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
|
/*
* Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef __NNFW_SRCN_COMMON_H__
#define __NNFW_SRCN_COMMON_H__
#include <string.h>
#include <limits>
#include <arm_neon.h>
#include "ncnn/srcn/conv_type.h"
namespace nnfw
{
namespace srcn
{
#define sizeof_RhsScalar 4
#define sizeof_LhsScalar 4
#define sizeof_ResScalar 4
#define MIN(a, b) (a) > (b) ? (b) : (a)
#define MAX(a, b) (a) > (b) ? (a) : (b)
enum shardType_t
{
shardByCol = 0,
shardByRow
};
#ifdef TIZEN
#define L1_CACHE_SIZE (16536 * 2)
#define L2_CACHE_SIZE (524288 * 2)
#define L3_CACHE_SIZE (0) // no L3
#define MAX_K (512)
// single-thread
#define GEN_COL (1440)
// multi-threads
#define MAX_COL (90)
#define MIN_COL (32)
#elif defined ANDROID
#define L1_CACHE_SIZE (16536 * 4)
#define L2_CACHE_SIZE (524288 * 8)
#define L3_CACHE_SIZE (0) //(524288 * 8) //no L3
#define MAX_K (512 * 2)
// single-thread
#define GEN_COL (1440)
// multi-threads
#if __aarch64__
#define MAX_COL (1024)
#else
#define MAX_COL (90)
#endif
#define MIN_COL (32)
#endif
enum
{
USE_COMMON_KENEL = 0,
USE_12BIT_KERNEL,
USE_NONZERO_KERENL
};
template <typename T> static T divup(const T &x, const T &y)
{
return static_cast<T>((x + y - 1) / y);
}
#ifdef NCNN
static inline size_t alignSize(size_t sz, int n) { return (sz + n - 1) / n * n; }
static inline size_t alignBy2(size_t sz) { return (sz + 1) & -2; }
#endif // NCNN
static inline int32_t BitNot(int32_t a) { return ~a; }
static inline int32_t MaskIfNonZero(int32_t a)
{
static int32_t zero = 0;
return a ? BitNot(zero) : zero;
}
static inline int32_t BitAnd(int32_t a, int32_t b) { return a & b; }
static inline int32_t ShiftRight(int32_t a, int offset) { return a >> offset; }
static inline int32_t MaskIfLessThan(int32_t a, int32_t b) { return MaskIfNonZero(a < b); }
static inline int32_t MaskIfGreaterThan(int32_t a, int32_t b) { return MaskIfNonZero(a > b); }
static inline int32_t Add(int32_t a, int32_t b) { return a + b; }
static inline int32_t RoundingDivideByPOT(int32_t x, int exponent)
{
const int32_t mask = (1ll << exponent) - 1;
const int32_t zero = 0;
const int32_t one = 1;
const int32_t remainder = BitAnd(x, mask);
const int32_t threshold = Add(ShiftRight(mask, 1), BitAnd(MaskIfLessThan(x, zero), one));
return Add(ShiftRight(x, exponent), BitAnd(MaskIfGreaterThan(remainder, threshold), one));
}
static inline int32_t SaturatingRoundingDoublingHighMul(int32_t a, int32_t b)
{
bool overflow = a == b && a == std::numeric_limits<int32_t>::min();
int64_t a_64(a);
int64_t b_64(b);
int64_t ab_64 = a_64 * b_64;
int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30));
int32_t ab_x2_high32 = static_cast<int32_t>((ab_64 + nudge) / (1ll << 31));
return overflow ? std::numeric_limits<int32_t>::max() : ab_x2_high32;
}
static inline int32_t MultiplyByQuantizedMultiplier(int32_t x, int32_t quantized_multiplier,
int shift)
{
int left_shift = shift > 0 ? shift : 0;
int right_shift = shift > 0 ? 0 : -shift;
return RoundingDivideByPOT(
SaturatingRoundingDoublingHighMul(x * (1 << left_shift), quantized_multiplier), right_shift);
}
static inline int32x4_t SaturatingRoundingDoublingHighMulV(int32x4_t a, int32x4_t b)
{
return vqrdmulhq_s32(a, b);
}
static inline int32x4_t RoundingDivideByPOTV(int32x4_t x, int exponent)
{
const int32x4_t shift_vec = vdupq_n_s32(-exponent);
const int32x4_t fixup = vshrq_n_s32(vandq_s32(x, shift_vec), 31);
const int32x4_t fixed_up_x = vqaddq_s32(x, fixup);
return vrshlq_s32(fixed_up_x, shift_vec);
}
static inline int32x4_t MultiplyByQuantizedMultiplierV(int32x4_t x, int32_t quantized_multiplier,
int shift)
{
int left_shift = shift > 0 ? shift : 0;
int right_shift = shift > 0 ? 0 : -shift;
return RoundingDivideByPOTV(
SaturatingRoundingDoublingHighMulV(vrshlq_s32(x, vdupq_n_s32(left_shift)),
vdupq_n_s32(quantized_multiplier)),
right_shift);
}
} // namespace srcn
} // namespace nnfw
#endif // __NNFW_SRCN_COMMON_H__
|
