Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "AdditionTestImpl.hpp" |
| 7 | |
| 8 | #include "ElementwiseTestImpl.hpp" |
| 9 | |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 10 | #include <QuantizeHelper.hpp> |
| 11 | |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 12 | template<> |
| 13 | std::unique_ptr<armnn::IWorkload> CreateWorkload<armnn::AdditionQueueDescriptor>( |
| 14 | const armnn::IWorkloadFactory& workloadFactory, |
| 15 | const armnn::WorkloadInfo& info, |
| 16 | const armnn::AdditionQueueDescriptor& descriptor) |
| 17 | { |
| 18 | return workloadFactory.CreateAddition(descriptor, info); |
| 19 | } |
| 20 | |
| 21 | LayerTestResult<float,4> AdditionTest( |
| 22 | armnn::IWorkloadFactory& workloadFactory, |
| 23 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 24 | { |
| 25 | unsigned int batchSize = 2u; |
| 26 | unsigned int channels = 2u; |
| 27 | unsigned int height = 2u; |
| 28 | unsigned int width = 3u; |
| 29 | |
| 30 | unsigned int shape[] = { batchSize, channels, height, width }; |
| 31 | |
| 32 | std::vector<float> input1 = |
| 33 | { |
| 34 | 0.0f, 2.0f, 1.0f, |
| 35 | 0.2f, 1.0f, 2.0f, |
| 36 | |
| 37 | 1.0f, 2.0f, 1.0f, |
| 38 | 0.2f, 1.0f, 2.0f, |
| 39 | |
| 40 | 0.0f, 2.0f, 1.0f, |
| 41 | 4.2f, 1.0f, 2.0f, |
| 42 | |
| 43 | 0.0f, 0.0f, 1.0f, |
| 44 | 0.2f, 1.0f, 2.0f, |
| 45 | }; |
| 46 | |
| 47 | std::vector<float> input2 = |
| 48 | { |
| 49 | 1.0f, 2.0f, 1.0f, |
| 50 | 0.0f, 1.0f, 2.0f, |
| 51 | |
| 52 | 1.0f, 2.0f, -2.0f, |
| 53 | 0.2f, 1.0f, 2.0f, |
| 54 | |
| 55 | 0.0f, 2.0f, 1.0f, |
| 56 | 4.2f, 0.0f, -3.0f, |
| 57 | |
| 58 | 0.0f, 0.0f, 1.0f, |
| 59 | 0.7f, 1.0f, 5.0f, |
| 60 | }; |
| 61 | |
| 62 | |
| 63 | std::vector<float> output |
| 64 | { |
| 65 | 1.0f, 4.0f, 2.0f, |
| 66 | 0.2f, 2.0f, 4.0f, |
| 67 | |
| 68 | 2.0f, 4.0f, -1.0f, |
| 69 | 0.4f, 2.0f, 4.0f, |
| 70 | |
| 71 | 0.0f, 4.0f, 2.0f, |
| 72 | 8.4f, 1.0f, -1.0f, |
| 73 | |
| 74 | 0.0f, 0.0f, 2.0f, |
| 75 | 0.9f, 2.0f, 7.0f, |
| 76 | }; |
| 77 | |
| 78 | return ElementwiseTestHelper<4, armnn::AdditionQueueDescriptor, armnn::DataType::Float32>( |
| 79 | workloadFactory, |
| 80 | memoryManager, |
| 81 | shape, |
| 82 | input1, |
| 83 | shape, |
| 84 | input2, |
| 85 | shape, |
| 86 | output); |
| 87 | } |
| 88 | |
| 89 | LayerTestResult<float, 5> Addition5dTest( |
| 90 | armnn::IWorkloadFactory& workloadFactory, |
| 91 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 92 | { |
| 93 | unsigned int depth = 2u; |
| 94 | unsigned int batchSize = 2u; |
| 95 | unsigned int channels = 2u; |
| 96 | unsigned int height = 2u; |
| 97 | unsigned int width = 3u; |
| 98 | |
| 99 | unsigned int shape[] = { depth, batchSize, channels, height, width }; |
| 100 | |
| 101 | std::vector<float> input1 = |
| 102 | { |
| 103 | 2.6f, 4.0f, 4.4f, 2.7f, 4.6f, 2.8f, |
| 104 | 2.3f, 1.9f, 3.4f, 2.9f, 2.2f, 4.5f, |
| 105 | |
| 106 | 2.8f, 1.9f, 2.3f, 2.6f, 4.7f, 3.5f, |
| 107 | 0.4f, 1.5f, 2.1f, 0.7f, 5.0f, 1.1f, |
| 108 | |
| 109 | |
| 110 | 1.0f, 2.7f, 0.0f, 0.6f, 0.8f, 0.9f, |
| 111 | 1.0f, 2.6f, 0.4f, 3.8f, 0.4f, 0.8f, |
| 112 | |
| 113 | 0.5f, 4.3f, 3.1f, 4.4f, 0.7f, 1.4f, |
| 114 | 0.4f, 4.4f, 0.7f, 0.6f, 4.7f, 1.2f, |
| 115 | |
| 116 | }; |
| 117 | |
| 118 | std::vector<float> input2 = |
| 119 | { |
| 120 | 4.4f, 3.0f, 1.0f, 0.0f, 3.9f, 3.1f, |
| 121 | 1.7f, 2.9f, 1.3f, 0.4f, 0.4f, 4.3f, |
| 122 | |
| 123 | 4.5f, 0.2f, 2.2f, 4.1f, 3.9f, 3.0f, |
| 124 | 0.1f, 2.5f, 4.1f, 4.6f, 1.5f, 0.0f, |
| 125 | |
| 126 | |
| 127 | 0.5f, 4.9f, 2.5f, 1.5f, 3.4f, 4.5f, |
| 128 | 2.0f, 3.0f, 4.9f, 1.6f, 2.4f, 3.4f, |
| 129 | |
| 130 | 3.6f, 1.8f, 1.3f, 2.6f, 2.1f, 4.8f, |
| 131 | 2.0f, 4.3f, 4.0f, 0.2f, 0.6f, 4.4f, |
| 132 | }; |
| 133 | |
| 134 | std::vector<float> output = |
| 135 | { |
| 136 | 7.0f, 7.0f, 5.4f, 2.7f, 8.5f, 5.9f, |
| 137 | 4.0f, 4.8f, 4.7f, 3.3f, 2.6f, 8.8f, |
| 138 | |
| 139 | 7.3f, 2.1f, 4.5f, 6.7f, 8.6f, 6.5f, |
| 140 | 0.5f, 4.0f, 6.2f, 5.3f, 6.5f, 1.1f, |
| 141 | |
| 142 | |
| 143 | 1.5f, 7.6f, 2.5f, 2.1f, 4.2f, 5.4f, |
| 144 | 3.0f, 5.6f, 5.3f, 5.4f, 2.8f, 4.2f, |
| 145 | |
| 146 | 4.1f, 6.1f, 4.4f, 7.0f, 2.8f, 6.2f, |
| 147 | 2.4f, 8.7f, 4.7f, 0.8f, 5.3f, 5.6f, |
| 148 | }; |
| 149 | |
| 150 | return ElementwiseTestHelper<5, armnn::AdditionQueueDescriptor, armnn::DataType::Float32>( |
| 151 | workloadFactory, |
| 152 | memoryManager, |
| 153 | shape, |
| 154 | input1, |
| 155 | shape, |
| 156 | input2, |
| 157 | shape, |
| 158 | output); |
| 159 | } |
| 160 | |
| 161 | template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> |
| 162 | LayerTestResult<T, 4> AdditionBroadcastTestImpl( |
| 163 | armnn::IWorkloadFactory& workloadFactory, |
| 164 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 165 | float qScale, |
| 166 | int32_t qOffset) |
| 167 | { |
| 168 | armnn::TensorInfo inputTensorInfo1 = armnn::TensorInfo({1, 3, 2, 1}, ArmnnType); |
| 169 | armnn::TensorInfo inputTensorInfo2 = armnn::TensorInfo({1, 1, 2, 3}, ArmnnType); |
| 170 | armnn::TensorInfo outputTensorInfo = armnn::TensorInfo({1, 3, 2, 3}, ArmnnType); |
| 171 | |
| 172 | if (armnn::IsQuantizedType<T>()) |
| 173 | { |
| 174 | inputTensorInfo1.SetQuantizationScale(qScale); |
| 175 | inputTensorInfo1.SetQuantizationOffset(qOffset); |
| 176 | inputTensorInfo2.SetQuantizationScale(qScale); |
| 177 | inputTensorInfo2.SetQuantizationOffset(qOffset); |
| 178 | outputTensorInfo.SetQuantizationScale(qScale); |
| 179 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 180 | } |
| 181 | |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 182 | auto input1 = MakeTensor<T, 4>(inputTensorInfo1, armnnUtils::QuantizedVector<T>( |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 183 | { |
| 184 | 0.0f, |
| 185 | 1.0f, |
| 186 | |
| 187 | 2.0f, |
| 188 | 3.0f, |
| 189 | |
| 190 | 4.0f, |
| 191 | 5.0f, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 192 | }, |
| 193 | qScale, qOffset)); |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 194 | |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 195 | auto input2 = MakeTensor<T, 4>(inputTensorInfo2, armnnUtils::QuantizedVector<T>( |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 196 | { |
| 197 | 0.5f, 1.5f, 2.5f, |
| 198 | 3.5f, 4.5f, 5.5f, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 199 | }, |
| 200 | qScale, qOffset)); |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 201 | |
| 202 | LayerTestResult<T,4> ret(outputTensorInfo); |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 203 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, armnnUtils::QuantizedVector<T>( |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 204 | { |
| 205 | 0.5f, 1.5f, 2.5f, |
| 206 | 4.5f, 5.5f, 6.5f, |
| 207 | |
| 208 | 2.5f, 3.5f, 4.5f, |
| 209 | 6.5f, 7.5f, 8.5f, |
| 210 | |
| 211 | 4.5f, 5.5f, 6.5f, |
| 212 | 8.5f, 9.5f, 10.5f, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 213 | }, |
| 214 | qScale, qOffset)); |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 215 | |
| 216 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 217 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 218 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 219 | |
| 220 | armnn::AdditionQueueDescriptor data; |
| 221 | armnn::WorkloadInfo info; |
| 222 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 223 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
| 224 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 225 | |
| 226 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info); |
| 227 | |
| 228 | inputHandle1->Allocate(); |
| 229 | inputHandle2->Allocate(); |
| 230 | outputHandle->Allocate(); |
| 231 | |
| 232 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 233 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]); |
| 234 | |
| 235 | workload->PostAllocationConfigure(); |
| 236 | workload->Execute(); |
| 237 | |
| 238 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 239 | |
| 240 | return ret; |
| 241 | } |
| 242 | |
| 243 | template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> |
| 244 | LayerTestResult<T, 4> AdditionBroadcast1ElementTestImpl( |
| 245 | armnn::IWorkloadFactory& workloadFactory, |
| 246 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 247 | float qScale, |
| 248 | int32_t qOffset) |
| 249 | { |
| 250 | armnn::TensorInfo inputTensorInfo1 = armnn::TensorInfo({1, 3, 2, 3}, ArmnnType); |
| 251 | armnn::TensorInfo inputTensorInfo2 = armnn::TensorInfo({1, 1, 1, 1}, ArmnnType); |
| 252 | armnn::TensorInfo outputTensorInfo = armnn::TensorInfo({1, 3, 2, 3}, ArmnnType); |
| 253 | |
| 254 | if (armnn::IsQuantizedType<T>()) |
| 255 | { |
| 256 | inputTensorInfo1.SetQuantizationScale(qScale); |
| 257 | inputTensorInfo1.SetQuantizationOffset(qOffset); |
| 258 | inputTensorInfo2.SetQuantizationScale(qScale); |
| 259 | inputTensorInfo2.SetQuantizationOffset(qOffset); |
| 260 | outputTensorInfo.SetQuantizationScale(qScale); |
| 261 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 262 | } |
| 263 | |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 264 | auto input1 = MakeTensor<T, 4>(inputTensorInfo1, armnnUtils::QuantizedVector<T>( |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 265 | { |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 266 | 0.0f, 1.0f, 2.0f, |
| 267 | 3.0f, 4.0f, 5.0f, |
| 268 | 6.0f, 7.0f, 8.0f, |
| 269 | 9.0f, 10.0f, 11.0f, |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 270 | 12.0f, 13.0f, 14.0f, |
| 271 | 15.0f, 16.0f, 17.0f, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 272 | }, |
| 273 | qScale, qOffset)); |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 274 | |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 275 | auto input2 = MakeTensor<T, 4>(inputTensorInfo2, armnnUtils::QuantizedVector<T>( |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 276 | { |
| 277 | 0.5f, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 278 | }, |
| 279 | qScale, qOffset)); |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 280 | |
| 281 | LayerTestResult<T,4> ret(outputTensorInfo); |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 282 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, armnnUtils::QuantizedVector<T>( |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 283 | { |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 284 | 0.5f, 1.5f, 2.5f, |
| 285 | 3.5f, 4.5f, 5.5f, |
| 286 | 6.5f, 7.5f, 8.5f, |
| 287 | 9.5f, 10.5f, 11.5f, |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 288 | 12.5f, 13.5f, 14.5f, |
| 289 | 15.5f, 16.5f, 17.5f, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 290 | }, |
| 291 | qScale, qOffset)); |
Aron Virginas-Tar | e89ebad | 2019-08-27 18:14:26 +0100 | [diff] [blame] | 292 | |
| 293 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 294 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 295 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 296 | |
| 297 | armnn::AdditionQueueDescriptor data; |
| 298 | armnn::WorkloadInfo info; |
| 299 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 300 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
| 301 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 302 | |
| 303 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info); |
| 304 | |
| 305 | inputHandle1->Allocate(); |
| 306 | inputHandle2->Allocate(); |
| 307 | outputHandle->Allocate(); |
| 308 | |
| 309 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 310 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]); |
| 311 | |
| 312 | workload->PostAllocationConfigure(); |
| 313 | workload->Execute(); |
| 314 | |
| 315 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 316 | |
| 317 | return ret; |
| 318 | } |
| 319 | |
| 320 | LayerTestResult<float, 4> AdditionBroadcastTest( |
| 321 | armnn::IWorkloadFactory& workloadFactory, |
| 322 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 323 | { |
| 324 | return AdditionBroadcastTestImpl<armnn::DataType::Float32>( |
| 325 | workloadFactory, memoryManager, 0.0f, 0); |
| 326 | } |
| 327 | |
| 328 | LayerTestResult<uint8_t, 4> AdditionBroadcastUint8Test( |
| 329 | armnn::IWorkloadFactory& workloadFactory, |
| 330 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 331 | { |
| 332 | return AdditionBroadcastTestImpl<armnn::DataType::QuantisedAsymm8>( |
| 333 | workloadFactory, memoryManager, 2.f, 0); |
| 334 | } |
| 335 | |
| 336 | LayerTestResult<int16_t, 4> AdditionBroadcastInt16Test( |
| 337 | armnn::IWorkloadFactory& workloadFactory, |
| 338 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 339 | { |
| 340 | return AdditionBroadcastTestImpl<armnn::DataType::QuantisedSymm16>( |
| 341 | workloadFactory, memoryManager, 2.f, 0); |
| 342 | } |
| 343 | |
| 344 | LayerTestResult<float, 4> AdditionBroadcast1ElementTest( |
| 345 | armnn::IWorkloadFactory& workloadFactory, |
| 346 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 347 | { |
| 348 | return AdditionBroadcast1ElementTestImpl<armnn::DataType::Float32>( |
| 349 | workloadFactory, memoryManager, 0.0f, 0); |
| 350 | } |
| 351 | |
| 352 | LayerTestResult<uint8_t, 4> AdditionBroadcast1ElementUint8Test( |
| 353 | armnn::IWorkloadFactory& workloadFactory, |
| 354 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 355 | { |
| 356 | return AdditionBroadcast1ElementTestImpl<armnn::DataType::QuantisedAsymm8>( |
| 357 | workloadFactory, memoryManager, 0.1333333f, 128); |
| 358 | } |
| 359 | |
| 360 | LayerTestResult<int16_t, 4> AdditionBroadcast1ElementInt16Test( |
| 361 | armnn::IWorkloadFactory& workloadFactory, |
| 362 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 363 | { |
| 364 | return AdditionBroadcast1ElementTestImpl<armnn::DataType::QuantisedSymm16>( |
| 365 | workloadFactory, memoryManager, 0.1333333f, 0); |
| 366 | } |
| 367 | |
| 368 | LayerTestResult<uint8_t, 4> AdditionUint8Test( |
| 369 | armnn::IWorkloadFactory& workloadFactory, |
| 370 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 371 | { |
| 372 | const unsigned int shape0[] = { 1, 2, 2, 3 }; |
| 373 | const unsigned int shape1[] = { 1, 2, 2, 3 }; |
| 374 | |
| 375 | std::vector<uint8_t> input0( |
| 376 | { |
| 377 | 63, 35, 77, 70, 56, 112, // 420, 224, 518, 469, 371, 763 |
| 378 | 203, 28, 252, 168, 245, 91 // 1400, 175, 1743, 1155, 1694, 616 |
| 379 | }); |
| 380 | |
| 381 | std::vector<uint8_t> input1( |
| 382 | { |
| 383 | 21, 7, 175, 231, 175, 210, // 126, 28, 1204, 1596, 1204, 1449 |
| 384 | 126, 161, 63, 21, 105, 126 // 861, 1106, 420, 126, 714, 861 |
| 385 | }); |
| 386 | |
| 387 | std::vector<uint8_t> output( |
| 388 | { |
| 389 | 81, 39, 249, 255, 228, 255, // 546, 252, 1722, 2065(clamped), 1575, 2212(clamped) |
| 390 | 255, 186, 255, 186, 255, 214, // 2261(clamped), 1281, 2163(clamped), 1281, 2408(clamped), 1477 |
| 391 | }); |
| 392 | |
| 393 | return ElementwiseTestHelper<4, armnn::AdditionQueueDescriptor, armnn::DataType::QuantisedAsymm8>( |
| 394 | workloadFactory, |
| 395 | memoryManager, |
| 396 | shape0, |
| 397 | input0, |
| 398 | 7.0f, |
| 399 | 3, |
| 400 | shape1, |
| 401 | input1, |
| 402 | 7.0f, |
| 403 | 3, |
| 404 | shape0, |
| 405 | output, |
| 406 | 7.0f, |
| 407 | 3); |
| 408 | } |
| 409 | |
| 410 | LayerTestResult<int16_t, 4> AdditionInt16Test( |
| 411 | armnn::IWorkloadFactory& workloadFactory, |
| 412 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 413 | { |
| 414 | const unsigned int shape0[] = { 1, 2, 2, 3 }; |
| 415 | const unsigned int shape1[] = { 1, 2, 2, 3 }; |
| 416 | |
| 417 | std::vector<int16_t> input0 = |
| 418 | { |
| 419 | 63, 35, 77, 70, 56, 112, // 441, 245, 539, 490, 392, 184 |
| 420 | 203, 28, 252, 168, 245, 91 // 1421, 196, 1764, 1176, 1715, 637 |
| 421 | }; |
| 422 | |
| 423 | std::vector<int16_t> input1 = |
| 424 | { |
| 425 | 21, 7, 175, 231, 175, 210, // 126, 28, 1204, 1596, 1204, 1449 |
| 426 | 126, 161, 63, 21, 105, 126 // 861, 1106, 420, 126, 714, 861 |
| 427 | }; |
| 428 | |
| 429 | std::vector<int16_t> output = |
| 430 | { |
| 431 | 84, 42, 252, 301, 231, 322, // 588, 294, 1764, 2107(clamped), 1617, 2254(clamped) |
| 432 | 329, 189, 315, 189, 350, 217, // 2303(clamped), 1323, 2205(clamped), 1323, 2450(clamped), 1519 |
| 433 | }; |
| 434 | |
| 435 | return ElementwiseTestHelper<4, armnn::AdditionQueueDescriptor, armnn::DataType::QuantisedSymm16>( |
| 436 | workloadFactory, |
| 437 | memoryManager, |
| 438 | shape0, |
| 439 | input0, |
| 440 | 7.0f, |
| 441 | 0, |
| 442 | shape1, |
| 443 | input1, |
| 444 | 7.0f, |
| 445 | 0, |
| 446 | shape0, |
| 447 | output, |
| 448 | 7.0f, |
| 449 | 0); |
| 450 | } |
| 451 | |
| 452 | LayerTestResult<float, 4> AdditionAfterMaxPoolTest( |
| 453 | armnn::IWorkloadFactory& workloadFactory, |
| 454 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 455 | { |
| 456 | // Create Initial Tensor |
| 457 | // 1, 2, 3 |
| 458 | // 4, 5, 6 |
| 459 | // 7, 8, 9 |
| 460 | |
| 461 | armnn::TensorInfo poolingInputTensorInfo({ 1, 1, 3, 3}, armnn::DataType::Float32); |
| 462 | armnn::TensorInfo poolingOutputTensorInfo({ 1, 1, 2, 2}, armnn::DataType::Float32); |
| 463 | |
| 464 | boost::multi_array<float, 4> poolingInput = MakeTensor<float,4>(poolingInputTensorInfo, |
| 465 | {1, 2, 3, |
| 466 | 4, 5, 6, |
| 467 | 7, 8, 9 |
| 468 | }); |
| 469 | |
| 470 | std::unique_ptr<armnn::ITensorHandle> poolingInputHandle = |
| 471 | workloadFactory.CreateTensorHandle(poolingInputTensorInfo); |
| 472 | std::unique_ptr<armnn::ITensorHandle> poolingOutputHandle = |
| 473 | workloadFactory.CreateTensorHandle(poolingOutputTensorInfo); |
| 474 | |
| 475 | // Apply MaxPool poolSize = 1x1, stride=2x2 |
| 476 | // Result = |
| 477 | // 1, 3 |
| 478 | // 7, 9 |
| 479 | armnn::Pooling2dDescriptor descriptor; |
| 480 | descriptor.m_PoolHeight = 1; |
| 481 | descriptor.m_PoolWidth = 1; |
| 482 | descriptor.m_StrideX = 2; |
| 483 | descriptor.m_StrideY = 2; |
| 484 | descriptor.m_PoolType = armnn::PoolingAlgorithm::Max; |
| 485 | |
| 486 | armnn::Pooling2dQueueDescriptor queueDescriptor; |
| 487 | queueDescriptor.m_Parameters = descriptor; |
| 488 | armnn::WorkloadInfo workloadInfo; |
| 489 | AddInputToWorkload(queueDescriptor, workloadInfo, poolingInputTensorInfo, poolingInputHandle.get()); |
| 490 | AddOutputToWorkload(queueDescriptor, workloadInfo, poolingOutputTensorInfo, poolingOutputHandle.get()); |
| 491 | |
| 492 | // Create the MaxPool |
| 493 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreatePooling2d(queueDescriptor, workloadInfo); |
| 494 | |
| 495 | //LayerTestResult<float, 4> result(poolingOutputTensorInfo); |
| 496 | auto shape( GetTensorShapeAsArray<4>(poolingOutputTensorInfo)); |
| 497 | boost::multi_array<float, 4> resultMaxPool; |
| 498 | resultMaxPool.resize(shape); |
| 499 | |
| 500 | |
| 501 | // Create addition with another tensor the same size |
| 502 | // This would be the result to apply a Conv2d with kernel ones(2) and stride 1x1 |
| 503 | // with the initial tensor. |
| 504 | // 12, 16 |
| 505 | // 24, 28 |
| 506 | |
| 507 | armnn::TensorInfo addInputTensorInfo({ 1,1,2,2}, armnn::DataType::Float32); |
| 508 | armnn::TensorInfo addOutputTensorInfo({ 1,1,2,2}, armnn::DataType::Float32); |
| 509 | |
| 510 | boost::multi_array<float, 4> addInput = MakeTensor<float,4>(addInputTensorInfo, |
| 511 | {12, 16, |
| 512 | 24, 28, |
| 513 | }); |
| 514 | |
| 515 | // Expected output tensor after MaxPool and Addition. |
| 516 | LayerTestResult<float,4> addRet(addOutputTensorInfo); |
| 517 | addRet.outputExpected = MakeTensor<float, 4>(addOutputTensorInfo, std::vector<float>( |
| 518 | { |
| 519 | 13, 19, |
| 520 | 31, 37 |
| 521 | })); |
| 522 | |
| 523 | std::unique_ptr<armnn::ITensorHandle> addInputHandle = workloadFactory.CreateTensorHandle(addInputTensorInfo); |
| 524 | std::unique_ptr<armnn::ITensorHandle> addOutputHandle = workloadFactory.CreateTensorHandle(addOutputTensorInfo); |
| 525 | |
| 526 | armnn::AdditionQueueDescriptor data; |
| 527 | armnn::WorkloadInfo info; |
| 528 | |
| 529 | // Add the output of the MaxPool and the new tensor |
| 530 | AddInputToWorkload(data, info, poolingOutputTensorInfo, poolingOutputHandle.get()); |
| 531 | AddInputToWorkload(data, info, addInputTensorInfo, addInputHandle.get()); |
| 532 | AddOutputToWorkload(data, info, addOutputTensorInfo, addOutputHandle.get()); |
| 533 | |
| 534 | std::unique_ptr<armnn::IWorkload> addWorkload = workloadFactory.CreateAddition(data, info); |
| 535 | |
| 536 | poolingInputHandle->Allocate(); |
| 537 | poolingOutputHandle->Allocate(); |
| 538 | addInputHandle->Allocate(); |
| 539 | addOutputHandle->Allocate(); |
| 540 | |
| 541 | CopyDataToITensorHandle(poolingInputHandle.get(), &poolingInput[0][0][0][0]); |
| 542 | CopyDataFromITensorHandle(&resultMaxPool[0][0][0][0], poolingOutputHandle.get()); |
| 543 | |
| 544 | CopyDataToITensorHandle(poolingOutputHandle.get(), &resultMaxPool[0][0][0][0]); |
| 545 | CopyDataToITensorHandle(addInputHandle.get(), &addInput[0][0][0][0]); |
| 546 | |
| 547 | workload->PostAllocationConfigure(); |
| 548 | workload->Execute(); |
| 549 | addWorkload->PostAllocationConfigure(); |
| 550 | addWorkload->Execute(); |
| 551 | |
| 552 | CopyDataFromITensorHandle(&addRet.output[0][0][0][0], addOutputHandle.get()); |
| 553 | |
| 554 | return addRet; |
| 555 | } |
| 556 | |
| 557 | LayerTestResult<float,4> CompareAdditionTest( |
| 558 | armnn::IWorkloadFactory& workloadFactory, |
| 559 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 560 | armnn::IWorkloadFactory& refWorkloadFactory) |
| 561 | { |
| 562 | unsigned int batchSize = 4; |
| 563 | unsigned int channels = 1; |
| 564 | unsigned int height = 2; |
| 565 | unsigned int width = 3; |
| 566 | |
| 567 | armnn::TensorInfo inputTensorInfo1, inputTensorInfo2; |
| 568 | armnn::TensorInfo outputTensorInfo; |
| 569 | |
| 570 | unsigned int shape[] = {batchSize, channels, height, width}; |
| 571 | |
| 572 | inputTensorInfo1 = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 573 | inputTensorInfo2 = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 574 | outputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 575 | |
| 576 | auto input1 = MakeRandomTensor<float, 4>(inputTensorInfo1, 1232); |
| 577 | auto input2 = MakeRandomTensor<float, 4>(inputTensorInfo2, 456); |
| 578 | |
| 579 | LayerTestResult<float,4> ret(outputTensorInfo); |
| 580 | |
| 581 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 582 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 583 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 584 | |
| 585 | std::unique_ptr<armnn::ITensorHandle> inputHandle1Ref = refWorkloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 586 | std::unique_ptr<armnn::ITensorHandle> inputHandle2Ref = refWorkloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 587 | std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo); |
| 588 | |
| 589 | armnn::AdditionQueueDescriptor data; |
| 590 | armnn::WorkloadInfo info; |
| 591 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 592 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
| 593 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 594 | |
| 595 | armnn::AdditionQueueDescriptor refData = data; |
| 596 | armnn::WorkloadInfo refInfo = info; |
| 597 | SetWorkloadInput(refData, refInfo, 0, inputTensorInfo1, inputHandle1Ref.get()); |
| 598 | SetWorkloadInput(refData, refInfo, 1, inputTensorInfo2, inputHandle2Ref.get()); |
| 599 | SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); |
| 600 | |
| 601 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info); |
| 602 | std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateAddition(refData, refInfo); |
| 603 | |
| 604 | inputHandle1->Allocate(); |
| 605 | inputHandle2->Allocate(); |
| 606 | outputHandle->Allocate(); |
| 607 | inputHandle1Ref->Allocate(); |
| 608 | inputHandle2Ref->Allocate(); |
| 609 | outputHandleRef->Allocate(); |
| 610 | |
| 611 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 612 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]); |
| 613 | CopyDataToITensorHandle(inputHandle1Ref.get(), &input1[0][0][0][0]); |
| 614 | CopyDataToITensorHandle(inputHandle2Ref.get(), &input2[0][0][0][0]); |
| 615 | |
| 616 | workload->PostAllocationConfigure(); |
| 617 | workload->Execute(); |
| 618 | workloadRef->PostAllocationConfigure(); |
| 619 | workloadRef->Execute(); |
| 620 | |
| 621 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 622 | CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get()); |
| 623 | |
| 624 | return ret; |
| 625 | } |