-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathspace_ip.h
282 lines (219 loc) · 8.17 KB
/
space_ip.h
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
#pragma once
#include "hnswlib.h"
namespace hnswlib {
static float
InnerProduct(const void *pVect1, const void *pVect2, const void *qty_ptr) {
size_t qty = *((size_t *) qty_ptr);
float res = 0;
for (unsigned i = 0; i < qty; i++) {
res += ((float *) pVect1)[i] * ((float *) pVect2)[i];
}
return (1.0f - res);
}
#if defined(USE_AVX)
// Favor using AVX if available.
static float
InnerProductSIMD4Ext(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
float PORTABLE_ALIGN32 TmpRes[8];
float *pVect1 = (float *) pVect1v;
float *pVect2 = (float *) pVect2v;
size_t qty = *((size_t *) qty_ptr);
size_t qty16 = qty / 16;
size_t qty4 = qty / 4;
const float *pEnd1 = pVect1 + 16 * qty16;
const float *pEnd2 = pVect1 + 4 * qty4;
__m256 sum256 = _mm256_set1_ps(0);
while (pVect1 < pEnd1) {
//_mm_prefetch((char*)(pVect2 + 16), _MM_HINT_T0);
__m256 v1 = _mm256_loadu_ps(pVect1);
pVect1 += 8;
__m256 v2 = _mm256_loadu_ps(pVect2);
pVect2 += 8;
sum256 = _mm256_add_ps(sum256, _mm256_mul_ps(v1, v2));
v1 = _mm256_loadu_ps(pVect1);
pVect1 += 8;
v2 = _mm256_loadu_ps(pVect2);
pVect2 += 8;
sum256 = _mm256_add_ps(sum256, _mm256_mul_ps(v1, v2));
}
__m128 v1, v2;
__m128 sum_prod = _mm_add_ps(_mm256_extractf128_ps(sum256, 0), _mm256_extractf128_ps(sum256, 1));
while (pVect1 < pEnd2) {
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
}
_mm_store_ps(TmpRes, sum_prod);
float sum = TmpRes[0] + TmpRes[1] + TmpRes[2] + TmpRes[3];;
return 1.0f - sum;
}
#elif defined(USE_SSE)
static float
InnerProductSIMD4Ext(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
float PORTABLE_ALIGN32 TmpRes[8];
float *pVect1 = (float *) pVect1v;
float *pVect2 = (float *) pVect2v;
size_t qty = *((size_t *) qty_ptr);
size_t qty16 = qty / 16;
size_t qty4 = qty / 4;
const float *pEnd1 = pVect1 + 16 * qty16;
const float *pEnd2 = pVect1 + 4 * qty4;
__m128 v1, v2;
__m128 sum_prod = _mm_set1_ps(0);
while (pVect1 < pEnd1) {
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
}
while (pVect1 < pEnd2) {
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
}
_mm_store_ps(TmpRes, sum_prod);
float sum = TmpRes[0] + TmpRes[1] + TmpRes[2] + TmpRes[3];
return 1.0f - sum;
}
#endif
#if defined(USE_AVX)
static float
InnerProductSIMD16Ext(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
float PORTABLE_ALIGN32 TmpRes[8];
float *pVect1 = (float *) pVect1v;
float *pVect2 = (float *) pVect2v;
size_t qty = *((size_t *) qty_ptr);
size_t qty16 = qty / 16;
const float *pEnd1 = pVect1 + 16 * qty16;
__m256 sum256 = _mm256_set1_ps(0);
while (pVect1 < pEnd1) {
//_mm_prefetch((char*)(pVect2 + 16), _MM_HINT_T0);
__m256 v1 = _mm256_loadu_ps(pVect1);
pVect1 += 8;
__m256 v2 = _mm256_loadu_ps(pVect2);
pVect2 += 8;
sum256 = _mm256_add_ps(sum256, _mm256_mul_ps(v1, v2));
v1 = _mm256_loadu_ps(pVect1);
pVect1 += 8;
v2 = _mm256_loadu_ps(pVect2);
pVect2 += 8;
sum256 = _mm256_add_ps(sum256, _mm256_mul_ps(v1, v2));
}
_mm256_store_ps(TmpRes, sum256);
float sum = TmpRes[0] + TmpRes[1] + TmpRes[2] + TmpRes[3] + TmpRes[4] + TmpRes[5] + TmpRes[6] + TmpRes[7];
return 1.0f - sum;
}
#elif defined(USE_SSE)
static float
InnerProductSIMD16Ext(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
float PORTABLE_ALIGN32 TmpRes[8];
float *pVect1 = (float *) pVect1v;
float *pVect2 = (float *) pVect2v;
size_t qty = *((size_t *) qty_ptr);
size_t qty16 = qty / 16;
const float *pEnd1 = pVect1 + 16 * qty16;
__m128 v1, v2;
__m128 sum_prod = _mm_set1_ps(0);
while (pVect1 < pEnd1) {
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
v1 = _mm_loadu_ps(pVect1);
pVect1 += 4;
v2 = _mm_loadu_ps(pVect2);
pVect2 += 4;
sum_prod = _mm_add_ps(sum_prod, _mm_mul_ps(v1, v2));
}
_mm_store_ps(TmpRes, sum_prod);
float sum = TmpRes[0] + TmpRes[1] + TmpRes[2] + TmpRes[3];
return 1.0f - sum;
}
#endif
#if defined(USE_SSE) || defined(USE_AVX)
static float
InnerProductSIMD16ExtResiduals(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
size_t qty = *((size_t *) qty_ptr);
size_t qty16 = qty >> 4 << 4;
float res = InnerProductSIMD16Ext(pVect1v, pVect2v, &qty16);
float *pVect1 = (float *) pVect1v + qty16;
float *pVect2 = (float *) pVect2v + qty16;
size_t qty_left = qty - qty16;
float res_tail = InnerProduct(pVect1, pVect2, &qty_left);
return res + res_tail - 1.0f;
}
static float
InnerProductSIMD4ExtResiduals(const void *pVect1v, const void *pVect2v, const void *qty_ptr) {
size_t qty = *((size_t *) qty_ptr);
size_t qty4 = qty >> 2 << 2;
float res = InnerProductSIMD4Ext(pVect1v, pVect2v, &qty4);
size_t qty_left = qty - qty4;
float *pVect1 = (float *) pVect1v + qty4;
float *pVect2 = (float *) pVect2v + qty4;
float res_tail = InnerProduct(pVect1, pVect2, &qty_left);
return res + res_tail - 1.0f;
}
#endif
class InnerProductSpace : public SpaceInterface<float> {
DISTFUNC<float> fstdistfunc_;
size_t data_size_;
size_t dim_;
public:
InnerProductSpace(size_t dim) {
fstdistfunc_ = InnerProduct;
#if defined(USE_AVX) || defined(USE_SSE)
if (dim % 16 == 0)
fstdistfunc_ = InnerProductSIMD16Ext;
else if (dim % 4 == 0)
fstdistfunc_ = InnerProductSIMD4Ext;
else if (dim > 16)
fstdistfunc_ = InnerProductSIMD16ExtResiduals;
else if (dim > 4)
fstdistfunc_ = InnerProductSIMD4ExtResiduals;
#endif
dim_ = dim;
data_size_ = dim * sizeof(float);
}
size_t get_data_size() {
return data_size_;
}
DISTFUNC<float> get_dist_func() {
return fstdistfunc_;
}
void *get_dist_func_param() {
return &dim_;
}
~InnerProductSpace() {}
};
}