1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
|
/*
* include/shard/VPTree.h
*
* Copyright (C) 2023 Douglas B. Rumbaugh <drumbaugh@psu.edu>
*
* Distributed under the Modified BSD License.
*
* A shard shim around a VPTree for high-dimensional metric similarity
* search.
*
* FIXME: Does not yet support the tombstone delete policy.
*
*/
#pragma once
#include <vector>
#include <unordered_map>
#include "framework/ShardRequirements.h"
#include "psu-ds/PriorityQueue.h"
using psudb::CACHELINE_SIZE;
using psudb::PriorityQueue;
using psudb::queue_record;
using psudb::byte;
namespace de {
template <NDRecordInterface R, size_t LEAFSZ=100, bool HMAP=false>
class VPTree {
private:
struct vpnode {
size_t start;
size_t stop;
bool leaf;
double radius;
vpnode *inside;
vpnode *outside;
vpnode() : start(0), stop(0), leaf(false), radius(0.0), inside(nullptr), outside(nullptr) {}
~vpnode() {
delete inside;
delete outside;
}
};
public:
VPTree(BufferView<R> buffer)
: m_reccnt(0), m_tombstone_cnt(0), m_root(nullptr), m_node_cnt(0) {
m_alloc_size = psudb::sf_aligned_alloc(CACHELINE_SIZE,
buffer.get_record_count() *
sizeof(Wrapped<R>),
(byte**) &m_data);
m_ptrs = new Wrapped<R>*[buffer.get_record_count()];
size_t offset = 0;
m_reccnt = 0;
// FIXME: will eventually need to figure out tombstones
// this one will likely require the multi-pass
// approach, as otherwise we'll need to sort the
// records repeatedly on each reconstruction.
for (size_t i=0; i<buffer.get_record_count(); i++) {
auto rec = buffer.get(i);
if (rec->is_deleted()) {
continue;
}
rec->header &= 3;
m_data[m_reccnt] = *rec;
m_ptrs[m_reccnt] = &m_data[m_reccnt];
m_reccnt++;
}
if (m_reccnt > 0) {
m_root = build_vptree();
build_map();
}
}
VPTree(std::vector<VPTree*> shards)
: m_reccnt(0), m_tombstone_cnt(0), m_root(nullptr), m_node_cnt(0) {
size_t attemp_reccnt = 0;
for (size_t i=0; i<shards.size(); i++) {
attemp_reccnt += shards[i]->get_record_count();
}
m_alloc_size = psudb::sf_aligned_alloc(CACHELINE_SIZE,
attemp_reccnt * sizeof(Wrapped<R>),
(byte **) &m_data);
m_ptrs = new Wrapped<R>*[attemp_reccnt];
// FIXME: will eventually need to figure out tombstones
// this one will likely require the multi-pass
// approach, as otherwise we'll need to sort the
// records repeatedly on each reconstruction.
for (size_t i=0; i<shards.size(); i++) {
for (size_t j=0; j<shards[i]->get_record_count(); j++) {
if (shards[i]->get_record_at(j)->is_deleted()) {
continue;
}
m_data[m_reccnt] = *shards[i]->get_record_at(j);
m_ptrs[m_reccnt] = &m_data[m_reccnt];
m_reccnt++;
}
}
if (m_reccnt > 0) {
m_root = build_vptree();
build_map();
}
}
~VPTree() {
free(m_data);
delete m_root;
delete[] m_ptrs;
}
Wrapped<R> *point_lookup(const R &rec, bool filter=false) {
if constexpr (HMAP) {
auto idx = m_lookup_map.find(rec);
if (idx == m_lookup_map.end()) {
return nullptr;
}
return m_data + idx->second;
} else {
vpnode *node = m_root;
while (!node->leaf && m_ptrs[node->start]->rec != rec) {
if (rec.calc_distance((m_ptrs[node->start]->rec)) >= node->radius) {
node = node->outside;
} else {
node = node->inside;
}
}
for (size_t i=node->start; i<=node->stop; i++) {
if (m_ptrs[i]->rec == rec) {
return m_ptrs[i];
}
}
return nullptr;
}
}
Wrapped<R>* get_data() const {
return m_data;
}
size_t get_record_count() const {
return m_reccnt;
}
size_t get_tombstone_count() const {
return m_tombstone_cnt;
}
const Wrapped<R>* get_record_at(size_t idx) const {
if (idx >= m_reccnt) return nullptr;
return m_data + idx;
}
size_t get_memory_usage() {
return m_node_cnt * sizeof(vpnode) + m_reccnt * sizeof(R*) + m_alloc_size;
}
size_t get_aux_memory_usage() {
// FIXME: need to return the size of the unordered_map
return 0;
}
void search(const R &point, size_t k, PriorityQueue<Wrapped<R>,
DistCmpMax<Wrapped<R>>> &pq) {
double farthest = std::numeric_limits<double>::max();
internal_search(m_root, point, k, pq, &farthest);
}
private:
Wrapped<R>* m_data;
Wrapped<R>** m_ptrs;
std::unordered_map<R, size_t, RecordHash<R>> m_lookup_map;
size_t m_reccnt;
size_t m_tombstone_cnt;
size_t m_node_cnt;
size_t m_alloc_size;
vpnode *m_root;
vpnode *build_vptree() {
if (m_reccnt == 0) {
return nullptr;
}
size_t lower = 0;
size_t upper = m_reccnt - 1;
auto rng = gsl_rng_alloc(gsl_rng_mt19937);
auto root = build_subtree(lower, upper, rng);
gsl_rng_free(rng);
return root;
}
void build_map() {
// Skip constructing the hashmap if disabled in the
// template parameters.
if constexpr (!HMAP) {
return;
}
for (size_t i=0; i<m_reccnt; i++) {
// FIXME: Will need to account for tombstones here too. Under
// tombstones, it is technically possible for two otherwise identical
// instances of the same record to exist within the same shard, so
// long as one of them is a tombstone. Because the table is currently
// using the unwrapped records for the key, it isn't possible for it
// to handle this case right now.
m_lookup_map.insert({m_data[i].rec, i});
}
}
vpnode *build_subtree(size_t start, size_t stop, gsl_rng *rng) {
// base-case: sometimes happens (probably because of the +1 and -1
// in the first recursive call)
if (start > stop) {
return nullptr;
}
// base-case: create a leaf node
if (stop - start <= LEAFSZ) {
vpnode *node = new vpnode();
node->start = start;
node->stop = stop;
node->leaf = true;
m_node_cnt++;
return node;
}
// select a random element to be the root of the
// subtree
auto i = start + gsl_rng_uniform_int(rng, stop - start + 1);
swap(start, i);
// partition elements based on their distance from the start,
// with those elements with distance falling below the median
// distance going into the left sub-array and those above
// the median in the right. This is easily done using QuickSelect.
auto mid = (start + 1 + stop) / 2;
quickselect(start + 1, stop, mid, m_ptrs[start], rng);
// Create a new node based on this partitioning
vpnode *node = new vpnode();
node->start = start;
// store the radius of the circle used for partitioning the node.
node->radius = m_ptrs[start]->rec.calc_distance(m_ptrs[mid]->rec);
// recursively construct the left and right subtrees
node->inside = build_subtree(start + 1, mid-1, rng);
node->outside = build_subtree(mid, stop, rng);
m_node_cnt++;
return node;
}
void quickselect(size_t start, size_t stop, size_t k, Wrapped<R> *p, gsl_rng *rng) {
if (start == stop) return;
auto pivot = partition(start, stop, p, rng);
if (k < pivot) {
quickselect(start, pivot - 1, k, p, rng);
} else if (k > pivot) {
quickselect(pivot + 1, stop, k, p, rng);
}
}
size_t partition(size_t start, size_t stop, Wrapped<R> *p, gsl_rng *rng) {
auto pivot = start + gsl_rng_uniform_int(rng, stop - start);
double pivot_dist = p->rec.calc_distance(m_ptrs[pivot]->rec);
swap(pivot, stop);
size_t j = start;
for (size_t i=start; i<stop; i++) {
if (p->rec.calc_distance(m_ptrs[i]->rec) < pivot_dist) {
swap(j, i);
j++;
}
}
swap(j, stop);
return j;
}
void swap(size_t idx1, size_t idx2) {
auto tmp = m_ptrs[idx1];
m_ptrs[idx1] = m_ptrs[idx2];
m_ptrs[idx2] = tmp;
}
void internal_search(vpnode *node, const R &point, size_t k, PriorityQueue<Wrapped<R>,
DistCmpMax<Wrapped<R>>> &pq, double *farthest) {
if (node == nullptr) return;
if (node->leaf) {
for (size_t i=node->start; i<=node->stop; i++) {
double d = point.calc_distance(m_ptrs[i]->rec);
if (d < *farthest) {
if (pq.size() == k) {
pq.pop();
}
pq.push(m_ptrs[i]);
if (pq.size() == k) {
*farthest = point.calc_distance(pq.peek().data->rec);
}
}
}
return;
}
double d = point.calc_distance(m_ptrs[node->start]->rec);
if (d < *farthest) {
if (pq.size() == k) {
auto t = pq.peek().data->rec;
pq.pop();
}
pq.push(m_ptrs[node->start]);
if (pq.size() == k) {
*farthest = point.calc_distance(pq.peek().data->rec);
}
}
if (d < node->radius) {
if (d - (*farthest) <= node->radius) {
internal_search(node->inside, point, k, pq, farthest);
}
if (d + (*farthest) >= node->radius) {
internal_search(node->outside, point, k, pq, farthest);
}
} else {
if (d + (*farthest) >= node->radius) {
internal_search(node->outside, point, k, pq, farthest);
}
if (d - (*farthest) <= node->radius) {
internal_search(node->inside, point, k, pq, farthest);
}
}
}
};
}
|