1 /*
2  * Copyright (C) 2019 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 #define LOG_TAG "Operations"
18 
19 #include <algorithm>
20 #include <vector>
21 
22 #include "OperationResolver.h"
23 #include "Tracing.h"
24 
25 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
26 #include <tensorflow/lite/kernels/internal/optimized/optimized_ops.h>
27 #include <tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h>
28 
29 #include "CpuOperationUtils.h"
30 #endif  // NN_INCLUDE_CPU_IMPLEMENTATION
31 
32 namespace android {
33 namespace nn {
34 namespace l2_norm {
35 
36 constexpr char kOperationName[] = "L2_NORMALIZATION";
37 
38 constexpr uint32_t kNumInputs = 2;
39 constexpr uint32_t kInputTensor = 0;
40 constexpr uint32_t kAxisScalar = 1;
41 
42 constexpr uint32_t kNumOutputs = 1;
43 constexpr uint32_t kOutputTensor = 0;
44 
45 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
46 namespace {
47 
l2normFloat32Impl(const float * inputData,const Shape & inputShape,int32_t axis,float * outputData,const Shape & outputShape)48 inline bool l2normFloat32Impl(const float* inputData, const Shape& inputShape, int32_t axis,
49                               float* outputData, const Shape& outputShape) {
50     NNTRACE_TRANS("l2normFloat32");
51     constexpr float kEpsilon = 1e-6f;
52     const uint32_t outerSize = getNumberOfElements(inputShape, 0, axis);
53     const uint32_t axisSize = getSizeOfDimension(inputShape, axis);
54     const uint32_t innerSize =
55             getNumberOfElements(inputShape, axis + 1, getNumberOfDimensions(inputShape));
56     for (uint32_t outer = 0; outer < outerSize; ++outer) {
57         const float* inputBeg = inputData + outer * axisSize * innerSize;
58         const float* inputEnd = inputBeg + axisSize * innerSize;
59         float* outputBeg = outputData + outer * axisSize * innerSize;
60         for (uint32_t inner = 0; inner < innerSize; ++inner, ++inputBeg, ++inputEnd, ++outputBeg) {
61             float sum = 0.0f;
62             for (const float* p = inputBeg; p < inputEnd; p += innerSize) {
63                 float val = *p;
64                 sum += val * val;
65             }
66             float l2_norm = std::max(std::sqrt(sum), kEpsilon);
67             float* pOut = outputBeg;
68             for (const float* p = inputBeg; p < inputEnd; p += innerSize, pOut += innerSize) {
69                 *pOut = *p / l2_norm;
70             }
71         }
72     }
73     return true;
74 }
75 
l2normQuant8Impl(const uint8_t * inputData,const Shape & inputShape,int32_t axis,uint8_t * outputData,const Shape & outputShape)76 inline bool l2normQuant8Impl(const uint8_t* inputData, const Shape& inputShape, int32_t axis,
77                              uint8_t* outputData, const Shape& outputShape) {
78     NNTRACE_TRANS("l2normQuant8");
79     const uint32_t outerSize = getNumberOfElements(inputShape, 0, axis);
80     const uint32_t axisSize = getSizeOfDimension(inputShape, axis);
81     const uint32_t innerSize =
82             getNumberOfElements(inputShape, axis + 1, getNumberOfDimensions(inputShape));
83     for (uint32_t outer = 0; outer < outerSize; ++outer) {
84         const uint8_t* inputBeg = inputData + outer * axisSize * innerSize;
85         const uint8_t* inputEnd = inputBeg + axisSize * innerSize;
86         uint8_t* outputBeg = outputData + outer * axisSize * innerSize;
87         for (uint32_t inner = 0; inner < innerSize; ++inner, ++inputBeg, ++inputEnd, ++outputBeg) {
88             int32_t sum = 0;
89             for (const uint8_t* p = inputBeg; p < inputEnd; p += innerSize) {
90                 int32_t val = static_cast<int32_t>(*p) - inputShape.offset;
91                 sum += val * val;
92             }
93             int32_t invMultiplier, invShift;
94             tflite::GetInvSqrtQuantizedMultiplierExp(sum, -1, &invMultiplier, &invShift);
95             uint8_t* pOut = outputBeg;
96             for (const uint8_t* p = inputBeg; p < inputEnd; p += innerSize, pOut += innerSize) {
97                 int32_t val = static_cast<int32_t>(*p) - inputShape.offset;
98                 int32_t scaledVal = tflite::MultiplyByQuantizedMultiplierSmallerThanOneExp(
99                                             val * 128, invMultiplier, invShift) +
100                                     128;
101                 *pOut = static_cast<uint8_t>(std::min(std::max(scaledVal, 0), 255));
102             }
103         }
104     }
105     return true;
106 }
107 
l2normQuant8SignedImpl(const int8_t * inputData,const Shape & inputShape,int32_t axis,int8_t * outputData,const Shape & outputShape)108 inline bool l2normQuant8SignedImpl(const int8_t* inputData, const Shape& inputShape, int32_t axis,
109                                    int8_t* outputData, const Shape& outputShape) {
110     NNTRACE_TRANS("l2normQuant8Signed");
111     const uint32_t outerSize = getNumberOfElements(inputShape, 0, axis);
112     const uint32_t axisSize = getSizeOfDimension(inputShape, axis);
113     const uint32_t innerSize =
114             getNumberOfElements(inputShape, axis + 1, getNumberOfDimensions(inputShape));
115     for (uint32_t outer = 0; outer < outerSize; ++outer) {
116         const int8_t* inputBeg = inputData + outer * axisSize * innerSize;
117         const int8_t* inputEnd = inputBeg + axisSize * innerSize;
118         int8_t* outputBeg = outputData + outer * axisSize * innerSize;
119         for (uint32_t inner = 0; inner < innerSize; ++inner, ++inputBeg, ++inputEnd, ++outputBeg) {
120             int32_t sum = 0;
121             for (const int8_t* p = inputBeg; p < inputEnd; p += innerSize) {
122                 int32_t val = static_cast<int32_t>(*p) - inputShape.offset;
123                 sum += val * val;
124             }
125             int32_t invMultiplier, invShift;
126             tflite::GetInvSqrtQuantizedMultiplierExp(sum, -1, &invMultiplier, &invShift);
127             int8_t* pOut = outputBeg;
128             for (const int8_t* p = inputBeg; p < inputEnd; p += innerSize, pOut += innerSize) {
129                 int32_t val = static_cast<int32_t>(*p) - inputShape.offset;
130                 int32_t scaledVal = tflite::MultiplyByQuantizedMultiplierSmallerThanOneExp(
131                         val * 128, invMultiplier, invShift);
132                 *pOut = static_cast<int8_t>(std::min(std::max(scaledVal, -128), 127));
133             }
134         }
135     }
136     return true;
137 }
138 
l2normFloat32(const float * inputData,const Shape & inputShape,int32_t axis,float * outputData,const Shape & outputShape)139 bool l2normFloat32(const float* inputData, const Shape& inputShape, int32_t axis, float* outputData,
140                    const Shape& outputShape) {
141     int32_t ndim = getNumberOfDimensions(inputShape);
142     NN_CHECK(handleNegativeAxis(inputShape, &axis));
143     // TFLite optimized implementation only supports computation along the last axis
144     if (axis == ndim - 1) {
145         NNTRACE_COMP("optimized_ops::L2Normalization::float");
146         tflite::L2NormalizationParams param = {.input_zero_point = 0};
147         tflite::optimized_ops::L2Normalization(param, convertShapeToTflshape(inputShape), inputData,
148                                                convertShapeToTflshape(outputShape), outputData);
149         return true;
150     } else {
151         return l2normFloat32Impl(inputData, inputShape, axis, outputData, outputShape);
152     }
153 }
154 
l2normFloat16(const _Float16 * inputData,const Shape & inputShape,int32_t axis,_Float16 * outputData,const Shape & outputShape)155 bool l2normFloat16(const _Float16* inputData, const Shape& inputShape, int32_t axis,
156                    _Float16* outputData, const Shape& outputShape) {
157     NNTRACE_TRANS("l2normFloat16");
158     std::vector<float> inputDataFloat32(getNumberOfElements(inputShape));
159     convertFloat16ToFloat32(inputData, &inputDataFloat32);
160     std::vector<float> outputDataFloat32(getNumberOfElements(outputShape));
161 
162     l2normFloat32(inputDataFloat32.data(), inputShape, axis, outputDataFloat32.data(), outputShape);
163     convertFloat32ToFloat16(outputDataFloat32, outputData);
164 
165     return true;
166 }
167 
l2normQuant8(const uint8_t * inputData,const Shape & inputShape,int32_t axis,uint8_t * outputData,const Shape & outputShape)168 bool l2normQuant8(const uint8_t* inputData, const Shape& inputShape, int32_t axis,
169                   uint8_t* outputData, const Shape& outputShape) {
170     int32_t ndim = getNumberOfDimensions(inputShape);
171     NN_CHECK(handleNegativeAxis(inputShape, &axis));
172     // TFLite optimized implementation only supports computation along the last axis
173     if (axis == ndim - 1) {
174         NNTRACE_COMP("optimized_ops::L2Normalization::uint8");
175         tflite::L2NormalizationParams param = {.input_zero_point = inputShape.offset};
176         tflite::optimized_ops::L2Normalization(param, convertShapeToTflshape(inputShape), inputData,
177                                                convertShapeToTflshape(outputShape), outputData);
178         return true;
179     } else {
180         return l2normQuant8Impl(inputData, inputShape, axis, outputData, outputShape);
181     }
182 }
183 
l2normQuant8Signed(const int8_t * inputData,const Shape & inputShape,int32_t axis,int8_t * outputData,const Shape & outputShape)184 bool l2normQuant8Signed(const int8_t* inputData, const Shape& inputShape, int32_t axis,
185                         int8_t* outputData, const Shape& outputShape) {
186     int32_t ndim = getNumberOfDimensions(inputShape);
187     NN_CHECK(handleNegativeAxis(inputShape, &axis));
188     // TFLite implementation only supports computation along the last axis
189     if (axis == ndim - 1) {
190         NNTRACE_COMP("reference_integer_ops::L2Normalization");
191         const int32_t outerSize = getNumberOfElements(inputShape, 0, axis);
192         const int32_t axisSize = getSizeOfDimension(inputShape, axis);
193         tflite::reference_integer_ops::L2Normalization(inputShape.offset, outerSize, axisSize,
194                                                        inputData, outputData);
195         return true;
196     } else {
197         return l2normQuant8SignedImpl(inputData, inputShape, axis, outputData, outputShape);
198     }
199 }
200 
201 }  // namespace
202 #endif  // NN_INCLUDE_CPU_IMPLEMENTATION
203 
validate(const IOperationValidationContext * context)204 Result<Version> validate(const IOperationValidationContext* context) {
205     NN_RET_CHECK(context->getNumInputs() == kNumInputs ||
206                  context->getNumInputs() == kNumInputs - 1);
207     NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
208 
209     const OperandType inputType = context->getInputType(kInputTensor);
210     std::vector<OperandType> inExpectedTypes = {inputType};
211     auto minSupportedVersion = Version::ANDROID_OC_MR1;
212     if (inputType == OperandType::TENSOR_FLOAT16 || inputType == OperandType::TENSOR_QUANT8_ASYMM) {
213         minSupportedVersion = Version::ANDROID_Q;
214     } else if (inputType == OperandType::TENSOR_FLOAT32) {
215         minSupportedVersion = Version::ANDROID_OC_MR1;
216     } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
217         minSupportedVersion = Version::ANDROID_R;
218     } else {
219         NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
220     }
221     if (context->getNumInputs() == kNumInputs) {
222         inExpectedTypes.push_back(OperandType::INT32);
223         minSupportedVersion = Version::ANDROID_Q;
224     } else if (context->getInputShape(kInputTensor).dimensions.size() != 4) {
225         minSupportedVersion = Version::ANDROID_Q;
226     }
227     const Shape& input = context->getInputShape(kInputTensor);
228     if (hasKnownRank(input)) {
229         NN_RET_CHECK_LE(getNumberOfDimensions(input), 4);
230     }
231     NN_RET_CHECK(validateInputTypes(context, inExpectedTypes));
232     NN_RET_CHECK(validateOutputTypes(context, {inputType}));
233     return minSupportedVersion;
234 }
235 
236 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
prepare(IOperationExecutionContext * context)237 bool prepare(IOperationExecutionContext* context) {
238     const Shape& input = context->getInputShape(kInputTensor);
239     int32_t numDimensions = getNumberOfDimensions(input);
240     int32_t axis = context->getNumInputs() == kNumInputs
241                            ? context->getInputValue<int32_t>(kAxisScalar)
242                            : -1;
243     NN_RET_CHECK_LE(numDimensions, 4);
244     NN_RET_CHECK_GE(axis, -numDimensions);
245     NN_RET_CHECK_LT(axis, numDimensions);
246     Shape output = context->getOutputShape(kOutputTensor);
247     output.type = input.type;
248     output.dimensions = input.dimensions;
249     if (output.type == OperandType::TENSOR_QUANT8_ASYMM) {
250         output.scale = 1.0f / 128.0f;
251         output.offset = 128;
252     } else if (output.type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
253         output.scale = 1.0f / 128.0f;
254         output.offset = 0;
255     } else {
256         output.scale = 0;
257         output.offset = 0;
258     }
259     return context->setOutputShape(kOutputTensor, output);
260 }
261 
execute(IOperationExecutionContext * context)262 bool execute(IOperationExecutionContext* context) {
263     int32_t axis = context->getNumInputs() == kNumInputs
264                            ? context->getInputValue<int32_t>(kAxisScalar)
265                            : -1;
266     NN_RET_CHECK(handleNegativeAxis(context->getInputShape(kInputTensor), &axis));
267     switch (context->getInputType(kInputTensor)) {
268         case OperandType::TENSOR_FLOAT32:
269             return l2normFloat32(context->getInputBuffer<float>(kInputTensor),
270                                  context->getInputShape(kInputTensor), axis,
271                                  context->getOutputBuffer<float>(kOutputTensor),
272                                  context->getOutputShape(kOutputTensor));
273         case OperandType::TENSOR_FLOAT16:
274             return l2normFloat16(context->getInputBuffer<_Float16>(kInputTensor),
275                                  context->getInputShape(kInputTensor), axis,
276                                  context->getOutputBuffer<_Float16>(kOutputTensor),
277                                  context->getOutputShape(kOutputTensor));
278         case OperandType::TENSOR_QUANT8_ASYMM:
279             return l2normQuant8(context->getInputBuffer<uint8_t>(kInputTensor),
280                                 context->getInputShape(kInputTensor), axis,
281                                 context->getOutputBuffer<uint8_t>(kOutputTensor),
282                                 context->getOutputShape(kOutputTensor));
283         case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
284             return l2normQuant8Signed(context->getInputBuffer<int8_t>(kInputTensor),
285                                       context->getInputShape(kInputTensor), axis,
286                                       context->getOutputBuffer<int8_t>(kOutputTensor),
287                                       context->getOutputShape(kOutputTensor));
288         default:
289             NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
290     }
291 }
292 #endif  // NN_INCLUDE_CPU_IMPLEMENTATION
293 
294 }  // namespace l2_norm
295 
296 NN_REGISTER_OPERATION(L2_NORMALIZATION, l2_norm::kOperationName, l2_norm::validate,
297                       l2_norm::prepare, l2_norm::execute);
298 
299 }  // namespace nn
300 }  // namespace android
301