8 //#define dprintf(...) fprintf(stderr, __VA_ARGS__);
11 #include "io_uring_engine.h"
12 #include "turbopfor-encode.h"
13 #include "turbopfor.h"
17 using namespace std::chrono;
21 int fd = open("plocate.db", O_RDONLY);
28 complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
30 unique_ptr<Trigram[]> ht(new Trigram[hdr.hashtable_size + hdr.extra_ht_slots + 1]);
31 complete_pread(fd, ht.get(), (hdr.hashtable_size + hdr.extra_ht_slots + 1) * sizeof(Trigram), hdr.hash_table_offset_bytes);
33 size_t posting_list_bytes = 0, own_posting_list_bytes = 0, total_elements = 0, most_bytes_pl = 0;
34 uint32_t longest_pl = 0;
35 vector<pair<string, unsigned>> posting_lists;
36 for (unsigned i = 0; i < hdr.hashtable_size + hdr.extra_ht_slots; ++i) {
37 if (ht[i].num_docids == 0) {
40 size_t len = ht[i + 1].offset - ht[i].offset;
43 complete_pread(fd, &str[0], len, ht[i].offset);
44 posting_lists.emplace_back(move(str), ht[i].num_docids);
45 longest_pl = std::max(ht[i].num_docids, longest_pl);
46 most_bytes_pl = std::max(len, most_bytes_pl);
47 posting_list_bytes += len;
48 total_elements += ht[i].num_docids;
51 fprintf(stderr, "Read %zu posting lists.\n", posting_lists.size());
54 encoded_pl.resize(longest_pl * 2 + 16384); // Lots of margin.
56 size_t num_decode_errors = 0, num_encode_errors = 0;
57 for (auto &[pl, num_docids] : posting_lists) {
58 //fprintf(stderr, "%zu bytes, %u docids\n", pl.size(), num_docids);
59 vector<uint32_t> out1, out2;
60 out1.resize(num_docids + 128);
61 out2.resize(num_docids + 128);
62 unsigned char *pldata = reinterpret_cast<unsigned char *>(&pl[0]);
63 p4nd1dec128v32(pldata, num_docids, &out1[0]);
64 decode_pfor_delta1<128>(pldata, num_docids, /*interleaved=*/true, &out2[0]);
65 for (unsigned i = 0; i < num_docids; ++i) {
66 if (out1[i] != out2[i]) {
67 if (++num_decode_errors < 10) {
68 fprintf(stderr, "Decode error:\n");
69 for (unsigned j = 0; j < num_docids; ++j) {
70 fprintf(stderr, "%3u: reference=%u ours=%u (diff=%d)\n", j, out1[j], out2[j], out1[j] - out2[j]);
77 // Test encoding, by encoding with out own implementation
78 // and checking that decoding with the reference gives
79 // the same result. We do not measure performance (we're slow).
81 unsigned char *ptr = reinterpret_cast<unsigned char *>(&encoded_pl[0]);
82 ptr = write_baseval(out1[0], ptr);
83 for (unsigned i = 1; i < num_docids; i += 128) {
84 unsigned num_docids_this_block = std::min(num_docids - i, 128u);
85 for (unsigned j = 0; j < num_docids_this_block; ++j) {
86 deltas[j] = out1[i + j] - out1[i + j - 1] - 1;
88 bool interleaved = (num_docids_this_block == 128);
89 ptr = encode_pfor_single_block<128>(deltas, num_docids_this_block, interleaved, ptr);
91 own_posting_list_bytes += ptr - reinterpret_cast<unsigned char *>(&encoded_pl[0]);
93 pldata = reinterpret_cast<unsigned char *>(&encoded_pl[0]);
94 p4nd1dec128v32(pldata, num_docids, &out2[0]);
95 for (unsigned i = 0; i < num_docids; ++i) {
96 if (out1[i] != out2[i]) {
97 if (++num_encode_errors < 10) {
98 fprintf(stderr, "Encode error:\n");
99 for (unsigned j = 0; j < num_docids; ++j) {
100 fprintf(stderr, "%3u: reference=%u ours=%u (diff=%d)\n", j, out1[j], out2[j], out1[j] - out2[j]);
107 fprintf(stderr, "%zu/%zu posting lists had errors in decoding.\n", num_decode_errors, posting_lists.size());
108 fprintf(stderr, "%zu/%zu posting lists had errors in encoding.\n", num_encode_errors, posting_lists.size());
111 vector<uint32_t> dummy;
112 dummy.resize(longest_pl + 128);
113 steady_clock::time_point start = steady_clock::now();
114 for (auto &[pl, num_docids] : posting_lists) {
115 unsigned char *pldata = reinterpret_cast<unsigned char *>(&pl[0]);
116 p4nd1dec128v32(pldata, num_docids, &dummy[0]);
118 steady_clock::time_point end = steady_clock::now();
119 double reference_sec = duration<double>(end - start).count();
120 fprintf(stderr, "Decoding with reference implementation: %.1f ms\n", 1e3 * reference_sec);
122 start = steady_clock::now();
123 for (auto &[pl, num_docids] : posting_lists) {
124 unsigned char *pldata = reinterpret_cast<unsigned char *>(&pl[0]);
125 decode_pfor_delta1<128>(pldata, num_docids, /*interleaved=*/true, &dummy[0]);
127 end = steady_clock::now();
128 double own_sec = duration<double>(end - start).count();
129 fprintf(stderr, "Decoding with own implementation: %.3f ms (%.2f%% speed)\n", 1e3 * own_sec, 100.0 * reference_sec / own_sec);
130 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));
132 // Three numbers giving rules of thumb for judging our own implementation:
134 // - Compressed speed is easy to compare to disk I/O, to see the relative importance
135 // - Uncompressed speed is easy to compare to intersection speeds and memory bandwidth
136 // (also very roughly comparable to the benchmark numbers in the TurboPFor README)
137 // - ns/element gives an absolute measure for plocate (e.g. if we can decompress at
138 // 1 ns/element, a 10k-element posting list goes by in 0.01 ms, which is way beyond
139 // instantaneous in practice).
140 fprintf(stderr, "%.1f MB/sec (compressed), %.1f MB/sec (uncompressed), %.1f ns/element\n", posting_list_bytes / own_sec / 1048576.0,
141 (total_elements * sizeof(uint32_t)) / own_sec / 1048576.0, 1e9 * own_sec / total_elements);