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
|
// Copyright Yahoo. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
#include "generic.h"
#include "private_helpers.hpp"
#include <cblas.h>
namespace vespalib::hwaccelrated {
namespace {
template <typename ACCUM, typename T, size_t UNROLL>
ACCUM
multiplyAdd(const T * a, const T * b, size_t sz)
{
ACCUM partial[UNROLL];
for (size_t i(0); i < UNROLL; i++) {
partial[i] = 0;
}
size_t i(0);
for (; i + UNROLL <= sz; i+= UNROLL) {
for (size_t j(0); j < UNROLL; j++) {
partial[j] += a[i+j] * b[i+j];
}
}
for (;i < sz; i++) {
partial[i%UNROLL] += a[i] * b[i];
}
ACCUM sum(0);
for (size_t j(0); j < UNROLL; j++) {
sum += partial[j];
}
return sum;
}
template <typename T, size_t UNROLL>
double
squaredEuclideanDistanceT(const T * a, const T * b, size_t sz)
{
T partial[UNROLL];
for (size_t i(0); i < UNROLL; i++) {
partial[i] = 0;
}
size_t i(0);
for (; i + UNROLL <= sz; i += UNROLL) {
for (size_t j(0); j < UNROLL; j++) {
T d = a[i+j] - b[i+j];
partial[j] += d * d;
}
}
for (;i < sz; i++) {
T d = a[i] - b[i];
partial[i%UNROLL] += d * d;
}
double sum(0);
for (size_t j(0); j < UNROLL; j++) {
sum += partial[j];
}
return sum;
}
template<size_t UNROLL, typename Operation>
void
bitOperation(Operation operation, void * aOrg, const void * bOrg, size_t bytes) {
const size_t sz(bytes/sizeof(uint64_t));
{
auto a(static_cast<uint64_t *>(aOrg));
auto b(static_cast<const uint64_t *>(bOrg));
size_t i(0);
for (; i + UNROLL <= sz; i += UNROLL) {
for (size_t j(0); j < UNROLL; j++) {
a[i + j] = operation(a[i + j], b[i + j]);
}
}
for (; i < sz; i++) {
a[i] = operation(a[i], b[i]);
}
}
auto a(static_cast<uint8_t *>(aOrg));
auto b(static_cast<const uint8_t *>(bOrg));
for (size_t i(sz*sizeof(uint64_t)); i < bytes; i++) {
a[i] = operation(a[i], b[i]);
}
}
}
float
GenericAccelrator::dotProduct(const float * a, const float * b, size_t sz) const
{
return cblas_sdot(sz, a, 1, b, 1);
}
double
GenericAccelrator::dotProduct(const double * a, const double * b, size_t sz) const
{
return cblas_ddot(sz, a, 1, b, 1);
}
int64_t
GenericAccelrator::dotProduct(const int8_t * a, const int8_t * b, size_t sz) const
{
return multiplyAdd<int64_t, int8_t, 8>(a, b, sz);
}
int64_t
GenericAccelrator::dotProduct(const int16_t * a, const int16_t * b, size_t sz) const
{
return multiplyAdd<int64_t, int16_t, 8>(a, b, sz);
}
int64_t
GenericAccelrator::dotProduct(const int32_t * a, const int32_t * b, size_t sz) const
{
return multiplyAdd<int64_t, int32_t, 8>(a, b, sz);
}
long long
GenericAccelrator::dotProduct(const int64_t * a, const int64_t * b, size_t sz) const
{
return multiplyAdd<long long, int64_t, 8>(a, b, sz);
}
void
GenericAccelrator::orBit(void * aOrg, const void * bOrg, size_t bytes) const
{
helper::bitOperation<8>([](uint64_t a, uint64_t b) { return a | b; }, aOrg, bOrg, bytes);
}
void
GenericAccelrator::andBit(void * aOrg, const void * bOrg, size_t bytes) const
{
helper::bitOperation<8>([](uint64_t a, uint64_t b) { return a & b; }, aOrg, bOrg, bytes);
}
void
GenericAccelrator::andNotBit(void * aOrg, const void * bOrg, size_t bytes) const
{
helper::bitOperation<8>([](uint64_t a, uint64_t b) { return a & ~b; }, aOrg, bOrg, bytes);
}
void
GenericAccelrator::notBit(void * aOrg, size_t bytes) const
{
auto a(static_cast<uint64_t *>(aOrg));
const size_t sz(bytes/sizeof(uint64_t));
for (size_t i(0); i < sz; i++) {
a[i] = ~a[i];
}
auto ac(static_cast<uint8_t *>(aOrg));
for (size_t i(sz*sizeof(uint64_t)); i < bytes; i++) {
ac[i] = ~ac[i];
}
}
size_t
GenericAccelrator::populationCount(const uint64_t *a, size_t sz) const {
return helper::populationCount(a, sz);
}
double
GenericAccelrator::squaredEuclideanDistance(const int8_t * a, const int8_t * b, size_t sz) const {
return helper::squaredEuclideanDistance(a, b, sz);
}
double
GenericAccelrator::squaredEuclideanDistance(const float * a, const float * b, size_t sz) const {
return squaredEuclideanDistanceT<float, 2>(a, b, sz);
}
double
GenericAccelrator::squaredEuclideanDistance(const double * a, const double * b, size_t sz) const {
return squaredEuclideanDistanceT<double, 2>(a, b, sz);
}
void
GenericAccelrator::and64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const {
helper::andChunks<16, 4>(offset, src, dest);
}
void
GenericAccelrator::or64(size_t offset, const std::vector<std::pair<const void *, bool>> &src, void *dest) const {
helper::orChunks<16,4>(offset, src, dest);
}
}
|
