5 #include <unordered_map>
12 #include <arpa/inet.h>
14 #include <sys/types.h>
20 #define P4NENC_BOUND(n) ((n+127)/128+(n+32)*sizeof(uint32_t))
22 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
25 using namespace std::chrono;
27 string zstd_compress(const string &src, string *tempbuf);
29 static inline uint32_t read_unigram(const string_view s, size_t idx)
32 return (unsigned char)s[idx];
38 static inline uint32_t read_trigram(const string_view s, size_t start)
40 return read_unigram(s, start) |
41 (read_unigram(s, start + 1) << 8) |
42 (read_unigram(s, start + 2) << 16);
47 DBE_NORMAL = 0, /* A non-directory file */
48 DBE_DIRECTORY = 1, /* A directory */
49 DBE_END = 2 /* End of directory contents; contains no name */
58 uint8_t check_visibility;
70 class PostingListBuilder {
72 void add_docid(uint32_t docid);
76 size_t num_docids = 0;
79 void write_header(uint32_t docid);
82 vector<uint32_t> pending_docids;
84 uint32_t last_block_end;
87 void PostingListBuilder::add_docid(uint32_t docid)
89 // Deduplicate against the last inserted value, if any.
90 if (pending_docids.empty()) {
91 if (encoded.empty()) {
95 last_block_end = docid;
97 } else if (docid == last_block_end) {
101 if (docid == pending_docids.back()) {
106 pending_docids.push_back(docid);
107 if (pending_docids.size() == 128) {
109 pending_docids.clear();
110 last_block_end = docid;
115 void PostingListBuilder::finish()
117 if (pending_docids.empty()) {
121 assert(!encoded.empty()); // write_header() should already have run.
123 // No interleaving for partial blocks.
124 unsigned char buf[P4NENC_BOUND(128)];
125 unsigned char *end = p4d1enc32(pending_docids.data(), pending_docids.size(), buf, last_block_end);
126 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
129 void PostingListBuilder::append_block()
131 unsigned char buf[P4NENC_BOUND(128)];
132 assert(pending_docids.size() == 128);
133 unsigned char *end = p4d1enc128v32(pending_docids.data(), 128, buf, last_block_end);
134 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
137 void PostingListBuilder::write_header(uint32_t docid)
139 unsigned char buf[P4NENC_BOUND(1)];
140 size_t bytes = p4nd1enc128v32(&docid, 1, buf);
141 encoded.append(reinterpret_cast<char *>(buf), bytes);
146 Corpus(size_t block_size) : block_size(block_size) {}
147 void add_file(string filename);
150 vector<string> filename_blocks;
151 unordered_map<uint32_t, PostingListBuilder> invindex;
152 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
155 string current_block;
157 const size_t block_size;
160 void Corpus::add_file(string filename)
163 if (!current_block.empty()) {
164 current_block.push_back('\0');
166 current_block += filename;
167 if (++num_files_in_block == block_size) {
172 void Corpus::flush_block()
174 if (current_block.empty()) {
178 uint32_t docid = num_blocks;
181 const char *ptr = current_block.c_str();
182 while (ptr < current_block.c_str() + current_block.size()) {
185 for (size_t j = 0; j < s.size() - 2; ++j) {
186 uint32_t trgm = read_trigram(s, j);
187 invindex[trgm].add_docid(docid);
193 // Compress and add the filename block.
194 filename_blocks.push_back(zstd_compress(current_block, &tempbuf));
196 current_block.clear();
197 num_files_in_block = 0;
201 const char *handle_directory(const char *ptr, Corpus *corpus)
203 ptr += sizeof(db_directory);
205 string dir_path = ptr;
206 ptr += dir_path.size() + 1;
207 if (dir_path == "/") {
212 uint8_t type = *ptr++;
213 if (type == DBE_NORMAL) {
214 string filename = ptr;
215 corpus->add_file(dir_path + "/" + filename);
216 ptr += filename.size() + 1;
217 } else if (type == DBE_DIRECTORY) {
218 string dirname = ptr;
219 corpus->add_file(dir_path + "/" + dirname);
220 ptr += dirname.size() + 1;
227 void read_mlocate(const char *filename, Corpus *corpus)
229 int fd = open(filename, O_RDONLY);
234 off_t len = lseek(fd, 0, SEEK_END);
239 const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
240 if (data == MAP_FAILED) {
245 const db_header *hdr = (const db_header *)data;
247 // TODO: Care about the base path.
248 string path = data + sizeof(db_header);
249 uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
251 const char *ptr = data + offset;
252 while (ptr < data + len) {
253 ptr = handle_directory(ptr, corpus);
256 munmap((void *)data, len);
260 string zstd_compress(const string &src, string *tempbuf)
262 size_t max_size = ZSTD_compressBound(src.size());
263 if (tempbuf->size() < max_size) {
264 tempbuf->resize(max_size);
266 size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
267 return string(tempbuf->data(), size);
270 void do_build(const char *infile, const char *outfile, int block_size)
272 //steady_clock::time_point start = steady_clock::now();
274 Corpus corpus(block_size);
276 read_mlocate(infile, &corpus);
277 if (false) { // To read a plain text file.
278 FILE *fp = fopen(infile, "r");
281 if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
285 if (s.back() == '\n') s.pop_back();
286 corpus.add_file(move(s));
290 corpus.flush_block();
291 dprintf("Read %zu files from %s\n", corpus.num_files, infile);
293 size_t trigrams = 0, longest_posting_list = 0;
294 size_t bytes_used = 0;
295 for (auto &[trigram, pl_builder] : corpus.invindex) {
297 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
298 trigrams += pl_builder.num_docids;
299 bytes_used += pl_builder.encoded.size();
301 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
302 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
303 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
305 //steady_clock::time_point end = steady_clock::now();
306 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(end - start).count());
308 vector<uint32_t> all_trigrams;
309 for (auto &[trigram, pl_builder] : corpus.invindex) {
310 all_trigrams.push_back(trigram);
312 sort(all_trigrams.begin(), all_trigrams.end());
315 FILE *outfp = fopen(outfile, "wb");
318 uint64_t num_trigrams = corpus.invindex.size(); // 64 to get alignment.
319 fwrite(&num_trigrams, sizeof(num_trigrams), 1, outfp);
321 // Find out where the posting lists will end.
322 uint64_t offset = sizeof(uint64_t) * 2 + 16 * all_trigrams.size();
323 uint64_t filename_index_offset = offset;
324 for (auto &[trigram, pl_builder] : corpus.invindex) {
325 filename_index_offset += pl_builder.encoded.size();
327 fwrite(&filename_index_offset, sizeof(filename_index_offset), 1, outfp);
329 size_t bytes_for_trigrams = 0, bytes_for_posting_lists = 0, bytes_for_filename_index = 0, bytes_for_filenames = 0;
330 for (uint32_t trgm : all_trigrams) {
331 // 16 bytes: trgm (4), number of docids (4), byte offset (8)
332 fwrite(&trgm, sizeof(trgm), 1, outfp); // One byte wasted, but OK.
333 const PostingListBuilder &pl_builder = corpus.invindex[trgm];
334 uint32_t num_docids = pl_builder.num_docids;
335 fwrite(&num_docids, sizeof(num_docids), 1, outfp);
336 fwrite(&offset, sizeof(offset), 1, outfp);
337 offset += pl_builder.encoded.size();
338 bytes_for_trigrams += 16;
341 // Write the actual posting lists.
342 for (uint32_t trgm : all_trigrams) {
343 const string &encoded = corpus.invindex[trgm].encoded;
344 fwrite(encoded.data(), encoded.size(), 1, outfp);
345 bytes_for_posting_lists += encoded.size();
348 // Stick an empty block at the end as sentinel.
349 corpus.filename_blocks.push_back("");
351 // Write the offsets to the filenames.
352 offset = filename_index_offset + corpus.filename_blocks.size() * sizeof(offset);
353 for (const string &filename : corpus.filename_blocks) {
354 fwrite(&offset, sizeof(offset), 1, outfp);
355 offset += filename.size();
356 bytes_for_filename_index += sizeof(offset);
357 bytes_for_filenames += filename.size();
360 // Write the actual filenames.
361 for (const string &filename : corpus.filename_blocks) {
362 fwrite(filename.data(), filename.size(), 1, outfp);
367 //size_t total_bytes = (bytes_for_trigrams + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
369 dprintf("Block size: %7d files\n", block_size);
370 dprintf("Trigrams: %'7.1f MB\n", bytes_for_trigrams / 1048576.0);
371 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
372 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
373 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
374 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);
378 int main(int argc, char **argv)
380 do_build(argv[1], argv[2], 32);