+#include <assert.h>
#include <stdio.h>
#include <string.h>
#include <algorithm>
#include <endian.h>
#include <sys/types.h>
#include <sys/stat.h>
+#include <zstd.h>
+#define VINT_IN
+#include "vint.h"
#include "vp4.h"
#define P4NENC_BOUND(n) ((n+127)/128+(n+32)*sizeof(uint32_t))
using namespace std;
using namespace std::chrono;
+
+string zstd_compress(const string &src, string *tempbuf);
-static inline uint32_t read_unigram(const string &s, size_t idx)
+static inline uint32_t read_unigram(const string_view s, size_t idx)
{
if (idx < s.size()) {
return (unsigned char)s[idx];
}
}
-static inline uint32_t read_trigram(const string &s, size_t start)
+static inline uint32_t read_trigram(const string_view s, size_t start)
{
return read_unigram(s, start) |
(read_unigram(s, start + 1) << 8) |
uint8_t pad[4];
};
-const char *handle_directory(const char *ptr, vector<string> *files)
+class PostingListBuilder {
+public:
+ void add_docid(uint32_t docid);
+ void finish();
+
+ string encoded;
+ size_t num_docids = 0;
+
+private:
+ void write_header(uint32_t docid);
+ void append_block();
+
+ uint32_t pending_docids[128];
+ unsigned num_pending_docids = 0;
+
+ uint32_t last_block_end;
+};
+
+void PostingListBuilder::add_docid(uint32_t docid)
+{
+ // Deduplicate against the last inserted value, if any.
+ if (num_pending_docids == 0) {
+ if (encoded.empty()) {
+ // Very first docid.
+ write_header(docid);
+ ++num_docids;
+ last_block_end = docid;
+ return;
+ } else if (docid == last_block_end) {
+ return;
+ }
+ } else {
+ if (docid == pending_docids[num_pending_docids - 1]) {
+ return;
+ }
+ }
+
+ pending_docids[num_pending_docids++] = docid;
+ if (num_pending_docids == 128) {
+ append_block();
+ num_pending_docids = 0;
+ last_block_end = docid;
+ }
+ ++num_docids;
+}
+
+void PostingListBuilder::finish()
+{
+ if (num_pending_docids == 0) {
+ return;
+ }
+
+ assert(!encoded.empty()); // write_header() should already have run.
+
+ // No interleaving for partial blocks.
+ unsigned char buf[P4NENC_BOUND(128)];
+ unsigned char *end = p4d1enc32(pending_docids, num_pending_docids, buf, last_block_end);
+ encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
+}
+
+void PostingListBuilder::append_block()
+{
+ unsigned char buf[P4NENC_BOUND(128)];
+ assert(num_pending_docids == 128);
+ unsigned char *end = p4d1enc128v32(pending_docids, 128, buf, last_block_end);
+ encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
+}
+
+void PostingListBuilder::write_header(uint32_t docid)
+{
+ unsigned char buf[P4NENC_BOUND(1)];
+ size_t bytes = p4nd1enc128v32(&docid, 1, buf);
+ encoded.append(reinterpret_cast<char *>(buf), bytes);
+}
+
+class Corpus {
+public:
+ Corpus(size_t block_size) : block_size(block_size) {}
+ void add_file(string filename);
+ void flush_block();
+
+ vector<string> filename_blocks;
+ unordered_map<uint32_t, PostingListBuilder> invindex;
+ size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
+
+private:
+ string current_block;
+ string tempbuf;
+ const size_t block_size;
+};
+
+void Corpus::add_file(string filename)
+{
+ ++num_files;
+ if (!current_block.empty()) {
+ current_block.push_back('\0');
+ }
+ current_block += filename;
+ if (++num_files_in_block == block_size) {
+ flush_block();
+ }
+}
+
+void Corpus::flush_block()
+{
+ if (current_block.empty()) {
+ return;
+ }
+
+ uint32_t docid = num_blocks;
+
+ // Create trigrams.
+ const char *ptr = current_block.c_str();
+ while (ptr < current_block.c_str() + current_block.size()) {
+ string_view s(ptr);
+ if (s.size() >= 3) {
+ for (size_t j = 0; j < s.size() - 2; ++j) {
+ uint32_t trgm = read_trigram(s, j);
+ invindex[trgm].add_docid(docid);
+ }
+ }
+ ptr += s.size() + 1;
+ }
+
+ // Compress and add the filename block.
+ filename_blocks.push_back(zstd_compress(current_block, &tempbuf));
+
+ current_block.clear();
+ num_files_in_block = 0;
+ ++num_blocks;
+}
+
+const char *handle_directory(const char *ptr, Corpus *corpus)
{
ptr += sizeof(db_directory);
uint8_t type = *ptr++;
if (type == DBE_NORMAL) {
string filename = ptr;
- files->push_back(dir_path + "/" + filename);
+ corpus->add_file(dir_path + "/" + filename);
ptr += filename.size() + 1;
} else if (type == DBE_DIRECTORY) {
string dirname = ptr;
- files->push_back(dir_path + "/" + dirname);
+ corpus->add_file(dir_path + "/" + dirname);
ptr += dirname.size() + 1;
} else {
return ptr;
}
}
-void read_mlocate(const char *filename, vector<string> *files)
+void read_mlocate(const char *filename, Corpus *corpus)
{
int fd = open(filename, O_RDONLY);
if (fd == -1) {
const char *ptr = data + offset;
while (ptr < data + len) {
- ptr = handle_directory(ptr, files);
+ ptr = handle_directory(ptr, corpus);
}
munmap((void *)data, len);
close(fd);
}
-void do_build(const char *infile, const char *outfile)
+string zstd_compress(const string &src, string *tempbuf)
+{
+ size_t max_size = ZSTD_compressBound(src.size());
+ if (tempbuf->size() < max_size) {
+ tempbuf->resize(max_size);
+ }
+ size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
+ return string(tempbuf->data(), size);
+}
+
+void do_build(const char *infile, const char *outfile, int block_size)
{
//steady_clock::time_point start = steady_clock::now();
- vector<string> files;
- read_mlocate(infile, &files);
+ Corpus corpus(block_size);
+
+ read_mlocate(infile, &corpus);
if (false) { // To read a plain text file.
FILE *fp = fopen(infile, "r");
while (!feof(fp)) {
}
string s(buf);
if (s.back() == '\n') s.pop_back();
- files.push_back(move(s));
+ corpus.add_file(move(s));
}
fclose(fp);
}
- dprintf("Read %zu files from %s\n", files.size(), infile);
-
- unordered_map<uint32_t, string> pl;
+ corpus.flush_block();
+ dprintf("Read %zu files from %s\n", corpus.num_files, infile);
+
size_t trigrams = 0, longest_posting_list = 0;
- unordered_map<uint32_t, vector<uint32_t>> invindex;
- for (size_t i = 0; i < files.size(); ++i) {
- const string &s = files[i];
- if (s.size() >= 3) {
- for (size_t j = 0; j < s.size() - 2; ++j) {
- uint32_t trgm = read_trigram(s, j);
- invindex[trgm].push_back(i);
- }
- }
- }
- string buf;
size_t bytes_used = 0;
- for (auto &[trigram, docids] : invindex) {
- auto last = unique(docids.begin(), docids.end());
- docids.erase(last, docids.end());
- longest_posting_list = max(longest_posting_list, docids.size());
- trigrams += docids.size();
-
- size_t bytes_needed = P4NENC_BOUND(docids.size());
- if (buf.size() < bytes_needed) buf.resize(bytes_needed);
- size_t bytes = p4nd1enc128v32(&docids[0], docids.size(), reinterpret_cast<unsigned char *>(&buf[0]));
- pl[trigram] = string(buf.data(), bytes);
- bytes_used += bytes;
+ for (auto &[trigram, pl_builder] : corpus.invindex) {
+ pl_builder.finish();
+ longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
+ trigrams += pl_builder.num_docids;
+ bytes_used += pl_builder.encoded.size();
}
dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
- files.size(), invindex.size(), trigrams, double(trigrams) / invindex.size(), longest_posting_list);
-
+ corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list);
dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
+
//steady_clock::time_point end = steady_clock::now();
dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(end - start).count());
vector<uint32_t> all_trigrams;
- for (auto &[trigram, docids] : invindex) {
+ for (auto &[trigram, pl_builder] : corpus.invindex) {
all_trigrams.push_back(trigram);
}
sort(all_trigrams.begin(), all_trigrams.end());
FILE *outfp = fopen(outfile, "wb");
// Write the header.
- uint64_t num_trigrams = invindex.size(); // 64 to get alignment.
+ uint64_t num_trigrams = corpus.invindex.size(); // 64 to get alignment.
fwrite(&num_trigrams, sizeof(num_trigrams), 1, outfp);
// Find out where the posting lists will end.
uint64_t offset = sizeof(uint64_t) * 2 + 16 * all_trigrams.size();
uint64_t filename_index_offset = offset;
- for (uint32_t trgm : all_trigrams) {
- filename_index_offset += pl[trgm].size();
+ for (auto &[trigram, pl_builder] : corpus.invindex) {
+ filename_index_offset += pl_builder.encoded.size();
}
fwrite(&filename_index_offset, sizeof(filename_index_offset), 1, outfp);
size_t bytes_for_trigrams = 0, bytes_for_posting_lists = 0, bytes_for_filename_index = 0, bytes_for_filenames = 0;
-
for (uint32_t trgm : all_trigrams) {
// 16 bytes: trgm (4), number of docids (4), byte offset (8)
fwrite(&trgm, sizeof(trgm), 1, outfp); // One byte wasted, but OK.
- uint32_t num_docids = invindex[trgm].size();
+ const PostingListBuilder &pl_builder = corpus.invindex[trgm];
+ uint32_t num_docids = pl_builder.num_docids;
fwrite(&num_docids, sizeof(num_docids), 1, outfp);
fwrite(&offset, sizeof(offset), 1, outfp);
- offset += pl[trgm].size();
+ offset += pl_builder.encoded.size();
bytes_for_trigrams += 16;
}
// Write the actual posting lists.
for (uint32_t trgm : all_trigrams) {
- fwrite(pl[trgm].data(), pl[trgm].size(), 1, outfp);
- bytes_for_posting_lists += pl[trgm].size();
+ const string &encoded = corpus.invindex[trgm].encoded;
+ fwrite(encoded.data(), encoded.size(), 1, outfp);
+ bytes_for_posting_lists += encoded.size();
}
+ // Stick an empty block at the end as sentinel.
+ corpus.filename_blocks.push_back("");
+
// Write the offsets to the filenames.
- offset = filename_index_offset + files.size() * sizeof(offset);
- for (const string &filename : files) {
+ offset = filename_index_offset + corpus.filename_blocks.size() * sizeof(offset);
+ for (const string &filename : corpus.filename_blocks) {
fwrite(&offset, sizeof(offset), 1, outfp);
- offset += filename.size() + 1;
+ offset += filename.size();
bytes_for_filename_index += sizeof(offset);
- bytes_for_filenames += filename.size() + 1;
+ bytes_for_filenames += filename.size();
}
// Write the actual filenames.
- for (const string &filename : files) {
- fwrite(filename.c_str(), filename.size() + 1, 1, outfp);
+ for (const string &filename : corpus.filename_blocks) {
+ fwrite(filename.data(), filename.size(), 1, outfp);
}
fclose(outfp);
+ //size_t total_bytes = (bytes_for_trigrams + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
+
+ dprintf("Block size: %7d files\n", block_size);
dprintf("Trigrams: %'7.1f MB\n", bytes_for_trigrams / 1048576.0);
dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
+ dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);
dprintf("\n");
}
int main(int argc, char **argv)
{
- do_build(argv[1], argv[2]);
+ do_build(argv[1], argv[2], 32);
exit(EXIT_SUCCESS);
}
projects / plocate / blobdiff