hash_longest_match_simd_inc.h (11759B)
1 /* NOLINT(build/header_guard) */ 2 /* Copyright 2010 Google Inc. All Rights Reserved. 3 Distributed under MIT license. 4 See file LICENSE for detail or copy at https://opensource.org/licenses/MIT 5 */ 6 /* template parameters: FN */ 7 /* A (forgetful) hash table to the data seen by the compressor, to 8 help create backward references to previous data. 9 This is a hash map of fixed size (bucket_size_) to a ring buffer of 10 fixed size (block_size_). The ring buffer contains the last block_size_ 11 index positions of the given hash key in the compressed data. */ 12 #define HashLongestMatch HASHER() 13 #define TAG_HASH_BITS 8 14 #define TAG_HASH_MASK ((1 << TAG_HASH_BITS) - 1) 15 static BROTLI_INLINE size_t FN(HashTypeLength)(void) { return 4; } 16 static BROTLI_INLINE size_t FN(StoreLookahead)(void) { return 4; } 17 /* HashBytes is the function that chooses the bucket to place the address in. */ 18 static uint32_t FN(HashBytes)( 19 const uint8_t* BROTLI_RESTRICT data, const int shift) { 20 uint32_t h = BROTLI_UNALIGNED_LOAD32LE(data) * kHashMul32; 21 /* The higher bits contain more mixture from the multiplication, 22 so we take our results from there. */ 23 return (uint32_t)(h >> shift); 24 } 25 typedef struct HashLongestMatch { 26 /* Number of hash buckets. */ 27 size_t bucket_size_; 28 /* Only block_size_ newest backward references are kept, 29 and the older are forgotten. */ 30 size_t block_size_; 31 /* Left-shift for computing hash bucket index from hash value. */ 32 int hash_shift_; 33 /* Mask for accessing entries in a block (in a ring-buffer manner). */ 34 uint32_t block_mask_; 35 int block_bits_; 36 int num_last_distances_to_check_; 37 /* Shortcuts. */ 38 HasherCommon* common_; 39 /* --- Dynamic size members --- */ 40 /* Number of entries in a particular bucket. */ 41 uint16_t* num_; /* uint16_t[bucket_size]; */ 42 uint8_t* tags_; 43 /* Buckets containing block_size_ of backward references. */ 44 uint32_t* buckets_; /* uint32_t[bucket_size * block_size]; */ 45 } HashLongestMatch; 46 static void FN(Initialize)( 47 HasherCommon* common, HashLongestMatch* BROTLI_RESTRICT self, 48 const BrotliEncoderParams* params) { 49 self->common_ = common; 50 BROTLI_UNUSED(params); 51 self->hash_shift_ = 32 - common->params.bucket_bits - TAG_HASH_BITS; 52 self->bucket_size_ = (size_t)1 << common->params.bucket_bits; 53 self->block_size_ = (size_t)1 << common->params.block_bits; 54 self->block_mask_ = (uint32_t)(self->block_size_ - 1); 55 self->num_ = (uint16_t*)common->extra[0]; 56 self->tags_ = (uint8_t*)common->extra[1]; 57 self->buckets_ = (uint32_t*)common->extra[2]; 58 self->block_bits_ = common->params.block_bits; 59 self->num_last_distances_to_check_ = 60 common->params.num_last_distances_to_check; 61 } 62 static void FN(Prepare)( 63 HashLongestMatch* BROTLI_RESTRICT self, BROTLI_BOOL one_shot, 64 size_t input_size, const uint8_t* BROTLI_RESTRICT data) { 65 uint16_t* BROTLI_RESTRICT num = self->num_; 66 /* Partial preparation is 100 times slower (per socket). */ 67 size_t partial_prepare_threshold = self->bucket_size_ >> 6; 68 if (one_shot && input_size <= partial_prepare_threshold) { 69 size_t i; 70 for (i = 0; i < input_size; ++i) { 71 const uint32_t hash = FN(HashBytes)(&data[i], self->hash_shift_); 72 const uint32_t key = hash >> TAG_HASH_BITS; 73 num[key] = 65535; 74 } 75 } else { 76 /* Set all the bytes of num to 255, which makes each uint16_t 65535. */ 77 memset(num, 255, self->bucket_size_ * sizeof(num[0])); 78 } 79 } 80 static BROTLI_INLINE void FN(HashMemAllocInBytes)( 81 const BrotliEncoderParams* params, BROTLI_BOOL one_shot, 82 size_t input_size, size_t* alloc_size) { 83 size_t bucket_size = (size_t)1 << params->hasher.bucket_bits; 84 size_t block_size = (size_t)1 << params->hasher.block_bits; 85 BROTLI_UNUSED(one_shot); 86 BROTLI_UNUSED(input_size); 87 alloc_size[0] = sizeof(uint16_t) * bucket_size; 88 alloc_size[1] = sizeof(uint8_t) * bucket_size * block_size; 89 alloc_size[2] = sizeof(uint32_t) * bucket_size * block_size; 90 } 91 /* Look at 4 bytes at &data[ix & mask]. 92 Compute a hash from these, and store the value of ix at that position. */ 93 static BROTLI_INLINE void FN(Store)( 94 HashLongestMatch* BROTLI_RESTRICT self, const uint8_t* BROTLI_RESTRICT data, 95 const size_t mask, const size_t ix) { 96 uint16_t* BROTLI_RESTRICT num = self->num_; 97 uint8_t* BROTLI_RESTRICT tags = self->tags_; 98 uint32_t* BROTLI_RESTRICT buckets = self->buckets_; 99 const size_t hash = FN(HashBytes)(&data[ix & mask], self->hash_shift_); 100 const size_t key = hash >> TAG_HASH_BITS; 101 const uint8_t tag = hash & TAG_HASH_MASK; 102 const size_t minor_ix = num[key] & self->block_mask_; 103 const size_t offset = minor_ix + (key << self->block_bits_); 104 --num[key]; 105 buckets[offset] = (uint32_t)ix; 106 tags[offset] = tag; 107 } 108 static BROTLI_INLINE void FN(StoreRange)(HashLongestMatch* BROTLI_RESTRICT self, 109 const uint8_t* BROTLI_RESTRICT data, const size_t mask, 110 const size_t ix_start, const size_t ix_end) { 111 size_t i; 112 for (i = ix_start; i < ix_end; ++i) { 113 FN(Store)(self, data, mask, i); 114 } 115 } 116 static BROTLI_INLINE void FN(StitchToPreviousBlock)( 117 HashLongestMatch* BROTLI_RESTRICT self, 118 size_t num_bytes, size_t position, const uint8_t* ringbuffer, 119 size_t ringbuffer_mask) { 120 if (num_bytes >= FN(HashTypeLength)() - 1 && position >= 3) { 121 /* Prepare the hashes for three last bytes of the last write. 122 These could not be calculated before, since they require knowledge 123 of both the previous and the current block. */ 124 FN(Store)(self, ringbuffer, ringbuffer_mask, position - 3); 125 FN(Store)(self, ringbuffer, ringbuffer_mask, position - 2); 126 FN(Store)(self, ringbuffer, ringbuffer_mask, position - 1); 127 } 128 } 129 static BROTLI_INLINE void FN(PrepareDistanceCache)( 130 HashLongestMatch* BROTLI_RESTRICT self, 131 int* BROTLI_RESTRICT distance_cache) { 132 PrepareDistanceCache(distance_cache, self->num_last_distances_to_check_); 133 } 134 135 /* Find a longest backward match of &data[cur_ix] up to the length of 136 max_length and stores the position cur_ix in the hash table. 137 REQUIRES: FN(PrepareDistanceCache) must be invoked for current distance cache 138 values; if this method is invoked repeatedly with the same distance 139 cache values, it is enough to invoke FN(PrepareDistanceCache) once. 140 Does not look for matches longer than max_length. 141 Does not look for matches further away than max_backward. 142 Writes the best match into |out|. 143 |out|->score is updated only if a better match is found. */ 144 static BROTLI_INLINE void FN(FindLongestMatch)( 145 HashLongestMatch* BROTLI_RESTRICT self, 146 const BrotliEncoderDictionary* dictionary, 147 const uint8_t* BROTLI_RESTRICT data, const size_t ring_buffer_mask, 148 const int* BROTLI_RESTRICT distance_cache, const size_t cur_ix, 149 const size_t max_length, const size_t max_backward, 150 const size_t dictionary_distance, const size_t max_distance, 151 HasherSearchResult* BROTLI_RESTRICT out) { 152 uint16_t* BROTLI_RESTRICT num = self->num_; 153 uint32_t* BROTLI_RESTRICT buckets = self->buckets_; 154 uint8_t* BROTLI_RESTRICT tags = self->tags_; 155 const size_t cur_ix_masked = cur_ix & ring_buffer_mask; 156 /* Don't accept a short copy from far away. */ 157 score_t min_score = out->score; 158 score_t best_score = out->score; 159 size_t best_len = out->len; 160 size_t i; 161 /* Precalculate the hash key and prefetch the bucket. */ 162 const uint32_t hash = 163 FN(HashBytes)(&data[cur_ix_masked], self->hash_shift_); 164 const uint32_t key = hash >> TAG_HASH_BITS; 165 uint32_t* BROTLI_RESTRICT bucket = &buckets[key << self->block_bits_]; 166 uint8_t* BROTLI_RESTRICT tag_bucket = &tags[key << self->block_bits_]; 167 PREFETCH_L1(bucket); 168 PREFETCH_L1(tag_bucket); 169 if (self->block_bits_ > 4) PREFETCH_L1(bucket + 16); 170 out->len = 0; 171 out->len_code_delta = 0; 172 173 BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask); 174 175 /* Try last distance first. */ 176 for (i = 0; i < (size_t)self->num_last_distances_to_check_; ++i) { 177 const size_t backward = (size_t)distance_cache[i]; 178 size_t prev_ix = (size_t)(cur_ix - backward); 179 if (prev_ix >= cur_ix) { 180 continue; 181 } 182 if (BROTLI_PREDICT_FALSE(backward > max_backward)) { 183 continue; 184 } 185 prev_ix &= ring_buffer_mask; 186 187 if (cur_ix_masked + best_len > ring_buffer_mask) { 188 break; 189 } 190 if (prev_ix + best_len > ring_buffer_mask || 191 data[cur_ix_masked + best_len] != data[prev_ix + best_len]) { 192 continue; 193 } 194 { 195 const size_t len = FindMatchLengthWithLimit(&data[prev_ix], 196 &data[cur_ix_masked], 197 max_length); 198 if (len >= 3 || (len == 2 && i < 2)) { 199 /* Comparing for >= 2 does not change the semantics, but just saves for 200 a few unnecessary binary logarithms in backward reference score, 201 since we are not interested in such short matches. */ 202 score_t score = BackwardReferenceScoreUsingLastDistance(len); 203 if (best_score < score) { 204 if (i != 0) score -= BackwardReferencePenaltyUsingLastDistance(i); 205 if (best_score < score) { 206 best_score = score; 207 best_len = len; 208 out->len = best_len; 209 out->distance = backward; 210 out->score = best_score; 211 } 212 } 213 } 214 } 215 } 216 /* we require matches of len >4, so increase best_len to 3, so we can compare 217 * 4 bytes all the time. */ 218 if (best_len < 3) { 219 best_len = 3; 220 } 221 { 222 const uint8_t tag = hash & TAG_HASH_MASK; 223 const size_t head = (num[key] + 1) & self->block_mask_; 224 uint64_t matches = 225 GetMatchingTagMask(self->block_size_ / 16, tag, tag_bucket, head); 226 /* Mask off any matches from uninitialized tags. */ 227 uint16_t n = 65535 - num[key]; 228 uint64_t block_has_unused_slots = self->block_size_ > n; 229 uint64_t mask = (block_has_unused_slots << (n & (64 - 1))) - 1; 230 matches &= mask; 231 for (; matches > 0; matches &= (matches - 1)) { 232 const size_t rb_index = 233 (head + (size_t)BROTLI_TZCNT64(matches)) & self->block_mask_; 234 size_t prev_ix = bucket[rb_index]; 235 const size_t backward = cur_ix - prev_ix; 236 if (BROTLI_PREDICT_FALSE(backward > max_backward)) { 237 break; 238 } 239 prev_ix &= ring_buffer_mask; 240 if (cur_ix_masked + best_len > ring_buffer_mask) { 241 break; 242 } 243 if (prev_ix + best_len > ring_buffer_mask || 244 /* compare 4 bytes ending at best_len + 1 */ 245 BrotliUnalignedRead32(&data[cur_ix_masked + best_len - 3]) != 246 BrotliUnalignedRead32(&data[prev_ix + best_len - 3])) { 247 continue; 248 } 249 { 250 const size_t len = FindMatchLengthWithLimit(&data[prev_ix], 251 &data[cur_ix_masked], 252 max_length); 253 if (len >= 4) { 254 /* Comparing for >= 3 does not change the semantics, but just saves 255 for a few unnecessary binary logarithms in backward reference 256 score, since we are not interested in such short matches. */ 257 score_t score = BackwardReferenceScore(len, backward); 258 if (best_score < score) { 259 best_score = score; 260 best_len = len; 261 out->len = best_len; 262 out->distance = backward; 263 out->score = best_score; 264 } 265 } 266 } 267 } 268 bucket[num[key] & self->block_mask_] = (uint32_t)cur_ix; 269 tag_bucket[num[key] & self->block_mask_] = tag; 270 --num[key]; 271 } 272 if (min_score == out->score) { 273 SearchInStaticDictionary(dictionary, 274 self->common_, &data[cur_ix_masked], max_length, dictionary_distance, 275 max_distance, out, BROTLI_FALSE); 276 } 277 } 278 #undef HashLongestMatch