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
|
/*
* Copyright (C) 2017 The Android Open Source Project
*
* 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.
*/
#include "Operations.h"
#include "CpuOperationUtils.h"
#include "tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h"
namespace android {
namespace nn {
// If possible we will use this static buffer for the tensor.
static constexpr size_t kStaticBufferSize = 1605632;
static char static_scratch_buffer[kStaticBufferSize];
#define ANDROID_NN_CONV_PARAMETERS(Type) \
uint32_t height = getSizeOfDimension(inputShape, 1); \
uint32_t width = getSizeOfDimension(inputShape, 2); \
uint32_t filterHeight = getSizeOfDimension(filterShape, 1); \
uint32_t filterWidth = getSizeOfDimension(filterShape, 2); \
uint32_t outHeight = getSizeOfDimension(outputShape, 1); \
uint32_t outWidth = getSizeOfDimension(outputShape, 2); \
uint32_t inDepth = getSizeOfDimension(inputShape, 3); \
\
uint32_t paddingHeight = (uint32_t)padding_top; \
uint32_t paddingWidth = (uint32_t)padding_left; \
\
tflite::Dims<4> im2colDim; \
im2colDim.sizes[3] = (int)getSizeOfDimension(outputShape, 0); \
im2colDim.sizes[2] = (int)getSizeOfDimension(outputShape, 1); \
im2colDim.sizes[1] = (int)getSizeOfDimension(outputShape, 2); \
im2colDim.sizes[0] = (int)inDepth * filterHeight * filterWidth; \
\
im2colDim.strides[0] = 1; \
for (int i=1; i<4; i++) { \
im2colDim.strides[i] = im2colDim.strides[i-1] * im2colDim.sizes[i-1]; \
} \
\
Type* im2colData = nullptr; \
uint64_t im2colByteSize = sizeof(Type); \
std::unique_ptr<Type[]> im2colGuard; \
for (int i=0; i<4; i++) { \
im2colByteSize *= im2colDim.sizes[i]; \
} \
\
if (sizeof(size_t) == 4 && (im2colByteSize / sizeof(Type)) > 0xFFFFFFFF) { \
LOG(ERROR) << "Conv size is too large, not enough memory"; \
return false; \
} \
if (im2colByteSize <= kStaticBufferSize) { \
im2colData = reinterpret_cast<Type *>(static_scratch_buffer); \
} else { \
im2colData = new (std::nothrow) Type[im2colByteSize / sizeof(Type)]; \
if (im2colData == nullptr) { \
LOG(ERROR) << "Conv size is too large, not enough memory"; \
return false; \
} \
im2colGuard.reset(im2colData); \
}
bool convFloat32(const float* inputData, const Shape& inputShape,
const float* filterData, const Shape& filterShape,
const float* biasData, const Shape& biasShape,
int32_t padding_left, int32_t padding_right,
int32_t padding_top, int32_t padding_bottom,
int32_t stride_width, int32_t stride_height,
int32_t activation,
float* outputData, const Shape& outputShape) {
ANDROID_NN_CONV_PARAMETERS(float)
float output_activation_min, output_activation_max;
CalculateActivationRangeFloat(activation, &output_activation_min,
&output_activation_max);
tflite::optimized_ops::Conv(
inputData, convertShapeToDims(inputShape),
filterData, convertShapeToDims(filterShape),
biasData, convertShapeToDims(biasShape),
stride_width, stride_height, paddingWidth, paddingHeight,
output_activation_min, output_activation_max,
outputData, convertShapeToDims(outputShape),
im2colData, im2colDim);
return true;
}
bool convQuant8(const uint8_t* inputData, const Shape& inputShape,
const uint8_t* filterData, const Shape& filterShape,
const int32_t* biasData, const Shape& biasShape,
int32_t padding_left, int32_t padding_right,
int32_t padding_top, int32_t padding_bottom,
int32_t stride_width, int32_t stride_height,
int32_t activation,
uint8_t* outputData, const Shape& outputShape) {
ANDROID_NN_CONV_PARAMETERS(uint8_t)
int32_t inputOffset = -inputShape.offset;
int32_t filterOffset = -filterShape.offset;
int32_t outputOffset = outputShape.offset;
float real_multiplier = 0.0;
int32_t output_multiplier = 0;
int32_t output_shift = 0;
int32_t output_activation_min = 0;
int32_t output_activation_max = 0;
if (!GetQuantizedConvolutionMultipler(inputShape, filterShape, biasShape,
outputShape, &real_multiplier) ||
!QuantizeMultiplierSmallerThanOne(real_multiplier, &output_multiplier,
&output_shift)){
return false;
}
CalculateActivationRangeUint8(activation, outputShape,
&output_activation_min,
&output_activation_max);
static gemmlowp::GemmContext gemm_context;
// Alow gemmlowp automatcally decide how many threads to use.
gemm_context.set_max_num_threads(0);
tflite::optimized_ops::Conv(
inputData, convertShapeToDims(inputShape), inputOffset,
filterData, convertShapeToDims(filterShape), filterOffset,
biasData, convertShapeToDims(biasShape),
stride_width, stride_height, paddingWidth, paddingHeight,
outputOffset, output_multiplier, output_shift,
output_activation_min, output_activation_max,
outputData, convertShapeToDims(outputShape),
im2colData, im2colDim, &gemm_context);
return true;
}
#undef ANDROID_NN_CONV_PARAMETERS
} // namespace nn
} // namespace android
|
