1 //
2 // Copyright © 2020, 2023 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #pragma once
7
8 #include "TestUtils.hpp"
9
10 #include <armnn_delegate.hpp>
11 #include <DelegateTestInterpreter.hpp>
12
13 #include <flatbuffers/flatbuffers.h>
14 #include <tensorflow/lite/kernels/register.h>
15 #include <tensorflow/lite/version.h>
16
17 #include <schema_generated.h>
18
19 #include <doctest/doctest.h>
20
21 namespace
22 {
23
CreateSplitTfLiteModel(tflite::TensorType tensorType,std::vector<int32_t> & axisTensorShape,std::vector<int32_t> & inputTensorShape,const std::vector<std::vector<int32_t>> & outputTensorShapes,std::vector<int32_t> & axisData,const int32_t numSplits,float quantScale=1.0f,int quantOffset=0)24 std::vector<char> CreateSplitTfLiteModel(tflite::TensorType tensorType,
25 std::vector<int32_t>& axisTensorShape,
26 std::vector<int32_t>& inputTensorShape,
27 const std::vector<std::vector<int32_t>>& outputTensorShapes,
28 std::vector<int32_t>& axisData,
29 const int32_t numSplits,
30 float quantScale = 1.0f,
31 int quantOffset = 0)
32 {
33 using namespace tflite;
34 flatbuffers::FlatBufferBuilder flatBufferBuilder;
35
36 std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
37 buffers.push_back(CreateBuffer(flatBufferBuilder));
38 buffers.push_back(CreateBuffer(flatBufferBuilder));
39 buffers.push_back(CreateBuffer(flatBufferBuilder,
40 flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(axisData.data()),
41 sizeof(int32_t) * axisData.size())));
42
43 auto quantizationParameters =
44 CreateQuantizationParameters(flatBufferBuilder,
45 0,
46 0,
47 flatBufferBuilder.CreateVector<float>({ quantScale }),
48 flatBufferBuilder.CreateVector<int64_t>({ quantOffset }));
49
50 std::array<flatbuffers::Offset<Tensor>, 4> tensors;
51 tensors[0] = CreateTensor(flatBufferBuilder,
52 flatBufferBuilder.CreateVector<int32_t>(axisTensorShape.data(),
53 axisTensorShape.size()),
54 ::tflite::TensorType_INT32,
55 2,
56 flatBufferBuilder.CreateString("axis"),
57 quantizationParameters);
58 tensors[1] = CreateTensor(flatBufferBuilder,
59 flatBufferBuilder.CreateVector<int32_t>(inputTensorShape.data(),
60 inputTensorShape.size()),
61 tensorType,
62 1,
63 flatBufferBuilder.CreateString("input"),
64 quantizationParameters);
65
66 // Create output tensor
67 for (unsigned int i = 0; i < outputTensorShapes.size(); ++i)
68 {
69 buffers.push_back(CreateBuffer(flatBufferBuilder));
70 tensors[i + 2] = CreateTensor(flatBufferBuilder,
71 flatBufferBuilder.CreateVector<int32_t>(outputTensorShapes[i].data(),
72 outputTensorShapes[i].size()),
73 tensorType,
74 (i+3),
75 flatBufferBuilder.CreateString("output"),
76 quantizationParameters);
77 }
78
79 // create operator. Mean uses ReducerOptions.
80 tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_SplitOptions;
81 flatbuffers::Offset<void> operatorBuiltinOptions = CreateSplitOptions(flatBufferBuilder, numSplits).Union();
82
83 const std::vector<int> operatorInputs{ {0, 1} };
84 const std::vector<int> operatorOutputs{ {2, 3} };
85 flatbuffers::Offset <Operator> controlOperator =
86 CreateOperator(flatBufferBuilder,
87 0,
88 flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
89 flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
90 operatorBuiltinOptionsType,
91 operatorBuiltinOptions);
92
93 const std::vector<int> subgraphInputs{ {0, 1} };
94 const std::vector<int> subgraphOutputs{ {2, 3} };
95 flatbuffers::Offset <SubGraph> subgraph =
96 CreateSubGraph(flatBufferBuilder,
97 flatBufferBuilder.CreateVector(tensors.data(), tensors.size()),
98 flatBufferBuilder.CreateVector<int32_t>(subgraphInputs.data(), subgraphInputs.size()),
99 flatBufferBuilder.CreateVector<int32_t>(subgraphOutputs.data(), subgraphOutputs.size()),
100 flatBufferBuilder.CreateVector(&controlOperator, 1));
101
102 flatbuffers::Offset <flatbuffers::String> modelDescription =
103 flatBufferBuilder.CreateString("ArmnnDelegate: SPLIT Operator Model");
104 flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, BuiltinOperator_SPLIT);
105
106 flatbuffers::Offset <Model> flatbufferModel =
107 CreateModel(flatBufferBuilder,
108 TFLITE_SCHEMA_VERSION,
109 flatBufferBuilder.CreateVector(&operatorCode, 1),
110 flatBufferBuilder.CreateVector(&subgraph, 1),
111 modelDescription,
112 flatBufferBuilder.CreateVector(buffers));
113
114 flatBufferBuilder.Finish(flatbufferModel, armnnDelegate::FILE_IDENTIFIER);
115
116 return std::vector<char>(flatBufferBuilder.GetBufferPointer(),
117 flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
118 }
119
120 template <typename T>
SplitTest(tflite::TensorType tensorType,std::vector<armnn::BackendId> & backends,std::vector<int32_t> & axisTensorShape,std::vector<int32_t> & inputTensorShape,std::vector<std::vector<int32_t>> & outputTensorShapes,std::vector<int32_t> & axisData,std::vector<T> & inputValues,std::vector<std::vector<T>> & expectedOutputValues,const int32_t numSplits,float quantScale=1.0f,int quantOffset=0)121 void SplitTest(tflite::TensorType tensorType,
122 std::vector<armnn::BackendId>& backends,
123 std::vector<int32_t>& axisTensorShape,
124 std::vector<int32_t>& inputTensorShape,
125 std::vector<std::vector<int32_t>>& outputTensorShapes,
126 std::vector<int32_t>& axisData,
127 std::vector<T>& inputValues,
128 std::vector<std::vector<T>>& expectedOutputValues,
129 const int32_t numSplits,
130 float quantScale = 1.0f,
131 int quantOffset = 0)
132 {
133 using namespace delegateTestInterpreter;
134 std::vector<char> modelBuffer = CreateSplitTfLiteModel(tensorType,
135 axisTensorShape,
136 inputTensorShape,
137 outputTensorShapes,
138 axisData,
139 numSplits,
140 quantScale,
141 quantOffset);
142 // Setup interpreter with just TFLite Runtime.
143 auto tfLiteInterpreter = DelegateTestInterpreter(modelBuffer);
144 CHECK(tfLiteInterpreter.AllocateTensors() == kTfLiteOk);
145 CHECK(tfLiteInterpreter.FillInputTensor<T>(inputValues, 1) == kTfLiteOk);
146 CHECK(tfLiteInterpreter.Invoke() == kTfLiteOk);
147
148 // Setup interpreter with Arm NN Delegate applied.
149 auto armnnInterpreter = DelegateTestInterpreter(modelBuffer, backends);
150 CHECK(armnnInterpreter.AllocateTensors() == kTfLiteOk);
151 CHECK(armnnInterpreter.FillInputTensor<T>(inputValues, 1) == kTfLiteOk);
152 CHECK(armnnInterpreter.Invoke() == kTfLiteOk);
153
154 // Compare output data
155 for (unsigned int i = 0; i < expectedOutputValues.size(); ++i)
156 {
157 std::vector<T> tfLiteOutputValues = tfLiteInterpreter.GetOutputResult<T>(i);
158 std::vector<int32_t> tfLiteOutputShape = tfLiteInterpreter.GetOutputShape(i);
159
160 std::vector<T> armnnOutputValues = armnnInterpreter.GetOutputResult<T>(i);
161 std::vector<int32_t> armnnOutputShape = armnnInterpreter.GetOutputShape(i);
162
163 armnnDelegate::CompareOutputData<T>(tfLiteOutputValues, armnnOutputValues, expectedOutputValues[i]);
164 armnnDelegate::CompareOutputShape(tfLiteOutputShape, armnnOutputShape, outputTensorShapes[i]);
165 }
166
167 tfLiteInterpreter.Cleanup();
168 armnnInterpreter.Cleanup();
169
170 } // End of SPLIT Test
171
CreateSplitVTfLiteModel(tflite::TensorType tensorType,std::vector<int32_t> & inputTensorShape,std::vector<int32_t> & splitsTensorShape,std::vector<int32_t> & axisTensorShape,const std::vector<std::vector<int32_t>> & outputTensorShapes,std::vector<int32_t> & splitsData,std::vector<int32_t> & axisData,const int32_t numSplits,float quantScale=1.0f,int quantOffset=0)172 std::vector<char> CreateSplitVTfLiteModel(tflite::TensorType tensorType,
173 std::vector<int32_t>& inputTensorShape,
174 std::vector<int32_t>& splitsTensorShape,
175 std::vector<int32_t>& axisTensorShape,
176 const std::vector<std::vector<int32_t>>& outputTensorShapes,
177 std::vector<int32_t>& splitsData,
178 std::vector<int32_t>& axisData,
179 const int32_t numSplits,
180 float quantScale = 1.0f,
181 int quantOffset = 0)
182 {
183 using namespace tflite;
184 flatbuffers::FlatBufferBuilder flatBufferBuilder;
185
186 std::array<flatbuffers::Offset<tflite::Buffer>, 3> buffers;
187 buffers[0] = CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({}));
188 buffers[1] = CreateBuffer(flatBufferBuilder,
189 flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(splitsData.data()),
190 sizeof(int32_t) * splitsData.size()));
191 buffers[2] = CreateBuffer(flatBufferBuilder,
192 flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(axisData.data()),
193 sizeof(int32_t) * axisData.size()));
194
195 auto quantizationParameters =
196 CreateQuantizationParameters(flatBufferBuilder,
197 0,
198 0,
199 flatBufferBuilder.CreateVector<float>({ quantScale }),
200 flatBufferBuilder.CreateVector<int64_t>({ quantOffset }));
201
202 std::array<flatbuffers::Offset<Tensor>, 5> tensors;
203 tensors[0] = CreateTensor(flatBufferBuilder,
204 flatBufferBuilder.CreateVector<int32_t>(inputTensorShape.data(),
205 inputTensorShape.size()),
206 tensorType,
207 0,
208 flatBufferBuilder.CreateString("input"),
209 quantizationParameters);
210 tensors[1] = CreateTensor(flatBufferBuilder,
211 flatBufferBuilder.CreateVector<int32_t>(splitsTensorShape.data(),
212 splitsTensorShape.size()),
213 ::tflite::TensorType_INT32,
214 1,
215 flatBufferBuilder.CreateString("splits"),
216 quantizationParameters);
217 tensors[2] = CreateTensor(flatBufferBuilder,
218 flatBufferBuilder.CreateVector<int32_t>(axisTensorShape.data(),
219 axisTensorShape.size()),
220 ::tflite::TensorType_INT32,
221 2,
222 flatBufferBuilder.CreateString("axis"),
223 quantizationParameters);
224
225 // Create output tensor
226 for (unsigned int i = 0; i < outputTensorShapes.size(); ++i)
227 {
228 tensors[i + 3] = CreateTensor(flatBufferBuilder,
229 flatBufferBuilder.CreateVector<int32_t>(outputTensorShapes[i].data(),
230 outputTensorShapes[i].size()),
231 tensorType,
232 0,
233 flatBufferBuilder.CreateString("output"),
234 quantizationParameters);
235 }
236
237 // create operator. Mean uses ReducerOptions.
238 tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_SplitVOptions;
239 flatbuffers::Offset<void> operatorBuiltinOptions = CreateSplitVOptions(flatBufferBuilder, numSplits).Union();
240
241 const std::vector<int> operatorInputs{ {0, 1, 2} };
242 const std::vector<int> operatorOutputs{ {3, 4} };
243 flatbuffers::Offset <Operator> controlOperator =
244 CreateOperator(flatBufferBuilder,
245 0,
246 flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
247 flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
248 operatorBuiltinOptionsType,
249 operatorBuiltinOptions);
250
251 const std::vector<int> subgraphInputs{ {0, 1, 2} };
252 const std::vector<int> subgraphOutputs{ {3, 4} };
253 flatbuffers::Offset <SubGraph> subgraph =
254 CreateSubGraph(flatBufferBuilder,
255 flatBufferBuilder.CreateVector(tensors.data(), tensors.size()),
256 flatBufferBuilder.CreateVector<int32_t>(subgraphInputs.data(), subgraphInputs.size()),
257 flatBufferBuilder.CreateVector<int32_t>(subgraphOutputs.data(), subgraphOutputs.size()),
258 flatBufferBuilder.CreateVector(&controlOperator, 1));
259
260 flatbuffers::Offset <flatbuffers::String> modelDescription =
261 flatBufferBuilder.CreateString("ArmnnDelegate: SPLIT_V Operator Model");
262 flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, BuiltinOperator_SPLIT_V);
263
264 flatbuffers::Offset <Model> flatbufferModel =
265 CreateModel(flatBufferBuilder,
266 TFLITE_SCHEMA_VERSION,
267 flatBufferBuilder.CreateVector(&operatorCode, 1),
268 flatBufferBuilder.CreateVector(&subgraph, 1),
269 modelDescription,
270 flatBufferBuilder.CreateVector(buffers.data(), buffers.size()));
271
272 flatBufferBuilder.Finish(flatbufferModel, armnnDelegate::FILE_IDENTIFIER);
273
274 return std::vector<char>(flatBufferBuilder.GetBufferPointer(),
275 flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
276 }
277
278 template <typename T>
SplitVTest(tflite::TensorType tensorType,std::vector<armnn::BackendId> & backends,std::vector<int32_t> & inputTensorShape,std::vector<int32_t> & splitsTensorShape,std::vector<int32_t> & axisTensorShape,std::vector<std::vector<int32_t>> & outputTensorShapes,std::vector<T> & inputValues,std::vector<int32_t> & splitsData,std::vector<int32_t> & axisData,std::vector<std::vector<T>> & expectedOutputValues,const int32_t numSplits,float quantScale=1.0f,int quantOffset=0)279 void SplitVTest(tflite::TensorType tensorType,
280 std::vector<armnn::BackendId>& backends,
281 std::vector<int32_t>& inputTensorShape,
282 std::vector<int32_t>& splitsTensorShape,
283 std::vector<int32_t>& axisTensorShape,
284 std::vector<std::vector<int32_t>>& outputTensorShapes,
285 std::vector<T>& inputValues,
286 std::vector<int32_t>& splitsData,
287 std::vector<int32_t>& axisData,
288 std::vector<std::vector<T>>& expectedOutputValues,
289 const int32_t numSplits,
290 float quantScale = 1.0f,
291 int quantOffset = 0)
292 {
293 using namespace delegateTestInterpreter;
294 std::vector<char> modelBuffer = CreateSplitVTfLiteModel(tensorType,
295 inputTensorShape,
296 splitsTensorShape,
297 axisTensorShape,
298 outputTensorShapes,
299 splitsData,
300 axisData,
301 numSplits,
302 quantScale,
303 quantOffset);
304
305 // Setup interpreter with just TFLite Runtime.
306 auto tfLiteInterpreter = DelegateTestInterpreter(modelBuffer);
307 CHECK(tfLiteInterpreter.AllocateTensors() == kTfLiteOk);
308 CHECK(tfLiteInterpreter.FillInputTensor<T>(inputValues, 0) == kTfLiteOk);
309 CHECK(tfLiteInterpreter.Invoke() == kTfLiteOk);
310
311 // Setup interpreter with Arm NN Delegate applied.
312 auto armnnInterpreter = DelegateTestInterpreter(modelBuffer, backends);
313 CHECK(armnnInterpreter.AllocateTensors() == kTfLiteOk);
314 CHECK(armnnInterpreter.FillInputTensor<T>(inputValues, 0) == kTfLiteOk);
315 CHECK(armnnInterpreter.Invoke() == kTfLiteOk);
316
317 // Compare output data
318 for (unsigned int i = 0; i < expectedOutputValues.size(); ++i)
319 {
320 std::vector<T> tfLiteOutputValues = tfLiteInterpreter.GetOutputResult<T>(i);
321 std::vector<int32_t> tfLiteOutputShape = tfLiteInterpreter.GetOutputShape(i);
322
323 std::vector<T> armnnOutputValues = armnnInterpreter.GetOutputResult<T>(i);
324 std::vector<int32_t> armnnOutputShape = armnnInterpreter.GetOutputShape(i);
325
326 armnnDelegate::CompareOutputData<T>(tfLiteOutputValues, armnnOutputValues, expectedOutputValues[i]);
327 armnnDelegate::CompareOutputShape(tfLiteOutputShape, armnnOutputShape, outputTensorShapes[i]);
328 }
329
330 tfLiteInterpreter.Cleanup();
331 armnnInterpreter.Cleanup();
332 } // End of SPLIT_V Test
333
334 } // anonymous namespace