->start(),
EstimateBitCostsForLiteralsUTF8(last_processed_pos_, bytes, mask,
literal_cost_mask_, data,
literal_cost_.get());
} else {
EstimateBitCostsForLiterals(input_pos_, input_size,
ringbuffer_->mask(), literal_cost_mask_,
ringbuffer_->start(), literal_cost_.get());
EstimateBitCostsForLiterals(last_processed_pos_, bytes, mask,
literal_cost_mask_,
data, literal_cost_.get());
}
}
int last_insert_len = 0;
int num_commands = 0;
double base_min_score = params_.enable_context_modeling ? 8.115 : 4.0;
CreateBackwardReferences(
input_size, input_pos_,
ringbuffer_->start(), ringbuffer_->mask(),
literal_cost_.get(), literal_cost_mask_,
CreateBackwardReferences(bytes, last_processed_pos_, data, mask,
literal_cost_.get(),
literal_cost_mask_,
max_backward_distance_,
base_min_score,
params_.quality,
hashers_.get(),
hash_type_,
dist_cache_,
&last_insert_len,
&commands[0],
&num_commands);
commands.resize(num_commands);
if (last_insert_len > 0) {
commands.push_back(Command(last_insert_len));
&last_insert_len_,
&commands_[num_commands_],
&num_commands_);
if (!is_last && !force_flush &&
num_commands_ + (input_block_size() >> 1) < cmd_buffer_size_ &&
input_pos_ + input_block_size() + 2 <= last_flush_pos_ + mask + 1) {
// Everything will happen later.
last_processed_pos_ = input_pos_;
*out_size = 0;
return true;
}
// Create the last insert-only command.
if (last_insert_len_ > 0) {
brotli::Command cmd(last_insert_len_);
commands_[num_commands_++] = cmd;
last_insert_len_ = 0;
}
int num_direct_distance_codes =
params_.mode == BrotliParams::MODE_FONT ? 12 : 0;
int distance_postfix_bits = params_.mode == BrotliParams::MODE_FONT ? 1 : 0;
int literal_context_mode = utf8_mode ? CONTEXT_UTF8 : CONTEXT_SIGNED;
const size_t max_out_size = 2 * input_size + 500;
return WriteMetaBlockInternal(is_last, utf8_mode, out_size, output);
}
bool BrotliCompressor::WriteMetaBlockInternal(const bool is_last,
const bool utf8_mode,
size_t* out_size,
uint8_t** output) {
const size_t bytes = input_pos_ - last_flush_pos_;
const uint8_t* data = ringbuffer_->start();
const size_t mask = ringbuffer_->mask();
const size_t max_out_size = 2 * bytes + 500;
uint8_t* storage = GetBrotliStorage(max_out_size);
storage[0] = last_byte_;
int storage_ix = last_byte_bits_;
bool uncompressed = false;
if (num_commands_ < (bytes >> 8) + 2) {
int num_literals = 0;
for (int i = 0; i < num_commands_; ++i) {
num_literals += commands_[i].insert_len_;
}
if (num_literals > 0.99 * bytes) {
int literal_histo[256] = { 0 };
static const int kSampleRate = 13;
static const double kMinEntropy = 7.92;
static const double kBitCostThreshold = bytes * kMinEntropy / kSampleRate;
for (int i = last_flush_pos_; i < input_pos_; i += kSampleRate) {
++literal_histo[data[i & mask]];
}
if (BitsEntropy(literal_histo, 256) > kBitCostThreshold) {
uncompressed = true;
}
}
}
if (bytes == 0) {
if (!StoreCompressedMetaBlockHeader(is_last, 0, &storage_ix, &storage[0])) {
return false;
}
storage_ix = (storage_ix + 7) & ~7;
} else if (uncompressed) {
if (!StoreUncompressedMetaBlock(is_last,
data, last_flush_pos_, mask, bytes,
&storage_ix,
&storage[0])) {
return false;
}
} else {
// Save the state of the distance cache in case we need to restore it for
// emitting an uncompressed block.
int saved_dist_cache[4];
memcpy(saved_dist_cache, dist_cache_, sizeof(dist_cache_));
int num_direct_distance_codes = 0;
int distance_postfix_bits = 0;
if (params_.quality > 9 && params_.mode == BrotliParams::MODE_FONT) {
num_direct_distance_codes = 12;
distance_postfix_bits = 1;
}
int literal_context_mode = utf8_mode ? CONTEXT_UTF8 : CONTEXT_SIGNED;
MetaBlockSplit mb;
if (!commands.empty()) {
if (params_.greedy_block_split) {
BuildMetaBlockGreedy(ringbuffer_->start(), input_pos_,
ringbuffer_->mask(),
commands.data(), commands.size(), params_.quality,
BuildMetaBlockGreedy(data, last_flush_pos_, mask,
commands_.get(), num_commands_,
params_.quality,
&mb);
} else {
RecomputeDistancePrefixes(&commands[0], commands.size(),
RecomputeDistancePrefixes(commands_.get(),
num_commands_,
num_direct_distance_codes,
distance_postfix_bits);
BuildMetaBlock(ringbuffer_->start(), input_pos_, ringbuffer_->mask(),
commands.data(), commands.size(),
BuildMetaBlock(data, last_flush_pos_, mask,
commands_.get(), num_commands_,
num_direct_distance_codes,
distance_postfix_bits,
literal_context_mode,
params_.enable_context_modeling,
&mb);
}
}
if (!StoreMetaBlock(ringbuffer_->start(), input_pos_, input_size,
ringbuffer_->mask(),
if (!StoreMetaBlock(data, last_flush_pos_, bytes, mask,
is_last, params_.quality,
num_direct_distance_codes,
distance_postfix_bits,
literal_context_mode,
commands.data(), commands.size(),
commands_.get(), num_commands_,
mb,
&storage_ix, storage)) {
&storage_ix,
&storage[0])) {
return false;
}
size_t output_size = storage_ix >> 3;
if (input_size + 4 < output_size) {
if (bytes + 4 < (storage_ix >> 3)) {
// Restore the distance cache and last byte.
memcpy(dist_cache_, saved_dist_cache, sizeof(dist_cache_));
storage[0] = last_byte_;
storage_ix = last_byte_bits_;
if (!StoreUncompressedMetaBlock(is_last,
ringbuffer_->start(), input_pos_,
ringbuffer_->mask(), input_size,
&storage_ix, storage)) {
if (!StoreUncompressedMetaBlock(is_last, data, last_flush_pos_, mask,
bytes, &storage_ix, &storage[0])) {
return false;
}
output_size = storage_ix >> 3;
}
if (output_size > *encoded_size) {
return false;
}
memcpy(encoded_buffer, storage, output_size);
*encoded_size = output_size;
last_byte_ = storage[output_size];
last_byte_ = storage[storage_ix >> 3];
last_byte_bits_ = storage_ix & 7;
input_pos_ += input_size;
last_flush_pos_ = input_pos_;
last_processed_pos_ = input_pos_;
num_commands_ = 0;
*output = &storage[0];
*out_size = storage_ix >> 3;
return true;
}
bool BrotliCompressor::WriteMetaBlock(const size_t input_size,
const uint8_t* input_buffer,
const bool is_last,
size_t* encoded_size,
uint8_t* encoded_buffer) {
CopyInputToRingBuffer(input_size, input_buffer);
size_t out_size = 0;
uint8_t* output;
if (!WriteBrotliData(is_last, /* force_flush = */ true, &out_size, &output) ||
out_size > *encoded_size) {
return false;
}
if (out_size > 0) {
memcpy(encoded_buffer, output, out_size);
}
*encoded_size = out_size;
return true;
}
@ -344,36 +480,69 @@ int BrotliCompressBuffer(BrotliParams params,
if (*encoded_size == 0) {
// Output buffer needs at least one byte.
return 0;
} else if (input_size == 0) {
encoded_buffer[0] = 6;
*encoded_size = 1;
return 1;
}
BrotliCompressor compressor(params);
const int max_block_size = compressor.input_block_size();
size_t max_output_size = *encoded_size;
const uint8_t* input_end = input_buffer + input_size;
*encoded_size = 0;
while (input_buffer < input_end) {
int block_size = max_block_size;
bool is_last = false;
if (block_size >= input_end - input_buffer) {
block_size = input_end - input_buffer;
is_last = true;
}
size_t output_size = max_output_size;
if (!compressor.WriteMetaBlock(block_size, input_buffer,
is_last, &output_size,
&encoded_buffer[*encoded_size])) {
BrotliMemIn in(input_buffer, input_size);
BrotliMemOut out(encoded_buffer, *encoded_size);
if (!BrotliCompress(params, &in, &out)) {
return 0;
}
input_buffer += block_size;
*encoded_size += output_size;
max_output_size -= output_size;
}
*encoded_size = out.position();
return 1;
}
size_t CopyOneBlockToRingBuffer(BrotliIn* r, BrotliCompressor* compressor) {
const size_t block_size = compressor->input_block_size();
size_t bytes_read = 0;
const uint8_t* data = reinterpret_cast<const uint8_t*>(
r->Read(block_size, &bytes_read));
if (data == NULL) {
return 0;
}
compressor->CopyInputToRingBuffer(bytes_read, data);
// Read more bytes until block_size is filled or an EOF (data == NULL) is
// received. This is useful to get deterministic compressed output for the
// same input no matter how r->Read splits the input to chunks.
for (size_t remaining = block_size - bytes_read; remaining > 0; ) {
size_t more_bytes_read = 0;
data = reinterpret_cast<const uint8_t*>(
r->Read(remaining, &more_bytes_read));
if (data == NULL) {
break;
}
compressor->CopyInputToRingBuffer(more_bytes_read, data);
bytes_read += more_bytes_read;
remaining -= more_bytes_read;
}
return bytes_read;
}
bool BrotliInIsFinished(BrotliIn* r) {
size_t read_bytes;
return r->Read(0, &read_bytes) == NULL;
}
int BrotliCompress(BrotliParams params, BrotliIn* in, BrotliOut* out) {
size_t in_bytes = 0;
size_t out_bytes = 0;
uint8_t* output;
bool final_block = false;
BrotliCompressor compressor(params);
while (!final_block) {
in_bytes = CopyOneBlockToRingBuffer(in, &compressor);
final_block = in_bytes == 0 || BrotliInIsFinished(in);
out_bytes = 0;
if (!compressor.WriteBrotliData(final_block,
/* force_flush = */ false,
&out_bytes, &output)) {
return false;
}
if (out_bytes > 0 && !out->Write(output, out_bytes)) {
return false;
}
}
return true;
}
} // namespace brotli

35
enc/encode.h
View File

@ -21,9 +21,11 @@
#include <stdint.h>
#include <string>
#include <vector>
#include "./command.h"
#include "./hash.h"
#include "./ringbuffer.h"
#include "./static_dict.h"
#include "./streams.h"
namespace brotli {
@ -90,6 +92,24 @@ class BrotliCompressor {
// an error and true otherwise.
bool FinishStream(size_t* encoded_size, uint8_t* encoded_buffer);
// Copies the given input data to the internal ring buffer of the compressor.
// No processing of the data occurs at this time and this function can be
// called multiple times before calling WriteBrotliData() to process the
// accumulated input. At most input_block_size() bytes of input data can be
// copied to the ring buffer, otherwise the next WriteBrotliData() will fail.
void CopyInputToRingBuffer(const size_t input_size,
const uint8_t* input_buffer);
// Processes the accumulated input data and sets *out_size to the length of
// the new output meta-block, or to zero if no new output meta-block was
// created (in this case the processed input data is buffered internally).
// If *out_size is positive, *output points to the start of the output data.
// Returns false if the size of the input data is larger than
// input_block_size() or if there was an error during writing the output.
// If is_last or force_flush is true, an output meta-block is always created.
bool WriteBrotliData(const bool is_last, const bool force_flush,
size_t* out_size, uint8_t** output);
// No-op, but we keep it here for API backward-compatibility.
void WriteStreamHeader() {}
@ -99,6 +119,11 @@ class BrotliCompressor {
uint8_t* GetBrotliStorage(size_t size);
bool WriteMetaBlockInternal(const bool is_last,
const bool utf8_mode,
size_t* out_size,
uint8_t** output);
BrotliParams params_;
int max_backward_distance_;
std::unique_ptr<Hashers> hashers_;
@ -107,6 +132,12 @@ class BrotliCompressor {
std::unique_ptr<RingBuffer> ringbuffer_;
std::unique_ptr<float[]> literal_cost_;
size_t literal_cost_mask_;
size_t cmd_buffer_size_;
std::unique_ptr<Command[]> commands_;
int num_commands_;
int last_insert_len_;
size_t last_flush_pos_;
size_t last_processed_pos_;
int dist_cache_[4];
uint8_t last_byte_;
uint8_t last_byte_bits_;
@ -124,6 +155,10 @@ int BrotliCompressBuffer(BrotliParams params,
size_t* encoded_size,
uint8_t* encoded_buffer);
// Same as above, but uses the specified input and output classes instead
// of reading from and writing to pre-allocated memory buffers.
int BrotliCompress(BrotliParams params, BrotliIn* in, BrotliOut* out);
} // namespace brotli
#endif // BROTLI_ENC_ENCODE_H_

26
tools/bro.cc
View File

@ -24,6 +24,7 @@
#include "../dec/decode.h"
#include "../enc/encode.h"
#include "../enc/streams.h"
static void ParseArgv(int argc, char **argv,
@ -133,32 +134,15 @@ int main(int argc, char** argv) {
exit(1);
}
} else {
const int max_block_size = 1 << 21;
const size_t max_output_size = 1 << 22;
uint8_t* input_buffer = new uint8_t[max_block_size];
uint8_t* output_buffer = new uint8_t[max_output_size];
bool input_end = false;
int block_size;
brotli::BrotliParams params;
brotli::BrotliCompressor compressor(params);
compressor.WriteStreamHeader();
while (!input_end) {
block_size = fread(input_buffer, 1, max_block_size, fin);
if (block_size == 0) {
input_end = true;
}
size_t output_size = max_output_size;
compressor.WriteMetaBlock(block_size, input_buffer, input_end,
&output_size, output_buffer);
if (fwrite(output_buffer, output_size, 1, fout) != 1) {
perror("fwrite");
brotli::BrotliFileIn in(fin, 1 << 16);
brotli::BrotliFileOut out(fout);
if (!BrotliCompress(params, &in, &out)) {
fprintf(stderr, "compression failed\n");
unlink(output_path);
exit(1);
}
}
delete[] input_buffer;
delete[] output_buffer;
}
if (fclose(fin) != 0) {
perror("fclose");
exit(1);