Use a hash table header and SIMD to speed up hash table operations (similar to [Swiss Tables](https://abseil.io/about/design/swisstables)).
PiperOrigin-RevId: 636018541
diff --git a/c/common/platform.h b/c/common/platform.h
index c18ff02..713119e 100644
--- a/c/common/platform.h
+++ b/c/common/platform.h
@@ -282,6 +282,11 @@
memcpy(&t, p, sizeof t);
return t;
}
+static BROTLI_INLINE size_t BrotliUnalignedReadSizeT(const void* p) {
+ size_t t;
+ memcpy(&t, p, sizeof t);
+ return t;
+}
static BROTLI_INLINE void BrotliUnalignedWrite64(void* p, uint64_t v) {
memcpy(p, &v, sizeof v);
}
@@ -489,11 +494,29 @@
/* Default brotli_free_func */
BROTLI_COMMON_API void BrotliDefaultFreeFunc(void* opaque, void* address);
+/* Circular logical rotates. */
+static BROTLI_INLINE uint16_t BrotliRotateRight16(uint16_t const value,
+ size_t count) {
+ count &= 0x0F; /* for fickle pattern recognition */
+ return (value >> count) | (uint16_t)(value << ((0U - count) & 0x0F));
+}
+static BROTLI_INLINE uint32_t BrotliRotateRight32(uint32_t const value,
+ size_t count) {
+ count &= 0x1F; /* for fickle pattern recognition */
+ return (value >> count) | (uint32_t)(value << ((0U - count) & 0x1F));
+}
+static BROTLI_INLINE uint64_t BrotliRotateRight64(uint64_t const value,
+ size_t count) {
+ count &= 0x3F; /* for fickle pattern recognition */
+ return (value >> count) | (uint64_t)(value << ((0U - count) & 0x3F));
+}
+
BROTLI_UNUSED_FUNCTION void BrotliSuppressUnusedFunctions(void) {
BROTLI_UNUSED(&BrotliSuppressUnusedFunctions);
BROTLI_UNUSED(&BrotliUnalignedRead16);
BROTLI_UNUSED(&BrotliUnalignedRead32);
BROTLI_UNUSED(&BrotliUnalignedRead64);
+ BROTLI_UNUSED(&BrotliUnalignedReadSizeT);
BROTLI_UNUSED(&BrotliUnalignedWrite64);
BROTLI_UNUSED(&BROTLI_UNALIGNED_LOAD16LE);
BROTLI_UNUSED(&BROTLI_UNALIGNED_LOAD32LE);
@@ -516,6 +539,9 @@
BROTLI_UNUSED(&brotli_max_uint8_t);
BROTLI_UNUSED(&BrotliDefaultAllocFunc);
BROTLI_UNUSED(&BrotliDefaultFreeFunc);
+ BROTLI_UNUSED(&BrotliRotateRight16);
+ BROTLI_UNUSED(&BrotliRotateRight32);
+ BROTLI_UNUSED(&BrotliRotateRight64);
#if BROTLI_ENABLE_DUMP
BROTLI_UNUSED(&BrotliDump);
#endif
@@ -534,6 +560,13 @@
# define PREFETCH_L1(ptr) do { (void)(ptr); } while (0) /* disabled */
# define PREFETCH_L2(ptr) do { (void)(ptr); } while (0) /* disabled */
#endif
+
+/* The SIMD matchers are only faster at certain quality levels. */
+#if defined(_M_X64) && defined(BROTLI_TZCNT64)
+#define BROTLI_MAX_SIMD_QUALITY 7
+#elif defined(BROTLI_TZCNT64)
+#define BROTLI_MAX_SIMD_QUALITY 6
+#endif
}
#endif /* BROTLI_COMMON_PLATFORM_H_ */
diff --git a/c/enc/backward_references.c b/c/enc/backward_references.c
index f600e64..2531ceb 100644
--- a/c/enc/backward_references.c
+++ b/c/enc/backward_references.c
@@ -116,6 +116,18 @@
#include "backward_references_inc.h"
#undef HASHER
+#if defined(BROTLI_MAX_SIMD_QUALITY)
+#define HASHER() H58
+/* NOLINTNEXTLINE(build/include) */
+#include "backward_references_inc.h"
+#undef HASHER
+
+#define HASHER() H68
+/* NOLINTNEXTLINE(build/include) */
+#include "backward_references_inc.h"
+#undef HASHER
+#endif
+
#undef ENABLE_COMPOUND_DICTIONARY
#undef PREFIX
#define PREFIX() D
@@ -149,6 +161,16 @@
/* NOLINTNEXTLINE(build/include) */
#include "backward_references_inc.h"
#undef HASHER
+#if defined(BROTLI_MAX_SIMD_QUALITY)
+#define HASHER() H58
+/* NOLINTNEXTLINE(build/include) */
+#include "backward_references_inc.h"
+#undef HASHER
+#define HASHER() H68
+/* NOLINTNEXTLINE(build/include) */
+#include "backward_references_inc.h"
+#undef HASHER
+#endif
#undef ENABLE_COMPOUND_DICTIONARY
#undef PREFIX
@@ -174,6 +196,10 @@
return;
CASE_(5)
CASE_(6)
+#if defined(BROTLI_MAX_SIMD_QUALITY)
+ CASE_(58)
+ CASE_(68)
+#endif
CASE_(40)
CASE_(41)
CASE_(42)
diff --git a/c/enc/hash.h b/c/enc/hash.h
index 9533b05..3ceb66c 100644
--- a/c/enc/hash.h
+++ b/c/enc/hash.h
@@ -22,6 +22,7 @@
#include "encoder_dict.h"
#include "fast_log.h"
#include "find_match_length.h"
+#include "matching_tag_mask.h"
#include "memory.h"
#include "quality.h"
#include "static_dict.h"
@@ -297,6 +298,16 @@
#include "hash_longest_match64_inc.h" /* NOLINT(build/include) */
#undef HASHER
+#if defined(BROTLI_MAX_SIMD_QUALITY)
+#define HASHER() H58
+#include "hash_longest_match_simd_inc.h" /* NOLINT(build/include) */
+#undef HASHER
+
+#define HASHER() H68
+#include "hash_longest_match64_simd_inc.h" /* NOLINT(build/include) */
+#undef HASHER
+#endif
+
#define BUCKET_BITS 15
#define NUM_LAST_DISTANCES_TO_CHECK 4
@@ -388,7 +399,13 @@
#undef CAT
#undef EXPAND_CAT
-#define FOR_SIMPLE_HASHERS(H) H(2) H(3) H(4) H(5) H(6) H(40) H(41) H(42) H(54)
+#if defined(BROTLI_MAX_SIMD_QUALITY)
+#define FOR_SIMPLE_HASHERS(H) \
+ H(2) H(3) H(4) H(5) H(6) H(40) H(41) H(42) H(54) H(58) H(68)
+#else
+#define FOR_SIMPLE_HASHERS(H) \
+ H(2) H(3) H(4) H(5) H(6) H(40) H(41) H(42) H(54)
+#endif
#define FOR_COMPOSITE_HASHERS(H) H(35) H(55) H(65)
#define FOR_GENERIC_HASHERS(H) FOR_SIMPLE_HASHERS(H) FOR_COMPOSITE_HASHERS(H)
#define FOR_ALL_HASHERS(H) FOR_GENERIC_HASHERS(H) H(10)
diff --git a/c/enc/hash_longest_match64_simd_inc.h b/c/enc/hash_longest_match64_simd_inc.h
new file mode 100644
index 0000000..cd4e7b1
--- /dev/null
+++ b/c/enc/hash_longest_match64_simd_inc.h
@@ -0,0 +1,300 @@
+/* NOLINT(build/header_guard) */
+/* Copyright 2010 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 */
+
+/* A (forgetful) hash table to the data seen by the compressor, to
+ help create backward references to previous data.
+
+ This is a hash map of fixed size (bucket_size_) to a ring buffer of
+ fixed size (block_size_). The ring buffer contains the last block_size_
+ index positions of the given hash key in the compressed data. */
+
+#define HashLongestMatch HASHER()
+
+#define TAG_HASH_BITS 8
+#define TAG_HASH_MASK ((1 << TAG_HASH_BITS) - 1)
+
+static BROTLI_INLINE size_t FN(HashTypeLength)(void) { return 8; }
+static BROTLI_INLINE size_t FN(StoreLookahead)(void) { return 8; }
+
+/* HashBytes is the function that chooses the bucket to place the address in. */
+static BROTLI_INLINE size_t FN(HashBytes)(const uint8_t* BROTLI_RESTRICT data,
+ uint64_t hash_mul) {
+ const uint64_t h = BROTLI_UNALIGNED_LOAD64LE(data) * hash_mul;
+ /* The higher bits contain more mixture from the multiplication,
+ so we take our results from there. */
+ return (size_t)(h >> (64 - 15 - TAG_HASH_BITS));
+}
+
+typedef struct HashLongestMatch {
+ /* Number of hash buckets. */
+ size_t bucket_size_;
+ /* Only block_size_ newest backward references are kept,
+ and the older are forgotten. */
+ size_t block_size_;
+ /* Hash multiplier tuned to match length. */
+ uint64_t hash_mul_;
+ /* Mask for accessing entries in a block (in a ring-buffer manner). */
+ uint32_t block_mask_;
+
+ int block_bits_;
+ int num_last_distances_to_check_;
+
+ /* Shortcuts. */
+ HasherCommon* common_;
+
+ /* --- Dynamic size members --- */
+
+ /* Number of entries in a particular bucket. */
+ uint16_t* num_; /* uint16_t[bucket_size]; */
+
+ uint8_t* tags_;
+
+ /* Buckets containing block_size_ of backward references. */
+ uint32_t* buckets_; /* uint32_t[bucket_size * block_size]; */
+} HashLongestMatch;
+
+static void FN(Initialize)(
+ HasherCommon* common, HashLongestMatch* BROTLI_RESTRICT self,
+ const BrotliEncoderParams* params) {
+ self->common_ = common;
+
+ BROTLI_UNUSED(params);
+ self->hash_mul_ = kHashMul64 << (64 - 5 * 8);
+ BROTLI_DCHECK(common->params.bucket_bits == 15);
+ self->bucket_size_ = (size_t)1 << common->params.bucket_bits;
+ self->block_bits_ = common->params.block_bits;
+ self->block_size_ = (size_t)1 << common->params.block_bits;
+ self->block_mask_ = (uint32_t)(self->block_size_ - 1);
+ self->num_last_distances_to_check_ =
+ common->params.num_last_distances_to_check;
+ self->num_ = (uint16_t*)common->extra[0];
+ self->tags_ = (uint8_t*)common->extra[1];
+ self->buckets_ = (uint32_t*)common->extra[2];
+}
+
+static void FN(Prepare)(
+ HashLongestMatch* BROTLI_RESTRICT self, BROTLI_BOOL one_shot,
+ size_t input_size, const uint8_t* BROTLI_RESTRICT data) {
+ uint16_t* BROTLI_RESTRICT num = self->num_;
+ /* Partial preparation is 100 times slower (per socket). */
+ size_t partial_prepare_threshold = self->bucket_size_ >> 6;
+ if (one_shot && input_size <= partial_prepare_threshold) {
+ size_t i;
+ for (i = 0; i < input_size; ++i) {
+ const size_t hash = FN(HashBytes)(&data[i], self->hash_mul_);
+ const size_t key = hash >> TAG_HASH_BITS;
+ num[key] = 65535;
+ }
+ } else {
+ /* Set all the bytes of num to 255, which makes each uint16_t 65535. */
+ memset(num, 255, self->bucket_size_ * sizeof(num[0]));
+ }
+}
+
+static BROTLI_INLINE void FN(HashMemAllocInBytes)(
+ const BrotliEncoderParams* params, BROTLI_BOOL one_shot,
+ size_t input_size, size_t* alloc_size) {
+ size_t bucket_size = (size_t)1 << params->hasher.bucket_bits;
+ size_t block_size = (size_t)1 << params->hasher.block_bits;
+ BROTLI_UNUSED(one_shot);
+ BROTLI_UNUSED(input_size);
+ alloc_size[0] = sizeof(uint16_t) * bucket_size;
+ alloc_size[1] = sizeof(uint8_t) * bucket_size * block_size;
+ alloc_size[2] = sizeof(uint32_t) * bucket_size * block_size;
+}
+
+/* Look at 4 bytes at &data[ix & mask].
+ Compute a hash from these, and store the value of ix at that position. */
+static BROTLI_INLINE void FN(Store)(
+ HashLongestMatch* BROTLI_RESTRICT self, const uint8_t* BROTLI_RESTRICT data,
+ const size_t mask, const size_t ix) {
+ uint16_t* BROTLI_RESTRICT num = self->num_;
+ uint8_t* BROTLI_RESTRICT tags = self->tags_;
+ uint32_t* BROTLI_RESTRICT buckets = self->buckets_;
+ const size_t hash = FN(HashBytes)(&data[ix & mask], self->hash_mul_);
+ const size_t key = hash >> TAG_HASH_BITS;
+ const uint8_t tag = hash & TAG_HASH_MASK;
+ const size_t minor_ix = num[key] & self->block_mask_;
+ const size_t offset = minor_ix + (key << self->block_bits_);
+ --num[key];
+ buckets[offset] = (uint32_t)ix;
+ tags[offset] = tag;
+}
+
+static BROTLI_INLINE void FN(StoreRange)(HashLongestMatch* BROTLI_RESTRICT self,
+ const uint8_t* BROTLI_RESTRICT data, const size_t mask,
+ const size_t ix_start, const size_t ix_end) {
+ size_t i;
+ for (i = ix_start; i < ix_end; ++i) {
+ FN(Store)(self, data, mask, i);
+ }
+}
+
+static BROTLI_INLINE void FN(StitchToPreviousBlock)(
+ HashLongestMatch* BROTLI_RESTRICT self,
+ size_t num_bytes, size_t position, const uint8_t* ringbuffer,
+ size_t ringbuffer_mask) {
+ if (num_bytes >= FN(HashTypeLength)() - 1 && position >= 3) {
+ /* Prepare the hashes for three last bytes of the last write.
+ These could not be calculated before, since they require knowledge
+ of both the previous and the current block. */
+ FN(Store)(self, ringbuffer, ringbuffer_mask, position - 3);
+ FN(Store)(self, ringbuffer, ringbuffer_mask, position - 2);
+ FN(Store)(self, ringbuffer, ringbuffer_mask, position - 1);
+ }
+}
+
+static BROTLI_INLINE void FN(PrepareDistanceCache)(
+ HashLongestMatch* BROTLI_RESTRICT self,
+ int* BROTLI_RESTRICT distance_cache) {
+ PrepareDistanceCache(distance_cache, self->num_last_distances_to_check_);
+}
+
+/* Find a longest backward match of &data[cur_ix] up to the length of
+ max_length and stores the position cur_ix in the hash table.
+
+ REQUIRES: FN(PrepareDistanceCache) must be invoked for current distance cache
+ values; if this method is invoked repeatedly with the same distance
+ cache values, it is enough to invoke FN(PrepareDistanceCache) once.
+
+ Does not look for matches longer than max_length.
+ Does not look for matches further away than max_backward.
+ Writes the best match into |out|.
+ |out|->score is updated only if a better match is found. */
+static BROTLI_INLINE void FN(FindLongestMatch)(
+ HashLongestMatch* BROTLI_RESTRICT self,
+ const BrotliEncoderDictionary* dictionary,
+ const uint8_t* BROTLI_RESTRICT data, const size_t ring_buffer_mask,
+ const int* BROTLI_RESTRICT distance_cache, const size_t cur_ix,
+ const size_t max_length, const size_t max_backward,
+ const size_t dictionary_distance, const size_t max_distance,
+ HasherSearchResult* BROTLI_RESTRICT out) {
+ uint16_t* BROTLI_RESTRICT num = self->num_;
+ uint32_t* BROTLI_RESTRICT buckets = self->buckets_;
+ uint8_t* BROTLI_RESTRICT tags = self->tags_;
+ const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
+ /* Don't accept a short copy from far away. */
+ score_t min_score = out->score;
+ score_t best_score = out->score;
+ size_t best_len = out->len;
+ size_t i;
+ /* Precalculate the hash key and prefetch the bucket. */
+ const size_t hash = FN(HashBytes)(&data[cur_ix_masked], self->hash_mul_);
+ const size_t key = hash >> TAG_HASH_BITS;
+ uint32_t* BROTLI_RESTRICT bucket = &buckets[key << self->block_bits_];
+ uint8_t* BROTLI_RESTRICT tag_bucket = &tags[key << self->block_bits_];
+ PREFETCH_L1(bucket);
+ PREFETCH_L1(tag_bucket);
+ if (self->block_bits_ > 4) PREFETCH_L1(bucket + 16);
+ out->len = 0;
+ out->len_code_delta = 0;
+
+ BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask);
+
+ /* Try last distance first. */
+ for (i = 0; i < (size_t)self->num_last_distances_to_check_; ++i) {
+ const size_t backward = (size_t)distance_cache[i];
+ size_t prev_ix = (size_t)(cur_ix - backward);
+ if (prev_ix >= cur_ix) {
+ continue;
+ }
+ if (BROTLI_PREDICT_FALSE(backward > max_backward)) {
+ continue;
+ }
+ prev_ix &= ring_buffer_mask;
+
+ if (cur_ix_masked + best_len > ring_buffer_mask ||
+ prev_ix + best_len > ring_buffer_mask ||
+ data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
+ continue;
+ }
+ {
+ const size_t len = FindMatchLengthWithLimit(&data[prev_ix],
+ &data[cur_ix_masked],
+ max_length);
+ if (len >= 3 || (len == 2 && i < 2)) {
+ /* Comparing for >= 2 does not change the semantics, but just saves for
+ a few unnecessary binary logarithms in backward reference score,
+ since we are not interested in such short matches. */
+ score_t score = BackwardReferenceScoreUsingLastDistance(len);
+ if (best_score < score) {
+ if (i != 0) score -= BackwardReferencePenaltyUsingLastDistance(i);
+ if (best_score < score) {
+ best_score = score;
+ best_len = len;
+ out->len = best_len;
+ out->distance = backward;
+ out->score = best_score;
+ }
+ }
+ }
+ }
+ }
+ /* we require matches of len >4, so increase best_len to 3, so we can compare
+ * 4 bytes all the time. */
+ if (best_len < 3) {
+ best_len = 3;
+ }
+ {
+ const uint8_t tag = hash & TAG_HASH_MASK;
+ const uint32_t first4 = BrotliUnalignedRead32(data + cur_ix_masked);
+ const size_t max_length_m4 = max_length - 4;
+ const size_t head = (num[key] + 1) & self->block_mask_;
+ uint64_t matches =
+ GetMatchingTagMask(self->block_size_ / 16, tag, tag_bucket, head);
+ /* Mask off any matches from uninitialized tags. */
+ uint16_t n = 65535 - num[key];
+ uint64_t block_has_unused_slots = self->block_size_ > n;
+ uint64_t mask = (block_has_unused_slots << (n & (64 - 1))) - 1;
+ matches &= mask;
+ for (; matches > 0; matches &= (matches - 1)) {
+ const size_t rb_index =
+ (head + (size_t)BROTLI_TZCNT64(matches)) & self->block_mask_;
+ size_t prev_ix = bucket[rb_index];
+ uint32_t current4;
+ const size_t backward = cur_ix - prev_ix;
+ if (BROTLI_PREDICT_FALSE(backward > max_backward)) {
+ break;
+ }
+ prev_ix &= ring_buffer_mask;
+ if (cur_ix_masked + best_len > ring_buffer_mask ||
+ prev_ix + best_len > ring_buffer_mask ||
+ /* compare 4 bytes ending at best_len + 1 */
+ BrotliUnalignedRead32(&data[cur_ix_masked + best_len - 3]) !=
+ BrotliUnalignedRead32(&data[prev_ix + best_len - 3])) {
+ continue;
+ }
+ current4 = BrotliUnalignedRead32(data + prev_ix);
+ if (first4 != current4) continue;
+ {
+ const size_t len = FindMatchLengthWithLimit(&data[prev_ix + 4],
+ &data[cur_ix_masked + 4],
+ max_length_m4) + 4;
+ const score_t score = BackwardReferenceScore(len, backward);
+ if (best_score < score) {
+ best_score = score;
+ best_len = len;
+ out->len = best_len;
+ out->distance = backward;
+ out->score = best_score;
+ }
+ }
+ }
+ bucket[num[key] & self->block_mask_] = (uint32_t)cur_ix;
+ tag_bucket[num[key] & self->block_mask_] = tag;
+ --num[key];
+ }
+ if (min_score == out->score) {
+ SearchInStaticDictionary(dictionary,
+ self->common_, &data[cur_ix_masked], max_length, dictionary_distance,
+ max_distance, out, BROTLI_FALSE);
+ }
+}
+
+#undef HashLongestMatch
diff --git a/c/enc/hash_longest_match_simd_inc.h b/c/enc/hash_longest_match_simd_inc.h
new file mode 100644
index 0000000..2bf9b6c
--- /dev/null
+++ b/c/enc/hash_longest_match_simd_inc.h
@@ -0,0 +1,274 @@
+/* NOLINT(build/header_guard) */
+/* Copyright 2010 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 */
+/* A (forgetful) hash table to the data seen by the compressor, to
+ help create backward references to previous data.
+ This is a hash map of fixed size (bucket_size_) to a ring buffer of
+ fixed size (block_size_). The ring buffer contains the last block_size_
+ index positions of the given hash key in the compressed data. */
+#define HashLongestMatch HASHER()
+#define TAG_HASH_BITS 8
+#define TAG_HASH_MASK ((1 << TAG_HASH_BITS) - 1)
+static BROTLI_INLINE size_t FN(HashTypeLength)(void) { return 4; }
+static BROTLI_INLINE size_t FN(StoreLookahead)(void) { return 4; }
+/* HashBytes is the function that chooses the bucket to place the address in. */
+static uint32_t FN(HashBytes)(
+ const uint8_t* BROTLI_RESTRICT data, const int shift) {
+ uint32_t h = BROTLI_UNALIGNED_LOAD32LE(data) * kHashMul32;
+ /* The higher bits contain more mixture from the multiplication,
+ so we take our results from there. */
+ return (uint32_t)(h >> shift);
+}
+typedef struct HashLongestMatch {
+ /* Number of hash buckets. */
+ size_t bucket_size_;
+ /* Only block_size_ newest backward references are kept,
+ and the older are forgotten. */
+ size_t block_size_;
+ /* Left-shift for computing hash bucket index from hash value. */
+ int hash_shift_;
+ /* Mask for accessing entries in a block (in a ring-buffer manner). */
+ uint32_t block_mask_;
+ int block_bits_;
+ int num_last_distances_to_check_;
+ /* Shortcuts. */
+ HasherCommon* common_;
+ /* --- Dynamic size members --- */
+ /* Number of entries in a particular bucket. */
+ uint16_t* num_; /* uint16_t[bucket_size]; */
+ uint8_t* tags_;
+ /* Buckets containing block_size_ of backward references. */
+ uint32_t* buckets_; /* uint32_t[bucket_size * block_size]; */
+} HashLongestMatch;
+static void FN(Initialize)(
+ HasherCommon* common, HashLongestMatch* BROTLI_RESTRICT self,
+ const BrotliEncoderParams* params) {
+ self->common_ = common;
+ BROTLI_UNUSED(params);
+ self->hash_shift_ = 32 - common->params.bucket_bits - TAG_HASH_BITS;
+ self->bucket_size_ = (size_t)1 << common->params.bucket_bits;
+ self->block_size_ = (size_t)1 << common->params.block_bits;
+ self->block_mask_ = (uint32_t)(self->block_size_ - 1);
+ self->num_ = (uint16_t*)common->extra[0];
+ self->tags_ = (uint8_t*)common->extra[1];
+ self->buckets_ = (uint32_t*)common->extra[2];
+ self->block_bits_ = common->params.block_bits;
+ self->num_last_distances_to_check_ =
+ common->params.num_last_distances_to_check;
+}
+static void FN(Prepare)(
+ HashLongestMatch* BROTLI_RESTRICT self, BROTLI_BOOL one_shot,
+ size_t input_size, const uint8_t* BROTLI_RESTRICT data) {
+ uint16_t* BROTLI_RESTRICT num = self->num_;
+ /* Partial preparation is 100 times slower (per socket). */
+ size_t partial_prepare_threshold = self->bucket_size_ >> 6;
+ if (one_shot && input_size <= partial_prepare_threshold) {
+ size_t i;
+ for (i = 0; i < input_size; ++i) {
+ const uint32_t hash = FN(HashBytes)(&data[i], self->hash_shift_);
+ const uint32_t key = hash >> TAG_HASH_BITS;
+ num[key] = 65535;
+ }
+ } else {
+ /* Set all the bytes of num to 255, which makes each uint16_t 65535. */
+ memset(num, 255, self->bucket_size_ * sizeof(num[0]));
+ }
+}
+static BROTLI_INLINE void FN(HashMemAllocInBytes)(
+ const BrotliEncoderParams* params, BROTLI_BOOL one_shot,
+ size_t input_size, size_t* alloc_size) {
+ size_t bucket_size = (size_t)1 << params->hasher.bucket_bits;
+ size_t block_size = (size_t)1 << params->hasher.block_bits;
+ BROTLI_UNUSED(one_shot);
+ BROTLI_UNUSED(input_size);
+ alloc_size[0] = sizeof(uint16_t) * bucket_size;
+ alloc_size[1] = sizeof(uint8_t) * bucket_size * block_size;
+ alloc_size[2] = sizeof(uint32_t) * bucket_size * block_size;
+}
+/* Look at 4 bytes at &data[ix & mask].
+ Compute a hash from these, and store the value of ix at that position. */
+static BROTLI_INLINE void FN(Store)(
+ HashLongestMatch* BROTLI_RESTRICT self, const uint8_t* BROTLI_RESTRICT data,
+ const size_t mask, const size_t ix) {
+ uint16_t* BROTLI_RESTRICT num = self->num_;
+ uint8_t* BROTLI_RESTRICT tags = self->tags_;
+ uint32_t* BROTLI_RESTRICT buckets = self->buckets_;
+ const size_t hash = FN(HashBytes)(&data[ix & mask], self->hash_shift_);
+ const size_t key = hash >> TAG_HASH_BITS;
+ const uint8_t tag = hash & TAG_HASH_MASK;
+ const size_t minor_ix = num[key] & self->block_mask_;
+ const size_t offset = minor_ix + (key << self->block_bits_);
+ --num[key];
+ buckets[offset] = (uint32_t)ix;
+ tags[offset] = tag;
+}
+static BROTLI_INLINE void FN(StoreRange)(HashLongestMatch* BROTLI_RESTRICT self,
+ const uint8_t* BROTLI_RESTRICT data, const size_t mask,
+ const size_t ix_start, const size_t ix_end) {
+ size_t i;
+ for (i = ix_start; i < ix_end; ++i) {
+ FN(Store)(self, data, mask, i);
+ }
+}
+static BROTLI_INLINE void FN(StitchToPreviousBlock)(
+ HashLongestMatch* BROTLI_RESTRICT self,
+ size_t num_bytes, size_t position, const uint8_t* ringbuffer,
+ size_t ringbuffer_mask) {
+ if (num_bytes >= FN(HashTypeLength)() - 1 && position >= 3) {
+ /* Prepare the hashes for three last bytes of the last write.
+ These could not be calculated before, since they require knowledge
+ of both the previous and the current block. */
+ FN(Store)(self, ringbuffer, ringbuffer_mask, position - 3);
+ FN(Store)(self, ringbuffer, ringbuffer_mask, position - 2);
+ FN(Store)(self, ringbuffer, ringbuffer_mask, position - 1);
+ }
+}
+static BROTLI_INLINE void FN(PrepareDistanceCache)(
+ HashLongestMatch* BROTLI_RESTRICT self,
+ int* BROTLI_RESTRICT distance_cache) {
+ PrepareDistanceCache(distance_cache, self->num_last_distances_to_check_);
+}
+
+/* Find a longest backward match of &data[cur_ix] up to the length of
+ max_length and stores the position cur_ix in the hash table.
+ REQUIRES: FN(PrepareDistanceCache) must be invoked for current distance cache
+ values; if this method is invoked repeatedly with the same distance
+ cache values, it is enough to invoke FN(PrepareDistanceCache) once.
+ Does not look for matches longer than max_length.
+ Does not look for matches further away than max_backward.
+ Writes the best match into |out|.
+ |out|->score is updated only if a better match is found. */
+static BROTLI_INLINE void FN(FindLongestMatch)(
+ HashLongestMatch* BROTLI_RESTRICT self,
+ const BrotliEncoderDictionary* dictionary,
+ const uint8_t* BROTLI_RESTRICT data, const size_t ring_buffer_mask,
+ const int* BROTLI_RESTRICT distance_cache, const size_t cur_ix,
+ const size_t max_length, const size_t max_backward,
+ const size_t dictionary_distance, const size_t max_distance,
+ HasherSearchResult* BROTLI_RESTRICT out) {
+ uint16_t* BROTLI_RESTRICT num = self->num_;
+ uint32_t* BROTLI_RESTRICT buckets = self->buckets_;
+ uint8_t* BROTLI_RESTRICT tags = self->tags_;
+ const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
+ /* Don't accept a short copy from far away. */
+ score_t min_score = out->score;
+ score_t best_score = out->score;
+ size_t best_len = out->len;
+ size_t i;
+ /* Precalculate the hash key and prefetch the bucket. */
+ const uint32_t hash =
+ FN(HashBytes)(&data[cur_ix_masked], self->hash_shift_);
+ const uint32_t key = hash >> TAG_HASH_BITS;
+ uint32_t* BROTLI_RESTRICT bucket = &buckets[key << self->block_bits_];
+ uint8_t* BROTLI_RESTRICT tag_bucket = &tags[key << self->block_bits_];
+ PREFETCH_L1(bucket);
+ PREFETCH_L1(tag_bucket);
+ if (self->block_bits_ > 4) PREFETCH_L1(bucket + 16);
+ out->len = 0;
+ out->len_code_delta = 0;
+
+ BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask);
+
+ /* Try last distance first. */
+ for (i = 0; i < (size_t)self->num_last_distances_to_check_; ++i) {
+ const size_t backward = (size_t)distance_cache[i];
+ size_t prev_ix = (size_t)(cur_ix - backward);
+ if (prev_ix >= cur_ix) {
+ continue;
+ }
+ if (BROTLI_PREDICT_FALSE(backward > max_backward)) {
+ continue;
+ }
+ prev_ix &= ring_buffer_mask;
+
+ if (cur_ix_masked + best_len > ring_buffer_mask ||
+ prev_ix + best_len > ring_buffer_mask ||
+ data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
+ continue;
+ }
+ {
+ const size_t len = FindMatchLengthWithLimit(&data[prev_ix],
+ &data[cur_ix_masked],
+ max_length);
+ if (len >= 3 || (len == 2 && i < 2)) {
+ /* Comparing for >= 2 does not change the semantics, but just saves for
+ a few unnecessary binary logarithms in backward reference score,
+ since we are not interested in such short matches. */
+ score_t score = BackwardReferenceScoreUsingLastDistance(len);
+ if (best_score < score) {
+ if (i != 0) score -= BackwardReferencePenaltyUsingLastDistance(i);
+ if (best_score < score) {
+ best_score = score;
+ best_len = len;
+ out->len = best_len;
+ out->distance = backward;
+ out->score = best_score;
+ }
+ }
+ }
+ }
+ }
+ /* we require matches of len >4, so increase best_len to 3, so we can compare
+ * 4 bytes all the time. */
+ if (best_len < 3) {
+ best_len = 3;
+ }
+ {
+ const uint8_t tag = hash & TAG_HASH_MASK;
+ const size_t head = (num[key] + 1) & self->block_mask_;
+ uint64_t matches =
+ GetMatchingTagMask(self->block_size_ / 16, tag, tag_bucket, head);
+ /* Mask off any matches from uninitialized tags. */
+ uint16_t n = 65535 - num[key];
+ uint64_t block_has_unused_slots = self->block_size_ > n;
+ uint64_t mask = (block_has_unused_slots << (n & (64 - 1))) - 1;
+ matches &= mask;
+ for (; matches > 0; matches &= (matches - 1)) {
+ const size_t rb_index =
+ (head + (size_t)BROTLI_TZCNT64(matches)) & self->block_mask_;
+ size_t prev_ix = bucket[rb_index];
+ const size_t backward = cur_ix - prev_ix;
+ if (BROTLI_PREDICT_FALSE(backward > max_backward)) {
+ break;
+ }
+ prev_ix &= ring_buffer_mask;
+ if (cur_ix_masked + best_len > ring_buffer_mask ||
+ prev_ix + best_len > ring_buffer_mask ||
+ /* compare 4 bytes ending at best_len + 1 */
+ BrotliUnalignedRead32(&data[cur_ix_masked + best_len - 3]) !=
+ BrotliUnalignedRead32(&data[prev_ix + best_len - 3])) {
+ continue;
+ }
+ {
+ const size_t len = FindMatchLengthWithLimit(&data[prev_ix],
+ &data[cur_ix_masked],
+ max_length);
+ if (len >= 4) {
+ /* Comparing for >= 3 does not change the semantics, but just saves
+ for a few unnecessary binary logarithms in backward reference
+ score, since we are not interested in such short matches. */
+ score_t score = BackwardReferenceScore(len, backward);
+ if (best_score < score) {
+ best_score = score;
+ best_len = len;
+ out->len = best_len;
+ out->distance = backward;
+ out->score = best_score;
+ }
+ }
+ }
+ }
+ bucket[num[key] & self->block_mask_] = (uint32_t)cur_ix;
+ tag_bucket[num[key] & self->block_mask_] = tag;
+ --num[key];
+ }
+ if (min_score == out->score) {
+ SearchInStaticDictionary(dictionary,
+ self->common_, &data[cur_ix_masked], max_length, dictionary_distance,
+ max_distance, out, BROTLI_FALSE);
+ }
+}
+#undef HashLongestMatch
diff --git a/c/enc/matching_tag_mask.h b/c/enc/matching_tag_mask.h
new file mode 100644
index 0000000..c166255
--- /dev/null
+++ b/c/enc/matching_tag_mask.h
@@ -0,0 +1,68 @@
+#ifndef THIRD_PARTY_BROTLI_ENC_MATCHING_TAG_MASK_H_
+#define THIRD_PARTY_BROTLI_ENC_MATCHING_TAG_MASK_H_
+
+#include "../common/platform.h"
+
+#if defined(__SSE2__) || defined(_M_AMD64) || (defined (_M_IX86) && defined(_M_IX86_FP) && (_M_IX86_FP >= 2))
+#define SUPPORTS_SSE_2
+#endif
+
+#if defined(SUPPORTS_SSE_2)
+#include <immintrin.h>
+#endif
+
+
+static BROTLI_INLINE uint64_t GetMatchingTagMask(
+ size_t chunk_count, const uint8_t tag,
+ const uint8_t* BROTLI_RESTRICT tag_bucket, const size_t head) {
+ uint64_t matches = 0;
+#if defined(SUPPORTS_SSE_2)
+ const __m128i comparison_mask = _mm_set1_epi8((char)tag);
+ size_t i;
+ for (i = 0; i < chunk_count && i < 4; i++) {
+ const __m128i chunk =
+ _mm_loadu_si128((const __m128i*)(const void*)(tag_bucket + 16 * i));
+ const __m128i equal_mask = _mm_cmpeq_epi8(chunk, comparison_mask);
+ matches |= (uint64_t)_mm_movemask_epi8(equal_mask) << 16 * i;
+ }
+#else
+ const int chunk_size = sizeof(size_t);
+ const size_t shift_amount = ((chunk_size * 8) - chunk_size);
+ const size_t xFF = ~((size_t)0);
+ const size_t x01 = xFF / 0xFF;
+ const size_t x80 = x01 << 7;
+ const size_t splat_char = tag * x01;
+ int i = ((int)chunk_count * 16) - chunk_size;
+ BROTLI_DCHECK((sizeof(size_t) == 4) || (sizeof(size_t) == 8));
+#if BROTLI_LITTLE_ENDIAN
+ const size_t extractMagic = (xFF / 0x7F) >> chunk_size;
+ do {
+ size_t chunk = BrotliUnalignedReadSizeT(&tag_bucket[i]);
+ chunk ^= splat_char;
+ chunk = (((chunk | x80) - x01) | chunk) & x80;
+ matches <<= chunk_size;
+ matches |= (chunk * extractMagic) >> shift_amount;
+ i -= chunk_size;
+ } while (i >= 0);
+#else
+ const size_t msb = xFF ^ (xFF >> 1);
+ const size_t extractMagic = (msb / 0x1FF) | msb;
+ do {
+ size_t chunk = BrotliUnalignedReadSizeT(&tag_bucket[i]);
+ chunk ^= splat_char;
+ chunk = (((chunk | x80) - x01) | chunk) & x80;
+ matches <<= chunk_size;
+ matches |= ((chunk >> 7) * extractMagic) >> shift_amount;
+ i -= chunk_size;
+ } while (i >= 0);
+#endif
+ matches = ~matches;
+#endif
+ if (chunk_count == 1) return BrotliRotateRight16((uint16_t)matches, head);
+ if (chunk_count == 2) return BrotliRotateRight32((uint32_t)matches, head);
+ return BrotliRotateRight64(matches, head);
+}
+
+#undef SUPPORTS_SSE_2
+
+#endif // THIRD_PARTY_BROTLI_ENC_MATCHING_TAG_MASK_H_
diff --git a/c/enc/quality.h b/c/enc/quality.h
index ffdfd72..4fe649e 100644
--- a/c/enc/quality.h
+++ b/c/enc/quality.h
@@ -128,16 +128,16 @@
- q04: h04 (longest_match_quickly), b17, l5
- q04: h54 (longest_match_quickly), b20, l7 | for large files
- - q05: h05 (longest_match ), b14, l4
- - q05: h06 (longest_match64 ), b15, l5 | for large files
+ - q05: h58 (longest_match_simd ), b14, l4
+ - q05: h68 (longest_match64_simd ), b15, l5 | for large files
- q05: h40 (forgetful_chain ), b15, l4 | for small window
- - q06: h05 (longest_match ), b14, l4
- - q06: h06 (longest_match64 ), b15, l5 | for large files
+ - q06: h58 (longest_match_simd ), b14, l4
+ - q06: h68 (longest_match64_simd ), b15, l5 | for large files
- q06: h40 (forgetful_chain ), b15, l4 | for small window
- - q07: h05 (longest_match ), b15, l4
- - q07: h06 (longest_match64 ), b15, l5 | for large files
+ - q07: h58 (longest_match_simd ), b15, l4
+ - q07: h68 (longest_match64_simd ), b15, l5 | for large files
- q07: h41 (forgetful_chain ), b15, l4 | for small window
- q08: h05 (longest_match ), b15, l4
@@ -165,7 +165,11 @@
} else if (params->lgwin <= 16) {
hparams->type = params->quality < 7 ? 40 : params->quality < 9 ? 41 : 42;
} else if (params->size_hint >= (1 << 20) && params->lgwin >= 19) {
+#if defined(BROTLI_MAX_SIMD_QUALITY)
+ hparams->type = params->quality <= BROTLI_MAX_SIMD_QUALITY ? 68 : 6;
+#else
hparams->type = 6;
+#endif
hparams->block_bits = params->quality - 1;
hparams->bucket_bits = 15;
hparams->num_last_distances_to_check =
@@ -173,7 +177,11 @@
} else {
/* TODO(eustas): often previous setting (H6) is faster and denser; consider
adding an option to use it. */
+#if defined(BROTLI_MAX_SIMD_QUALITY)
+ hparams->type = params->quality <= BROTLI_MAX_SIMD_QUALITY ? 58 : 5;
+#else
hparams->type = 5;
+#endif
hparams->block_bits = params->quality - 1;
hparams->bucket_bits = params->quality < 7 ? 14 : 15;
hparams->num_last_distances_to_check =
