15 #include <sys/types.h>
17 #include <unordered_map>
21 #include "turbopfor-encode.h"
23 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
25 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
27 #define NUM_TRIGRAMS 16777216
30 using namespace std::chrono;
32 string zstd_compress(const string &src, string *tempbuf);
34 constexpr unsigned num_overflow_slots = 16;
36 static inline uint32_t read_unigram(const string_view s, size_t idx)
39 return (unsigned char)s[idx];
45 static inline uint32_t read_trigram(const string_view s, size_t start)
47 return read_unigram(s, start) |
48 (read_unigram(s, start + 1) << 8) |
49 (read_unigram(s, start + 2) << 16);
53 DBE_NORMAL = 0, /* A non-directory file */
54 DBE_DIRECTORY = 1, /* A directory */
55 DBE_END = 2 /* End of directory contents; contains no name */
63 uint8_t check_visibility;
74 class PostingListBuilder {
76 inline void add_docid(uint32_t docid);
80 size_t num_docids = 0;
83 void write_header(uint32_t docid);
86 vector<uint32_t> pending_deltas;
88 uint32_t last_block_end, last_docid = -1;
91 void PostingListBuilder::add_docid(uint32_t docid)
93 // Deduplicate against the last inserted value, if any.
94 if (docid == last_docid) {
98 if (num_docids == 0) {
102 last_block_end = last_docid = docid;
106 pending_deltas.push_back(docid - last_docid - 1);
108 if (pending_deltas.size() == 128) {
110 pending_deltas.clear();
111 last_block_end = docid;
116 void PostingListBuilder::finish()
118 if (pending_deltas.empty()) {
122 assert(!encoded.empty()); // write_header() should already have run.
124 // No interleaving for partial blocks.
125 unsigned char buf[P4NENC_BOUND(128)];
126 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
127 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
130 void PostingListBuilder::append_block()
132 unsigned char buf[P4NENC_BOUND(128)];
133 assert(pending_deltas.size() == 128);
134 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
135 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
138 void PostingListBuilder::write_header(uint32_t docid)
140 unsigned char buf[P4NENC_BOUND(1)];
141 unsigned char *end = write_baseval(docid, buf);
142 encoded.append(reinterpret_cast<char *>(buf), end - buf);
147 Corpus(FILE *outfp, size_t block_size)
148 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size) {}
149 void add_file(string filename);
152 vector<uint64_t> filename_blocks;
153 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
154 bool seen_trigram(uint32_t trgm)
156 return invindex[trgm] != nullptr;
158 PostingListBuilder &get_pl_builder(uint32_t trgm)
160 if (invindex[trgm] == nullptr) {
161 invindex[trgm] = new PostingListBuilder;
163 return *invindex[trgm];
167 unique_ptr<PostingListBuilder *[]> invindex;
169 string current_block;
171 const size_t block_size;
174 void Corpus::add_file(string filename)
177 if (!current_block.empty()) {
178 current_block.push_back('\0');
180 current_block += filename;
181 if (++num_files_in_block == block_size) {
186 void Corpus::flush_block()
188 if (current_block.empty()) {
192 uint32_t docid = num_blocks;
195 const char *ptr = current_block.c_str();
196 while (ptr < current_block.c_str() + current_block.size()) {
199 for (size_t j = 0; j < s.size() - 2; ++j) {
200 uint32_t trgm = read_trigram(s, j);
201 get_pl_builder(trgm).add_docid(docid);
207 // Compress and add the filename block.
208 filename_blocks.push_back(ftell(outfp));
209 string compressed = zstd_compress(current_block, &tempbuf);
210 if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
215 current_block.clear();
216 num_files_in_block = 0;
220 string read_cstr(FILE *fp)
236 void handle_directory(FILE *fp, Corpus *corpus)
239 if (fread(&dummy, sizeof(dummy), 1, fp) != 1) {
247 string dir_path = read_cstr(fp);
248 if (dir_path == "/") {
254 if (type == DBE_NORMAL) {
255 string filename = read_cstr(fp);
256 corpus->add_file(dir_path + "/" + filename);
257 } else if (type == DBE_DIRECTORY) {
258 string dirname = read_cstr(fp);
259 corpus->add_file(dir_path + "/" + dirname);
261 return; // Probably end.
266 void read_mlocate(const char *filename, Corpus *corpus)
268 FILE *fp = fopen(filename, "rb");
275 if (fread(&hdr, sizeof(hdr), 1, fp) != 1) {
276 perror("short read");
280 // TODO: Care about the base path.
281 string path = read_cstr(fp);
283 handle_directory(fp, corpus);
288 string zstd_compress(const string &src, string *tempbuf)
290 size_t max_size = ZSTD_compressBound(src.size());
291 if (tempbuf->size() < max_size) {
292 tempbuf->resize(max_size);
294 size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
295 return string(tempbuf->data(), size);
298 bool is_prime(uint32_t x)
300 if ((x % 2) == 0 || (x % 3) == 0) {
303 uint32_t limit = ceil(sqrt(x));
304 for (uint32_t factor = 5; factor <= limit; ++factor) {
305 if ((x % factor) == 0) {
312 uint32_t next_prime(uint32_t x)
317 while (!is_prime(x)) {
323 unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
325 unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
326 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
327 ht[i].trgm = uint32_t(-1);
328 ht[i].num_docids = 0;
331 for (uint32_t trgm : all_trigrams) {
332 // We don't know offset yet, so set it to zero.
333 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
335 uint32_t bucket = hash_trigram(trgm, ht_size);
336 unsigned distance = 0;
337 while (ht[bucket].num_docids != 0) {
338 // Robin Hood hashing; reduces the longest distance by a lot.
339 unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
340 if (distance > other_distance) {
341 swap(to_insert, ht[bucket]);
342 distance = other_distance;
345 ++bucket, ++distance;
346 if (distance > num_overflow_slots) {
350 ht[bucket] = to_insert;
355 void do_build(const char *infile, const char *outfile, int block_size)
357 steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
360 FILE *outfp = fopen(outfile, "wb");
364 memcpy(hdr.magic, "\0plocate", 8);
365 hdr.version = -1; // Mark as broken.
366 hdr.hashtable_size = 0; // Not known yet.
367 hdr.extra_ht_slots = num_overflow_slots;
368 hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet.
369 hdr.filename_index_offset_bytes = -1;
370 fwrite(&hdr, sizeof(hdr), 1, outfp);
372 Corpus corpus(outfp, block_size);
374 read_mlocate(infile, &corpus);
375 if (false) { // To read a plain text file.
376 FILE *fp = fopen(infile, "r");
379 if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
383 if (s.back() == '\n')
385 corpus.add_file(move(s));
389 corpus.flush_block();
390 dprintf("Read %zu files from %s\n", corpus.num_files, infile);
391 hdr.num_docids = corpus.filename_blocks.size();
393 // Stick an empty block at the end as sentinel.
394 corpus.filename_blocks.push_back(ftell(outfp));
395 const size_t bytes_for_filenames = corpus.filename_blocks.back() - corpus.filename_blocks.front();
397 // Write the offsets to the filenames.
398 hdr.filename_index_offset_bytes = ftell(outfp);
399 const size_t bytes_for_filename_index = corpus.filename_blocks.size() * sizeof(uint64_t);
400 fwrite(corpus.filename_blocks.data(), corpus.filename_blocks.size(), sizeof(uint64_t), outfp);
401 corpus.filename_blocks.clear();
402 corpus.filename_blocks.shrink_to_fit();
404 // Finish up encoding the posting lists.
405 size_t trigrams = 0, longest_posting_list = 0;
406 size_t bytes_for_posting_lists = 0;
407 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
408 if (!corpus.seen_trigram(trgm))
410 PostingListBuilder &pl_builder = corpus.get_pl_builder(trgm);
412 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
413 trigrams += pl_builder.num_docids;
414 bytes_for_posting_lists += pl_builder.encoded.size();
416 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
417 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
418 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
420 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - start).count());
422 // Find the used trigrams.
423 vector<uint32_t> all_trigrams;
424 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
425 if (corpus.seen_trigram(trgm)) {
426 all_trigrams.push_back(trgm);
430 // Create the hash table.
431 unique_ptr<Trigram[]> hashtable;
432 uint32_t ht_size = next_prime(all_trigrams.size());
434 hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots);
435 if (hashtable == nullptr) {
436 dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
437 ht_size = next_prime(ht_size * 1.05);
439 dprintf("Created hash table of size %u.\n\n", ht_size);
444 // Find the offsets for each posting list.
445 size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
446 uint64_t offset = ftell(outfp) + bytes_for_hashtable;
447 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
448 hashtable[i].offset = offset; // Needs to be there even for empty slots.
449 if (hashtable[i].num_docids == 0) {
453 const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
454 offset += encoded.size();
457 // Write the hash table.
458 hdr.hash_table_offset_bytes = ftell(outfp);
459 hdr.hashtable_size = ht_size;
460 fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
462 // Write the actual posting lists.
463 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
464 if (hashtable[i].num_docids == 0) {
467 const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
468 fwrite(encoded.data(), encoded.size(), 1, outfp);
471 // Rewind, and write the updated header.
473 fseek(outfp, 0, SEEK_SET);
474 fwrite(&hdr, sizeof(hdr), 1, outfp);
477 size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
479 dprintf("Block size: %7d files\n", block_size);
480 dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
481 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
482 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
483 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
484 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);
488 int main(int argc, char **argv)
490 do_build(argv[1], argv[2], 32);