AuroraMiddleware/brotli
1
0
Fork 0
mirror of https://github.com/google/brotli.git synced 2024-09-19 14:19:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update encoder (#504)

Browse Source 
* pull `BROTLI_MAX_BACKWARD_LIMIT` to constants
 * split generic and Zopfli backward references code
 * pull hashers init and stitch invocation to encoder
 * make `dictionary_hash` a compilation unit
 * add `size hint` parameter
 * add new hasher
 * use `size hint` to pick new hasher for q4
 * modernize clz guard (fix #495)
 * move `hash to binary tree` to separate file
 * add `Initialize` and `Cleanup` to all hashers
 * do not raise OOM if malloc(0) == NULL (fix #500)
This commit is contained in:
Eugene Kliuchnikov 2017-02-06 14:20:43 +01:00 committed by GitHub
parent 8d3fdc1dfe
commit 11df843cf0
25 changed files with 2483 additions and 2307 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
CMakeLists.txt
View File

@ -134,12 +134,14 @@ add_library(brotlidec
dec/state.c)
add_library(brotlienc
enc/backward_references.c
enc/backward_references_hq.c
enc/bit_cost.c
enc/block_splitter.c
enc/brotli_bit_stream.c
enc/cluster.c
enc/compress_fragment.c
enc/compress_fragment_two_pass.c
enc/dictionary_hash.c
enc/encode.c
enc/entropy_encode.c
enc/histogram.c

1
common/constants.h
View File

@ -50,5 +50,6 @@
/* Number of slack bytes for window size. Don't confuse
with BROTLI_NUM_DISTANCE_SHORT_CODES. */
#define BROTLI_WINDOW_GAP 16
#define BROTLI_MAX_BACKWARD_LIMIT(W) (((size_t)1 << (W)) - BROTLI_WINDOW_GAP)
#endif /* BROTLI_COMMON_CONSTANTS_H_ */

790
enc/backward_references.c
View File

@ -8,205 +8,17 @@
#include "./backward_references.h"
#include <string.h> /* memcpy, memset */
#include "../common/constants.h"
#include <brotli/types.h>
#include "./command.h"
#include "./fast_log.h"
#include "./find_match_length.h"
#include "./literal_cost.h"
#include "./memory.h"
#include "./port.h"
#include "./prefix.h"
#include "./quality.h"
#if defined(__cplusplus) || defined(c_plusplus)
extern "C" {
#endif
static const float kInfinity = 1.7e38f; /* ~= 2 ^ 127 */
void BrotliInitZopfliNodes(ZopfliNode* array, size_t length) {
ZopfliNode stub;
size_t i;
stub.length = 1;
stub.distance = 0;
stub.insert_length = 0;
stub.u.cost = kInfinity;
for (i = 0; i < length; ++i) array[i] = stub;
}
static BROTLI_INLINE uint32_t ZopfliNodeCopyLength(const ZopfliNode* self) {
return self->length & 0xffffff;
}
static BROTLI_INLINE uint32_t ZopfliNodeLengthCode(const ZopfliNode* self) {
const uint32_t modifier = self->length >> 24;
return ZopfliNodeCopyLength(self) + 9u - modifier;
}
static BROTLI_INLINE uint32_t ZopfliNodeCopyDistance(const ZopfliNode* self) {
return self->distance & 0x1ffffff;
}
static BROTLI_INLINE uint32_t ZopfliNodeDistanceCode(const ZopfliNode* self) {
const uint32_t short_code = self->distance >> 25;
return short_code == 0 ?
ZopfliNodeCopyDistance(self) + BROTLI_NUM_DISTANCE_SHORT_CODES - 1 :
short_code - 1;
}
static BROTLI_INLINE uint32_t ZopfliNodeCommandLength(const ZopfliNode* self) {
return ZopfliNodeCopyLength(self) + self->insert_length;
}
/* Histogram based cost model for zopflification. */
typedef struct ZopfliCostModel {
/* The insert and copy length symbols. */
float cost_cmd_[BROTLI_NUM_COMMAND_SYMBOLS];
float cost_dist_[BROTLI_NUM_DISTANCE_SYMBOLS];
/* Cumulative costs of literals per position in the stream. */
float* literal_costs_;
float min_cost_cmd_;
size_t num_bytes_;
} ZopfliCostModel;
static void InitZopfliCostModel(
MemoryManager* m, ZopfliCostModel* self, size_t num_bytes) {
self->num_bytes_ = num_bytes;
self->literal_costs_ = BROTLI_ALLOC(m, float, num_bytes + 2);
if (BROTLI_IS_OOM(m)) return;
}
static void CleanupZopfliCostModel(MemoryManager* m, ZopfliCostModel* self) {
BROTLI_FREE(m, self->literal_costs_);
}
static void SetCost(const uint32_t* histogram, size_t histogram_size,
float* cost) {
size_t sum = 0;
float log2sum;
size_t i;
for (i = 0; i < histogram_size; i++) {
sum += histogram[i];
}
log2sum = (float)FastLog2(sum);
for (i = 0; i < histogram_size; i++) {
if (histogram[i] == 0) {
cost[i] = log2sum + 2;
continue;
}
/* Shannon bits for this symbol. */
cost[i] = log2sum - (float)FastLog2(histogram[i]);
/* Cannot be coded with less than 1 bit */
if (cost[i] < 1) cost[i] = 1;
}
}
static void ZopfliCostModelSetFromCommands(ZopfliCostModel* self,
size_t position,
const uint8_t* ringbuffer,
size_t ringbuffer_mask,
const Command* commands,
size_t num_commands,
size_t last_insert_len) {
uint32_t histogram_literal[BROTLI_NUM_LITERAL_SYMBOLS];
uint32_t histogram_cmd[BROTLI_NUM_COMMAND_SYMBOLS];
uint32_t histogram_dist[BROTLI_NUM_DISTANCE_SYMBOLS];
float cost_literal[BROTLI_NUM_LITERAL_SYMBOLS];
size_t pos = position - last_insert_len;
float min_cost_cmd = kInfinity;
size_t i;
float* cost_cmd = self->cost_cmd_;
memset(histogram_literal, 0, sizeof(histogram_literal));
memset(histogram_cmd, 0, sizeof(histogram_cmd));
memset(histogram_dist, 0, sizeof(histogram_dist));
for (i = 0; i < num_commands; i++) {
size_t inslength = commands[i].insert_len_;
size_t copylength = CommandCopyLen(&commands[i]);
size_t distcode = commands[i].dist_prefix_;
size_t cmdcode = commands[i].cmd_prefix_;
size_t j;
histogram_cmd[cmdcode]++;
if (cmdcode >= 128) histogram_dist[distcode]++;
for (j = 0; j < inslength; j++) {
histogram_literal[ringbuffer[(pos + j) & ringbuffer_mask]]++;
}
pos += inslength + copylength;
}
SetCost(histogram_literal, BROTLI_NUM_LITERAL_SYMBOLS, cost_literal);
SetCost(histogram_cmd, BROTLI_NUM_COMMAND_SYMBOLS, cost_cmd);
SetCost(histogram_dist, BROTLI_NUM_DISTANCE_SYMBOLS, self->cost_dist_);
for (i = 0; i < BROTLI_NUM_COMMAND_SYMBOLS; ++i) {
min_cost_cmd = BROTLI_MIN(float, min_cost_cmd, cost_cmd[i]);
}
self->min_cost_cmd_ = min_cost_cmd;
{
float* literal_costs = self->literal_costs_;
size_t num_bytes = self->num_bytes_;
literal_costs[0] = 0.0;
for (i = 0; i < num_bytes; ++i) {
literal_costs[i + 1] = literal_costs[i] +
cost_literal[ringbuffer[(position + i) & ringbuffer_mask]];
}
}
}
static void ZopfliCostModelSetFromLiteralCosts(ZopfliCostModel* self,
size_t position,
const uint8_t* ringbuffer,
size_t ringbuffer_mask) {
float* literal_costs = self->literal_costs_;
float* cost_dist = self->cost_dist_;
float* cost_cmd = self->cost_cmd_;
size_t num_bytes = self->num_bytes_;
size_t i;
BrotliEstimateBitCostsForLiterals(position, num_bytes, ringbuffer_mask,
ringbuffer, &literal_costs[1]);
literal_costs[0] = 0.0;
for (i = 0; i < num_bytes; ++i) {
literal_costs[i + 1] += literal_costs[i];
}
for (i = 0; i < BROTLI_NUM_COMMAND_SYMBOLS; ++i) {
cost_cmd[i] = (float)FastLog2(11 + (uint32_t)i);
}
for (i = 0; i < BROTLI_NUM_DISTANCE_SYMBOLS; ++i) {
cost_dist[i] = (float)FastLog2(20 + (uint32_t)i);
}
self->min_cost_cmd_ = (float)FastLog2(11);
}
static BROTLI_INLINE float ZopfliCostModelGetCommandCost(
const ZopfliCostModel* self, uint16_t cmdcode) {
return self->cost_cmd_[cmdcode];
}
static BROTLI_INLINE float ZopfliCostModelGetDistanceCost(
const ZopfliCostModel* self, size_t distcode) {
return self->cost_dist_[distcode];
}
static BROTLI_INLINE float ZopfliCostModelGetLiteralCosts(
const ZopfliCostModel* self, size_t from, size_t to) {
return self->literal_costs_[to] - self->literal_costs_[from];
}
static BROTLI_INLINE float ZopfliCostModelGetMinCostCmd(
const ZopfliCostModel* self) {
return self->min_cost_cmd_;
}
static BROTLI_INLINE size_t ComputeDistanceCode(size_t distance,
size_t max_distance,
const int* dist_cache) {
@ -231,462 +43,6 @@ static BROTLI_INLINE size_t ComputeDistanceCode(size_t distance,
return distance + BROTLI_NUM_DISTANCE_SHORT_CODES - 1;
}
/* REQUIRES: len >= 2, start_pos <= pos */
/* REQUIRES: cost < kInfinity, nodes[start_pos].cost < kInfinity */
/* Maintains the "ZopfliNode array invariant". */
static BROTLI_INLINE void UpdateZopfliNode(ZopfliNode* nodes, size_t pos,
size_t start_pos, size_t len, size_t len_code, size_t dist,
size_t short_code, float cost) {
ZopfliNode* next = &nodes[pos + len];
next->length = (uint32_t)(len | ((len + 9u - len_code) << 24));
next->distance = (uint32_t)(dist | (short_code << 25));
next->insert_length = (uint32_t)(pos - start_pos);
next->u.cost = cost;
}
typedef struct PosData {
size_t pos;
int distance_cache[4];
float costdiff;
float cost;
} PosData;
/* Maintains the smallest 8 cost difference together with their positions */
typedef struct StartPosQueue {
PosData q_[8];
size_t idx_;
} StartPosQueue;
static BROTLI_INLINE void InitStartPosQueue(StartPosQueue* self) {
self->idx_ = 0;
}
static size_t StartPosQueueSize(const StartPosQueue* self) {
return BROTLI_MIN(size_t, self->idx_, 8);
}
static void StartPosQueuePush(StartPosQueue* self, const PosData* posdata) {
size_t offset = ~(self->idx_++) & 7;
size_t len = StartPosQueueSize(self);
size_t i;
PosData* q = self->q_;
q[offset] = *posdata;
/* Restore the sorted order. In the list of |len| items at most |len - 1|
adjacent element comparisons / swaps are required. */
for (i = 1; i < len; ++i) {
if (q[offset & 7].costdiff > q[(offset + 1) & 7].costdiff) {
BROTLI_SWAP(PosData, q, offset & 7, (offset + 1) & 7);
}
++offset;
}
}
static const PosData* StartPosQueueAt(const StartPosQueue* self, size_t k) {
return &self->q_[(k - self->idx_) & 7];
}
/* Returns the minimum possible copy length that can improve the cost of any */
/* future position. */
static size_t ComputeMinimumCopyLength(const float start_cost,
const ZopfliNode* nodes,
const size_t num_bytes,
const size_t pos) {
/* Compute the minimum possible cost of reaching any future position. */
float min_cost = start_cost;
size_t len = 2;
size_t next_len_bucket = 4;
size_t next_len_offset = 10;
while (pos + len <= num_bytes && nodes[pos + len].u.cost <= min_cost) {
/* We already reached (pos + len) with no more cost than the minimum
possible cost of reaching anything from this pos, so there is no point in
looking for lengths <= len. */
++len;
if (len == next_len_offset) {
/* We reached the next copy length code bucket, so we add one more
extra bit to the minimum cost. */
min_cost += 1.0f;
next_len_offset += next_len_bucket;
next_len_bucket *= 2;
}
}
return len;
}
/* REQUIRES: nodes[pos].cost < kInfinity
REQUIRES: nodes[0..pos] satisfies that "ZopfliNode array invariant". */
static uint32_t ComputeDistanceShortcut(const size_t block_start,
const size_t pos,
const size_t max_backward,
const ZopfliNode* nodes) {
const size_t clen = ZopfliNodeCopyLength(&nodes[pos]);
const size_t ilen = nodes[pos].insert_length;
const size_t dist = ZopfliNodeCopyDistance(&nodes[pos]);
/* Since |block_start + pos| is the end position of the command, the copy part
starts from |block_start + pos - clen|. Distances that are greater than
this or greater than |max_backward| are static dictionary references, and
do not update the last distances. Also distance code 0 (last distance)
does not update the last distances. */
if (pos == 0) {
return 0;
} else if (dist + clen <= block_start + pos &&
dist <= max_backward &&
ZopfliNodeDistanceCode(&nodes[pos]) > 0) {
return (uint32_t)pos;
} else {
return nodes[pos - clen - ilen].u.shortcut;
}
}
/* Fills in dist_cache[0..3] with the last four distances (as defined by
Section 4. of the Spec) that would be used at (block_start + pos) if we
used the shortest path of commands from block_start, computed from
nodes[0..pos]. The last four distances at block_start are in
starting_dist_cache[0..3].
REQUIRES: nodes[pos].cost < kInfinity
REQUIRES: nodes[0..pos] satisfies that "ZopfliNode array invariant". */
static void ComputeDistanceCache(const size_t pos,
const int* starting_dist_cache,
const ZopfliNode* nodes,
int* dist_cache) {
int idx = 0;
size_t p = nodes[pos].u.shortcut;
while (idx < 4 && p > 0) {
const size_t ilen = nodes[p].insert_length;
const size_t clen = ZopfliNodeCopyLength(&nodes[p]);
const size_t dist = ZopfliNodeCopyDistance(&nodes[p]);
dist_cache[idx++] = (int)dist;
/* Because of prerequisite, p >= clen + ilen >= 2. */
p = nodes[p - clen - ilen].u.shortcut;
}
for (; idx < 4; ++idx) {
dist_cache[idx] = *starting_dist_cache++;
}
}
/* Maintains "ZopfliNode array invariant" and pushes node to the queue, if it
is eligible. */
static void EvaluateNode(
const size_t block_start, const size_t pos, const size_t max_backward_limit,
const int* starting_dist_cache, const ZopfliCostModel* model,
StartPosQueue* queue, ZopfliNode* nodes) {
/* Save cost, because ComputeDistanceCache invalidates it. */
float node_cost = nodes[pos].u.cost;
nodes[pos].u.shortcut = ComputeDistanceShortcut(
block_start, pos, max_backward_limit, nodes);
if (node_cost <= ZopfliCostModelGetLiteralCosts(model, 0, pos)) {
PosData posdata;
posdata.pos = pos;
posdata.cost = node_cost;
posdata.costdiff = node_cost -
ZopfliCostModelGetLiteralCosts(model, 0, pos);
ComputeDistanceCache(
pos, starting_dist_cache, nodes, posdata.distance_cache);
StartPosQueuePush(queue, &posdata);
}
}
/* Returns longest copy length. */
static size_t UpdateNodes(
const size_t num_bytes, const size_t block_start, const size_t pos,
const uint8_t* ringbuffer, const size_t ringbuffer_mask,
const BrotliEncoderParams* params, const size_t max_backward_limit,
const int* starting_dist_cache, const size_t num_matches,
const BackwardMatch* matches, const ZopfliCostModel* model,
StartPosQueue* queue, ZopfliNode* nodes) {
const size_t cur_ix = block_start + pos;
const size_t cur_ix_masked = cur_ix & ringbuffer_mask;
const size_t max_distance = BROTLI_MIN(size_t, cur_ix, max_backward_limit);
const size_t max_len = num_bytes - pos;
const size_t max_zopfli_len = MaxZopfliLen(params);
const size_t max_iters = MaxZopfliCandidates(params);
size_t min_len;
size_t result = 0;
size_t k;
EvaluateNode(block_start, pos, max_backward_limit, starting_dist_cache, model,
queue, nodes);
{
const PosData* posdata = StartPosQueueAt(queue, 0);
float min_cost = (posdata->cost + ZopfliCostModelGetMinCostCmd(model) +
ZopfliCostModelGetLiteralCosts(model, posdata->pos, pos));
min_len = ComputeMinimumCopyLength(min_cost, nodes, num_bytes, pos);
}
/* Go over the command starting positions in order of increasing cost
difference. */
for (k = 0; k < max_iters && k < StartPosQueueSize(queue); ++k) {
const PosData* posdata = StartPosQueueAt(queue, k);
const size_t start = posdata->pos;
const uint16_t inscode = GetInsertLengthCode(pos - start);
const float start_costdiff = posdata->costdiff;
const float base_cost = start_costdiff + (float)GetInsertExtra(inscode) +
ZopfliCostModelGetLiteralCosts(model, 0, pos);
/* Look for last distance matches using the distance cache from this
starting position. */
size_t best_len = min_len - 1;
size_t j = 0;
for (; j < BROTLI_NUM_DISTANCE_SHORT_CODES && best_len < max_len; ++j) {
const size_t idx = kDistanceCacheIndex[j];
const size_t backward =
(size_t)(posdata->distance_cache[idx] + kDistanceCacheOffset[j]);
size_t prev_ix = cur_ix - backward;
if (prev_ix >= cur_ix) {
continue;
}
if (BROTLI_PREDICT_FALSE(backward > max_distance)) {
continue;
}
prev_ix &= ringbuffer_mask;
if (cur_ix_masked + best_len > ringbuffer_mask ||
prev_ix + best_len > ringbuffer_mask ||
ringbuffer[cur_ix_masked + best_len] !=
ringbuffer[prev_ix + best_len]) {
continue;
}
{
const size_t len =
FindMatchLengthWithLimit(&ringbuffer[prev_ix],
&ringbuffer[cur_ix_masked],
max_len);
const float dist_cost = base_cost +
ZopfliCostModelGetDistanceCost(model, j);
size_t l;
for (l = best_len + 1; l <= len; ++l) {
const uint16_t copycode = GetCopyLengthCode(l);
const uint16_t cmdcode =
CombineLengthCodes(inscode, copycode, j == 0);
const float cost = (cmdcode < 128 ? base_cost : dist_cost) +
(float)GetCopyExtra(copycode) +
ZopfliCostModelGetCommandCost(model, cmdcode);
if (cost < nodes[pos + l].u.cost) {
UpdateZopfliNode(nodes, pos, start, l, l, backward, j + 1, cost);
result = BROTLI_MAX(size_t, result, l);
}
best_len = l;
}
}
}
/* At higher iterations look only for new last distance matches, since
looking only for new command start positions with the same distances
does not help much. */
if (k >= 2) continue;
{
/* Loop through all possible copy lengths at this position. */
size_t len = min_len;
for (j = 0; j < num_matches; ++j) {
BackwardMatch match = matches[j];
size_t dist = match.distance;
BROTLI_BOOL is_dictionary_match = TO_BROTLI_BOOL(dist > max_distance);
/* We already tried all possible last distance matches, so we can use
normal distance code here. */
size_t dist_code = dist + BROTLI_NUM_DISTANCE_SHORT_CODES - 1;
uint16_t dist_symbol;
uint32_t distextra;
uint32_t distnumextra;
float dist_cost;
size_t max_match_len;
PrefixEncodeCopyDistance(dist_code, 0, 0, &dist_symbol, &distextra);
distnumextra = distextra >> 24;
dist_cost = base_cost + (float)distnumextra +
ZopfliCostModelGetDistanceCost(model, dist_symbol);
/* Try all copy lengths up until the maximum copy length corresponding
to this distance. If the distance refers to the static dictionary, or
the maximum length is long enough, try only one maximum length. */
max_match_len = BackwardMatchLength(&match);
if (len < max_match_len &&
(is_dictionary_match || max_match_len > max_zopfli_len)) {
len = max_match_len;
}
for (; len <= max_match_len; ++len) {
const size_t len_code =
is_dictionary_match ? BackwardMatchLengthCode(&match) : len;
const uint16_t copycode = GetCopyLengthCode(len_code);
const uint16_t cmdcode = CombineLengthCodes(inscode, copycode, 0);
const float cost = dist_cost + (float)GetCopyExtra(copycode) +
ZopfliCostModelGetCommandCost(model, cmdcode);
if (cost < nodes[pos + len].u.cost) {
UpdateZopfliNode(nodes, pos, start, len, len_code, dist, 0, cost);
result = BROTLI_MAX(size_t, result, len);
}
}
}
}
}
return result;
}
static size_t ComputeShortestPathFromNodes(size_t num_bytes,
ZopfliNode* nodes) {
size_t index = num_bytes;
size_t num_commands = 0;
while (nodes[index].insert_length == 0 && nodes[index].length == 1) --index;
nodes[index].u.next = BROTLI_UINT32_MAX;
while (index != 0) {
size_t len = ZopfliNodeCommandLength(&nodes[index]);
index -= len;
nodes[index].u.next = (uint32_t)len;
num_commands++;
}
return num_commands;
}
void BrotliZopfliCreateCommands(const size_t num_bytes,
const size_t block_start,
const size_t max_backward_limit,
const ZopfliNode* nodes,
int* dist_cache,
size_t* last_insert_len,
Command* commands,
size_t* num_literals) {
size_t pos = 0;
uint32_t offset = nodes[0].u.next;
size_t i;
for (i = 0; offset != BROTLI_UINT32_MAX; i++) {
const ZopfliNode* next = &nodes[pos + offset];
size_t copy_length = ZopfliNodeCopyLength(next);
size_t insert_length = next->insert_length;
pos += insert_length;
offset = next->u.next;
if (i == 0) {
insert_length += *last_insert_len;
*last_insert_len = 0;
}
{
size_t distance = ZopfliNodeCopyDistance(next);
size_t len_code = ZopfliNodeLengthCode(next);
size_t max_distance =
BROTLI_MIN(size_t, block_start + pos, max_backward_limit);
BROTLI_BOOL is_dictionary = TO_BROTLI_BOOL(distance > max_distance);
size_t dist_code = ZopfliNodeDistanceCode(next);
InitCommand(
&commands[i], insert_length, copy_length, len_code, dist_code);
if (!is_dictionary && dist_code > 0) {
dist_cache[3] = dist_cache[2];
dist_cache[2] = dist_cache[1];
dist_cache[1] = dist_cache[0];
dist_cache[0] = (int)distance;
}
}
*num_literals += insert_length;
pos += copy_length;
}
*last_insert_len += num_bytes - pos;
}
static size_t ZopfliIterate(size_t num_bytes,
size_t position,
const uint8_t* ringbuffer,
size_t ringbuffer_mask,
const BrotliEncoderParams* params,
const size_t max_backward_limit,
const int* dist_cache,
const ZopfliCostModel* model,
const uint32_t* num_matches,
const BackwardMatch* matches,
ZopfliNode* nodes) {
const size_t max_zopfli_len = MaxZopfliLen(params);
StartPosQueue queue;
size_t cur_match_pos = 0;
size_t i;
nodes[0].length = 0;
nodes[0].u.cost = 0;
InitStartPosQueue(&queue);
for (i = 0; i + 3 < num_bytes; i++) {
size_t skip = UpdateNodes(num_bytes, position, i, ringbuffer,
ringbuffer_mask, params, max_backward_limit, dist_cache,
num_matches[i], &matches[cur_match_pos], model, &queue, nodes);
if (skip < BROTLI_LONG_COPY_QUICK_STEP) skip = 0;
cur_match_pos += num_matches[i];
if (num_matches[i] == 1 &&
BackwardMatchLength(&matches[cur_match_pos - 1]) > max_zopfli_len) {
skip = BROTLI_MAX(size_t,
BackwardMatchLength(&matches[cur_match_pos - 1]), skip);
}
if (skip > 1) {
skip--;
while (skip) {
i++;
if (i + 3 >= num_bytes) break;
EvaluateNode(
position, i, max_backward_limit, dist_cache, model, &queue, nodes);
cur_match_pos += num_matches[i];
skip--;
}
}
}
return ComputeShortestPathFromNodes(num_bytes, nodes);
}
size_t BrotliZopfliComputeShortestPath(MemoryManager* m,
size_t num_bytes,
size_t position,
const uint8_t* ringbuffer,
size_t ringbuffer_mask,
const BrotliEncoderParams* params,
const size_t max_backward_limit,
const int* dist_cache,
H10* hasher,
ZopfliNode* nodes) {
const size_t max_zopfli_len = MaxZopfliLen(params);
ZopfliCostModel model;
StartPosQueue queue;
BackwardMatch matches[MAX_NUM_MATCHES_H10];
const size_t store_end = num_bytes >= StoreLookaheadH10() ?
position + num_bytes - StoreLookaheadH10() + 1 : position;
size_t i;
nodes[0].length = 0;
nodes[0].u.cost = 0;
InitZopfliCostModel(m, &model, num_bytes);
if (BROTLI_IS_OOM(m)) return 0;
ZopfliCostModelSetFromLiteralCosts(
&model, position, ringbuffer, ringbuffer_mask);
InitStartPosQueue(&queue);
for (i = 0; i + HashTypeLengthH10() - 1 < num_bytes; i++) {
const size_t pos = position + i;
const size_t max_distance = BROTLI_MIN(size_t, pos, max_backward_limit);
size_t num_matches = FindAllMatchesH10(hasher, ringbuffer, ringbuffer_mask,
pos, num_bytes - i, max_distance, params, matches);
size_t skip;
if (num_matches > 0 &&
BackwardMatchLength(&matches[num_matches - 1]) > max_zopfli_len) {
matches[0] = matches[num_matches - 1];
num_matches = 1;
}
skip = UpdateNodes(num_bytes, position, i, ringbuffer, ringbuffer_mask,
params, max_backward_limit, dist_cache, num_matches, matches, &model,
&queue, nodes);
if (skip < BROTLI_LONG_COPY_QUICK_STEP) skip = 0;
if (num_matches == 1 && BackwardMatchLength(&matches[0]) > max_zopfli_len) {
skip = BROTLI_MAX(size_t, BackwardMatchLength(&matches[0]), skip);
}
if (skip > 1) {
/* Add the tail of the copy to the hasher. */
StoreRangeH10(hasher, ringbuffer, ringbuffer_mask, pos + 1, BROTLI_MIN(
size_t, pos + skip, store_end));
skip--;
while (skip) {
i++;
if (i + HashTypeLengthH10() - 1 >= num_bytes) break;
EvaluateNode(
position, i, max_backward_limit, dist_cache, &model, &queue, nodes);
skip--;
}
}
}
CleanupZopfliCostModel(m, &model);
return ComputeShortestPathFromNodes(num_bytes, nodes);
}
#define EXPAND_CAT(a, b) CAT(a, b)
#define CAT(a, b) a ## b
#define FN(X) EXPAND_CAT(X, HASHER())
@ -746,134 +102,17 @@ size_t BrotliZopfliComputeShortestPath(MemoryManager* m,
#include "./backward_references_inc.h"
#undef HASHER
#define HASHER() H54
/* NOLINTNEXTLINE(build/include) */
#include "./backward_references_inc.h"
#undef HASHER
#undef FN
#undef CAT
#undef EXPAND_CAT
static BROTLI_NOINLINE void CreateZopfliBackwardReferences(
MemoryManager* m, size_t num_bytes, size_t position, BROTLI_BOOL is_last,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, H10* hasher, int* dist_cache,
size_t* last_insert_len, Command* commands, size_t* num_commands,
size_t* num_literals) {
const size_t max_backward_limit = MaxBackwardLimit(params->lgwin);
ZopfliNode* nodes;
InitH10(m, hasher, ringbuffer, params, position, num_bytes, is_last);
if (BROTLI_IS_OOM(m)) return;
StitchToPreviousBlockH10(hasher, num_bytes, position,
ringbuffer, ringbuffer_mask);
nodes = BROTLI_ALLOC(m, ZopfliNode, num_bytes + 1);
if (BROTLI_IS_OOM(m)) return;
BrotliInitZopfliNodes(nodes, num_bytes + 1);
*num_commands += BrotliZopfliComputeShortestPath(m, num_bytes, position,
ringbuffer, ringbuffer_mask, params, max_backward_limit,
dist_cache, hasher, nodes);
if (BROTLI_IS_OOM(m)) return;
BrotliZopfliCreateCommands(num_bytes, position, max_backward_limit, nodes,
dist_cache, last_insert_len, commands, num_literals);
BROTLI_FREE(m, nodes);
}
static BROTLI_NOINLINE void CreateHqZopfliBackwardReferences(
MemoryManager* m, size_t num_bytes, size_t position, BROTLI_BOOL is_last,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, H10* hasher, int* dist_cache,
size_t* last_insert_len, Command* commands, size_t* num_commands,
size_t* num_literals) {
const size_t max_backward_limit = MaxBackwardLimit(params->lgwin);
uint32_t* num_matches = BROTLI_ALLOC(m, uint32_t, num_bytes);
size_t matches_size = 4 * num_bytes;
const size_t store_end = num_bytes >= StoreLookaheadH10() ?
position + num_bytes - StoreLookaheadH10() + 1 : position;
size_t cur_match_pos = 0;
size_t i;
size_t orig_num_literals;
size_t orig_last_insert_len;
int orig_dist_cache[4];
size_t orig_num_commands;
ZopfliCostModel model;
ZopfliNode* nodes;
BackwardMatch* matches = BROTLI_ALLOC(m, BackwardMatch, matches_size);
if (BROTLI_IS_OOM(m)) return;
InitH10(m, hasher, ringbuffer, params, position, num_bytes, is_last);
if (BROTLI_IS_OOM(m)) return;
StitchToPreviousBlockH10(hasher, num_bytes, position,
ringbuffer, ringbuffer_mask);
for (i = 0; i + HashTypeLengthH10() - 1 < num_bytes; ++i) {
const size_t pos = position + i;
size_t max_distance = BROTLI_MIN(size_t, pos, max_backward_limit);
size_t max_length = num_bytes - i;
size_t num_found_matches;
size_t cur_match_end;
size_t j;
/* Ensure that we have enough free slots. */
BROTLI_ENSURE_CAPACITY(m, BackwardMatch, matches, matches_size,
cur_match_pos + MAX_NUM_MATCHES_H10);
if (BROTLI_IS_OOM(m)) return;
num_found_matches = FindAllMatchesH10(hasher, ringbuffer, ringbuffer_mask,
pos, max_length, max_distance, params, &matches[cur_match_pos]);
cur_match_end = cur_match_pos + num_found_matches;
for (j = cur_match_pos; j + 1 < cur_match_end; ++j) {
assert(BackwardMatchLength(&matches[j]) <
BackwardMatchLength(&matches[j + 1]));
assert(matches[j].distance > max_distance ||
matches[j].distance <= matches[j + 1].distance);
}
num_matches[i] = (uint32_t)num_found_matches;
if (num_found_matches > 0) {
const size_t match_len = BackwardMatchLength(&matches[cur_match_end - 1]);
if (match_len > MAX_ZOPFLI_LEN_QUALITY_11) {
const size_t skip = match_len - 1;
matches[cur_match_pos++] = matches[cur_match_end - 1];
num_matches[i] = 1;
/* Add the tail of the copy to the hasher. */
StoreRangeH10(hasher, ringbuffer, ringbuffer_mask, pos + 1,
BROTLI_MIN(size_t, pos + match_len, store_end));
memset(&num_matches[i + 1], 0, skip * sizeof(num_matches[0]));
i += skip;
} else {
cur_match_pos = cur_match_end;
}
}
}
orig_num_literals = *num_literals;
orig_last_insert_len = *last_insert_len;
memcpy(orig_dist_cache, dist_cache, 4 * sizeof(dist_cache[0]));
orig_num_commands = *num_commands;
nodes = BROTLI_ALLOC(m, ZopfliNode, num_bytes + 1);
if (BROTLI_IS_OOM(m)) return;
InitZopfliCostModel(m, &model, num_bytes);
if (BROTLI_IS_OOM(m)) return;
for (i = 0; i < 2; i++) {
BrotliInitZopfliNodes(nodes, num_bytes + 1);
if (i == 0) {
ZopfliCostModelSetFromLiteralCosts(
&model, position, ringbuffer, ringbuffer_mask);
} else {
ZopfliCostModelSetFromCommands(&model, position, ringbuffer,
ringbuffer_mask, commands, *num_commands - orig_num_commands,
orig_last_insert_len);
}
*num_commands = orig_num_commands;
*num_literals = orig_num_literals;
*last_insert_len = orig_last_insert_len;
memcpy(dist_cache, orig_dist_cache, 4 * sizeof(dist_cache[0]));
*num_commands += ZopfliIterate(num_bytes, position, ringbuffer,
ringbuffer_mask, params, max_backward_limit, dist_cache,
&model, num_matches, matches, nodes);
BrotliZopfliCreateCommands(num_bytes, position, max_backward_limit,
nodes, dist_cache, last_insert_len, commands, num_literals);
}
CleanupZopfliCostModel(m, &model);
BROTLI_FREE(m, nodes);
BROTLI_FREE(m, matches);
BROTLI_FREE(m, num_matches);
}
void BrotliCreateBackwardReferences(MemoryManager* m,
size_t num_bytes,
void BrotliCreateBackwardReferences(size_t num_bytes,
size_t position,
BROTLI_BOOL is_last,
const uint8_t* ringbuffer,
size_t ringbuffer_mask,
const BrotliEncoderParams* params,
@ -883,24 +122,10 @@ void BrotliCreateBackwardReferences(MemoryManager* m,
Command* commands,
size_t* num_commands,
size_t* num_literals) {
if (params->quality == ZOPFLIFICATION_QUALITY) {
CreateZopfliBackwardReferences(
m, num_bytes, position, is_last, ringbuffer, ringbuffer_mask,
params, hashers->h10, dist_cache,
last_insert_len, commands, num_commands, num_literals);
return;
} else if (params->quality == HQ_ZOPFLIFICATION_QUALITY) {
CreateHqZopfliBackwardReferences(
m, num_bytes, position, is_last, ringbuffer, ringbuffer_mask,
params, hashers->h10, dist_cache,
last_insert_len, commands, num_commands, num_literals);
return;
}
switch (ChooseHasher(params)) {
#define CASE_(N) \
case N: \
CreateBackwardReferencesH ## N(m, num_bytes, position, is_last, \
CreateBackwardReferencesH ## N(num_bytes, position, \
ringbuffer, ringbuffer_mask, params, hashers->h ## N, dist_cache, \
last_insert_len, commands, num_commands, num_literals); \
break;
@ -909,7 +134,6 @@ void BrotliCreateBackwardReferences(MemoryManager* m,
default:
break;
}
if (BROTLI_IS_OOM(m)) return;
}
#if defined(__cplusplus) || defined(c_plusplus)

64
enc/backward_references.h
View File

@ -13,7 +13,6 @@
#include <brotli/types.h>
#include "./command.h"
#include "./hash.h"
#include "./memory.h"
#include "./port.h"
#include "./quality.h"
@ -26,73 +25,12 @@ extern "C" {
CreateBackwardReferences calls, and must be incremented by the amount written
by this call. */
BROTLI_INTERNAL void BrotliCreateBackwardReferences(
MemoryManager* m, size_t num_bytes, size_t position, BROTLI_BOOL is_last,
size_t num_bytes, size_t position,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, Hashers* hashers, int* dist_cache,
size_t* last_insert_len, Command* commands, size_t* num_commands,
size_t* num_literals);
typedef struct ZopfliNode {
/* best length to get up to this byte (not including this byte itself)
highest 8 bit is used to reconstruct the length code */
uint32_t length;
/* distance associated with the length
highest 7 bit contains distance short code + 1 (or zero if no short code)
*/
uint32_t distance;
/* number of literal inserts before this copy */
uint32_t insert_length;
/* This union holds information used by dynamic-programming. During forward
pass |cost| it used to store the goal function. When node is processed its
|cost| is invalidated in favor of |shortcut|. On path back-tracing pass
|next| is assigned the offset to next node on the path. */
union {
/* Smallest cost to get to this byte from the beginning, as found so far. */
float cost;
/* Offset to the next node on the path. Equals to command_length() of the
next node on the path. For last node equals to BROTLI_UINT32_MAX */
uint32_t next;
/* Node position that provides next distance for distance cache. */
uint32_t shortcut;
} u;
} ZopfliNode;
BROTLI_INTERNAL void BrotliInitZopfliNodes(ZopfliNode* array, size_t length);
/* Computes the shortest path of commands from position to at most
position + num_bytes.
On return, path->size() is the number of commands found and path[i] is the
length of the i-th command (copy length plus insert length).
Note that the sum of the lengths of all commands can be less than num_bytes.
On return, the nodes[0..num_bytes] array will have the following
"ZopfliNode array invariant":
For each i in [1..num_bytes], if nodes[i].cost < kInfinity, then
(1) nodes[i].copy_length() >= 2
(2) nodes[i].command_length() <= i and
(3) nodes[i - nodes[i].command_length()].cost < kInfinity */
BROTLI_INTERNAL size_t BrotliZopfliComputeShortestPath(
MemoryManager* m, size_t num_bytes, size_t position,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, const size_t max_backward_limit,
const int* dist_cache, H10* hasher, ZopfliNode* nodes);
BROTLI_INTERNAL void BrotliZopfliCreateCommands(const size_t num_bytes,
const size_t block_start,
const size_t max_backward_limit,
const ZopfliNode* nodes,
int* dist_cache,
size_t* last_insert_len,
Command* commands,
size_t* num_literals);
/* Maximum distance, see section 9.1. of the spec. */
static BROTLI_INLINE size_t MaxBackwardLimit(int lgwin) {
return (1u << lgwin) - BROTLI_WINDOW_GAP;
}
#if defined(__cplusplus) || defined(c_plusplus)
} /* extern "C" */
#endif

780
enc/backward_references_hq.c Executable file
View File

@ -0,0 +1,780 @@
/* Copyright 2013 Google Inc. All Rights Reserved.
Distributed under MIT license.
See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
*/
/* Function to find backward reference copies. */
#include "./backward_references_hq.h"
#include <string.h> /* memcpy, memset */
#include "../common/constants.h"
#include <brotli/types.h>
#include "./command.h"
#include "./fast_log.h"
#include "./find_match_length.h"
#include "./literal_cost.h"
#include "./memory.h"
#include "./port.h"
#include "./prefix.h"
#include "./quality.h"
#if defined(__cplusplus) || defined(c_plusplus)
extern "C" {
#endif
static const float kInfinity = 1.7e38f; /* ~= 2 ^ 127 */
void BrotliInitZopfliNodes(ZopfliNode* array, size_t length) {
ZopfliNode stub;
size_t i;
stub.length = 1;
stub.distance = 0;
stub.insert_length = 0;
stub.u.cost = kInfinity;
for (i = 0; i < length; ++i) array[i] = stub;
}
static BROTLI_INLINE uint32_t ZopfliNodeCopyLength(const ZopfliNode* self) {
return self->length & 0xffffff;
}
static BROTLI_INLINE uint32_t ZopfliNodeLengthCode(const ZopfliNode* self) {
const uint32_t modifier = self->length >> 24;
return ZopfliNodeCopyLength(self) + 9u - modifier;
}
static BROTLI_INLINE uint32_t ZopfliNodeCopyDistance(const ZopfliNode* self) {
return self->distance & 0x1ffffff;
}
static BROTLI_INLINE uint32_t ZopfliNodeDistanceCode(const ZopfliNode* self) {
const uint32_t short_code = self->distance >> 25;
return short_code == 0 ?
ZopfliNodeCopyDistance(self) + BROTLI_NUM_DISTANCE_SHORT_CODES - 1 :
short_code - 1;
}
static BROTLI_INLINE uint32_t ZopfliNodeCommandLength(const ZopfliNode* self) {
return ZopfliNodeCopyLength(self) + self->insert_length;
}
/* Histogram based cost model for zopflification. */
typedef struct ZopfliCostModel {
/* The insert and copy length symbols. */
float cost_cmd_[BROTLI_NUM_COMMAND_SYMBOLS];
float cost_dist_[BROTLI_NUM_DISTANCE_SYMBOLS];
/* Cumulative costs of literals per position in the stream. */
float* literal_costs_;
float min_cost_cmd_;
size_t num_bytes_;
} ZopfliCostModel;
static void InitZopfliCostModel(
MemoryManager* m, ZopfliCostModel* self, size_t num_bytes) {
self->num_bytes_ = num_bytes;
self->literal_costs_ = BROTLI_ALLOC(m, float, num_bytes + 2);
if (BROTLI_IS_OOM(m)) return;
}
static void CleanupZopfliCostModel(MemoryManager* m, ZopfliCostModel* self) {
BROTLI_FREE(m, self->literal_costs_);
}
static void SetCost(const uint32_t* histogram, size_t histogram_size,
float* cost) {
size_t sum = 0;
float log2sum;
size_t i;
for (i = 0; i < histogram_size; i++) {
sum += histogram[i];
}
log2sum = (float)FastLog2(sum);
for (i = 0; i < histogram_size; i++) {
if (histogram[i] == 0) {
cost[i] = log2sum + 2;
continue;
}
/* Shannon bits for this symbol. */
cost[i] = log2sum - (float)FastLog2(histogram[i]);
/* Cannot be coded with less than 1 bit */
if (cost[i] < 1) cost[i] = 1;
}
}
static void ZopfliCostModelSetFromCommands(ZopfliCostModel* self,
size_t position,
const uint8_t* ringbuffer,
size_t ringbuffer_mask,
const Command* commands,
size_t num_commands,
size_t last_insert_len) {
uint32_t histogram_literal[BROTLI_NUM_LITERAL_SYMBOLS];
uint32_t histogram_cmd[BROTLI_NUM_COMMAND_SYMBOLS];
uint32_t histogram_dist[BROTLI_NUM_DISTANCE_SYMBOLS];
float cost_literal[BROTLI_NUM_LITERAL_SYMBOLS];
size_t pos = position - last_insert_len;
float min_cost_cmd = kInfinity;
size_t i;
float* cost_cmd = self->cost_cmd_;
memset(histogram_literal, 0, sizeof(histogram_literal));
memset(histogram_cmd, 0, sizeof(histogram_cmd));
memset(histogram_dist, 0, sizeof(histogram_dist));
for (i = 0; i < num_commands; i++) {
size_t inslength = commands[i].insert_len_;
size_t copylength = CommandCopyLen(&commands[i]);
size_t distcode = commands[i].dist_prefix_;
size_t cmdcode = commands[i].cmd_prefix_;
size_t j;
histogram_cmd[cmdcode]++;
if (cmdcode >= 128) histogram_dist[distcode]++;
for (j = 0; j < inslength; j++) {
histogram_literal[ringbuffer[(pos + j) & ringbuffer_mask]]++;
}
pos += inslength + copylength;
}
SetCost(histogram_literal, BROTLI_NUM_LITERAL_SYMBOLS, cost_literal);
SetCost(histogram_cmd, BROTLI_NUM_COMMAND_SYMBOLS, cost_cmd);
SetCost(histogram_dist, BROTLI_NUM_DISTANCE_SYMBOLS, self->cost_dist_);
for (i = 0; i < BROTLI_NUM_COMMAND_SYMBOLS; ++i) {
min_cost_cmd = BROTLI_MIN(float, min_cost_cmd, cost_cmd[i]);
}
self->min_cost_cmd_ = min_cost_cmd;
{
float* literal_costs = self->literal_costs_;
size_t num_bytes = self->num_bytes_;
literal_costs[0] = 0.0;
for (i = 0; i < num_bytes; ++i) {
literal_costs[i + 1] = literal_costs[i] +
cost_literal[ringbuffer[(position + i) & ringbuffer_mask]];
}
}
}
static void ZopfliCostModelSetFromLiteralCosts(ZopfliCostModel* self,
size_t position,
const uint8_t* ringbuffer,
size_t ringbuffer_mask) {
float* literal_costs = self->literal_costs_;
float* cost_dist = self->cost_dist_;
float* cost_cmd = self->cost_cmd_;
size_t num_bytes = self->num_bytes_;
size_t i;
BrotliEstimateBitCostsForLiterals(position, num_bytes, ringbuffer_mask,
ringbuffer, &literal_costs[1]);
literal_costs[0] = 0.0;
for (i = 0; i < num_bytes; ++i) {
literal_costs[i + 1] += literal_costs[i];
}
for (i = 0; i < BROTLI_NUM_COMMAND_SYMBOLS; ++i) {
cost_cmd[i] = (float)FastLog2(11 + (uint32_t)i);
}
for (i = 0; i < BROTLI_NUM_DISTANCE_SYMBOLS; ++i) {
cost_dist[i] = (float)FastLog2(20 + (uint32_t)i);
}
self->min_cost_cmd_ = (float)FastLog2(11);
}
static BROTLI_INLINE float ZopfliCostModelGetCommandCost(
const ZopfliCostModel* self, uint16_t cmdcode) {
return self->cost_cmd_[cmdcode];
}
static BROTLI_INLINE float ZopfliCostModelGetDistanceCost(
const ZopfliCostModel* self, size_t distcode) {
return self->cost_dist_[distcode];
}
static BROTLI_INLINE float ZopfliCostModelGetLiteralCosts(
const ZopfliCostModel* self, size_t from, size_t to) {
return self->literal_costs_[to] - self->literal_costs_[from];
}
static BROTLI_INLINE float ZopfliCostModelGetMinCostCmd(
const ZopfliCostModel* self) {
return self->min_cost_cmd_;
}
/* REQUIRES: len >= 2, start_pos <= pos */
/* REQUIRES: cost < kInfinity, nodes[start_pos].cost < kInfinity */
/* Maintains the "ZopfliNode array invariant". */
static BROTLI_INLINE void UpdateZopfliNode(ZopfliNode* nodes, size_t pos,
size_t start_pos, size_t len, size_t len_code, size_t dist,
size_t short_code, float cost) {
ZopfliNode* next = &nodes[pos + len];
next->length = (uint32_t)(len | ((len + 9u - len_code) << 24));
next->distance = (uint32_t)(dist | (short_code << 25));
next->insert_length = (uint32_t)(pos - start_pos);
next->u.cost = cost;
}
typedef struct PosData {
size_t pos;
int distance_cache[4];
float costdiff;
float cost;
} PosData;
/* Maintains the smallest 8 cost difference together with their positions */
typedef struct StartPosQueue {
PosData q_[8];
size_t idx_;
} StartPosQueue;
static BROTLI_INLINE void InitStartPosQueue(StartPosQueue* self) {
self->idx_ = 0;
}
static size_t StartPosQueueSize(const StartPosQueue* self) {
return BROTLI_MIN(size_t, self->idx_, 8);
}
static void StartPosQueuePush(StartPosQueue* self, const PosData* posdata) {
size_t offset = ~(self->idx_++) & 7;
size_t len = StartPosQueueSize(self);
size_t i;
PosData* q = self->q_;
q[offset] = *posdata;
/* Restore the sorted order. In the list of |len| items at most |len - 1|
adjacent element comparisons / swaps are required. */
for (i = 1; i < len; ++i) {
if (q[offset & 7].costdiff > q[(offset + 1) & 7].costdiff) {
BROTLI_SWAP(PosData, q, offset & 7, (offset + 1) & 7);
}
++offset;
}
}
static const PosData* StartPosQueueAt(const StartPosQueue* self, size_t k) {
return &self->q_[(k - self->idx_) & 7];
}
/* Returns the minimum possible copy length that can improve the cost of any */
/* future position. */
static size_t ComputeMinimumCopyLength(const float start_cost,
const ZopfliNode* nodes,
const size_t num_bytes,
const size_t pos) {
/* Compute the minimum possible cost of reaching any future position. */
float min_cost = start_cost;
size_t len = 2;
size_t next_len_bucket = 4;
size_t next_len_offset = 10;
while (pos + len <= num_bytes && nodes[pos + len].u.cost <= min_cost) {
/* We already reached (pos + len) with no more cost than the minimum
possible cost of reaching anything from this pos, so there is no point in
looking for lengths <= len. */
++len;
if (len == next_len_offset) {
/* We reached the next copy length code bucket, so we add one more
extra bit to the minimum cost. */
min_cost += 1.0f;
next_len_offset += next_len_bucket;
next_len_bucket *= 2;
}
}
return len;
}
/* REQUIRES: nodes[pos].cost < kInfinity
REQUIRES: nodes[0..pos] satisfies that "ZopfliNode array invariant". */
static uint32_t ComputeDistanceShortcut(const size_t block_start,
const size_t pos,
const size_t max_backward,
const ZopfliNode* nodes) {
const size_t clen = ZopfliNodeCopyLength(&nodes[pos]);
const size_t ilen = nodes[pos].insert_length;
const size_t dist = ZopfliNodeCopyDistance(&nodes[pos]);
/* Since |block_start + pos| is the end position of the command, the copy part
starts from |block_start + pos - clen|. Distances that are greater than
this or greater than |max_backward| are static dictionary references, and
do not update the last distances. Also distance code 0 (last distance)
does not update the last distances. */
if (pos == 0) {
return 0;
} else if (dist + clen <= block_start + pos &&
dist <= max_backward &&
ZopfliNodeDistanceCode(&nodes[pos]) > 0) {
return (uint32_t)pos;
} else {
return nodes[pos - clen - ilen].u.shortcut;
}
}
/* Fills in dist_cache[0..3] with the last four distances (as defined by
Section 4. of the Spec) that would be used at (block_start + pos) if we
used the shortest path of commands from block_start, computed from
nodes[0..pos]. The last four distances at block_start are in
starting_dist_cache[0..3].
REQUIRES: nodes[pos].cost < kInfinity
REQUIRES: nodes[0..pos] satisfies that "ZopfliNode array invariant". */
static void ComputeDistanceCache(const size_t pos,
const int* starting_dist_cache,
const ZopfliNode* nodes,
int* dist_cache) {
int idx = 0;
size_t p = nodes[pos].u.shortcut;
while (idx < 4 && p > 0) {
const size_t ilen = nodes[p].insert_length;
const size_t clen = ZopfliNodeCopyLength(&nodes[p]);
const size_t dist = ZopfliNodeCopyDistance(&nodes[p]);
dist_cache[idx++] = (int)dist;
/* Because of prerequisite, p >= clen + ilen >= 2. */
p = nodes[p - clen - ilen].u.shortcut;
}
for (; idx < 4; ++idx) {
dist_cache[idx] = *starting_dist_cache++;
}
}
/* Maintains "ZopfliNode array invariant" and pushes node to the queue, if it
is eligible. */
static void EvaluateNode(
const size_t block_start, const size_t pos, const size_t max_backward_limit,
const int* starting_dist_cache, const ZopfliCostModel* model,
StartPosQueue* queue, ZopfliNode* nodes) {
/* Save cost, because ComputeDistanceCache invalidates it. */
float node_cost = nodes[pos].u.cost;
nodes[pos].u.shortcut = ComputeDistanceShortcut(
block_start, pos, max_backward_limit, nodes);
if (node_cost <= ZopfliCostModelGetLiteralCosts(model, 0, pos)) {
PosData posdata;
posdata.pos = pos;
posdata.cost = node_cost;
posdata.costdiff = node_cost -
ZopfliCostModelGetLiteralCosts(model, 0, pos);
ComputeDistanceCache(
pos, starting_dist_cache, nodes, posdata.distance_cache);
StartPosQueuePush(queue, &posdata);
}
}
/* Returns longest copy length. */
static size_t UpdateNodes(
const size_t num_bytes, const size_t block_start, const size_t pos,
const uint8_t* ringbuffer, const size_t ringbuffer_mask,
const BrotliEncoderParams* params, const size_t max_backward_limit,
const int* starting_dist_cache, const size_t num_matches,
const BackwardMatch* matches, const ZopfliCostModel* model,
StartPosQueue* queue, ZopfliNode* nodes) {
const size_t cur_ix = block_start + pos;
const size_t cur_ix_masked = cur_ix & ringbuffer_mask;
const size_t max_distance = BROTLI_MIN(size_t, cur_ix, max_backward_limit);
const size_t max_len = num_bytes - pos;
const size_t max_zopfli_len = MaxZopfliLen(params);
const size_t max_iters = MaxZopfliCandidates(params);
size_t min_len;
size_t result = 0;
size_t k;
EvaluateNode(block_start, pos, max_backward_limit, starting_dist_cache, model,
queue, nodes);
{
const PosData* posdata = StartPosQueueAt(queue, 0);
float min_cost = (posdata->cost + ZopfliCostModelGetMinCostCmd(model) +
ZopfliCostModelGetLiteralCosts(model, posdata->pos, pos));
min_len = ComputeMinimumCopyLength(min_cost, nodes, num_bytes, pos);
}
/* Go over the command starting positions in order of increasing cost
difference. */
for (k = 0; k < max_iters && k < StartPosQueueSize(queue); ++k) {
const PosData* posdata = StartPosQueueAt(queue, k);
const size_t start = posdata->pos;
const uint16_t inscode = GetInsertLengthCode(pos - start);
const float start_costdiff = posdata->costdiff;
const float base_cost = start_costdiff + (float)GetInsertExtra(inscode) +
ZopfliCostModelGetLiteralCosts(model, 0, pos);
/* Look for last distance matches using the distance cache from this
starting position. */
size_t best_len = min_len - 1;
size_t j = 0;
for (; j < BROTLI_NUM_DISTANCE_SHORT_CODES && best_len < max_len; ++j) {
const size_t idx = kDistanceCacheIndex[j];
const size_t backward =
(size_t)(posdata->distance_cache[idx] + kDistanceCacheOffset[j]);
size_t prev_ix = cur_ix - backward;
if (prev_ix >= cur_ix) {
continue;
}
if (BROTLI_PREDICT_FALSE(backward > max_distance)) {
continue;
}
prev_ix &= ringbuffer_mask;
if (cur_ix_masked + best_len > ringbuffer_mask ||
prev_ix + best_len > ringbuffer_mask ||
ringbuffer[cur_ix_masked + best_len] !=
ringbuffer[prev_ix + best_len]) {
continue;
}
{
const size_t len =
FindMatchLengthWithLimit(&ringbuffer[prev_ix],
&ringbuffer[cur_ix_masked],
max_len);
const float dist_cost = base_cost +
ZopfliCostModelGetDistanceCost(model, j);
size_t l;
for (l = best_len + 1; l <= len; ++l) {
const uint16_t copycode = GetCopyLengthCode(l);
const uint16_t cmdcode =
CombineLengthCodes(inscode, copycode, j == 0);
const float cost = (cmdcode < 128 ? base_cost : dist_cost) +
(float)GetCopyExtra(copycode) +
ZopfliCostModelGetCommandCost(model, cmdcode);
if (cost < nodes[pos + l].u.cost) {
UpdateZopfliNode(nodes, pos, start, l, l, backward, j + 1, cost);
result = BROTLI_MAX(size_t, result, l);
}
best_len = l;
}
}
}
/* At higher iterations look only for new last distance matches, since
looking only for new command start positions with the same distances
does not help much. */
if (k >= 2) continue;
{
/* Loop through all possible copy lengths at this position. */
size_t len = min_len;
for (j = 0; j < num_matches; ++j) {
BackwardMatch match = matches[j];
size_t dist = match.distance;
BROTLI_BOOL is_dictionary_match = TO_BROTLI_BOOL(dist > max_distance);
/* We already tried all possible last distance matches, so we can use
normal distance code here. */
size_t dist_code = dist + BROTLI_NUM_DISTANCE_SHORT_CODES - 1;
uint16_t dist_symbol;
uint32_t distextra;
uint32_t distnumextra;
float dist_cost;
size_t max_match_len;
PrefixEncodeCopyDistance(dist_code, 0, 0, &dist_symbol, &distextra);
distnumextra = distextra >> 24;
dist_cost = base_cost + (float)distnumextra +
ZopfliCostModelGetDistanceCost(model, dist_symbol);
/* Try all copy lengths up until the maximum copy length corresponding
to this distance. If the distance refers to the static dictionary, or
the maximum length is long enough, try only one maximum length. */
max_match_len = BackwardMatchLength(&match);
if (len < max_match_len &&
(is_dictionary_match || max_match_len > max_zopfli_len)) {
len = max_match_len;
}
for (; len <= max_match_len; ++len) {
const size_t len_code =
is_dictionary_match ? BackwardMatchLengthCode(&match) : len;
const uint16_t copycode = GetCopyLengthCode(len_code);
const uint16_t cmdcode = CombineLengthCodes(inscode, copycode, 0);
const float cost = dist_cost + (float)GetCopyExtra(copycode) +
ZopfliCostModelGetCommandCost(model, cmdcode);
if (cost < nodes[pos + len].u.cost) {
UpdateZopfliNode(nodes, pos, start, len, len_code, dist, 0, cost);
result = BROTLI_MAX(size_t, result, len);
}
}
}
}
}
return result;
}
static size_t ComputeShortestPathFromNodes(size_t num_bytes,
ZopfliNode* nodes) {
size_t index = num_bytes;
size_t num_commands = 0;
while (nodes[index].insert_length == 0 && nodes[index].length == 1) --index;
nodes[index].u.next = BROTLI_UINT32_MAX;
while (index != 0) {
size_t len = ZopfliNodeCommandLength(&nodes[index]);
index -= len;
nodes[index].u.next = (uint32_t)len;
num_commands++;
}
return num_commands;
}
void BrotliZopfliCreateCommands(const size_t num_bytes,
const size_t block_start,
const size_t max_backward_limit,
const ZopfliNode* nodes,
int* dist_cache,
size_t* last_insert_len,
Command* commands,
size_t* num_literals) {
size_t pos = 0;
uint32_t offset = nodes[0].u.next;
size_t i;
for (i = 0; offset != BROTLI_UINT32_MAX; i++) {
const ZopfliNode* next = &nodes[pos + offset];
size_t copy_length = ZopfliNodeCopyLength(next);
size_t insert_length = next->insert_length;
pos += insert_length;
offset = next->u.next;
if (i == 0) {
insert_length += *last_insert_len;
*last_insert_len = 0;
}
{
size_t distance = ZopfliNodeCopyDistance(next);
size_t len_code = ZopfliNodeLengthCode(next);
size_t max_distance =
BROTLI_MIN(size_t, block_start + pos, max_backward_limit);
BROTLI_BOOL is_dictionary = TO_BROTLI_BOOL(distance > max_distance);
size_t dist_code = ZopfliNodeDistanceCode(next);
InitCommand(
&commands[i], insert_length, copy_length, len_code, dist_code);
if (!is_dictionary && dist_code > 0) {
dist_cache[3] = dist_cache[2];
dist_cache[2] = dist_cache[1];
dist_cache[1] = dist_cache[0];
dist_cache[0] = (int)distance;
}
}
*num_literals += insert_length;
pos += copy_length;
}
*last_insert_len += num_bytes - pos;
}
static size_t ZopfliIterate(size_t num_bytes,
size_t position,
const uint8_t* ringbuffer,
size_t ringbuffer_mask,
const BrotliEncoderParams* params,
const size_t max_backward_limit,
const int* dist_cache,
const ZopfliCostModel* model,
const uint32_t* num_matches,
const BackwardMatch* matches,
ZopfliNode* nodes) {
const size_t max_zopfli_len = MaxZopfliLen(params);
StartPosQueue queue;
size_t cur_match_pos = 0;
size_t i;
nodes[0].length = 0;
nodes[0].u.cost = 0;
InitStartPosQueue(&queue);
for (i = 0; i + 3 < num_bytes; i++) {
size_t skip = UpdateNodes(num_bytes, position, i, ringbuffer,
ringbuffer_mask, params, max_backward_limit, dist_cache,
num_matches[i], &matches[cur_match_pos], model, &queue, nodes);
if (skip < BROTLI_LONG_COPY_QUICK_STEP) skip = 0;
cur_match_pos += num_matches[i];
if (num_matches[i] == 1 &&
BackwardMatchLength(&matches[cur_match_pos - 1]) > max_zopfli_len) {
skip = BROTLI_MAX(size_t,
BackwardMatchLength(&matches[cur_match_pos - 1]), skip);
}
if (skip > 1) {
skip--;
while (skip) {
i++;
if (i + 3 >= num_bytes) break;
EvaluateNode(
position, i, max_backward_limit, dist_cache, model, &queue, nodes);
cur_match_pos += num_matches[i];
skip--;
}
}
}
return ComputeShortestPathFromNodes(num_bytes, nodes);
}
size_t BrotliZopfliComputeShortestPath(MemoryManager* m,
size_t num_bytes,
size_t position,
const uint8_t* ringbuffer,
size_t ringbuffer_mask,
const BrotliEncoderParams* params,
const size_t max_backward_limit,
const int* dist_cache,
H10* hasher,
ZopfliNode* nodes) {
const size_t max_zopfli_len = MaxZopfliLen(params);
ZopfliCostModel model;
StartPosQueue queue;
BackwardMatch matches[MAX_NUM_MATCHES_H10];
const size_t store_end = num_bytes >= StoreLookaheadH10() ?
position + num_bytes - StoreLookaheadH10() + 1 : position;
size_t i;
nodes[0].length = 0;
nodes[0].u.cost = 0;
InitZopfliCostModel(m, &model, num_bytes);
if (BROTLI_IS_OOM(m)) return 0;
ZopfliCostModelSetFromLiteralCosts(
&model, position, ringbuffer, ringbuffer_mask);
InitStartPosQueue(&queue);
for (i = 0; i + HashTypeLengthH10() - 1 < num_bytes; i++) {
const size_t pos = position + i;
const size_t max_distance = BROTLI_MIN(size_t, pos, max_backward_limit);
size_t num_matches = FindAllMatchesH10(hasher, ringbuffer, ringbuffer_mask,
pos, num_bytes - i, max_distance, params, matches);
size_t skip;
if (num_matches > 0 &&
BackwardMatchLength(&matches[num_matches - 1]) > max_zopfli_len) {
matches[0] = matches[num_matches - 1];
num_matches = 1;
}
skip = UpdateNodes(num_bytes, position, i, ringbuffer, ringbuffer_mask,
params, max_backward_limit, dist_cache, num_matches, matches, &model,
&queue, nodes);
if (skip < BROTLI_LONG_COPY_QUICK_STEP) skip = 0;
if (num_matches == 1 && BackwardMatchLength(&matches[0]) > max_zopfli_len) {
skip = BROTLI_MAX(size_t, BackwardMatchLength(&matches[0]), skip);
}
if (skip > 1) {
/* Add the tail of the copy to the hasher. */
StoreRangeH10(hasher, ringbuffer, ringbuffer_mask, pos + 1, BROTLI_MIN(
size_t, pos + skip, store_end));
skip--;
while (skip) {
i++;
if (i + HashTypeLengthH10() - 1 >= num_bytes) break;
EvaluateNode(
position, i, max_backward_limit, dist_cache, &model, &queue, nodes);
skip--;
}
}
}
CleanupZopfliCostModel(m, &model);
return ComputeShortestPathFromNodes(num_bytes, nodes);
}
void BrotliCreateZopfliBackwardReferences(
MemoryManager* m, size_t num_bytes, size_t position,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, H10* hasher, int* dist_cache,
size_t* last_insert_len, Command* commands, size_t* num_commands,
size_t* num_literals) {
const size_t max_backward_limit = BROTLI_MAX_BACKWARD_LIMIT(params->lgwin);
ZopfliNode* nodes;
nodes = BROTLI_ALLOC(m, ZopfliNode, num_bytes + 1);
if (BROTLI_IS_OOM(m)) return;
BrotliInitZopfliNodes(nodes, num_bytes + 1);
*num_commands += BrotliZopfliComputeShortestPath(m, num_bytes, position,
ringbuffer, ringbuffer_mask, params, max_backward_limit,
dist_cache, hasher, nodes);
if (BROTLI_IS_OOM(m)) return;
BrotliZopfliCreateCommands(num_bytes, position, max_backward_limit, nodes,
dist_cache, last_insert_len, commands, num_literals);
BROTLI_FREE(m, nodes);
}
void BrotliCreateHqZopfliBackwardReferences(
MemoryManager* m, size_t num_bytes, size_t position,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, H10* hasher, int* dist_cache,
size_t* last_insert_len, Command* commands, size_t* num_commands,
size_t* num_literals) {
const size_t max_backward_limit = BROTLI_MAX_BACKWARD_LIMIT(params->lgwin);
uint32_t* num_matches = BROTLI_ALLOC(m, uint32_t, num_bytes);
size_t matches_size = 4 * num_bytes;
const size_t store_end = num_bytes >= StoreLookaheadH10() ?
position + num_bytes - StoreLookaheadH10() + 1 : position;
size_t cur_match_pos = 0;
size_t i;
size_t orig_num_literals;
size_t orig_last_insert_len;
int orig_dist_cache[4];
size_t orig_num_commands;
ZopfliCostModel model;
ZopfliNode* nodes;
BackwardMatch* matches = BROTLI_ALLOC(m, BackwardMatch, matches_size);
if (BROTLI_IS_OOM(m)) return;
for (i = 0; i + HashTypeLengthH10() - 1 < num_bytes; ++i) {
const size_t pos = position + i;
size_t max_distance = BROTLI_MIN(size_t, pos, max_backward_limit);
size_t max_length = num_bytes - i;
size_t num_found_matches;
size_t cur_match_end;
size_t j;
/* Ensure that we have enough free slots. */
BROTLI_ENSURE_CAPACITY(m, BackwardMatch, matches, matches_size,
cur_match_pos + MAX_NUM_MATCHES_H10);
if (BROTLI_IS_OOM(m)) return;
num_found_matches = FindAllMatchesH10(hasher, ringbuffer, ringbuffer_mask,
pos, max_length, max_distance, params, &matches[cur_match_pos]);
cur_match_end = cur_match_pos + num_found_matches;
for (j = cur_match_pos; j + 1 < cur_match_end; ++j) {
assert(BackwardMatchLength(&matches[j]) <
BackwardMatchLength(&matches[j + 1]));
assert(matches[j].distance > max_distance ||
matches[j].distance <= matches[j + 1].distance);
}
num_matches[i] = (uint32_t)num_found_matches;
if (num_found_matches > 0) {
const size_t match_len = BackwardMatchLength(&matches[cur_match_end - 1]);
if (match_len > MAX_ZOPFLI_LEN_QUALITY_11) {
const size_t skip = match_len - 1;
matches[cur_match_pos++] = matches[cur_match_end - 1];
num_matches[i] = 1;
/* Add the tail of the copy to the hasher. */
StoreRangeH10(hasher, ringbuffer, ringbuffer_mask, pos + 1,
BROTLI_MIN(size_t, pos + match_len, store_end));
memset(&num_matches[i + 1], 0, skip * sizeof(num_matches[0]));
i += skip;
} else {
cur_match_pos = cur_match_end;
}
}
}
orig_num_literals = *num_literals;
orig_last_insert_len = *last_insert_len;
memcpy(orig_dist_cache, dist_cache, 4 * sizeof(dist_cache[0]));
orig_num_commands = *num_commands;
nodes = BROTLI_ALLOC(m, ZopfliNode, num_bytes + 1);
if (BROTLI_IS_OOM(m)) return;
InitZopfliCostModel(m, &model, num_bytes);
if (BROTLI_IS_OOM(m)) return;
for (i = 0; i < 2; i++) {
BrotliInitZopfliNodes(nodes, num_bytes + 1);
if (i == 0) {
ZopfliCostModelSetFromLiteralCosts(
&model, position, ringbuffer, ringbuffer_mask);
} else {
ZopfliCostModelSetFromCommands(&model, position, ringbuffer,
ringbuffer_mask, commands, *num_commands - orig_num_commands,
orig_last_insert_len);
}
*num_commands = orig_num_commands;
*num_literals = orig_num_literals;
*last_insert_len = orig_last_insert_len;
memcpy(dist_cache, orig_dist_cache, 4 * sizeof(dist_cache[0]));
*num_commands += ZopfliIterate(num_bytes, position, ringbuffer,
ringbuffer_mask, params, max_backward_limit, dist_cache,
&model, num_matches, matches, nodes);
BrotliZopfliCreateCommands(num_bytes, position, max_backward_limit,
nodes, dist_cache, last_insert_len, commands, num_literals);
}
CleanupZopfliCostModel(m, &model);
BROTLI_FREE(m, nodes);
BROTLI_FREE(m, matches);
BROTLI_FREE(m, num_matches);
}
#if defined(__cplusplus) || defined(c_plusplus)
} /* extern "C" */
#endif

98
enc/backward_references_hq.h Executable file
View File

@ -0,0 +1,98 @@
/* Copyright 2013 Google Inc. All Rights Reserved.
Distributed under MIT license.
See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
*/
/* Function to find backward reference copies. */
#ifndef BROTLI_ENC_BACKWARD_REFERENCES_HQ_H_
#define BROTLI_ENC_BACKWARD_REFERENCES_HQ_H_
#include "../common/constants.h"
#include <brotli/types.h>
#include "./command.h"
#include "./hash.h"
#include "./memory.h"
#include "./port.h"
#include "./quality.h"
#if defined(__cplusplus) || defined(c_plusplus)
extern "C" {
#endif
BROTLI_INTERNAL void BrotliCreateZopfliBackwardReferences(
MemoryManager* m, size_t num_bytes, size_t position,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, H10* hasher, int* dist_cache,
size_t* last_insert_len, Command* commands, size_t* num_commands,
size_t* num_literals);
BROTLI_INTERNAL void BrotliCreateHqZopfliBackwardReferences(
MemoryManager* m, size_t num_bytes, size_t position,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, H10* hasher, int* dist_cache,
size_t* last_insert_len, Command* commands, size_t* num_commands,
size_t* num_literals);
typedef struct ZopfliNode {
/* best length to get up to this byte (not including this byte itself)
highest 8 bit is used to reconstruct the length code */
uint32_t length;
/* distance associated with the length
highest 7 bit contains distance short code + 1 (or zero if no short code)
*/
uint32_t distance;
/* number of literal inserts before this copy */
uint32_t insert_length;
/* This union holds information used by dynamic-programming. During forward
pass |cost| it used to store the goal function. When node is processed its
|cost| is invalidated in favor of |shortcut|. On path back-tracing pass
|next| is assigned the offset to next node on the path. */
union {
/* Smallest cost to get to this byte from the beginning, as found so far. */
float cost;
/* Offset to the next node on the path. Equals to command_length() of the
next node on the path. For last node equals to BROTLI_UINT32_MAX */
uint32_t next;
/* Node position that provides next distance for distance cache. */
uint32_t shortcut;
} u;
} ZopfliNode;
BROTLI_INTERNAL void BrotliInitZopfliNodes(ZopfliNode* array, size_t length);
/* Computes the shortest path of commands from position to at most
position + num_bytes.
On return, path->size() is the number of commands found and path[i] is the
length of the i-th command (copy length plus insert length).
Note that the sum of the lengths of all commands can be less than num_bytes.
On return, the nodes[0..num_bytes] array will have the following
"ZopfliNode array invariant":
For each i in [1..num_bytes], if nodes[i].cost < kInfinity, then
(1) nodes[i].copy_length() >= 2
(2) nodes[i].command_length() <= i and
(3) nodes[i - nodes[i].command_length()].cost < kInfinity */
BROTLI_INTERNAL size_t BrotliZopfliComputeShortestPath(
MemoryManager* m, size_t num_bytes, size_t position,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, const size_t max_backward_limit,
const int* dist_cache, H10* hasher, ZopfliNode* nodes);
BROTLI_INTERNAL void BrotliZopfliCreateCommands(const size_t num_bytes,
const size_t block_start,
const size_t max_backward_limit,
const ZopfliNode* nodes,
int* dist_cache,
size_t* last_insert_len,
Command* commands,
size_t* num_literals);
#if defined(__cplusplus) || defined(c_plusplus)
} /* extern "C" */
#endif
#endif /* BROTLI_ENC_BACKWARD_REFERENCES_HQ_H_ */

9
enc/backward_references_inc.h
View File

@ -10,13 +10,13 @@
#define Hasher HASHER()
static BROTLI_NOINLINE void FN(CreateBackwardReferences)(
MemoryManager* m, size_t num_bytes, size_t position, BROTLI_BOOL is_last,
size_t num_bytes, size_t position,
const uint8_t* ringbuffer, size_t ringbuffer_mask,
const BrotliEncoderParams* params, Hasher* hasher, int* dist_cache,
size_t* last_insert_len, Command* commands, size_t* num_commands,
size_t* num_literals) {
/* Set maximum distance, see section 9.1. of the spec. */
const size_t max_backward_limit = MaxBackwardLimit(params->lgwin);
const size_t max_backward_limit = BROTLI_MAX_BACKWARD_LIMIT(params->lgwin);
const Command* const orig_commands = commands;
size_t insert_length = *last_insert_len;
@ -32,11 +32,6 @@ static BROTLI_NOINLINE void FN(CreateBackwardReferences)(
/* Minimum score to accept a backward reference. */
const score_t kMinScore = BROTLI_SCORE_BASE + 400;
FN(Init)(m, hasher, ringbuffer, params, position, num_bytes, is_last);
if (BROTLI_IS_OOM(m)) return;
FN(StitchToPreviousBlock)(hasher, num_bytes, position,
ringbuffer, ringbuffer_mask);
while (position + FN(HashTypeLength)() < pos_end) {
size_t max_length = pos_end - position;
size_t max_distance = BROTLI_MIN(size_t, position, max_backward_limit);

3
enc/compress_fragment.c
View File

@ -31,8 +31,7 @@
extern "C" {
#endif
/* Same as MaxBackwardLimit(18) */
#define MAX_DISTANCE ((1 << 18) - BROTLI_WINDOW_GAP)
#define MAX_DISTANCE (long)BROTLI_MAX_BACKWARD_LIMIT(18)
/* kHashMul32 multiplier has these properties:
* The multiplier must be odd. Otherwise we may lose the highest bit.

2
enc/compress_fragment.h
View File

@ -40,7 +40,7 @@ extern "C" {
REQUIRES: All elements in "table[0..table_size-1]" are initialized to zero.
REQUIRES: "table_size" is an odd (9, 11, 13, 15) power of two
OUTPUT: maximal copy distance <= |input_size|
OUTPUT: maximal copy distance <= MaxBackwardLimit(18) */
OUTPUT: maximal copy distance <= BROTLI_MAX_BACKWARD_LIMIT(18) */
BROTLI_INTERNAL void BrotliCompressFragmentFast(MemoryManager* m,
const uint8_t* input,
size_t input_size,

3
enc/compress_fragment_two_pass.c
View File

@ -30,8 +30,7 @@
extern "C" {
#endif
/* Same as MaxBackwardLimit(18) */
#define MAX_DISTANCE ((1 << 18) - BROTLI_WINDOW_GAP)
#define MAX_DISTANCE (long)BROTLI_MAX_BACKWARD_LIMIT(18)
/* kHashMul32 multiplier has these properties:
* The multiplier must be odd. Otherwise we may lose the highest bit.

2
enc/compress_fragment_two_pass.h
View File

@ -34,7 +34,7 @@ static const size_t kCompressFragmentTwoPassBlockSize = 1 << 17;
REQUIRES: All elements in "table[0..table_size-1]" are initialized to zero.
REQUIRES: "table_size" is a power of two
OUTPUT: maximal copy distance <= |input_size|
OUTPUT: maximal copy distance <= MaxBackwardLimit(18) */
OUTPUT: maximal copy distance <= BROTLI_MAX_BACKWARD_LIMIT(18) */
BROTLI_INTERNAL void BrotliCompressFragmentTwoPass(MemoryManager* m,
const uint8_t* input,
size_t input_size,

1120
enc/dictionary_hash.c Executable file
View File

File diff suppressed because it is too large Load Diff

1102
enc/dictionary_hash.h
View File

File diff suppressed because it is too large Load Diff

65
enc/encode.c
View File

@ -13,6 +13,7 @@
#include "../common/version.h"
#include "./backward_references.h"
#include "./backward_references_hq.h"
#include "./bit_cost.h"
#include "./brotli_bit_stream.h"
#include "./compress_fragment.h"
@ -156,6 +157,10 @@ BROTLI_BOOL BrotliEncoderSetParameter(
state->params.disable_literal_context_modeling = TO_BROTLI_BOOL(!!value);
return BROTLI_TRUE;
case BROTLI_PARAM_SIZE_HINT:
state->params.size_hint = value;
return BROTLI_TRUE;
default: return BROTLI_FALSE;
}
}
@ -583,13 +588,17 @@ static BROTLI_BOOL EnsureInitialized(BrotliEncoderState* s) {
return BROTLI_TRUE;
}
static void BrotliEncoderInitState(BrotliEncoderState* s) {
s->params.mode = BROTLI_DEFAULT_MODE;
s->params.quality = BROTLI_DEFAULT_QUALITY;
s->params.lgwin = BROTLI_DEFAULT_WINDOW;
s->params.lgblock = 0;
s->params.disable_literal_context_modeling = BROTLI_FALSE;
static void BrotliEncoderInitParams(BrotliEncoderParams* params) {
params->mode = BROTLI_DEFAULT_MODE;
params->quality = BROTLI_DEFAULT_QUALITY;
params->lgwin = BROTLI_DEFAULT_WINDOW;
params->lgblock = 0;
params->size_hint = 0;
params->disable_literal_context_modeling = BROTLI_FALSE;
}
static void BrotliEncoderInitState(BrotliEncoderState* s) {
BrotliEncoderInitParams(&s->params);
s->input_pos_ = 0;
s->num_commands_ = 0;
s->num_literals_ = 0;
@ -741,7 +750,7 @@ static void CopyInputToRingBuffer(BrotliEncoderState* s,
void BrotliEncoderSetCustomDictionary(BrotliEncoderState* s, size_t size,
const uint8_t* dict) {
size_t max_dict_size = MaxBackwardLimit(s->params.lgwin);
size_t max_dict_size = BROTLI_MAX_BACKWARD_LIMIT(s->params.lgwin);
size_t dict_size = size;
MemoryManager* m = &s->memory_manager_;
@ -884,17 +893,29 @@ static BROTLI_BOOL EncodeData(
}
}
BrotliCreateBackwardReferences(m, bytes, wrapped_last_processed_pos,
is_last, data, mask,
&s->params,
&s->hashers_,
s->dist_cache_,
&s->last_insert_len_,
&s->commands_[s->num_commands_],
&s->num_commands_,
&s->num_literals_);
InitOrStitchToPreviousBlock(m, &s->hashers_, data, mask, &s->params,
wrapped_last_processed_pos, bytes, is_last);
if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
if (s->params.quality == ZOPFLIFICATION_QUALITY) {
BrotliCreateZopfliBackwardReferences(
m, bytes, wrapped_last_processed_pos, data, mask,
&s->params, s->hashers_.h10, s->dist_cache_, &s->last_insert_len_,
&s->commands_[s->num_commands_], &s->num_commands_, &s->num_literals_);
if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
} else if (s->params.quality == HQ_ZOPFLIFICATION_QUALITY) {
BrotliCreateHqZopfliBackwardReferences(
m, bytes, wrapped_last_processed_pos, data, mask,
&s->params, s->hashers_.h10, s->dist_cache_, &s->last_insert_len_,
&s->commands_[s->num_commands_], &s->num_commands_, &s->num_literals_);
if (BROTLI_IS_OOM(m)) return BROTLI_FALSE;
} else {
BrotliCreateBackwardReferences(
bytes, wrapped_last_processed_pos, data, mask,
&s->params, &s->hashers_, s->dist_cache_, &s->last_insert_len_,
&s->commands_[s->num_commands_], &s->num_commands_, &s->num_literals_);
}
{
const size_t max_length = MaxMetablockSize(&s->params);
const size_t max_literals = max_length / 8;
@ -1008,7 +1029,7 @@ static BROTLI_BOOL BrotliCompressBufferQuality10(
MemoryManager* m = &memory_manager;
const size_t mask = BROTLI_SIZE_MAX >> 1;
const size_t max_backward_limit = MaxBackwardLimit(lgwin);
const size_t max_backward_limit = BROTLI_MAX_BACKWARD_LIMIT(lgwin);
int dist_cache[4] = { 4, 11, 15, 16 };
int saved_dist_cache[4] = { 4, 11, 15, 16 };
BROTLI_BOOL ok = BROTLI_TRUE;
@ -1032,11 +1053,9 @@ static BROTLI_BOOL BrotliCompressBufferQuality10(
uint8_t prev_byte = 0;
uint8_t prev_byte2 = 0;
params.mode = BROTLI_DEFAULT_MODE;
BrotliEncoderInitParams(&params);
params.quality = 10;
params.lgwin = lgwin;
params.lgblock = 0;
params.disable_literal_context_modeling = BROTLI_FALSE;
SanitizeParams(&params);
params.lgblock = ComputeLgBlock(&params);
max_block_size = (size_t)1 << params.lgblock;
@ -1308,6 +1327,7 @@ BROTLI_BOOL BrotliEncoderCompress(
BrotliEncoderSetParameter(s, BROTLI_PARAM_QUALITY, (uint32_t)quality);
BrotliEncoderSetParameter(s, BROTLI_PARAM_LGWIN, (uint32_t)lgwin);
BrotliEncoderSetParameter(s, BROTLI_PARAM_MODE, (uint32_t)mode);
BrotliEncoderSetParameter(s, BROTLI_PARAM_SIZE_HINT, (uint32_t)input_size);
result = BrotliEncoderCompressStream(s, BROTLI_OPERATION_FINISH,
&available_in, &next_in, &available_out, &next_out, &total_out);
if (!BrotliEncoderIsFinished(s)) result = 0;
@ -1572,6 +1592,11 @@ BROTLI_BOOL BrotliEncoderCompressStream(
BrotliEncoderState* s, BrotliEncoderOperation op, size_t* available_in,
const uint8_t** next_in, size_t* available_out,uint8_t** next_out,
size_t* total_out) {
/* If we don't have any size hint, set it based on the size of the first
input chunk. */
if (s->params.size_hint == 0) {
s->params.size_hint = (uint32_t)*available_in;
}
if (!EnsureInitialized(s)) return BROTLI_FALSE;
/* Unfinished metadata block; check requirements. */

2
enc/fast_log.h
View File

@ -19,7 +19,7 @@ extern "C" {
#endif
static BROTLI_INLINE uint32_t Log2FloorNonZero(size_t n) {
#ifdef __GNUC__
#if BROTLI_MODERN_COMPILER || __has_builtin(__builtin_clz)
return 31u ^ (uint32_t)__builtin_clz((uint32_t)n);
#else
uint32_t result = 0;

368
enc/hash.h
View File

@ -27,8 +27,6 @@
extern "C" {
#endif
#define MAX_TREE_SEARCH_DEPTH 64
#define MAX_TREE_COMP_LENGTH 128
#define score_t size_t
static const uint32_t kDistanceCacheIndex[] = {
@ -62,6 +60,7 @@ typedef struct DictionarySearchStatictics {
for this use.
* The number has been tuned heuristically against compression benchmarks. */
static const uint32_t kHashMul32 = 0x1e35a7bd;
static const uint64_t kHashMul64 = BROTLI_MAKE_UINT64_T(0x1e35a7bd, 0x1e35a7bd);
static BROTLI_INLINE uint32_t Hash14(const uint8_t* data) {
uint32_t h = BROTLI_UNALIGNED_LOAD32(data) * kHashMul32;
@ -221,323 +220,17 @@ static BROTLI_INLINE size_t BackwardMatchLengthCode(const BackwardMatch* self) {
#define CAT(a, b) a ## b
#define FN(X) EXPAND_CAT(X, HASHER())
#define MAX_NUM_MATCHES_H10 (64 + MAX_TREE_SEARCH_DEPTH)
#define HASHER() H10
#define HashToBinaryTree HASHER()
#define BUCKET_BITS 17
#define BUCKET_SIZE (1 << BUCKET_BITS)
static size_t FN(HashTypeLength)(void) { return 4; }
static size_t FN(StoreLookahead)(void) { return MAX_TREE_COMP_LENGTH; }
static uint32_t FN(HashBytes)(const uint8_t *data) {
uint32_t h = BROTLI_UNALIGNED_LOAD32(data) * kHashMul32;
/* The higher bits contain more mixture from the multiplication,
so we take our results from there. */
return h >> (32 - BUCKET_BITS);
}
/* A (forgetful) hash table where each hash bucket contains a binary tree of
sequences whose first 4 bytes share the same hash code.
Each sequence is MAX_TREE_COMP_LENGTH long and is identified by its starting
position in the input data. The binary tree is sorted by the lexicographic
order of the sequences, and it is also a max-heap with respect to the
starting positions. */
typedef struct HashToBinaryTree {
/* The window size minus 1 */
size_t window_mask_;
/* Hash table that maps the 4-byte hashes of the sequence to the last
position where this hash was found, which is the root of the binary
tree of sequences that share this hash bucket. */
uint32_t buckets_[BUCKET_SIZE];
/* The union of the binary trees of each hash bucket. The root of the tree
corresponding to a hash is a sequence starting at buckets_[hash] and
the left and right children of a sequence starting at pos are
forest_[2 * pos] and forest_[2 * pos + 1]. */
uint32_t* forest_;
/* A position used to mark a non-existent sequence, i.e. a tree is empty if
its root is at invalid_pos_ and a node is a leaf if both its children
are at invalid_pos_. */
uint32_t invalid_pos_;
size_t forest_size_;
BROTLI_BOOL is_dirty_;
} HashToBinaryTree;
static void FN(Reset)(HashToBinaryTree* self) {
self->is_dirty_ = BROTLI_TRUE;
}
static void FN(Initialize)(HashToBinaryTree* self) {
self->forest_ = NULL;
self->forest_size_ = 0;
FN(Reset)(self);
}
static void FN(Cleanup)(MemoryManager* m, HashToBinaryTree* self) {
BROTLI_FREE(m, self->forest_);
}
static void FN(Init)(
MemoryManager* m, HashToBinaryTree* self, const uint8_t* data,
const BrotliEncoderParams* params, size_t position, size_t bytes,
BROTLI_BOOL is_last) {
if (self->is_dirty_) {
uint32_t invalid_pos;
size_t num_nodes;
uint32_t i;
BROTLI_UNUSED(data);
self->window_mask_ = (1u << params->lgwin) - 1u;
invalid_pos = (uint32_t)(0 - self->window_mask_);
self->invalid_pos_ = invalid_pos;
for (i = 0; i < BUCKET_SIZE; i++) {
self->buckets_[i] = invalid_pos;
}
num_nodes = (position == 0 && is_last) ? bytes : self->window_mask_ + 1;
if (num_nodes > self->forest_size_) {
BROTLI_FREE(m, self->forest_);
self->forest_ = BROTLI_ALLOC(m, uint32_t, 2 * num_nodes);
if (BROTLI_IS_OOM(m)) return;
self->forest_size_ = num_nodes;
}
self->is_dirty_ = BROTLI_FALSE;
}
}
static BROTLI_INLINE size_t FN(LeftChildIndex)(HashToBinaryTree* self,
const size_t pos) {
return 2 * (pos & self->window_mask_);
}
static BROTLI_INLINE size_t FN(RightChildIndex)(HashToBinaryTree* self,
const size_t pos) {
return 2 * (pos & self->window_mask_) + 1;
}
/* Stores the hash of the next 4 bytes and in a single tree-traversal, the
hash bucket's binary tree is searched for matches and is re-rooted at the
current position.
If less than MAX_TREE_COMP_LENGTH data is available, the hash bucket of the
current position is searched for matches, but the state of the hash table
is not changed, since we can not know the final sorting order of the
current (incomplete) sequence.
This function must be called with increasing cur_ix positions. */
static BROTLI_INLINE BackwardMatch* FN(StoreAndFindMatches)(
HashToBinaryTree* self, const uint8_t* const BROTLI_RESTRICT data,
const size_t cur_ix, const size_t ring_buffer_mask, const size_t max_length,
const size_t max_backward, size_t* const BROTLI_RESTRICT best_len,
BackwardMatch* BROTLI_RESTRICT matches) {
const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
const size_t max_comp_len =
BROTLI_MIN(size_t, max_length, MAX_TREE_COMP_LENGTH);
const BROTLI_BOOL should_reroot_tree =
TO_BROTLI_BOOL(max_length >= MAX_TREE_COMP_LENGTH);
const uint32_t key = FN(HashBytes)(&data[cur_ix_masked]);
size_t prev_ix = self->buckets_[key];
/* The forest index of the rightmost node of the left subtree of the new
root, updated as we traverse and re-root the tree of the hash bucket. */
size_t node_left = FN(LeftChildIndex)(self, cur_ix);
/* The forest index of the leftmost node of the right subtree of the new
root, updated as we traverse and re-root the tree of the hash bucket. */
size_t node_right = FN(RightChildIndex)(self, cur_ix);
/* The match length of the rightmost node of the left subtree of the new
root, updated as we traverse and re-root the tree of the hash bucket. */
size_t best_len_left = 0;
/* The match length of the leftmost node of the right subtree of the new
root, updated as we traverse and re-root the tree of the hash bucket. */
size_t best_len_right = 0;
size_t depth_remaining;
if (should_reroot_tree) {
self->buckets_[key] = (uint32_t)cur_ix;
}
for (depth_remaining = MAX_TREE_SEARCH_DEPTH; ; --depth_remaining) {
const size_t backward = cur_ix - prev_ix;
const size_t prev_ix_masked = prev_ix & ring_buffer_mask;
if (backward == 0 || backward > max_backward || depth_remaining == 0) {
if (should_reroot_tree) {
self->forest_[node_left] = self->invalid_pos_;
self->forest_[node_right] = self->invalid_pos_;
}
break;
}
{
const size_t cur_len = BROTLI_MIN(size_t, best_len_left, best_len_right);
size_t len;
assert(cur_len <= MAX_TREE_COMP_LENGTH);
len = cur_len +
FindMatchLengthWithLimit(&data[cur_ix_masked + cur_len],
&data[prev_ix_masked + cur_len],
max_length - cur_len);
assert(0 == memcmp(&data[cur_ix_masked], &data[prev_ix_masked], len));
if (matches && len > *best_len) {
*best_len = len;
InitBackwardMatch(matches++, backward, len);
}
if (len >= max_comp_len) {
if (should_reroot_tree) {
self->forest_[node_left] =
self->forest_[FN(LeftChildIndex)(self, prev_ix)];
self->forest_[node_right] =
self->forest_[FN(RightChildIndex)(self, prev_ix)];
}
break;
}
if (data[cur_ix_masked + len] > data[prev_ix_masked + len]) {
best_len_left = len;
if (should_reroot_tree) {
self->forest_[node_left] = (uint32_t)prev_ix;
}
node_left = FN(RightChildIndex)(self, prev_ix);
prev_ix = self->forest_[node_left];
} else {
best_len_right = len;
if (should_reroot_tree) {
self->forest_[node_right] = (uint32_t)prev_ix;
}
node_right = FN(LeftChildIndex)(self, prev_ix);
prev_ix = self->forest_[node_right];
}
}
}
return matches;
}
/* Finds all backward matches of &data[cur_ix & ring_buffer_mask] up to the
length of max_length and stores the position cur_ix in the hash table.
Sets *num_matches to the number of matches found, and stores the found
matches in matches[0] to matches[*num_matches - 1]. The matches will be
sorted by strictly increasing length and (non-strictly) increasing
distance. */
static BROTLI_INLINE size_t FN(FindAllMatches)(HashToBinaryTree* self,
const uint8_t* data, const size_t ring_buffer_mask, const size_t cur_ix,
const size_t max_length, const size_t max_backward,
const BrotliEncoderParams* params, BackwardMatch* matches) {
BackwardMatch* const orig_matches = matches;
const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
size_t best_len = 1;
const size_t short_match_max_backward =
params->quality != HQ_ZOPFLIFICATION_QUALITY ? 16 : 64;
size_t stop = cur_ix - short_match_max_backward;
uint32_t dict_matches[BROTLI_MAX_STATIC_DICTIONARY_MATCH_LEN + 1];
size_t i;
if (cur_ix < short_match_max_backward) { stop = 0; }
for (i = cur_ix - 1; i > stop && best_len <= 2; --i) {
size_t prev_ix = i;
const size_t backward = cur_ix - prev_ix;
if (BROTLI_PREDICT_FALSE(backward > max_backward)) {
break;
}
prev_ix &= ring_buffer_mask;
if (data[cur_ix_masked] != data[prev_ix] ||
data[cur_ix_masked + 1] != data[prev_ix + 1]) {
continue;
}
{
const size_t len =
FindMatchLengthWithLimit(&data[prev_ix], &data[cur_ix_masked],
max_length);
if (len > best_len) {
best_len = len;
InitBackwardMatch(matches++, backward, len);
}
}
}
if (best_len < max_length) {
matches = FN(StoreAndFindMatches)(self, data, cur_ix, ring_buffer_mask,
max_length, max_backward, &best_len, matches);
}
for (i = 0; i <= BROTLI_MAX_STATIC_DICTIONARY_MATCH_LEN; ++i) {
dict_matches[i] = kInvalidMatch;
}
{
size_t minlen = BROTLI_MAX(size_t, 4, best_len + 1);
if (BrotliFindAllStaticDictionaryMatches(&data[cur_ix_masked], minlen,
max_length, &dict_matches[0])) {
size_t maxlen = BROTLI_MIN(
size_t, BROTLI_MAX_STATIC_DICTIONARY_MATCH_LEN, max_length);
size_t l;
for (l = minlen; l <= maxlen; ++l) {
uint32_t dict_id = dict_matches[l];
if (dict_id < kInvalidMatch) {
InitDictionaryBackwardMatch(matches++,
max_backward + (dict_id >> 5) + 1, l, dict_id & 31);
}
}
}
}
return (size_t)(matches - orig_matches);
}
/* Stores the hash of the next 4 bytes and re-roots the binary tree at the
current sequence, without returning any matches.
REQUIRES: ix + MAX_TREE_COMP_LENGTH <= end-of-current-block */
static BROTLI_INLINE void FN(Store)(HashToBinaryTree* self, const uint8_t *data,
const size_t mask, const size_t ix) {
/* Maximum distance is window size - 16, see section 9.1. of the spec. */
const size_t max_backward = self->window_mask_ - BROTLI_WINDOW_GAP + 1;
FN(StoreAndFindMatches)(self, data, ix, mask, MAX_TREE_COMP_LENGTH,
max_backward, NULL, NULL);
}
static BROTLI_INLINE void FN(StoreRange)(HashToBinaryTree* self,
const uint8_t *data, const size_t mask, const size_t ix_start,
const size_t ix_end) {
size_t i = ix_start;
size_t j = ix_start;
if (ix_start + 63 <= ix_end) {
i = ix_end - 63;
}
if (ix_start + 512 <= i) {
for (; j < i; j += 8) {
FN(Store)(self, data, mask, j);
}
}
for (; i < ix_end; ++i) {
FN(Store)(self, data, mask, i);
}
}
static BROTLI_INLINE void FN(StitchToPreviousBlock)(HashToBinaryTree* self,
size_t num_bytes, size_t position, const uint8_t* ringbuffer,
size_t ringbuffer_mask) {
if (num_bytes >= FN(HashTypeLength)() - 1 &&
position >= MAX_TREE_COMP_LENGTH) {
/* Store the last `MAX_TREE_COMP_LENGTH - 1` positions in the hasher.
These could not be calculated before, since they require knowledge
of both the previous and the current block. */
const size_t i_start = position - MAX_TREE_COMP_LENGTH + 1;
const size_t i_end = BROTLI_MIN(size_t, position, i_start + num_bytes);
size_t i;
for (i = i_start; i < i_end; ++i) {
/* Maximum distance is window size - 16, see section 9.1. of the spec.
Furthermore, we have to make sure that we don't look further back
from the start of the next block than the window size, otherwise we
could access already overwritten areas of the ring-buffer. */
const size_t max_backward =
self->window_mask_ - BROTLI_MAX(size_t,
BROTLI_WINDOW_GAP - 1,
position - i);
/* We know that i + MAX_TREE_COMP_LENGTH <= position + num_bytes, i.e. the
end of the current block and that we have at least
MAX_TREE_COMP_LENGTH tail in the ring-buffer. */
FN(StoreAndFindMatches)(self, ringbuffer, i, ringbuffer_mask,
MAX_TREE_COMP_LENGTH, max_backward, NULL, NULL);
}
}
}
#undef BUCKET_SIZE
#define MAX_TREE_SEARCH_DEPTH 64
#define MAX_TREE_COMP_LENGTH 128
#include "./hash_to_binary_tree_inc.h" /* NOLINT(build/include) */
#undef MAX_TREE_SEARCH_DEPTH
#undef MAX_TREE_COMP_LENGTH
#undef BUCKET_BITS
#undef HASHER
/* MAX_NUM_MATCHES == 64 + MAX_TREE_SEARCH_DEPTH */
#define MAX_NUM_MATCHES_H10 128
/* For BUCKET_SWEEP == 1, enabling the dictionary lookup makes compression
a little faster (0.5% - 1%) and it compresses 0.15% better on small text
@ -546,6 +239,7 @@ static BROTLI_INLINE void FN(StitchToPreviousBlock)(HashToBinaryTree* self,
#define HASHER() H2
#define BUCKET_BITS 16
#define BUCKET_SWEEP 1
#define HASH_LEN 5
#define USE_DICTIONARY 1
#include "./hash_longest_match_quickly_inc.h" /* NOLINT(build/include) */
#undef BUCKET_SWEEP
@ -567,6 +261,7 @@ static BROTLI_INLINE void FN(StitchToPreviousBlock)(HashToBinaryTree* self,
#define USE_DICTIONARY 1
#include "./hash_longest_match_quickly_inc.h" /* NOLINT(build/include) */
#undef USE_DICTIONARY
#undef HASH_LEN
#undef BUCKET_SWEEP
#undef BUCKET_BITS
#undef HASHER
@ -641,12 +336,24 @@ static BROTLI_INLINE void FN(StitchToPreviousBlock)(HashToBinaryTree* self,
#undef BUCKET_BITS
#define HASHER() H54
#define BUCKET_BITS 20
#define BUCKET_SWEEP 4
#define HASH_LEN 7
#define USE_DICTIONARY 0
#include "./hash_longest_match_quickly_inc.h" /* NOLINT(build/include) */
#undef USE_DICTIONARY
#undef HASH_LEN
#undef BUCKET_SWEEP
#undef BUCKET_BITS
#undef HASHER
#undef FN
#undef CAT
#undef EXPAND_CAT
#define FOR_GENERIC_HASHERS(H) H(2) H(3) H(4) H(5) H(6) H(7) H(8) H(9) \
H(40) H(41) H(42)
H(40) H(41) H(42) H(54)
#define FOR_ALL_HASHERS(H) FOR_GENERIC_HASHERS(H) H(10)
typedef struct Hashers {
@ -662,8 +369,8 @@ static BROTLI_INLINE void InitHashers(Hashers* self) {
}
static BROTLI_INLINE void DestroyHashers(MemoryManager* m, Hashers* self) {
if (self->h10) CleanupH10(m, self->h10);
#define CLEANUP_(N) BROTLI_FREE(m, self->h ## N)
#define CLEANUP_(N) if (self->h ## N) CleanupH ## N(m, self->h ## N); \
BROTLI_FREE(m, self->h ## N);
FOR_ALL_HASHERS(CLEANUP_)
#undef CLEANUP_
}
@ -686,7 +393,12 @@ static BROTLI_INLINE void HashersSetup(
default: break;
}
if (BROTLI_IS_OOM(m)) return;
if (type == 10) InitializeH10(self->h10);
switch (type) {
#define INITIALIZE_(N) case N: InitializeH ## N(self->h ## N); break;
FOR_ALL_HASHERS(INITIALIZE_);
#undef INITIALIZE_
default: break;
}
HashersReset(self, type);
}
@ -720,6 +432,24 @@ static BROTLI_INLINE void HashersPrependCustomDictionary(
if (BROTLI_IS_OOM(m)) return;
}
static BROTLI_INLINE void InitOrStitchToPreviousBlock(
MemoryManager* m, Hashers* self, const uint8_t* data, size_t mask,
const BrotliEncoderParams* params, size_t position,
size_t bytes, BROTLI_BOOL is_last) {
int hasher_type = ChooseHasher(params);
switch (hasher_type) {
#define INIT_(N) \
case N: \
InitH ## N(m, self->h ## N, data, params, position, bytes, is_last); \
if (BROTLI_IS_OOM(m)) return; \
StitchToPreviousBlockH ## N(self->h ## N, bytes, position, data, mask); \
break;
FOR_ALL_HASHERS(INIT_)
#undef INIT_
default: break;
}
}
#if defined(__cplusplus) || defined(c_plusplus)
} /* extern "C" */

9
enc/hash_forgetful_chain_inc.h
View File

@ -56,6 +56,15 @@ typedef struct HashForgetfulChain {
size_t max_hops;
} HashForgetfulChain;
static void FN(Initialize)(HashForgetfulChain* self) {
BROTLI_UNUSED(self);
}
static void FN(Cleanup)(MemoryManager* m, HashForgetfulChain* self) {
BROTLI_UNUSED(m);
BROTLI_UNUSED(self);
}
static void FN(Reset)(HashForgetfulChain* self) {
self->is_dirty_ = BROTLI_TRUE;
DictionarySearchStaticticsReset(&self->dict_search_stats_);

9
enc/hash_longest_match_inc.h
View File

@ -55,6 +55,15 @@ typedef struct HashLongestMatch {
DictionarySearchStatictics dict_search_stats_;
} HashLongestMatch;
static void FN(Initialize)(HashLongestMatch* self) {
BROTLI_UNUSED(self);
}
static void FN(Cleanup)(MemoryManager* m, HashLongestMatch* self) {
BROTLI_UNUSED(m);
BROTLI_UNUSED(self);
}
static void FN(Reset)(HashLongestMatch* self) {
self->is_dirty_ = BROTLI_TRUE;
DictionarySearchStaticticsReset(&self->dict_search_stats_);

18
enc/hash_longest_match_quickly_inc.h
View File

@ -5,7 +5,9 @@
See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
*/
/* template parameters: FN, BUCKET_BITS, BUCKET_SWEEP, USE_DICTIONARY */
/* template parameters: FN, BUCKET_BITS, BUCKET_SWEEP, HASH_LEN,
USE_DICTIONARY
*/
#define HashLongestMatchQuickly HASHER()
@ -20,9 +22,8 @@ static BROTLI_INLINE size_t FN(StoreLookahead)(void) { return 8; }
the address in. The HashLongestMatch and HashLongestMatchQuickly
classes have separate, different implementations of hashing. */
static uint32_t FN(HashBytes)(const uint8_t *data) {
/* Computing a hash based on 5 bytes works much better for
qualities 1 and 3, where the next hash value is likely to replace */
uint64_t h = (BROTLI_UNALIGNED_LOAD64(data) << 24) * kHashMul32;
const uint64_t h = ((BROTLI_UNALIGNED_LOAD64(data) << (64 - 8 * HASH_LEN)) *
kHashMul64);
/* The higher bits contain more mixture from the multiplication,
so we take our results from there. */
return (uint32_t)(h >> (64 - BUCKET_BITS));
@ -40,6 +41,15 @@ typedef struct HashLongestMatchQuickly {
DictionarySearchStatictics dict_search_stats_;
} HashLongestMatchQuickly;
static void FN(Initialize)(HashLongestMatchQuickly* self) {
BROTLI_UNUSED(self);
}
static void FN(Cleanup)(MemoryManager* m, HashLongestMatchQuickly* self) {
BROTLI_UNUSED(m);
BROTLI_UNUSED(self);
}
static void FN(Reset)(HashLongestMatchQuickly* self) {
self->is_dirty_ = BROTLI_TRUE;
DictionarySearchStaticticsReset(&self->dict_search_stats_);

323
enc/hash_to_binary_tree_inc.h Executable file
View File

@ -0,0 +1,323 @@
/* NOLINT(build/header_guard) */
/* Copyright 2016 Google Inc. All Rights Reserved.
Distributed under MIT license.
See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
*/
/* template parameters: FN, BUCKET_BITS, MAX_TREE_COMP_LENGTH,
MAX_TREE_SEARCH_DEPTH */
/* A (forgetful) hash table where each hash bucket contains a binary tree of
sequences whose first 4 bytes share the same hash code.
Each sequence is MAX_TREE_COMP_LENGTH long and is identified by its starting
position in the input data. The binary tree is sorted by the lexicographic
order of the sequences, and it is also a max-heap with respect to the
starting positions. */
#define HashToBinaryTree HASHER()
#define BUCKET_SIZE (1 << BUCKET_BITS)
static size_t FN(HashTypeLength)(void) { return 4; }
static size_t FN(StoreLookahead)(void) { return MAX_TREE_COMP_LENGTH; }
static uint32_t FN(HashBytes)(const uint8_t *data) {
uint32_t h = BROTLI_UNALIGNED_LOAD32(data) * kHashMul32;
/* The higher bits contain more mixture from the multiplication,
so we take our results from there. */
return h >> (32 - BUCKET_BITS);
}
typedef struct HashToBinaryTree {
/* The window size minus 1 */
size_t window_mask_;
/* Hash table that maps the 4-byte hashes of the sequence to the last
position where this hash was found, which is the root of the binary
tree of sequences that share this hash bucket. */
uint32_t buckets_[BUCKET_SIZE];
/* The union of the binary trees of each hash bucket. The root of the tree
corresponding to a hash is a sequence starting at buckets_[hash] and
the left and right children of a sequence starting at pos are
forest_[2 * pos] and forest_[2 * pos + 1]. */
uint32_t* forest_;
/* A position used to mark a non-existent sequence, i.e. a tree is empty if
its root is at invalid_pos_ and a node is a leaf if both its children
are at invalid_pos_. */
uint32_t invalid_pos_;
size_t forest_size_;
BROTLI_BOOL is_dirty_;
} HashToBinaryTree;
static void FN(Reset)(HashToBinaryTree* self) {
self->is_dirty_ = BROTLI_TRUE;
}
static void FN(Initialize)(HashToBinaryTree* self) {
self->forest_ = NULL;
self->forest_size_ = 0;
FN(Reset)(self);
}
static void FN(Cleanup)(MemoryManager* m, HashToBinaryTree* self) {
BROTLI_FREE(m, self->forest_);
}
static void FN(Init)(
MemoryManager* m, HashToBinaryTree* self, const uint8_t* data,
const BrotliEncoderParams* params, size_t position, size_t bytes,
BROTLI_BOOL is_last) {
if (self->is_dirty_) {
uint32_t invalid_pos;
size_t num_nodes;
uint32_t i;
BROTLI_UNUSED(data);
self->window_mask_ = (1u << params->lgwin) - 1u;
invalid_pos = (uint32_t)(0 - self->window_mask_);
self->invalid_pos_ = invalid_pos;
for (i = 0; i < BUCKET_SIZE; i++) {
self->buckets_[i] = invalid_pos;
}
num_nodes = (position == 0 && is_last) ? bytes : self->window_mask_ + 1;
if (num_nodes > self->forest_size_) {
BROTLI_FREE(m, self->forest_);
self->forest_ = BROTLI_ALLOC(m, uint32_t, 2 * num_nodes);
if (BROTLI_IS_OOM(m)) return;
self->forest_size_ = num_nodes;
}
self->is_dirty_ = BROTLI_FALSE;
}
}
static BROTLI_INLINE size_t FN(LeftChildIndex)(HashToBinaryTree* self,
const size_t pos) {
return 2 * (pos & self->window_mask_);
}
static BROTLI_INLINE size_t FN(RightChildIndex)(HashToBinaryTree* self,
const size_t pos) {
return 2 * (pos & self->window_mask_) + 1;
}
/* Stores the hash of the next 4 bytes and in a single tree-traversal, the
hash bucket's binary tree is searched for matches and is re-rooted at the
current position.
If less than MAX_TREE_COMP_LENGTH data is available, the hash bucket of the
current position is searched for matches, but the state of the hash table
is not changed, since we can not know the final sorting order of the
current (incomplete) sequence.
This function must be called with increasing cur_ix positions. */
static BROTLI_INLINE BackwardMatch* FN(StoreAndFindMatches)(
HashToBinaryTree* self, const uint8_t* const BROTLI_RESTRICT data,
const size_t cur_ix, const size_t ring_buffer_mask, const size_t max_length,
const size_t max_backward, size_t* const BROTLI_RESTRICT best_len,
BackwardMatch* BROTLI_RESTRICT matches) {
const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
const size_t max_comp_len =
BROTLI_MIN(size_t, max_length, MAX_TREE_COMP_LENGTH);
const BROTLI_BOOL should_reroot_tree =
TO_BROTLI_BOOL(max_length >= MAX_TREE_COMP_LENGTH);
const uint32_t key = FN(HashBytes)(&data[cur_ix_masked]);
size_t prev_ix = self->buckets_[key];
/* The forest index of the rightmost node of the left subtree of the new
root, updated as we traverse and re-root the tree of the hash bucket. */
size_t node_left = FN(LeftChildIndex)(self, cur_ix);
/* The forest index of the leftmost node of the right subtree of the new
root, updated as we traverse and re-root the tree of the hash bucket. */
size_t node_right = FN(RightChildIndex)(self, cur_ix);
/* The match length of the rightmost node of the left subtree of the new
root, updated as we traverse and re-root the tree of the hash bucket. */
size_t best_len_left = 0;
/* The match length of the leftmost node of the right subtree of the new
root, updated as we traverse and re-root the tree of the hash bucket. */
size_t best_len_right = 0;
size_t depth_remaining;
if (should_reroot_tree) {
self->buckets_[key] = (uint32_t)cur_ix;
}
for (depth_remaining = MAX_TREE_SEARCH_DEPTH; ; --depth_remaining) {
const size_t backward = cur_ix - prev_ix;
const size_t prev_ix_masked = prev_ix & ring_buffer_mask;
if (backward == 0 || backward > max_backward || depth_remaining == 0) {
if (should_reroot_tree) {
self->forest_[node_left] = self->invalid_pos_;
self->forest_[node_right] = self->invalid_pos_;
}
break;
}
{
const size_t cur_len = BROTLI_MIN(size_t, best_len_left, best_len_right);
size_t len;
assert(cur_len <= MAX_TREE_COMP_LENGTH);
len = cur_len +
FindMatchLengthWithLimit(&data[cur_ix_masked + cur_len],
&data[prev_ix_masked + cur_len],
max_length - cur_len);
assert(0 == memcmp(&data[cur_ix_masked], &data[prev_ix_masked], len));
if (matches && len > *best_len) {
*best_len = len;
InitBackwardMatch(matches++, backward, len);
}
if (len >= max_comp_len) {
if (should_reroot_tree) {
self->forest_[node_left] =
self->forest_[FN(LeftChildIndex)(self, prev_ix)];
self->forest_[node_right] =
self->forest_[FN(RightChildIndex)(self, prev_ix)];
}
break;
}
if (data[cur_ix_masked + len] > data[prev_ix_masked + len]) {
best_len_left = len;
if (should_reroot_tree) {
self->forest_[node_left] = (uint32_t)prev_ix;
}
node_left = FN(RightChildIndex)(self, prev_ix);
prev_ix = self->forest_[node_left];
} else {
best_len_right = len;
if (should_reroot_tree) {
self->forest_[node_right] = (uint32_t)prev_ix;
}
node_right = FN(LeftChildIndex)(self, prev_ix);
prev_ix = self->forest_[node_right];
}
}
}
return matches;
}
/* Finds all backward matches of &data[cur_ix & ring_buffer_mask] up to the
length of max_length and stores the position cur_ix in the hash table.
Sets *num_matches to the number of matches found, and stores the found
matches in matches[0] to matches[*num_matches - 1]. The matches will be
sorted by strictly increasing length and (non-strictly) increasing
distance. */
static BROTLI_INLINE size_t FN(FindAllMatches)(HashToBinaryTree* self,
const uint8_t* data, const size_t ring_buffer_mask, const size_t cur_ix,
const size_t max_length, const size_t max_backward,
const BrotliEncoderParams* params, BackwardMatch* matches) {
BackwardMatch* const orig_matches = matches;
const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
size_t best_len = 1;
const size_t short_match_max_backward =
params->quality != HQ_ZOPFLIFICATION_QUALITY ? 16 : 64;
size_t stop = cur_ix - short_match_max_backward;
uint32_t dict_matches[BROTLI_MAX_STATIC_DICTIONARY_MATCH_LEN + 1];
size_t i;
if (cur_ix < short_match_max_backward) { stop = 0; }
for (i = cur_ix - 1; i > stop && best_len <= 2; --i) {
size_t prev_ix = i;
const size_t backward = cur_ix - prev_ix;
if (BROTLI_PREDICT_FALSE(backward > max_backward)) {
break;
}
prev_ix &= ring_buffer_mask;
if (data[cur_ix_masked] != data[prev_ix] ||
data[cur_ix_masked + 1] != data[prev_ix + 1]) {
continue;
}
{
const size_t len =
FindMatchLengthWithLimit(&data[prev_ix], &data[cur_ix_masked],
max_length);
if (len > best_len) {
best_len = len;
InitBackwardMatch(matches++, backward, len);
}
}
}
if (best_len < max_length) {
matches = FN(StoreAndFindMatches)(self, data, cur_ix, ring_buffer_mask,
max_length, max_backward, &best_len, matches);
}
for (i = 0; i <= BROTLI_MAX_STATIC_DICTIONARY_MATCH_LEN; ++i) {
dict_matches[i] = kInvalidMatch;
}
{
size_t minlen = BROTLI_MAX(size_t, 4, best_len + 1);
if (BrotliFindAllStaticDictionaryMatches(&data[cur_ix_masked], minlen,
max_length, &dict_matches[0])) {
size_t maxlen = BROTLI_MIN(
size_t, BROTLI_MAX_STATIC_DICTIONARY_MATCH_LEN, max_length);
size_t l;
for (l = minlen; l <= maxlen; ++l) {
uint32_t dict_id = dict_matches[l];
if (dict_id < kInvalidMatch) {
InitDictionaryBackwardMatch(matches++,
max_backward + (dict_id >> 5) + 1, l, dict_id & 31);
}
}
}
}
return (size_t)(matches - orig_matches);
}
/* Stores the hash of the next 4 bytes and re-roots the binary tree at the
current sequence, without returning any matches.
REQUIRES: ix + MAX_TREE_COMP_LENGTH <= end-of-current-block */
static BROTLI_INLINE void FN(Store)(HashToBinaryTree* self, const uint8_t *data,
const size_t mask, const size_t ix) {
/* Maximum distance is window size - 16, see section 9.1. of the spec. */
const size_t max_backward = self->window_mask_ - BROTLI_WINDOW_GAP + 1;
FN(StoreAndFindMatches)(self, data, ix, mask, MAX_TREE_COMP_LENGTH,
max_backward, NULL, NULL);
}
static BROTLI_INLINE void FN(StoreRange)(HashToBinaryTree* self,
const uint8_t *data, const size_t mask, const size_t ix_start,
const size_t ix_end) {
size_t i = ix_start;
size_t j = ix_start;
if (ix_start + 63 <= ix_end) {
i = ix_end - 63;
}
if (ix_start + 512 <= i) {
for (; j < i; j += 8) {
FN(Store)(self, data, mask, j);
}
}
for (; i < ix_end; ++i) {
FN(Store)(self, data, mask, i);
}
}
static BROTLI_INLINE void FN(StitchToPreviousBlock)(HashToBinaryTree* self,
size_t num_bytes, size_t position, const uint8_t* ringbuffer,
size_t ringbuffer_mask) {
if (num_bytes >= FN(HashTypeLength)() - 1 &&
position >= MAX_TREE_COMP_LENGTH) {
/* Store the last `MAX_TREE_COMP_LENGTH - 1` positions in the hasher.
These could not be calculated before, since they require knowledge
of both the previous and the current block. */
const size_t i_start = position - MAX_TREE_COMP_LENGTH + 1;
const size_t i_end = BROTLI_MIN(size_t, position, i_start + num_bytes);
size_t i;
for (i = i_start; i < i_end; ++i) {
/* Maximum distance is window size - 16, see section 9.1. of the spec.
Furthermore, we have to make sure that we don't look further back
from the start of the next block than the window size, otherwise we
could access already overwritten areas of the ring-buffer. */
const size_t max_backward =
self->window_mask_ - BROTLI_MAX(size_t,
BROTLI_WINDOW_GAP - 1,
position - i);
/* We know that i + MAX_TREE_COMP_LENGTH <= position + num_bytes, i.e. the
end of the current block and that we have at least
MAX_TREE_COMP_LENGTH tail in the ring-buffer. */
FN(StoreAndFindMatches)(self, ringbuffer, i, ringbuffer_mask,
MAX_TREE_COMP_LENGTH, max_backward, NULL, NULL);
}
}
}
#undef BUCKET_SIZE
#undef HashToBinaryTree

3
enc/memory.h
View File

@ -39,7 +39,8 @@ BROTLI_INTERNAL void BrotliInitMemoryManager(
void* opaque);
BROTLI_INTERNAL void* BrotliAllocate(MemoryManager* m, size_t n);
#define BROTLI_ALLOC(M, T, N) ((T*)BrotliAllocate((M), (N) * sizeof(T)))
#define BROTLI_ALLOC(M, T, N) \
((N) ? ((T*)BrotliAllocate((M), (N) * sizeof(T))) : NULL)
BROTLI_INTERNAL void BrotliFree(MemoryManager* m, void* p);
#define BROTLI_FREE(M, P) { \

3
enc/quality.h
View File

@ -37,6 +37,7 @@ typedef struct BrotliEncoderParams {
int quality;
int lgwin;
int lgblock;
size_t size_hint;
BROTLI_BOOL disable_literal_context_modeling;
} BrotliEncoderParams;
@ -124,6 +125,8 @@ static BROTLI_INLINE size_t LiteralSpreeLengthForSparseSearch(
static BROTLI_INLINE int ChooseHasher(const BrotliEncoderParams* params) {
if (params->quality > 9) {
return 10;
} else if (params->quality == 4 && params->size_hint >= (1 << 20)) {
return 54;
} else if (params->quality < 5) {
return params->quality;
} else if (params->lgwin <= 16) {

8
include/brotli/encode.h
View File

@ -171,7 +171,13 @@ typedef enum BrotliEncoderParameter {
*
* This flag is a "decoding-speed vs compression ratio" trade-off.
*/
BROTLI_PARAM_DISABLE_LITERAL_CONTEXT_MODELING = 4
BROTLI_PARAM_DISABLE_LITERAL_CONTEXT_MODELING = 4,
/**
* Estimated total input size for all ::BrotliEncoderCompressStream calls.
*
* The default value is 0, which means that the total input size is unknown.
*/
BROTLI_PARAM_SIZE_HINT = 5
} BrotliEncoderParameter;
/**

2
include/brotli/port.h
View File

@ -36,7 +36,7 @@
#if defined(BROTLI_BUILD_MODERN_COMPILER)
#define BROTLI_MODERN_COMPILER 1
#elif BROTLI_GCC_VERSION > 300 || BROTLI_ICC_VERSION >= 1600
#elif BROTLI_GCC_VERSION >= 304 || BROTLI_ICC_VERSION >= 1600
#define BROTLI_MODERN_COMPILER 1
#else
#define BROTLI_MODERN_COMPILER 0

4
setup.py
View File

@ -170,12 +170,14 @@ EXT_MODULES = [
'dec/huffman.c',
'dec/state.c',
'enc/backward_references.c',
'enc/backward_references_hq.c',
'enc/bit_cost.c',
'enc/block_splitter.c',
'enc/brotli_bit_stream.c',
'enc/cluster.c',
'enc/compress_fragment.c',
'enc/compress_fragment_two_pass.c',
'enc/dictionary_hash.c',
'enc/encode.c',
'enc/entropy_encode.c',
'enc/histogram.c',
@ -199,6 +201,7 @@ EXT_MODULES = [
'dec/streams.h',
'dec/transform.h',
'enc/backward_references.h',
'enc/backward_references_hq.h',
'enc/backward_references_inc.h',
'enc/bit_cost.h',
'enc/bit_cost_inc.h',
@ -217,6 +220,7 @@ EXT_MODULES = [
'enc/fast_log.h',
'enc/find_match_length.h',
'enc/hash.h',
'enc/hash_to_binary_tree_inc.h',
'enc/hash_longest_match_inc.h',
'enc/hash_longest_match_quickly_inc.h',
'enc/histogram.h',