git.sesse.net Git - plocate/blob - database-builder.cpp

   1 #include "database-builder.h"
   2
   3 #include "dprintf.h"
   4 #include "turbopfor-encode.h"
   5
   6 #include <algorithm>
   7 #include <assert.h>
   8 #include <fcntl.h>
   9 #include <string.h>
  10 #include <string_view>
  11 #include <sys/stat.h>
  12 #include <sys/time.h>
  13 #include <sys/types.h>
  14 #include <zdict.h>
  15 #include <zstd.h>
  16
  17 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
  18
  19 #define NUM_TRIGRAMS 16777216
  20
  21 using namespace std;
  22 using namespace std::chrono;
  23
  24 constexpr unsigned num_overflow_slots = 16;
  25
  26 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf);
  27
  28 static inline uint32_t read_unigram(const string_view s, size_t idx)
  29 {
  30         if (idx < s.size()) {
  31                 return (unsigned char)s[idx];
  32         } else {
  33                 return 0;
  34         }
  35 }
  36
  37 static inline uint32_t read_trigram(const string_view s, size_t start)
  38 {
  39         return read_unigram(s, start) |
  40                 (read_unigram(s, start + 1) << 8) |
  41                 (read_unigram(s, start + 2) << 16);
  42 }
  43
  44 class PostingListBuilder {
  45 public:
  46         inline void add_docid(uint32_t docid);
  47         void finish();
  48
  49         string encoded;
  50         size_t num_docids = 0;
  51
  52 private:
  53         void write_header(uint32_t docid);
  54         void append_block();
  55
  56         vector<uint32_t> pending_deltas;
  57
  58         uint32_t last_block_end, last_docid = -1;
  59 };
  60
  61 void PostingListBuilder::add_docid(uint32_t docid)
  62 {
  63         // Deduplicate against the last inserted value, if any.
  64         if (docid == last_docid) {
  65                 return;
  66         }
  67
  68         if (num_docids == 0) {
  69                 // Very first docid.
  70                 write_header(docid);
  71                 ++num_docids;
  72                 last_block_end = last_docid = docid;
  73                 return;
  74         }
  75
  76         pending_deltas.push_back(docid - last_docid - 1);
  77         last_docid = docid;
  78         if (pending_deltas.size() == 128) {
  79                 append_block();
  80                 pending_deltas.clear();
  81                 last_block_end = docid;
  82         }
  83         ++num_docids;
  84 }
  85
  86 void PostingListBuilder::finish()
  87 {
  88         if (pending_deltas.empty()) {
  89                 return;
  90         }
  91
  92         assert(!encoded.empty());  // write_header() should already have run.
  93
  94         // No interleaving for partial blocks.
  95         unsigned char buf[P4NENC_BOUND(128)];
  96         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
  97         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
  98 }
  99
 100 void PostingListBuilder::append_block()
 101 {
 102         unsigned char buf[P4NENC_BOUND(128)];
 103         assert(pending_deltas.size() == 128);
 104         unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
 105         encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
 106 }
 107
 108 void PostingListBuilder::write_header(uint32_t docid)
 109 {
 110         unsigned char buf[P4NENC_BOUND(1)];
 111         unsigned char *end = write_baseval(docid, buf);
 112         encoded.append(reinterpret_cast<char *>(buf), end - buf);
 113 }
 114
 115 void DictionaryBuilder::add_file(string filename)
 116 {
 117         if (keep_current_block) {  // Only bother saving the filenames if we're actually keeping the block.
 118                 if (!current_block.empty()) {
 119                         current_block.push_back('\0');
 120                 }
 121                 current_block += filename;
 122         }
 123         if (++num_files_in_block == block_size) {
 124                 flush_block();
 125         }
 126 }
 127
 128 void DictionaryBuilder::flush_block()
 129 {
 130         if (keep_current_block) {
 131                 if (slot_for_current_block == -1) {
 132                         lengths.push_back(current_block.size());
 133                         sampled_blocks.push_back(move(current_block));
 134                 } else {
 135                         lengths[slot_for_current_block] = current_block.size();
 136                         sampled_blocks[slot_for_current_block] = move(current_block);
 137                 }
 138         }
 139         current_block.clear();
 140         num_files_in_block = 0;
 141         ++block_num;
 142
 143         if (block_num < blocks_to_keep) {
 144                 keep_current_block = true;
 145                 slot_for_current_block = -1;
 146         } else {
 147                 // Keep every block with equal probability (reservoir sampling).
 148                 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
 149                 keep_current_block = (idx < blocks_to_keep);
 150                 slot_for_current_block = idx;
 151         }
 152 }
 153
 154 string DictionaryBuilder::train(size_t buf_size)
 155 {
 156         string dictionary_buf;
 157         sort(sampled_blocks.begin(), sampled_blocks.end());  // Seemingly important for decompression speed.
 158         for (const string &block : sampled_blocks) {
 159                 dictionary_buf += block;
 160         }
 161
 162         string buf;
 163         buf.resize(buf_size);
 164         size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
 165         if (ret == size_t(-1)) {
 166                 return "";
 167         }
 168         dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
 169         buf.resize(ret);
 170
 171         sampled_blocks.clear();
 172         lengths.clear();
 173
 174         return buf;
 175 }
 176
 177 Corpus::Corpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict)
 178         : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), cdict(cdict)
 179 {
 180         fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
 181 }
 182
 183 Corpus::~Corpus()
 184 {
 185         for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
 186                 delete invindex[i];
 187         }
 188 }
 189
 190 PostingListBuilder &Corpus::get_pl_builder(uint32_t trgm)
 191 {
 192         if (invindex[trgm] == nullptr) {
 193                 invindex[trgm] = new PostingListBuilder;
 194         }
 195         return *invindex[trgm];
 196 }
 197
 198 void Corpus::add_file(string filename)
 199 {
 200         ++num_files;
 201         if (!current_block.empty()) {
 202                 current_block.push_back('\0');
 203         }
 204         current_block += filename;
 205         if (++num_files_in_block == block_size) {
 206                 flush_block();
 207         }
 208 }
 209
 210 void Corpus::flush_block()
 211 {
 212         if (current_block.empty()) {
 213                 return;
 214         }
 215
 216         uint32_t docid = num_blocks;
 217
 218         // Create trigrams.
 219         const char *ptr = current_block.c_str();
 220         while (ptr < current_block.c_str() + current_block.size()) {
 221                 string_view s(ptr);
 222                 if (s.size() >= 3) {
 223                         for (size_t j = 0; j < s.size() - 2; ++j) {
 224                                 uint32_t trgm = read_trigram(s, j);
 225                                 get_pl_builder(trgm).add_docid(docid);
 226                         }
 227                 }
 228                 ptr += s.size() + 1;
 229         }
 230
 231         // Compress and add the filename block.
 232         filename_blocks.push_back(ftell(outfp));
 233         string compressed = zstd_compress(current_block, cdict, &tempbuf);
 234         if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
 235                 perror("fwrite()");
 236                 exit(1);
 237         }
 238
 239         current_block.clear();
 240         num_files_in_block = 0;
 241         ++num_blocks;
 242 }
 243
 244 void Corpus::finish()
 245 {
 246         flush_block();
 247         string tempbuf;
 248         string compressed = zstd_compress(directory_data, nullptr, &tempbuf);
 249         fprintf(stderr, "dirdata: %zu bytes (compressed: %zu)\n", directory_data.size(), compressed.size());
 250 }
 251
 252 size_t Corpus::num_trigrams() const
 253 {
 254         size_t num = 0;
 255         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 256                 if (invindex[trgm] != nullptr) {
 257                         ++num;
 258                 }
 259         }
 260         return num;
 261 }
 262
 263 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
 264 {
 265         static ZSTD_CCtx *ctx = nullptr;
 266         if (ctx == nullptr) {
 267                 ctx = ZSTD_createCCtx();
 268         }
 269
 270         size_t max_size = ZSTD_compressBound(src.size());
 271         if (tempbuf->size() < max_size) {
 272                 tempbuf->resize(max_size);
 273         }
 274         size_t size;
 275         if (cdict == nullptr) {
 276                 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
 277         } else {
 278                 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
 279         }
 280         return string(tempbuf->data(), size);
 281 }
 282
 283 bool is_prime(uint32_t x)
 284 {
 285         if ((x % 2) == 0 || (x % 3) == 0) {
 286                 return false;
 287         }
 288         uint32_t limit = ceil(sqrt(x));
 289         for (uint32_t factor = 5; factor <= limit; ++factor) {
 290                 if ((x % factor) == 0) {
 291                         return false;
 292                 }
 293         }
 294         return true;
 295 }
 296
 297 uint32_t next_prime(uint32_t x)
 298 {
 299         if ((x % 2) == 0) {
 300                 ++x;
 301         }
 302         while (!is_prime(x)) {
 303                 x += 2;
 304         }
 305         return x;
 306 }
 307
 308 unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
 309 {
 310         unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]);  // 1 for the sentinel element at the end.
 311         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 312                 ht[i].trgm = uint32_t(-1);
 313                 ht[i].num_docids = 0;
 314                 ht[i].offset = 0;
 315         }
 316         for (uint32_t trgm : all_trigrams) {
 317                 // We don't know offset yet, so set it to zero.
 318                 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
 319
 320                 uint32_t bucket = hash_trigram(trgm, ht_size);
 321                 unsigned distance = 0;
 322                 while (ht[bucket].num_docids != 0) {
 323                         // Robin Hood hashing; reduces the longest distance by a lot.
 324                         unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
 325                         if (distance > other_distance) {
 326                                 swap(to_insert, ht[bucket]);
 327                                 distance = other_distance;
 328                         }
 329
 330                         ++bucket, ++distance;
 331                         if (distance > num_overflow_slots) {
 332                                 return nullptr;
 333                         }
 334                 }
 335                 ht[bucket] = to_insert;
 336         }
 337         return ht;
 338 }
 339
 340 DatabaseBuilder::DatabaseBuilder(const char *outfile, int block_size, string dictionary)
 341         : block_size(block_size)
 342 {
 343         umask(0027);
 344         outfp = fopen(outfile, "wb");
 345         if (outfp == nullptr) {
 346                 perror(outfile);
 347                 exit(1);
 348         }
 349
 350         // Write the header.
 351         memcpy(hdr.magic, "\0plocate", 8);
 352         hdr.version = -1;  // Mark as broken.
 353         hdr.hashtable_size = 0;  // Not known yet.
 354         hdr.extra_ht_slots = num_overflow_slots;
 355         hdr.num_docids = 0;
 356         hdr.hash_table_offset_bytes = -1;  // We don't know these offsets yet.
 357         hdr.max_version = 1;
 358         hdr.filename_index_offset_bytes = -1;
 359         hdr.zstd_dictionary_length_bytes = -1;
 360         fwrite(&hdr, sizeof(hdr), 1, outfp);
 361
 362         if (dictionary.empty()) {
 363                 hdr.zstd_dictionary_offset_bytes = 0;
 364                 hdr.zstd_dictionary_length_bytes = 0;
 365         } else {
 366                 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
 367                 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
 368                 hdr.zstd_dictionary_length_bytes = dictionary.size();
 369                 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
 370         }
 371 }
 372
 373 Corpus *DatabaseBuilder::start_corpus()
 374 {
 375         corpus_start = steady_clock::now();
 376         corpus = new Corpus(outfp, block_size, cdict);
 377         return corpus;
 378 }
 379
 380 void DatabaseBuilder::finish_corpus()
 381 {
 382         corpus->finish();
 383         hdr.num_docids = corpus->filename_blocks.size();
 384
 385         // Stick an empty block at the end as sentinel.
 386         corpus->filename_blocks.push_back(ftell(outfp));
 387         const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
 388
 389         // Write the offsets to the filenames.
 390         hdr.filename_index_offset_bytes = ftell(outfp);
 391         const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
 392         fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
 393         corpus->filename_blocks.clear();
 394         corpus->filename_blocks.shrink_to_fit();
 395
 396         // Finish up encoding the posting lists.
 397         size_t trigrams = 0, longest_posting_list = 0;
 398         size_t bytes_for_posting_lists = 0;
 399         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 400                 if (!corpus->seen_trigram(trgm))
 401                         continue;
 402                 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
 403                 pl_builder.finish();
 404                 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
 405                 trigrams += pl_builder.num_docids;
 406                 bytes_for_posting_lists += pl_builder.encoded.size();
 407         }
 408         size_t num_trigrams = corpus->num_trigrams();
 409         dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
 410                 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
 411         dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
 412
 413         dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
 414
 415         // Find the used trigrams.
 416         vector<uint32_t> all_trigrams;
 417         for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
 418                 if (corpus->seen_trigram(trgm)) {
 419                         all_trigrams.push_back(trgm);
 420                 }
 421         }
 422
 423         // Create the hash table.
 424         unique_ptr<Trigram[]> hashtable;
 425         uint32_t ht_size = next_prime(all_trigrams.size());
 426         for (;;) {
 427                 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
 428                 if (hashtable == nullptr) {
 429                         dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
 430                         ht_size = next_prime(ht_size * 1.05);
 431                 } else {
 432                         dprintf("Created hash table of size %u.\n\n", ht_size);
 433                         break;
 434                 }
 435         }
 436
 437         // Find the offsets for each posting list.
 438         size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
 439         uint64_t offset = ftell(outfp) + bytes_for_hashtable;
 440         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 441                 hashtable[i].offset = offset;  // Needs to be there even for empty slots.
 442                 if (hashtable[i].num_docids == 0) {
 443                         continue;
 444                 }
 445
 446                 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 447                 offset += encoded.size();
 448         }
 449
 450         // Write the hash table.
 451         hdr.hash_table_offset_bytes = ftell(outfp);
 452         hdr.hashtable_size = ht_size;
 453         fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 454
 455         // Write the actual posting lists.
 456         for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
 457                 if (hashtable[i].num_docids == 0) {
 458                         continue;
 459                 }
 460                 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
 461                 fwrite(encoded.data(), encoded.size(), 1, outfp);
 462         }
 463
 464         // Rewind, and write the updated header.
 465         hdr.version = 1;
 466         fseek(outfp, 0, SEEK_SET);
 467         fwrite(&hdr, sizeof(hdr), 1, outfp);
 468         fclose(outfp);
 469
 470         size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
 471
 472         dprintf("Block size:     %7d files\n", block_size);
 473         dprintf("Dictionary:     %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
 474         dprintf("Hash table:     %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
 475         dprintf("Posting lists:  %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
 476         dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
 477         dprintf("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
 478         dprintf("Total:          %'7.1f MB\n", total_bytes / 1048576.0);
 479         dprintf("\n");
 480 }