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 uint32_t pending_docids[128];
85 unsigned num_pending_docids = 0;
87 uint32_t last_block_end;
90 void PostingListBuilder::add_docid(uint32_t docid)
92 // Deduplicate against the last inserted value, if any.
93 if (num_pending_docids == 0) {
94 if (encoded.empty()) {
98 last_block_end = docid;
100 } else if (docid == last_block_end) {
104 if (docid == pending_docids[num_pending_docids - 1]) {
109 pending_docids[num_pending_docids++] = docid;
110 if (num_pending_docids == 128) {
112 num_pending_docids = 0;
113 last_block_end = docid;
118 void PostingListBuilder::finish()
120 if (num_pending_docids == 0) {
124 assert(!encoded.empty()); // write_header() should already have run.
126 // No interleaving for partial blocks.
127 unsigned char buf[P4NENC_BOUND(128)];
128 unsigned char *end = p4d1enc32(pending_docids, num_pending_docids, buf, last_block_end);
129 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
132 void PostingListBuilder::append_block()
134 unsigned char buf[P4NENC_BOUND(128)];
135 assert(num_pending_docids == 128);
136 unsigned char *end = p4d1enc128v32(pending_docids, 128, buf, last_block_end);
137 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
140 void PostingListBuilder::write_header(uint32_t docid)
142 unsigned char buf[P4NENC_BOUND(1)];
143 size_t bytes = p4nd1enc128v32(&docid, 1, buf);
144 encoded.append(reinterpret_cast<char *>(buf), bytes);
149 Corpus(size_t block_size) : block_size(block_size) {}
150 void add_file(string filename);
153 vector<string> filename_blocks;
154 unordered_map<uint32_t, PostingListBuilder> invindex;
155 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
158 string current_block;
160 const size_t block_size;
163 void Corpus::add_file(string filename)
166 if (!current_block.empty()) {
167 current_block.push_back('\0');
169 current_block += filename;
170 if (++num_files_in_block == block_size) {
175 void Corpus::flush_block()
177 if (current_block.empty()) {
181 uint32_t docid = num_blocks;
184 const char *ptr = current_block.c_str();
185 while (ptr < current_block.c_str() + current_block.size()) {
188 for (size_t j = 0; j < s.size() - 2; ++j) {
189 uint32_t trgm = read_trigram(s, j);
190 invindex[trgm].add_docid(docid);
196 // Compress and add the filename block.
197 filename_blocks.push_back(zstd_compress(current_block, &tempbuf));
199 current_block.clear();
200 num_files_in_block = 0;
204 const char *handle_directory(const char *ptr, Corpus *corpus)
206 ptr += sizeof(db_directory);
208 string dir_path = ptr;
209 ptr += dir_path.size() + 1;
210 if (dir_path == "/") {
215 uint8_t type = *ptr++;
216 if (type == DBE_NORMAL) {
217 string filename = ptr;
218 corpus->add_file(dir_path + "/" + filename);
219 ptr += filename.size() + 1;
220 } else if (type == DBE_DIRECTORY) {
221 string dirname = ptr;
222 corpus->add_file(dir_path + "/" + dirname);
223 ptr += dirname.size() + 1;
230 void read_mlocate(const char *filename, Corpus *corpus)
232 int fd = open(filename, O_RDONLY);
237 off_t len = lseek(fd, 0, SEEK_END);
242 const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
243 if (data == MAP_FAILED) {
248 const db_header *hdr = (const db_header *)data;
250 // TODO: Care about the base path.
251 string path = data + sizeof(db_header);
252 uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
254 const char *ptr = data + offset;
255 while (ptr < data + len) {
256 ptr = handle_directory(ptr, corpus);
259 munmap((void *)data, len);
263 string zstd_compress(const string &src, string *tempbuf)
265 size_t max_size = ZSTD_compressBound(src.size());
266 if (tempbuf->size() < max_size) {
267 tempbuf->resize(max_size);
269 size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
270 return string(tempbuf->data(), size);
273 void do_build(const char *infile, const char *outfile, int block_size)
275 //steady_clock::time_point start = steady_clock::now();
277 Corpus corpus(block_size);
279 read_mlocate(infile, &corpus);
280 if (false) { // To read a plain text file.
281 FILE *fp = fopen(infile, "r");
284 if (fgets(buf, 1024, fp) == nullptr || feof(fp)) {
288 if (s.back() == '\n') s.pop_back();
289 corpus.add_file(move(s));
293 corpus.flush_block();
294 dprintf("Read %zu files from %s\n", corpus.num_files, infile);
296 size_t trigrams = 0, longest_posting_list = 0;
297 size_t bytes_used = 0;
298 for (auto &[trigram, pl_builder] : corpus.invindex) {
300 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
301 trigrams += pl_builder.num_docids;
302 bytes_used += pl_builder.encoded.size();
304 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
305 corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
306 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
308 //steady_clock::time_point end = steady_clock::now();
309 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(end - start).count());
311 vector<uint32_t> all_trigrams;
312 for (auto &[trigram, pl_builder] : corpus.invindex) {
313 all_trigrams.push_back(trigram);
315 sort(all_trigrams.begin(), all_trigrams.end());
318 FILE *outfp = fopen(outfile, "wb");
321 uint64_t num_trigrams = corpus.invindex.size(); // 64 to get alignment.
322 fwrite(&num_trigrams, sizeof(num_trigrams), 1, outfp);
324 // Find out where the posting lists will end.
325 uint64_t offset = sizeof(uint64_t) * 2 + 16 * all_trigrams.size();
326 uint64_t filename_index_offset = offset;
327 for (auto &[trigram, pl_builder] : corpus.invindex) {
328 filename_index_offset += pl_builder.encoded.size();
330 fwrite(&filename_index_offset, sizeof(filename_index_offset), 1, outfp);
332 size_t bytes_for_trigrams = 0, bytes_for_posting_lists = 0, bytes_for_filename_index = 0, bytes_for_filenames = 0;
333 for (uint32_t trgm : all_trigrams) {
334 // 16 bytes: trgm (4), number of docids (4), byte offset (8)
335 fwrite(&trgm, sizeof(trgm), 1, outfp); // One byte wasted, but OK.
336 const PostingListBuilder &pl_builder = corpus.invindex[trgm];
337 uint32_t num_docids = pl_builder.num_docids;
338 fwrite(&num_docids, sizeof(num_docids), 1, outfp);
339 fwrite(&offset, sizeof(offset), 1, outfp);
340 offset += pl_builder.encoded.size();
341 bytes_for_trigrams += 16;
344 // Write the actual posting lists.
345 for (uint32_t trgm : all_trigrams) {
346 const string &encoded = corpus.invindex[trgm].encoded;
347 fwrite(encoded.data(), encoded.size(), 1, outfp);
348 bytes_for_posting_lists += encoded.size();
351 // Stick an empty block at the end as sentinel.
352 corpus.filename_blocks.push_back("");
354 // Write the offsets to the filenames.
355 offset = filename_index_offset + corpus.filename_blocks.size() * sizeof(offset);
356 for (const string &filename : corpus.filename_blocks) {
357 fwrite(&offset, sizeof(offset), 1, outfp);
358 offset += filename.size();
359 bytes_for_filename_index += sizeof(offset);
360 bytes_for_filenames += filename.size();
363 // Write the actual filenames.
364 for (const string &filename : corpus.filename_blocks) {
365 fwrite(filename.data(), filename.size(), 1, outfp);
370 //size_t total_bytes = (bytes_for_trigrams + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
372 dprintf("Block size: %7d files\n", block_size);
373 dprintf("Trigrams: %'7.1f MB\n", bytes_for_trigrams / 1048576.0);
374 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
375 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
376 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
377 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);
381 int main(int argc, char **argv)
383 do_build(argv[1], argv[2], 32);