Aron Virginas-Tar | 00d306e | 2019-08-28 18:08:46 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "BatchNormalizationTestImpl.hpp" |
| 7 | |
| 8 | #include <DataLayoutIndexed.hpp> |
| 9 | #include <ResolveType.hpp> |
| 10 | |
| 11 | #include <armnn/ArmNN.hpp> |
| 12 | |
| 13 | #include <backendsCommon/CpuTensorHandle.hpp> |
| 14 | #include <backendsCommon/IBackendInternal.hpp> |
| 15 | #include <backendsCommon/WorkloadFactory.hpp> |
| 16 | |
| 17 | #include <backendsCommon/test/QuantizeHelper.hpp> |
| 18 | #include <backendsCommon/test/TensorCopyUtils.hpp> |
| 19 | #include <backendsCommon/test/WorkloadTestUtils.hpp> |
| 20 | |
| 21 | #include <test/TensorHelpers.hpp> |
| 22 | |
| 23 | namespace |
| 24 | { |
| 25 | |
| 26 | template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> |
| 27 | LayerTestResult<T, 4> BatchNormTestImpl( |
| 28 | armnn::IWorkloadFactory& workloadFactory, |
| 29 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 30 | const armnn::TensorShape& inputOutputTensorShape, |
| 31 | const std::vector<float>& inputValues, |
| 32 | const std::vector<float>& expectedOutputValues, |
| 33 | float qScale, |
| 34 | int32_t qOffset, |
| 35 | armnn::DataLayout dataLayout) |
| 36 | { |
| 37 | armnn::TensorInfo inputTensorInfo(inputOutputTensorShape, ArmnnType); |
| 38 | armnn::TensorInfo outputTensorInfo(inputOutputTensorShape, ArmnnType); |
| 39 | |
| 40 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(dataLayout); |
| 41 | |
| 42 | armnn::TensorInfo tensorInfo({ inputOutputTensorShape[dataLayoutIndexed.GetChannelsIndex()] }, |
| 43 | ArmnnType); |
| 44 | |
| 45 | // Set quantization parameters if the requested type is a quantized type. |
| 46 | if (armnn::IsQuantizedType<T>()) |
| 47 | { |
| 48 | inputTensorInfo.SetQuantizationScale(qScale); |
| 49 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 50 | outputTensorInfo.SetQuantizationScale(qScale); |
| 51 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 52 | tensorInfo.SetQuantizationScale(qScale); |
| 53 | tensorInfo.SetQuantizationOffset(qOffset); |
| 54 | } |
| 55 | |
| 56 | auto inputTensor = MakeTensor<T, 4>(inputTensorInfo, |
| 57 | QuantizedVector<T>(qScale, qOffset, inputValues)); |
| 58 | |
| 59 | // These values are per-channel of the input. |
| 60 | auto mean = MakeTensor<T, 1>(tensorInfo, QuantizedVector<T>(qScale, qOffset, {3, -2})); |
| 61 | auto variance = MakeTensor<T, 1>(tensorInfo, QuantizedVector<T>(qScale, qOffset, {4, 9})); |
| 62 | auto beta = MakeTensor<T, 1>(tensorInfo, QuantizedVector<T>(qScale, qOffset, {3, 2})); |
| 63 | auto gamma = MakeTensor<T, 1>(tensorInfo, QuantizedVector<T>(qScale, qOffset, {2, 1})); |
| 64 | |
| 65 | LayerTestResult<T, 4> result(outputTensorInfo); |
| 66 | |
| 67 | result.outputExpected = MakeTensor<T, 4>(inputTensorInfo, |
| 68 | QuantizedVector<T>(qScale, qOffset, expectedOutputValues)); |
| 69 | |
| 70 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 71 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 72 | |
| 73 | armnn::ScopedCpuTensorHandle meanTensor(tensorInfo); |
| 74 | armnn::ScopedCpuTensorHandle varianceTensor(tensorInfo); |
| 75 | armnn::ScopedCpuTensorHandle betaTensor(tensorInfo); |
| 76 | armnn::ScopedCpuTensorHandle gammaTensor(tensorInfo); |
| 77 | |
| 78 | armnn::BatchNormalizationQueueDescriptor descriptor; |
| 79 | descriptor.m_Mean = &meanTensor; |
| 80 | descriptor.m_Variance = &varianceTensor; |
| 81 | descriptor.m_Beta = &betaTensor; |
| 82 | descriptor.m_Gamma = &gammaTensor; |
| 83 | descriptor.m_Parameters.m_Eps = 0.0f; |
| 84 | descriptor.m_Parameters.m_DataLayout = dataLayout; |
| 85 | armnn::WorkloadInfo info; |
| 86 | |
| 87 | AllocateAndCopyDataToITensorHandle(&meanTensor, &mean[0]); |
| 88 | AllocateAndCopyDataToITensorHandle(&varianceTensor, &variance[0]); |
| 89 | AllocateAndCopyDataToITensorHandle(&betaTensor, &beta[0]); |
| 90 | AllocateAndCopyDataToITensorHandle(&gammaTensor, &gamma[0]); |
| 91 | |
| 92 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 93 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 94 | |
| 95 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateBatchNormalization(descriptor, info); |
| 96 | |
| 97 | inputHandle->Allocate(); |
| 98 | outputHandle->Allocate(); |
| 99 | |
| 100 | CopyDataToITensorHandle(inputHandle.get(), &inputTensor[0][0][0][0]); |
| 101 | |
| 102 | workload->Execute(); |
| 103 | |
| 104 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 105 | |
| 106 | return result; |
| 107 | } |
| 108 | |
| 109 | template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> |
| 110 | LayerTestResult<T,4> BatchNormTestNhwcImpl( |
| 111 | armnn::IWorkloadFactory& workloadFactory, |
| 112 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 113 | float qScale, |
| 114 | int32_t qOffset) |
| 115 | { |
| 116 | const unsigned int width = 2; |
| 117 | const unsigned int height = 3; |
| 118 | const unsigned int channels = 2; |
| 119 | const unsigned int num = 1; |
| 120 | |
| 121 | armnn::TensorInfo inputTensorInfo({num, height, width, channels}, ArmnnType); |
| 122 | armnn::TensorInfo outputTensorInfo({num, height, width, channels}, ArmnnType); |
| 123 | armnn::TensorInfo tensorInfo({channels}, ArmnnType); |
| 124 | |
| 125 | // Set quantization parameters if the requested type is a quantized type. |
| 126 | if(armnn::IsQuantizedType<T>()) |
| 127 | { |
| 128 | inputTensorInfo.SetQuantizationScale(qScale); |
| 129 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 130 | outputTensorInfo.SetQuantizationScale(qScale); |
| 131 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 132 | tensorInfo.SetQuantizationScale(qScale); |
| 133 | tensorInfo.SetQuantizationOffset(qOffset); |
| 134 | } |
| 135 | |
| 136 | auto input = MakeTensor<T, 4>(inputTensorInfo, |
| 137 | QuantizedVector<T>(qScale, qOffset, |
| 138 | { |
| 139 | 1.f, 1.f, 4.f, 1.f, |
| 140 | 4.f, 4.f, 2.f, 1.f, |
| 141 | 1.f, -2.f, 6.f, 4.f |
| 142 | })); |
| 143 | // These values are per-channel of the input. |
| 144 | auto mean = MakeTensor<T, 1>(tensorInfo, QuantizedVector<T>(qScale, qOffset, {3, -2})); |
| 145 | auto variance = MakeTensor<T, 1>(tensorInfo, QuantizedVector<T>(qScale, qOffset, {4, 9})); |
| 146 | auto beta = MakeTensor<T, 1>(tensorInfo, QuantizedVector<T>(qScale, qOffset, {3, 2})); |
| 147 | auto gamma = MakeTensor<T, 1>(tensorInfo, QuantizedVector<T>(qScale, qOffset, {2, 1})); |
| 148 | LayerTestResult<T,4> ret(outputTensorInfo); |
| 149 | |
| 150 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 151 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 152 | |
| 153 | armnn::BatchNormalizationQueueDescriptor data; |
| 154 | armnn::WorkloadInfo info; |
| 155 | armnn::ScopedCpuTensorHandle meanTensor(tensorInfo); |
| 156 | armnn::ScopedCpuTensorHandle varianceTensor(tensorInfo); |
| 157 | armnn::ScopedCpuTensorHandle betaTensor(tensorInfo); |
| 158 | armnn::ScopedCpuTensorHandle gammaTensor(tensorInfo); |
| 159 | |
| 160 | AllocateAndCopyDataToITensorHandle(&meanTensor, &mean[0]); |
| 161 | AllocateAndCopyDataToITensorHandle(&varianceTensor, &variance[0]); |
| 162 | AllocateAndCopyDataToITensorHandle(&betaTensor, &beta[0]); |
| 163 | AllocateAndCopyDataToITensorHandle(&gammaTensor, &gamma[0]); |
| 164 | |
| 165 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 166 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 167 | data.m_Mean = &meanTensor; |
| 168 | data.m_Variance = &varianceTensor; |
| 169 | data.m_Beta = &betaTensor; |
| 170 | data.m_Gamma = &gammaTensor; |
| 171 | data.m_Parameters.m_Eps = 0.0f; |
| 172 | data.m_Parameters.m_DataLayout = armnn::DataLayout::NHWC; |
| 173 | |
| 174 | // For each channel: |
| 175 | // substract mean, divide by standard deviation (with an epsilon to avoid div by 0), |
| 176 | // multiply by gamma and add beta |
| 177 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, |
| 178 | QuantizedVector<T>(qScale, qOffset, |
| 179 | { |
| 180 | 1.f, 3.f, 4.f, 3.f, |
| 181 | 4.f, 4.f, 2.f, 3.f, |
| 182 | 1.f, 2.f, 6.f, 4.f |
| 183 | })); |
| 184 | |
| 185 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateBatchNormalization(data, info); |
| 186 | |
| 187 | inputHandle->Allocate(); |
| 188 | outputHandle->Allocate(); |
| 189 | |
| 190 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 191 | |
| 192 | workload->Execute(); |
| 193 | |
| 194 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 195 | |
| 196 | return ret; |
| 197 | } |
| 198 | |
| 199 | } // anonymous namespace |
| 200 | |
| 201 | LayerTestResult<float, 4> BatchNormFloatTest( |
| 202 | armnn::IWorkloadFactory& workloadFactory, |
| 203 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 204 | { |
| 205 | // BatchSize: 1 |
| 206 | // Channels: 2 |
| 207 | // Height: 3 |
| 208 | // Width: 2 |
| 209 | |
| 210 | const armnn::TensorShape inputOutputShape{ 1, 2, 3, 2 }; |
| 211 | std::vector<float> inputValues |
| 212 | { |
| 213 | // Batch 0, Channel 0, Height (3) x Width (2) |
| 214 | 1.f, 4.f, |
| 215 | 4.f, 2.f, |
| 216 | 1.f, 6.f, |
| 217 | |
| 218 | // Batch 0, Channel 1, Height (3) x Width (2) |
| 219 | 1.f, 1.f, |
| 220 | 4.f, 1.f, |
| 221 | -2.f, 4.f |
| 222 | }; |
| 223 | std::vector<float> expectedOutputValues |
| 224 | { |
| 225 | // Batch 0, Channel 0, Height (3) x Width (2) |
| 226 | 1.f, 4.f, |
| 227 | 4.f, 2.f, |
| 228 | 1.f, 6.f, |
| 229 | |
| 230 | // Batch 0, Channel 1, Height (3) x Width (2) |
| 231 | 3.f, 3.f, |
| 232 | 4.f, 3.f, |
| 233 | 2.f, 4.f |
| 234 | }; |
| 235 | |
| 236 | return BatchNormTestImpl<armnn::DataType::Float32>( |
| 237 | workloadFactory, |
| 238 | memoryManager, |
| 239 | inputOutputShape, |
| 240 | inputValues, |
| 241 | expectedOutputValues, |
| 242 | 0.f, |
| 243 | 0, |
| 244 | armnn::DataLayout::NCHW); |
| 245 | } |
| 246 | |
| 247 | LayerTestResult<float, 4> BatchNormFloatNhwcTest( |
| 248 | armnn::IWorkloadFactory& workloadFactory, |
| 249 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 250 | { |
| 251 | // BatchSize: 1 |
| 252 | // Height: 3 |
| 253 | // Width: 2 |
| 254 | // Channels: 2 |
| 255 | |
| 256 | const armnn::TensorShape inputOutputShape{ 1, 3, 2, 2 }; |
| 257 | std::vector<float> inputValues |
| 258 | { |
| 259 | // Batch 0, Height 0, Width (2) x Channel (2) |
| 260 | 1.f, 1.f, |
| 261 | 4.f, 1.f, |
| 262 | |
| 263 | // Batch 0, Height 1, Width (2) x Channel (2) |
| 264 | 4.f, 4.f, |
| 265 | 2.f, 1.f, |
| 266 | |
| 267 | // Batch 0, Height 2, Width (2) x Channel (2) |
| 268 | 1.f, -2.f, |
| 269 | 6.f, 4.f |
| 270 | }; |
| 271 | std::vector<float> expectedOutputValues |
| 272 | { |
| 273 | // Batch 0, Height 0, Width (2) x Channel (2) |
| 274 | 1.f, 3.f, |
| 275 | 4.f, 3.f, |
| 276 | |
| 277 | // Batch 0, Height 1, Width (2) x Channel (2) |
| 278 | 4.f, 4.f, |
| 279 | 2.f, 3.f, |
| 280 | |
| 281 | // Batch 0, Height 2, Width (2) x Channel (2) |
| 282 | 1.f, 2.f, |
| 283 | 6.f, 4.f |
| 284 | }; |
| 285 | |
| 286 | return BatchNormTestImpl<armnn::DataType::Float32>( |
| 287 | workloadFactory, |
| 288 | memoryManager, |
| 289 | inputOutputShape, |
| 290 | inputValues, |
| 291 | expectedOutputValues, |
| 292 | 0.f, |
| 293 | 0, |
| 294 | armnn::DataLayout::NHWC); |
| 295 | } |
| 296 | |
| 297 | LayerTestResult<uint8_t, 4> BatchNormUint8Test( |
| 298 | armnn::IWorkloadFactory& workloadFactory, |
| 299 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 300 | { |
| 301 | // BatchSize: 1 |
| 302 | // Channels: 2 |
| 303 | // Height: 3 |
| 304 | // Width: 2 |
| 305 | |
| 306 | const armnn::TensorShape inputOutputShape{ 1, 2, 3, 2 }; |
| 307 | std::vector<float> inputValues |
| 308 | { |
| 309 | // Batch 0, Channel 0, Height (3) x Width (2) |
| 310 | 1.f, 4.f, |
| 311 | 4.f, 2.f, |
| 312 | 1.f, 6.f, |
| 313 | |
| 314 | // Batch 0, Channel 1, Height (3) x Width (2) |
| 315 | 1.f, 1.f, |
| 316 | 4.f, 1.f, |
| 317 | -2.f, 4.f |
| 318 | }; |
| 319 | std::vector<float> expectedOutputValues |
| 320 | { |
| 321 | // Batch 0, Channel 0, Height (3) x Width (2) |
| 322 | 1.f, 4.f, |
| 323 | 4.f, 2.f, |
| 324 | 1.f, 6.f, |
| 325 | |
| 326 | // Batch 0, Channel 1, Height (3) x Width (2) |
| 327 | 3.f, 3.f, |
| 328 | 4.f, 3.f, |
| 329 | 2.f, 4.f |
| 330 | }; |
| 331 | |
| 332 | return BatchNormTestImpl<armnn::DataType::QuantisedAsymm8>( |
| 333 | workloadFactory, |
| 334 | memoryManager, |
| 335 | inputOutputShape, |
| 336 | inputValues, |
| 337 | expectedOutputValues, |
| 338 | 1.f / 20.f, |
| 339 | 50, |
| 340 | armnn::DataLayout::NCHW); |
| 341 | } |
| 342 | |
| 343 | LayerTestResult<uint8_t, 4> BatchNormUint8NhwcTest( |
| 344 | armnn::IWorkloadFactory& workloadFactory, |
| 345 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 346 | { |
| 347 | // BatchSize: 1 |
| 348 | // Height: 3 |
| 349 | // Width: 2 |
| 350 | // Channels: 2 |
| 351 | |
| 352 | const armnn::TensorShape inputOutputShape{ 1, 3, 2, 2 }; |
| 353 | std::vector<float> inputValues |
| 354 | { |
| 355 | // Batch 0, Height 0, Width (2) x Channel (2) |
| 356 | 1.f, 1.f, |
| 357 | 4.f, 1.f, |
| 358 | |
| 359 | // Batch 0, Height 1, Width (2) x Channel (2) |
| 360 | 4.f, 4.f, |
| 361 | 2.f, 1.f, |
| 362 | |
| 363 | // Batch 0, Height 2, Width (2) x Channel (2) |
| 364 | 1.f, -2.f, |
| 365 | 6.f, 4.f |
| 366 | }; |
| 367 | std::vector<float> expectedOutputValues |
| 368 | { |
| 369 | // Batch 0, Height 0, Width (2) x Channel (2) |
| 370 | 1.f, 3.f, |
| 371 | 4.f, 3.f, |
| 372 | |
| 373 | // Batch 0, Height 1, Width (2) x Channel (2) |
| 374 | 4.f, 4.f, |
| 375 | 2.f, 3.f, |
| 376 | |
| 377 | // Batch 0, Height 2, Width (2) x Channel (2) |
| 378 | 1.f, 2.f, |
| 379 | 6.f, 4.f |
| 380 | }; |
| 381 | |
| 382 | return BatchNormTestImpl<armnn::DataType::QuantisedAsymm8>( |
| 383 | workloadFactory, |
| 384 | memoryManager, |
| 385 | inputOutputShape, inputValues, expectedOutputValues, |
| 386 | 1.f/20.f, 50, armnn::DataLayout::NHWC); |
| 387 | } |
| 388 | |
| 389 | LayerTestResult<int16_t, 4> BatchNormInt16Test( |
| 390 | armnn::IWorkloadFactory& workloadFactory, |
| 391 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 392 | { |
| 393 | // BatchSize: 1 |
| 394 | // Channels: 2 |
| 395 | // Height: 3 |
| 396 | // Width: 2 |
| 397 | |
| 398 | const armnn::TensorShape inputOutputShape{ 1, 2, 3, 2 }; |
| 399 | std::vector<float> inputValues |
| 400 | { |
| 401 | // Batch 0, Channel 0, Height (3) x Width (2) |
| 402 | 1.f, 4.f, |
| 403 | 4.f, 2.f, |
| 404 | 1.f, 6.f, |
| 405 | |
| 406 | // Batch 0, Channel 1, Height (3) x Width (2) |
| 407 | 1.f, 1.f, |
| 408 | 4.f, 1.f, |
| 409 | -2.f, 4.f |
| 410 | }; |
| 411 | std::vector<float> expectedOutputValues |
| 412 | { |
| 413 | // Batch 0, Channel 0, Height (3) x Width (2) |
| 414 | 1.f, 4.f, |
| 415 | 4.f, 2.f, |
| 416 | 1.f, 6.f, |
| 417 | |
| 418 | // Batch 0, Channel 1, Height (3) x Width (2) |
| 419 | 3.f, 3.f, |
| 420 | 4.f, 3.f, |
| 421 | 2.f, 4.f |
| 422 | }; |
| 423 | |
| 424 | return BatchNormTestImpl<armnn::DataType::QuantisedSymm16>( |
| 425 | workloadFactory, |
| 426 | memoryManager, |
| 427 | inputOutputShape, |
| 428 | inputValues, |
| 429 | expectedOutputValues, |
| 430 | 1.f / 20.f, |
| 431 | 50, |
| 432 | armnn::DataLayout::NCHW); |
| 433 | } |
| 434 | |
| 435 | LayerTestResult<int16_t, 4> BatchNormInt16NhwcTest( |
| 436 | armnn::IWorkloadFactory& workloadFactory, |
| 437 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager) |
| 438 | { |
| 439 | // BatchSize: 1 |
| 440 | // Height: 3 |
| 441 | // Width: 2 |
| 442 | // Channels: 2 |
| 443 | |
| 444 | const armnn::TensorShape inputOutputShape{ 1, 3, 2, 2 }; |
| 445 | std::vector<float> inputValues |
| 446 | { |
| 447 | // Batch 0, Height 0, Width (2) x Channel (2) |
| 448 | 1.f, 1.f, |
| 449 | 4.f, 1.f, |
| 450 | |
| 451 | // Batch 0, Height 1, Width (2) x Channel (2) |
| 452 | 4.f, 4.f, |
| 453 | 2.f, 1.f, |
| 454 | |
| 455 | // Batch 0, Height 2, Width (2) x Channel (2) |
| 456 | 1.f, -2.f, |
| 457 | 6.f, 4.f |
| 458 | }; |
| 459 | std::vector<float> expectedOutputValues |
| 460 | { |
| 461 | // Batch 0, Height 0, Width (2) x Channel (2) |
| 462 | 1.f, 3.f, |
| 463 | 4.f, 3.f, |
| 464 | |
| 465 | // Batch 0, Height 1, Width (2) x Channel (2) |
| 466 | 4.f, 4.f, |
| 467 | 2.f, 3.f, |
| 468 | |
| 469 | // Batch 0, Height 2, Width (2) x Channel (2) |
| 470 | 1.f, 2.f, |
| 471 | 6.f, 4.f |
| 472 | }; |
| 473 | |
| 474 | return BatchNormTestImpl<armnn::DataType::QuantisedSymm16>( |
| 475 | workloadFactory, |
| 476 | memoryManager, |
| 477 | inputOutputShape, |
| 478 | inputValues, |
| 479 | expectedOutputValues, |
| 480 | 1.f / 20.f, |
| 481 | 50, |
| 482 | armnn::DataLayout::NHWC); |
| 483 | } |
| 484 | |
| 485 | LayerTestResult<float,4> CompareBatchNormTest( |
| 486 | armnn::IWorkloadFactory& workloadFactory, |
| 487 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 488 | armnn::IWorkloadFactory& refWorkloadFactory) |
| 489 | { |
| 490 | const unsigned int width = 2; |
| 491 | const unsigned int height = 3; |
| 492 | const unsigned int channels = 5; |
| 493 | const unsigned int batchSize = 3; |
| 494 | |
| 495 | armnn::TensorInfo inputTensorInfo; |
| 496 | armnn::TensorInfo outputTensorInfo; |
| 497 | armnn::TensorInfo tensorInfo; |
| 498 | |
| 499 | constexpr unsigned int shape[] = {batchSize, channels, height, width}; |
| 500 | constexpr unsigned int tensorShape[] = {channels}; |
| 501 | |
| 502 | inputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 503 | outputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 504 | tensorInfo = armnn::TensorInfo(1, tensorShape, armnn::DataType::Float32); |
| 505 | |
| 506 | auto input = MakeRandomTensor<float, 4>(inputTensorInfo, 21312); |
| 507 | |
| 508 | auto mean = MakeRandomTensor<float, 1>(tensorInfo, 123); |
| 509 | auto variance = MakeRandomTensor<float, 1>(tensorInfo, 234, 0.0f); |
| 510 | auto beta = MakeRandomTensor<float, 1>(tensorInfo, 123); |
| 511 | auto gamma = MakeRandomTensor<float, 1>(tensorInfo, 345); |
| 512 | |
| 513 | LayerTestResult<float,4> ret(outputTensorInfo); |
| 514 | |
| 515 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 516 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 517 | |
| 518 | std::unique_ptr<armnn::ITensorHandle> inputHandleRef = refWorkloadFactory.CreateTensorHandle(inputTensorInfo); |
| 519 | std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo); |
| 520 | |
| 521 | armnn::BatchNormalizationQueueDescriptor data; |
| 522 | armnn::WorkloadInfo info; |
| 523 | armnn::ScopedCpuTensorHandle meanTensor(tensorInfo); |
| 524 | armnn::ScopedCpuTensorHandle varianceTensor(tensorInfo); |
| 525 | armnn::ScopedCpuTensorHandle betaTensor(tensorInfo); |
| 526 | armnn::ScopedCpuTensorHandle gammaTensor(tensorInfo); |
| 527 | |
| 528 | AllocateAndCopyDataToITensorHandle(&meanTensor, &mean[0]); |
| 529 | AllocateAndCopyDataToITensorHandle(&varianceTensor, &variance[0]); |
| 530 | AllocateAndCopyDataToITensorHandle(&betaTensor, &beta[0]); |
| 531 | AllocateAndCopyDataToITensorHandle(&gammaTensor, &gamma[0]); |
| 532 | |
| 533 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 534 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 535 | data.m_Mean = &meanTensor; |
| 536 | data.m_Variance = &varianceTensor; |
| 537 | data.m_Beta = &betaTensor; |
| 538 | data.m_Gamma = &gammaTensor; |
| 539 | data.m_Parameters.m_Eps = 0.01f; |
| 540 | |
| 541 | armnn::BatchNormalizationQueueDescriptor refData = data; |
| 542 | armnn::WorkloadInfo refInfo = info; |
| 543 | SetWorkloadInput(refData, refInfo, 0, inputTensorInfo, inputHandleRef.get()); |
| 544 | SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); |
| 545 | |
| 546 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateBatchNormalization(data, info); |
| 547 | std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateBatchNormalization(refData, refInfo); |
| 548 | |
| 549 | inputHandle->Allocate(); |
| 550 | outputHandle->Allocate(); |
| 551 | inputHandleRef->Allocate(); |
| 552 | outputHandleRef->Allocate(); |
| 553 | |
| 554 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 555 | CopyDataToITensorHandle(inputHandleRef.get(), &input[0][0][0][0]); |
| 556 | |
| 557 | workload->PostAllocationConfigure(); |
| 558 | workload->Execute(); |
| 559 | workloadRef->PostAllocationConfigure(); |
| 560 | workloadRef->Execute(); |
| 561 | |
| 562 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 563 | CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get()); |
| 564 | |
| 565 | return ret; |
| 566 | } |