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
|
/*
* Copyright (c) Facebook, Inc. and its affiliates.
* All rights reserved.
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
*/
#pragma once
#include <string>
#include <type_traits>
#include "FbgemmBuild.h"
#include "UtilsAvx2.h"
namespace fbgemm {
/**
* @brief Helper struct to type specialize for uint8 and int8 together.
*/
template <typename T>
struct is_8bit {
static constexpr bool value =
std::is_same<T, int8_t>::value || std::is_same<T, uint8_t>::value;
};
/**
* @brief Typed enum to specify matrix operations.
*/
enum class matrix_op_t { NoTranspose, Transpose };
/**
* @brief Typed enum for supported instruction sets.
*/
enum class inst_set_t { anyarch, avx2, avx512 };
/**
* @brief Typed enum for optimized paths for convolutions
*/
enum class optimized_conv_t { depthwise, groupwise, im2col };
/**
* @brief Typed enum for implementation type.
*
* ref is reference and opt is optimized.
*/
enum class impl_type_t { ref, opt };
/**
* @brief A function to compare data in two buffers for closeness/equality.
*/
template <typename T>
FBGEMM_API int compare_buffers(
const T* ref,
const T* test,
int m,
int n,
int ld,
int max_mismatches_to_report,
float atol = 1e-3);
/**
* @brief Debugging helper.
*/
template <typename T>
void printMatrix(
matrix_op_t trans,
const T* inp,
size_t R,
size_t C,
size_t ld,
std::string name);
/**
* @brief Transpose a matrix.
*
* @param M the number of rows of input matrix
* @param N the number of columns of input matrix
*/
void transpose_simd(
int M,
int N,
const float* src,
int ld_src,
float* dst,
int ld_dst);
/**
* @brief Are we running on a AVX512 supported cpu?
*/
FBGEMM_API bool fbgemmHasAvx512Support();
/**
* @brief Are we running on a AVX2 supported cpu?
*/
FBGEMM_API bool fbgemmHasAvx2Support();
/**
* @brief Helper struct to enable autotuning of FBGEMM packing and kernels.
*
* This structure is optional. If not used, the default values for these
* parameters are picked up from PackingTraits-inl.h. Please see this
* file for details on these parameters.
*/
struct FBGEMM_API BlockingFactors {
int MR;
int NR;
int NR_MIN;
int ROW_INTERLEAVE;
int MCB;
int KCB;
int NCB;
};
template <typename accT = std::int32_t>
FBGEMM_API bool isValidBlockingFactor(BlockingFactors* param) {
constexpr bool is_32bit = std::is_same<accT, int32_t>::value;
constexpr bool is_16bit = std::is_same<accT, int16_t>::value;
if (is_32bit) {
if (param->ROW_INTERLEAVE != 4)
return false;
if (fbgemmHasAvx512Support()) {
if (param->NR != 16)
return false;
} else if (fbgemmHasAvx2Support()) {
if (param->NR != 8)
return false;
}
} else if (is_16bit) {
if (param->ROW_INTERLEAVE != 2)
return false;
if (fbgemmHasAvx512Support()) {
if (param->NR != 32)
return false;
} else if (fbgemmHasAvx2Support()) {
if (param->NR != 16)
return false;
}
}
if (param->MCB % param->MR)
return false;
if (param->NCB % param->NR)
return false;
if (fbgemmHasAvx512Support()) {
if (param->MR * (param->NCB / param->NR) > 24)
return false;
} else if (fbgemmHasAvx2Support()) {
if (param->MR * (param->NCB / param->NR) > 16)
return false;
}
return true;
}
} // namespace fbgemm
|
