5 #include <unordered_map>
12 #include <arpa/inet.h>
14 #include <sys/types.h>
22 #define P4NENC_BOUND(n) ((n+127)/128+(n+32)*sizeof(uint32_t))
24 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
27 using namespace std::chrono;
29 string zstd_compress(const string &src, string *tempbuf);
31 static inline uint32_t read_unigram(const string_view s, size_t idx)
34 return (unsigned char)s[idx];
40 static inline uint32_t read_trigram(const string_view s, size_t start)
42 return read_unigram(s, start) |
43 (read_unigram(s, start + 1) << 8) |
44 (read_unigram(s, start + 2) << 16);
49 DBE_NORMAL = 0, /* A non-directory file */
50 DBE_DIRECTORY = 1, /* A directory */
51 DBE_END = 2 /* End of directory contents; contains no name */
60 uint8_t check_visibility;
72 class PostingListBuilder {
74 void add_docid(uint32_t docid);
78 size_t num_docids = 0;
81 void write_header(uint32_t docid);
84 vector<uint32_t> pending_docids;
86 uint32_t last_block_end;
89 void PostingListBuilder::add_docid(uint32_t docid)
91 // Deduplicate against the last inserted value, if any.
92 if (pending_docids.empty()) {
93 if (encoded.empty()) {
97 last_block_end = docid;
99 } else if (docid == last_block_end) {
103 if (docid == pending_docids.back()) {
108 pending_docids.push_back(docid);
109 if (pending_docids.size() == 128) {
111 pending_docids.clear();
112 last_block_end = docid;
117 void PostingListBuilder::finish()
119 if (pending_docids.empty()) {
123 assert(!encoded.empty()); // write_header() should already have run.
125 // No interleaving for partial blocks.
126 unsigned char buf[P4NENC_BOUND(128)];
127 unsigned char *end = p4d1enc32(pending_docids.data(), pending_docids.size(), buf, last_block_end);
128 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
131 void PostingListBuilder::append_block()
133 unsigned char buf[P4NENC_BOUND(128)];
134 assert(pending_docids.size() == 128);
135 unsigned char *end = p4d1enc128v32(pending_docids.data(), 128, buf, last_block_end);
136 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
139 void PostingListBuilder::write_header(uint32_t docid)
141 unsigned char buf[P4NENC_BOUND(1)];
142 size_t bytes = p4nd1enc128v32(&docid, 1, buf);
143 encoded.append(reinterpret_cast<char *>(buf), bytes);
148 Corpus(size_t block_size) : block_size(block_size) {}
149 void add_file(string filename);
152 vector<string> filename_blocks;
153 unordered_map<uint32_t, PostingListBuilder> invindex;
154 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
157 string current_block;
159 const size_t block_size;
162 void Corpus::add_file(string filename)
165 if (!current_block.empty()) {
166 current_block.push_back('\0');
168 current_block += filename;
169 if (++num_files_in_block == block_size) {
174 void Corpus::flush_block()
176 if (current_block.empty()) {
180 uint32_t docid = num_blocks;
183 const char *ptr = current_block.c_str();
184 while (ptr < current_block.c_str() + current_block.size()) {
187 for (size_t j = 0; j < s.size() - 2; ++j) {
188 uint32_t trgm = read_trigram(s, j);
189 invindex[trgm].add_docid(docid);
195 // Compress and add the filename block.
196 filename_blocks.push_back(zstd_compress(current_block, &tempbuf));
198 current_block.clear();
199 num_files_in_block = 0;
203 const char *handle_directory(const char *ptr, Corpus *corpus)
205 ptr += sizeof(db_directory);
207 string dir_path = ptr;
208 ptr += dir_path.size() + 1;
209 if (dir_path == "/") {
214 uint8_t type = *ptr++;
215 if (type == DBE_NORMAL) {
216 string filename = ptr;
217 corpus->add_file(dir_path + "/" + filename);
218 ptr += filename.size() + 1;
219 } else if (type == DBE_DIRECTORY) {
220 string dirname = ptr;
221 corpus->add_file(dir_path + "/" + dirname);
222 ptr += dirname.size() + 1;
229 void read_mlocate(const char *filename, Corpus *corpus)
231 int fd = open(filename, O_RDONLY);
236 off_t len = lseek(fd, 0, SEEK_END);
241 const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
242 if (data == MAP_FAILED) {
247 const db_header *hdr = (const db_header *)data;
249 // TODO: Care about the base path.
250 string path = data + sizeof(db_header);
251 uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
253 const char *ptr = data + offset;
254 while (ptr < data + len) {
255 ptr = handle_directory(ptr, corpus);
258 munmap((void *)data, len);
262 string zstd_compress(const string &src, string *tempbuf)
264 size_t max_size = ZSTD_compressBound(src.size());
265 if (tempbuf->size() < max_size) {
266 tempbuf->resize(max_size);
268 size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
269 return string(tempbuf->data(), size);
272 void do_build(const char *infile, const char *outfile, int block_size)
274 //steady_clock::time_point start = steady_clock::now();
276 Corpus corpus(block_size);
278 read_mlocate(infile, &corpus);
279 if (false) { // To read a plain text file.
280 FILE *fp = fopen(infile, "r");
283 if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
287 if (s.back() == '\n') s.pop_back();
288 corpus.add_file(move(s));
292 corpus.flush_block();
293 dprintf("Read %zu files from %s\n", corpus.num_files, infile);
295 size_t trigrams = 0, longest_posting_list = 0;
296 size_t bytes_used = 0;
297 for (auto &[trigram, pl_builder] : corpus.invindex) {
299 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
300 trigrams += pl_builder.num_docids;
301 bytes_used += pl_builder.encoded.size();
303 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
304 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
305 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
307 //steady_clock::time_point end = steady_clock::now();
308 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(end - start).count());
310 vector<uint32_t> all_trigrams;
311 for (auto &[trigram, pl_builder] : corpus.invindex) {
312 all_trigrams.push_back(trigram);
314 sort(all_trigrams.begin(), all_trigrams.end());
317 FILE *outfp = fopen(outfile, "wb");
320 uint64_t num_trigrams = corpus.invindex.size(); // 64 to get alignment.
321 fwrite(&num_trigrams, sizeof(num_trigrams), 1, outfp);
323 // Find out where the posting lists will end.
324 uint64_t offset = sizeof(uint64_t) * 2 + 16 * all_trigrams.size();
325 uint64_t filename_index_offset = offset;
326 for (auto &[trigram, pl_builder] : corpus.invindex) {
327 filename_index_offset += pl_builder.encoded.size();
329 fwrite(&filename_index_offset, sizeof(filename_index_offset), 1, outfp);
331 size_t bytes_for_trigrams = 0, bytes_for_posting_lists = 0, bytes_for_filename_index = 0, bytes_for_filenames = 0;
332 for (uint32_t trgm : all_trigrams) {
333 // 16 bytes: trgm (4), number of docids (4), byte offset (8)
334 fwrite(&trgm, sizeof(trgm), 1, outfp); // One byte wasted, but OK.
335 const PostingListBuilder &pl_builder = corpus.invindex[trgm];
336 uint32_t num_docids = pl_builder.num_docids;
337 fwrite(&num_docids, sizeof(num_docids), 1, outfp);
338 fwrite(&offset, sizeof(offset), 1, outfp);
339 offset += pl_builder.encoded.size();
340 bytes_for_trigrams += 16;
343 // Write the actual posting lists.
344 for (uint32_t trgm : all_trigrams) {
345 const string &encoded = corpus.invindex[trgm].encoded;
346 fwrite(encoded.data(), encoded.size(), 1, outfp);
347 bytes_for_posting_lists += encoded.size();
350 // Stick an empty block at the end as sentinel.
351 corpus.filename_blocks.push_back("");
353 // Write the offsets to the filenames.
354 offset = filename_index_offset + corpus.filename_blocks.size() * sizeof(offset);
355 for (const string &filename : corpus.filename_blocks) {
356 fwrite(&offset, sizeof(offset), 1, outfp);
357 offset += filename.size();
358 bytes_for_filename_index += sizeof(offset);
359 bytes_for_filenames += filename.size();
362 // Write the actual filenames.
363 for (const string &filename : corpus.filename_blocks) {
364 fwrite(filename.data(), filename.size(), 1, outfp);
369 //size_t total_bytes = (bytes_for_trigrams + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
371 dprintf("Block size: %7d files\n", block_size);
372 dprintf("Trigrams: %'7.1f MB\n", bytes_for_trigrams / 1048576.0);
373 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
374 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
375 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
376 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);
380 int main(int argc, char **argv)
382 do_build(argv[1], argv[2], 32);