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
|
/*
* 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 <algorithm>
#include <cmath>
#include "tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h"
#include "Tracing.h"
namespace android {
namespace nn {
inline bool l2normFloat32Impl(const float* inputData, const Shape& inputShape, int32_t axis,
float* outputData, const Shape& outputShape) {
NNTRACE_TRANS("l2normFloat32");
const uint32_t outerSize = getNumberOfElements(inputShape, 0, axis);
const uint32_t axisSize = getSizeOfDimension(inputShape, axis);
const uint32_t innerSize =
getNumberOfElements(inputShape, axis + 1, getNumberOfDimensions(inputShape));
for (uint32_t outer = 0; outer < outerSize; ++outer) {
const float* inputBeg = inputData + outer * axisSize * innerSize;
const float* inputEnd = inputBeg + axisSize * innerSize;
float* outputBeg = outputData + outer * axisSize * innerSize;
for (uint32_t inner = 0; inner < innerSize; ++inner, ++inputBeg, ++inputEnd, ++outputBeg) {
float sum = 0.0f;
for (const float* p = inputBeg; p < inputEnd; p += innerSize) {
float val = *p;
sum += val * val;
}
float l2_norm = std::sqrt(sum);
float* pOut = outputBeg;
for (const float* p = inputBeg; p < inputEnd; p += innerSize, pOut += innerSize) {
*pOut = *p / l2_norm;
}
}
}
return true;
}
bool l2normFloat32(const float* inputData, const Shape& inputShape, int32_t axis, float* outputData,
const Shape& outputShape) {
int32_t ndim = getNumberOfDimensions(inputShape);
axis = getDimensionIndex(inputShape, axis);
// TFLite optimized implementation only supports computation along the last axis
if (axis == ndim - 1) {
NNTRACE_COMP("optimized_ops::L2Normalization::float");
tflite::L2NormalizationParams param = {.input_zero_point = 0};
tflite::optimized_ops::L2Normalization(param, convertShapeToTflshape(inputShape), inputData,
convertShapeToTflshape(outputShape), outputData);
return true;
} else {
return l2normFloat32Impl(inputData, inputShape, axis, outputData, outputShape);
}
}
inline bool localResponseNormFloat32Impl(const float* inputData, const Shape& inputShape,
int32_t radius, float bias, float alpha, float beta,
int32_t axis, float* outputData,
const Shape& outputShape) {
NNTRACE_TRANS("localResponseNormFloat32");
const uint32_t outerSize = getNumberOfElements(inputShape, 0, axis);
const uint32_t axisSize = getSizeOfDimension(inputShape, axis);
const uint32_t innerSize =
getNumberOfElements(inputShape, axis + 1, getNumberOfDimensions(inputShape));
for (uint32_t outer = 0; outer < outerSize; ++outer) {
const float* inputBase = inputData + outer * axisSize * innerSize;
float* outputBase = outputData + outer * axisSize * innerSize;
for (uint32_t inner = 0; inner < innerSize; ++inner, ++inputBase, ++outputBase) {
for (int32_t i = 0; i < axisSize; i++) {
const int32_t dBegin = std::max(0, i - radius);
// Add 1 on dEnd to comply with optimized_ops in TFLite
const int32_t dEnd = std::min(static_cast<int32_t>(axisSize), i + radius + 1);
float sum = 0.0f;
for (int32_t d = dBegin; d < dEnd; d++) {
float val = inputBase[d * innerSize];
sum += val * val;
}
float multiplier = std::pow(bias + alpha * sum, -beta);
outputBase[i * innerSize] = inputBase[i * innerSize] * multiplier;
}
}
}
return true;
}
bool localResponseNormFloat32(const float* inputData, const Shape& inputShape, int32_t radius,
float bias, float alpha, float beta, int32_t axis, float* outputData,
const Shape& outputShape) {
int32_t ndim = getNumberOfDimensions(inputShape);
axis = getDimensionIndex(inputShape, axis);
// TFLite optimized implementation only supports computation along the last axis
if (axis == ndim - 1) {
NNTRACE_COMP("optimized_ops::LocalResponseNormalization::float");
tflite::LocalResponseNormalizationParams param = {
.range = radius, .bias = bias, .alpha = alpha, .beta = beta};
tflite::optimized_ops::LocalResponseNormalization(
param, convertShapeToTflshape(inputShape), inputData,
convertShapeToTflshape(outputShape), outputData);
return true;
} else {
return localResponseNormFloat32Impl(inputData, inputShape, radius, bias, alpha, beta, axis,
outputData, outputShape);
}
}
} // namespace nn
} // namespace android
|
