-
Notifications
You must be signed in to change notification settings - Fork 26
/
fast.cpp
226 lines (186 loc) · 6.17 KB
/
fast.cpp
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
/*
Fast Architecture Sensitive Tree layout for binary search trees
(Kim et. al, SIGMOD 2010)
implementation by Viktor Leis, TUM, 2012
notes:
-keys are 4 byte integers
-SSE instructions are used for comparisons
-huge memory pages (2MB)
-page blocks store 4 levels of cacheline blocks
-cacheline blocks store 15 keys and are 64-byte aligned
-the parameter K results in a tree size of (2^(16+K*4))
*/
#include <sys/mman.h>
#include <stdlib.h>
#include <sys/time.h>
#include <stdint.h>
#include <iostream>
#include <emmintrin.h>
#include <cassert>
#include <climits>
#include <string.h>
#include <algorithm>
#include <vector>
#include <random>
#include <utility>
const unsigned K=3;
struct LeafEntry {
int32_t key;
uint64_t value;
};
void* malloc_huge(size_t size) {
void* p=mmap(NULL,size,PROT_READ|PROT_WRITE,MAP_PRIVATE|MAP_ANONYMOUS,-1,0);
#if __linux__
madvise(p,size,MADV_HUGEPAGE);
#endif
return p;
}
inline unsigned pow16(unsigned exponent) {
// 16^exponent
return 1<<(exponent<<2);
}
inline unsigned median(unsigned i,unsigned j) {
return i+(j-1-i)/2;
}
inline void storeSIMDblock(int32_t v[],unsigned k,LeafEntry l[],unsigned i,unsigned j) {
unsigned m=median(i,j);
v[k+0]=l[m].key;
v[k+1]=l[median(i,m)].key;
v[k+2]=l[median(1+m,j)].key;
}
inline unsigned storeCachelineBlock(int32_t v[],unsigned k,LeafEntry l[],unsigned i,unsigned j) {
storeSIMDblock(v,k+3*0,l,i,j);
unsigned m=median(i,j);
storeSIMDblock(v,k+3*1,l,i,median(i,m));
storeSIMDblock(v,k+3*2,l,median(i,m)+1,m);
storeSIMDblock(v,k+3*3,l,m+1,median(m+1,j));
storeSIMDblock(v,k+3*4,l,median(m+1,j)+1,j);
return k+16;
}
unsigned storeFASTpage(int32_t v[],unsigned offset,LeafEntry l[],unsigned i,unsigned j,unsigned levels) {
for (unsigned level=0;level<levels;level++) {
unsigned chunk=(j-i)/pow16(level);
for (unsigned cl=0;cl<pow16(level);cl++)
offset=storeCachelineBlock(v,offset,l,i+cl*chunk,i+(cl+1)*chunk);
}
return offset;
}
int32_t* buildFAST(LeafEntry l[],unsigned len) {
// create array of appropriate size
unsigned n=0;
for (unsigned i=0; i<K+4; i++)
n+=pow16(i);
n=n*64/4;
int32_t* v=(int32_t*)malloc_huge(sizeof(int32_t)*n);
// build FAST
unsigned offset=storeFASTpage(v,0,l,0,len,4);
unsigned chunk=len/(1<<16);
for (unsigned i=0;i<(1<<16);i++)
offset=storeFASTpage(v,offset,l,i*chunk,(i+1)*chunk,K);
assert(offset==n);
return v;
}
inline unsigned maskToIndex(unsigned bitmask) {
static unsigned table[8]={0,9,1,2,9,9,9,3};
return table[bitmask&7];
}
unsigned scale=0;
unsigned search(int32_t v[],int32_t key_q) {
__m128i xmm_key_q=_mm_set1_epi32(key_q);
unsigned page_offset=0;
unsigned level_offset=0;
// first page
for (unsigned cl_level=1; cl_level<=4; cl_level++) {
// first SIMD block
__m128i xmm_tree=_mm_loadu_si128((__m128i*) (v+page_offset+level_offset*16));
__m128i xmm_mask=_mm_cmpgt_epi32(xmm_key_q,xmm_tree);
unsigned index=_mm_movemask_ps(_mm_castsi128_ps(xmm_mask));
unsigned child_index=maskToIndex(index);
// second SIMD block
xmm_tree=_mm_loadu_si128((__m128i*) (v+page_offset+level_offset*16+3+3*child_index));
xmm_mask=_mm_cmpgt_epi32(xmm_key_q,xmm_tree);
index=_mm_movemask_ps(_mm_castsi128_ps(xmm_mask));
unsigned cache_offset=child_index*4 + maskToIndex(index);
level_offset=level_offset*16 + cache_offset;
page_offset+=pow16(cl_level);
}
unsigned pos=level_offset;
unsigned offset=69904+level_offset*scale;
page_offset=0;
level_offset=0;
// second page
for (unsigned cl_level=1; cl_level<=K; cl_level++) {
// first SIMD block
__m128i xmm_tree=_mm_loadu_si128((__m128i*) (v+offset+page_offset+level_offset*16));
__m128i xmm_mask=_mm_cmpgt_epi32(xmm_key_q,xmm_tree);
unsigned index=_mm_movemask_ps(_mm_castsi128_ps(xmm_mask));
unsigned child_index=maskToIndex(index);
// second SIMD block
xmm_tree=_mm_loadu_si128((__m128i*) (v+offset+page_offset+level_offset*16+3+3*child_index));
xmm_mask=_mm_cmpgt_epi32(xmm_key_q,xmm_tree);
index=_mm_movemask_ps(_mm_castsi128_ps(xmm_mask));
unsigned cache_offset=child_index*4 + maskToIndex(index);
level_offset=level_offset*16 + cache_offset;
page_offset+=pow16(cl_level);
}
return (pos<<(K*4))|level_offset;
}
#define BUF_SIZE 2048
std::vector<int> read_data(const char *path) {
std::vector<int> vec;
FILE *fin = fopen(path, "rb");
int buf[BUF_SIZE];
while (true) {
size_t num_read = fread(buf, sizeof(int), BUF_SIZE, fin);
for (int i = 0; i < num_read; i++) {
vec.push_back(buf[i]);
}
if (num_read < BUF_SIZE) break;
}
fclose(fin);
return vec;
}
#define QUERIES_PER_TRIAL (50 * 1000 * 1000)
int main(int argc,char** argv) {
if (argc < 2) {
std::cerr << "Usage: " << argv[0] << " DATA_PATH" << std::endl;
exit(1);
}
std::vector<int> keys = read_data(argv[1]);
printf("num elements: %lu\n", keys.size());
// Clone vec so we don't bring pages from it into cache when selecting random keys
std::vector<int> keys_clone(keys.begin(), keys.end());
keys.push_back(INT_MAX);
unsigned n=(1<<(16+(K*4))); // note: padding to power of 2 for FAST
for (unsigned i = keys.size(); i < n; i++) {
keys.push_back(INT_MAX);
}
printf("num elements (padded): %d\n", n);
LeafEntry* leaves = new LeafEntry[n];
for (unsigned i = 0; i < n; i++) {
leaves[i].key = keys[i];
leaves[i].value = i;
}
int32_t *fast = buildFAST(leaves, n);
for (unsigned i = 0; i < K; i++)
scale += pow16(i);
scale *= 16;
uint32_t seed = std::random_device()();
std::mt19937 rng;
std::uniform_int_distribution<> dist(0, keys_clone.size()-1);
std::vector<int> queries(QUERIES_PER_TRIAL);
rng.seed(seed);
for (int &query : queries) {
query = keys_clone[dist(rng)];
}
long check = 0;
auto start = clock();
for (const int& key : queries) {
check += search(fast, key);
}
auto end = clock();
printf("FAST average time taken: %lf ns\n",
double(end - start) / CLOCKS_PER_SEC / queries.size() * 1e9);
printf("FAST checksum (can be different from other range search baselines) = %ld\n", check);
return 0;
}