Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1 | // |
Kevin May | 4a54daa | 2023-07-04 16:10:55 +0100 | [diff] [blame] | 2 | // Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved. |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include "CanonicalUtils.hpp" |
| 9 | |
| 10 | #include <armnn/ArmNN.hpp> |
| 11 | #include <armnn/BackendHelper.hpp> |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 12 | #include <armnn/Exceptions.hpp> |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 13 | #include <armnn/utility/Assert.hpp> |
| 14 | #include <armnn/utility/IgnoreUnused.hpp> |
| 15 | #include <armnn/utility/NumericCast.hpp> |
| 16 | |
| 17 | #include <armnnUtils/DataLayoutIndexed.hpp> |
| 18 | #include <armnnUtils/Transpose.hpp> |
| 19 | |
| 20 | #include <ActivationFunctor.h> |
| 21 | #include <CpuExecutor.h> |
| 22 | #include <OperationsUtils.h> |
| 23 | |
| 24 | #include <armnnUtils/FloatingPointComparison.hpp> |
| 25 | |
| 26 | #include <log/log.h> |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 27 | #include <sstream> |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 28 | #include <vector> |
| 29 | |
| 30 | inline const android::nn::Model::Subgraph& getMainModel(const android::nn::Model& model) { return model.main; } |
| 31 | |
| 32 | namespace armnn_driver |
| 33 | { |
| 34 | |
| 35 | /// |
| 36 | /// Helper classes |
| 37 | /// |
| 38 | |
| 39 | #include <nnapi/OperandTypes.h> |
| 40 | #include <nnapi/Result.h> |
| 41 | #include <nnapi/TypeUtils.h> |
| 42 | #include <nnapi/Types.h> |
| 43 | #include <nnapi/Validation.h> |
| 44 | |
| 45 | using Model = ::android::nn::Model; |
| 46 | using Operand = ::android::nn::Operand; |
| 47 | using OperandLifeTime = ::android::nn::Operand::LifeTime; |
| 48 | using OperandType = ::android::nn::OperandType; |
| 49 | using Operation = ::android::nn::Operation; |
| 50 | using OperationType = ::android::nn::OperationType; |
| 51 | using ErrorStatus = ::android::nn::ErrorStatus; |
| 52 | |
| 53 | struct ConversionData |
| 54 | { |
| 55 | ConversionData(const std::vector<armnn::BackendId>& backends) |
| 56 | : m_Backends(backends) |
| 57 | , m_Network(nullptr, nullptr) |
| 58 | , m_DynamicInputsEncountered(false) |
| 59 | {} |
| 60 | |
| 61 | const std::vector<armnn::BackendId> m_Backends; |
| 62 | armnn::INetworkPtr m_Network; |
| 63 | std::vector<armnn::IOutputSlot*> m_OutputSlotForOperand; |
| 64 | std::vector<::android::nn::RunTimePoolInfo> m_MemPools; |
| 65 | bool m_DynamicInputsEncountered; |
| 66 | }; |
| 67 | |
| 68 | class LayerInputHandle |
| 69 | { |
| 70 | public: |
| 71 | LayerInputHandle(); |
| 72 | LayerInputHandle(bool valid, armnn::IOutputSlot* outputSlot, armnn::TensorInfo tensorInfo); |
| 73 | |
| 74 | bool IsValid() const; |
| 75 | |
| 76 | void Connect(armnn::IInputSlot& inputSlot); |
| 77 | |
| 78 | void Disconnect(armnn::IInputSlot& inputSlot); |
| 79 | |
| 80 | const armnn::TensorInfo& GetTensorInfo() const; |
| 81 | |
| 82 | void SanitizeQuantizationScale(LayerInputHandle& weight, LayerInputHandle& input); |
| 83 | |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 84 | armnn::IOutputSlot* GetOutputSlot() const; |
| 85 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 86 | private: |
| 87 | armnn::IOutputSlot* m_OutputSlot; |
| 88 | bool m_Valid; |
| 89 | armnn::TensorInfo m_TensorInfo; |
| 90 | }; |
| 91 | |
| 92 | class ConstTensorPin |
| 93 | { |
| 94 | public: |
| 95 | // Creates an invalid tensor pin (can be used to signal errors) |
| 96 | // The optional flag can be set to indicate the tensor values were missing, but it was otherwise valid |
| 97 | ConstTensorPin(bool optional = false); |
| 98 | |
| 99 | // @param tensorInfo TensorInfo associated with the tensor. |
| 100 | // @param valueStart Start address of tensor data. Belongs to one of the memory pools associated with |
| 101 | // the model being converted. |
| 102 | // @param numBytes Number of bytes for the tensor data. |
| 103 | ConstTensorPin(armnn::TensorInfo& tensorInfo, const void* valueStart, uint32_t numBytes, |
| 104 | const armnn::PermutationVector& mappings); |
| 105 | |
| 106 | ConstTensorPin(const ConstTensorPin& other) = delete; |
| 107 | ConstTensorPin(ConstTensorPin&& other) = default; |
| 108 | |
| 109 | bool IsValid() const; |
| 110 | bool IsOptional() const; |
| 111 | |
| 112 | const armnn::ConstTensor& GetConstTensor() const; |
| 113 | const armnn::ConstTensor* GetConstTensorPtr() const; |
| 114 | |
| 115 | private: |
| 116 | armnn::ConstTensor m_ConstTensor; |
| 117 | |
| 118 | // Owned memory for swizzled tensor data, only required if the tensor needed |
| 119 | // swizzling. Otherwise, @ref m_ConstTensor will reference memory from one of |
| 120 | // the pools associated with the model being converted. |
| 121 | std::vector<uint8_t> m_SwizzledTensorData; |
| 122 | |
| 123 | // optional flag to indicate that an invalid tensor pin is not an error, but the optional values were not given |
| 124 | bool m_Optional; |
| 125 | }; |
| 126 | |
| 127 | enum class ConversionResult |
| 128 | { |
| 129 | Success, |
| 130 | ErrorMappingPools, |
| 131 | UnsupportedFeature |
| 132 | }; |
| 133 | |
| 134 | } // namespace armnn_driver |
| 135 | |
| 136 | /// |
| 137 | /// Utility functions |
| 138 | /// |
| 139 | |
| 140 | namespace |
| 141 | { |
| 142 | using namespace armnn_driver; |
| 143 | |
| 144 | // Convenience function to log the reason for failing to convert a model. |
| 145 | // @return Always returns false (so that it can be used by callers as a quick way to signal an error and return) |
| 146 | template<class... Args> |
| 147 | static bool Fail(const char* formatStr, Args&&... args) |
| 148 | { |
| 149 | ALOGD(formatStr, std::forward<Args>(args)...); |
| 150 | return false; |
| 151 | } |
| 152 | |
| 153 | // Convenience macro to call an Is*Supported function and log caller name together with reason for lack of support. |
| 154 | // Called as: FORWARD_LAYER_SUPPORT_FUNC(__func__, Is*Supported, backends, a, b, c, d, e) |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 155 | #define FORWARD_LAYER_SUPPORT_FUNC(funcName, func, backends, supported, setBackend, ...) \ |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 156 | try \ |
| 157 | { \ |
| 158 | for (auto&& backendId : backends) \ |
| 159 | { \ |
| 160 | auto layerSupportObject = armnn::GetILayerSupportByBackendId(backendId); \ |
| 161 | if (layerSupportObject.IsBackendRegistered()) \ |
| 162 | { \ |
| 163 | std::string reasonIfUnsupported; \ |
| 164 | supported = \ |
| 165 | layerSupportObject.func(__VA_ARGS__, armnn::Optional<std::string&>(reasonIfUnsupported)); \ |
| 166 | if (supported) \ |
| 167 | { \ |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 168 | setBackend = backendId; \ |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 169 | break; \ |
| 170 | } \ |
| 171 | else \ |
| 172 | { \ |
| 173 | if (reasonIfUnsupported.size() > 0) \ |
| 174 | { \ |
| 175 | VLOG(DRIVER) << funcName << ": not supported by armnn: " << reasonIfUnsupported.c_str(); \ |
| 176 | } \ |
| 177 | else \ |
| 178 | { \ |
| 179 | VLOG(DRIVER) << funcName << ": not supported by armnn"; \ |
| 180 | } \ |
| 181 | } \ |
| 182 | } \ |
| 183 | else \ |
| 184 | { \ |
| 185 | VLOG(DRIVER) << funcName << ": backend not registered: " << backendId.Get().c_str(); \ |
| 186 | } \ |
| 187 | } \ |
| 188 | if (!supported) \ |
| 189 | { \ |
| 190 | VLOG(DRIVER) << funcName << ": not supported by any specified backend"; \ |
| 191 | } \ |
| 192 | } \ |
| 193 | catch (const armnn::InvalidArgumentException &e) \ |
| 194 | { \ |
| 195 | throw armnn::InvalidArgumentException(e, "Failed to check layer support", CHECK_LOCATION()); \ |
| 196 | } |
| 197 | |
| 198 | inline armnn::TensorShape GetTensorShapeForOperand(const Operand& operand) |
| 199 | { |
| 200 | return armnn::TensorShape(operand.dimensions.size(), operand.dimensions.data()); |
| 201 | } |
| 202 | |
| 203 | // Support within the 1.3 driver for specific tensor data types |
| 204 | inline bool IsOperandTypeSupportedForTensors(OperandType type) |
| 205 | { |
| 206 | return type == OperandType::BOOL || |
| 207 | type == OperandType::TENSOR_BOOL8 || |
| 208 | type == OperandType::TENSOR_FLOAT16 || |
| 209 | type == OperandType::TENSOR_FLOAT32 || |
| 210 | type == OperandType::TENSOR_QUANT8_ASYMM || |
| 211 | type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED || |
| 212 | type == OperandType::TENSOR_QUANT8_SYMM || |
| 213 | type == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL || |
| 214 | type == OperandType::TENSOR_QUANT16_SYMM || |
| 215 | type == OperandType::TENSOR_INT32; |
| 216 | } |
| 217 | |
| 218 | inline bool IsBool(Operand operand) |
| 219 | { |
| 220 | return operand.type == OperandType::BOOL; |
| 221 | } |
| 222 | |
| 223 | inline bool Is12OrLaterOperand(Operand) |
| 224 | { |
| 225 | return true; |
| 226 | } |
| 227 | |
| 228 | |
| 229 | template<typename LayerHandleType> |
| 230 | armnn::IConnectableLayer& AddReshapeLayer(armnn::INetwork& network, |
| 231 | LayerHandleType& inputLayer, |
| 232 | armnn::TensorInfo reshapeInfo) |
| 233 | { |
| 234 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 235 | reshapeDescriptor.m_TargetShape = reshapeInfo.GetShape(); |
| 236 | |
| 237 | armnn::IConnectableLayer* reshapeLayer = network.AddReshapeLayer(reshapeDescriptor); |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 238 | if (reshapeLayer == nullptr) |
| 239 | { |
| 240 | throw armnn::Exception("failed to add reshape layer to network"); |
| 241 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 242 | |
| 243 | // Attach the input layer to the reshape layer |
| 244 | inputLayer.Connect(reshapeLayer->GetInputSlot(0)); |
| 245 | reshapeLayer->GetOutputSlot(0).SetTensorInfo(reshapeInfo); |
| 246 | |
| 247 | return *reshapeLayer; |
| 248 | } |
| 249 | |
| 250 | |
| 251 | armnn::TensorShape FlattenFullyConnectedInput(const armnn::TensorShape& inputShape, |
| 252 | const armnn::TensorShape& weightsShape) |
| 253 | { |
| 254 | if (inputShape.GetNumDimensions() > 2U) |
| 255 | { |
| 256 | unsigned int totalInputElements = inputShape.GetNumElements(); |
| 257 | unsigned int inputSize = weightsShape[1]; |
| 258 | |
| 259 | unsigned int batchSize = totalInputElements / inputSize; |
| 260 | |
| 261 | if(totalInputElements % batchSize != 0) |
| 262 | { |
| 263 | throw std::runtime_error("Failed to deduce tensor shape"); |
| 264 | } |
| 265 | |
| 266 | return armnn::TensorShape({batchSize, inputSize}); |
| 267 | } |
| 268 | else |
| 269 | { |
| 270 | return inputShape; |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | inline bool VerifyFullyConnectedShapes(const armnn::TensorShape& inputShape, |
| 275 | const armnn::TensorShape& weightsShape, |
| 276 | const armnn::TensorShape& outputShape, |
| 277 | bool transposeWeightMatrix) |
| 278 | { |
| 279 | unsigned int dimIdx = transposeWeightMatrix ? 0 : 1; |
| 280 | return (inputShape[0] == outputShape[0] && weightsShape[dimIdx] == outputShape[1]); |
| 281 | } |
| 282 | |
| 283 | bool BroadcastTensor(LayerInputHandle& input0, |
| 284 | LayerInputHandle& input1, |
| 285 | armnn::IConnectableLayer* startLayer, |
| 286 | ConversionData& data) |
| 287 | { |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 288 | if (startLayer == nullptr) |
| 289 | { |
| 290 | throw armnn::InvalidArgumentException("BroadcastTensor: startLayer pointer handed in is null"); |
| 291 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 292 | |
| 293 | const armnn::TensorInfo& inputInfo0 = input0.GetTensorInfo(); |
| 294 | const armnn::TensorInfo& inputInfo1 = input1.GetTensorInfo(); |
| 295 | |
| 296 | unsigned int inputDimensions0 = inputInfo0.GetNumDimensions(); |
| 297 | unsigned int inputDimensions1 = inputInfo1.GetNumDimensions(); |
| 298 | |
| 299 | if (inputDimensions0 == inputDimensions1) |
| 300 | { |
| 301 | // The inputs have the same number of dimensions, simply connect them to the given layer as they are |
| 302 | input0.Connect(startLayer->GetInputSlot(0)); |
| 303 | input1.Connect(startLayer->GetInputSlot(1)); |
| 304 | |
| 305 | return true; |
| 306 | } |
| 307 | |
| 308 | // Since the number of dimensions do not match then we need to add degenerate dimensions |
| 309 | // to the "smaller" tensor using a reshape, while keeping the order of the inputs. |
| 310 | |
| 311 | unsigned int maxInputDimensions = std::max(inputDimensions0, inputDimensions1); |
| 312 | unsigned int sizeDifference = std::abs(armnn::numeric_cast<int>(inputDimensions0) - |
| 313 | armnn::numeric_cast<int>(inputDimensions1)); |
| 314 | |
| 315 | bool input0IsSmaller = inputDimensions0 < inputDimensions1; |
| 316 | LayerInputHandle& smallInputHandle = input0IsSmaller ? input0 : input1; |
| 317 | const armnn::TensorInfo& smallInfo = smallInputHandle.GetTensorInfo(); |
| 318 | |
| 319 | const armnn::TensorShape& smallShape = smallInfo.GetShape(); |
| 320 | std::vector<unsigned int> reshapedDimensions(maxInputDimensions, 1); |
| 321 | for (unsigned int i = sizeDifference; i < maxInputDimensions; i++) |
| 322 | { |
| 323 | reshapedDimensions[i] = smallShape[i - sizeDifference]; |
| 324 | } |
| 325 | |
| 326 | armnn::TensorInfo reshapedInfo = smallInfo; |
| 327 | reshapedInfo.SetShape(armnn::TensorShape{ armnn::numeric_cast<unsigned int>(reshapedDimensions.size()), |
| 328 | reshapedDimensions.data() }); |
| 329 | |
| 330 | // RehsapeDescriptor that is ignored in the IsReshapeSupported function |
| 331 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 332 | |
| 333 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 334 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 335 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 336 | IsReshapeSupported, |
| 337 | data.m_Backends, |
| 338 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 339 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 340 | smallInfo, |
| 341 | reshapedInfo, |
| 342 | reshapeDescriptor); |
| 343 | if (!isSupported) |
| 344 | { |
| 345 | return false; |
| 346 | } |
| 347 | |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 348 | if (data.m_Network == nullptr) |
| 349 | { |
| 350 | throw armnn::InvalidArgumentException( |
| 351 | "BroadcastTensor: the conversion data handed in has a null network pointer"); |
| 352 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 353 | armnn::IConnectableLayer& reshapeLayer = AddReshapeLayer(*data.m_Network, smallInputHandle, reshapedInfo); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 354 | reshapeLayer.SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 355 | |
| 356 | if (input0IsSmaller) |
| 357 | { |
| 358 | // Input0 is the "smaller" tensor, connect the reshape layer as follows: |
| 359 | // |
| 360 | // Input0 Input1 |
| 361 | // | | |
| 362 | // Reshape | |
| 363 | // \ / |
| 364 | // StartLayer |
| 365 | |
| 366 | reshapeLayer.GetOutputSlot(0).Connect(startLayer->GetInputSlot(0)); |
| 367 | input1.Connect(startLayer->GetInputSlot(1)); |
| 368 | } |
| 369 | else |
| 370 | { |
| 371 | // Input1 is the "smaller" tensor, connect the reshape layer as follows: |
| 372 | // |
| 373 | // Input0 Input1 |
| 374 | // | | |
| 375 | // | Reshape |
| 376 | // \ / |
| 377 | // StartLayer |
| 378 | |
| 379 | input0.Connect(startLayer->GetInputSlot(0)); |
| 380 | reshapeLayer.GetOutputSlot(0).Connect(startLayer->GetInputSlot(1)); |
| 381 | } |
| 382 | |
| 383 | return true; |
| 384 | } |
| 385 | |
| 386 | void CalcPadding(uint32_t input, |
| 387 | uint32_t kernel, |
| 388 | uint32_t stride, |
| 389 | uint32_t& outPadHead, |
| 390 | uint32_t& outPadTail, |
| 391 | PaddingScheme scheme) |
| 392 | { |
| 393 | int32_t padHead; |
| 394 | int32_t padTail; |
| 395 | calculateExplicitPadding(input, stride, kernel, scheme, &padHead, &padTail); |
| 396 | outPadHead = armnn::numeric_cast<uint32_t>(padHead); |
| 397 | outPadTail = armnn::numeric_cast<uint32_t>(padTail); |
| 398 | } |
| 399 | |
| 400 | void CalcPadding(uint32_t input, uint32_t kernel, uint32_t stride, uint32_t dilation, uint32_t& outPadHead, |
| 401 | uint32_t& outPadTail, ::android::nn::PaddingScheme scheme) |
| 402 | { |
| 403 | int32_t padHead; |
| 404 | int32_t padTail; |
| 405 | calculateExplicitPadding(input, stride, dilation, kernel, scheme, &padHead, &padTail); |
| 406 | outPadHead = armnn::numeric_cast<uint32_t>(padHead); |
| 407 | outPadTail = armnn::numeric_cast<uint32_t>(padTail); |
| 408 | } |
| 409 | |
| 410 | inline void CalcPaddingTransposeConv(uint32_t output, uint32_t kernel, int32_t stride, int32_t& outPadHead, |
| 411 | int32_t& outPadTail, ::android::nn::PaddingScheme scheme) |
| 412 | { |
| 413 | calculateExplicitPaddingTransposeConv(output, stride, kernel, scheme, &outPadHead, &outPadTail); |
| 414 | } |
| 415 | |
| 416 | Shape GetOperandShape(const Operand& operand) |
| 417 | { |
| 418 | Shape shape; |
| 419 | shape.type = OperandType(operand.type); |
| 420 | shape.dimensions = operand.dimensions; |
| 421 | shape.scale = operand.scale; |
| 422 | shape.offset = operand.zeroPoint; |
| 423 | return shape; |
| 424 | } |
| 425 | |
| 426 | |
| 427 | // ArmNN requires the bias scale to be equal to the product of the weight and input scales, which is also |
| 428 | // what AndroidNN requires. However for some of the AndroidNN tests the values don't exactly match so |
| 429 | // we accept some tolerance. We don't want ArmNN itself to accept these inconsistencies as it is up to the |
| 430 | // user (us, in this case) to ensure they match. |
| 431 | void SanitizeBiasQuantizationScale(armnn::TensorInfo& biasInfo, |
| 432 | const armnn::TensorInfo& weightInfo, |
| 433 | const armnn::TensorInfo& inputInfo) |
| 434 | { |
| 435 | if (weightInfo.HasPerAxisQuantization()) |
| 436 | { |
| 437 | // NOTE: Bias scale is always set to 0 for per-axis quantization and |
| 438 | // it needs to be calculated: scale[i] = input_scale * weight_scale[i] |
| 439 | auto UpdateBiasScaleValue = [&inputInfo](float biasScale) -> float |
| 440 | { |
| 441 | return biasScale * inputInfo.GetQuantizationScale(); |
| 442 | }; |
| 443 | |
| 444 | std::vector<float> biasScales(weightInfo.GetQuantizationScales()); |
| 445 | std::transform(biasScales.begin(), biasScales.end(), biasScales.begin(), UpdateBiasScaleValue); |
| 446 | |
| 447 | biasInfo.SetQuantizationScales(biasScales); |
| 448 | // bias is expected to be a 1d tensor, set qdim=0 |
| 449 | biasInfo.SetQuantizationDim(0); |
| 450 | |
| 451 | VLOG(DRIVER) << "Bias quantization params have been updated for per-axis quantization"; |
| 452 | } |
| 453 | else |
| 454 | { |
| 455 | const float expectedBiasScale = weightInfo.GetQuantizationScale() * inputInfo.GetQuantizationScale(); |
| 456 | if (biasInfo.GetQuantizationScale() != expectedBiasScale) |
| 457 | { |
| 458 | if (armnnUtils::within_percentage_tolerance(biasInfo.GetQuantizationScale(), expectedBiasScale, 1.0f)) |
| 459 | { |
| 460 | VLOG(DRIVER) << "Bias quantization scale has been modified to match input * weights"; |
| 461 | biasInfo.SetQuantizationScale(expectedBiasScale); |
| 462 | } |
| 463 | } |
| 464 | } |
| 465 | } |
| 466 | |
| 467 | // 4D Tensor Permutations |
| 468 | const armnn::PermutationVector IdentityPermutation4D({ 0U, 1U, 2U, 3U }); |
| 469 | const armnn::PermutationVector IdentityPermutation3D({ 0U, 1U, 2U }); |
Kevin May | 4a54daa | 2023-07-04 16:10:55 +0100 | [diff] [blame] | 470 | const armnn::PermutationVector SwapDim2And3({ 0U, 1U, 3U, 2U }); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 471 | |
| 472 | // 3D Permutation Vectors |
| 473 | const armnn::PermutationVector RotateTensorLeft({ 1U, 2U, 0U }); |
| 474 | const armnn::PermutationVector RotateTensorRight({ 2U, 0U, 1U }); |
| 475 | |
| 476 | template<typename OSlot> |
| 477 | armnn::IConnectableLayer& AddTransposeLayer(armnn::INetwork& network, OSlot& input, |
| 478 | const armnn::PermutationVector& mappings) |
| 479 | { |
| 480 | // Add swizzle layer |
| 481 | armnn::IConnectableLayer* const layer = network.AddTransposeLayer(mappings); |
| 482 | |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 483 | if (layer == nullptr) |
| 484 | { |
| 485 | throw armnn::Exception("failed to add transpose layer to network"); |
| 486 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 487 | |
| 488 | // Connect input to swizzle layer |
| 489 | input.Connect(layer->GetInputSlot(0)); |
| 490 | |
| 491 | // Setup swizzled output |
| 492 | const armnn::TensorInfo outInfo = armnnUtils::TransposeTensorShape(input.GetTensorInfo(), mappings); |
| 493 | layer->GetOutputSlot(0).SetTensorInfo(outInfo); |
| 494 | |
| 495 | return *layer; |
| 496 | } |
| 497 | |
| 498 | bool ValidateConcatOutputShape(const std::vector<armnn::TensorShape> & inputShapes, |
| 499 | const armnn::TensorShape & outputShape, |
| 500 | uint32_t concatDim) |
| 501 | { |
| 502 | // Validate the output shape is correct given the input shapes (which have just been validated) |
| 503 | unsigned int numDimensions = inputShapes[0].GetNumDimensions(); |
| 504 | if (outputShape.GetNumDimensions() != numDimensions) |
| 505 | { |
| 506 | return Fail("%s: Output shape has wrong number of dimensions", __func__); |
| 507 | } |
| 508 | |
| 509 | unsigned int outputSizeAlongConcatenatedDimension = 0; |
| 510 | for (unsigned int i = 0; i < inputShapes.size(); i++) |
| 511 | { |
| 512 | outputSizeAlongConcatenatedDimension += inputShapes[i][concatDim]; |
| 513 | } |
| 514 | |
| 515 | for (unsigned int i = 0; i < numDimensions; ++i) |
| 516 | { |
| 517 | if (i == concatDim) |
| 518 | { |
| 519 | if (outputShape[i] != outputSizeAlongConcatenatedDimension) |
| 520 | { |
| 521 | return Fail( |
| 522 | "%s: Invalid output shape for dimension %d (%d != %d)", |
| 523 | __func__, |
| 524 | i, |
| 525 | outputShape[i], |
| 526 | outputSizeAlongConcatenatedDimension); |
| 527 | } |
| 528 | } |
| 529 | else |
| 530 | { |
| 531 | if (outputShape[i] != inputShapes[0][i]) |
| 532 | { |
| 533 | return Fail("%s: Invalid output shape", __func__); |
| 534 | } |
| 535 | } |
| 536 | } |
| 537 | |
| 538 | return true; |
| 539 | } |
| 540 | |
| 541 | inline bool RequiresReshape(armnn::TensorShape & inputShape) |
| 542 | { |
| 543 | return inputShape.GetNumDimensions() < 3; |
| 544 | } |
| 545 | |
| 546 | inline void SwizzleInputs(armnn::INetwork& network, |
| 547 | std::vector<LayerInputHandle>& inputs, |
| 548 | std::vector<armnn::TensorShape>& inputShapes, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 549 | const armnn::PermutationVector& mapping, |
| 550 | std::vector<armnn::BackendId>& setBackends) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 551 | { |
| 552 | if (!mapping.IsEqual(IdentityPermutation4D)) |
| 553 | { |
| 554 | size_t nInputs = inputs.size(); |
| 555 | for (size_t i=0; i<nInputs; ++i) |
| 556 | { |
| 557 | // add swizzle layer |
| 558 | armnn::IConnectableLayer& swizzleLayer = AddTransposeLayer(network, inputs[i], mapping); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 559 | swizzleLayer.SetBackendId(setBackends[i]); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 560 | auto& outputSlot = swizzleLayer.GetOutputSlot(0); |
| 561 | auto& outputInfo = outputSlot.GetTensorInfo(); |
| 562 | // replace inputs with the swizzled ones |
| 563 | inputs[i] = LayerInputHandle(true, &outputSlot, outputInfo); |
| 564 | inputShapes[i] = inputs[i].GetTensorInfo().GetShape(); |
| 565 | } |
| 566 | } |
| 567 | } |
| 568 | |
| 569 | bool TransposeInputTensors(ConversionData& data, |
| 570 | std::vector<LayerInputHandle>& inputs, |
| 571 | std::vector<armnn::TensorShape>& inputShapes, |
| 572 | const armnn::PermutationVector& mapping) |
| 573 | { |
| 574 | // If we have a IdentityPermutation4D or IdentityPermutation3D then we are not permuting |
| 575 | if (!mapping.IsEqual(IdentityPermutation4D) && !mapping.IsEqual(IdentityPermutation3D)) |
| 576 | { |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 577 | std::vector<armnn::BackendId> setBackendsVec; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 578 | armnn::TensorInfo outputTransposeInfo; |
| 579 | size_t nInputs = inputs.size(); |
| 580 | for (size_t i=0; i<nInputs; ++i) |
| 581 | { |
| 582 | // check permute layer |
| 583 | armnn::TransposeDescriptor transposeDesc; |
| 584 | transposeDesc.m_DimMappings = mapping; |
| 585 | outputTransposeInfo = armnnUtils::TransposeTensorShape(inputs[i].GetTensorInfo(), mapping); |
| 586 | |
| 587 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 588 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 589 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 590 | IsTransposeSupported, |
| 591 | data.m_Backends, |
| 592 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 593 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 594 | inputs[i].GetTensorInfo(), |
| 595 | outputTransposeInfo, |
| 596 | transposeDesc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 597 | setBackendsVec.push_back(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 598 | if (!isSupported) |
| 599 | { |
| 600 | return false; |
| 601 | } |
| 602 | |
| 603 | } |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 604 | SwizzleInputs(*data.m_Network, inputs, inputShapes, mapping, setBackendsVec); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 605 | } |
| 606 | return true; |
| 607 | } |
| 608 | |
| 609 | bool CreateConcatPermutationParameters(const unsigned int numberOfDimensions, |
| 610 | int32_t & concatDimension, |
| 611 | std::pair<armnn::PermutationVector, armnn::PermutationVector> & permutationPair) |
| 612 | { |
| 613 | bool needPermute = false; |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 614 | if (numberOfDimensions < 3) |
| 615 | { |
| 616 | throw armnn::InvalidArgumentException( |
| 617 | "CreateConcatPermutationParameters: numberOfDimensions handed in cannot be less than three"); |
| 618 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 619 | |
| 620 | // ArmNN uses Compute Library subtensors to perform concatenation |
| 621 | // This only works when concatenating along dimension 0, 1 or 3 for a 4-D tensor, |
| 622 | // or along dimension 0 or 2 for a 3-D tensor. |
| 623 | if (numberOfDimensions == 4 && concatDimension == 2) |
| 624 | { |
Kevin May | 4a54daa | 2023-07-04 16:10:55 +0100 | [diff] [blame] | 625 | concatDimension = 3; |
| 626 | permutationPair = std::make_pair(SwapDim2And3, SwapDim2And3); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 627 | needPermute = true; |
| 628 | } |
| 629 | else if (numberOfDimensions == 3 && concatDimension == 1) |
| 630 | { |
| 631 | concatDimension = 0; |
| 632 | permutationPair = std::make_pair(RotateTensorLeft, RotateTensorRight); |
| 633 | needPermute = true; |
| 634 | } |
| 635 | // If the tensor is 3-D and the concat dimension is 2 then we don't need to permute but we do need to change the |
| 636 | // permutation identity to only have 3 dimensions |
| 637 | else if (numberOfDimensions == 3 && concatDimension == 2) |
| 638 | { |
| 639 | permutationPair = std::make_pair(IdentityPermutation3D, IdentityPermutation3D); |
| 640 | } |
| 641 | return needPermute; |
| 642 | } |
| 643 | |
| 644 | } // anonymous namespace |
| 645 | |
| 646 | namespace armnn_driver |
| 647 | { |
| 648 | using namespace android::nn; |
| 649 | |
| 650 | //// Creates an ArmNN activation layer and connects it to the given layer, if the |
| 651 | //// passed in AndroidNN activation function requires so. |
| 652 | //// @return The end layer of the sequence of layers built for the given AndroidNN |
| 653 | //// activation function or nullptr if an error occurred (e.g. unsupported activation). |
| 654 | //// Note that the end layer matches the input layer if no activation is required |
| 655 | //// (the sequence of layers has length 1). |
| 656 | armnn::IConnectableLayer* ProcessActivation(const armnn::TensorInfo& tensorInfo, |
| 657 | ActivationFn activation, |
| 658 | armnn::IConnectableLayer* prevLayer, |
| 659 | ConversionData& data); |
| 660 | |
| 661 | |
| 662 | inline const Operand* GetInputOperand(const Operation& operation, |
| 663 | uint32_t inputIndex, |
| 664 | const Model& model, |
| 665 | bool failOnIndexOutOfBounds = true) |
| 666 | { |
| 667 | if (inputIndex >= operation.inputs.size()) |
| 668 | { |
| 669 | if (failOnIndexOutOfBounds) |
| 670 | { |
| 671 | Fail("%s: invalid input index: %i out of %i", __func__, inputIndex, operation.inputs.size()); |
| 672 | } |
| 673 | return nullptr; |
| 674 | } |
| 675 | |
| 676 | // Model should have been validated beforehand |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 677 | if (!(operation.inputs[inputIndex] < getMainModel(model).operands.size())) |
| 678 | { |
| 679 | std::ostringstream os; |
| 680 | os << "GetInputOperand: inputIndex [" << inputIndex << "]"; |
| 681 | os << " is too large. The number of main model operands is ["; |
| 682 | os << getMainModel(model).operands.size() << "]"; |
| 683 | throw armnn::InvalidArgumentException(os.str()); |
| 684 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 685 | return &getMainModel(model).operands[operation.inputs[inputIndex]]; |
| 686 | } |
| 687 | |
| 688 | inline const Operand* GetOutputOperand(const Operation& operation, |
| 689 | uint32_t outputIndex, |
| 690 | const Model& model) |
| 691 | { |
| 692 | if (outputIndex >= operation.outputs.size()) |
| 693 | { |
| 694 | Fail("%s: invalid output index: %i out of %i", __func__, outputIndex, operation.outputs.size()); |
| 695 | return nullptr; |
| 696 | } |
| 697 | |
| 698 | // Model should have been validated beforehand |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 699 | if (!(operation.outputs[outputIndex] < getMainModel(model).operands.size())) |
| 700 | { |
| 701 | std::ostringstream os; |
| 702 | os << "GetOutputOperand: outputIndex [" << outputIndex << "]"; |
| 703 | os << " is too large. The number of main model operands is ["; |
| 704 | os << getMainModel(model).operands.size() << "]"; |
| 705 | throw armnn::InvalidArgumentException(os.str()); |
| 706 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 707 | |
| 708 | return &getMainModel(model).operands[operation.outputs[outputIndex]]; |
| 709 | } |
| 710 | |
| 711 | const void* GetOperandValueReadOnlyAddress(const Operand& operand, |
| 712 | const Model& model, |
| 713 | const ConversionData& data, |
| 714 | bool optional = false); |
| 715 | |
| 716 | inline bool GetOperandType(const Operation& operation, |
| 717 | uint32_t inputIndex, |
| 718 | const Model& model, |
| 719 | OperandType& type) |
| 720 | { |
| 721 | const Operand* operand = GetInputOperand(operation, inputIndex, model); |
| 722 | if (!operand) |
| 723 | { |
| 724 | return Fail("%s: invalid input operand at index %i", __func__, inputIndex); |
| 725 | } |
| 726 | |
| 727 | type = operand->type; |
| 728 | return true; |
| 729 | } |
| 730 | |
| 731 | inline bool IsOperandConstant(const Operand& operand) |
| 732 | { |
| 733 | OperandLifeTime lifetime = operand.lifetime; |
| 734 | |
| 735 | return lifetime == OperandLifeTime::CONSTANT_COPY || |
| 736 | lifetime == OperandLifeTime::CONSTANT_REFERENCE || |
| 737 | lifetime == OperandLifeTime::POINTER || |
| 738 | lifetime == OperandLifeTime::NO_VALUE; |
| 739 | } |
| 740 | |
Kevin May | 7fbf810 | 2023-08-23 10:07:26 +0100 | [diff] [blame] | 741 | bool IsWeightsValid(const Operation& operation, uint32_t inputIndex, const Model& model, const bool IsOptional); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 742 | |
| 743 | ConstTensorPin ConvertOperandToConstTensorPin(const Operand& operand, |
| 744 | const Model& model, |
| 745 | const ConversionData& data, |
| 746 | const armnn::PermutationVector& dimensionMappings = g_DontPermute, |
| 747 | const armnn::TensorShape* overrideTensorShape = nullptr, |
Sadik Armagan | 1e276f3 | 2022-07-19 12:37:20 +0100 | [diff] [blame] | 748 | bool optional = false, |
| 749 | const armnn::DataType* overrideDataType = nullptr); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 750 | |
| 751 | inline ConstTensorPin ConvertOperationInputToConstTensorPin( |
| 752 | const Operation& operation, |
| 753 | uint32_t inputIndex, |
| 754 | const Model& model, |
| 755 | const ConversionData& data, |
| 756 | const armnn::PermutationVector& dimensionMappings = g_DontPermute, |
| 757 | const armnn::TensorShape* overrideTensorShape = nullptr, |
| 758 | bool optional = false) |
| 759 | { |
| 760 | const Operand* operand = GetInputOperand(operation, inputIndex, model); |
| 761 | if (!operand) |
| 762 | { |
| 763 | Fail("%s: failed to get input operand: index=%u", __func__, inputIndex); |
| 764 | return ConstTensorPin(); |
| 765 | } |
| 766 | return ConvertOperandToConstTensorPin(*operand, |
| 767 | model, |
| 768 | data, |
| 769 | dimensionMappings, |
| 770 | overrideTensorShape, |
| 771 | optional); |
| 772 | } |
| 773 | |
| 774 | template <typename OutputType> |
| 775 | bool GetInputScalar(const Operation& operation, |
| 776 | uint32_t inputIndex, |
| 777 | OperandType type, |
| 778 | OutputType& outValue, |
| 779 | const Model& model, |
| 780 | const ConversionData& data, |
| 781 | bool optional = false) |
| 782 | { |
| 783 | const Operand* operand = GetInputOperand(operation, inputIndex, model); |
| 784 | if (!optional && !operand) |
| 785 | { |
| 786 | return Fail("%s: invalid input operand at index %i", __func__, inputIndex); |
| 787 | } |
| 788 | |
| 789 | if (!optional && operand->type != type) |
| 790 | { |
| 791 | VLOG(DRIVER) << __func__ << ": unexpected operand type: " << operand->type << " should be: " << type; |
| 792 | return false; |
| 793 | } |
| 794 | |
| 795 | if (!optional && operand->location.length != sizeof(OutputType)) |
| 796 | { |
| 797 | return Fail("%s: incorrect operand location length: %i (should be %i)", |
| 798 | __func__, operand->location.length, sizeof(OutputType)); |
| 799 | } |
| 800 | |
| 801 | const void* valueAddress = GetOperandValueReadOnlyAddress(*operand, model, data); |
| 802 | if (!optional && !valueAddress) |
| 803 | { |
| 804 | return Fail("%s: failed to get address for operand", __func__); |
| 805 | } |
| 806 | |
| 807 | if(!optional) |
| 808 | { |
| 809 | outValue = *(static_cast<const OutputType*>(valueAddress)); |
| 810 | } |
| 811 | |
| 812 | return true; |
| 813 | } |
| 814 | |
| 815 | inline bool GetInputInt32(const Operation& operation, |
| 816 | uint32_t inputIndex, |
| 817 | int32_t& outValue, |
| 818 | const Model& model, |
| 819 | const ConversionData& data) |
| 820 | { |
| 821 | return GetInputScalar(operation, inputIndex, OperandType::INT32, outValue, model, data); |
| 822 | } |
| 823 | |
| 824 | inline bool GetInputFloat32(const Operation& operation, |
| 825 | uint32_t inputIndex, |
| 826 | float& outValue, |
| 827 | const Model& model, |
| 828 | const ConversionData& data) |
| 829 | { |
| 830 | return GetInputScalar(operation, inputIndex, OperandType::FLOAT32, outValue, model, data); |
| 831 | } |
| 832 | |
| 833 | inline bool GetInputActivationFunctionImpl(const Operation& operation, |
| 834 | uint32_t inputIndex, |
| 835 | OperandType type, |
| 836 | ActivationFn& outActivationFunction, |
| 837 | const Model& model, |
| 838 | const ConversionData& data) |
| 839 | { |
| 840 | if (type != OperandType::INT32 && type != OperandType::TENSOR_INT32) |
| 841 | { |
| 842 | VLOG(DRIVER) << __func__ << ": unexpected operand type: " << type |
| 843 | << " should be OperandType::INT32 or OperandType::TENSOR_INT32"; |
| 844 | return false; |
| 845 | } |
| 846 | |
| 847 | int32_t activationFunctionAsInt; |
| 848 | if (!GetInputScalar(operation, inputIndex, type, activationFunctionAsInt, model, data)) |
| 849 | { |
| 850 | return Fail("%s: failed to get activation input value", __func__); |
| 851 | } |
| 852 | outActivationFunction = static_cast<ActivationFn>(activationFunctionAsInt); |
| 853 | return true; |
| 854 | } |
| 855 | |
| 856 | inline bool GetInputActivationFunction(const Operation& operation, |
| 857 | uint32_t inputIndex, |
| 858 | ActivationFn& outActivationFunction, |
| 859 | const Model& model, |
| 860 | const ConversionData& data) |
| 861 | { |
| 862 | return GetInputActivationFunctionImpl(operation, |
| 863 | inputIndex, |
| 864 | OperandType::INT32, |
| 865 | outActivationFunction, |
| 866 | model, |
| 867 | data); |
| 868 | } |
| 869 | |
| 870 | inline bool GetInputActivationFunctionFromTensor(const Operation& operation, |
| 871 | uint32_t inputIndex, |
| 872 | ActivationFn& outActivationFunction, |
| 873 | const Model& model, |
| 874 | const ConversionData& data) |
| 875 | { |
| 876 | // This only accepts a 1-D tensor of size 1 |
| 877 | return GetInputActivationFunctionImpl(operation, |
| 878 | inputIndex, |
| 879 | OperandType::INT32, |
| 880 | outActivationFunction, |
| 881 | model, |
| 882 | data); |
| 883 | } |
| 884 | |
| 885 | |
| 886 | inline bool GetOptionalInputActivation(const Operation& operation, |
| 887 | uint32_t inputIndex, |
| 888 | ActivationFn& activationFunction, |
| 889 | const Model& model, |
| 890 | const ConversionData& data) |
| 891 | { |
| 892 | if (operation.inputs.size() <= inputIndex) |
| 893 | { |
| 894 | activationFunction = ActivationFn::kActivationNone; |
| 895 | } |
| 896 | else |
| 897 | { |
| 898 | if (!GetInputActivationFunction(operation, inputIndex, activationFunction, model, data)) |
| 899 | { |
| 900 | return Fail("%s: Operation has invalid inputs", __func__); |
| 901 | } |
| 902 | } |
| 903 | return true; |
| 904 | } |
| 905 | |
| 906 | template<typename ConvolutionDescriptor> |
| 907 | bool GetOptionalConvolutionDilationParams(const Operation& operation, |
| 908 | uint32_t dilationXIndex, |
| 909 | ConvolutionDescriptor& descriptor, |
| 910 | const Model& model, |
| 911 | const ConversionData& data) |
| 912 | { |
| 913 | bool success = true; |
| 914 | if (operation.inputs.size() >= dilationXIndex + 2) |
| 915 | { |
| 916 | success &= GetInputScalar(operation, |
| 917 | dilationXIndex, |
| 918 | OperandType::INT32, |
| 919 | descriptor.m_DilationX, |
| 920 | model, |
| 921 | data); |
| 922 | success &= GetInputScalar(operation, |
| 923 | dilationXIndex + 1, |
| 924 | OperandType::INT32, |
| 925 | descriptor.m_DilationY, |
| 926 | model, |
| 927 | data); |
| 928 | } |
| 929 | |
| 930 | return success; |
| 931 | } |
| 932 | |
| 933 | inline bool GetOptionalBool(const Operation& operation, |
| 934 | uint32_t inputIndex, |
| 935 | const Model& model, |
| 936 | const ConversionData& data) |
| 937 | { |
| 938 | const Operand* operand = GetInputOperand(operation, inputIndex, model); |
| 939 | if (!operand) |
| 940 | { |
| 941 | return false; |
| 942 | } |
| 943 | |
| 944 | if (!IsBool(*operand)) |
| 945 | { |
| 946 | return false; |
| 947 | } |
| 948 | |
| 949 | const void* valueAddress = GetOperandValueReadOnlyAddress(*operand, model, data); |
| 950 | if (!valueAddress) |
| 951 | { |
| 952 | return false; |
| 953 | } |
| 954 | |
| 955 | return *(static_cast<const bool*>(valueAddress)); |
| 956 | } |
| 957 | |
| 958 | bool GetTensorInt32Values(const Operand& operand, |
| 959 | std::vector<int32_t>& outValues, |
| 960 | const Model& model, |
| 961 | const ConversionData& data); |
| 962 | |
| 963 | bool GetInputPaddingScheme(const Operation& operation, |
| 964 | uint32_t inputIndex, |
| 965 | PaddingScheme& outPaddingScheme, |
| 966 | const Model& model, |
| 967 | const ConversionData& data); |
| 968 | |
| 969 | LayerInputHandle ConvertToLayerInputHandle(const Operation& operation, |
| 970 | uint32_t inputIndex, |
| 971 | const Model& model, |
| 972 | ConversionData& data, |
Sadik Armagan | 1e276f3 | 2022-07-19 12:37:20 +0100 | [diff] [blame] | 973 | const armnn::PermutationVector& dimensionMappings = g_DontPermute, |
| 974 | const LayerInputHandle* inputHandle = nullptr); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 975 | |
| 976 | bool SetupAndTrackLayerOutputSlot(const Operation& operation, |
| 977 | uint32_t operationOutputIndex, |
| 978 | armnn::IConnectableLayer& layer, |
| 979 | uint32_t layerOutputIndex, |
| 980 | const Model& model, |
| 981 | ConversionData& data, |
| 982 | const armnn::TensorInfo* overrideOutputInfo = nullptr, |
| 983 | const std::function <void (const armnn::TensorInfo&, bool&)>& validateFunc = nullptr, |
| 984 | const ActivationFn& activationFunction = ActivationFn::kActivationNone, |
| 985 | bool inferOutputShapes = false); |
| 986 | |
| 987 | armnn::DataLayout OptionalDataLayout(const Operation& operation, |
| 988 | uint32_t inputIndex, |
| 989 | const Model& model, |
| 990 | ConversionData& data); |
| 991 | |
| 992 | inline bool SetupAndTrackLayerOutputSlot( |
| 993 | const Operation& operation, |
| 994 | uint32_t outputIndex, |
| 995 | armnn::IConnectableLayer& layer, |
| 996 | const Model& model, |
| 997 | ConversionData& data, |
| 998 | const armnn::TensorInfo* overrideOutputInfo = nullptr, |
| 999 | const std::function <void (const armnn::TensorInfo&, bool&)>& validateFunc = nullptr, |
| 1000 | const ActivationFn& activationFunction = ActivationFn::kActivationNone) |
| 1001 | { |
| 1002 | return SetupAndTrackLayerOutputSlot(operation, |
| 1003 | outputIndex, |
| 1004 | layer, |
| 1005 | outputIndex, |
| 1006 | model, |
| 1007 | data, |
| 1008 | overrideOutputInfo, |
| 1009 | validateFunc, |
| 1010 | activationFunction); |
| 1011 | } |
| 1012 | |
| 1013 | bool ConvertToActivation(const Operation& operation, |
| 1014 | const char* operationName, |
| 1015 | const armnn::ActivationDescriptor& activationDesc, |
| 1016 | const Model& model, |
| 1017 | ConversionData& data); |
| 1018 | |
| 1019 | bool ConvertPaddings(const Operation& operation, |
| 1020 | const Model& model, |
| 1021 | ConversionData& data, |
| 1022 | unsigned int rank, |
| 1023 | armnn::PadDescriptor& padDescriptor); |
| 1024 | bool ConvertReduce(const Operation& operation, |
| 1025 | const Model& model, |
| 1026 | ConversionData& data, |
| 1027 | armnn::ReduceOperation reduceOperation); |
| 1028 | |
| 1029 | bool ConvertPooling2d(const Operation& operation, |
| 1030 | const char* operationName, |
| 1031 | armnn::PoolingAlgorithm poolType, |
| 1032 | const Model& model, |
| 1033 | ConversionData& data); |
| 1034 | |
| 1035 | inline bool IsQSymm8(const Operand& operand) |
| 1036 | { |
| 1037 | return operand.type == OperandType::TENSOR_QUANT8_SYMM; |
| 1038 | } |
| 1039 | |
| 1040 | enum class DequantizeStatus |
| 1041 | { |
| 1042 | SUCCESS, |
| 1043 | NOT_REQUIRED, |
| 1044 | INVALID_OPERAND |
| 1045 | }; |
| 1046 | |
| 1047 | using DequantizeResult = std::tuple<std::unique_ptr<float[]>, size_t, armnn::TensorInfo, DequantizeStatus>; |
| 1048 | |
| 1049 | DequantizeResult DequantizeIfRequired(size_t operand_index, |
| 1050 | const Operation& operation, |
| 1051 | const Model& model, |
| 1052 | const ConversionData& data); |
| 1053 | |
| 1054 | ConstTensorPin DequantizeAndMakeConstTensorPin(const Operation& operation, |
| 1055 | const Model& model, |
| 1056 | const ConversionData& data, |
| 1057 | size_t operandIndex, |
| 1058 | bool optional = false); |
| 1059 | |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1060 | bool IsConnectedToDequantize(armnn::IOutputSlot* ioutputSlot); |
| 1061 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1062 | } // namespace armnn_driver |