1 /*
2 * Copyright (C) 2017 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 // Contains the implementation of the operations.
18
19 #define LOG_TAG "Operations"
20
21 #include <vector>
22
23 #include "CpuOperationUtils.h"
24 #include "Operations.h"
25
26 #include <tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h>
27 #include <tensorflow/lite/kernels/internal/reference/reference_ops.h>
28
29 #include "Tracing.h"
30
31 namespace android {
32 namespace nn {
33
copyData(const void * inputData,const Shape & inputShape,void * outputData,const Shape & outputShape)34 bool copyData(const void* inputData, const Shape& inputShape, void* outputData,
35 const Shape& outputShape) {
36 NNTRACE_COMP("copyData");
37 size_t count = nonExtensionOperandSizeOfData(inputShape.type, inputShape.dimensions);
38 memcpy(outputData, inputData, count);
39 return true;
40 }
41
42 template <typename T>
depthToSpaceGeneric(const T * inputData,const Shape & inputShape,int32_t blockSize,T * outputData,const Shape & outputShape)43 bool depthToSpaceGeneric(const T* inputData, const Shape& inputShape, int32_t blockSize,
44 T* outputData, const Shape& outputShape) {
45 NNTRACE_COMP("optimized_ops::DepthToSpace");
46 tflite::optimized_ops::DepthToSpace(inputData, convertShapeToDims(inputShape), blockSize,
47 outputData, convertShapeToDims(outputShape));
48 return true;
49 }
50 template bool depthToSpaceGeneric<float>(const float* inputData, const Shape& inputShape,
51 int32_t blockSize, float* outputData,
52 const Shape& outputShape);
53 template bool depthToSpaceGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
54 int32_t blockSize, _Float16* outputData,
55 const Shape& outputShape);
56 template bool depthToSpaceGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
57 int32_t blockSize, uint8_t* outputData,
58 const Shape& outputShape);
59 template bool depthToSpaceGeneric<int8_t>(const int8_t* inputData, const Shape& inputShape,
60 int32_t blockSize, int8_t* outputData,
61 const Shape& outputShape);
62
63 template <typename T>
spaceToDepthGeneric(const T * inputData,const Shape & inputShape,int32_t blockSize,T * outputData,const Shape & outputShape)64 bool spaceToDepthGeneric(const T* inputData, const Shape& inputShape, int32_t blockSize,
65 T* outputData, const Shape& outputShape) {
66 NNTRACE_COMP("optimized_ops::SpaceToDepth");
67 tflite::optimized_ops::SpaceToDepth(inputData, convertShapeToDims(inputShape), blockSize,
68 outputData, convertShapeToDims(outputShape));
69 return true;
70 }
71 template bool spaceToDepthGeneric<float>(const float* inputData, const Shape& inputShape,
72 int32_t blockSize, float* outputData,
73 const Shape& outputShape);
74 template bool spaceToDepthGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
75 int32_t blockSize, _Float16* outputData,
76 const Shape& outputShape);
77 template bool spaceToDepthGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
78 int32_t blockSize, uint8_t* outputData,
79 const Shape& outputShape);
80 template bool spaceToDepthGeneric<int8_t>(const int8_t* inputData, const Shape& inputShape,
81 int32_t blockSize, int8_t* outputData,
82 const Shape& outputShape);
83
84 template <typename T>
padGeneric(const T * inputData,const Shape & inputShape,const int32_t * paddings,T padValue,T * outputData,const Shape & outputShape)85 bool padGeneric(const T* inputData, const Shape& inputShape, const int32_t* paddings, T padValue,
86 T* outputData, const Shape& outputShape) {
87 NNTRACE_TRANS("padGeneric");
88
89 // Based on
90 // http://google3/third_party/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h?l=6194&rcl=213557260
91
92 // TFLite runtime calls are currently fixed at 4 dimensions. Copy inputs so
93 // we can pad them to 4 dims (yes, we are "padding the padding").
94 int32_t numInputDims = static_cast<int32_t>(getNumberOfDimensions(inputShape));
95 NN_OPS_CHECK(numInputDims <= 4);
96 std::vector<int> leftPaddings(4 - numInputDims, 0);
97 std::vector<int> rightPaddings(4 - numInputDims, 0);
98 for (int32_t i = 0; i < numInputDims; ++i) {
99 leftPaddings.push_back(paddings[i * 2]);
100 rightPaddings.push_back(paddings[i * 2 + 1]);
101 }
102 const int leftBPadding = leftPaddings[0];
103 const int leftHPadding = leftPaddings[1];
104 const int leftWPadding = leftPaddings[2];
105 const int leftDPadding = leftPaddings[3];
106 const int rightBPadding = rightPaddings[0];
107 const int rightHPadding = rightPaddings[1];
108 const int rightWPadding = rightPaddings[2];
109 const int rightDPadding = rightPaddings[3];
110
111 const auto extInputShape =
112 tflite::RuntimeShape::ExtendedShape(4, convertShapeToTflshape(inputShape));
113 const auto extOutputShape =
114 tflite::RuntimeShape::ExtendedShape(4, convertShapeToTflshape(outputShape));
115
116 const int outputBatch = extOutputShape.Dims(0);
117 const int outputHeight = extOutputShape.Dims(1);
118 const int outputWidth = extOutputShape.Dims(2);
119 const int outputDepth = extOutputShape.Dims(3);
120
121 const int inputDepth = extInputShape.Dims(3);
122
123 NNTRACE_COMP_SWITCH("padGeneric");
124
125 if (leftBPadding != 0) {
126 tflite::optimized_ops::TypedMemset<T>(
127 outputData, padValue, leftBPadding * outputHeight * outputWidth * outputDepth);
128 }
129 for (int outB = leftBPadding; outB < outputBatch - rightBPadding; ++outB) {
130 if (leftHPadding != 0) {
131 tflite::optimized_ops::TypedMemset<T>(
132 outputData + tflite::Offset(extOutputShape, outB, 0, 0, 0), padValue,
133 leftHPadding * outputWidth * outputDepth);
134 }
135 for (int outH = leftHPadding; outH < outputHeight - rightHPadding; ++outH) {
136 if (leftWPadding != 0) {
137 tflite::optimized_ops::TypedMemset<T>(
138 outputData + tflite::Offset(extOutputShape, outB, outH, 0, 0), padValue,
139 leftWPadding * outputDepth);
140 }
141 for (int outW = leftWPadding; outW < outputWidth - rightWPadding; ++outW) {
142 if (leftDPadding != 0) {
143 tflite::optimized_ops::TypedMemset<T>(
144 outputData + tflite::Offset(extOutputShape, outB, outH, outW, 0),
145 padValue, leftDPadding);
146 }
147
148 T* out =
149 outputData + tflite::Offset(extOutputShape, outB, outH, outW, leftDPadding);
150 const T* in =
151 inputData + tflite::Offset(extInputShape, outB - leftBPadding,
152 outH - leftHPadding, outW - leftWPadding, 0);
153 memcpy(out, in, inputDepth * sizeof(T));
154
155 if (rightDPadding != 0) {
156 tflite::optimized_ops::TypedMemset<T>(
157 outputData + tflite::Offset(extOutputShape, outB, outH, outW,
158 outputDepth - rightDPadding),
159 padValue, rightDPadding);
160 }
161 }
162 if (rightWPadding != 0) {
163 tflite::optimized_ops::TypedMemset<T>(
164 outputData + tflite::Offset(extOutputShape, outB, outH,
165 outputWidth - rightWPadding, 0),
166 padValue, rightWPadding * outputDepth);
167 }
168 }
169 if (rightHPadding != 0) {
170 tflite::optimized_ops::TypedMemset<T>(
171 outputData + tflite::Offset(extOutputShape, outB, outputHeight - rightHPadding,
172 0, 0),
173 padValue, rightHPadding * outputWidth * outputDepth);
174 }
175 }
176 if (rightBPadding != 0) {
177 tflite::optimized_ops::TypedMemset<T>(
178 outputData + tflite::Offset(extOutputShape, outputBatch - rightBPadding, 0, 0, 0),
179 padValue, rightBPadding * outputHeight * outputWidth * outputDepth);
180 }
181
182 return true;
183 }
184 template bool padGeneric<float>(const float* inputData, const Shape& inputShape,
185 const int32_t* paddings, float padValue, float* outputData,
186 const Shape& outputShape);
187 template bool padGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
188 const int32_t* paddings, _Float16 padValue, _Float16* outputData,
189 const Shape& outputShape);
190 template bool padGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
191 const int32_t* paddings, uint8_t padValue, uint8_t* outputData,
192 const Shape& outputShape);
193 template bool padGeneric<int8_t>(const int8_t* inputData, const Shape& inputShape,
194 const int32_t* paddings, int8_t padValue, int8_t* outputData,
195 const Shape& outputShape);
196
197 template <typename T>
batchToSpaceGeneric(const T * inputData,const Shape & inputShape,const int32_t * blockSize,T * outputData,const Shape & outputShape)198 bool batchToSpaceGeneric(const T* inputData, const Shape& inputShape, const int32_t* blockSize,
199 T* outputData, const Shape& outputShape) {
200 // Needed by low level implementation, but not really used.
201 tflite::Dims<4> blockSizeDim, cropsDim;
202 const int32 crops[4] = {0, 0, 0, 0};
203 NNTRACE_COMP("optimized_ops::BatchToSpaceND");
204 tflite::optimized_ops::BatchToSpaceND(inputData, convertShapeToDims(inputShape), blockSize,
205 blockSizeDim, crops, cropsDim, outputData,
206 convertShapeToDims(outputShape));
207 return true;
208 }
209 template bool batchToSpaceGeneric<float>(const float* inputData, const Shape& inputShape,
210 const int32_t* blockSize, float* outputData,
211 const Shape& outputShape);
212 template bool batchToSpaceGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
213 const int32_t* blockSize, _Float16* outputData,
214 const Shape& outputShape);
215 template bool batchToSpaceGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
216 const int32_t* blockSize, uint8_t* outputData,
217 const Shape& outputShape);
218 template bool batchToSpaceGeneric<int8_t>(const int8_t* inputData, const Shape& inputShape,
219 const int32_t* blockSize, int8_t* outputData,
220 const Shape& outputShape);
221
222 template <typename T>
spaceToBatchGeneric(const T * inputData,const Shape & inputShape,const int32_t * blockSize,const int32_t * padding,const Shape & paddingShape,T * outputData,const Shape & outputShape)223 bool spaceToBatchGeneric(const T* inputData, const Shape& inputShape, const int32_t* blockSize,
224 const int32_t* padding, const Shape& paddingShape, T* outputData,
225 const Shape& outputShape) {
226 // Needed by low level implementation, but not really used.
227 tflite::RuntimeShape blockSizeDim;
228 NNTRACE_COMP("optimized_ops::SpaceToBatchND");
229 tflite::optimized_ops::SpaceToBatchND(
230 {.output_offset = outputShape.offset}, convertShapeToTflshape(inputShape), inputData,
231 blockSizeDim, blockSize, convertShapeToTflshape(paddingShape), padding,
232 convertShapeToTflshape(outputShape), outputData);
233 return true;
234 }
235 template bool spaceToBatchGeneric<float>(const float* inputData, const Shape& inputShape,
236 const int32_t* blockSize, const int32_t* padding,
237 const Shape& paddingShape, float* outputData,
238 const Shape& outputShape);
239 template bool spaceToBatchGeneric<_Float16>(const _Float16* inputData, const Shape& inputShape,
240 const int32_t* blockSize, const int32_t* padding,
241 const Shape& paddingShape, _Float16* outputData,
242 const Shape& outputShape);
243 template bool spaceToBatchGeneric<uint8_t>(const uint8_t* inputData, const Shape& inputShape,
244 const int32_t* blockSize, const int32_t* padding,
245 const Shape& paddingShape, uint8_t* outputData,
246 const Shape& outputShape);
247 template bool spaceToBatchGeneric<int8_t>(const int8_t* inputData, const Shape& inputShape,
248 const int32_t* blockSize, const int32_t* padding,
249 const Shape& paddingShape, int8_t* outputData,
250 const Shape& outputShape);
251
252 } // namespace nn
253 } // namespace android
254