Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "ConversionUtils.hpp" |
| 7 | #include <armnnUtils/Permute.hpp> |
| 8 | |
| 9 | /// |
| 10 | /// Helper classes |
| 11 | /// |
| 12 | |
| 13 | namespace armnn_driver |
| 14 | { |
| 15 | |
| 16 | LayerInputHandle::LayerInputHandle() |
| 17 | : m_OutputSlot(nullptr) |
| 18 | , m_Valid(false) |
| 19 | {} |
| 20 | |
| 21 | LayerInputHandle::LayerInputHandle(bool valid, armnn::IOutputSlot* outputSlot, armnn::TensorInfo tensorInfo) |
| 22 | : m_OutputSlot(outputSlot) |
| 23 | , m_Valid(valid) |
| 24 | , m_TensorInfo(tensorInfo) |
| 25 | {} |
| 26 | |
| 27 | bool LayerInputHandle::IsValid() const |
| 28 | { |
| 29 | return m_Valid; |
| 30 | } |
| 31 | |
| 32 | void LayerInputHandle::Connect(armnn::IInputSlot& inputSlot) |
| 33 | { |
| 34 | ARMNN_ASSERT(IsValid()); |
| 35 | if (m_OutputSlot) |
| 36 | { |
| 37 | m_OutputSlot->Connect(inputSlot); |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | void LayerInputHandle::Disconnect(armnn::IInputSlot& inputSlot) |
| 42 | { |
| 43 | ARMNN_ASSERT(IsValid()); |
| 44 | if (m_OutputSlot) |
| 45 | { |
| 46 | m_OutputSlot->Disconnect(inputSlot); |
| 47 | } |
| 48 | } |
| 49 | |
| 50 | const armnn::TensorInfo& LayerInputHandle::GetTensorInfo() const |
| 51 | { |
| 52 | return m_TensorInfo; |
| 53 | } |
| 54 | |
| 55 | void LayerInputHandle::SanitizeQuantizationScale(LayerInputHandle& weight, LayerInputHandle& input) |
| 56 | { |
| 57 | if (m_OutputSlot) |
| 58 | { |
| 59 | armnn::TensorInfo weightInfo = weight.GetTensorInfo(); |
| 60 | armnn::TensorInfo inputInfo = input.GetTensorInfo(); |
| 61 | armnn::TensorInfo biasInfo = GetTensorInfo(); |
| 62 | |
| 63 | SanitizeBiasQuantizationScale(biasInfo, weightInfo, inputInfo); |
| 64 | |
| 65 | m_TensorInfo = biasInfo; |
| 66 | m_OutputSlot->SetTensorInfo(biasInfo); |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | ConstTensorPin::ConstTensorPin(bool optional) |
| 71 | : m_Optional(optional) |
| 72 | {} |
| 73 | |
| 74 | ConstTensorPin::ConstTensorPin(armnn::TensorInfo& tensorInfo, |
| 75 | const void* valueStart, |
| 76 | uint32_t numBytes, |
| 77 | const armnn::PermutationVector& mappings) |
| 78 | : m_Optional(false) |
| 79 | { |
| 80 | armnn::IgnoreUnused(numBytes); |
| 81 | if (tensorInfo.GetNumBytes() != numBytes) |
| 82 | { |
| 83 | VLOG(DRIVER) << "The size of ConstTensor does not match its TensorInfo."; |
| 84 | } |
| 85 | |
| 86 | const bool needsSwizzling = (mappings.GetSize() > 0); |
| 87 | if (needsSwizzling) |
| 88 | { |
| 89 | m_SwizzledTensorData.resize(tensorInfo.GetNumBytes()); |
| 90 | SwizzleAndroidNn4dTensorToArmNn(tensorInfo, valueStart, m_SwizzledTensorData.data(), mappings); |
| 91 | |
| 92 | m_ConstTensor = armnn::ConstTensor(tensorInfo, m_SwizzledTensorData.data()); |
| 93 | } |
| 94 | else |
| 95 | { |
| 96 | m_ConstTensor = armnn::ConstTensor(tensorInfo, valueStart); |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | bool ConstTensorPin::IsValid() const |
| 101 | { |
| 102 | return m_ConstTensor.GetMemoryArea() != nullptr; |
| 103 | } |
| 104 | |
| 105 | bool ConstTensorPin::IsOptional() const |
| 106 | { |
| 107 | return m_Optional; |
| 108 | } |
| 109 | |
| 110 | const armnn::ConstTensor& ConstTensorPin::GetConstTensor() const |
| 111 | { |
| 112 | return m_ConstTensor; |
| 113 | } |
| 114 | |
| 115 | const armnn::ConstTensor* ConstTensorPin::GetConstTensorPtr() const |
| 116 | { |
| 117 | if (IsValid() && m_ConstTensor.GetNumElements() > 0) |
| 118 | { |
| 119 | return &m_ConstTensor; |
| 120 | } |
| 121 | // tensor is either invalid, or has no elements (indicating an optional tensor that was not provided) |
| 122 | return nullptr; |
| 123 | } |
| 124 | |
| 125 | /// |
| 126 | /// Utility functions |
| 127 | /// |
| 128 | |
| 129 | bool IsWeightsValid(const Operation& operation, |
| 130 | uint32_t inputIndex, |
| 131 | const Model& model) |
| 132 | { |
| 133 | const Operand* operand = GetInputOperand(operation, inputIndex, model); |
| 134 | if (!operand) |
| 135 | { |
| 136 | Fail("%s: failed to get input operand %i", __func__, inputIndex); |
| 137 | return false; |
| 138 | } |
| 139 | |
| 140 | if (operand->lifetime != OperandLifeTime::CONSTANT_COPY |
| 141 | && operand->lifetime != OperandLifeTime::CONSTANT_REFERENCE |
| 142 | && operand->lifetime != OperandLifeTime::NO_VALUE) |
| 143 | { |
| 144 | return false; |
| 145 | } |
| 146 | return true; |
| 147 | } |
| 148 | |
| 149 | ConstTensorPin ConvertOperandToConstTensorPin(const Operand& operand, |
| 150 | const Model& model, |
| 151 | const ConversionData& data, |
| 152 | const armnn::PermutationVector& dimensionMappings, |
| 153 | const armnn::TensorShape* overrideTensorShape, |
Sadik Armagan | 5b1f539 | 2022-07-19 12:37:20 +0100 | [diff] [blame^] | 154 | bool optional, |
| 155 | const armnn::DataType* overrideDataType) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 156 | { |
| 157 | if (!IsOperandTypeSupportedForTensors(operand.type)) |
| 158 | { |
| 159 | VLOG(DRIVER) << __func__ << ": unsupported operand type for tensor" << operand.type; |
| 160 | return ConstTensorPin(); |
| 161 | } |
| 162 | |
| 163 | if (!optional && !IsOperandConstant(operand)) |
| 164 | { |
| 165 | VLOG(DRIVER) << __func__ << ": lifetime for input tensor: r" << operand.lifetime; |
| 166 | return ConstTensorPin(); |
| 167 | } |
| 168 | |
| 169 | const void* const valueStart = GetOperandValueReadOnlyAddress(operand, model, data, optional); |
| 170 | if (!valueStart) |
| 171 | { |
| 172 | if (optional) |
| 173 | { |
| 174 | // optional tensor with no values is not really an error; return it as invalid, but marked as optional |
| 175 | return ConstTensorPin(true); |
| 176 | } |
| 177 | // mandatory tensor with no values |
| 178 | Fail("%s: failed to get operand address", __func__); |
| 179 | return ConstTensorPin(); |
| 180 | } |
| 181 | |
| 182 | armnn::TensorInfo tensorInfo = GetTensorInfoForOperand(operand); |
| 183 | |
Sadik Armagan | 5b1f539 | 2022-07-19 12:37:20 +0100 | [diff] [blame^] | 184 | if (overrideTensorShape) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 185 | { |
| 186 | tensorInfo.SetShape(*overrideTensorShape); |
| 187 | } |
Sadik Armagan | 5b1f539 | 2022-07-19 12:37:20 +0100 | [diff] [blame^] | 188 | |
| 189 | if (overrideDataType) |
| 190 | { |
| 191 | tensorInfo.SetDataType(*overrideDataType); |
| 192 | } |
| 193 | |
| 194 | // Make sure isConstant flag is set. |
| 195 | tensorInfo.SetConstant(); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 196 | return ConstTensorPin(tensorInfo, valueStart, operand.location.length, dimensionMappings); |
| 197 | } |
| 198 | |
| 199 | LayerInputHandle ConvertToLayerInputHandle(const Operation& operation, |
| 200 | uint32_t inputIndex, |
| 201 | const Model& model, |
| 202 | ConversionData& data, |
Sadik Armagan | 5b1f539 | 2022-07-19 12:37:20 +0100 | [diff] [blame^] | 203 | const armnn::PermutationVector& dimensionMappings, |
| 204 | const LayerInputHandle* inputHandle) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 205 | { |
| 206 | |
| 207 | const Operand* operand = GetInputOperand(operation, inputIndex, model); |
| 208 | if (!operand) |
| 209 | { |
| 210 | Fail("%s: failed to get input operand %i", __func__, inputIndex); |
| 211 | return LayerInputHandle(); |
| 212 | } |
| 213 | |
| 214 | if (!IsOperandTypeSupportedForTensors(operand->type)) |
| 215 | { |
| 216 | VLOG(DRIVER) << __func__ << ": unsupported operand type for tensor: " << operand->type; |
| 217 | return LayerInputHandle(); |
| 218 | } |
| 219 | |
| 220 | try |
| 221 | { |
| 222 | armnn::TensorInfo operandTensorInfo = GetTensorInfoForOperand(*operand); |
| 223 | |
| 224 | if (IsDynamicTensor(operandTensorInfo)) |
| 225 | { |
| 226 | data.m_DynamicInputsEncountered = true; |
| 227 | |
| 228 | const uint32_t operandIndex = operation.inputs[inputIndex]; |
| 229 | |
| 230 | // Check if the dynamic input tensors have been inferred by one of the previous layers |
| 231 | // If not we can't support them |
| 232 | if (data.m_OutputSlotForOperand.size() >= operandIndex && data.m_OutputSlotForOperand[operandIndex]) |
| 233 | { |
| 234 | operandTensorInfo = data.m_OutputSlotForOperand[operandIndex]->GetTensorInfo(); |
| 235 | } |
| 236 | else |
| 237 | { |
| 238 | Fail("%s: Type 2 dynamic input tensors are not supported", __func__); |
| 239 | return LayerInputHandle(); |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | switch (operand->lifetime) |
| 244 | { |
| 245 | case OperandLifeTime::SUBGRAPH_INPUT: |
| 246 | { |
| 247 | // NOTE: We must check whether we can support the input tensor on at least one |
| 248 | // of the provided backends; otherwise we cannot convert the operation |
| 249 | bool isInputSupported = false; |
| 250 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 251 | IsInputSupported, |
| 252 | data.m_Backends, |
| 253 | isInputSupported, |
| 254 | operandTensorInfo); |
| 255 | |
| 256 | if (!isInputSupported) |
| 257 | { |
| 258 | Fail("%s: unsupported input tensor", __func__); |
| 259 | return LayerInputHandle(); |
| 260 | } |
| 261 | |
| 262 | [[clang::fallthrough]]; // intentional fallthrough |
| 263 | } |
| 264 | case OperandLifeTime::TEMPORARY_VARIABLE: // intentional fallthrough |
| 265 | case OperandLifeTime::SUBGRAPH_OUTPUT: |
| 266 | { |
| 267 | // The tensor is either an operand internal to the model, or a model input. |
| 268 | // It can be associated with an ArmNN output slot for an existing layer. |
| 269 | |
| 270 | // m_OutputSlotForOperand[...] can be nullptr if the previous layer could not be converted |
| 271 | const uint32_t operandIndex = operation.inputs[inputIndex]; |
| 272 | return LayerInputHandle(true, data.m_OutputSlotForOperand[operandIndex], operandTensorInfo); |
| 273 | } |
| 274 | case OperandLifeTime::CONSTANT_COPY: // intentional fallthrough |
| 275 | case OperandLifeTime::POINTER: |
| 276 | case OperandLifeTime::CONSTANT_REFERENCE: |
| 277 | { |
Sadik Armagan | 5b1f539 | 2022-07-19 12:37:20 +0100 | [diff] [blame^] | 278 | auto constantTensorDataType = operandTensorInfo.GetDataType(); |
| 279 | if (inputHandle) |
| 280 | { |
| 281 | if ((inputHandle->GetTensorInfo().GetDataType() == armnn::DataType::Float32 |
| 282 | || inputHandle->GetTensorInfo().GetDataType() == armnn::DataType::Float16) |
| 283 | && (operandTensorInfo.GetDataType() == armnn::DataType::QAsymmU8 |
| 284 | || operandTensorInfo.GetDataType() == armnn::DataType::QAsymmS8)) |
| 285 | { |
| 286 | constantTensorDataType = inputHandle->GetTensorInfo().GetDataType(); |
| 287 | } |
| 288 | } |
| 289 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 290 | // The tensor has an already known constant value, and can be converted into an ArmNN Constant layer. |
Sadik Armagan | 5b1f539 | 2022-07-19 12:37:20 +0100 | [diff] [blame^] | 291 | ConstTensorPin tensorPin = ConvertOperandToConstTensorPin(*operand, |
| 292 | model, |
| 293 | data, |
| 294 | dimensionMappings, |
| 295 | nullptr, |
| 296 | false, |
| 297 | &constantTensorDataType); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 298 | if (tensorPin.IsValid()) |
| 299 | { |
| 300 | bool isSupported = false; |
| 301 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 302 | IsConstantSupported, |
| 303 | data.m_Backends, |
| 304 | isSupported, |
| 305 | tensorPin.GetConstTensor().GetInfo()); |
| 306 | if (!isSupported) |
| 307 | { |
| 308 | return LayerInputHandle(); |
| 309 | } |
| 310 | |
| 311 | armnn::IConnectableLayer* constantLayer = |
| 312 | data.m_Network->AddConstantLayer(tensorPin.GetConstTensor()); |
| 313 | armnn::IOutputSlot& outputSlot = constantLayer->GetOutputSlot(0); |
Sadik Armagan | b9570c1 | 2022-06-21 11:39:41 +0100 | [diff] [blame] | 314 | armnn::TensorInfo constantTensorInfo = tensorPin.GetConstTensor().GetInfo(); |
| 315 | outputSlot.SetTensorInfo(constantTensorInfo); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 316 | |
Sadik Armagan | b9570c1 | 2022-06-21 11:39:41 +0100 | [diff] [blame] | 317 | return LayerInputHandle(true, &outputSlot, constantTensorInfo); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 318 | } |
| 319 | else |
| 320 | { |
| 321 | Fail("%s: invalid operand tensor", __func__); |
| 322 | return LayerInputHandle(); |
| 323 | } |
| 324 | break; |
| 325 | } |
| 326 | default: |
| 327 | { |
| 328 | VLOG(DRIVER) << __func__ << ": unsupported lifetime for input tensor: " << operand->lifetime; |
| 329 | return LayerInputHandle(); |
| 330 | } |
| 331 | } |
| 332 | } |
| 333 | catch (UnsupportedOperand<OperandType>& e) |
| 334 | { |
| 335 | VLOG(DRIVER) << __func__ << ": Operand type: " << e.m_type << " not supported in ArmnnDriver"; |
| 336 | return LayerInputHandle(); |
| 337 | } |
| 338 | } |
| 339 | |
| 340 | bool ConvertPaddings(const Operation& operation, |
| 341 | const Model& model, |
| 342 | ConversionData& data, |
| 343 | unsigned int rank, |
| 344 | armnn::PadDescriptor& padDescriptor) |
| 345 | { |
| 346 | const Operand* paddingsOperand = GetInputOperand(operation, 1, model); |
| 347 | if (!paddingsOperand) |
| 348 | { |
| 349 | return Fail("%s: Could not read paddings operand", __func__); |
| 350 | } |
| 351 | |
| 352 | armnn::TensorShape paddingsOperandShape = GetTensorShapeForOperand(*paddingsOperand); |
| 353 | if (paddingsOperandShape.GetNumDimensions() != 2 || paddingsOperandShape.GetNumElements() != rank * 2) |
| 354 | { |
| 355 | return Fail("%s: Operation has invalid paddings operand: expected shape [%d, 2]", __func__, rank); |
| 356 | } |
| 357 | |
| 358 | std::vector<int32_t> paddings; |
| 359 | if (!GetTensorInt32Values(*paddingsOperand, paddings, model, data)) |
| 360 | { |
| 361 | return Fail("%s: Operation has invalid or unsupported paddings operand", __func__); |
| 362 | } |
| 363 | |
| 364 | // add padding for each dimension of input tensor. |
| 365 | for (unsigned int i = 0; i < paddings.size() - 1; i += 2) |
| 366 | { |
| 367 | int paddingBeforeInput = paddings[i]; |
| 368 | int paddingAfterInput = paddings[i + 1]; |
| 369 | |
| 370 | if (paddingBeforeInput < 0 || paddingAfterInput < 0) |
| 371 | { |
| 372 | return Fail("%s: Operation has invalid paddings operand, invalid padding values.", __func__); |
| 373 | } |
| 374 | |
| 375 | padDescriptor.m_PadList.emplace_back((unsigned int) paddingBeforeInput, (unsigned int) paddingAfterInput); |
| 376 | } |
| 377 | |
| 378 | return true; |
| 379 | } |
| 380 | |
| 381 | |
| 382 | bool ConvertPooling2d(const Operation& operation, |
| 383 | const char* operationName, |
| 384 | armnn::PoolingAlgorithm poolType, |
| 385 | const Model& model, |
| 386 | ConversionData& data) |
| 387 | { |
| 388 | |
| 389 | VLOG(DRIVER) << "Converter::ConvertL2Pool2d()"; |
| 390 | |
| 391 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 392 | if (!input.IsValid()) |
| 393 | { |
| 394 | return Fail("%s: Operation Could not read input 0", operationName); |
| 395 | } |
| 396 | |
| 397 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 398 | if (!output) |
| 399 | { |
| 400 | return Fail("%s: Could not read output 0", __func__); |
| 401 | } |
| 402 | |
| 403 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 404 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 405 | |
| 406 | armnn::Pooling2dDescriptor desc; |
| 407 | desc.m_PoolType = poolType; |
| 408 | desc.m_OutputShapeRounding = armnn::OutputShapeRounding::Floor; |
| 409 | desc.m_DataLayout = armnn::DataLayout::NHWC; |
| 410 | |
| 411 | ActivationFn activation; |
| 412 | |
| 413 | auto inputSize = operation.inputs.size(); |
| 414 | |
| 415 | if (inputSize >= 10) |
| 416 | { |
| 417 | // one input, 9 parameters (padding l r t b, stridex, stridey, width, height, activation type) |
| 418 | if (!GetInputScalar(operation, 1, OperandType::INT32, desc.m_PadLeft, model, data) || |
| 419 | !GetInputScalar(operation, 2, OperandType::INT32, desc.m_PadRight, model, data) || |
| 420 | !GetInputScalar(operation, 3, OperandType::INT32, desc.m_PadTop, model, data) || |
| 421 | !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PadBottom, model, data) || |
| 422 | !GetInputScalar(operation, 5, OperandType::INT32, desc.m_StrideX, model, data) || |
| 423 | !GetInputScalar(operation, 6, OperandType::INT32, desc.m_StrideY, model, data) || |
| 424 | !GetInputScalar(operation, 7, OperandType::INT32, desc.m_PoolWidth, model, data) || |
| 425 | !GetInputScalar(operation, 8, OperandType::INT32, desc.m_PoolHeight, model, data) || |
| 426 | !GetInputActivationFunction(operation, 9, activation, model, data)) |
| 427 | { |
| 428 | return Fail("%s: Operation has invalid inputs", operationName); |
| 429 | } |
| 430 | |
| 431 | if (Is12OrLaterOperand(*output)) |
| 432 | { |
| 433 | desc.m_DataLayout = OptionalDataLayout(operation, 10, model, data); |
| 434 | } |
| 435 | } |
| 436 | else |
| 437 | { |
| 438 | // one input, 6 parameters (padding, stridex, stridey, width, height, activation type) |
| 439 | ::android::nn::PaddingScheme scheme; |
| 440 | if (!GetInputPaddingScheme(operation, 1, scheme, model, data) || |
| 441 | !GetInputScalar(operation, 2, OperandType::INT32, desc.m_StrideX, model, data) || |
| 442 | !GetInputScalar(operation, 3, OperandType::INT32, desc.m_StrideY, model, data) || |
| 443 | !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PoolWidth, model, data) || |
| 444 | !GetInputScalar(operation, 5, OperandType::INT32, desc.m_PoolHeight, model, data) || |
| 445 | !GetInputActivationFunction(operation, 6, activation, model, data)) |
| 446 | { |
| 447 | return Fail("%s: Operation has invalid inputs", operationName); |
| 448 | } |
| 449 | |
| 450 | if (Is12OrLaterOperand(*output)) |
| 451 | { |
| 452 | desc.m_DataLayout = OptionalDataLayout(operation, 7, model, data); |
| 453 | } |
| 454 | |
| 455 | const armnnUtils::DataLayoutIndexed dataLayout(desc.m_DataLayout); |
| 456 | const unsigned int inputWidth = inputInfo.GetShape()[dataLayout.GetWidthIndex()]; |
| 457 | const unsigned int inputHeight = inputInfo.GetShape()[dataLayout.GetHeightIndex()]; |
| 458 | |
| 459 | CalcPadding(inputWidth, desc.m_PoolWidth, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, scheme); |
| 460 | CalcPadding(inputHeight, desc.m_PoolHeight, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, scheme); |
| 461 | } |
| 462 | |
| 463 | bool isSupported = false; |
| 464 | |
| 465 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 466 | { |
| 467 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 468 | IsPooling2dSupported, |
| 469 | data.m_Backends, |
| 470 | isSupported, |
| 471 | inputInfo, |
| 472 | outputInfo, |
| 473 | desc); |
| 474 | |
| 475 | }; |
| 476 | |
| 477 | if(IsDynamicTensor(outputInfo)) |
| 478 | { |
| 479 | isSupported = AreDynamicTensorsSupported(); |
| 480 | } |
| 481 | else |
| 482 | { |
| 483 | validateFunc(outputInfo, isSupported); |
| 484 | } |
| 485 | |
| 486 | if (!isSupported) |
| 487 | { |
| 488 | return false; |
| 489 | } |
| 490 | |
| 491 | armnn::IConnectableLayer* pooling2dLayer = data.m_Network->AddPooling2dLayer(desc); |
| 492 | if (!pooling2dLayer) |
| 493 | { |
| 494 | return Fail("%s: AddPooling2dLayer failed", __func__); |
| 495 | } |
| 496 | |
| 497 | input.Connect(pooling2dLayer->GetInputSlot(0)); |
| 498 | |
| 499 | if (!isSupported) |
| 500 | { |
| 501 | return false; |
| 502 | } |
| 503 | |
| 504 | return SetupAndTrackLayerOutputSlot(operation, 0, *pooling2dLayer, model, |
| 505 | data, nullptr, validateFunc, activation); |
| 506 | } |
| 507 | |
| 508 | bool ConvertReduce(const Operation& operation, |
| 509 | const Model& model, |
| 510 | ConversionData& data, |
| 511 | armnn::ReduceOperation reduceOperation) |
| 512 | { |
| 513 | armnn::ReduceDescriptor descriptor; |
| 514 | descriptor.m_ReduceOperation = reduceOperation; |
| 515 | |
| 516 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 517 | if (!input.IsValid()) |
| 518 | { |
| 519 | return Fail("%s: Operation has invalid inputs", __func__); |
| 520 | } |
| 521 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 522 | |
| 523 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 524 | if (!output) |
| 525 | { |
| 526 | return Fail("%s: Could not read output 0", __func__); |
| 527 | } |
| 528 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 529 | |
| 530 | const Operand* axisOperand = GetInputOperand(operation, 1, model); |
| 531 | if (!axisOperand) |
| 532 | { |
| 533 | return Fail("%s: Could not read input 1", __func__); |
| 534 | } |
| 535 | std::vector<int32_t> axis; |
| 536 | if (!GetTensorInt32Values(*axisOperand, axis, model, data)) |
| 537 | { |
| 538 | return Fail("%s: Input 1 has invalid values", __func__); |
| 539 | } |
| 540 | |
| 541 | // Convert the axis to unsigned int and remove duplicates. |
| 542 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 543 | std::set<unsigned int> uniqueAxis; |
| 544 | std::transform(axis.begin(), axis.end(), |
| 545 | std::inserter(uniqueAxis, uniqueAxis.begin()), |
| 546 | [rank](int i) -> unsigned int { return (i + rank) % rank; }); |
| 547 | descriptor.m_vAxis.assign(uniqueAxis.begin(), uniqueAxis.end()); |
| 548 | |
| 549 | // Get the "keep dims" flag. |
| 550 | if (!GetInputScalar(operation, 2, OperandType::BOOL, descriptor.m_KeepDims, model, data)) |
| 551 | { |
| 552 | return Fail("%s: Could not read input 2", __func__); |
| 553 | } |
| 554 | |
| 555 | bool isSupported = false; |
| 556 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 557 | { |
| 558 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 559 | IsReduceSupported, |
| 560 | data.m_Backends, |
| 561 | isSupported, |
| 562 | inputInfo, |
| 563 | outputInfo, |
| 564 | descriptor); |
| 565 | }; |
| 566 | |
| 567 | if(!IsDynamicTensor(outputInfo)) |
| 568 | { |
| 569 | validateFunc(outputInfo, isSupported); |
| 570 | } |
| 571 | else |
| 572 | { |
| 573 | isSupported = AreDynamicTensorsSupported(); |
| 574 | } |
| 575 | |
| 576 | if (!isSupported) |
| 577 | { |
| 578 | return false; |
| 579 | } |
| 580 | |
| 581 | armnn::IConnectableLayer* const layer = data.m_Network->AddReduceLayer(descriptor); |
| 582 | assert(layer != nullptr); |
| 583 | input.Connect(layer->GetInputSlot(0)); |
| 584 | |
| 585 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 586 | } |
| 587 | |
| 588 | |
| 589 | bool ConvertToActivation(const Operation& operation, |
| 590 | const char* operationName, |
| 591 | const armnn::ActivationDescriptor& activationDesc, |
| 592 | const Model& model, |
| 593 | ConversionData& data) |
| 594 | { |
| 595 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 596 | if (!input.IsValid()) |
| 597 | { |
| 598 | return Fail("%s: Input 0 is invalid", operationName); |
| 599 | } |
| 600 | |
| 601 | const Operand* outputOperand = GetOutputOperand(operation, 0, model); |
| 602 | if (!outputOperand) |
| 603 | { |
| 604 | return false; |
| 605 | } |
| 606 | |
| 607 | const armnn::TensorInfo& outInfo = GetTensorInfoForOperand(*outputOperand); |
| 608 | |
| 609 | bool isSupported = false; |
| 610 | |
| 611 | auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported) |
| 612 | { |
| 613 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 614 | IsActivationSupported, |
| 615 | data.m_Backends, |
| 616 | isSupported, |
| 617 | input.GetTensorInfo(), |
| 618 | outInfo, |
| 619 | activationDesc); |
| 620 | }; |
| 621 | |
| 622 | if(IsDynamicTensor(outInfo)) |
| 623 | { |
| 624 | isSupported = AreDynamicTensorsSupported(); |
| 625 | } |
| 626 | else |
| 627 | { |
| 628 | validateFunc(outInfo, isSupported); |
| 629 | } |
| 630 | |
| 631 | if (!isSupported) |
| 632 | { |
| 633 | return false; |
| 634 | } |
| 635 | |
| 636 | armnn::IConnectableLayer* layer = data.m_Network->AddActivationLayer(activationDesc); |
| 637 | ARMNN_ASSERT(layer != nullptr); |
| 638 | input.Connect(layer->GetInputSlot(0)); |
| 639 | |
| 640 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 641 | } |
| 642 | |
| 643 | DequantizeResult DequantizeIfRequired(size_t operand_index, |
| 644 | const Operation& operation, |
| 645 | const Model& model, |
| 646 | const ConversionData& data) |
| 647 | { |
| 648 | const Operand* weightsOperand = GetInputOperand(operation, operand_index, model); |
| 649 | if (!weightsOperand) |
| 650 | { |
| 651 | return { nullptr, 0, armnn::TensorInfo(), DequantizeStatus::INVALID_OPERAND }; |
| 652 | } |
| 653 | |
| 654 | if (IsOperandConstant(*weightsOperand)) |
| 655 | { |
| 656 | // Weights are already constant |
| 657 | return { nullptr, 0, armnn::TensorInfo(), DequantizeStatus::NOT_REQUIRED }; |
| 658 | } |
| 659 | |
| 660 | const size_t weightsInputIndex = operation.inputs[operand_index]; |
| 661 | |
| 662 | // The weights are a non const tensor, this indicates they might be the output of a dequantize op. |
| 663 | // Iterate over the nodes and find the previous operation which should be DEQUANTIZE |
| 664 | for (uint32_t operationIdx = 0; operationIdx < getMainModel(model).operations.size(); ++operationIdx) |
| 665 | { |
| 666 | // Search for the DEQUANTIZE op which has the operand with index equal to operandIndex |
| 667 | const auto& operationIt = getMainModel(model).operations[operationIdx]; |
| 668 | if (operationIt.type != OperationType::DEQUANTIZE) |
| 669 | { |
| 670 | continue; |
| 671 | } |
| 672 | |
| 673 | size_t outOpIndex = weightsInputIndex + 1; |
| 674 | for (size_t i = 0; outOpIndex != weightsInputIndex && i < operationIt.outputs.size(); ++i) |
| 675 | { |
| 676 | outOpIndex = operationIt.outputs[i]; |
| 677 | } |
| 678 | |
| 679 | if (outOpIndex != weightsInputIndex) |
| 680 | { |
| 681 | continue; |
| 682 | } |
| 683 | |
| 684 | const Operand* operand = GetInputOperand(operationIt, 0, model); |
| 685 | ARMNN_ASSERT(operand); |
| 686 | |
| 687 | if (!IsQSymm8(*operand)) |
| 688 | { |
| 689 | // Only supporting dequantize from QSYMM8 to FLOAT |
| 690 | break; |
| 691 | } |
| 692 | |
| 693 | // Allocate a new buffer for the dequantized data and manually dequantize |
| 694 | const void* startValue = GetOperandValueReadOnlyAddress(*operand, model, data); |
| 695 | if (!startValue) |
| 696 | { |
| 697 | // Failed to get the operand address |
| 698 | break; |
| 699 | } |
| 700 | |
| 701 | const uint8_t* quantizedBuffer = reinterpret_cast<const uint8_t*>(startValue); |
| 702 | size_t dequantizedBufferLength = operand->location.length; |
| 703 | const float quantizationScale = operand->scale; |
| 704 | |
| 705 | auto dequantizedBuffer = std::make_unique<float[]>(dequantizedBufferLength + 1); |
| 706 | for (size_t i = 0; i < dequantizedBufferLength; ++i) |
| 707 | { |
| 708 | float* dstPtr = dequantizedBuffer.get(); |
| 709 | ARMNN_ASSERT(dstPtr); |
| 710 | *dstPtr++ = quantizedBuffer[i] * quantizationScale; |
| 711 | } |
| 712 | |
| 713 | // Construct tensor info for dequantized ConstTensor |
| 714 | armnn::TensorInfo tensorInfo(operand->dimensions.size(), |
| 715 | operand->dimensions.data(), |
| 716 | armnn::DataType::Float32); |
| 717 | |
| 718 | return { std::move(dequantizedBuffer), dequantizedBufferLength * sizeof(float), |
| 719 | std::move(tensorInfo), |
| 720 | DequantizeStatus::SUCCESS }; |
| 721 | } |
| 722 | |
| 723 | return { nullptr, 0, armnn::TensorInfo() , DequantizeStatus::NOT_REQUIRED}; |
| 724 | } |
| 725 | |
| 726 | ConstTensorPin DequantizeAndMakeConstTensorPin(const Operation& operation, |
| 727 | const Model& model, |
| 728 | const ConversionData& data, |
| 729 | size_t operandIndex, |
| 730 | bool optional) |
| 731 | { |
| 732 | DequantizeResult dequantized = DequantizeIfRequired(operandIndex,operation, model, data); |
| 733 | |
| 734 | DequantizeStatus status = std::get<3>(dequantized); |
| 735 | switch (status) |
| 736 | { |
| 737 | case DequantizeStatus::INVALID_OPERAND: |
| 738 | { |
| 739 | // return invalid const tensor pin |
| 740 | return ConstTensorPin(); |
| 741 | } |
| 742 | case DequantizeStatus::NOT_REQUIRED: |
| 743 | { |
| 744 | return ConvertOperationInputToConstTensorPin( |
| 745 | operation, operandIndex, model, data, g_DontPermute, nullptr, optional); |
| 746 | } |
| 747 | case DequantizeStatus::SUCCESS: |
| 748 | default: |
| 749 | { |
| 750 | return ConstTensorPin( |
| 751 | std::get<2>(dequantized), std::get<0>(dequantized).get(), std::get<1>(dequantized), g_DontPermute); |
| 752 | } |
| 753 | } |
| 754 | } |
| 755 | |
| 756 | bool GetInputPaddingScheme(const Operation& operation, |
| 757 | uint32_t inputIndex, |
| 758 | PaddingScheme& outPaddingScheme, |
| 759 | const Model& model, |
| 760 | const ConversionData& data) |
| 761 | { |
| 762 | int32_t paddingSchemeAsInt; |
| 763 | if (!GetInputInt32(operation, inputIndex, paddingSchemeAsInt, model, data)) |
| 764 | { |
| 765 | return Fail("%s: failed to get padding scheme input value", __func__); |
| 766 | } |
| 767 | |
| 768 | outPaddingScheme = static_cast<::android::nn::PaddingScheme>(paddingSchemeAsInt); |
| 769 | return true; |
| 770 | } |
| 771 | |
| 772 | const void* GetOperandValueReadOnlyAddress(const Operand& operand, |
| 773 | const Model& model, |
| 774 | const ConversionData& data, |
| 775 | bool optional) |
| 776 | { |
| 777 | const void* valueStart = nullptr; |
| 778 | switch (operand.lifetime) |
| 779 | { |
| 780 | case OperandLifeTime::CONSTANT_COPY: |
| 781 | { |
| 782 | valueStart = model.operandValues.data() + operand.location.offset; |
| 783 | break; |
| 784 | } |
| 785 | case OperandLifeTime::POINTER: |
| 786 | { |
| 787 | // Pointer specified in the model |
| 788 | valueStart = std::get<const void*>(operand.location.pointer); |
| 789 | break; |
| 790 | } |
| 791 | case OperandLifeTime::CONSTANT_REFERENCE: |
| 792 | { |
| 793 | // Constant specified via a Memory object |
| 794 | valueStart = GetMemoryFromPool(operand.location, data.m_MemPools); |
| 795 | break; |
| 796 | } |
| 797 | case OperandLifeTime::NO_VALUE: |
| 798 | { |
| 799 | // An optional input tensor with no values is not an error so should not register as a fail |
| 800 | if (optional) |
| 801 | { |
| 802 | valueStart = nullptr; |
| 803 | break; |
| 804 | } |
| 805 | [[fallthrough]]; |
| 806 | } |
| 807 | default: |
| 808 | { |
| 809 | VLOG(DRIVER) << __func__ << ": unsupported/invalid operand lifetime:: " << operand.lifetime; |
| 810 | valueStart = nullptr; |
| 811 | } |
| 812 | } |
| 813 | |
| 814 | return valueStart; |
| 815 | } |
| 816 | |
| 817 | bool GetTensorInt32Values(const Operand& operand, |
| 818 | std::vector<int32_t>& outValues, |
| 819 | const Model& model, |
| 820 | const ConversionData& data) |
| 821 | { |
| 822 | if (operand.type != OperandType::TENSOR_INT32) |
| 823 | { |
| 824 | VLOG(DRIVER) << __func__ << ": invalid operand type: " << operand.type; |
| 825 | return false; |
| 826 | } |
| 827 | |
| 828 | const void* startAddress = GetOperandValueReadOnlyAddress(operand, model, data); |
| 829 | if (!startAddress) |
| 830 | { |
| 831 | VLOG(DRIVER) << __func__ << ": failed to get operand address " << operand.type; |
| 832 | return false; |
| 833 | } |
| 834 | |
| 835 | // Check number of bytes is sensible |
| 836 | const uint32_t numBytes = operand.location.length; |
| 837 | if (numBytes % sizeof(int32_t) != 0) |
| 838 | { |
| 839 | return Fail("%s: invalid number of bytes: %i, expected to be a multiple of %i", |
| 840 | __func__, numBytes, sizeof(int32_t)); |
| 841 | } |
| 842 | |
| 843 | outValues.resize(numBytes / sizeof(int32_t)); |
| 844 | memcpy(outValues.data(), startAddress, numBytes); |
| 845 | return true; |
| 846 | } |
| 847 | |
| 848 | armnn::DataLayout OptionalDataLayout(const Operation& operation, |
| 849 | uint32_t inputIndex, |
| 850 | const Model& model, |
| 851 | ConversionData& data) |
| 852 | { |
| 853 | const Operand* operand = GetInputOperand(operation, inputIndex, model); |
| 854 | if (!operand) |
| 855 | { |
| 856 | return armnn::DataLayout::NHWC; |
| 857 | } |
| 858 | |
| 859 | if (!IsBool(*operand)) |
| 860 | { |
| 861 | return armnn::DataLayout::NHWC; |
| 862 | } |
| 863 | |
| 864 | const void* valueAddress = GetOperandValueReadOnlyAddress(*operand, model, data); |
| 865 | if (!valueAddress) |
| 866 | { |
| 867 | return armnn::DataLayout::NHWC; |
| 868 | } |
| 869 | |
| 870 | if (*(static_cast<const bool*>(valueAddress))) |
| 871 | { |
| 872 | return armnn::DataLayout::NCHW; |
| 873 | } |
| 874 | else |
| 875 | { |
| 876 | return armnn::DataLayout::NHWC; |
| 877 | } |
| 878 | } |
| 879 | |
| 880 | armnn::IConnectableLayer* ProcessActivation(const armnn::TensorInfo& tensorInfo, |
| 881 | ActivationFn activation, |
| 882 | armnn::IConnectableLayer* prevLayer, |
| 883 | ConversionData& data) |
| 884 | { |
| 885 | ARMNN_ASSERT(prevLayer->GetNumOutputSlots() == 1); |
| 886 | |
| 887 | prevLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 888 | |
| 889 | armnn::IConnectableLayer* activationLayer = prevLayer; |
| 890 | |
| 891 | if (activation != ActivationFn::kActivationNone) |
| 892 | { |
| 893 | armnn::ActivationDescriptor activationDesc; |
| 894 | switch (activation) |
| 895 | { |
| 896 | case ActivationFn::kActivationRelu: |
| 897 | { |
| 898 | activationDesc.m_Function = armnn::ActivationFunction::ReLu; |
| 899 | break; |
| 900 | } |
| 901 | case ActivationFn::kActivationRelu1: |
| 902 | { |
| 903 | activationDesc.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 904 | activationDesc.m_A = 1.0f; |
| 905 | activationDesc.m_B = -1.0f; |
| 906 | break; |
| 907 | } |
| 908 | case ActivationFn::kActivationRelu6: |
| 909 | { |
| 910 | activationDesc.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 911 | activationDesc.m_A = 6.0f; |
| 912 | break; |
| 913 | } |
| 914 | case ActivationFn::kActivationSigmoid: |
| 915 | { |
| 916 | activationDesc.m_Function = armnn::ActivationFunction::Sigmoid; |
| 917 | break; |
| 918 | } |
| 919 | case ActivationFn::kActivationTanh: |
| 920 | { |
| 921 | activationDesc.m_Function = armnn::ActivationFunction::TanH; |
| 922 | activationDesc.m_A = 1.0f; |
| 923 | activationDesc.m_B = 1.0f; |
| 924 | break; |
| 925 | } |
| 926 | default: |
| 927 | { |
| 928 | Fail("%s: Invalid activation enum value %i", __func__, activation); |
| 929 | return nullptr; |
| 930 | } |
| 931 | } |
| 932 | |
| 933 | bool isSupported = false; |
| 934 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 935 | IsActivationSupported, |
| 936 | data.m_Backends, |
| 937 | isSupported, |
| 938 | prevLayer->GetOutputSlot(0).GetTensorInfo(), |
| 939 | tensorInfo, |
| 940 | activationDesc); |
| 941 | if (!isSupported) |
| 942 | { |
| 943 | return nullptr; |
| 944 | } |
| 945 | |
| 946 | activationLayer = data.m_Network->AddActivationLayer(activationDesc); |
| 947 | |
| 948 | prevLayer->GetOutputSlot(0).Connect(activationLayer->GetInputSlot(0)); |
| 949 | activationLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 950 | } |
| 951 | |
| 952 | return activationLayer; |
| 953 | } |
| 954 | |
| 955 | bool SetupAndTrackLayerOutputSlot(const Operation& operation, |
| 956 | uint32_t operationOutputIndex, |
| 957 | armnn::IConnectableLayer& layer, |
| 958 | uint32_t layerOutputIndex, |
| 959 | const Model& model, |
| 960 | ConversionData& data, |
| 961 | const armnn::TensorInfo* overrideOutputInfo, |
| 962 | const std::function <void (const armnn::TensorInfo&, bool&)>& validateFunc, |
| 963 | const ActivationFn& activationFunction, |
| 964 | bool inferOutputShapes) |
| 965 | { |
| 966 | const Operand* outputOperand = GetOutputOperand(operation, operationOutputIndex, model); |
| 967 | if ((outputOperand == nullptr) || (operationOutputIndex >= layer.GetNumOutputSlots())) |
| 968 | { |
| 969 | return false; |
| 970 | } |
| 971 | |
| 972 | armnn::IOutputSlot& outputSlot = layer.GetOutputSlot(layerOutputIndex); |
| 973 | if (overrideOutputInfo == nullptr) |
| 974 | { |
| 975 | outputSlot.SetTensorInfo(GetTensorInfoForOperand(*outputOperand)); |
| 976 | } |
| 977 | else |
| 978 | { |
| 979 | outputSlot.SetTensorInfo(*overrideOutputInfo); |
| 980 | } |
| 981 | |
| 982 | bool isSupported = false; |
| 983 | if (validateFunc && (IsDynamicTensor(outputSlot.GetTensorInfo()) || inferOutputShapes)) |
| 984 | { |
| 985 | // Type one dynamic tensors require the previous layer's output shape for inference |
| 986 | for (unsigned int inputSlotIndex = 0; inputSlotIndex < layer.GetNumInputSlots(); ++inputSlotIndex) |
| 987 | { |
| 988 | if(!layer.GetInputSlot(inputSlotIndex).GetConnection()) |
| 989 | { |
| 990 | return false; |
| 991 | } |
| 992 | } |
| 993 | // IsTensorInfoSet will infer the dynamic output shape |
| 994 | outputSlot.IsTensorInfoSet(); |
| 995 | // Once the shape is inferred we can validate it |
| 996 | validateFunc(outputSlot.GetTensorInfo(), isSupported); |
| 997 | |
| 998 | if(!isSupported) |
| 999 | { |
| 1000 | for (unsigned int inputSlotIndex = 0; inputSlotIndex < layer.GetNumInputSlots(); ++inputSlotIndex) |
| 1001 | { |
| 1002 | layer.GetInputSlot(inputSlotIndex).GetConnection()->Disconnect(layer.GetInputSlot(inputSlotIndex)); |
| 1003 | } |
| 1004 | return false; |
| 1005 | } |
| 1006 | } |
| 1007 | |
| 1008 | const uint32_t operandIndex = operation.outputs[operationOutputIndex]; |
| 1009 | |
| 1010 | if (activationFunction != ActivationFn::kActivationNone) |
| 1011 | { |
| 1012 | const armnn::TensorInfo& activationOutputInfo = outputSlot.GetTensorInfo(); |
| 1013 | armnn::IConnectableLayer* const endLayer = ProcessActivation(activationOutputInfo, activationFunction, |
| 1014 | &layer, data); |
| 1015 | |
| 1016 | if (!endLayer) |
| 1017 | { |
| 1018 | return Fail("%s: ProcessActivation failed", __func__); |
| 1019 | } |
| 1020 | |
| 1021 | armnn::IOutputSlot& activationOutputSlot = endLayer->GetOutputSlot(layerOutputIndex); |
| 1022 | data.m_OutputSlotForOperand[operandIndex] = &activationOutputSlot; |
| 1023 | } |
| 1024 | else |
| 1025 | { |
| 1026 | data.m_OutputSlotForOperand[operandIndex] = &outputSlot; |
| 1027 | } |
| 1028 | |
| 1029 | return true; |
| 1030 | } |
| 1031 | |
| 1032 | } // namespace armnn_driver |