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 "FullyConnectedTestImpl.hpp" |
| 7 | |
| 8 | #include <armnn/ArmNN.hpp> |
| 9 | |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 10 | #include <QuantizeHelper.hpp> |
| 11 | |
Aron Virginas-Tar | 00d306e | 2019-08-28 18:08:46 +0100 | [diff] [blame] | 12 | #include <backendsCommon/CpuTensorHandle.hpp> |
| 13 | |
| 14 | #include <backendsCommon/test/DataTypeUtils.hpp> |
| 15 | #include <backendsCommon/test/TensorCopyUtils.hpp> |
| 16 | #include <backendsCommon/test/WorkloadTestUtils.hpp> |
| 17 | |
| 18 | #include <test/TensorHelpers.hpp> |
| 19 | |
| 20 | // |
| 21 | // Implementation templates |
| 22 | // |
| 23 | |
| 24 | template<typename T, typename B> |
| 25 | LayerTestResult<T, 2> SimpleFullyConnectedTestImpl( |
| 26 | armnn::IWorkloadFactory& workloadFactory, |
| 27 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 28 | armnn::TensorInfo inputTensorInfo, |
| 29 | armnn::TensorInfo outputTensorInfo, |
| 30 | armnn::TensorInfo weightsDesc, |
| 31 | armnn::TensorInfo biasesDesc, |
| 32 | boost::multi_array<T, 2>& weights, |
| 33 | boost::multi_array<B, 1>& bias, |
| 34 | boost::multi_array<T, 4>& input, |
| 35 | bool biasEnabled, |
| 36 | bool transposeWeights) |
| 37 | { |
Derek Lamberti | c374ff0 | 2019-12-10 21:57:35 +0000 | [diff] [blame^] | 38 | boost::ignore_unused(memoryManager); |
Aron Virginas-Tar | 00d306e | 2019-08-28 18:08:46 +0100 | [diff] [blame] | 39 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 40 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 41 | |
| 42 | armnn::FullyConnectedQueueDescriptor data; |
| 43 | armnn::WorkloadInfo info; |
| 44 | armnn::ScopedCpuTensorHandle weightsTensor(weightsDesc); |
| 45 | armnn::ScopedCpuTensorHandle biasTensor(biasesDesc); |
| 46 | |
| 47 | AllocateAndCopyDataToITensorHandle(&weightsTensor, &weights[0][0]); |
| 48 | AllocateAndCopyDataToITensorHandle(&biasTensor, &bias[0]); |
| 49 | |
| 50 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 51 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 52 | data.m_Weight = &weightsTensor; |
| 53 | data.m_Bias = &biasTensor; |
| 54 | data.m_Parameters.m_BiasEnabled = biasEnabled; |
| 55 | data.m_Parameters.m_TransposeWeightMatrix = transposeWeights; |
| 56 | |
| 57 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateFullyConnected(data, info); |
| 58 | LayerTestResult<T, 2> result(outputTensorInfo); |
| 59 | |
| 60 | inputHandle->Allocate(); |
| 61 | outputHandle->Allocate(); |
| 62 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 63 | |
| 64 | ExecuteWorkload(*workload, memoryManager); |
| 65 | |
| 66 | CopyDataFromITensorHandle(&result.output[0][0], outputHandle.get()); |
| 67 | |
| 68 | return result; |
| 69 | } |
| 70 | |
| 71 | template<armnn::DataType ArmnnType, typename T> |
| 72 | LayerTestResult<T, 2> FullyConnectedTest( |
| 73 | armnn::IWorkloadFactory& workloadFactory, |
| 74 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 75 | bool biasEnabled) |
| 76 | { |
| 77 | constexpr static unsigned int inputWidth = 3u; |
| 78 | constexpr static unsigned int inputHeight = 2u; |
| 79 | constexpr static unsigned int inputChannels = 1u; |
| 80 | |
| 81 | constexpr static unsigned int inputSize = inputWidth * inputHeight * inputChannels; |
| 82 | |
| 83 | constexpr static unsigned int outputChannels = 2u; |
| 84 | |
| 85 | armnn::TensorInfo inputTensorInfo({ 1, inputChannels, inputHeight, inputWidth }, ArmnnType); |
| 86 | inputTensorInfo.SetQuantizationScale(0.1f); |
| 87 | inputTensorInfo.SetQuantizationOffset(63); |
| 88 | |
| 89 | armnn::TensorInfo outputTensorInfo({ 1, outputChannels }, ArmnnType); |
| 90 | outputTensorInfo.SetQuantizationScale(5.f); |
| 91 | outputTensorInfo.SetQuantizationOffset(biasEnabled ? -50 : 10); |
| 92 | |
| 93 | armnn::TensorInfo weightsDesc({ outputChannels, inputSize }, ArmnnType); |
| 94 | weightsDesc.SetQuantizationScale(0.2f); |
| 95 | weightsDesc.SetQuantizationOffset(93); |
| 96 | |
| 97 | armnn::TensorInfo biasesDesc({ outputChannels }, GetBiasTypeFromWeightsType(weightsDesc.GetDataType()).value()); |
| 98 | biasesDesc.SetQuantizationScale(inputTensorInfo.GetQuantizationScale() * weightsDesc.GetQuantizationScale()); |
| 99 | biasesDesc.SetQuantizationOffset(0); |
| 100 | |
| 101 | LayerTestResult<T, 2> result(outputTensorInfo); |
| 102 | |
| 103 | auto input = MakeTensor<T, 4>(inputTensorInfo, ConvertToDataType<ArmnnType>( |
| 104 | { |
| 105 | -1.2f, 6.1f, -3.5f, |
| 106 | 18.8f, -5.5f, 2.9f |
| 107 | }, |
| 108 | inputTensorInfo)); |
| 109 | |
| 110 | auto weights = MakeTensor<T, 2>(weightsDesc, ConvertToDataType<ArmnnType>( |
| 111 | { |
| 112 | -8.4f, 20.0f, -10.4f, -8, 16.4f, -11.8f, |
| 113 | 23.4f, 10.4f, -14.0f, -3.8f, -11.8f, 11.4f |
| 114 | }, |
| 115 | weightsDesc)); |
| 116 | |
| 117 | auto bias = MakeTensor<int32_t, 1>(biasesDesc, std::vector<int32_t>{9250, 67500}); |
| 118 | |
| 119 | result = SimpleFullyConnectedTestImpl<T>( |
| 120 | workloadFactory, |
| 121 | memoryManager, |
| 122 | inputTensorInfo, outputTensorInfo, |
| 123 | weightsDesc, biasesDesc, |
| 124 | weights, bias, input, |
| 125 | biasEnabled, true |
| 126 | ); |
| 127 | |
| 128 | if (biasEnabled) |
| 129 | { |
| 130 | result.outputExpected = MakeTensor<T, 2>(outputTensorInfo, |
| 131 | ConvertToDataType<ArmnnType>({80.f, 1460.f}, outputTensorInfo)); |
| 132 | } |
| 133 | else |
| 134 | { |
| 135 | result.outputExpected = MakeTensor<T, 2>(outputTensorInfo, |
| 136 | ConvertToDataType<ArmnnType>({-107.04f, 110.f}, outputTensorInfo)); |
| 137 | } |
| 138 | |
| 139 | return result; |
| 140 | } |
| 141 | |
| 142 | // |
| 143 | // ArmNN variant of the AndroidNN fully_connected_float_large test. |
| 144 | // |
| 145 | // Tests the fully connected layer with large values, optionally transposing weights. |
| 146 | // Note this is templated for consistency, but the nature of this tests makes it unlikely to be useful in Uint8 mode. |
| 147 | // |
| 148 | template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> |
| 149 | LayerTestResult<T, 2> FullyConnectedLargeTestCommon( |
| 150 | armnn::IWorkloadFactory& workloadFactory, |
| 151 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 152 | bool transposeWeights, |
| 153 | float qScale = 0.0f, |
| 154 | int32_t qOffset = 0) |
| 155 | { |
| 156 | unsigned int inputWidth = 1; |
| 157 | unsigned int inputHeight = 1; |
| 158 | unsigned int inputChannels = 5; |
| 159 | unsigned int inputNum = 1; |
| 160 | |
| 161 | unsigned int outputChannels = 1; |
| 162 | unsigned int outputNum = 1; |
| 163 | |
| 164 | // Define the tensor descriptors. |
| 165 | armnn::TensorInfo inputTensorInfo; |
| 166 | armnn::TensorInfo outputTensorInfo; |
| 167 | armnn::TensorInfo weightsDesc; |
| 168 | armnn::TensorInfo biasesDesc; |
| 169 | |
| 170 | unsigned int inputShape[] = { inputNum, inputChannels, inputHeight, inputWidth }; |
| 171 | unsigned int outputShape[] = { outputNum, outputChannels }; |
| 172 | unsigned int weightsShape[] = { inputChannels, outputChannels }; |
| 173 | if (transposeWeights) |
| 174 | { |
| 175 | std::swap(weightsShape[0], weightsShape[1]); |
| 176 | } |
| 177 | |
| 178 | unsigned int biasShape[] = { outputChannels }; |
| 179 | |
| 180 | inputTensorInfo = armnn::TensorInfo(4, inputShape, ArmnnType); |
| 181 | outputTensorInfo = armnn::TensorInfo(2, outputShape, ArmnnType); |
| 182 | weightsDesc = armnn::TensorInfo(2, weightsShape, ArmnnType); |
| 183 | biasesDesc = armnn::TensorInfo(1, biasShape, ArmnnType); |
| 184 | |
| 185 | // Set quantization parameters if the requested type is a quantized type. |
| 186 | if(armnn::IsQuantizedType<T>()) |
| 187 | { |
| 188 | inputTensorInfo.SetQuantizationScale(qScale); |
| 189 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 190 | outputTensorInfo.SetQuantizationScale(qScale); |
| 191 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 192 | } |
| 193 | |
| 194 | LayerTestResult<T, 2> result(outputTensorInfo); |
| 195 | |
| 196 | boost::multi_array<T, 4> input = MakeTensor<T, 4>(inputTensorInfo, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 197 | armnnUtils::QuantizedVector<T>({ |
Aron Virginas-Tar | 00d306e | 2019-08-28 18:08:46 +0100 | [diff] [blame] | 198 | 1.0f, 10.0f, 100.0f, 1000.0f, 10000.0f, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 199 | }, |
| 200 | qScale, qOffset) |
Aron Virginas-Tar | 00d306e | 2019-08-28 18:08:46 +0100 | [diff] [blame] | 201 | ); |
| 202 | |
| 203 | boost::multi_array<T, 2> weights = MakeTensor<T, 2>(weightsDesc, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 204 | armnnUtils::QuantizedVector<T>({ |
Aron Virginas-Tar | 00d306e | 2019-08-28 18:08:46 +0100 | [diff] [blame] | 205 | 2.0f, 3.0f, 4.0f, 5.0f, 6.0f |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 206 | }, |
| 207 | qScale, qOffset) |
Aron Virginas-Tar | 00d306e | 2019-08-28 18:08:46 +0100 | [diff] [blame] | 208 | ); |
| 209 | |
| 210 | std::vector<T> biasValues({900000.f}); |
| 211 | boost::multi_array<T, 1> bias = MakeTensor<T, 1>(biasesDesc, biasValues); |
| 212 | |
| 213 | result = SimpleFullyConnectedTestImpl<T>( |
| 214 | workloadFactory, |
| 215 | memoryManager, |
| 216 | inputTensorInfo, outputTensorInfo, |
| 217 | weightsDesc, biasesDesc, |
| 218 | weights, bias, input, |
| 219 | true, transposeWeights |
| 220 | ); |
| 221 | |
| 222 | result.outputExpected = MakeTensor<T, 2>(outputTensorInfo, |
Aron Virginas-Tar | 48623a0 | 2019-10-22 10:00:28 +0100 | [diff] [blame] | 223 | armnnUtils::QuantizedVector<T>({ 965432.0f }, qScale, qOffset)); |
Aron Virginas-Tar | 00d306e | 2019-08-28 18:08:46 +0100 | [diff] [blame] | 224 | |
| 225 | return result; |
| 226 | } |
| 227 | |
| 228 | // |
| 229 | // Explicit template specializations |
| 230 | // |
| 231 | |
| 232 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QuantisedAsymm8>, 2> |
| 233 | FullyConnectedTest<armnn::DataType::QuantisedAsymm8>( |
| 234 | armnn::IWorkloadFactory& workloadFactory, |
| 235 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 236 | bool biasEnabled); |
| 237 | |
| 238 | template LayerTestResult<armnn::ResolveType<armnn::DataType::QuantisedSymm16>, 2> |
| 239 | FullyConnectedTest<armnn::DataType::QuantisedSymm16>( |
| 240 | armnn::IWorkloadFactory& workloadFactory, |
| 241 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 242 | bool biasEnabled); |
| 243 | |
| 244 | // |
| 245 | // Implementation functions |
| 246 | // |
| 247 | |
| 248 | LayerTestResult<float, 2> FullyConnectedFloat32Test( |
| 249 | armnn::IWorkloadFactory& workloadFactory, |
| 250 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 251 | bool biasEnabled, |
| 252 | bool transposeWeights) |
| 253 | { |
| 254 | unsigned int inputWidth = 1; |
| 255 | unsigned int inputHeight = 1; |
| 256 | unsigned int inputChannels = 5; |
| 257 | unsigned int inputNum = 2; |
| 258 | |
| 259 | unsigned int outputChannels = 3; |
| 260 | unsigned int outputNum = 2; |
| 261 | |
| 262 | // Define the tensor descriptors. |
| 263 | armnn::TensorInfo inputTensorInfo; |
| 264 | armnn::TensorInfo outputTensorInfo; |
| 265 | armnn::TensorInfo weightsDesc; |
| 266 | armnn::TensorInfo biasesDesc; |
| 267 | |
| 268 | unsigned int inputShape[] = { inputNum, inputChannels, inputHeight, inputWidth }; |
| 269 | unsigned int outputShape[] = { outputNum, outputChannels }; |
| 270 | unsigned int weightsShape[] = { inputChannels, outputChannels }; |
| 271 | |
| 272 | if (transposeWeights) |
| 273 | { |
| 274 | std::swap(weightsShape[0], weightsShape[1]); |
| 275 | } |
| 276 | |
| 277 | unsigned int biasShape[] = { outputChannels }; |
| 278 | |
| 279 | inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::DataType::Float32); |
| 280 | outputTensorInfo = armnn::TensorInfo(2, outputShape, armnn::DataType::Float32); |
| 281 | weightsDesc = armnn::TensorInfo(2, weightsShape, armnn::DataType::Float32); |
| 282 | biasesDesc = armnn::TensorInfo(1, biasShape, armnn::DataType::Float32); |
| 283 | |
| 284 | LayerTestResult<float, 2> result(outputTensorInfo); |
| 285 | |
| 286 | boost::multi_array<float, 4> input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>( |
| 287 | { |
| 288 | 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, |
| 289 | |
| 290 | 5.0f, 4.0f, 3.0f, 2.0f, 1.0f |
| 291 | }) |
| 292 | ); |
| 293 | |
| 294 | boost::multi_array<float, 2> weights = MakeTensor<float, 2>(weightsDesc, std::vector<float>( |
| 295 | { |
| 296 | .5f, 2.f, .5f, |
| 297 | .5f, 2.f, 1.f, |
| 298 | .5f, 2.f, 2.f, |
| 299 | .5f, 2.f, 3.f, |
| 300 | .5f, 2.f, 4.f |
| 301 | })); |
| 302 | |
| 303 | if (transposeWeights) |
| 304 | { |
| 305 | weights = MakeTensor<float, 2>(weightsDesc, std::vector<float>( |
| 306 | { |
| 307 | .5f, .5f, .5f, .5f, .5f, |
| 308 | 2.f, 2.f, 2.f, 2.f, 2.f, |
| 309 | .5f, 1.f, 2.f, 3.f, 4.f |
| 310 | })); |
| 311 | } |
| 312 | |
| 313 | |
| 314 | std::vector<float> biasValues({0.f, 0.f, 0.f}); |
| 315 | if (biasEnabled) |
| 316 | { |
| 317 | biasValues = std::vector<float>({10.f, 20.f, 30.f}); |
| 318 | } |
| 319 | boost::multi_array<float, 1> bias = MakeTensor<float, 1>(biasesDesc, biasValues); |
| 320 | |
| 321 | result = SimpleFullyConnectedTestImpl<float>( |
| 322 | workloadFactory, |
| 323 | memoryManager, |
| 324 | inputTensorInfo, outputTensorInfo, |
| 325 | weightsDesc, biasesDesc, |
| 326 | weights, bias, input, |
| 327 | biasEnabled, transposeWeights |
| 328 | ); |
| 329 | |
| 330 | result.outputExpected = MakeTensor<float, 2>(outputTensorInfo, std::vector<float>( |
| 331 | { |
| 332 | 0.5f + 1.0f + 1.5f + 2.0f + 2.5f + biasValues[0], |
| 333 | 2.0f + 4.0f + 6.0f + 8.0f + 10.f + biasValues[1], |
| 334 | 0.5f + 2.0f + 6.0f + 12.f + 20.f + biasValues[2], |
| 335 | |
| 336 | 2.5f + 2.0f + 1.5f + 1.0f + 0.5f + biasValues[0], |
| 337 | 10.0f + 8.0f + 6.0f + 4.0f + 2.f + biasValues[1], |
| 338 | 2.5f + 4.0f + 6.0f + 6.f + 4.f + biasValues[2] |
| 339 | }) |
| 340 | ); |
| 341 | |
| 342 | return result; |
| 343 | } |
| 344 | |
| 345 | LayerTestResult<float, 2> FullyConnectedLargeTest( |
| 346 | armnn::IWorkloadFactory& workloadFactory, |
| 347 | const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, |
| 348 | bool transposeWeights) |
| 349 | { |
| 350 | return FullyConnectedLargeTestCommon<armnn::DataType::Float32>(workloadFactory, memoryManager, transposeWeights); |
| 351 | } |