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
|
#include "fbgemm_pack_matrix_cache.h"
#include <map>
#include <memory>
#include <mutex>
using namespace std;
namespace caffe2 {
template <typename ACC_T>
shared_ptr<fbgemm::PackBMatrix<int8_t, ACC_T>> GetOrCreateFbgemmPackBMatrix(
fbgemm::matrix_op_t trans,
int32_t m,
int32_t n,
const void* orig_data,
const int8_t* quantized_data,
int32_t ld) {
static std::map<
std::tuple<int, int, const void*>,
weak_ptr<fbgemm::PackBMatrix<int8_t, ACC_T>>>
cache;
static mutex cache_mutex;
// Create a new packed matrix and compare with cached one if there's any.
// Note that a cache miss is as expensive as a cache hit here, the purpose of
// this cache is only to deduplicate the quantized tensors for improved
// memory bandwidth if different nets share copies of the same operator.
// TODO: make this cheaper by computing hash of fdata.
auto new_packed = make_shared<fbgemm::PackBMatrix<int8_t, ACC_T>>(
trans,
m,
n,
quantized_data,
ld,
nullptr, // pmat
1); // groups
std::tuple<int, int, const void*> key(m, n, orig_data);
std::shared_ptr<fbgemm::PackBMatrix<int8_t, ACC_T>> cache_entry;
{
lock_guard<mutex> lock(cache_mutex);
auto itr = cache.find(key);
if (itr != cache.end()) {
cache_entry = itr->second.lock();
}
} // release lock here during expensive equals()
if (!cache_entry || !cache_entry->metaEquals(*new_packed) ||
!cache_entry->equals(*new_packed)) {
// cache miss
lock_guard<mutex> lock(cache_mutex);
cache[key] = new_packed;
return new_packed;
} else {
return cache_entry;
}
}
template shared_ptr<fbgemm::PackBMatrix<int8_t, int16_t>>
GetOrCreateFbgemmPackBMatrix<int16_t>(
fbgemm::matrix_op_t trans,
int32_t m,
int32_t n,
const void* orig_data,
const int8_t* quantized_data,
int32_t ld);
template shared_ptr<fbgemm::PackBMatrix<int8_t, int32_t>>
GetOrCreateFbgemmPackBMatrix<int32_t>(
fbgemm::matrix_op_t trans,
int32_t m,
int32_t n,
const void* orig_data,
const int8_t* quantized_data,
int32_t ld);
} // namespace caffe2
|
