9 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
12 #include "io_uring_engine.h"
13 #include "turbopfor-encode.h"
14 #include "turbopfor.h"
18 using namespace std::chrono;
20 bool use_debug = false;
24 int fd = open("plocate.db", O_RDONLY);
31 complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
33 unique_ptr<Trigram[]> ht(new Trigram[hdr.hashtable_size + hdr.extra_ht_slots + 1]);
34 complete_pread(fd, ht.get(), (hdr.hashtable_size + hdr.extra_ht_slots + 1) * sizeof(Trigram), hdr.hash_table_offset_bytes);
36 size_t posting_list_bytes = 0, own_posting_list_bytes = 0, total_elements = 0, most_bytes_pl = 0;
37 uint32_t longest_pl = 0;
38 vector<pair<string, unsigned>> posting_lists;
39 for (unsigned i = 0; i < hdr.hashtable_size + hdr.extra_ht_slots; ++i) {
40 if (ht[i].num_docids == 0) {
43 size_t len = ht[i + 1].offset - ht[i].offset;
46 complete_pread(fd, &str[0], len, ht[i].offset);
47 posting_lists.emplace_back(move(str), ht[i].num_docids);
48 longest_pl = std::max(ht[i].num_docids, longest_pl);
49 most_bytes_pl = std::max(len, most_bytes_pl);
50 posting_list_bytes += len;
51 total_elements += ht[i].num_docids;
54 fprintf(stderr, "Read %zu posting lists.\n", posting_lists.size());
57 encoded_pl.resize(longest_pl * 2 + 16384); // Lots of margin.
59 size_t num_decode_errors = 0, num_encode_errors = 0;
60 for (auto &[pl, num_docids] : posting_lists) {
61 //fprintf(stderr, "%zu bytes, %u docids\n", pl.size(), num_docids);
62 vector<uint32_t> out1, out2;
63 out1.resize(num_docids + 128);
64 out2.resize(num_docids + 128);
65 unsigned char *pldata = reinterpret_cast<unsigned char *>(&pl[0]);
66 p4nd1dec128v32(pldata, num_docids, &out1[0]);
67 decode_pfor_delta1_128(pldata, num_docids, /*interleaved=*/true, &out2[0]);
68 for (unsigned i = 0; i < num_docids; ++i) {
69 if (out1[i] != out2[i]) {
70 if (++num_decode_errors < 10) {
71 fprintf(stderr, "Decode error:\n");
72 for (unsigned j = 0; j < num_docids; ++j) {
73 fprintf(stderr, "%3u: reference=%u ours=%u (diff=%d)\n", j, out1[j], out2[j], out1[j] - out2[j]);
80 // Test encoding, by encoding with out own implementation
81 // and checking that decoding with the reference gives
82 // the same result. We do not measure performance (we're slow).
84 unsigned char *ptr = reinterpret_cast<unsigned char *>(&encoded_pl[0]);
85 ptr = write_baseval(out1[0], ptr);
86 for (unsigned i = 1; i < num_docids; i += 128) {
87 unsigned num_docids_this_block = std::min(num_docids - i, 128u);
88 for (unsigned j = 0; j < num_docids_this_block; ++j) {
89 deltas[j] = out1[i + j] - out1[i + j - 1] - 1;
91 bool interleaved = (num_docids_this_block == 128);
92 ptr = encode_pfor_single_block<128>(deltas, num_docids_this_block, interleaved, ptr);
94 own_posting_list_bytes += ptr - reinterpret_cast<unsigned char *>(&encoded_pl[0]);
96 pldata = reinterpret_cast<unsigned char *>(&encoded_pl[0]);
97 p4nd1dec128v32(pldata, num_docids, &out2[0]);
98 for (unsigned i = 0; i < num_docids; ++i) {
99 if (out1[i] != out2[i]) {
100 if (++num_encode_errors < 10) {
101 fprintf(stderr, "Encode error:\n");
102 for (unsigned j = 0; j < num_docids; ++j) {
103 fprintf(stderr, "%3u: reference=%u ours=%u (diff=%d)\n", j, out1[j], out2[j], out1[j] - out2[j]);
110 fprintf(stderr, "%zu/%zu posting lists had errors in decoding.\n", num_decode_errors, posting_lists.size());
111 fprintf(stderr, "%zu/%zu posting lists had errors in encoding.\n", num_encode_errors, posting_lists.size());
114 vector<uint32_t> dummy;
115 dummy.resize(longest_pl + 128);
116 steady_clock::time_point start = steady_clock::now();
117 for (auto &[pl, num_docids] : posting_lists) {
118 unsigned char *pldata = reinterpret_cast<unsigned char *>(&pl[0]);
119 p4nd1dec128v32(pldata, num_docids, &dummy[0]);
121 steady_clock::time_point end = steady_clock::now();
122 double reference_sec = duration<double>(end - start).count();
123 fprintf(stderr, "Decoding with reference implementation: %.1f ms\n", 1e3 * reference_sec);
125 start = steady_clock::now();
126 for (auto &[pl, num_docids] : posting_lists) {
127 unsigned char *pldata = reinterpret_cast<unsigned char *>(&pl[0]);
128 decode_pfor_delta1_128(pldata, num_docids, /*interleaved=*/true, &dummy[0]);
130 end = steady_clock::now();
131 double own_sec = duration<double>(end - start).count();
132 fprintf(stderr, "Decoding with own implementation: %.3f ms (%.2f%% speed)\n", 1e3 * own_sec, 100.0 * reference_sec / own_sec);
133 fprintf(stderr, "Size with own implementation: %.1f MB (%.2f%% of reference, %+d bytes)\n", own_posting_list_bytes / 1048576.0, 100.0 * own_posting_list_bytes / posting_list_bytes, int(own_posting_list_bytes) - int(posting_list_bytes));
135 // Three numbers giving rules of thumb for judging our own implementation:
137 // - Compressed speed is easy to compare to disk I/O, to see the relative importance
138 // - Uncompressed speed is easy to compare to intersection speeds and memory bandwidth
139 // (also very roughly comparable to the benchmark numbers in the TurboPFor README)
140 // - ns/element gives an absolute measure for plocate (e.g. if we can decompress at
141 // 1 ns/element, a 10k-element posting list goes by in 0.01 ms, which is way beyond
142 // instantaneous in practice).
143 fprintf(stderr, "%.1f MB/sec (compressed), %.1f MB/sec (uncompressed), %.1f ns/element\n", posting_list_bytes / own_sec / 1048576.0,
144 (total_elements * sizeof(uint32_t)) / own_sec / 1048576.0, 1e9 * own_sec / total_elements);